Open Collections

UBC Theses and Dissertations

UBC Theses Logo

UBC Theses and Dissertations

Evaluating how neighbourhood housing diversity relates with residential location choice, residential… Machler, Leonard 2015

Your browser doesn't seem to have a PDF viewer, please download the PDF to view this item.

Item Metadata

Download

Media
24-ubc_2016_february_machler_leonard.pdf [ 7.11MB ]
Metadata
JSON: 24-1.0223165.json
JSON-LD: 24-1.0223165-ld.json
RDF/XML (Pretty): 24-1.0223165-rdf.xml
RDF/JSON: 24-1.0223165-rdf.json
Turtle: 24-1.0223165-turtle.txt
N-Triples: 24-1.0223165-rdf-ntriples.txt
Original Record: 24-1.0223165-source.json
Full Text
24-1.0223165-fulltext.txt
Citation
24-1.0223165.ris

Full Text

  EVALUATING HOW NEIGHBOURHOOD HOUSING DIVERSITY RELATES WITH RESIDENTIAL LOCATION CHOICE, RESIDENTIAL SATISFACTION AND HEALTH   by  Leonard Machler  BSc, The University of Waterloo, 2005 M.A., Arizona State University, 2010   A THESIS SUBMITTED IN PARTIAL FULFILLMENT OF THE REQUIREMENTS FOR THE DEGREE OF  DOCTOR OF PHILOSOPHY in THE FACULTY OF GRADUATE AND POSTDOCTORAL STUDIES (Planning)  THE UNIVERSITY OF BRITISH COLUMBIA (Vancouver)   December 2015 © Leonard Machler, 2015 ii  Abstract Urban planners have long advocated strategies that enable a broad spectrum of the population to live in their preferred communities.  In particular, planning researchers emphasize the importance of enabling households with preferences for Smart Growth communities to match.  Living in Smart Growth communities – characterized by higher densities, more mixed land uses, and better access to transport alternatives to the car - has been empirically linked with improved health, environmental and economic outcomes.  One widely cited neighbourhood matching strategy is to increase the level of housing mix - or the diversity and distribution of different housing typologies within a neighbourhood - to permit households of lesser financial means to trade living space for an opportunity to live in their desired communities.  However, no empirical study has investigated whether increased neighbourhood housing mix is associated with higher levels of neighbourhood matching in the population. The purpose of this dissertation is to evaluate the effectiveness of housing mix as a planning strategy. Using data obtained from a residential preference survey of 1,186 Vancouver area households, this project investigates the association between neighbourhood housing mix and the ability for households to match into their preferred neighbourhood type.  The project also tests the association between neighbourhood match and neighbourhood satisfaction as well as the association between neighbourhood match and two measures of health: self-reported health status and body mass index (BMI).  Neighbourhood match is defined two ways: based on a survey respondent’s subjective interpretation of their actual neighbourhood design compared to their preferences (i.e. “subjective match), and a comparison of the respondent’s survey-indicated preference versus an objective assessment of their community based on measurable features of the built environment (i.e. “objective match”).  Findings reveal that housing mix only iii  significantly predicts objective match, and significant associations are limited to owner-occupiers and respondents under the age of 60.  Objective match is not a significant predictor of neighbourhood satisfaction or health.  This dissertation concludes that housing mix is not an effective planning strategy for enabling households with Smart Growth preferences to live in their desired community.    iv  Preface While I was responsible for the design of the research project and the analysis of all data, this project relied on datasets that were collected and developed by other researchers. The 2011 walkability surface was developed by Dr. Lawrence Frank in collaboration with the Health and Community Design Lab in the School of Population and Public Health at the University of British Columbia.  Parcel data on residential properties was developed by the British Columbia Assessment Authority and was purchased by the Health and Community Design Lab.  The Coalition Linking Action Science and Prevention residential preference survey was developed by Dr. Lawrence Frank in collaboration with Toronto Public Health.  As a Graduate Research Assistant in the Health and Community Design Lab, I was granted permission to use these datasets for my dissertation. No part of this dissertation has been previously published or is in the process of being published.  The student, Leonard Machler, identified the research question, developed all the variables used in the analysis, performed the analysis, and prepared the manuscript.  All images found within this dissertation were created by Leonard Machler unless otherwise noted.   v  Table of Contents Abstract .......................................................................................................................................... ii Preface ........................................................................................................................................... iv Table of Contents ...........................................................................................................................v List of Tables ................................................................................................................................ xi List of Figures ............................................................................................................................. xvi List of Abbreviations ................................................................................................................. xix Acknowledgements .................................................................................................................... xxi Dedication .................................................................................................................................. xxii Chapter 1: Introduction ................................................................................................................1 1.1 Problem Background and Literature Review .................................................................. 1 1.2 Research Questions ......................................................................................................... 3 1.3 Methods........................................................................................................................... 5 1.4 Results ............................................................................................................................. 7 1.5 Organization of Dissertation ......................................................................................... 10 Chapter 2: Problem Background ...............................................................................................12 2.1 Smart Growth Neighbourhoods .................................................................................... 13 2.1.1 Definition of Smart Growth ...................................................................................... 13 2.1.2 Evidence in Support of Smart Growth ...................................................................... 14 2.1.3 Evidence in Critique of Smart Growth ..................................................................... 17 2.1.4 Demand for Smart Growth Neighbourhoods ............................................................ 18 2.1.5 Why does Mismatch occur?  New Directions for Research ..................................... 20 2.2 Theories of Community Supply and Community Choice............................................. 21 vi  2.2.1 Bid-Rent Theory ....................................................................................................... 22 2.2.2 Public Goods Sorting ................................................................................................ 26 2.3 The Role of Housing Mix: Theories and Evidence ...................................................... 29 2.3.1 Housing Mix as a Means to Achieve Place Diversity............................................... 30 2.3.2 Housing Mix to Support Social Mix ......................................................................... 32 2.3.3 Housing Mix as a Means to Facilitate Residential Matching ................................... 33 2.4 Conclusions ................................................................................................................... 34 Chapter 3: Literature Review .....................................................................................................36 3.1 Smart Growth Preferences and Neighbourhood Matching ........................................... 36 3.2 Approaches to Measuring Smart Growth Preferences .................................................. 37 3.2.1 Preference for Smart Growth Neighbourhoods: Evidence ....................................... 41 3.2.2 Three Studies of Residential Preference and Mismatch ........................................... 44 3.2.3 Research Gaps ........................................................................................................... 48 3.3 Neighbourhood Satisfaction.......................................................................................... 49 3.3.1 Definitions of Satisfaction and Neighbourhood Satisfaction .................................... 49 3.3.2 Existing Research...................................................................................................... 52 3.3.3 Research Gaps ........................................................................................................... 55 3.4 The Relationship between Neighbourhood Match, Satisfaction and Health Outcomes 56 3.4.1 Self-Reported Health Status ...................................................................................... 56 3.4.2 Body Mass Index ...................................................................................................... 58 3.4.3 Implications of investigating Research Question 3 .................................................. 61 Chapter 4: Methods and Research Design ................................................................................63 4.1 Research Questions ....................................................................................................... 63 vii  4.2 Metro Vancouver as a Research Context ...................................................................... 68 4.3 Data Sources ................................................................................................................. 73 4.3.1 CLASP Residential Preference Survey ..................................................................... 74 4.3.1.1 Residential preference questions....................................................................... 76 4.3.2 2011 Walkability Surface ......................................................................................... 77 4.3.3 Parcel Data ................................................................................................................ 81 4.3.4 2011 National Household Survey ............................................................................. 83 4.4 Variable Construction Techniques ................................................................................ 85 4.4.1 Network Buffers for Spatial Variables ..................................................................... 85 4.4.1.1 Rationale of Using Buffers to Define “Neighbourhoods” ................................ 85 4.4.1.2 Buffer Design .................................................................................................... 88 4.4.2 Housing Mix ............................................................................................................. 92 4.4.2.1 Categorizing Housing Typologies – Exploratory Work ................................... 93 4.4.2.2 Variable Construction ....................................................................................... 98 4.4.3 Neighbourhood Match .............................................................................................. 99 4.4.3.1 Theoretical Issues............................................................................................ 100 4.4.3.2 Measuring Smart Growth environments ......................................................... 101 4.4.3.3 Defining Neighbourhood Match ..................................................................... 109 4.4.4 Neighbourhood Importance Factors and Built Form Component .......................... 111 4.4.5 Outlier Removal ...................................................................................................... 115 4.4.6 Other Predictor Variables ....................................................................................... 117 4.5 Models Used ............................................................................................................... 119 4.5.1 Research Question 1 ............................................................................................... 119 viii  4.5.2 Research Question 2 ............................................................................................... 123 4.5.3 Research Question 3 ............................................................................................... 124 4.5.3.1 Self-reported Health Status ............................................................................. 124 4.5.3.2 Body Mass Index ............................................................................................ 125 Chapter 5: Descriptive Results and Sample Characteristics .................................................127 5.1 Sample Characteristics ................................................................................................ 127 5.1.1 Geographic Distribution of Survey Responses ....................................................... 128 5.1.2 Socioeconomic and Demographic Representativeness ........................................... 132 5.2 Variable Characteristics .............................................................................................. 138 5.2.1 Housing Mix ........................................................................................................... 139 5.2.2 Neighbourhood Preferences .................................................................................... 147 5.2.3 Neighbourhood Match ............................................................................................ 149 5.2.3.1 Categories of Neighbourhood Match .............................................................. 154 5.2.4 Neighbourhood Satisfaction.................................................................................... 159 5.2.5 Health Outcome Variables ...................................................................................... 163 5.2.5.1 Self-Reported Health Status ............................................................................ 163 5.2.5.2 BMI measures ................................................................................................. 167 5.2.6 Conclusions ............................................................................................................. 170 Chapter 6: Model Results ..........................................................................................................173 6.1 Introduction ................................................................................................................. 173 6.2 Statistical Models Answering Research Question 1 ................................................... 173 6.2.1 Simple Models Predicting Neighbourhood Match ................................................. 174 6.2.2 Simple Models Predicting Neighbourhood Matching Category ............................. 179 ix  6.2.3 Discrepancies in Matching Outcomes at Different Spatial Scales .......................... 185 6.2.4 Multivariate Models Predicting Neighbourhood Matching Category .................... 188 6.2.4.1 The Insignificance of Household Income ....................................................... 194 6.2.4.2 Stratified Multivariate Models by Population Subgroup ................................ 198 6.2.4.2.1 Renters versus Owners .............................................................................. 199 6.2.4.2.2 Stratification Results by Age Group: Under 60 and Over 60 ................... 202 6.2.5 Summary of Findings .............................................................................................. 205 6.3 Statistical Models answering Research Question 2 .................................................... 206 6.3.1 Simple models predicting Neighbourhood Satisfaction ......................................... 206 6.3.2 Multivariate Models of Neighbourhood Satisfaction ............................................. 210 6.3.3 Summary of Findings .............................................................................................. 218 6.4 Statistical Models answering Research Question 3 .................................................... 218 6.4.1 Models predicting self-reported health status ......................................................... 219 6.4.2 Models predicting BMI ........................................................................................... 223 6.4.3 Summary of Findings .............................................................................................. 227 6.5 Conclusions ................................................................................................................. 227 Chapter 7: Conclusions .............................................................................................................231 7.1 Introduction ................................................................................................................. 231 7.2 Limitations of Research Design .................................................................................. 231 7.2.1 Potential Issues with the Sampling Strategy ........................................................... 232 7.2.2 Potential Issues with the Design of the Survey Instrument .................................... 234 7.2.3 Potential Issues from Survey Responses................................................................. 238 7.2.4 Potential Issue with the Construction of Variables ................................................. 239 x  7.2.5 Potential Issues with the Statistical Models ............................................................ 243 7.2.6 Summary of Limitations ......................................................................................... 244 7.3 The Unique Context of Metro Vancouver .................................................................. 245 7.4 Is Housing Mix Useful Policy? ................................................................................... 250 7.5 Future Policy Directions and Future Research Directions .......................................... 251 7.5.1 Policy Alternatives .................................................................................................. 251 7.5.1.1 Smart Growth Initiatives to Consider ............................................................. 258 7.5.2 Future Research Directions ..................................................................................... 260 Bibliography ...............................................................................................................................262 Appendices ..................................................................................................................................275 Appendix A Sensitivity Analysis ............................................................................................ 275 Appendix B CLASP Survey Instrument ................................................................................. 276  xi  List of Tables Table 1: Research questions............................................................................................................ 3 Table 2: Selected references to neighbourhood housing mix in planning documents .................. 31 Table 3: Approaches to measuring preference for residential environments ............................... 37 Table 4: Selected stated preference surveys of Smart Growth environments .............................. 42 Table 5: Existing studies of built environment predictors of neighbourhood satisfaction ........... 53 Table 6: Previously studied Smart Growth built environment predictors of neighbourhood satisfaction .................................................................................................................................... 54 Table 7: Spatial measures created using network buffers ............................................................. 90 Table 8: Housing categories used to construct housing mix variable ........................................... 94 Table 9: Frequencies of different housing categories, Metro Vancouver ..................................... 96 Table 10: CLASP survey trade-off questions ............................................................................. 102 Table 11: Principal components analysis for neighbourhood match, total variance explained .. 104 Table 12: Principal components analysis for neighbourhood match, factor loadings for component 1 ................................................................................................................................ 105 Table 13: Correlation coefficients between walkability index scores and current and neighbourhood preference scores ............................................................................................... 108 Table 14: Logistic regression results predicting odds of reporting top quartile of neighbourhood preference scores ......................................................................................................................... 108 Table 15: Factor analysis for neighbourhood importance questions, total variance explained .. 112 Table 16: Neighbourhood importance factors, factor loadings .................................................. 113 Table 17: Principal components analysis for built form factor, total variance explained .......... 114 Table 18: Principal components analysis for built form factor, factor loadings ......................... 115 xii  Table 19: Predictor variables used .............................................................................................. 117 Table 20: Neighbourhood match categories used in multinomial logistic regression models ... 120 Table 21: Neighbourhood satisfaction – frequency of scores ..................................................... 124 Table 22: Self-reported health status, frequency of scores ......................................................... 125 Table 23: Municipal representation in survey sample ................................................................ 131 Table 24: Walkability index, descriptive statistics ..................................................................... 132 Table 25: Demographic characteristics of study respondents ..................................................... 134 Table 26: Socioeconomic characteristics of study respondents .................................................. 135 Table 27: Housing mix defined around home postal code, descriptive statistics ....................... 141 Table 28: Housing mix defined around workplace postal code, descriptive statistics ............... 141 Table 29: Neighbourhood preference scores for selected population groups ............................. 148 Table 30: Neighbourhood match score, descriptive statistics ..................................................... 150 Table 31: Neighbourhood match score, descriptive statistics, renters vs. owners ...................... 152 Table 32: Neighbourhood match score, descriptive statistics, by age of respondent ................. 153 Table 33: Neighbourhood match score, descriptive statistics, by family type ........................... 153 Table 34: Neighbourhood match score, descriptive statistics, East Asian vs. non-East Asian respondents ................................................................................................................................. 154 Table 35: Neighbourhood match categories, frequency of observations .................................... 155 Table 36: Neighbourhood satisfaction, frequency of scores ....................................................... 160 Table 37: Neighbourhood satisfaction, descriptive statistics, subjective neighbourhood match categories .................................................................................................................................... 162 Table 38: Neighbourhood satisfaction, descriptive statistics, objective neighbourhood match categories .................................................................................................................................... 163 xiii  Table 39: Self-reported health status, frequency of scores ......................................................... 165 Table 40: Self-reported health status, descriptive statistics, subjective neighbourhood match categories .................................................................................................................................... 166 Table 41: Self-reported health status, descriptive statistics, objective neighbourhood match categories .................................................................................................................................... 166 Table 42: BMI variables, descriptive statistics ........................................................................... 167 Table 43: BMI, descriptive statistics, subjective neighbourhood match categories ................... 168 Table 44: BMI, descriptive statistics, objective neighbourhood match categories ..................... 169 Table 45: Binary logistic regression predicting subjective neighbourhood match, housing mix defined around home postal code ............................................................................................... 175 Table 46: Binary logistic regression predicting subjective neighbourhood match, housing mix defined around workplace postal code ........................................................................................ 176 Table 47: Binary logistic regression predicting objective neighbourhood match, housing mix defined around home postal code ............................................................................................... 177 Table 48: Binary logistic regression predicting objective neighbourhood match, housing mix defined around workplace postal code ........................................................................................ 177 Table 49: Multinomial logistic regression predicting subjective neighbourhood match, housing mix defined around home postal code ........................................................................................ 181 Table 50: Multinomial logistic regression predicting subjective neighbourhood match categories, housing mix defined around workplace postal code ................................................................... 182 Table 51: Multinomial logistic regression predicting objective neighbourhood match categories, housing mix defined around home postal code........................................................................... 183 xiv  Table 52: Multinomial logistic regression predicting objective neighbourhood match categories, housing mix defined around workplace postal code ................................................................... 184 Table 53: Multinomial logistic regression results predicting odds of subjectively matching neighbourhood to Smart Growth preferences ............................................................................. 190 Table 54: Multinomial logistic regression results predicting odds of objectively matching neighbourhood to Smart Growth preferences ............................................................................. 192 Table 55: Multinomial logistic regression results predicting odds of subjectively matching neighbourhood to Smart Growth preferences, effect of household income only ....................... 194 Table 56: Multinomial logistic regression results predicting odds of objectively matching neighbourhood to Smart Growth preferences, effect of household income only ....................... 195 Table 57: Tenure and household income, cross-tabulation of frequencies ................................. 196 Table 58: Multinomial logistic regression results predicting odds of matching neighbourhood to Smart Growth preferences, stratified by tenure .......................................................................... 199 Table 59: Multinomial logistic regression results predicting odds of matching neighbourhood to Smart Growth preferences, stratified by age ............................................................................... 202 Table 60: Ordered logit results predicting degree of neighbourhood dissatisfaction, subjective vs. objective neighbourhood match .................................................................................................. 207 Table 61: Ordered logit results predicting degree of neighbourhood dissatisfaction, neighbourhood match categories ................................................................................................ 208 Table 62: Ordered logit results predicting degree of neighbourhood dissatisfaction, neighbourhood match categories, socioeconomic and demographic covariates ........................ 210 Table 63: Ordered logit results predicting degree of neighbourhood dissatisfaction, neighbourhood match categories, built form covariates ............................................................. 214 xv  Table 64: Ordered logit results predicting degree of neighbourhood dissatisfaction, neighbourhood match categories,  satisfaction questions ........................................................... 217 Table 65: Ordered logistic regression predicting self-reported health status, subjective neighbourhood match categories ................................................................................................ 220 Table 66: Ordered logistic regression predicting self-reported health status, objective neighbourhood match categories ................................................................................................ 221 Table 67: Ordinary least squares regression predicting BMI, subjective neighbourhood match categories .................................................................................................................................... 225 Table 68: Ordinary least squares regression predicting BMI, objective neighbourhood match categories .................................................................................................................................... 226 Table 69: Walkability quartile and housing mix quartile, cross-tabulation of frequencies ........ 233 Table 70: Walkability quartile and tenure, cross-tabulation of frequencies ............................... 238  xvi  List of Figures Figure 1: Conceptual framework .................................................................................................... 4 Figure 2: Housing mix at 2 km, Metro Vancouver ......................................................................... 8 Figure 3: Outcomes pathway controlling for SES and demographic covariates ............................ 9 Figure 4: Neighbourhood satisfaction as it relates to overall life satisfaction .............................. 51 Figure 5: Conceptual framework .................................................................................................. 66 Figure 6: Metro Vancouver municipalities ................................................................................... 69 Figure 7: Distribution of walkability index scores, Metro Vancouver ......................................... 71 Figure 8: Street network buffer used for the creation of the 2011 walkability surface ................ 79 Figure 9: Buffer size comparisons around postal code V5V 2C6 ................................................ 91 Figure 10: Buffer size comparisons around postal code V4R 1S4 ............................................... 91 Figure 11: Distribution of subjective neighbourhood match scores ........................................... 110 Figure 12: Distribution of subjective neighbourhood match as a continuous variable ............... 110 Figure 13: Neighbourhood match categories visualized ............................................................. 121 Figure 14: BMI distribution in CLASP sample .......................................................................... 126 Figure 15: CLASP survey responses, Metro Vancouver ............................................................ 129 Figure 16: Walkability index and distribution of CLASP survey respondents .......................... 130 Figure 17: Household income of CLASP respondents compared to DA medians, Metro Vancouver ................................................................................................................................... 136 Figure 18: Household income of CLASP respondents compared to DA medians, central Metro Vancouver ................................................................................................................................... 138 Figure 19: Housing mix scores at 2km, Metro Vancouver ......................................................... 143 Figure 20: Housing mix at 2 km and walkability, central Metro Vancouver ............................. 144 xvii  Figure 21: Housing mix at 2km, spatial clustering, Metro Vancouver ....................................... 145 Figure 22: Housing mix clusters and median DA household income, central Metro Vancouver..................................................................................................................................................... 146 Figure 23: Subjective neighbourhood match categories, Metro Vancouver............................... 155 Figure 24: Objective neighbourhood match categories, Metro Vancouver ................................ 156 Figure 25: Neighbourhood match categories, subjective and objective conflicts ...................... 158 Figure 26: Neighbourhood satisfaction responses, Metro Vancouver ........................................ 159 Figure 27: Neighbourhood satisfaction responses, central Metro Vancouver ............................ 161 Figure 28: Self-reported health scores, Metro Vancouver .......................................................... 164 Figure 29: Self-reported health scores and median DA household income ................................ 165 Figure 30: BMI observations, Metro Vancouver ........................................................................ 170 Figure 31: Housing mix at 500 m, central Vancouver ................................................................ 187 Figure 32: Housing mix at 1 km, central Vancouver .................................................................. 187 Figure 33: Income category of respondents, Metro Vancouver ................................................. 196 Figure 34: Tenure of respondents, Metro Vancouver ................................................................. 197 Figure 35: Renters by objective neighbourhood match categories, central Metro Vancouver ... 200 Figure 36: Renters by subjective neighbourhood match categories, central Metro Vancouver . 201 Figure 37: Respondents over 60 years of age by objective neighbourhood match categories, Metro Vancouver ........................................................................................................................ 204 Figure 38: Respondents over 60 years of age by objective neighbourhood match categories, central Metro Vancouver ............................................................................................................ 205 Figure 39: Outcomes pathway, no control variables .................................................................. 228 Figure 40: Outcomes pathway, controlling for SES and demographic variables ....................... 229 xviii  Figure 41:Trade-off question 5 from CLASP survey ................................................................. 235 Figure 42: Housing mix and property values, Metro Vancouver ............................................... 248 Figure 43: Housing mix and property values, central Metro Vancouver ................................... 249  xix  List of Abbreviations BC – British Columbia BMI – Body Mass Index CAC – Community Amenity Contribution CBD – Central Business District CLASP – Coalition Linking Action Science and Prevention CNS – Current Neighbourhood Score CPR – Canadian Pacific Railway DA – Dissemination Area DCL – Development Cost Levy FSA – Forward Sortation Area GIS – Geographic Information Systems HOPE – Home Ownership for People Everywhere  LCA – Latent Class Analysis LEM – Location Efficient Mortgage MAUP – Modifiable Area Unit Problem NHS – National Household Survey NPS – Neighbourhood Preference Score PCA – Principal Component Analysis RCS – Regional Context Statement RGS – Regional Growth Strategy SES – Socioeconomic Status SG – Smart Growth xx  TAZ – Traffic Analysis Zone TMM – Target Market Method VIF – Variance Inflection Factor   xxi  Acknowledgements I would like to begin by thanking my supervisor, Dr. Lawrence Frank for all of his mentorship and support.  Dr. Frank was unfailingly generous to me during my time as his PhD student, and provided me with both funding and the opportunity to work in his research lab, as well as thoughtful but critical advice on how to successfully navigate my PhD. I would also like to thank my committee members, Dr. Mark Stevens and Dr. Craig Tsuriel Somerville, who offered different and very welcome perspectives on my research project as it evolved.  I would like to express my gratitude toward a number of different faculty members, post-doctoral researchers, staff and fellow students both at the University and abroad for their guidance, inspiration, support and camaraderie. These include, in no particular order, Dr. Joshua Van Loon, Dr. James White, Dr. Ren Thomas, Jacopo Miro, Magdalena Ugarte, Dr. Nathan Schiff, Dr. Judith Innes, Dr. Aaron Golub, Dr. Dan Milz, Sean Bohle, and many others.    I would like to acknowledge the financial support I received from the Social Sciences Humanities and Research Council of Canada’s doctoral fellowship, as well as the University of British Columbia’s Four Year Fellowship.  Finally, I would like to thank my friends and family for their enduring support throughout these past few years.   xxii  Dedication       For Arabelle 1  Chapter 1: Introduction 1.1 Problem Background and Literature Review The overarching goal of this dissertation is to investigate effective planning strategies for housing people into the residential communities that they prefer.  In particular, this project examines strategies that effectively house individuals that prefer Smart Growth neighbourhoods –characterized by higher densities, mixed land uses and a more pedestrian-focused design (Downs 2005).  This goal is informed by a belief supported by considerable evidence that Smart Growth communities perform better environmentally, socially, and economically than the suburban development patterns that predominated in North America since the end of the Second World War. Residing in a Smart Growth community has been empirically linked with numerous environmental, economic, and health benefits for both individuals and society over living in low density, automobile-oriented “suburban” neighbourhood environments (Burchell et al. 2002; Frumkin, Frank, and Jackson 2004; Newman and Kenworthy 1999). This is a study of “housing mix”: the diversity and distribution of different housing types, such as single family homes, duplexes, townhomes, and multifamily apartments, among others, within a defined geographic area.  Different housing types are expected to cater to households with different space needs and varying economic means.  Therefore, providing a greater variety and distribution of housing types within a neighbourhood is often cited as a strategy for ensuring that a broader spectrum of the population is housed in the communities of their choice (Levine and Inam 2004; Danielsen, Lang, and Fulton 1999). Greater housing mix is a strategy that informs official community plans and strategic growth documents in countless North American municipalities and regions (City of Vancouver 2013; Metro Vancouver 2011; City of Toronto 2010; City of Los Angeles 2009). Smart Growth 2  advocates, in particular, have embraced the concept as one of the guiding principles for urban development (Nelson 2013; Talen and Knaap 2003; Leinberger 2008). Very little empirical evidence has been gathered on the effectiveness of housing mix as a planning strategy.  In particular, no study has identified whether insufficient neighbourhood housing mix is a barrier to “residential matching” that is, how successfully households are able to live in their preferred type of community.  This is a particularly crucial research gap to fill, not only given the benefits associated with Smart Growth, but also given that demand for Smart Growth neighbourhoods is on the rise (National Association of Realtors 2011) with particularly strong preferences from the large and aging baby boom cohort (1946-1964) (Myers and Ryu 2008; Myers and Gearin 2001) and Generation Y cohort (1981-1995) (Logan 2013).  Retrofitting the urban landscape and residential stock to accommodate the demand for Smart Growth may pose one of the most significant challenges to the real estate industry and planning profession in North America in the coming decades (Nelson 2013, 2006). Several residential preference studies have suggested that a considerable number of households with preferences for Smart Growth are unable to locate in their preferred environments (Schwanen and Mokhtarian 2005; Schwanen and Mokhtarian 2004; Levine and Frank 2006; Levine, Inam, and Torng 2005).  Many of these studies have inferred that there is an insufficient supply of homes to meet the growing demand for Smart Growth communities (Levine, Inam, and Torng 2005; Levine and Frank 2006; cf. Cao 2008). Regions that lack a sufficiently diverse range of Smart Growth housing options may risk restricting the benefits of these communities to a small, privileged segment of their population.  This problem may be particularly acute if metropolitan housing costs are high.    3  1.2 Research Questions  This dissertation examines the effectiveness of housing mix as a planning strategy in achieving better residential matching into Smart Growth neighbourhoods.  It further examines whether better Smart Growth matching leads to beneficial outcomes for individuals. As such, this dissertation poses three research questions (Table 1) and a conceptual framework (Figure 1) guides the design of the research project. Table 1: Research questions Question # Research Question   1st (Primary) Controlling for relevant personal characteristics, is there a significant association between the degree of neighbourhood housing mix and the ability of households to match into the neighbourhood type of their preference?   2nd Is an individual’s ability to match their neighbourhood type with their preference significantly associated with their self-reported  level of neighbourhood satisfaction?   3rd Is an individual’s ability to match their neighbourhood type with their preference significantly associated with favourable health outcomes?     “Housing mix” refers to a measure of how diverse and evenly distributed the residential stock of dwellings is among different housing typologies, such as single family homes, townhouses, and multifamily apartments. “Neighbourhood” is a defined spatial unit of analysis extending from between 500 m to 5 km around a residential or workplace address, and “neighbourhood type” specifically refers to the degree to which the physical environment of the neighbourhood embodies attributes of Smart Growth development.  “Match” refers to a measured value of how closely individuals are able to align the neighbourhood type that they reside in with that of their stated preference. 4  Figure 1: Conceptual framework  Note: solid lines indicate relationships that are expected to be significant; dashed lines indicate relationships that are expected to be insignificant.  Controlling for relevant personal and neighbourhood characteristics, it is hypothesized that a higher degree of housing mix is significantly and positively associated with ‘neighbourhood match’. However, housing mix, on its own, is not anticipated to be a significant positive predictor of neighbourhood satisfaction nor of favourable health outcomes.  Instead, it is anticipated that the pathway between housing mix and neighbourhood satisfaction is moderated through the ability to “neighbourhood match”.  Respondents who live in neighbourhoods with a high degree of housing mix are expected to more closely match their neighbourhood preferences, and respondents who are better able to match their neighbourhood preferences will be more satisfied with their neighbourhood environments.  Neighbourhood satisfaction, in turn, Personal Characteristicsof IndividualSelf-reported Satisfaction with Attributes of Home and NeighbourhoodNeighbourhood SatisfactionNeighbourhood MatchDegree of Neighbourhood Housing MixHealth OutcomesOther, Unobserved Factors(Macroeconomic, Personal)Measurable Attributes of the Home and Neighbourhood EnvironmentPRIMARY RESEARCH QUESTION SECONDARY RESEARCH QUESTIONS5  moderates the relationship between neighbourhood match and health outcomes.  While neighbourhood match is not expected to be a significant predictor of favourable health outcomes on its own, it is expected that neighbourhood match will significantly predict neighbourhood satisfaction, and that neighbourhood satisfaction will be positively and significantly associated with improved health outcomes.    1.3 Methods These relationships were tested in the Greater Vancouver Regional District (informally known, and henceforth referred to, as Metro Vancouver) using data obtained from several sources, all from 2011.  Personal data on residential preferences, neighbourhood satisfaction, health, socioeconomic status (SES) and demographic characteristics were obtained from the Coalition Linking Action Science and Prevention (CLASP) survey – a residential preference survey of 1,186 Metro Vancouver households (Frank et al. 2012; Frank et al. 2014).  Parcel-level data on the structural attributes of 744,457 Metro Vancouver residential properties was obtained from the British Columbia Assessment Authority (BC Assessment).  Objectively-measured attributes of neighbourhood environments were obtained at the postal code level (n = 61,299) from a 2011 Walkability surface developed by the University of British Columbia’s Health and Community Design Lab (Frank et al 2011).  Finally, neighbourhood-level SES and demographic data was supplied by the 2011 National Household Survey (NHS) at the level of the dissemination area (n = 3,438).   Using GIS shapefiles of all Metro Vancouver properties supplied by BC Assessment and Metro Vancouver road network shapefiles supplied by DMTI Spatial, network buffers were created in ArcMap at 500m, 1 km, 2km, 3km and 5km distances around CLASP survey 6  respondents’ postal codes (n=1,118).  Housing mix was then calculated by applying an entropy calculation (Shannon and Weaver 1949) on the prevalence of housing units within 7 pre-defined housing categories within the area encompassed by the 5 different network buffers.  Neighbourhood preferences were measured by running a principal component analysis (PCA) on the results from five illustrated questions on the CLASP survey asking respondents to rate their preference for a Smart Growth versus a suburban environment on an 11-point ordinal scale. From this, respondent-specific “neighbourhood preference scores” (NPS) were elicited, with positive scores indicating greater preferences for Smart Growth over suburban environments. The same five questions also asked respondents to assess their own neighbourhood using the same ordinal scale, so a “current neighbourhood” score (CNS) was estimated by multiplying the factor loadings from the preference PCA onto the respondent’s current neighbourhood responses.  Additionally, the Smart Growth attributes of the current neighbourhood were objectively measured around each respondent’s postal code using the 2011 walkability surface.  For this dissertation, the walkability surface for the entire region was recalibrated to use the same scale as the NPS. “Neighbourhood match” – the ability for a household to sort into their desired neighbourhood environment – was measured two ways:  (1) Subjective neighbourhood match proxied for a respondent’s own assessment of their match situation.  Subjective neighbourhood match was defined by subtracting the CNS from the NPS and assigning all observations within a half standard deviation from 0 (i.e. perfect match) as “matched”. (2) Objective neighbourhood match proxied for a respondent’s actual ability to match their Smart Growth preferences to the actual neighbourhood environment.  Objective 7  neighbourhood match was defined by subtracting the walkability index score from the 2011 walkability surface from the NPS and assigning all observations within a half standard deviation from 0 as “matched”. Neighbourhood satisfaction was measured as an 11 point ordinal question on the CLASP survey.  Measured health outcomes included “self-reported health status”, in which CLASP survey respondents were asked to evaluate their own health compared to people of the same age on a 5 point ordinal scale, and body mass index (BMI), which was calculated from the height and weight measurements that respondents supplied on the survey. The effect of housing mix on neighbourhood match (Research question 1) was tested using a series of logistic regressions and multinomial logit models.  The effect of housing mix and neighbourhood match on neighbourhood satisfaction was tested using ordered logit models.  Ordered logit models were also used to predict self-reported health status, while BMI was predicted using an ordinary least squares regression.  1.4 Results The findings support previous conclusions that Smart Growth neighbourhoods are more popular overall than their suburban counterparts.  In particular, renters express greater preferences for Smart Growth than owners, and respondents over the age of 60 express greater preferences than respondents under the age of 60 highlighting the importance of smart growth environments for older Canadians.  Areas of high housing mix are dispersed throughout the region, although some significant clustering occurs in the inner city and inner suburban neighbourhoods to the east of downtown Vancouver.  Low housing mix is clustered in the 8  downtown peninsula.  Figure 2 shows the housing mix scores, defined using a 2 km network buffer, for the 1,118 unique postal codes in the CLASP dataset. Figure 2: Housing mix at 2 km, Metro Vancouver  Housing mix was a significant predictor of objective but not subjective neighbourhood match.  Predictive ability peaks when housing mix is defined at a 2 km distance from a respondent’s home postal code.  Additionally, housing mix was found to significantly increase the odds that households with preferences for Smart Growth neighbourhoods would objectively match into these neighbourhood types.  However, this relationship is not significant for renters or respondents over the age of 60.  Neither subjective nor objective neighbourhood match was significantly associated with neighbourhood satisfaction. The degree of housing mix within a neighbourhood was significantly associated with lowered neighbourhood satisfaction levels..  9  Neighbourhood satisfaction and being subjectively matched into a Smart Growth environment were positively associated with self-reported health status.  However, being objectively matched was not associated with self-reported health status.  Neither subjective nor objective match with Smart Growth predicts BMI levels, nor does neighbourhood satisfaction. The pathway between these relationships is illustrated in Figure 3.  From these findings, we may conclude that housing mix does not enable households to sort into the neighbourhoods they prefer, and is not an effective policy for ensuring better residential matching into Smart Growth.  These conclusions are underscored by the finding that housing mix only significantly predicts one’s ability to objectively match into Smart Growth, and then only for owner occupiers and people under 60.   Figure 3: Outcomes pathway controlling for SES and demographic covariates  The possibility that areas of high housing mix are associated with high property values and, thus, out of the financial reach of many of the households it is intended to assist is Housing MixObjectively Matched into SGSubjectively Matched into SGNeighb.SatisfactionHealth StatusBMISignificant and positive (p <0.05)Significant and negative (p <0.05)No significance10  dismissed.  Findings reveal that housing mix and property values are uncorrelated;  areas of high housing mix can be found across the region among an assortment  of suburban and Smart Growth neighbourhoods featuring a range of median property values.  1.5 Organization of Dissertation The dissertation is organized as follows: Chapter 2 introduces the problem and provides some theoretical background on Smart Growth, theories of community choice and supply, and housing mix. Chapter 3 reviews the state of academic literature on measuring preferences for Smart Growth environments as well as neighbourhood satisfaction and the link between neighbourhood matching and neighbourhood satisfaction and self-reported health status and BMI.  Chapter 4 describes the research design of the dissertation and the methods used in the analysis.  The chapter begins by outlining the research questions and conceptual framework guiding the dissertation as well as some conceptual issues in measurement that might arise (Section 3.1).  An overview of the Metro Vancouver region, where the study is conducted, and its planning history is briefly introduced, while the suitability of Metro Vancouver as a research context is examined (Section 3.2).  Data sources are introduced (Section 3.4), and the techniques used to construct the variables for analysis are described in detail (Section 3.5).  Finally, the choice of statistical models used to analyze the research questions are offered in Section 3.6. Chapter 5 uses descriptive statistics and maps to provide an overview of the sample’s representativeness and the characteristics of the major variables.  Chapter 6 presents the results of the statistical models, in the order of the research questions. Finally, the dissertation concludes with a discussion chapter (Chapter 7).  This chapter begins by examining the various limitations to the research design, beginning with the 11  sampling strategy and ending with the choice of statistical models (Section 6.2).  The appropriateness of Metro Vancouver as a region in which to study housing mix is considered (Section 6.3), and the usefulness of housing mix as a planning strategy to enable greater residential matching is contemplated (Section 6.4).  In the final section, the viability of policy alternatives are examined and future research opportunities are identified (Section 6.5).   12  Chapter 2: Problem Background  In this chapter, the research problem is introduced, and some theoretical background is provided to build an argument in support of pursuing this dissertation.  The chapter is organized into three sections.  The first section explores the concept of Smart Growth neighbourhoods.  After defining the term, Smart Growth neighbourhood designs are demonstrated to support beneficial environmental, economic and health outcomes and are also increasingly in demand.  Finding approaches to house the growing number of people who prefer these normatively beneficial environments poses a challenge for planners.  Housing mix, which enables individuals of various economic means to forego living space in exchange for the opportunity to reside in neighbourhoods beyond their means, is considered. The second section  examines two urban economic theories of how residential communities are chosen and supplied: bid-rent theory and public goods sorting. In the third major section, three normative theories of the role of housing mix are presented.  Academic evidence on the role of housing mix plays in achieving place diversity by supporting social mixing and facilitating residential matching are considered. From these three sections, a general theme emerges: housing mix is presented as a means to overcome barriers to living in a smart growth neighbourhood for those who prefer it, but may not have the means to live there.  Housing mix is also shown to conceptually align with bid-rent and public goods sorting theories of community choice and community supply and build off of existing studies on matching Smart Growth preferences.  Additionally, this review argues that the role of housing mix in facilitating greater residential matching remains understudied.  Despite this lack of evidence, neighbourhood housing mix is widely touted as a housing strategy in many planning documents.  Studying housing mix’s ability to facilitate Smart Growth sorting is an important gap in planning research that needs to be addressed. 13  2.1 Smart Growth Neighbourhoods The aim of this section is to introduce the concept of Smart Growth neighbourhood environments and use existing academic evidence to argue for the continued supply of Smart Growth communities.  The term “Smart Growth” is defined before two motivations for supplying more Smart Growth communities are examined: these include studies that reveal both normative benefits to society and individuals, and also the increasing popularity of communities of this type.  Approaches to measuring the preference of neighbourhood environments are introduced before evidence of the preference for Smart Growth neighbourhood types is presented.  This leads to a review of three, particularly influential studies that informed the design of this dissertation, and examines their approaches to investigating whether or not homes in Smart Growth communities were undersupplied relative to demand.    2.1.1 Definition of Smart Growth Along with words such as “sustainability” and “resilience”, the term “Smart Growth” has become something of a buzzword in urban planning (Resnik 2010).  What is Smart Growth, exactly? Experts themselves may not agree: Anthony Downs once wryly observed that “Smart Growth does not mean the same thing for everyone” (Downs 2005).  It is also unclear where the term originated (Knaap 2004), but the word “Smart Growth” began to emerge in the 1990s when a host of influential policy programs were initiated with these words in the title (Burchell, Listokin, and Galley 2000).  From its inception, Smart Growth was a “big tent” movement incorporating a variety of different planning issues and their associated advocacy groups, so the meaning of the term may never have emerged from a consensus position to begin with (Goetz 2004).  Still, most advocates of Smart Growth would generally agree that communities with 14  “Smart Growth” characteristics include, but are not excluded to, compact, higher density development patterns with mixed land uses that support the viability of travel alternatives to the private automobile (Downs 2005). More importantly, they would also agree on what Smart Growth communities are not: low density, automobile-oriented environments characterized by segregated land uses that have dominated urban development in North America since the end of the Second World War (ibid).  For both its supporters and its detractors, the “Smart Growth” concept has been effectively positioned as a reaction against “suburban” forms of development, and it is not unusual to see the two development patterns presented as dichotomies (viz. Danielsen, Lang, and Fulton 1999; Easterbrook 1999).  For the remainder of this dissertation these two neighbourhood conceptions will be referred to as “Smart Growth” and “suburbia”, respectively. .  It is also important to recognize that while they refer to different practices, terms such as “New Urbanism” (Congress of the New Urbanism 2001)1, “Walkable Urbanism” (Leinberger 2008), “walkable” (vs. “car-dependent”) (Frank et al. 2007), and “traditional neighbourhood design” (as opposed to “suburban”) (Cao 2008; Lovejoy, Handy, and Mokhtarian 2010) connote similar ideas and are used interchangeably with the term “Smart Growth”.    2.1.2 Evidence in Support of Smart Growth Is the development of Smart Growth a worthy policy goal?  Many studies have demonstrated that communities with Smart Growth characteristics confer many economic, environmental and social benefits to both the individuals who reside in them and to society at large.  Evidence gathered to support the development of these communities may be divided                                                  1 See, for example, The Smart Growth Manual (Duany, Speck, and Lydon 2010), a book co-written by Andres Duany, a co-founder of the Congress for New Urbanism, largely referencing New Urbanist principles. 15  between those studies that analyze the social, environmental and economic benefits of Smart Growth residential developments on societal well-being and those that measure the benefits of the Smart Growth built form on individual well-being.  It is important to recognize that a large volume of research have been performed linking Smart Growth to normative benefits for both individuals and society, and this review will only cover a small selection of these studies. Briefly, Smart Growth built forms, with their compact designs and emphasis on providing transport alternatives to the car have been associated with environmental benefits to all of society in the form of less natural ecosystem damage from sprawling, lower density urban development (see an extensive review by Johnson (2001)) and  lower greenhouse gas emissions (Hankey and Marshall 2010).  In terms of economic benefits, the efficiency of more compact development forms over automobile-oriented “sprawl” has been analyzed most thoroughly in The Costs of Sprawl, a multi-volume report compiled by Burchell et al. (2002). That report found significant cost savings to municipalities from more dense development patterns, particularly on infrastructure construction and maintenance expenditures related to water and sewer utilities and roadways.  However, other economic benefits to society, such as monetizing the health benefits of Smart Growth versus conventional suburban developments, remain understudied, as does a quantification of how the amenities offered in Smart Growth communities are valued by residents (Song and Stevens 2012).  As such, affixing a monetary value to the welfare benefits of Smart Growth and the welfare costs of sprawl may be difficult, and designing policies to more properly capitalize these costs and benefits into the price of walkable urban housing products and their suburban counterparts remains an area for further study. Studies that link the benefits of Smart Growth environments to individual wellbeing have overwhelmingly been attributed to reductions in automobile use and ownership associated with 16  living in Smart Growth communities.  Due to increasing car ownership and reliance on cars for personal transportation, travel already accounts for the second highest - and fastest growing -component of discretionary household spending in the US after housing (Pisarski 2006), and combined household spending on housing and transportation now consumes over half of the budget of an average working family (Lipman 2006). Controlling for income, household size, and other demographic factors, Giuliano and Dargay (2006) found that the per capita vehicle ownership rate of US households living in dense neighbourhoods was 61% lower than those households living in low density areas. Meanwhile, Frank et al. (2006) demonstrated that residing in Smart Growth communities was significantly linked to decreases in vehicle miles traveled and vehicular emissions. Residing in suburban communities is also significantly associated with the time individuals spent in cars (Frank et al. 2007), which may fragment the amount of time spent in other areas of life, and have significant ramifications for the development of one’s social networks (Putnam 1995, 214).  As the name implies, walkable neighbourhood environments are more oriented toward walking activity and individuals who live in these communities are significantly more likely to receive the recommended level of daily exercise Even if individuals hold positive attitudes toward urban living and enjoy walking, the car-dependency of their neighbourhood may negative impact their level of walking activity. Frank et al. (2007) revealed that individuals with preferences for walkable neighbourhoods walked less, traveled further in cars, and were more likely to be obese if they lived in low walkability residential environments compared to their counterparts who had successfully sorted into high walkability neighbourhoods. These findings underscore the importance of matching households with preferences for walking and walkable communities into neighbourhoods that possess those built form characteristics. 17  2.1.3 Evidence in Critique of Smart Growth Smart Growth policy, having attracted considerable attention, has also attracted its share of critics.  Some, such as Briffault (2002) and Downs (2005), are not critical of the aims of Smart Growth, per se, but doubt the feasibility of implementing one particular principle of Smart Growth: coordinating regional land use planning (particularly growth management) among competing local governments.  Perhaps most famously within planning circles, Gordon and Richardson (1997) were critical of Smart Growth’s aims, questioning whether compact cities were desirable planning goals.  Their critique primarily scrutinized the economic efficiency of Smart Growth policy using evidence from other studies. Bruegmann (2006) used historical analysis to suggest that the long term trend in human settlement was toward ever-decreasing residential densities and ever-more segregated land uses, especially as societies increased their wealth and purchasing power.  Bruegmann concluded that sprawl has always existed and is heavily desired by upwardly mobile households throughout the world; efforts to curb sprawl and build compact communities are dismissed as flawed interventions to counteract a natural process.  Echoing Gordon and Richardson (1997), Heikkila et al. (1989) and Giuliano et al. (2010)2 provide evidence that the increasingly dispersed location of jobs and amenities in metropolitan regions undermines the value and importance of traditional downtowns to a large portion of the population, and suggest that efforts to spend considerable resources to revitalize and direct growth to downtown areas may be misguided.  There is also considerable debate as to whether the travel health benefits of living in Smart Growth environments are, in fact, the result of people with a predisposition for walking and healthy lifestyles self-selecting into Smart Growth                                                  2 Note that Peter Gordon was an author on both these articles. 18  communities that can facilitate this behaviour (Handy and Clifton 2001; Bagley and Mokhtarian 2002).  Longitudinal studies that control for attitudes and compare behaviour as individuals move from suburban to Smart Growth environments (and vice versa) will inform this debate.    2.1.4 Demand for Smart Growth Neighbourhoods Another motivation for building Smart Growth communities is that they are becoming increasingly popular (National Association of Realtors 2011).  While preference studies show that Smart Growth neighbourhoods still remain less favoured than suburban neighbourhoods (see Section 3.2.1), Smart Growth communities still represent a product of considerable size. In the United States where approximately 2 million homes are constructed every year, if only a quarter of new home purchasers preferred Smart Growth neighbourhoods, this would still translate to a demand for over 500,000 new homes built in these types of communities per year (Nelson 2013).  More importantly, demand for Smart Growth is growing, and much of that growth may be motivated by demographic change.  Stronger than average preferences for Smart Growth neighbourhoods have been uncovered among influential demographic groups, such as the large Generation Y cohort (born 1981-1995) that are making their first home purchases and family formation decisions (Logan 2013) but, most importantly, among the aging baby boom cohort (born 1946-1964) (Myers and Gearin 2001) that, comprising over a quarter of the American population, represents the largest generational cohort in US3 history (Nelson 2013, 23).  The residential preferences of the baby boom cohort present planners with two impending challenges:                                                  3 Canada’s population breakdown is more even: in 2014, 26.3% of the population was between 50 and 69 years of age (roughly corresponding to Canada’s baby boom definition of 1946-1965), while 27.3% of the population was between 30 and 49.  However, these figures reflect a lower population among people aged 65 to 69 (i.e. born 1945-1949) compared to other five year cohorts (Statistics Canada 2014). 19  first to accommodate a rising demand for housing stock in communities with Smart Growth attributes and - because of the prospect of elderly people downsizing - a second issue of meeting the rising demand for smaller dwelling units or alternatives to single family homes within those communities (Myers and Ryu 2008; Myers and Gearin 2001).  The American homebuilding industry may not be adequately responding to these changing circumstances; in 2010, over 85% of the homes built in the US were still single family detached homes, and more than a quarter of those were built on lots that were over half an acre in size (Nelson 2013, 54).  Myers and Ryu (2008) went as far as to suggest that the continued construction of large, suburban homes just as baby boomers begin to downsize out of this product, simply adds to the oversupply of dwelling types that a future market will not be able to absorb.  They predict a large demand discrepancy between highly sought-after smaller housing typologies in Smart Growth communities compared to a glut of large homes in suburban and exurban automobile-oriented areas (ibid).  As Nelson (2006) remarked in an influential piece in the Journal of the American Planning Association, the challenge of meeting the demand for alternative housing types in Smart Growth communities, and the institutional apparatus that must be set up to accommodate this type of land use planning, represents nothing less than  a “new era” in American planning.  Concern over the gap between the supply and demand for homes within Smart Growth neighbourhoods has spawned a series of studies that have sought to measure how many households may be mismatched away from their Smart Growth preferences (Schwanen and Mokhtarian 2004; Levine, Inam, and Torng 2005; Levine and Frank 2006; Cao 2008)4.                                                     4 These will be reviewed in much more detail in the following chapter. 20  2.1.5 Why does Mismatch occur?  New Directions for Research Regardless of the degree of mismatch, why do households with preference for Smart Growth (or suburban) areas end up mismatched?  Research into the factors that lead households with preferences for urban living to be mismatched remains understudied. Schwanen and Mokhtarian (2004) focused on the personal characteristics of respondents who are shown to be mismatched, revealing that households with fewer workers relative to household members (presumably families with children), and those who own fewer cars, were among the suburban-dwelling respondents who were more likely to be mismatched with their urban preferences. Their findings also suggest that households with fewer space constraints (e.g. the elderly) were better able to sort themselves in their preferred neighbourhoods than those that had space constraints (e.g. families with children). However, the authors focus on the household as a unit of analysis - testing differences in self-reported personal characteristics that might lead to mismatch - and do not examine how measurable attributes of the property or the neighbourhood characteristics may lead to residential mismatch. Unlike property and neighbourhood attributes, personal characteristics cannot be modified – at least directly - through planning intervention and urban policy.   Of course, the most obvious reason for why people may not be able to sort themselves into the neighbourhood of their choice is their ability to pay to live in homes in those communities.  This assumption is taken for granted in most studies, and a direct question on housing costs – if not income - is not queried in many residential preference surveys, including the one used in this dissertation.  However, if we accept the premise that people cannot live in neighbourhoods that they cannot afford, we must also accept the premise that individuals are free to make personal sacrifices, if they are possible, to live in communities that would otherwise be 21  beyond their financial means (Stone 2006).  These sacrifices, or trade-offs, could involve anything from curtailing household spending set aside for other expenditures to borrowing larger sums to choosing homes that are inferior in certain qualities (Harkness and Newman 2005; Pickvance and Pickvance 1994). One of the most important sacrifices that people can make to sort themselves into their preferred communities is the amount of living space available.  One of the few mechanisms planners have to augment the amount of living space available for residents is to regulate the diversity and distribution of housing typologies permitted.  As the next section will illustrate, increasing the diversity of housing typologies at the neighbourhood level has been championed in many official plans and growth strategies, and remains a major policy plank of Smart Growth and New Urbanist advocates.  Despite the importance given to this concept, no evidence has been amassed that greater housing mix at the neighbourhood level increases the ability for a greater proportion of the population to sort themselves into the neighbourhoods of their choice. If finding ways to house people who prefer and derive health and other benefits from Smart Growth communities represents the challenge that will define the new era of planning (Nelson 2006), we must understand whether housing mix is an effective policy lever to achieve this.   2.2 Theories of Community Supply and Community Choice In order to appreciate how homes enable individuals to sort themselves into the communities of their choice it is critical to link this behaviour to existing microeconomic theories of how neighbourhood types are supplied and how individuals choose them, thereby putting demand for these neighbourhoods into action.  Two very old theories in urban economics are explored and, it is argued, are intertwined.  “Bid rent theory” provides both a supply and 22  demand-side model of how land is priced, and homes of different sizes supplied, as a function of travel costs and distance to an implied centre.  Public goods, or “Tiebout”, sorting proposes that individuals spatially arrange their settlement around a metropolitan area based on their affinity to amenities that are differentially provided by different communities.  An attempt is made to tie both theories together by arguing that residents who seek to live in certain communities because of their amenities (e.g. neighbourhoods with Smart Growth attributes) must choose from homes that were supplied largely under bid-rent conditions.  Because the supply of homes reflects political and societal intervention as much as unencumbered land economics, it is proposed that the types of homes supplied in communities present barriers to community choice.  Based on this logic, a diversity of housing typologies within a community is hypothesized to offer a “second best” choice to Pareto-optimal neighbourhood sorting, which is deemed to be practically impossible.  2.2.1 Bid-Rent Theory What factors determine which housing types are supplied, and where they are distributed within a metropolitan region?  While no universal theory can provide an explanation for all contexts, perhaps the most general answer comes from one of the oldest and most prominent theories in urban economics (Davidoff 2008).  Known as “bid-rent theory”, or the bid-rent model, the concept holds that the central business district (CBD) commands the greatest location efficiency for firms who compete fiercely for land in this area.  With increasing distance from the CBD, competition falls, and the price of real estate per unit of area decreases.  The concept harkens back to Johann Von Thünen’s 1826 theories of agricultural land rent being governed by the distance, and therefore costs, of transporting crops to a market (Alonso 1964, 3).  However, 23  housing economists are more likely to identify with adaptations to bid rent theory associated with the work of Alonso (1964), Mills (1967) and Muth (1969).  According to Muth (1969), the marginal utility per dollar spent is the same for all households, and housing prices must decline with distance from the CBD if the marginal cost of transportation is positive.  These conditions imply that identical households increase their consumption of housing services if they locate progressively further from the CBD.  In Muth’s model, specifically, when moving to housing in the suburban periphery, households trade the convenience of proximity to the central core for greater living space. The bid-rent model neglects many important dimensions of housing supply, demand, and urban growth.  Among other things, the model ignores the decline in the location efficiency of access to the CBD in regions with dispersed employment and travel patterns (Heikkila et al. 1989), political and cultural forces that shape the location and scale of development in a metropolitan area (Fischel 2004), and that it relies on the erroneous assumption that residential location decisions are made by individuals with a perfect knowledge of the entire housing market (Jones and Watkins 2009; MacLennan and Tu 1996).  Nevertheless, the bid-rent model is instructive in explaining a general phenomenon observed in rapidly growing regions with high land prices and desirable CBDs.  If we assume that housing developments on the suburban periphery offer the least access to the core, they will presumably offer the greatest amount of space in the form of a low density, single family detached home environments.  However, the outer suburbs of previous generations have, by dint of outward urban growth, become today’s middle-ring suburbs.  In rapidly growing regions, low density, single family neighbourhoods may even constitute today’s inner city (Kelly, Weidmann, and Walsh 2011).  This phenomenon may be particularly apparent in young metropolitan areas, such as Vancouver, which began 24  developing in a low density, suburban fashion almost right after settlement began (Punter 2003; Pettit 1993).  An interesting built environment pattern might be produced where, relative to the region as whole, residents of inner city neighbourhoods live in low density, single family home neighbourhoods built by previous generations, and consequently neither trade CBD proximity nor space.  In an extreme case, the marginal cost of land on the urban periphery today may be considerably higher than the marginal cost of land on the urban periphery of several generations ago.  This could result in contemporary outer suburban housing developments being built on smaller lots, or with a greater proportion of multifamily housing typologies, than more centrally-located former suburbs (Kelly, Weidmann, and Walsh 2011).  Bid rent theory also infers that the value of a former suburban location will increase as the city grows outward and the area gains accessibility to the core relative to newer, more peripheral development.  As land values increase in more central locations, developers of higher density homes would outcompete bids for single-family homes, leading to a declining density gradient from the central city to the outermost suburbs.  In practice, however, the density (or living space) of infill developments are profoundly affected by more than just the bid rent theory.  Some commentators have argued that land development is often constrained by land use planning practices that serve the desire of established homeowners (Fischel 2004, 1999; Epple, Thomas, and Filimon 1988).  Existing homeowners may shape land use decisions by voting for politicians who promise to preserve existing urban forms, using land use controls as an exclusionary means for discouraging new, less affluent arrivals (Ihlanfeldt 2004).  Apart from securing votes, municipalities can use exclusionary measures such as these to their advantage by reducing the level of services needed for a more dependent population group, and keep property taxes low (Fischel 2004, 1992). While land use regulations such as large lot zoning have typically been synonymous with exclusionary 25  land use motives (Pendall 2000), in built-out jurisdictions land use restrictions for exclusionary purposes may also extend to the construction of alternative housing typologies – often multifamily housing alternatives to existing single family homes5 (Schuetz, Meltzer, and Been 2011).  When centrally-located single family home neighbourhoods are protected, bid rent theory proposes that property values in these communities will rise to very high levels, prompting affluent home owners to capture as much value from their properties as possible (Kelly, Weidmann, and Walsh 2011).  For example, property owners may capture greater value by replacing modest-sized single family homes with much larger, more expensive homes that still fit within the parameters of existing zoning by-laws (Pettit 1993).  On the other hand, in the handful of inner city locations where building to higher densities is permissible, the dramatic increase in central city land value may outprice bids from all but the developers of the highest density residential products, such as high-rise condominium towers containing small multifamily dwelling units (Kelly, Weidmann, and Walsh 2011).  What may result, over time, is a bifurcation of the central city housing stock toward dwelling units at the very top and bottom of the dwelling space spectrum. Under such a scenario, the ability to live in the central city – where the majority of neighbourhoods with Smart Growth characteristics are typically located – may become restricted to either the very affluent, or to those households who are willing to trade almost all living space for the benefit of access.  In other words, Smart Growth communities, with their associated health, economic and environmental benefits and their growing popularity, may be increasingly restricted to a shrinking segment of the population.                                                    5 See, for example, Pennsylvania’s Girsh appeal of 1969 which stipulates that municipalities provide realistically available zones for all reasonable types of housing (Mitchell 2004). 26  2.2.2 Public Goods Sorting Apart from reducing travel costs, households may derive utility from a residential location based on the amenities available in neighbourhoods irrespective of their distance from a CBD.  While urban amenities are non-exclusionary goods6, the distribution and relative provision of these amenities may vary spatially across a metropolitan region.  As a result, households searching for a place to live may sort themselves in different neighbourhoods based on the utility they derive from the public goods that different communities provide.  This line of thinking is most closely associated with the early work of Charles Tiebout (1956).  Tiebout postulated that a metropolitan area, unencumbered by national borders and citizenship rules, provided relatively frictionless searching and moving costs, enabling households to sort themselves in a Pareto optimal fashion across the region.  Tiebout’s theory relied on a number of assumptions that seem incompatible with the real world.  Apart from the assumption that moving is frictionless and that amenities are of equal price (Hamilton 1975), households were assumed to have a perfect knowledge of the different choices that were available to them in the metropolitan housing market.  Tiebout (1956) also assumed that decisions were made based on the provision of one dominant good, rather than the unique bundle of preferences that each individual desires (Levine 2006; Rothenberg et al. 1991).  In particular, Tiebout neglected to appreciate the infinitely heterogeneous nature of housing, with each property providing a unique bundle of “housing services” (Malpezzi 2003) – not just the attributes of the property but the property’s location, and the amenities of the surrounding neighbourhood,  that add up to the utility the property provides for different households (Rothenberg et al. 1991, 210).  While people may                                                  6 That is, non-rival, non-excludable public goods where one person’s use does not diminish the availability of that good to others. 27  seek out neighbourhoods or locations, they ultimately put their preferences into action by buying or renting homes. Choosing a neighbourhood is inextricably linked to choosing among the selection of homes that are available within them, so the utility of a neighbourhood’s location or nearby amenities are bound into a “bundle of housing services” that, along with the structural attributes of the home, define the value and utility of a property (Malpezzi 2003).  In theory, an ideal way of achieving Pareto optimal residential matching would be to supply exactly those homes with housing service bundles that align exactly with the preference bundles representative of the individuals in a metropolitan area. That is, homes in exactly the ideal locations with the ideal structural and neighbourhood attributes (among other desirable factors) to exactly meet each and every consumer’s unique wishes. In practice, of course, such “Tiebout sorting” is impossible. Additionally, the matching of households with Smart Growth preferences to Smart Growth neighbourhoods may represent a special case where public goods sorting intersect with bid-rent theory.  Smart growth represents a public good, with Smart Growth attributes commanding more utility for some households than for others. Residents that place a high value on Smart Growth characteristics self-select into these communities, but their ability to settle in these neighbourhoods may be influenced by supply conditions described by bid-rent theory.  For example, many neighbourhoods with Smart Growth characteristics are older, pre-war areas of a metropolitan region that predate mass suburbanization and automobile-oriented development, and are therefore closer to the region’s core.  Many Smart Growth communities are therefore subject to many of the land development constraints experienced by areas with established homeowners exercising political influence over the feasibility of new development. 28  For people with preferences for Smart Growth communities, the barrier to residential matching may not lie in the total number of homes in these neighbourhoods, nor in the number of walkable neighbourhoods, but how well the attributes of the homes in these neighbourhoods align with their housing preferences.  Until now, studies that have measured Smart Growth mismatch, such as Levine and Frank (2006); Schwanen and Mokhtarian (2004); Cao (2008), and Levine, Inam, and Torng (2005) have focused on the supply of Smart Growth neighbourhoods, but not on the supply of different housing types within those neighbourhoods. While it will not ensure ideal matching conditions, providing a greater degree of housing mix in all neighbourhoods may represent a “second best” approach to Tiebout sorting.  It may be instructive to consider the search process as two separate processes: one where households first seek to maximize neighbourhood utility and, thereafter, seek to maximize dwelling utility.  Evidence from Clark and Smith (1982) and Smith and Clark (1982) demonstrate that most households search for homes in this two-step process.  Under a scenario where a metropolitan area contains a diverse assortment of neighbourhoods, each with a diverse assortment of housing types, households first sort among the communities that offer the closest match to their neighbourhood preference bundle and then select homes within these neighbourhoods based on their preference for property-specific attributes.  While households may not be able to find homes that perfectly match their preference bundle for property attributes, a greater diversity and distribution of housing types would enable households to trade off between a home’s living space or privacy and its affordability.  This action may enable a greater proportion of interested households to settle in the community of their choice.  Under a scenario where there is a diverse assortment of neighbourhoods, but little diversity in housing types, households still encounter no difficulties finding their preferred 29  neighbourhood choice, but are restricted in their ability to match homes within those neighbourhoods.  In many cases, the trade-off between space and affordability may not even present itself.  Returning to the bid-rent theory scenario where housing options in Smart Growth areas are split between very small and very large dwelling units, settlement may be out of reach for all but those small segments of the population who are willing to trade nearly all space and privacy or willing to pay any price.  With the rising demand for Smart Growth communities (Nelson 2013, 2006; Leinberger 2008) and the associated health and social benefits that Smart Growth communities confer, it is critical to understand structural barriers to matching people into these types of communities.  Housing mix may be one barrier, and is one of the few barriers to residential matching that planners can control.  The next section will explore how housing mix has been understood and conceptualized in planning research.  2.3 The Role of Housing Mix: Theories and Evidence The previous section suggested that housing mix at the neighbourhood level - provided a sufficient mix of housing typologies existed - could serve as a means by which people could sort themselves into the communities of their choice. What is known about “housing mix”, conceptually?  In this review of the academic literature, it is argued that housing mix – or, the diversity and distribution of housing typologies within a defined spatial area – while under-theorized, has long been tied to the idea of “place diversity” as a social good.  The concept of place diversity is reviewed before exploring the twin social equity goals of housing mix:  as a means to enable social mixing across disparate population groups, and as a means to enable residential preference sorting irrespective of social interaction.    30  2.3.1 Housing Mix as a Means to Achieve Place Diversity The idea that people of various backgrounds and enterprises of various functions should be spatially mixed remains a fundamental tenet of urban planning (Talen 2006).  This idea of “place diversity” (ibid) is certainly not new, and some of the most celebrated thinkers in urban planning and urban studies have espoused the idea.  Jane Jacobs devoted nearly a third of her seminal book, the Death and Life of Great American Cities to the goal of achieving diversity in city neighbourhoods (Jacobs 1961).  Lewis Mumford wrote extensively about the importance of creating social and economic mixing within communities (Mumford 1954), while Richard Florida has suggested that a diversity of human talents within metropolitan regions drives economic innovation and growth (Florida 2002).  “Diversity” is almost universally accepted as a positive force, and the construction of different housing typologies within a community have been considered as a means to achieve diversity for more than a century (Sarkissian and Heine 1978).  In a comprehensive review, Talen (2006) offers that securing place diversity for the purposes of achieving social equity may be divided into two different social goals: the goal of facilitating social mixing between diverse groups as an end in itself, and the goal of enabling people of different economic means to access the same place-based opportunities.  Housing mix has been offered as a way to achieve both social equity goals of place diversity in various policy documents, including countless municipal official community plans and growth strategies, as well as in various highly influential policy papers.  For example, greater housing mix at the neighbourhood level has been cited as a central plank of Metro Vancouver’s (2011) Regional Growth Strategy, the Smart Growth Network’s (2013) list of principles, and as one of the key principles that the Congress for New Urbanism (2001) lists in their Charter of the New Urbanism.  As Table 2, shows, the rationale for housing mix invokes the two, separate social 31  goals of place diversity.  But does housing mix lead to either social mixing or to better residential preference matching?  The academic evidence is considered in the next two sections.  Table 2: Selected references to neighbourhood housing mix in planning documents Source Aspect of housing mix invoked Quote Metro Vancouver (2011) Regional Growth Strategy Housing mix to support residential preference matching  “Complete communities are walkable, mixed use, transit-oriented communities where people can: find an appropriate place to live at all stages of their lives, earn a living, access the services they need, and enjoy social, cultural, educational and recreational pursuits. A diverse mix of housing types is fundamental to creating complete communities. This includes a mix of housing types and tenures that respond to an aging population, changing family and household characteristics and the full range of household incomes and needs across the region.”  Smart Growth Network (2013) Housing mix to support residential preference matching  “No single type of housing can serve the varied needs of today's diverse households. Smart growth represents an opportunity for local communities to increase housing choice not only by modifying land-use patterns on newly developed land, but also by increasing housing supply in existing neighborhoods and on land served by existing infrastructure. Integrating single- and multi-family structures in new housing developments can support a more diverse population and allow more equitable distribution of households of all income levels.”   Charter of the Congress of New Urbanism (2001) Housing mix to support social mixing “Within neighborhoods, a broad range of housing types and price levels can bring people of diverse ages, races, and incomes into daily interaction, strengthening the personal and civic bonds essential to an authentic community”       32  2.3.2 Housing Mix to Support Social Mix One argument for greater housing mix is to bring people of various economic and lifestyle backgrounds in close proximity with one another, thus allowing social linkages and personal interaction to form between individuals across diverse communities (Bridge, Butler, and Lees 2012).  Using different housing typologies to allow different groups – particularly groups of different incomes or social classes - to settle in the same community for the purpose of greater social interaction - is not new (Sarkissian 1976).  Master planned communities of the 19th century, such as the worker’s village created at Bournville, England by the Cadbury Company, sought to use different housing and tenure types to enable greater social mix between different classes of workers (Sarkissian and Heine 1978).  Cadbury’s intention was to help lower class workers imbibe some of the more “morally acceptable” lifestyle practices of their more upper class counterparts.  The idea that social mix between different classes was desirable persisted into the late 20th century, although it would no longer be explicitly grounded in terms that were as politically incorrect or patronizing.  The goal of social mixing would gain new currency with the rise of sociological studies showing that people with more social networks – i.e. “social capital” – had better opportunities to advance in their careers or professional development (Granovetter 1973), or in many different facets of life (Putnam 2000).  These arguments, among others, may have informed the move from predominantly low income housing projects to mixed income housing projects in the 1980s and 1990s, such as the Home Ownership for People Everywhere (HOPE) VI federal program in the United States (Joseph, Chaskin, and Webber 2007).  However, the evidence that social mixing occurs, or that the close proximity of different income groups leads to better social outcomes for lower income people remains thin.  Herbert Gans (1961), who observed the lives of residents in one of the most famously homogeneous and 33  unmixed suburban communities of its day - Levittown, New York - first cast doubt on the idea that social mix was possible in communities where housing developments were structured to enable a variety of residents of different economic means to reside there. Perrin and Grant (2014) sought to understand whether housing mix at the neighbourhood scale contributed to social mixing across income lines, and conducted semi-structured interviews with residents of diverse economic backgrounds living in a mixed typology housing development outside Halifax.  Their findings showed that residents spoke favourably about the principle of social mixing, but revealed that they, themselves, did not interact with people of different social classes on all but the most superficial levels.  Owners, in particular, expressed wariness of renters living in their midst.  These findings echo those found by Joseph and Chaskin (2010) in their own semi-structured interviews with residents of a mixed-income housing project in Chicago.  There, respondents of all income levels expressed frustration at the level of “coolness” between neighbours of different income levels.  In an example that would refute the notion of housing mix leading to social mix, one interview subject observed that the best way to be a “good neighbour” amidst people of radically different income and social backgrounds was to “keep to yourself” and avoid unnecessary contact.  The subject succinctly concluded to the interviewers “I don’t really think that’s what the objective of this whole project is” (ibid).    2.3.3 Housing Mix as a Means to Facilitate Residential Matching In contrast to housing mix to enable social mix, housing mix as a means to allow people to settle into communities remains both under-theorized and under-investigated.  Although the bid-rent theory and the theory of public goods sorting presented previously seems to underpin the idea that a greater range of housing options within neighbourhoods enables people of lesser 34  economic means to sort into their desired communities, almost no studies reference these concepts.  Few academic papers study the phenomenon of housing mix, even though it is cited as a development strategy in many planning documents (Metro Vancouver 2011; City of Vancouver 2013; City of Los Angeles 2009; City of Toronto 2010; Seattle 2005).  Analysis by Aurand (2010) demonstrates an association between housing mix and the prevalence of affordable housing units, but no study demonstrates an association between housing mix and actual settlement behaviour. Is housing mix, or the lack thereof, a real barrier to households locating into their desired neighbourhood? Is the lack of sufficient housing diversity a particular barrier to individuals wishing to settle in Smart Growth neighbourhoods, with their demonstrated health and social benefits?  These questions are addressed in this dissertation for the first time.  2.4 Conclusions Considerable evidence shows that Smart Growth communities confer numerous social, environmental and economic benefits to both the people who reside in them and to society, at large.  Beyond these normative benefits, Smart Growth communities are increasingly popular, driven by changing preferences in key demographic groups, such as the large and aging baby boom cohort.  Despite the benefits and popularity of Smart Growth communities, some commentators have expressed concern that the supply of appropriate homes within Smart Growth communities has not kept up with demand.  The normative benefits and growing demand of Smart Growth residential environments underscore the importance of matching people with preferences for these neighbourhoods into these communities.  Barriers to residential preference matching remain under-researched.  However, it is assumed that the ability for people to move into the communities of their choice is largely determined by their ability to afford to live in a 35  community, which includes a consideration of the trade-offs that are available to them in their decision-making.  Living space represents one such trade-off, and one of the few elements of a household’s decision-making calculus that can be directly impacted by planning.  Assuring that a neighbourhood has an adequate diversity and distribution of different housing typologies may enable a broad spectrum of the population to trade off living space in exchange for the ability to live in a Smart Growth neighbourhood with its demonstrated benefits and favourable preferences. Many planning documents endorse housing mix as an approach to facilitate better residential matching.  No studies have explored whether housing mix enables households to sort into the neighbourhood of preference.  This is the research gap this dissertation intends to help fill. 36  Chapter 3: Literature Review The previous chapter identified the research problem and provided theoretical context for why the degree of housing mix at the neighbourhood level was a possible, and understudied, predictor of the ability for households with Smart Growth preferences to sort into these neighbourhood types.  This chapter reviews the existing academic literature in the three areas of planning that support the three research questions of this dissertation.  Section 3.1, Smart Growth Preferences and Neighbourhood Matching, reviews the current state of research on measuring Smart Growth preference and neighbourhood matching.  The findings from these studies directly informed the main research question of this dissertation.  Section 3.2, Neighbourhood Satisfaction, reviews theories of neighbourhood satisfaction and provides an overview of studies that investigate the relationship between the urban built environment and neighbourhood satisfaction.  The findings from these studies helped inform the design of the second research question.  The third section, Neighbourhood Match, Neighbourhood Satisfaction and Health Outcomes (Section 3.3), reviews literature connecting neighbourhood matching and satisfaction levels to two measures of health: self-reported health status and BMI.  3.1 Smart Growth Preferences and Neighbourhood Matching The first and most important research question of this dissertation seeks to test the relationship between housing mix at the neighbourhood level and a household’s ability to sort into their preferred neighbourhood types.  In particular, finding strategies for households with Smart Growth preferences to match has been identified as a major research problem for which housing mix may be a potential, and as of yet, unstudied, approach.  This section reviews the literature on Smart Growth preferences and matching.  It begins by examining approaches to 37  measuring preference for neighbourhood environments before highlighting the findings of major studies that measured the degree to which Smart Growth neighbourhoods are preferred.  Finally, three studies of Smart Growth preference and neighbourhood match that were instrumental in informing the design of this dissertation are reviewed in more detail.  The section concludes with a discussion of how this dissertation builds on this existing body of knowledge.  3.2 Approaches to Measuring Smart Growth Preferences In order to understand the ability for an individual to settle in a Smart Growth neighbourhood, we must first measure the demand or preference individuals indicate for certain neighbourhood types over others.  Broadly speaking, there are at least three ways of estimating the preference for different residential environments (Table 3).  Table 3: Approaches to measuring preference for residential environments Method Methodological Approach Advantages Disadvantages Stated Preference Studies Household surveys of residential preferences Ability to gauge interest in a product that does not yet exist; Ability to more clearly understand personal motivations for choosing a product.  Misleading responses in metros with few examples of existing, middle class urban communities;  Survey design can lead to misleading responses  Revealed Preference Studies Hedonic Price Analysis of property attributes Ability to understand revealed preferences; what people actually choose, as opposed to what they say.   Ignores latent demand for products which may be undersupplied due to ‘market failure’. Target Market Method See, for example, Danielsen, Lang, and Fulton (1999) and Volk and Zimmerman (2000) Ability to predict preference in situations where products may not exist. Labour intensive; based on assumptions of behaviour.     38  The two most common approaches are to examine revealed preferences by analyzing existing market demand and to inquire about stated preferences through surveys (Bartholomew and Ewing 2011).  Each poses its advantages and disadvantages to measuring preference. Revealed preference techniques measure the willingness of people to pay for goods, ex post, by drawing conclusions from transactions that individuals have already made (ibid). In housing economics, revealed preferences can be measured by employing hedonic regression techniques of property transactions to disentangle the value of measurable housing services (i.e. attributes) from the aggregate value of the property (Rosen 1974). If, among a sample of properties, a measurable attribute of Smart Growth is demonstrated to be significant and positively associated with the value of the property, one can infer that Smart Growth is in demand (Bartholomew and Ewing 2011). This conclusion can be strengthened if repeat observations are made over time and the coefficient of the Smart Growth variable on property value is shown to increase faster (i.e. inflate) than that of the other variables. The primary advantage of revealed preference studies over stated preference surveys is that they provide evidence of actual choices made under real world circumstances, not merely evidence of what individuals say they might choose without being placed in a position of constraint (ibid). However, while revealed preference studies provide evidence that a particular product is valued; they cannot uncover the reasons for why that product was chosen. An individual may pay a premium for a home in a Smart Growth neighbourhood, but the motivation for paying that premium may be completely unrelated to the Smart Growth characteristics of the surrounding area. For example, the home may also be near the respondent’s relatives, or perhaps the individual secured a favorable lending rate on his or her mortgage. These factors – which affect attitudes of the individual and not attributes of the property - could be as likely, or more likely, to influence the value of a property than the Smart 39  Growth character of the neighbourhood. More importantly, revealed preference studies cannot estimate preferences for products that do not yet exist. Revealed preference studies ideally require products to be part of well-functioning markets in which the range of choices are fully available to all consumers (Francescato 2002). However, the premise that underlies this dissertation is that the supply of different housing types in Smart Growth neighbourhoods is neither well-functioning nor fully available to all consumers (see Section 2.2.2), and there may be considerable latent demand for homes in Smart Growth neighbourhoods made inaccessible through insufficient neighbourhood housing mix. But revealed preference studies cannot, by their nature, capture latent demand (Bartholomew and Ewing 2011; Volk and Zimmerman 2000). Stated preference surveys have the advantage of being able to measure preferences for products that do not meet ideal market conditions. Moreover, personal motivations for choosing walkable neighbourhoods can be more easily queried. The primary disadvantages of stated preference studies are that they do not provide evidence of what people choose under real world conditions where costs and constraints are imposed, and that, without familiarity with the product being queried, individuals are likely to provide misleading responses (Bartholomew and Ewing 2011). The latter consideration is particularly important in trying to estimate preferences for walkable urban housing products in the United States where, until recently, there were few examples of desirable walkable residential areas catering to the middle class outside of a handful of major cities (Volk and Zimmerman 2000). While these disadvantages to stated preference studies persist, they have been ameliorated to some extent by advancements to survey design as well as to sampling strategies. Familiarity with walkable communities may be addressed by locating the study in a metropolitan region where urban living has a long history of desirability, and where a range of housing types 40  exist across a wide geographical area to suit a diverse assortment of tastes. The costs and constraints inherent in making choices can be accounted for by phrasing questions of preference to incorporate a plausible trade off involved between selecting an urban versus a suburban product. For example, homes in more central city, walkable locations that are equivalently priced to their suburban counterparts often involve a trade-off in space or privacy. More sophisticated studies, such as Levine and Frank (2006), which informed the design of the survey instrument used in this study, use choice-based conjoint analysis techniques (Boyle et al. 2001), attempting to understand the combination of different home and neighbourhood attributes that affect overall preference for a neighbourhood product. A series of trade off questions are posed to inquire into the preferences for different attributes, and following a correlation test, the variables are reduced to a principal component that proxies for overall preferences for walkable home-neighbourhood products. In contrast, studies like Allison (1997), National Association of Homebuilders (1999) or National Association of Realtors (2011) draw their neighbourhood preference conclusions from a single trade-off question, and therefore might only inquire into one facet of neighbourhood preference (e.g. interior space, but not access to parks or recreational facilities, nearby retail options, etc.). An alternative to incorporating trade-offs is to avoid direct questions of preference for neighbourhoods, and to ask more concealed questions about a respondent’s lifestyle and a respondent’s attitude toward characteristics implied by urban neighbourhoods (Schwanen and Mokhtarian 2005; Schwanen and Mokhtarian 2004; Cao 2008). A third - and by far the least common – approach to measuring residential preference is to gather information on the socioeconomic, demographic and migration characteristics of households in a region and predict the market potential of a neighbourhood type based on the revealed behaviour of people with similar characteristics in other contexts (Volk and 41  Zimmerman 2000; Danielsen, Lang, and Fulton 1999). Proponents of this “target market method” (TMM) argue that this technique avoids some of the pitfalls of revealed preference surveys, such as the ability to infer demand for a product that does not yet exist, as well as some of the disadvantages of stated preference surveys, such as the inability for respondents to make a judgment on an unfamiliar product. In the opinion of this researcher TMM poses numerous drawbacks. The approach infers behaviour of a group based on a reference group’s behaviour in another context, rather than observing the behaviour of respondents directly. TMM is also very data intensive; Volk and Zimmerman’s approach relies on measuring over 500 variables of a household’s socioeconomic status, demographics and culture (ibid).    3.2.1 Preference for Smart Growth Neighbourhoods: Evidence Table 4 lists a selection of stated residential preference surveys that have sought to examine the preference for walkable urban environments, arranged in chronological order. From casual observation, the table demonstrates several trends in residential preference research for walkable environments. First, that there has been some movement toward increased preferences for Smart Growth alternatives to suburban residential neighbourhoods in the American population over time. Secondly, the higher rates of preference for Smart Growth areas in regions where urban living has been historically desirable, such as Boston and Vancouver. Most importantly, however, survey design and sampling strategies are increasingly more rigorous compared with earlier studies from the 1990s.  Nelessen’s (1994) series of visual preference surveys, for example, reported that 88% of over 50,000 respondents across the US showed a preference for Smart Growth over suburban communities.   42  Table 4: Selected stated preference surveys of Smart Growth environments Study Location1 N Smart Growth Preference (%) Trade-off questions posed? Mismatch investigated?2  Nelessen (1994)  U.S.  50,000+  88%  No  No  American LIVES (1995)  California, Texas, Florida, Colorado and Washington   566  30%  No  No Allison (1997) Vancouver, Canada  611 56% Yes No Professional Builder (1998)  U.S.  Unknown 37% Yes No National Association of Homebuilders (1999)  U.S. 2,000 17% Yes No Schwanen and Mokhtarian (2004) San Francisco Bay Area  1,358 25% n/a Yes Levine, Inam and Torng (2005)  Boston and Atlanta 1,607 40% (Boston); 29% (Atlanta) Yes Yes Levine and Frank (2006)  Atlanta 1,455 25%3 Yes Yes Cao (2008) 8 Northern California communities  594 n/a n/a Yes National Association of Realtors (2011) U.S. 2,071   56% Yes No Frank et al. (2012) Vancouver and Toronto, Canada 2,748 83% (Vancouver); 82% (Toronto) Yes Yes       1 Cities listed are metropolitan areas unless otherwise specified. 2 Study investigates whether respondents’ neighbourhood of residence is mismatched with their preferences. 3 Unweighted averages of several preference questions.  43    Those surveys, however, did not correct for the self-selection of respondents in their sampling strategies, and so an abnormally high number of people with pre-existing preferences for Smart Growth may have been attracted to filling out the survey and may have encouraged like-minded peers to do so as well (Leinberger 2008, 93).  Similarly, several residential preference studies were collected in cities where Smart Growth already predominated, such as Portland, Oregon, Midtown Atlanta, and Aurora, Colorado where the respondents, though randomized, may have been biased toward Smart Growth, having already settled in these communities (Malizia and Exline 2000).  Conversely, large, national studies such as the “Smart Growth: Building Better Places to Live” survey (National Association of Homebuilders 1999),  may have over-sampled single family homeowners  and, as a result, reported attributes that may more accurately reflect this demographic, such as a greater willingness to sacrifice shorter commuting times than a smaller lot in exchange for greater affordability.   In some visual surveys, such as one administered in Fort Collins, Colorado (not shown in Table 4) participants responded most positively to the landscaping and greenspace features depicted in the visual examples of Smart Growth neighbourhoods, rather than to features that planners typically ascribe to the movement, such as enhanced accessibility to stores and services, pedestrian design, higher densities, or improved access to public transit (Malizia and Exline 2000).  These findings underscore two additional problems with these studies. First, respondents may inadvertently be drawn to suburban elements presented in visual cues.  More importantly, though, many stated preference surveys are biased because they fail to present complete information on the consequences of choosing one neighbourhood alternative over another (Bartholomew and Ewing 2011).  Considerable evidence has been amassed in the field of transportation planning that decent public transit services and the variety of retail and service 44  opportunities that are demanded in Smart Growth environments can only be viable if a neighbourhood has sufficiently population density to support these amenities (see, for example, Cervero and Kockelman (1997), but also many others).  Suggesting that a neighbourhood could support all the amenities and transit access of a Smart Growth neighbourhood, but not invariably involve trade-offs in property size and privacy would be to present a false choice. A more recent trend in stated preference surveys was initiated by Schwanen and Mokhtarian (2004) who not only measured the degree of household preferences for walkable neighbourhoods, but also identified whether households who reported preferences for one type of environment were ‘mismatched’ and lived in another. Their study revealed that 25% of suburbanites actually reported preferences for Smart Growth communities (ibid). Three additional studies have quantified the mismatch between Smart Growth preferences and actual choices, and have attempted to infer whether there is an undersupply of housing available in these neighbourhoods (Levine, Inam, and Torng 2005; Levine and Frank 2006; Cao 2008). Levine, Inam, and Torng (2005) and Levine and Frank (2006) suggest that there is an insufficient supply of homes, while Cao (2008) refutes this.  Since these three studies directly inform the research design of this dissertation, their methods and results will now be examined in some detail.   3.2.2 Three Studies of Residential Preference and Mismatch Three academic studies eliciting preferences for Smart Growth stand out for their robustness, incorporating trade-offs and more sophisticated analyses than many of the professional consultant studies that predate them.  In addition, these studies investigated the 45  degree to which households who lived in one neighbourhood type – Smart Growth or suburban –  were ‘mismatched’ with the neighbourhood type of their preference.  Levine, Inam, and Torng (2005) used a phone survey of 800 households in metropolitan Boston and Atlanta to understand preferences for residential neighbourhood types, and to determine if households settle in neighbourhoods that are mismatched with their stated preferences.  Respondents were classified according to the traffic analysis zone (TAZ) they resided in, and cluster analysis was used to categorize each TAZ as one of 5 neighbourhood types, each considered by the authors to possess different degrees of measurable Smart Growth attributes.  Survey questions queried respondents on their choice of one neighbourhood over the other, and were phrased as inevitable trade-offs, typically involving a compromise in accessibility to transit and walkable stores and services in exchange for greater dwelling and garden space (ibid).  To see how individual households chose from identified neighbourhood types, and whether their actual place of residence was mismatched with their neighbourhood preferences, the authors ran a series of multinomial logit models predicting the likelihood of residing in the 4 most Smart Growth-like neighbourhood  types relative to the most suburban environment, controlling for neighbourhood specific constants (i.e. no data on neighbourhood preferences) and people’s stated preference scores for Smart Growth.  The model tested the “closeness of fit” between people’s preferences for Smart Growth and the neighbourhood environment in which they actually resided; a negative coefficient could thus be interpreted as a greater preference for Smart Growth.  All models but the non-white Boston sample of respondents carried negative coefficients; the incorporation of preference data to the model added greater explanatory power to the Boston models than the Atlanta models, suggesting that Bostonians with preferences for Smart Growth had a better ability to match themselves to the 46  environments they desired than comparable Atlantans (ibid).  Levine, Inam, and Torng (2005) conclude that if metropolitan Atlanta - a rapidly growing city with the vast majority of its residential stock built during the postwar, automobile-oriented age – possessed more housing stock in Smart Growth neighbourhoods, there is a likelihood that many households with preferences for these types of environments would not be as mismatched as they currently are. Levine and Frank (2006) also tested the mismatch between respondents’ stated preferences and their actual neighbourhood environments, although their approach differed in several key respects.  Instead of a phone survey of binary choices, data was collected through a paper-based survey of 1,455 Atlanta area households.  The survey presented subjects with a series of trade-off scenarios between Smart Growth and suburban residential environments, and asked them to both indicate their preferred neighbourhood type as well as to indicate whether the neighbourhood they would hope to move to was more Smart Growth or suburban-oriented than their current community.  Both questions were posed using identical 10 point ordinal scales with the midpoint (i.e. a value of 5) set to indicate either indifference to Smart Growth versus suburban environments or no desire to change neighbourhood type in a hypothetical move.  The scores from these two questions were subtracted from one another to calculate each respondent’s “direction of move score”.  A direction of move score of 0, in other words, indicated no desire for change in neighbourhood orientation.  Instead of a discrete choice model testing the likelihood of choosing one environment over another, the authors gathered response scores to the preference questions and ran a principal component analysis to extract a single factor,  labeled the “neighbourhood preference” factor, with which all the other variables were highly correlated.  This factor was then normalized to set neutral preferences for Smart Growth versus suburban neighbourhoods to zero (such that a negative score indicated greater preference for Smart 47  Growth), and the distribution of “neighbourhood preference” scores was tabulated by decile.  In a region with sufficient housing options in Smart Growth areas, the point at which “direction of move” score is lowest should theoretically occur at the 50th percentile of neighbourhood preference scores in the population7.  Instead, this occurs at the 30th percentile (or 70th percentile in favour of automobile-oriented environments), and this finding is replicated among whites, non-whites, low-income earners and high-income earners (ibid).  Levine and Frank (2006) use this as evidence that housing in Smart Growth neighbourhoods in the Atlanta area is underserved relative to demand. Finally Cao’s (2008) survey of 594 recent movers to urban and suburban neighbourhoods in Northern California cities provides further evidence of whether homes in Smart Growth neighbourhoods are undersupplied.  Arguing that stated preference surveys of generic households may bias results towards the environment in which people already reside, Cao chose to survey recent movers on their preferences for neighbourhood environments and to compare how closely those built form characteristics existed in the neighbourhood they moved to (ibid). The mismatch between preferences and perceptions for neighbourhood characteristics was calculated by taking the difference between the two t-tests; if the value was positive (i.e. if preference was greater than perception), it was argued that recent movers with a preference for Smart Growth environments had experienced a mismatch.  Cao reports the paired t-test comparisons for all variables where differences were significant at p=0.10 or greater, showing that, in almost all cases - and for both suburban and urban movers -perceptions of the built environment exceeded preferences.  He uses this as evidence that Smart Growth communities are                                                  7 That is, neutral preference toward Smart Growth or suburbia. 48  not undersupplied, at least in Northern California (ibid).  However, the choice of urban communities studied in Cao (2008)  included more central, pre-war areas of Mountain View, Sacramento, Santa Rosa and Modesto, California; bedroom communities of the Bay Area.  While these neighbourhoods contain measurable built form attributes that align with Smart Growth principles (viz. Smart Growth Network 2013), in a non-technical sense, how well do they proxy for the latent qualities of “big city lifestyles” that attract many people to urban neighbourhoods? It is likely that the qualities that attract people to an older neighbourhood of Modesto are considerably different from the qualities that attract people to a neighbourhood in San Francisco or Manhattan.  This raises philosophical questions about whether the degree to which the qualities that make communities desirable to certain groups of people can be measured through their built attributes alone.  The selection of residents who moved to the neighbourhoods in Cao (2008) may also not have been representative of the kind of residents who make deliberate choices to trade space and affordability to reside in Smart Growth neighbourhoods in larger cities.    3.2.3 Research Gaps As was described in Chapter 2, the studies reviewed in the previous section measured the degree to which people were mismatched away from their neighbourhood preferences, but did not investigate possible explanations as to why people were mismatched.  This dissertation not only examines housing mix as a potential barrier to matching, but also examines whether housing mix is a barrier to the ability for individuals to match into what they perceive to be a Smart Growth environment versus their ability to match into an objectively-defined Smart Growth environment.  A comparison between the perceived and actual built environment is 49  crucial, not only because objectively-measured Smart Growth environments have been associated with measurable health benefits (see review, section 1.4), but because objective definitions of the environment based on measurable attributes can guide planning practice.  Moreover, separate outcomes of housing mix on subjective and objective neighbourhood match can be followed with an analysis of how these different measures of match affect individual wellbeing.  In this way, the consequences of mismatch can be measured directly and the consequences of housing mix policies can be measured indirectly.  The first measure of wellbeing tested in this dissertation is neighbourhood satisfaction, which is explored in the next section.  3.3 Neighbourhood Satisfaction A further purpose of this dissertation is to examine how neighbourhood housing mix, moderated by the ability for people to match their neighbourhood preferences, is associated with the satisfaction people report with their neighbourhood of residence.  The rationale is to understand whether the neighbourhood housing mix actually has a bearing on an individual’s wellbeing.  In this section, the concept of “satisfaction” and “neighbourhood satisfaction” are defined, and existing research on neighbourhood satisfaction is introduced.  This section concludes by examining the contribution that this dissertation makes to planning research on individual satisfaction with residential environments.  3.3.1 Definitions of Satisfaction and Neighbourhood Satisfaction What does the term ‘satisfaction’ imply? ”Satisfaction” with one’s environment has been variously described as the degree to which people “feel” the environment helps them achieve 50  their “goals” (Canter and Rees 1982); as the “meaning” that people “obtain” from their everyday experiences (Barthes 1967); as an “affective response” to the environment (Weidemann and Anderson 1985), and as one’s “attitude toward” their residential environment (Francescato, Weidemann, and Anderson 1989). Each of these descriptions imply the act of assessing one’s surroundings, but do not hint at how those assessments are made, nor what psychological feelings are registered in the act of assessing, nor how one’s satisfaction with neighbourhood is part of one’s overall satisfaction with life.  Perhaps the most in-depth analysis of the concept of satisfaction is the comprehensive study undertaken on American households by Campbell, Converse, and Rodgers (1976) in the late 1960s.   Campbell’s study made significant contributions to uncovering the psychology behind satisfaction, to research methods for measuring satisfaction, and to understanding how neighbourhood satisfaction impacted overall life satisfaction.  While Campbell’s study was chiefly concerned with defining individual life satisfaction, much of his empirical evidence was, quite conveniently for this research project, based on his examination of personal satisfaction derived from residential environments.  Through a battery of interviews with over 200 subjects, Campbell concluded that ‘satisfaction’ may be defined as the psychological state that results from the difference of a respondent’s comparison of their current situation with the situation they aspire to.  Aspiration, in turn, is defined as a respondent’s ideal situation, constructed using multiple, personal frames of reference regarding what individuals might constitute as an ‘ideal’.  Among these frames of reference, comparison to one’s own ‘best experience’ in memory is the strongest predictor; comparison to the situation of close relatives, the “average” American, and close friends were also strong predictors.  In Campbell’s conceptualization, neighbourhood satisfaction is just one of at least three realms that contribute to residential satisfaction, itself just 51  one contributor among several domains to life satisfaction (See Figure 4).  However, Campbell argued that the direction of influence is not linear, and one’s life satisfaction may influence one’s residential satisfaction or neighbourhood satisfaction. Campbell’s definition of residential satisfaction has been influential on many subsequent studies of neighbourhood satisfaction (Basolo and Strong 2002; Lovejoy, Handy, and Mokhtarian 2010; Parkes, Kearns, and Atkinson 2002; Sirgy and Cornwell 2002), and is the definition that informs this study.   Figure 4: Neighbourhood satisfaction as it relates to overall life satisfaction   Notes: Adapted from Campbell, Converse, and Rodgers (1976) The relationship tested in the second research question of this dissertation is whether an individual’s neighbourhood match – which measures the difference between one’s current Life SatisfactionResidential SatisfactionCommunity SatisfactionDwelling SatisfactionOther FactorsNeighborhood Satisfaction52  neighbourhood situation and the neighbourhood they aspire to live in – predicts their neighbourhood satisfaction, controlling for one’s satisfaction with other attributes of the residential environment.  The second research question directly tests Campbell’s theory, but also considers the impact that satisfaction with attributes of the property and the neighbourhood may play in overall neighbourhood satisfaction.   3.3.2 Existing Research While numerous studies have examined social determinants of neighbourhood satisfaction, such as perceived crime (Robinson et al. 2003) and social capital (Dassopoulos et al. 2012), built environment predictors of neighbourhood satisfaction are less well explored (Hur, Nasar, and Chun 2010).  An extensive literature search was conducted, filtering for peer-reviewed academic studies that:  1) Used quantitative methods to measure attributes of the built environment, or transport access to environments;  2) Were published in the last 30 years (since 1985);  3) Treated satisfaction with one’s neighbourhood of residence as the dependent variable, and, 4)  Were investigated in English-speaking countries (US, UK, Canada, Australia, New Zealand, Ireland) because of their long histories of private homeownership and low density suburban residential development.   Based on these criteria, only 14 studies were identified (Table 5).     53  Table 5: Existing studies of built environment predictors of neighbourhood satisfaction Study Location N Basolo and Strong (2002) New Orleans, Louisiana        125  Buys and Miller (2012) Brisbane, Australia         638  Chapman and Lombard (2006) United States (American Housing Survey, 2003)     55,000  Cook (1988) Metropolitan Minneapolis, Minnesota         449  Ellis, Lee, and Kweon (2006) College Station, Texas         122  Gruber and Shelton (1987) 11  North Carolina counties         305  Ha and Weber (1994) United States (not specified)      1,041  Howley (2009) Dublin, Ireland         270  Hur and Morrow-Jones (2008) Franklin County, Ohio      2,060  Hur, Nasar, and Chun (2010) Franklin County, Ohio         725  Lovejoy, Handy, and Mokhtarian (2010) 8 Northern California communities      1,682  Lu (1999) United States (American Housing Survey, 1989)     55,000  Mohan and Twigg (2007) 335 United Kingdom communities     13,390  Parkes, Kearns, and Atkinson (2002) England (English Housing Survey)     18,361      Many of these papers focused on metrics such as the quality of the housing structure (Basolo and Strong 2002; Cook 1988; Ha and Weber 1994; Howley 2009; Lu 1999) or the role of neighbourhood aesthetics (Chapman and Lombard 2006).  Measurable features of Smart Growth remain relatively understudied, with only one study focused explicitly on measuring whether a Smart Growth residential environment is a significant predictor of neighbourhood satisfaction (Lovejoy, Handy, and Mokhtarian 2010).  Table 6 lists those studies that measure how attributes of the built environment typically championed by Smart Growth impact neighbourhood satisfaction, and the relationship uncovered in their analysis.   54  Table 6: Previously studied Smart Growth built environment predictors of neighbourhood satisfaction Study Land Use Mix Quality of Pedestrian Facilities Housing Type of Resp-ondent Residential Density (Subjective) Residential Density (Objective) Yard Size Smart Growth Character         Buys and Miller (2012)  (+)      Gruber and Shelton (1987)    (–)  (+)  Hur and Morrow-Jones (2008)    (–)    Hur, Nasar, and Chun (2010)    (–)    Lovejoy, Handy, and Mokhtarian (2010) NS    (+) NS (+) Mohan and Twigg (2007)   (–)1     Parkes, Kearns, and Atkinson (2002)   NS             (+) = significant and positive relationship with satisfaction; (–) = significant and negative relationship with satisfaction; NS = not significant.  1. Reference group is single family homes.   As expected, the quality of pedestrian facilities (sidewalks, road crossings, benches, etc.) was a significant and positive predictor of neighbourhood satisfaction in an Australian study conducted by Buys and Miller (2012).  In a British study, residents of alternative housing types to the single family home, such as rowhouses, reported significantly lower levels of neighbourhood satisfaction (Mohan and Twigg 2007), although this was not replicated in an earlier study, also in the United Kingdom (Parkes, Kearns, and Atkinson 2002).  Three studies reveal that respondents who perceive that they live in dense neighbourhoods report lower levels of neighbourhood satisfaction (Hur, Nasar, and Chun 2010; Hur and Morrow-Jones 2008; Gruber and Shelton 1987).  However, when residential densities are measured objectively, rather than inquired of residents, respondents who lived in denser neighbourhoods actually reported higher 55  levels of neighbourhood satisfaction (Lovejoy, Handy, and Mokhtarian 2010).  Yard size – a feature associated with suburban living, as opposed to Smart Growth residential environments – was significantly related to satisfaction in an early study by Gruber and Shelton (1987), but was shown to be insignificant in a more recent study by Lovejoy, Handy, and Mokhtarian (2010).  In both cases, yard size was measured subjectively by inquiring with respondents about their satisfaction with the size of their yard.  Finally, residents who lived in communities with Smart Growth characteristics were significantly more satisfied with their neighbourhood than residents of suburban communities (ibid).    3.3.3 Research Gaps Despite the volume of advocacy for Smart Growth planning, understanding whether households who live in Smart Growth are more satisfied with their environment remains relatively understudied.  Apart from research conducted by Lovejoy, Handy, and Mokhtarian (2010), no other studies were encountered that explicitly investigate the difference in satisfaction levels of households who live in Smart Growth communities versus suburban ones.    This dissertation attempts to make several contributions to the field of neighbourhood satisfaction research, building off existing studies – particularly that of Lovejoy, Handy, and Mokhtarian (2010).  Primarily, the relationship between the ability to match neighbourhood preferences to one’s neighbourhood of actual residence and neighbourhood satisfaction is tested.  Additionally, this relationship is tested on both a subjective and an objective measure of neighbourhood match, so the discrepancy between an individual’s perception of their environment and the actual environment (based on objective measures of built form) is compared to see if discrepancies in satisfaction levels appear.  Previous studies have either studied the 56  objective environment’s effect on satisfaction levels (ibid) or the subjective environment’s effect (Hur, Nasar, and Chun 2010; Hur and Morrow-Jones 2008; Gruber and Shelton 1987), but the two have not been tested on the same sample.  Additionally, this dissertation tests another facet of Smart Growth and its effects on satisfaction levels.  Housing mix – which, as demonstrated in Chapter 2, is a major goal of numerous Smart Growth policies – has nevertheless never been analyzed as a predictor of neighbourhood satisfaction.    3.4 The Relationship between Neighbourhood Match, Satisfaction and Health Outcomes The final research question explores the pathway from neighbourhood match to two health outcomes: self-reported health status and BMI.  This relationship is tested directly, but is also tested with neighbourhood satisfaction serving as a moderating variable.  Unlike the first two research questions, this dissertation is not expected to make any contribution to the existing body of knowledge on how neighbourhood environments shape health outcomes.  As will be described, the rationale for including a question on downstream health outcomes is to provide evidence to support (or refute) whether housing mix should be pursued by policymakers.  It is argued that health benefits of housing mix should be more convincing than satisfaction itself because it provides an argument for a greater public benefit.  Existing research linking neighbourhood match and neighbourhood satisfaction to these two health outcomes are explored in this section.  3.4.1 Self-Reported Health Status “Self-reported health”, also known as “self-rated health”, is an indicator used by Statistics Canada in its Canadian Community Health Survey (Employment and Social Development 57  Canada 2014) that asks respondents to rate their own health compared to people of the same age on a 5 point ordinal scale from “poor” to “excellent”.  It does not distinguish between physical and mental health, but is meant to complement more objective health measures, such as mortality rates and life expectancy, by accounting for the incidence and severity of unmeasured diseases (ibid).  The self-reported health indicator has been demonstrated to yield reliable and consistent results in surveys (Miilunpalo et al. 1997) and has been shown to significantly predict objectively measurable health events such as mortality (DeSalvo et al. 2006; Burström and Fredlund 2001).  The self-reported health measure is not without flaws, however.  Controlling for socioeconomic, demographic and health factors, subjects with poor literacy skills have been shown to rate their health significantly lower (Baker et al. 1997).  Meanwhile, there is psychological evidence that respondents may be overly optimistic in assessing their health status compared to their peers (see a review by Dunning, Heath, and Suls (2004)), although this phenomenon has not explicitly been tested in a study of the self-reported health indicator. Given that “neighbourhood matching” – conceived as the ability to match one’s neighbourhood type to one’s preference along a continuum of Smart Growth characteristics – is only a concept that has gained recent traction in the field of urban planning, it is not surprising that the direct link between neighbourhood match and self-reported health status has not been studied in the health sciences.  This dissertation proposes that this relationship is insignificant unless moderated through neighbourhood satisfaction; individuals who match are expected to report better neighbourhood satisfaction outcomes and, in turn, report better self-reported health outcomes.   The association between neighbourhood satisfaction and one’s perception of health rests on considerably more evidence.  In a study conducted in Metro Vancouver, Collins, Hayes, and 58  Oliver (2009) found that neighbourhood satisfaction was a significant predictor of self-reported health status.  Stronegger, Titze, and Oja (2010) also identified a significant relationship between an individual’s satisfaction with the built infrastructure in their community and their self-rated health, and inferred that residents who were more satisfied with local infrastructure, such as recreational facilities and public transport, engaged in higher levels of physical activity than those that held low opinions of such amenities.  Leslie and Cerin (2008) found a significant association between individuals’ satisfaction with different neighbourhood attributes and their self-reported mental health.  Another study, conducted in the City of Vancouver, found no significant relationship between neighbourhood or housing satisfaction and either mental health or self-reported health when personal characteristics such as stress levels, education or length of residence were introduced into the analysis (Dunn 2002).  3.4.2 Body Mass Index BMI, first developed in the 19th century by French Mathematician Adolphe Quetelet, has become the most widespread measure of obesity in both individuals and the population (Gallagher et al. 1996).  Unlike alternative obesity indicators such as the quantity adipose tissue mass, BMI benefits from being extremely straightforward to measure: BMI is simply the weight of an individual (kg) divided by their height, in meters, squared (i.e. kg/m2).  For these reasons, a measure can be inquired of people in the general population simply by asking patients or survey respondents to report their height and weight (ibid).  By 1998, after prominent multiyear studies tracking BMI revealed a growing obesity epidemic in the American population (Flegal et al. 1998), a BMI of 25.0 – 29.9 kg/m2 became established as the definition of an “overweight” 59  individual and a BMI of 30 kg/m2 and above became established as the definition of “obesity” in standard American medical practice (National Heart Lung and Blood Institute 1998). Given the considerable attention that has been given to researching contributors to obesity, many studies have already examined the link between the urban environment in which people reside and their BMI outcomes.  Briefly, lower neighbourhood socioeconomic status has been associated with higher BMI scores in Canada (Matheson, Moineddin, and Glazier 2008) and the US (Wang et al. 2007).  Perceptions of neighbourhood safety have also been significantly linked to BMI, with individuals who perceive their neighbourhood to be less safe reporting higher BMIs (Fish et al. 2010) and parents reporting their neighbourhood as unsafe linked with higher BMIs in children (Lumeng et al. 2006).  Access to nutritious food in urban neighbourhoods also may play a role in obesity; Morland, Roux, and Wing (2006) found that the absence of supermarkets and the presence of small grocery stores (i.e. convenience stores) explained obesity rates in older adults.  The Smart Growth or suburban built form of neighbourhoods also has been extensively studied for its association with BMI scores.  These associations relate to the amount of physical activity people engage in as part of their travel patterns in different neighbourhood environments.  Frank et al. (2006) showed that measurable increases in neighbourhood walkability (i.e. Smart Growth character of the neighbourhoods)  in which people resided were significantly and strongly related to the amount of time people spent walking and, therefore, measurable reductions in their BMI.  These findings were echoed in a study by Brown et al. (2009), who looked at the mix of land uses and proximity to rapid transit stations and found these, along with the age of the neighbourhood (i.e. pre-automobile design) to be associated with BMI.  Refuting these findings somewhat, Lovasi et al. (2008) found higher population densities, land use mixing and greater transit access to be associated with lower BMI, 60  but only among traditionally privileged segments of the population  such as non-Hispanic Whites and individuals with more education and higher incomes.  They conclude that there may be other social and economic barriers to leading active lifestyles among more marginalized groups that supersede the built form of the communities in which they live (ibid). Most importantly for this study, Frank et al. (2007) were the only researchers to study the effect of matching neighbourhood types (suburban and Smart Growth) to preferences on BMI.  Individuals who preferred Smart Growth and lived in Smart Growth (i.e. Smart Growth matched) reported the lowest rates of obesity, followed by individuals who preferred suburbia but lived in Smart Growth.  Individuals who lived in suburbia, regardless of whether they preferred suburbia (i.e. matched) or Smart Growth (i.e. mismatched) were nearly twice as likely to be obese as individuals who were Smart Growth matched, although the difference between people living in suburbia based on their preferences was minor.  Frank et al. (2007) suggest that their study provides some evidence that the built environment people find themselves in, rather than their preferences, may better predict walking behaviour and health outcomes, although they acknowledge that causality can only be predicted using a longitudinal study design.  Although their study did not incorporate a health measure, Schwanen and Mokhtarian (2004) also found that the neighbourhood environment, rather than preference, explained travel behaviour and the amount of time spent driving.   Studies linking neighbourhood satisfaction directly to BMI were not encountered in this review, although significant contributors to neighbourhood satisfaction such as perceptions of crime (Buys and Miller 2012; Howley 2009; Basolo and Strong 2002) and access to recreational facilities (Hur, Nasar, and Chun 2010; Chapman and Lombard 2006) are also significant predictors of BMI outcomes (Fish et al. 2010; Lumeng et al. 2006; Björk et al. 2008).  For this 61  dissertation, there is an expectation that neighbourhood satisfaction will have a significant, negative, association with BMI but that this relationship is almost entirely based on predictors of neighbourhood satisfaction such as safety and access to recreational facilities, rather than on satisfaction itself.    3.4.3 Implications of investigating Research Question 3  The link between neighbourhood matching and BMI and the link between neighbourhood satisfaction and self-reported health status have already been investigated.  The link between neighbourhood satisfaction and BMI and neighbourhood match and self-reported health status have not been investigated, but given that these relationships are either expected to be insignificant or significant because of upstream, already-investigated factors, these may be rather banal research questions.  If a research gap is not being resolved, why are health outcomes studied in this dissertation?  The justification for including health measures as dependent variables is that, if a significant pathway can be established between the degree of housing mix in a respondent’s neighbourhood and positive health outcomes, normative support is lent to policies that encourage greater housing mix.  As indicators of well-being, self-reported health status and BMI may be more morally persuasive than “neighbourhood satisfaction”.  On its own, “dissatisfaction” might be interpreted as one’s negative reaction to the inability to fulfill demand for a luxury product.  Subsidizing exotic vacations and sports cars may raise societal “well-being”, but are difficult for policymakers to justify under conditions of finite resources.  Levels of “neighbourhood satisfaction” among individuals who successfully sort into the urban, walkable neighbourhoods of their choice is a weak argument for supporting Smart Growth policy measures, unless 62  measures of satisfaction can be tethered to a more normatively convincing argument. “Good health”, in contrast to satisfaction, is almost universally recognized as a basic need or a fundamental human right (Gruskin, Mills, and Tarantola 2007; UN General Assembly 1948), and so the inclusion of its association, if demonstrated to be significant, strengthens the argument for greater housing mix.  63  Chapter 4: Methods and Research Design This chapter details the methods and research approaches used to answer the three research questions.  This chapter is structured as follows: the research questions are first introduced along with the conceptual framework that guides the dissertation.  To offer the reader some context into the choice of study location, a brief history of Metro Vancouver’s housing affordability situation is presented in section 2.  In section 3, the data sources used in this study are described, with particular detail given to describing the household residential preference survey and walkability index that form the backbone of this study.  In section 4, the variables used in the models are discussed and the variable construction process is provided.  Finally, section 5 outlines the choice of statistical models used to answer the three research questions.   4.1 Research Questions  This dissertation attempts to measure the effectiveness of neighbourhood housing mix as a Smart Growth policy.  Of primary interest is an assessment of how successful neighbourhood housing mix is at achieving its intention of matching a greater proportion of households into neighbourhood types of their preference.  The first research question of this dissertation is therefore posed as follows:  Research Question 1: “Controlling for relevant personal characteristics, is there a significant association between the degree of neighbourhood housing mix and the ability of households to match the neighbourhood type of their residence with that of their preference?”  64  Neighbourhood housing mix policies were designed to enable a greater proportion of a region’s population to sort into the neighbourhoods of their preference.  However, since the idea of neighbourhood housing mix has been at the forefront of Smart Growth and New Urbanist principles, housing mix as a strategy to enable greater Smart Growth matching will be the particular focus of this part of the study. Additionally, the proposed dissertation seeks to examine the broader social implications of housing mix beyond the ability to match neighbourhood preferences to actual choices.  Potential indirect effects of housing mix are measured by examining the association between the ability for households to match their Smart Growth preferences with the satisfaction they report with their neighbourhoods and with favourable health outcomes.  These secondary research questions are posed as follows:  Research Question 2:  “Is an individual’s ability to match their neighbourhood preference with their choice significantly associated with their satisfaction with their neighbourhood?”   Research Question 3: “Is an individual’s ability to match their neighbourhood preference with their choice significantly associated with favourable health outcomes?”  A conceptual framework is shown illustrating the theorized pathway that is to be tested (Figure 5).  The analysis is divided into the primary research question, concerning the relationship between housing mix and an individual’s ability to match the neighbourhood of their preference, 65  and secondary research questions that test the hypothesized pathway between housing mix and the two areas of well-being.  In addition, a number of measurable covariates are used in the analysis, representing various attributes of the respondent’s residential property and neighbourhood, their satisfaction with those attributes, and the  socio-economic and demographic characteristics of the respondent and his/her neighbourhood.  Controlling for all covariates, it is hypothesized that a higher degree of housing mix is significantly and positively associated with ‘neighbourhood match’, but that housing mix is not likely to be a significant positive predictor of neighbourhood satisfaction or favourable health outcomes.  Instead, it is anticipated that the pathway between housing mix and neighbourhood satisfaction is moderated through “neighbourhood match”.  While greater housing diversity at the neighbourhood level is not expected to be a positive and significant predictor of neighbourhood satisfaction, respondents who live in neighbourhoods with a high degree of housing mix are expected to more closely match their neighbourhood preferences, and respondents who are better able to match their neighbourhood preferences will be more satisfied with their neighbourhood environments.  Neighbourhood satisfaction, in turn, moderates the relationship between neighbourhood match and health outcomes.  While neighbourhood match is not expected to be a significant predictor of favourable health outcomes on its own, it is expected that neighbourhood match will significantly predict neighbourhood satisfaction, and that neighbourhood satisfaction will have a positive and significant association with improved health outcomes.       66  Figure 5: Conceptual framework  Note: solid lines indicate relationships that are expected to be significant; dashed lines indicate relationships that are expected to be insignificant.  Apart from the identified relationships, many latent, unobserved factors are expected to impart significant effects on the various outcomes that are being measured.  These latent factors include empirically verified predictors of housing, satisfaction and health outcomes that were not measured for this project.  For example, real interest rates strongly predict housing prices (Harris 1989) and the favourability of mortgages lent to individuals may strongly determine which homes and neighbourhoods they may be able to choose from.  However, latent factors also include unknown economic, personal and cultural factors that may have profound effects on individual housing, satisfaction and health outcomes.  It is expected that these latent factors will have even more impact on matching and satisfaction than the cumulative effect of the variables Personal Characteristicsof IndividualSelf-reported Satisfaction with Attributes of Home and NeighbourhoodNeighbourhood SatisfactionNeighbourhood MatchDegree of Neighbourhood Housing MixHealth OutcomesOther, Unobserved Factors(Macroeconomic, Personal)Measurable Attributes of the Home and Neighbourhood EnvironmentPRIMARY RESEARCH QUESTION SECONDARY RESEARCH QUESTIONS67  that are included and controlled for.  As a result, R2 values of model fit are not expected to be high.   Another source of measurement bias will likely result from the approach to which phenomena are measured.  In certain cases, what is measurable may not be an accurate representation of the essence of what that measure is supposed to represent.  For example, what qualifies as a neighbourhood with Smart Growth “character” or suburban “character” may vary from individual to individual, and may be comprised of other factors that are not captured in the available data.  For example, Smart Growth character is objectively measured using the built form attributes summarized by the Metro Vancouver walkability surface (Frank et al. 2013) (see section 4.3.2).  These measurable attributes, which include the commercial floor area ratio, land use mix, intersection density and net residential density of a set geographical area, only capture a fraction of the many qualities – both tangible and intangible – of what we may call “Smart Growth” character in neighbourhood design.  Similarly, both personal preferences and a personal assessment of a neighbourhood’s character are measured by asking just 8 questions about the built attributes of a neighbourhood (see Section 4.4.3.2).  It is likely impossible to design a survey instrument or to isolate the built form measures that perfectly capture the idiosyncracies of Smart Growth character, and each person’s assessment of what qualifies as “Smart Growth” will vary from individual to individual.  It is inevitable that a model will be a simplification of “reality”, but it is unclear how, and to what extent, the findings from a simplified model deviate from actual behaviour.    68  4.2 Metro Vancouver as a Research Context Metro Vancouver (formally the Greater Vancouver Regional District) is an agglomeration of 23 municipalities with a combined population of approximately 2.5 million inhabitants, located in the province of British Columbia (BC) on the Pacific coast (Figure 6).  The region is centered on the City of Vancouver (population 600,000), which was founded in 1885 as the western terminus of the Canadian Pacific Railroad (CPR), Canada’s transcontinental railway, and, as such, Canada’s maritime gateway to the Orient. Metro Vancouver’s economy has gradually shifted over the past century from one dominated by natural resource extraction and management to one dominated by the service sector (Hutton 2011).  Nevertheless, real estate development has always played a large role in the region’s economy, beginning almost immediately after settlement with the private sale and subdivision of almost 6,500 acres of land originally granted to the CPR by the Canadian Federal government (Harris 2012).   69  Figure 6: Metro Vancouver municipalities  Notes: Map created by Jacopo Miro, University of British Columbia.  Reproduced with permission. Along with having a large sector of the economy structured around real estate development, Metro Vancouver is also limited by a constricted geography for outward urban development (Tomalty 2002). Additionally, a strong cultural desire to preserve the region’s dramatic natural setting through aesthetic control has always been a part of local planning culture (Punter 2002; 2003).  (Punter 2002; Tomalty 1997; Punter 2003).  For these reasons, among others, the Vancouver region – and the City of Vancouver, in particular – may place a greater emphasis on land use planning and urban design than any comparably-sized North American city or region. 70  Historically, pre-war urban growth was suburban and low density in character, led by a Garden City plan drafted by Harland Bartholomew in 1923 (Pettit 1993). Unlike many North American cities, however, Vancouver’s downtown core never lost its desirability to middle class residents, and higher density housing alternatives to the single family home have perhaps been viewed more favourably in Vancouver than elsewhere (Punter 2002). Vancouverites’ familiarity with attractive urban neighbourhoods as well as the lack of stigma associated with non single family home types makes Vancouver an ideal location in which to survey residential preferences for Smart Growth communities.  Situating the study in Vancouver avoids some of the pitfalls of non-familiarity leading to misleading responses that may plague stated preference studies in other contexts (Volk and Zimmerman 2000).  At the same time, the region also contains a sufficient proportion of homes in low density, suburban environments.  Walkability index scores (Frank et al. 2013) proxy for the degree of Smart Growth character neighbourhoods possess (see Section 4.3.2).  A histogram (Figure 7) of the walkability index scores of Metro Vancouver’s 61,299 postal codes shows an approximately normal distribution8 of neighbourhood environments ranging from extremely automobile-oriented, low density suburban  (low walkability, left side of chart) to neighbourhoods embodying Smart Growth characteristics (high walkability, right side of chart).                                                    8 Albeit slightly right skewed (i.e. towards more Smart Growth) (Skewness: 1.714). 71  Figure 7: Distribution of walkability index scores, Metro Vancouver   This distribution of neighbourhood types is likely very uncharacteristic of many North American metropolitan regions, where car-oriented neighbourhoods tend to predominate and walkability index scores would be, subsequently, quite negatively skewed.  However, Metro Vancouver provides an almost ideal sampling of neighbourhoods with different degrees of Smart Growth characteristics to conduct a residential preference study. Housing diversity is also likely higher across neighbourhood environments of all types in Metro Vancouver than it would be in other, comparably-sized North American metropolitan areas where single family homes would likely predominate.  In 2011, only 34% of all dwelling units in the region were classified as “single family detached” homes, while roughly 40% and 23% of residents lived in multifamily apartments and other, ground-oriented attached homes, 72  respectively9 (Metro Vancouver 2015).  The diversity of housing options likely applies across all municipalities and all neighbourhood types.  With the exception of the tiny suburban municipalities of Anmore, Belcarra, Bowen Island and Lion’s Bay, the proportion of the dwelling stock comprised of single family detached homes does not exceed 64% for any municipality in Metro Vancouver (ibid). Metro Vancouver is also infamous for its housing affordability problem (Vancouver Foundation 2010), that may be particularly pronounced in the most central neighbourhoods of the region, which is also where the majority of the neighbourhoods with Smart Growth characteristics are located (Frank et al. 2013).  Commentators have suggested that a lack of housing mix in these neighbourhoods may be leading to rising housing prices (Lee et al. 2008; Villagomez et al. 2012).  In particular, these commentators cited concerns that suitably-sized homes for families with children in more central parts of the region had appreciated beyond the means of even members of the middle and professional classes (viz. Villagomez 2011), a concern that was echoed in municipally-administered household surveys (City of Vancouver 2013; Corporation of Delta 2010).  Despite fears of an insufficient diversity of housing types in key areas, housing mix is considered an important policy lever among Metro Vancouver policymakers. Strategies to encourage greater housing mix at the neighbourhood level, or even within development projects, have received considerable attention (MacDonald 2005; Punter 2003).  Accommodating a range of family types and income groups using housing alternatives to the single family home has been raised by both the City (City of Vancouver 2013, 1992) and the Region (Metro Vancouver                                                  9 Also see Table 26 in Chapter 4  73  2007).  The diversification of housing types within walkable neighbourhoods has also been established as a central policy goal of Metro Vancouver’s strategic long-term growth plan, the Regional Growth Strategy (RGS) (Metro Vancouver 2011).  In addition to selecting Vancouver as an ideal context to empirically study housing mix and Smart Growth preferences, the findings of this dissertation also align very closely with the interests of regional policymakers.  However, it is also possible that Metro Vancouver’s housing costs might be so high that even purchasing townhouses and multifamily housing units – typically thought of as being more “affordable” alternatives to single family homes – might be beyond the financial reach of many home seekers.  Drawing conclusions about housing mix’s ability to enable better neighbourhood matching might be challenging if smaller or denser housing typologies are still too expensive for most home purchasers or renters.  4.3 Data Sources The analysis relies on four previously assembled data sources.  Census data are reported at the dissemination area or sub-tract level; CLASP data are reported at the postal code level, and walkability and road network data are reported at the parcel level and summarized  at the postal level.  All data sources were spatially linked at the postal code level.   Road network and postal code centroid data, up to date as of 2013, was obtained from DMTI spatial and used to construct the network buffers (see Section 4.4.1) which defined the geography in which a variety of spatial variables were constructed.  In the next section, the four main datasets will be described in some detail.  74  4.3.1 CLASP Residential Preference Survey  The primary dataset used in this study was obtained through the Coalition Linking Action Science and Prevention (CLASP) residential preference survey, initially prepared for Toronto Public Health and designed by Urban Design 4 Health Ltd.  Potential participants were recruited through private survey firm Ipsos-Reid Public Affairs in early 2011, primarily via online advertising and through supplemental telephone recruiting among Ipsos-Reid’s pre-recruited panel of Canadian households10. Eligible participants were restricted to individuals, 25 years of age or older, who had moved within the past seven years, or who were planning to move within the next seven years, living in the 23 municipalities of Metro Vancouver or the municipalities of the Greater Toronto Area.  Only one eligible participant per household was allowed.  Participants completed the survey online through an Ipsos-Reid website. A total of 10,350 participants in Metro Vancouver and 26,344 participants in Greater Toronto were recruited, of which 1,223 Metro Vancouver and 1,525 Greater Toronto participants successfully completed the survey; a completion rate of 12.1% and 5.8%, respectively.  Because of the inability to obtain neighbourhood-level built form measures for all of Greater Toronto, and the inability to obtain property-specific data for any Greater Toronto municipality, only survey responses from Metro Vancouver residents are used in this dissertation. To ensure that the survey respondents were drawn from a representative range of neighbourhood types and income levels, a sampling strategy was undertaken where respondents                                                  10 Ipsos-Reid Public Affairs has pre-recruited over 240,000 individual panelists in unique households across Canada, of which 28,000 reside in the Greater Toronto Area.  The number of Ipsos-Reid panelists who reside in Metro Vancouver was not disclosed. 75  were recruited from Forward Sortation Areas  (FSA)11 that were stratified according to their walkability index score quartile (henceforth: walkability quartile) or one of three income categories. The 2006 Walkability Surface (the predecessor of the 2011 walkability surface used to measure neighbourhood-level built environment attributes in this study) and 2006 Census household median income levels were used to assign each FSA by walkability quartile and income category.  Initially, sampling goals called for a minimum count of 50 completed surveys in each of the 12 walkability/income cells. However, as shown in Table 8, the distribution of potential recruits was not equal across all 12 walkability-income cells, and two cells (households earning more than $70,000 in the top 2 walkability quartiles) did not contain any potential recruits.  While there were no potential recruits in the top-most walkability and income FSAs, this does not imply that no households were recruited with these characteristics12.  Table 8:  Potential CLASP survey recruits, stratified by income and walkability score at the FSA level, Metro Vancouver   Annual Household Income  Walkabilty Quartile Less than $50,000 $50,000 - $70,000 Over $70,000 Total       1st (lowest)  967 3,412 1,804 6,183       2nd  873 1,826 72 2,771       3rd  660 537 0 1,197       4th (highest)  131 68 0 199       Total  2,631 5,843 1,876 10,350                                                   11 FSAs are geographic regions denoted by the first 3 characters of a 6 character postal code (e.g. the FSA for the postal code V5T 1T9 is “V5T”).  Canada Post uses the FSA to sort mail at a sublevel below that of the metropolitan region.  There are 93 FSAs in Metro Vancouver, or approximately one for every 9,000 households. 12 Based on the postal codes respondents indicated for their home address, Figure 14 in Chapter 5 illustrates where the participants who successfully completed the survey resided in Metro Vancouver.   76   Of the 1,223 households who completed the study, 1,186 were retained for analysis.  The primary reasons for eliminating responses were erroneously-supplied postal codes – typically postal codes outside of the region or postal codes that were associated with exclusively with non-residential addresses. 4.3.1.1 Residential preference questions An excerpt of the survey instrument containing the questions used to design the variables used in the proposed dissertation is included in the Appendix.  The design of the CLASP survey instrument incorporates a number of considerations that are meant to measure residential preference with more rigour than previous generations of stated preference survey instruments (see review in Section 3.2).  These include: (1) Options presented as a forced trade-off between two equivalently-priced properties, one suburban and one urban; (2) The accompaniment of worded questions with illustrations to visually present the options; (3) Using choice-based conjoint analysis techniques involving a series of questions to determine which combination of attributes contribute to overall preference for walkable neighbourhood environments.   In all, eight neighbourhood and property-specific scenarios are posed, and preference is conceived as a principal component extracted from responses to five of these questions, rather than a direct answer to a single question. While the techniques used in the CLASP study are considerably more refined than previous generations of residential preference studies, the design of the survey still poses methodological limitations that should be acknowledged.  The drawings 77  that accompany the questions are stylized and were used in a previous, Atlanta-based study (Levine and Frank 2006).  They may not reflect the built environment or aesthetics of either Smart Growth or suburban communities in Metro Vancouver very accurately.  Secondly, some of the illustrations may mislead survey respondents; as Malizia and Exline (2000) recall, many respondents filling out a survey using visual cues in Fort Collins, Colorado inadvertently chose neighbourhoods based on features such as the pleasance of the landscaping that the illustrator had drawn, and ignored many of the Smart Growth trade-offs – such as higher densities and reduced living space – that the researcher wanted to inquire about.  Finally, the choice set presented in the eight questions likely represents only a fraction of the choices that enter the decision-making calculus of households when they choose residential neighbourhoods.  To some extent this is ameliorated by the inclusion of question B1 (see Appendix) in both the survey and in some of the models used in this analysis that inquires into a range of other factors –from noise to affordability to proximity to family – that respondents may have considered when they moved to their current neighbourhood.  However, it is very unlikely that this list of considerations is exhaustive, and the built environment – as it is surveyed – may not even be a major consideration of households when they choose neighbourhoods and homes.  These limitations are hard to account for, but should be acknowledged when interpreting final results.  4.3.2 2011 Walkability Surface The Smart Growth characteristics of the built environment are measured through data obtained from walkability index scores provided in the 2011 ‘Walkability Surface’ for Metro Vancouver (Frank et al. 2013).  This data was based on the methods described elsewhere in detail (Frank et al. 2010) is used to develop the various “objective neighbourhood match” family 78  of dependent variables used in all the research questions of this dissertation (see Section 4.4.3.3). Walkability index scores are calculated using the sum of the Z scores of the net residential density, degree of land use mix, intersection connectivity and commercial floor area ratio measured around a 1 kilometer network buffer of the geographical centre (henceforth: the “centroid”) of each of the region’s 61,299 postal codes (Figure 8).  This network buffer represents a refinement over the previous method (Frank et al. 2005), in that the 1 km distance conforms only to lots directly accessible from the existing street system, while the 2005 surface cuts across property lines.     79  Figure 8: Street network buffer used for the creation of the 2011 walkability surface  From Frank et al. (2013).  Reproduced with permission.  Of the 1,186 observations (survey respondents) used in this analysis, 1,105 had an existing walkability index score assigned to their postal code, while 85 lived in newly-generated postal codes that had not been updated in the most recent walkability surface.  Using Google Street View, the built form characteristics of adjacent postal codes were reviewed, and the walkability index scores of these postal codes were assigned to the 85 missing observations.  The walkability surface of Frank et al. (2013) is just one among a family of indices that objectively measure “Smart Growth” characteristics in urban environments.  Forsyth et al. (2008) 80  divided the Minneapolis-St. Paul metropolitan region into a series of arbitrarily positioned 805 X 805 meter grids and measured street patterns and residential densities. Levine, Inam, and Torng (2005) used roughly 900 and 600 TAZs covering the Boston and Atlanta metropolitan areas, respectively, measuring their performance on 13 indicators of Smart Growth characteristics (such as street design, population design and land use intensity) and used cluster analysis to categorize these TAZs into 5 Smart Growth levels. Li et al. (2005) measured  neighbourhood design in 56 politically-defined neighbourhoods in Portland, Oregon by their degree of street connectivity, resident population, employment density and the percent of the area covered in open space.  Song and Knaap (2003) developed a quantitative measure of neighbourhood design by measuring 21 built environment attributes around a quarter mile crow-fly buffer of new residential developments in Metropolitan Portland.  These variables were then reduced by factor analysis to 7 uncorrelated factors and a cluster analysis was run to identify neighbourhood groupings and measure the quantity of residential stock within each cluster.  An indicator developed by Handy et al. (2004) and used in Cao (2008) did not study street patterns and road design directly, but inferred a neighbourhood’s Smart Growth characteristics based on the proximity of stores and services.  Their approach measured the number and diversity of stores and services accessible within 3 miles from a respondent’s home addresses along the area’s road network (ibid). Beyond pragmatic reasons of existing data availability, an advantage of choosing the walkability surface of Frank et al. (2013)  is that the definition of “neighbourhood” in which to measure characteristics is centered on the address of the respondent, rather than within a more formal definition of a “neighbourhood” with set boundaries.  This avoids the “scale issue” of the modifiable areal unit problem (MAUP) (Openshaw and Taylor 1979), in which the boundaries of 81  arbitrarily-defined areal units, such as census tracts or other politically-designated neighbourhoods, can be modified to result in vastly different associations of the predictor with the dependent variable.  For example, a respondent living on the edge of an urban census tract may reside in a completely different neighbourhood environment than someone living closer to the centre, but would nevertheless be aggregated to this same unit of areal analysis. Her neighbour across the street might be assigned to another census tract, but live amidst the same built environment attributes. Measuring the neighbourhood environment from a postal code centroid along the street network results in a unique measurement of neighbourhood environment for almost every household in this study’s sample.  Postal codes are very small units of geography – often comprising no more than a hundred households – and large multifamily buildings often comprise their own postal code.   4.3.3 Parcel Data Parcel data on residential properties was obtained from the British Columbia Assessment Authority’s (BC Assessment) 2011 database.  This database provided property-level information on 744,577 dwelling units within 677,679 residential properties in Metro Vancouver.  For the purposes of this dissertation, the data collected from the BC Assessment database included assessed property values13, interior size (square footage), and the number of bedrooms per unit.  Additionally, the BC Assessment database assigned “Actual Use codes” –an identifier of  the primary purpose for which a property is held (BC Assessment 2015) – to each property, which                                                  13 The reported assessed property values reflect “total improvement values” that are the sum of the unimproved value of the land plus the improved value of the structure. 82  was used to categorize housing types and build the housing mix measure (see Section 4.4.2) as well as to define which properties may have had the presence of privately-owned yards.   The BC Assessment database provides perhaps the most comprehensive spatially-referenced14 database on the attributes of residential properties in Metro Vancouver, but it is not without limitations for the purposes of this study.  Rental apartment buildings are treated as a single unit, so the interior floor space, number of bedrooms per unit, or the assessed value of individual apartment units could not be ascertained.  Property values, the number of bedrooms per unit, and the interior sizes of rental apartments were coded as “missing” data.  This may affect the interpretation of results in areas of the region which contain a high concentration of purpose-built, large rental stock, such as more central areas of the City of Vancouver, as well as North Vancouver and Burnaby.  Given that there are over 200,000 renter households living in multifamily buildings, many living  in purpose-built rental buildings (Metro Vancouver 2015), verifying the interior size, bedroom count and “values” of individual rental apartments was considered to be too onerous for this project.  Fortunately, BC Assessment does report the number of units within a rental building, so housing mix, itself, could be estimated in all areas.    Secondly, the method with which BC Assessment assesses property values may introduce some bias into the analysis.  BC Assessment derives property values by first calculating the depreciation and replacement cost of the dwelling unit, and then using the sale prices of nearby properties with similar structural characteristics to estimate the final, “improved” value of the property, which includes the building unit as well as the value of the land itself (BC Assessment,                                                  14 The alternative would be to use the 2011 National Household Survey, which cannot provide dwelling count information below the census metropolitan area level, or to use Canada Mortgage and Housing Corporation data, which is not reported below the level of the municipality.  In either case, these data sources are not spatially reference, and could not be entered into ArcGIS for any spatial analysis (i.e. defining average attribute values within a network buffer). 83  Personal Communication).  This method of property valuation may be somewhat  biased since assessed values – ostensibly imputed using hedonic regressions of nearby sales15 – are often not the same as the sale price of homes when they enter the market.  The final sale price is, arguably, the only monetary measure of property value that matters in a study of residential preference and household decision-making, since the asking price is the value that home seekers must pay.  However, since not all homes in an area undergo a transaction, and since houses with certain features in an area may sell more frequently than others, the actual “price” of all housing units is not known (Case and Quigley 1991).  ‘Hybrid’ approaches that use sales of nearby properties to then run hedonic regressions and impute property values on unsold nearby properties may not be ideal, but offer a useful alternative method to valuation (ibid).  4.3.4 2011 National Household Survey Finally, neighbourhood data on income levels, the age of the primary maintainer and the population of ethnic groups was supplied by the 2011 National Household Survey (NHS).  This data was collected at the level of 3,438 “dissemination areas” (DA).  DAs are the smallest unit of census geography in Canada, and each dissemination area in Metro Vancouver had, on average, a population of 673 residents, 259 private dwellings occupied by usual residents, and an area of 0.83 square kilometers.  DA data was aggregated to the level of the census tract, and only census tracts in which the 1,186 CLASP respondents resided were used for analysis.  As a result, the final sample consists of data linked to 1,355 dissemination areas that make up 457 census tracts.                                                   15 The methodological details of BC Assessment’s approach to property valuation are regrettably not very transparent.  An official at BC Assessment was unable to clarify whether hedonic regression techniques were used to estimate property values, even when asked by the researcher directly.   84  These 457 census tracts contain, on average, a population of 5,062 residents, 1,950 private dwellings occupied by usual residents, and an area of 6.37 square kilometers16 The voluntary 2011 NHS controversially replaced the mandatory long form census that had been used for census recordkeeping in Canada until 2006.  Prominent social scientists have rejected the validity of the census data in light of low response rates in certain areas, especially compared to previous censuses where response was mandatory (Hulchanski et al. 2013).  Statistics Canada, the government agency in charge of adminsitering the NHS, used a two phase sampling strategy that first assigned the voluntary NHS to 30% of Canadian households and, in a follow-up phase several weeks after mail-out, targeted a random subsample of non-responding households for follow-up (Verret 2013).  43.2% of households in the targeted sub-sample eventually completed the questionnaire, leading to a 77.4% response rate in total; this compares unfavourably with the 94% response rate from the 2006 long form census (ibid).  Statistics Canada suppressed the reporting of data from dissemination areas where the non-response rate for the NHS exceeded 50% of targeted households (Statistics Canada 2013a).  For the purposes of this dissertation, out of a total of 3,438 dissemination areas, data was suppressed (i.e. missing) from 25 dissemination areas with respect to ethnicity, 29 dissemination areas with respect to median age of the primary maintainer, and 66 dissemination areas with respect to household income.  Despite the shortcomings of the NHS, a decision was made to use the 2011 data over 2006 results, since collection of the 2011 NHS coincided with the collection of CLASP survey data, as well as collection of built form data used to build the 2011 walkability surface and 2011 BC assessment database.                                                    16 3 of these census tracts are over 100 km2 and contain large areas of uninhabited parkland.  85  4.4 Variable Construction Techniques  This section describes the construction of the key variables used in the analyses.  Both the approach and rationale to the design of these variables are discussed, beginning with a discussion on measuring geographically-defined variables using spatial buffers.  The two variables of interest: housing mix and neighbourhood match are then described in considerable detail.  The section ends with a much more brief description of the other variables used in this study.  4.4.1  Network Buffers for Spatial Variables A number of variables used in the analysis, most prominently housing mix, are measured based on observations within a prescribed spatial area known as a “buffer” (Oliver et al 2007).  This subsection describes the rationale for using buffers to measure spatial areas and explores different approaches to buffer design.  4.4.1.1 Rationale of Using Buffers to Define “Neighbourhoods” One of the main conceptual problems in this dissertation is to approximate the geographical boundaries of the “neighbourhood” which households identify with.   Because this dissertation seeks to investigate whether housing mix at the neighbourhood level predicts the ability for people to sort themselves into that same neighbourhood, this geographical area would, ideally, perfectly encapsulate both the region in which households first search for homes and, later, the region in which households render judgements about the built environment surrounding the homes they chose.  Understandably, this optimal definition of neighbourhood is nearly impossible to operationalize.  The psycho-spatial boundaries of “neighbourhoods”, and people’s 86  identification of “neighbourhoods”, themselves, are not exogenous factors that predict the housing search process.  Households consume “neighbourhoods” by purchasing a bundle of housing services linked to location, but they also produce them (Galster 2001); positive and negative attributes associated with certain neighbourhoods, such as good schools, cultural character but also crime or voluntary ethnic segregation – admittedly many things over which planners have very little control – are reinforced by the initial image of a neighbourhood and the type of people who self-select into them.  This, in turn, affects the ability of people to sort into these communities, but also the spatial boundaries of the “community” itself.  Neighbourhood spatial definition is complex, if not impossible, but a geographic area that approximates the boundaries of a region in which decisions related to various aspects of housing search are made must be defined in order to conduct further analysis. As mentioned in the description of the walkability surface (Section 4.3.2), a simple way to define the “neighbourhood” is to use existing, politically-drawn neighbourhood boundaries such as census tracts or municipally-designated neighbourhoods or wards.  This approach has been criticized for contributing to the “scale issue” of the MAUP (Openshaw and Taylor 1979; Jelinski and Wu 1996) that was discussed in Section 4.3.217.  An alternative to using pre-defined neighbourhood boundaries is to define a “buffer” – a zone centered around a particular point, such as a respondent’s address – and measure the environment encompassed within that area.  The simplest buffer design takes the form of a “crow fly” buffer that imposes a circular area of a set radius around a centroid.  However, a crow fly buffer assumes that there are no impediments                                                  17 However, because households both consume and produce neighbourhoods, formal definitions of neighbourhood may reinforce and segregate spatially distinct behaviours.  For example, Bourassa et al (2003) show that formal real estate boundaries within cities form fairly robust housing submarkets, with significant price differences for structurally similar homes across these borders. 87  to accessing the environment, and ignores obstacles – both natural and man-made – which likely impede access to different points defined within the buffer (Oliver, Schuurman, and Hall 2007).  For phenomena that are not impeded by ground features, such as cellular network signals (Balakrishnan, Seshan, and Katz 1995), a circular buffer may suffice.  However, for phenomena which approximate the ability for humans to navigate an urban environment, a “network buffer” defined by the distance traveled from a central point along an existing road network, may be more appropriate (Oliver, Schuurman, and Hall 2007).  For these reasons, a network buffer centered on the postal code of each survey respondent’s address18 was chosen.  Furthermore, network buffers were drawn at various distances around a respondent’s address, at 500 m, 1 km, 2 km, 3km, and 5 km from the respondent’s postal code centroid.  The rationale behind using varying buffer sizes in the analysis was to examine the scale at which spatially-defined predictor variables – primarily housing mix– best predicted the ability for households to match their neighbourhood preferences; in other words, to approximate the geographic area in which households search for homes.  It is expected that housing mix will increase in predictive ability up to a scale where spatial heterogeneity will be lost and the variance among the sample of observations will decrease (Jelinski and Wu 1996).  A final note on the selection of buffers is the dilemma of reconciling an objective spatial measurement with a survey question that is not objectively spatial.  While network buffers of various sizes are used to define the search area in which households search for homes, most of the survey questions on neighbourhood preference do not explicitly reference those distances                                                  18 To maintain the confidentiality of the survey participants, the CLASP survey only asked respondents to indicate the postal code in which they lived, rather than to provide their exact address.  Nevertheless, postal codes are very small units of geography.  The 1,118 postal codes used for analysis contained, on average, 52 “points of call”, or separate addresses18, and the residential environment within a postal code is likely to be the same throughout.   88  (see Appendix).  This potentially introduces a major source of bias, since the “neighbourhood” in which survey respondents render their judgments may vary considerably in size and will almost certainly not be exactly coterminous with a network buffer drawn by a researcher. A 1km network buffer was first reasoned to approximate the area in which households search for homes since question 1 of the visual preference component of the survey19 encourages respondents to consider the area “within 1 kilometer or ½ mile of [their] home”. Question 2, which was used to construct the neighbourhood match variable (see Appendix), also prompts respondents to consider the neighbourhood surrounding a “1 kilometer/half mile or 10 minute walk” from their home.  In summary, it is unclear whether the network buffers used approximate the actual conception that respondents hold of their “neighbourhood”, but a consistently-defined spatial area should be employed to measure spatial variables.  4.4.1.2 Buffer Design  To define the “neighbourhood” of search, network buffers were drawn at 500m to 5 km distances from a respondent’s address along Metro Vancouver’s road network.  This network buffer was used not only to define housing mix, but also to define various other built environment aspects of the neighbourhood that were used to create other spatial variables (Table 7).  The design of the network buffer followed the approach used by Frank et al. (2013) to develop the walkability surface for Metro Vancouver. A network buffer was created in ArcGIS along the existing road network, and the resulting polygon was trimmed to within 25 meters of the edge to strictly include only those properties that were directly accessible from the street                                                  19 This question was omitted from the construction of the neighbourhood match variable because of explicit references to housing mix 89  network (Oliver, Schuurman, and Hall 2007).  Using regional road network and postal code local delivery unit shapefiles obtained from DMTI Spatial’s CanMap Suite, network buffers were created in ArcMap around the 1,118 unique postal code centroids in which CLASP respondents resided.   An objection may be raised that a network buffer in a dense neighbourhood with many streets might include many more properties than a network buffer within a low density neighbourhood with fewer streets.  For example, the properties encompassed by the five network buffer distances are compared between a postal code in an urban area - postal code V5V 2C6 in the City of Vancouver (Figure 9) - and a postal code in an exurban area - postal code V4R 1S4 in Maple Ridge (Figure 10).  The two maps are presented at identical scales (1:100,000) and the difference in buffer coverage is immediately apparent.  The exurban community, with its long distances between adjacent streets and its paucity of intersections, only permits the buffer to capture properties along a handful of cul de sacs and rural roads.  Smaller buffer distances, such as 500 meters to 2 km are almost invisible, since very few intersections are encountered within those distances.  The urban community, with its dense street grid, allows the buffer to cover an extensive area, with each buffer distance clearly visible as concentric diamonds.    90  Table 7: Spatial measures created using network buffers Measure  Description Network Buffer Size(s) Used Variables Derived from this Measure     Housing Mix Diversity and distribution of dwelling units according to 7 different categories within a prescribed geographic boundary.   See section 4.4.2 From respondent’s residential postal code centroid: 500m, 1 km, 2 km, 3 km, 5 km  From respondent’s workplace postal code centroid: 1 km, 2 km, 3 km HsgMix; HsgMix_workplace     Bedroom Mix Diversity and distribution of homes with different bedroom quantities (0 to 6 or more bedrooms) within a 2 km network buffer. Calculated using the same formula as housing mix. 2 km from respondent’s residential postal code centroid. BuiltForm_2km_Factor Bedroom Quantity Mean number of bedrooms per unit for all properties within a 2 km network buffer. 2 km from respondent’s residential postal code centroid. BuiltForm_2km_Factor Interior Size Mean residential interior size (ft2) of all properties within a 2 km network buffer. 2 km from respondent’s residential postal code centroid. BuiltForm_2km_Factor Property Value Mean property value (in $) of all properties within a 2 km network buffer. 2 km from respondent’s residential postal code centroid. PropVal_2km Walkability Index  Sum of the Z scores of net residential density, commercial floor area ratio, intersection density, and degree of land use mix.  See Section 4.4.2. 1 km from respondent’s residential postal code centroid. WalkIndex; Objective Neighbourhood Match Score Yard Presence Proportion of residential units with a yard within a 2 km network buffer. 2 km from respondent’s residential postal code centroid. BuiltForm_2km_Factor         91  Figure 9: Buffer size comparisons around postal code V5V 2C6  Figure 10: Buffer size comparisons around postal code V4R 1S4  92  Aside from the area encompassed by the buffer, the net residential density within the urban neighbourhood is considerably higher than in the exurban neighbourhood.  This would entail that neighbourhoods are being compared that vary considerably in the number of homes within them, which would theoretically affect both the degree of housing mix within these areas and the choice that home seekers have within them.  However, an assumption is made that individuals are more apt to define ‘neighbourhoods’ - and therefore define their residential search - to comparable geographical areas rather than areas with a comparable number of homes.  For example, it seems illogical that a person choosing between a potential residence in a dense, Smart Growth community and a sprawling suburban community would restrict their search to a tiny geographical section of the Smart Growth neighbourhood and a much larger region of the suburban area because both areas contain exactly 1,000 dwelling units.  Nevertheless, this discrepancy should be acknowledged and particular attention should be paid to smaller network buffer distances, where the size discrepancy between an urban network buffer and a suburban network buffer is likely to be especially pronounced.    4.4.2 Housing Mix The predictor variable of interest for the first research question is “housing mix”, which proxies for the diversity and distribution of different housing typologies at different spatial definitions of neighbourhood.  Just as a suitable geographic definition of “neighbourhood” raises conceptual questions, further issues are raised as to how to categorize housing typologies.  Unlike neighbourhood definition, where researchers have recourse to an array of empirically tested techniques such as different buffers or the use of formal neighbourhood boundaries, “housing mix” has been understudied and there is little empirical guidance on how to segment 93  housing products by type.  Aurand (2010) segments housing typologies into 5 types: “single-family detached homes, single-family attached homes, small multi-unit structures of 2–4 units, large multi-unit structures of 5 or more units and other housing types such as mobile homes” (ibid). Regional housing policymakers use Canadian census data to classify housing types into single family detached, apartment, and “other ground oriented20” dwellings (Metro Vancouver 2015).  However, in both cases, there is no supporting rationale for why housing units were categorized into these typologies. In the absence of any empirical guidance, three options for defining housing categories, and thus housing mix, were considered during the exploratory phase of the research.   4.4.2.1 Categorizing Housing Typologies – Exploratory Work The first two options involved using parcel data from BC Assessment to manually assign housing typologies to different numbers of categories.  In the first iteration properties were assigned to 7 categories according to the ‘actual use code’ and ‘manual use code’ supplied by BC Assessment (Table 8).  This classification scheme was largely informed by Aurand’s (2010) 5-class categorization, but introduced new categories that were assumed to reflect Metro Vancouver housing culture and policymaking.                                                    20 This category includes semi-detached, apartment-duplex, row house and “other single detached” as well as movable dwellings (Metro Vancouver 2015). 94  Table 8: Housing categories used to construct housing mix variable Category Actual Use Code Manual Use Code Property Description (BC Assessment)     Single Family Detached 0  “Single Family Dwelling”  2  “Single Family Dwelling – Property Subject to Section 19(8)”  60  “Single Family Dwelling, Duplex – 2 acres or more”     Single Family Detached – with suite 32  “Single Family Dwelling with Basement Suite”     Single Family Attached 39 D701 “Row Housing (Single Unit Ownership) – Strata Ownership”  52 D733 “Multi-family (Garden Apartment & Rowhousing) - Townhouse     Ground-oriented Attached 33  “Duplex (Single Unit Occupancy – Front)”  34  “Duplex (Single Unit Occupancy – Bottom)”  35  “Duplex (Single Unit Occupancy – Side)”  47  “Triplex”  49  “Fourplex”  52 D734 “Multi-family (Garden Apartment & Rowhousing) – Rowhouse Tenant  53  “Multi-family Conversion”     Multifamily – below 5 stories 30 D702 “Strata-lot Residence (Condominium) – Strata Apartment – Frame”  50 D291 “Multi-family (Apartment Block) – Apt Walk up – Owner Pays Heat”  50  D292 “Multi-family (Apartment Block) – Apartment – Tenant Pays Heat”  50 D293 “Multi-family (Apartment Block) – Apartment – Owner Pays Heat”  50 D297 “Multi-family (Apartment Block) – Apartment – Frame”  52 D288 “Multi-family (Garden Apartment & Rowhousing) – Apartment over Commercial”  52 D289 “Multi-family (Garden Apartment & Rowhousing) –Apartment with Elevator”  52 D291 “Multi-family (Garden Apartment & Rowhousing)  – Apt Walk up – Owner Pays Heat”  52 D292 “Multi-family (Garden Apartment & Rowhousing)  – Apt Walk up – Tenant Pays Heat”  52 D297 “Multi-family (Garden Apartment & Rowhousing)  – Apartment – Frame”  55 C288 “Multi-family (Minimal Commercial) – Apartment over Commercial”  55 C293 “Multi-family (Minimal Commercial) – Apartment – Owner Pays Heat”  55 D288 “Multi-family (Minimal Commercial) – Apartment over Commercial”      95  (Table 8 Continued)  Category Actual Use Code Manual Use Code Property Description (BC Assessment)     Multifamily – below 5 stories 55 D352 “Multi-family (Minimal Commercial) – Multiple Residence”  58 D702 “Stratified Rental Apartment (Frame Construction) – Strata Apartment – Frame”  202 C288 “Store(s) and Living Quarters – Apartment over Commercial”  202 D288 “Store(s) and Living Quarters – Apartment over Commercial”     Multifamily – 5 stories and above 30 B705 “Strata-lot Residence (Condominium) – Strata Apartment – Hi-Rise”  54 B299 “Multi-family (High Rise) – Apartment (Reinforced Concrete)”  54 B300 “Multi-family (High Rise) – Apartment (High Rise)”  55 B300 “Multi-family (Minimal Commercial) – Apartment (High Rise)”  59 B705 “Stratified Rental Apartment (Hi-Rise Construction) – Strata Apartment – Hi-Rise”     Other Housing Types 37  “Manufactured Home (Within Manufactured Home Park)”  38  “Manufactured Home (Not In Manufactured Home Park)”  56  “Multi-Family (Residential Hotel)”      For example, building codes that historically limited cheaper wood-framed construction to a height of 4 storeys due to fire safety concerns may have served as a supply side deterrent to the construction of taller buildings, and provides an intuitive point of reference to divide multifamily properties between “high rise” and “low rise” types.  “Duplex”, “Triplex”, “Fourplex” and “Garden style apartments”  are separately categorized from larger multifamily dwellings because the former comply with “as-of-right” zoning in most Metro Vancouver municipalities (e.g. ‘RM’ zoning bylaws in the City of Vancouver), while the latter often require much more onerous discretionary review.  In other cases, demand-side assumptions are used to assign Actual Class Codes to different housing type categories.  Single family detached homes are separated into conventional single family detached homes, and those that possess a legal secondary suite.  The justification for this separation is that the presence of a secondary suite 96  may alter the decision-making calculus of households interested in settling in the neighbourhood. A secondary suite may provide an opportunity for homeowners to derive further income from renting out a second unit, encourage aging empty nesters to ‘downsize’ within their existing property, or open up settlement in a primarily single family home neighbourhood to people who may only have the means to afford a secondary suite.  The presence of secondary suites within single family home neighbourhoods may also alter the desirability of a community for certain groups of people, whether negatively or positively.  While this categorization scheme may rely heavily on anecdotal assumptions, previous studies of housing mix, such as Aurand (2010), as well as regional housing policy documents (Metro Vancouver 2015) and census classifications also do not provide theoretical or empirical justifications for the way in which housing types are classified.  A breakdown of the number of dwelling units in Metro Vancouver in 2011 according to this 7 class categorization scheme is outlined in Table 9. Table 9: Frequencies of different housing categories, Metro Vancouver Housing Category N %    Single Family Detached 304,127 40.8% Single Family Detached – with suite 63,645 8.5% Single Family Attached 83,514 11.2% Ground-oriented Attached 28,782 3.9% Multifamily – below 5 stories 156,749 21.1% Multifamily – 5 stories and above 102,629 13.8% Other Housing Types 5,131 0.7%    TOTAL 744,577 100.0%     A second alternative was to categorize housing types into three simple, arbitrary categories based loosely on Canadian census definitions: single family homes, multifamily homes (apartments and strata condominiums with common areas) and other “ground-oriented” 97  single family homes.  This is loosely the same categorization scheme used by Metro Vancouver housing policymakers (Metro Vancouver 2015).  The categorization of properties into these three classes involved aggregating the 7 existing categories: the two multifamily housing categories were combined into one; single family detached homes and single family detached homes with suites were also combined into a single category, and ground-oriented attached and single family attached were combined into a new “ground oriented” category; the 0.7% of the residential stock that was classified as “other” was removed in this second classification scheme.  In the end, this second simplified alternative was abandoned from further analysis for two reasons.  Firstly, over 15% of the postal codes recorded no housing mix at all when the 3 category scheme was used.  Moreover, it was felt that a 3 category classification scheme may have been too blunt and failed to capture some of the nuances between different housing typologies in the search process.   A final option was to perform a latent class analysis (LCA) to empirically define housing types based on their structural features.  Using data from BC Assessment on the number of bedrooms, square footage, and assessed value, each property was classified according to membership in 3, 4 or 5 latent classes using SPSS Amos software. Unfortunately, the practical application of this technique was limited by several factors.  Firstly, the computing program and computer power available to conduct LCAs of this size was not sufficient to investigate beyond 5 latent classes.  With nearly 800,000 properties to classify, even with just 3 input parameters calculations in SPSS Amos took nearly 24 hours and were prone to crash.  Although SPSS Amos is capable of performing LCAs, it is not among the software programs recommended for LCA in research use (Haughton, Legrand, and Woolford 2012).  With the suggested programs being proprietary and beyond the budget allocated for this research project, SPSS Amos was used as an alternative.  However, SPSS Amos lacked basic features that would have been instrumental in 98  evaluating the effectiveness of LCA; most notably, SPSS Amos lacked “goodness of fit” measures for latent class membership, such as the Aikake Information Criterion (Bozdogan 1987), so interpreting the appropriateness of choosing a 3 versus a 4 or 5 member latent class structure was not possible.  Regardless of computing limitations, an LCA-defined housing mix measure was eventually abandoned for a much more pragmatic reason:  latent classes are unique statistical constructs that cannot be translated into actual policy.  Often homes with identical numbers of bedrooms and very similar interior sizes were segmented into separate latent classes.  However, housing units are developed and built according to established dwelling categorization schemes found in zoning bylaws and other planning documents.  .    4.4.2.2 Variable Construction After selecting the categories by which housing “types” were defined, the final step was to choose a formula that created a continuous measure – or index – of the diversity of mixing amongst housing types in a specified area. As a theoretical concept, “diversity” should incorporate both considerations of the number of classes represented in a subsample, as well as the distribution of those elements across the classes (McDonald and Dimmick 2003).  For example, a neighbourhood with 3 housing types split among a residential stock of 100 homes in the proportion 98:1:1 may have the same number of classes as a neighbourhood with 3 housing types split 33:34:33, but the latter offers, theoretically, a greater ability for households with preferences for housing types other than the first class to sort themselves into that neighbourhood.  Likewise, a neighbourhood with a 50:50 split between only two housing types may provide less choice to a sample of interested home seekers than a neighbourhood with more 99  housing types represented, but a less even distribution between the types.  Housing mix [1] is measured using Shannon’s entropy index (Shannon and Weaver 1949):   [1]  HousingMix = - Pi X ln Pi  Where Pi is the proportion of homes of type ‘i’ among all residential units within the 1 km postal code buffer.  The choice of Shannon’s entropy index was based on its previous use in measuring housing mix (Aurand 2010) and based on an empirical comparison of 12 measures of diversity undertaken by McDonald and Dimmick (2003).  In that study, Simpson’s D and Shannon’s entropy index were both considered to offer the greatest flexibility and sensitivity among diversity measurements, with Shannon’s entropy index being especially sensitive to changes in the number of categories (i.e. housing types) between observations (ibid).  For this study, this attribute of Shannon’s entropy index was deemed especially useful in recognition of the possibility that the number of housing types represented might vary widely from neighbourhood to neighbourhood.    4.4.3 Neighbourhood Match The outcome of interest for the primary research question is to measure whether greater housing mix allows households to “neighbourhood match” – that is, to live in the type of neighbourhood they prefer, according to its Smart Growth attributes.  Two different approaches to measuring “neighbourhood match” were considered:  (3) Subjective Neighbourhood Match attempts to measure the congruence between the neighbourhood that respondents live in and the neighbourhood they prefer, based entirely 100  on the self-assessment of their current neighbourhood using the visual preference questions in the CLASP survey. (4) Objective Neighbourhood Match attempts to measure the congruence between the neighbourhood that people desire, as indicated on the survey, and the actual neighbourhood environment.  The Smart Growth attributes of the actual environment are objectively measured using the 2011 walkability surface (Frank et al. 2013). Separate models were run predicting subjective and objective neighbourhood match.  The two approaches are meant to capture differences between perceived and objective environments, and to investigate, in the second and third research questions, whether these different conceptions lead to different behavioural outcomes.     4.4.3.1 Theoretical Issues Before developing a measure of “neighbourhood match”, at least three theoretical issues need to be considered.  The first issue is to capture Smart Growth attributes of residential neighbourhoods in both the measure of neighbourhood preference and the appraisal of the current neighbourhood.  That is, to ensure that the attributes that are under consideration by the respondent are specifically aspects typically associated with Smart Growth, and not other neighbourhood features that may be somewhat unrelated, such as school quality, proximity to friends, family and ethnic groups, or other aspects of urban environments that may not fall under the rather large umbrella of Smart Growth characteristics.  The second challenge is to ensure that measures of neighbourhood preference are comparable with measures of the actual neighbourhood environment.  If these two measures are not at least somewhat comparable, or based on comparable attributes, then we cannot claim that household are able to match their 101  residential preferences with their choices.  Finally, when is a respondent considered to be “matched” to their neighbourhood type?  Unless a household finds a home with the exact bundle of housing services (including neighbourhood and location) that are exactly congruent with their preferences, they will not be ideally matched.  In practice, almost no household will be ideally matched, and so it is important to justify where a boundary might be drawn between respondents who are “matched” and those who are not.  These three theoretical issues were operationalized in the construction of the various neighbourhood match variables.  4.4.3.2 Measuring Smart Growth environments To measure the neighbourhood environment according to its Smart Growth attributes, survey respondents answered 8 questions on the CLASP residential preference survey (see Appendix).  These questions queried respondents on various built environment features typically associated with Smart Growth.  The questions were presented as “trade offs” between an environment that was analogous to a Smart Growth community, and a community that was automobile-oriented and suburban in design.  Respondents not only ranked their preference on an 11 point (0 to 10) ordinal scale for what community over the other, but were asked to evaluate their current neighbourhood on the same features, using the same scale.      102  Table 10: CLASP survey trade-off questions Ques-tion Trade off Posed Wording of Smart Growth Option1 Wording of Suburban Option1 Use in Variable Constr-uction? 1 Lot size, proximity of commercial services, travel options, commute distance, transit options (See Appendix) (See Appendix) NO 2 Walkability and proximity of commercial services Where houses and commercial areas are within a 1 km/half mile or 10 minute walk of each other so that I can walk to stores, libraries or restaurants. Where the commercial areas are kept separate (over 2 km/1.5 miles or more than a 30 minute walk away) from the houses, even if this means that I cannot walk to stores, libraries or restaurants. YES 3 Level of activity and mix of housing With lots of services and activities nearby, even if this means it has a mixture of single family houses, townhouses and apartment buildings that are close together on various sized lots with less private backyard space. With single family houses farther apart on lots 10 meters (35 feet) wide or more with more private backyard space, even if this means it is not an area with services or activities nearby. NO 4 Home size and travel options Where I can walk, cycle, or take public transit for some of my trips because commercial areas are nearby within a 1 km/half mile or 10 minute-walk), even if this means the homes are smaller with less interior living space. With larger homes with more interior living space, where the commercial areas are distant (over 5 km/3 miles) or more than a 45 minute walk away) form the houses, even if this means I have to drive for all m trips. YES 5 Lot size and commute distance Within 5 km or 3 miles (10-15 minute drive) of work, school or my other important destinations, even if this means that houses are close together – on smaller lots approximately 6 meters (20 feet) wide. With houses farther apart – on large lots 15 meters (50 or more feet) wide – even if this means traveling more than 25 km or 15 miles (over 30 minutes) to work, school or my other important destinations. YES 6 Street design and travel options Where I can walk, cycle or take public transit for some of my trips, even if it has through streets and people from other neighbourhoods walking or driving on them. With cul-de-sacs and few people from other neighbourhoods walking or driving on them, even if this means I must drive for all my trips. YES 7 Public recreation opportunities and lot size Where within a short walk there is lots of public recreation and green space for swimming, walking , jogging, running trails, social interaction, sports and playgrounds even though there is little space for recreational activities on my own property Where there is lots of space on my own property for recreational activities, but there is little public recreation and green space for swimming, walking, jogging, running trails, social interaction, sports and playgrounds within a short walk. YES 8 Access to, and size of, food outlets Where I could easily walk to a wide range of small to medium sized grocery stores, fruit and vegetable stands, butchers, bakers and specialty food stores. With a few food stores within walking distance but several very large supermarkets within a 10 minute drive. YES      1 All questions prefaced with the phrase “If I were to move, I’d like to live in a neighbourhood…” 103   Table 10 outlines the 8 questions, and the tradeoffs that were posed.  As described in Section 3.2, earlier generations of residential preference surveys suffered from proxying residential preference, or representing neighbourhood built environment choices, with a single measure often derived from a single survey question.  While Smart Growth environments are multi-faceted and incorporate several different, but probably related, design elements in order to develop a measure of “preference”, these attributes have to be distilled into a single measure.   To accommodate this, and because  many of the responses to the questions were highly correlated with one another, a single measure representing a respondent’s preference for “Smart Growth attributes” was isolated by running a principal component analysis (PCA) on the residential preference questions.   Since two of the questions – questions 1 and 3 – explicitly reference the concept of housing mix in their phrasing, these two questions were omitted from the PCA, to avoid introducing endogeneity into the dependent variable21.  The PCA was run in SPSS on neighbourhood preference questions for the 6 questions identified in Table 10 using a Varimax rotation, with a maximum of 25 iterations, selecting for components with Eigenvalues above 122.  Bartlett’s (1954) test of sphericity revealed that the correlation matrix was non-random (p<0.001), while a Kaiser-Meyer-Olkin statistic of 0.853 was above the recommended level to conduct factor analyses (Kline 2014).  The choice of using an orthogonal rotation method such as Varimax, as opposed to an oblique rotation, was informed by the need to create a                                                  21 Neighbourhood match is derived from the preference score, and the first research question tests whether housing mix predicts neighbourhood match.  22 Eigenvalues are analogous to the proportion of variance accounted for by each component (DiStefano et al 2009); an Eigenvalue below 1 represents a factor that explains less variance than a single variable, and it is assumed that, for practical reasons, no single component should explain less than a variable’s worth of the total variance. 104  single independent component – one representing Smart Growth preference - as opposed to correlated factors (Field 2005).    Table 11 shows the outcome of the PCA, with the total variance explained by each component.  Table 11 also reveals that a single component – component 1 – reported an Eigenvalue above 1, and was retained to create the measure of preference and, eventually, the measure of neighbourhood match.  This component explained over 56% of the total variance in the data.   Table 11: Principal components analysis for neighbourhood match, total variance explained Component Eigenvalue Variance Explained (%) Cumulative Variance Explained (%)     1 3.368 56.1 56.1 2 0.733 12.2 68.4 3 0.641 10.7 79.0 4 0.487 8.1 87.2 5 0.414 6.9 94.1 6 0.357 5.9 100.0    105  Table 12: Principal components analysis for neighbourhood match, factor loadings for component 1 Question Trade-off posed Factor loading on component 1    2 Walkability and proximity of commercial services  .784702 4 Home size and travel options .842392 5 Lot size and commute distance -.750513 6 Street design and travel options .751287 7 Public recreation opportunities and lot size -.730686 8 Access to, and size of, food outlets -.617180  The factor loadings from this component (Table 12) were then multiplied by the original responses to the 6 preference questions and summed to create the “neighbourhood preference score” (NPS) [2] – a measure of a respondent’s preference for Smart Growth communities -  and “current neighbourhood score” (CNS) [3] – a measure of a respondent’s appraisal of his/her own community for its Smart Growth attributes:  [2]  NPS = 0.784702 X Q2a  + 0.842392 X Q4b  - 0.750513 X Q5a + 0.751287 X Q6a 0.730686 X Q7a - 0.617180 X Q8a.  [3]  CNS = 0.784702 X Q2b  + 0.842392 X Q4b  - 0.750513 X Q5b + 0.751287 X Q6b 0.730686 X Q7b - 0.617180 X Q8b.  Where Q2a through Q8a represent ordinal scale responses to survey questions on neighbourhood preferences, and Q2b through Q8b represent the same survey questions prhased 106  to query respondents on the design of the neighbourhood they currently reside in (see Appendix).  Thus, because they are multiplied by identical factor loadings, NPS and CNS are intended to measure the same phenomena and can be directly compared.  The difference between NPS and CNS is the “subjective neighbourhood match score” (SubjNM_score) [4], or the degree to which a respondent’s preference for smart growth matches the smart growth “attributes” they perceive in their residential neighbourhood:  [4] SubjNM_score = NPS – CNS.   A subjective neighbourhood match score of 0 indicates a “perfect” match between one’s neighbourhood and one’s preferences23.  A positive NPS indicates that a respondent prefers Smart Growth environments, so a positive subjective neighbourhood match score indicates that a respondent is “mismatched” into a neighbourhood that he/she perceives to possess fewer Smart Growth characteristics than their preferences, and a negative score indicates that a respondent is “mismatched” into a neighbourhood that possesses more Smart Growth characteristics than they prefer.   Similarly, an objective neighbourhood match score (ObjNM_score) [5], was derived by taking the difference of the walkability surface score for the respondent’s postal code from the NPS.  NPS and the walkability surface were calibrated such that they were on the same scale.  This was achieved by taking the difference between the observed scores at the 5th and 95th                                                  23 To derive this, a questionnaire was entered with a score of 5 on an ordinal scale of 0 to 10 for all of the neighbourhood preference questions, indicating total indifference to Smart Growth or suburban neighbourhood design.  This resulted in a NPS of 1.4001.  This value was subtracted from all observations of NPS and CNS so that 0 indicated total indifference. 107  percentile (i.e. within two standard deviations from the mean) for both walkability and NPS, dividing these two values and multiplying the walkability surface score by this value24.    [5] ObjNM_score = NPS – WalkabilityScore.  How comparable are subjective and objective neighbourhood match?  More specifically, how closely does the objectively measured walkability surface compare to respondents’ subjectively indicated “current neighbourhood score”?  If the two are not roughly substitutable, there is a greater likelihood that each measure  may proxy for different phenomena; a person’s appraisal of their current neighbourhood according to a survey’s visual cues  may not be the same as the objectively-defined measure of that same environment (and vice versa).  To investigate this, a simple Pearson’s correlation test was run between neighbourhood preference score, current neighbourhood score and the walkability surface values for all CLASP respondents (Table 13).  Following this, a binary logistic regression was performed to compare how current neighbourhood score and the walkability surface separately predict the odds of a respondent indicating the top quartile of the neighbourhood preference score25 on the survey (Table 14).                                                    24 The difference between the 5th and 95th percentile for NPS was 34.8801 and the difference between the 5th and 95th percentile for the uncalibrated walkability surface was 8.10046.  As such the walkability surface was multiplied by 34.8801/8.10046, or 4.3060 to yield a newly calibrated walkability surface score on the same scale as NPS. 25 “Neighbourhood preference score” was not normally distributed, and various transformations did not yield normal distributions.  As a result, a decision was made to isolate the top quartile of neighbourhood preference scores and to run a logistic regression. 108  Table 13: Correlation coefficients between walkability index scores and current and neighbourhood preference scores  CNS Walkability Index Scores NPS     CNS 1 .590*** .611**     Walkability index score - 1 .413**     NPS - - 1     Coefficients reported are Pearson’s r. * Significant at p <0.05  ** Significant at p<0.01  ***Significant at p<0.001  Table 14: Logistic regression results predicting odds of reporting top quartile of neighbourhood preference scores  Model 1 Model 2    Constant .038*** .179***    CNS 1.200***     Walkability Index  1.257***    Pseudo-R2 (Nagelkerke) .411 .134 * Significant at p <0.05  ** Significant at p<0.01  ***Significant at p<0.001   The correlation analysis reveals that current neighbourhood scores and the walkability surface are strongly and positively correlated with one another and with neighbourhood preference scores.  The result from this logistic regression showed that the odds of indicating a high neighbourhood preference score were almost the same (25.7% for each unit of increase in walkability; 20.0% for each unit of increase in “current neighbourhood score”) for the two measures of the Smart Growth attributes of the existing residential environments. From this rudimentary analysis, it can at least be qualitatively inferred that respondent’s perceptions of their neighbourhood and the objectively-measured walkability surface proxy for comparable 109  aspects of the environment, and that subjective and objective neighbourhood match therefore evaluate similar phenomena.  4.4.3.3 Defining Neighbourhood Match While the subjective and objective neighbourhood match scores indicate the degree to which households are mismatched away from a theoretical “perfect” match score of 0, the likelihood of a household perfectly matching their neighbourhood to their preferences was considered to be extremely low.  Indeed, no survey respondent reported a match score of 0, measured either subjectively or objectively.  At the same time, measuring “neighbourhood match” on a continuous scale from “unmatched” to “matched”  was ruled out because the development of such a scale - as well as any easily interpretable transformation, such as a log transformation - did not yield a normal distribution.  Figures 11 and 12 demonstrate the challenge of creating a continuous variable of neighbourhood match with a normal distribution.  While match scores are relatively normal, households that are ostensibly “matched” lie in the middle of the distribution, clustered around a score of 0 (Figure 11).  To define neighbourhood match would involve taking the inverse of the absolute value of subjective and objective neighbourhood match scores, resulting in an extremely skewed distribution that would not be suitable for linear regression (Figure 12).    110  Figure 11: Distribution of subjective neighbourhood match scores  Figure 12: Distribution of subjective neighbourhood match as a continuous variable  111   To avoid this, neighbourhood “match” was defined as a binary outcome, where survey respondents were either classified as “matched” or “unmatched” with their neighbourhood preferences.  In the absence of any prior empirical work determining what constitutes a “matched” neighbourhood preference, the distribution of subjective and objective neighbourhood match scores were plotted, and observations within one, one half, one third and one quarter of a standard deviation on either side of a neighbourhood match score of 0 (i.e. “perfect” match) were considered.  A sensitivity analysis was run by running a hierarchical logistic regression of the various neighbourhood match alternatives on the 5 importance factors and subsequently on housing mix measured at 5 search geographies: 500m, 1 km, 2km, 3km and 5 km network buffers from a respondent’s postal code.  The sensitivity analysis was evaluated based on which value improved overall prediction the most, based on the largest change in Nagelkerke’s Pseudo-R2. Based on these findings (see Appendix), observations defined within half a standard deviation of a neighbourhood match score of 0 was selected to represent  individuals who were neighbourhood matched, for both subjective and objective measures of the variable.  4.4.4 Neighbourhood Importance Factors and Built Form Component  Two other families of variables will be discussed in further detail because of the complexity of their design. The five “neighbourhood importance factors” (NIF) measure the importance respondents assigned to various aspects of their dwelling and neighbourhood when they were searching for homes.  In all 22 separate questions were posed on a 4 point ordinal scale (See Question B1 in the Survey Appendix), querying respondents on how important each aspect was in selecting their current residence.  Because the responses to these questions are 112  likely very highly correlated, a factor analysis was performed using a Varimax rotation to create uncorrelated factors.  Based on the analysis, five factors with Eigenvalues above 1 were isolated (Table 15) proxying for various underlying aspects of residential search; these 5 factors explained over 56% of the cumulative variance in the responses to the 22 questions.  These factors were renamed based on the factor loadings of these components on the various aspects of neighbourhood importance that respondents were asked to consider (Table 16). Table 15: Factor analysis for neighbourhood importance questions, total variance explained Factor Eigenvalue Variance Explained (%) Cumulative Variance Explained (%)     1 5.183 24.7 24.7 2 2.628 12.5 37.2 3 1.554 7.4 44.6 4 1.359 6.5 51.1 5 1.081 5.1 56.2     6 0.940 4.5 60.7 7 0.829 3.9 64.6 8 0.792 3.8 68.4 9 0.765 3.6 72.1 10 0.730 3.5 75.5 11 0.701 3.3 78.9 12 0.642 3.1 81.9 13 0.575 2.7 84.7 14 0.549 2.6 87.3 15 0.484 2.3 89.6 16 0.439 2.1 91.7 17 0.427 2.0 93.7 18 0.410 2.0 95.7 19 0.364 1.7 97.4 20 0.332 1.6 99.0 21 0.215 1.0 100.0 Bolded factors retained for variable construction.    113  Table 16: Neighbourhood importance factors, factor loadings  Factor Loadings Factor 1 2 3 4 5 Name “Neighbourhood Amenities” “Regional Access” “Jobs and School Access” “Social Support” “Home Attributes” Abbreviated Name NeighbAm_ factor RegAccess_ factor JobSchool_ factor Social_ factor HomeAtts_  factor Importance question:      Closeness to public open space .749     Ease of walking .730     Closeness to public recreation space .716     Closeness to a wide range of small to medium sized grocery stores .603     Closeness to shops and services .568 .322    Closeness to restaurants .521   .493  Closeness to cultural/ entertainment venues .519   .487  Ease of bicycling .506  .341   Convenient access to work and other destinations on public transit  .779    Closeness to a bus stop  .768    Closeness to a rail station  .768    Quality of schools   .868   Closeness to child care or school   .863   The size of your yard   .566  .338 Closeness to job or school  .311 .520   Closeness to particular cultural/ethnic community    .705  Closeness to friends and family    .610  The amount of interior space in your home     .696 Affordability/Value  .333   .651 The noise from traffic .312    .588 Highway/freeway access from your home    .362 .422        114  Similarly, a variable was created to represent objectively-measured features of the dwellings and neighbourhood built environments surrounding a respondent’s address.  This measure was intended to represent the various inducements and obstacles to residential location search, including median area home prices, a home’s interior size, availability of a yard, the average number of bedrooms, and the “bedroom mix”26 encountered in the search process.  Because exploratory analysis (see section 6.2.1) revealed that housing mix peaked in prediction at a 2km search buffer, the same 2 km search buffer was used to measure these 5 variables, and a PCA was performed using Varimax rotation to isolate a single component representing the built form measures of the home and neighbourhood environment.  Much like the PCA to define neighbourhood match, the built form PCA was deemed to meet the criteria for factor analyses (Kaiser-Meyer-Olin statistic of 0.669; Bartlett’s test significant at p<0.001), and yielded a single component that explained over 68% of the total variance (Table 17).  This component was renamed the “built form factor” (Built_Form_Factor_2km). Table 17: Principal components analysis for built form factor, total variance explained Component Eigenvalue Variance Explained (%) Cumulative Variance Explained (%)     1 3.433 68.7 68.7 2 1.021 20.4 89.1 3 0.474 9.5 98.6 4 0.044 0.9 99.4 5 0.028 0.6 100.0                                                     26 A measure of the distribution and diversity of dwelling units with different numbers of bedrooms calculated using the same entropy index used for housing mix 115  Table 18: Principal components analysis for built form factor, factor loadings Built Form Measure Factor loading on component 1   Bedroom Quantity .987 Yard Presence .966 Interior Size .941 Bedroom Mix .799 Property Value No loading  Interestingly, property value was not found to be highly correlated with the other built form measures and was dropped from the construction of the principal component. Because logistic regression is especially sensitive to multicollinearity (Leech, Barrett, and Morgan 2005), a multicollinearity test was undertaken by linearly regressing housing mix (the main predictor variable) on the 5 neighbourhood importance factors and the built form factor.  The results of this regression demonstrated that none of the variance inflection factors (VIF) exceeded 1.399, which is well below the maximum VIF of 4.0 recommended by Pan and Jackson (2008), or the 5.0 recommended by Rogerson (2010).    4.4.5 Outlier Removal To remove distortions in statistical estimation, outliers were removed from observations of housing mix (at all network buffers), subjective and objective neighbourhood match scores, BMI, as well as the variables used to construct the Built Form principle component (i.e. mean interior size and bedroom count within a 2 km network buffer of a respondent’s address).  These 116  variables were chosen for outlier removal because prior assessment revealed that each of these contained observations that were more than 3 standard deviations away from the sample mean.  Some of these outliers may have been the result of respondent’s misreporting values (e.g. a BMI of 73 – 19 points, or more than 4 standard deviations, higher than the next highest observation), while others reflect extremities inherent in the structural attributes of residential properties.  For example, extremely large homes being concentrated near one another  due to land use regulations and zoning bylaws (e.g. one area having a median home size of 4,365 ft2 – nearly 3 times the regional median).  These observations could have introduced substantial bias to model results, and may even have led to completely different conclusions (Osborne and Overbay 2004).  For example, the correlation between BMI and Housing Mix goes from being insignificant to significant at p<0.10 if just 5 outliers (out of a sample size of 1,186) are retained.  To remove outliers, the following formulas were applied following the method prescribed by Hoaglin, Iglewicz, and Tukey (1986):    Upper Limit = Q3 + [(Q3 – Q1) * 2.2]   Lower Limit = Q1 - [(Q3 – Q1) * 2.2]   Where the “upper limit” and “lower limit” are the upper and lower boundaries of observations used in the model, Q1 is the observation found at the 25th percentile, and Q3 is the observation found at the 75th percentile.  117  4.4.6 Other Predictor Variables A number of predictor variables are used in the various models used in all three research questions.  Table 19 lists them as well as the data sources from which they are obtained. Table 19: Predictor variables used Abbreviation Variable Name/Description Type Data Source     Age Age Continuous CLASP Q S1     Age_under60 Respondent is under 60 years of age Dummy  CLASP Q S1     BuiltForm_2km_Factor Principal component of various built form factors (see Section 4.4.4) measured around a 2 km network buffer of a respondent’s postal code centroid Continuous BCAA database     Commute_BelowMedian Respondent reports an average daily commute time under the CLASP sample median of 20 minutes Dummy  CLASP Q F14      Cul_de_Sac Respondent lives on a cul-de-sac Dummy CLASP Q A9     Ethn_Canadian Respondent reports “Canadian” ancestry  Dummy  CLASP Q F20     Ethn_European Respondent reports European ancestry (i.e. English/ Scottish/ Irish; French; Italian; German; Greek; Polish; Portuguese; Dutch; Ukrainian) Dummy  CLASP Q F20     Ethn_EastAsian Respondent reports East Asian ancestry (i.e. Chinese, Korean, Vietnamese, Filipino) Dummy CLASP Q F20     Ethnicity_allothers All other ethnic groups (i.e. all other selections in question F20 apart from those already listed above) Dummy  CLASP Q F20     Ethnicity_NotMatched Ethnicity of respondent is different from largest group in census tract Dummy  CLASP Q F20, 2011 NHS     Gender_Male Gender of respondent is Male Dummy  CLASP Q S2     HHInc_under$40k Respondent’s household income is below $40,000/year Dummy  CLASP Q F23     HHInc_$40_$100k Respondent’s household income is between $40,000 and less than $100,000/year Dummy  CLASP Q F23             118      Table 19 (continued)     Abbreviation Variable Name/Description Type Data Source     HHInc_over$100k Respondent’s household income is at least $100,000/year or above Dummy  CLASP Q F23     HHInc_NotMatched Respondent’s household income category is not the same category as the median household income of the census tract Dummy CLASP Q F23, 2011 NHS     HomeAtts_  factor “Home attributes” factor – one of the 5 neighbourhood importance factors (see Section 3.4.4) Continuous CLASP Q B1     HsgType_GroundAttached Respondent lives in a ground-oriented attached dwelling (i.e. semi-detached house, row/townhouse, duplex/triplex/quadplex ) Dummy CLASP Q A1     HsgType_MFHhigh Respondent lives in a multifamily home (apartment or condo), 5 stories of above in height Dummy CLASP Q A1     HsgType_MFHlow Respondent lives in a multifamily home (apartment or condo), 4 stories or lower in height Dummy CLASP Q A1     HsgType_SingleFam Respondent lives in a single family detached home Dummy CLASP Q A1 JobSchool_ factor “Jobs and School Access” factor– one of the 5 neighbourhood importance factors (see Section 3.4.4) Continuous CLASP Q B1     Length_of_Res Number of years respondent has lived in present residence Continuous CLASP Q A4     NeighbAm_ factor “Neighbourhood Amenities” factor – one of the 5 neighbourhood importance factors (see Section 3.4.4) Continuous CLASP Q B1     NeighbSat_very_sat Respondent’s overall satisfaction with their neighbourhood is rated as 10/10, or “very satisfied” Dummy CLASP Q A11     NeighbSat_very_unsat Respondent’s overall satisfaction with their neighbourhood is rated as 5/10, or lower (very unsatisfied) Dummy CLASP Q A11     RegAccess_ factor “Regional Access” factor – one of the 5 neighbourhood importance factors (see Section 3.4.4)  Continuous CLASP Q B1     119  Table 19 (continued)     Abbreviation Variable Name/Description Type Data Source     Tenure_Rent Respondent currently rents their home Dummy CLASP Q A3     WalkIndex Walkability Index Score (see Section 4.3.2) measured around a 1 km network buffer of a respondent’s postal code centroid Continuous 2011 Walkability Surface      4.5 Models Used This section provides an overview of the statistical models utilized to answer the three research questions.  Both the choice of regression analysis and the selection of variables used in the models will be discussed.  A power calculation was run using G*power software to determine the minimum sample sizes to be used in the models.  Cohen (1992) recommends a power of .80 to detect a medium effect (d=0.15) at 0.05.   Given that research questions 1 through 3 (self-reported health status) use logistic regression models, the desired effect size was converted to an odds ratio of 1.31 following the approach advocated by Chinn (2000).  Given an odds ratio of 1.31, with a default probability of H0 of 0.2, a power of .80 and a non-normal (manual) distribution of X, the recommended minimum sample size was estimated to be 145.    4.5.1 Research Question 1 The first research question seeks to understand whether housing mix is a significant predictor of a household’s ability to match their current residential environment to that of their preference.  Two sets of models were used to answer this question. The first series of models employed binary logistic regression to predict how an increase in neighbourhood housing mix surrounding a respondent’s address increased (or decreased) their probability of subjectively or 120  objectively matching their neighbourhood with their preferences.  A second series of model employed multinomial logistic regression to compare the probability of respondents belonging to different matching outcomes (Table 20).    Table 20: Neighbourhood match categories used in multinomial logistic regression models Match Category Title Description Notes     “Prefers SG and Matched” Respondent is “matched”, either objectively or subjectively, and prefers Smart Growth residential environments (i.e. neighbourhood preference score > 0). Main outcome of interest; always reported    “Prefers Suburbia and Matched”  Respondent is “matched”, either objectively or subjectively, and prefers a suburban residential environment (i.e. neighbourhood preference score < 0).     “Mismatched into higher SG” Respondent is “not matched”, either objectively or subjectively, and they reside in a community with more Smart Growth attributes than their preference (i.e. a neighbourhood match score below half a standard deviation below 0).     “Mismatched into lower SG” Respondent is “not matched”, either objectively or subjectively, and they reside in a community with fewer Smart Growth attributes than their preference (i.e. a neighbourhood match score above half a standard deviation above 0). Reference group.       The reference group for all multinomial logit models are respondents who are “Mismatched into lower SG” – those who live in communities that are more than half a standard deviation less “Smart Growth”-oriented or more than half a standard deviation less walkable than they desire, and represented by the right tail of the neighbourhood match score distributions (see Figure 13).   121  Figure 13: Neighbourhood match categories visualized  Since this study is primarily interested in understanding whether housing mix enables people with Smart Growth preferences to sort into their preferred neighbourhood environment, the main comparison of interest is between the first category (“Prefers SG and Matched”) and the reference group.  However, the reference group should not be misconstrued to represent all people with Smart Growth preferences who happen to be mismatched into suburban environments (i.e. communities with a walkability index score below 0, or a current ½ Standard Deviation below 0½ Standard Deviation  above 0MatchedMismatched into higher Smart GrowthMismatched into lower Smart Growth(Smart Growth or suburbia)122  neighbourhood score below 0).  People who belong to this reference group may also include those with a slight suburban preference mismatched into a very suburban environment and also people with preferences for very high density Smart Growth environments who live in areas with only a moderate number of Smart Growth attributes.  The reason this approach was taken - rather than simply defining mismatch by those who prefer Smart Growth versus those who prefer suburbia - is that a large number of respondents with Smart Growth preferences were mismatched into areas that had even more Smart Growth attributes than they preferred, most often in the very high density neighbourhoods of the downtown Vancouver peninsula.  This phenomenon is somewhat unusual for a North American city; in Levine and Frank’s (2006) analysis of Boston and Atlanta, for example, most households with Smart Growth preferences who were mismatched found themselves in low density suburban areas.  The unusual phenomenon of mismatch into higher smart growth environments – and what this says about Vancouver’s housing market and the geographic distribution of its rental stock – will be analyzed in considerable detail in Chapters 5 and 6. Collinearity diagnostics revealed that the maximum variance inflection factor (VIF) between variables was 2.41427, with the mean VIF of all variables being 1.520.  While there is no definitive guidance on what VIF indicates a maximum permissible level of multicollinearity, this is well below the VIF of 4.0 recommended by Pan and Jackson (2008), or the 5.0 recommended by Rogerson (2010).                                                    27 Between the variables “length of residence in years” and “household income between $40,000 - $59,999/year”.   123  4.5.2 Research Question 2 The dependent variable used for the second research question was the degree to which respondents rated their satisfaction with their neighbourhood.  This variable was directly measured from question A11in the CLASP survey instrument using an 11 point ordinal scale ranging from 0 (extremely unsatisfied) to 10 (extremely satisfied) (See Appendix).  Neighbourhood satisfaction is measured on an ordinal scale, where each successive level is implied to represent a higher value (i.e. greater satisfaction) than the level before it.  However, since these values are self-reported and individuals interpret satisfaction and levels of satisfaction differently, the values are not scalar and the value of each unit of satisfaction is somewhat arbitrary.  Earlier generations of residential satisfaction surveys, such as those conducted by Ha and Ha and Weber (1994) and Cook (1988) used linear regression techniques to predict an outcome that was non-scalar and likely provided spurious relationships (Lu 1999).  Since then, linear models have been largely eschewed in favour of logistic models (Basolo and Strong 2002; Buys and Miller 2012; Howley 2009; Hur and Morrow-Jones 2008; Lovejoy, Handy, and Mokhtarian 2010; Parkes, Kearns, and Atkinson 2002; Lu 1999).  For these reasons, an ordered logit model was chosen for this analysis.   The majority of survey respondents were satisfied with their neighbourhood, with nearly a third of the sample reporting the highest level of satisfaction (i.e. 10/10) with their communities (Table 12).  In contrast, only 4 respondents rated their neighbourhood extremely negatively (i.e. 0/10).  Because the size of the lowest category was incredibly small, a decision was made to reverse the order of the ratings such that extreme satisfaction (i.e. 10/10) was the reference group.  Instead, the logit model was structured to predict the changing odds of reporting one level lower of satisfaction associated with the change in a predictor variable.  124  Therefore, the model actually predicts the level of dissatisfaction with a neighbourhood, and the interpretation of an odds ratio higher than 1 implies a negative response (i.e. increasing the odds that respondents are more dissatisfied with their neighbourhood). Table 21: Neighbourhood satisfaction – frequency of scores Satisfaction Score (0=very unsatisfied; 10 = very satisfied) N Percent of Total 0 4 0.3 1 8 0.7 2 5 0.4 3 19 1.6 4 17 1.4 5 57 4.8 6 74 6.2 7 170 14.3 8 264 22.3 9 185 15.6 10 383 32.3 TOTAL 1,186 100.0 Question posed: “How satisfied are you with your neighbourhood (0 = dislike very much, 10=like very much)”  4.5.3 Research Question 3 The third, and final, research question investigates the effect of neighbourhood matching on two health outcomes self-reported health status and obesity.   4.5.3.1 Self-reported Health Status The direct relationship between the housing mix of an individual’s neighbourhood and their perceived health condition is expected to be insignificant, but moderated first through neighbourhood match and, in turn, by neighbourhood satisfaction, a significant pathway is expected between the two variables.  Self-reported health status is measured directly from question F6 in the CLASP survey (see Appendix), asking respondents to rate their health status compared to other people of their age group on a 5 point ordinal scale, with 1 indicating “poor” 125  health status and 5 indicating “excellent” health status.  Given the ordinal nature of the data, an ordered logit model was selected to predict self-reported health status.  The reference group in these series of models were people with the lowest self-reported health rating.  Since less than 4 percent of the sample reported “poor” health ratings, these values were combined with those people rating their health as “fair” for a reference group that contained 285 observations.  Table 22: Self-reported health status, frequency of scores  Self-Reported Health Status  N Percent of Total 1 = “poor” 46 3.9 2 = “fair” 239 20.2 3 = “good” 428 36.1 4 = “very good” 329 27.7 5 = “excellent” 138 11.6 6 = “don’t know/prefer not to say” 6 0.5 TOTAL 1,186 100.0 Question posed: “In general, compared to people your own age, would you say your health is…(select one)”  4.5.3.2 Body Mass Index Respondents were asked to specify their height (Question F3) and weight (Question F4) in the CLASP survey.  From this, BMI was calculated by dividing the respondent’s reported weight (in kg) by their height squared (m2).  A descriptive analysis of the normality of the natural logarithm of BMI revealed that the variable was approximately normal in distribution, with a skewness statistic of .308 (standard error: .071) and a Kurtosis of -.013 (standard error: .149).  Figure 14 illustrates the normality of the distribution of BMI scores in the sample using a histogram.  Given the normality of the BMI distribution, ordinary least squares regression was used to predict the natural logarithm of BMI.  A log transformation was performed, allowing the coefficients to be easily interpreted as a percent change in BMI for every unit change in the predictor variable.  The recommended sample size to achieve statistical power was estimated 126  using G*power software.  The following parameters were inputted: effect size = 0.15 = .05; power = .80; 6 predictors.  Based on these inputs, G*power recommended a minimum sample size of 146.   Figure 14: BMI distribution in CLASP sample  127  Chapter 5: Descriptive Results and Sample Characteristics This chapter profiles the characteristics of the study’s sample versus the region’s population and provides descriptive statistics of the main variables used in the models.  As such, it is organized into two sections. The first section, “sample characteristics”, describes how representative the survey sample is to the Metro Vancouver region. The second section, “variable characteristics”, uses descriptive statistics and maps to describe the basic features of the main variables in the study.    5.1 Sample Characteristics How representative is the CLASP sample compared to the Metro Vancouver region which is under study?  If the sample under investigation differs substantially from the general population in key characteristics, we may not be able to infer that the outcomes studied would apply.  From an academic standpoint, this would violate one of the key assumptions of positivist, deductive reasoning that findings are universal (Creswell 2008).  Practically, study outcomes that cannot be tied to behaviour in the general population are meaningless for planning and policymaking.  For example, if housing mix is shown to be a significant determinant of neighbourhood match in the sample, but the sample was highly unrepresentative of the population, we may not be able to infer that housing mix would be linked to better residential matching across all of the region, and therefore we would not be able to provide evidence to guide regional planners.  Two aspects of representation are considered: geographic representation and socioeconomic and demographic representation.   128  5.1.1 Geographic Distribution of Survey Responses One important consideration of conducting residential preference and location choice research is to sample a diverse complement of neighbourhood environments and neighbourhood environment choice sets of individuals.  This would range from respondents residing in very suburban, low density environments to those residing in very high density, Smart Growth-oriented communities and a diverse assortment of preferences for this same range of neighbourhood designs.  However, the distribution of different residential environments and residential preferences along a “Smart Growth spectrum” is very spatially determined.  As described in Chapter 2, urban economic theory suggests that areas characterized by either Smart Growth attributes and/or greater housing mix could either be located closer to the region’s CBD, or in select communities where residential intensification and Smart Growth characteristics would be favoured (or, at least, not discouraged) by existing home owners and enshrined in local land use planning laws. Figure 15 shows the geographical distribution of CLASP survey respondents across Metro Vancouver’s constituent municipalities.  As Table 23 reveals, the distribution of survey respondents corresponds reasonably closely to the distribution of households in Metro Vancouver by municipality, with a slight overrepresentation in the City of Vancouver.  As the overrepresentation in the City of Vancouver might suggest, respondents in the CLASP sample reside in neighbourhoods that are more walkable (i.e. have more Smart Growth attributes) than the regional average (Table 24).  Over half of the sample lived in areas in the highest walkability quartile, for example.  Nevertheless, this may not necessarily be indicative of a non-representative sample.  129  Figure 15: CLASP survey responses, Metro Vancouver    The highest and lowest walkability index score observations in the sample closely match the highest and lowest observations in the region. Additionally, the walkability quartiles represent the number of postal codes in the region, broken down by their walkability index score and not the population of individuals residing within these quartiles. Unfortunately, since postal code boundaries are different from census geography boundaries, it is difficult to obtain accurate population statistics at the postal code level, so the population breakdown by walkability quartile cannot be ascertained.    130  Figure 16 shows the distribution of respondents mapped over the walkability quartiles of all postal codes in Metro Vancouver.  The map underscores one of the unique attributes of Metro Vancouver among North American metropolitan regions: that areas of both high and low walkability (i.e. high walkability index scores) are distributed across nearly every municipality in the region.  The map also demonstrates that CLASP survey respondents are clustered in a manner that is expected given the population densities of different neighbourhoods in the region.  Since the walkability index is estimated using net residential densities as one component, we would expect population densities to increase in areas with a high walkability index score and, hence, the distribution of observations would appear to cluster. Figure 16: Walkability index and distribution of CLASP survey respondents  131  Table 23: Municipal representation in survey sample  Metro Vancouver (2011 NHS) CLASP Sample Municipality Population % of Total Number of Households % of Total N % of Total Sample over-representation         Anmore  2,092  0.1%  628  0.1% 2 0.2% 0.1% Belcarra  644  0.0%  268  0.0% 1 0.1% 0.1% Burnaby  223,218  9.6%  86,839  9.7% 63 5.3% -4.4% Coquitlam  126,456  5.5%  45,553  5.1% 27 2.3% -2.8% Delta  99,863  4.3%  34,755  3.9% 43 3.6% -0.3% Langley1   129,258  5.6%  48,552  5.4% 41 3.5% -2.0% Maple Ridge  76,052  3.3%  28,044  3.1% 25 2.1% -1.0% New Westminster  65,976  2.9%  30,586  3.4% 56 4.7% 1.3% North Vancouver1   132,608  5.7%  53,342  6.0% 89 7.5% 1.5% Pitt Meadows  17,736  0.8%  6,718  0.8% 5 0.4% -0.3% Port Coquitlam  56,342  2.4%  20,651  2.3% 19 1.6% -0.7% Port Moody  32,975  1.4%  12,628  1.4% 23 1.9% 0.5% Richmond  190,473  8.2%  67,976  7.6% 51 4.3% -3.3% Surrey  468,251  20.2%  152,847  17.1% 198 16.7% -0.5% Vancouver2  616,537  26.7%  269,614  30.2% 508 42.8% 12.6% West Vancouver  42,694  1.8%  17,074  1.9% 23 1.9% 0.0% White Rock  19,339  0.8%  9,866  1.1% 10 0.8% -0.3% Other   12,788  0.6%  5,395  0.6% 2 0.2% -0.4%         TOTAL 2,313,328  891,310  1,186   1. Includes both City and District Municipality/Township figures 2. Includes Electoral District A    132  Table 24: Walkability index, descriptive statistics  Metro Vancouver CLASP Sample Statistic   % of Total     N1  61,299 1,118      Mean 0.331 2.473  Median -0.054 1.699  Standard Deviation 2.881 4.055  Minimum -6.11 -5.40  Maximum 16.38 15.84      Quartiles    1st  146 12.3% 2nd  167 14.1% 3rd  258 21.8% 4th  615 51.9%     1Number of unique postal codes represented  5.1.2 Socioeconomic and Demographic Representativeness If the addresses of CLASP respondents are geographically representative of the region and its built form, how representative are the respondents, themselves?  Table 25 provides a demographic snapshot of the respondents in the sample compared to census figures for the region.  By age group, the sample is somewhat older than the region’s population.  One of the requirements for eligibility was to be older than 25 years of age, which explains why people under the age of 25 are not at all present in the sample.  Apart from this omission, all age groups are reasonably well represented.  While there are disproportionately more female respondents than in the regional population, this does not suggest that there were fewer males in the households queried.  Unfortunately, the survey only queried one member of each household about their residential preferences, while research has shown that decisions to move homes and the appraisal of homes and neighbourhoods often involve multiple household members (Levy, 133  Murphy, and Lee 2008).  While this is a flaw in the design of the survey instrument, it was justified by the logistical difficulties of asking multiple members of a household to fill out an online survey.   CLASP survey respondents are more educated than the regional average, which is a typical sampling issue encountered in research surveys (Holbrook, Krosnick, and Pfent 2007).  Additionally, fewer survey respondents have children than the regional average; the underrepresentation of families with children may pose a challenge to connecting the evidence gathered from this study to regional policymaking since families with children encounter space and financial constraints that may dictate which homes, neighbourhoods and tenure options are accessible to them.  Finally, the ethnicity of respondents does not closely match the ethnic makeup of the region; fewer East Asian households are represented than the regional average, while “all other ethnicities” – a catch-all for households that do not identify with “European”, “Canadian” or “East Asian” ethnic groupings – is overrepresented.  It should be acknowledged that making comparisons between the ethnic representation in the sample and the ethnic representation in the population are fraught with difficulties.  The 22 ethnicity categories available to respondents in the CLASP survey differed from the categories available in the 2011 NHS, of which respondents can select from any combination of over 200 nationalities (Statistics Canada 2013b).  In both the CLASP and NHS survey instrument, membership in multiple categories was allowed, making it difficult to assign individuals to specific ethnic groups.  Not only may it be difficult to gauge the representativeness of the sample to the population, but the interpretation of variables of ethnicity may be difficult in statistical models.   134  Table 25: Demographic characteristics of study respondents  Metro Vancouver (2011 NHS) CLASP Sample  N % of Total N % of Total Number of Households 891,310       1,186        Age Group (years)1     Under 25 28,800 3.2% n/a n/a 25 to 34 130,170 14.6% 196 16.5% 35 to 44 175,255 19.7% 234 19.7% 45 to 54 209,620 23.5% 232 19.6% 55 to 64 168,710 18.9% 302 25.5% 65 to 74 95,985 10.8% 170 14.3% 75 and over 82,770 9.3% 52 4.4%      Gender     Male 1,116,840 49.0% 473 39.9% Female 1,163,860 51.0% 713 60.1%      Education Level (Highest Level of Schooling Completed) High School or Less 796,340 41.3% 234 19.7% Less than High School  280,575  14.6% 9 0.8% High School  515,765  26.8% 225 19.0% Community College 313,630 16.3% 292 24.6% Some University n/a  177 14.9% Completed University 339,435 17.6% 303 25.5% Graduate Degree 194,215 10.1% 180 15.2%      Children in Household     Yes 300,640 33.7% 222 18.7% No 590,670 66.3% 964 81.3%      Ethnicity     European  953,130  41.2% 505 42.6% Canadian  327,910  14.2% 93 7.8% East Asian  668,300  28.9% 125 10.5% All Other Ethnicities n/a 15.7% 463 39.0% 1. In the 2011 NHS survey this is the age of the primary maintainer of the household.  Table 26 compares some of the socioeconomic characteristics of the sample alongside census statistics for the region.  Importantly for this study, with the exception of duplex/triplex/quadplex residents, the dwelling typologies that respondents lived in closely 135  mirrored the stock of properties within Metro Vancouver.  The tenure breakdown – whether households own or rent the homes in which they live - is also roughly similar in the sample compared to the population.  In general, household income is well represented in the sample.  Households only reported incomes according to the categories listed, and did not specify their exact dollar income, but just over half of the sample reported incomes less than $60,000, compared to the median household income of $63,347 for Metro Vancouver in 2011.   Table 26: Socioeconomic characteristics of study respondents   Metro Vancouver (2011 NHS) CLASP Sample  N % of Total N % of Total Number of Households 891,310       1,186        Commute Time  Median (minutes) 25.6  20.0       Household Income     less than $20,000  131,475  14.8% 128 10.8% $20,000 to less than $40,000  148,300  16.6% 217 18.3% $40,000 to less than $60,000  142,870  16.0% 305 25.7% $60,000 to less than $80,000  215,360  24.2% 205 17.3% $80,000 to less than $100,000  94,970  10.7% 120 10.1% $100,000 to less than $120,000 n/a  87 7.3% $120,000 or more n/a  124 10.5%      Housing Type Respondent Lives ina   Detached house  301,140  33.8% 459 38.7% Semi-detached house  n/a   9 0.8% Row/townhouses  80,500  9.0% 124 10.5% Duplex/triplex/quadplex  126,600  14.2% 32 2.7% Apartment/condo              (1 to 4 stories)  228,555  25.6% 344 29.0% Apartment/condo                 (5 or more stories)  129,255  14.5% 192 16.2% Other  n/a   26 2.2%      Tenurea     Own 529,090 65.1% 708 59.7% Rent/other 283,755 34.9% 478 40.3% a. Source: Metro Vancouver (2015) 136  While the household income distribution in the sample closely approximates the distribution in the population, it is interesting to examine the spatial distribution of household incomes in the sample compared to the median household income in each of the region’s census DAs (Figure 17).   Figure 17: Household income of CLASP respondents compared to DA medians, Metro Vancouver  Figure 17 seems to suggest that there is a high degree of income mixing both in the sample and the region, with low income areas adjacent to high income areas, and with pockets of poverty and wealth distributed throughout Metro Vancouver.  A closer examination of the most central part of the region (Figure 18) shows a more distinct wealth divide at a finer scale.  Incomes are still divided between the traditionally more affluent west side of the City of Vancouver, compared to the more working class neighbourhoods of East Vancouver, and traditionally low income Strathcona.  The downtown peninsula is a microcosm of the region as a 137  whole, with incomes ranging from some of the poorest DAs (#59150759 in the Downtown East Side with a median household income of just $11,113 /year) to some of the richest (#59153833 in Yaletown, at $141,414/year) within a kilometer of each other.  Figure 18 also highlights the diverse incomes of respondents in downtown Vancouver and the innermost neighbourhoods.  While more high income respondents live west of the main north-south rapid transit line on the city’s more affluent West Side, there is substantial mixing of incomes nearly everywhere, with lower income respondents living in high income dissemination areas, and vice versa.  A cursory glance at this map would seem to suggest that there are fewer financial barriers to settling in the most central parts of Vancouver than many commentators would imply (cf. Villagomez 2011).   138   Figure 18: Household income of CLASP respondents compared to DA medians, central Metro Vancouver   5.2 Variable Characteristics If the first section provided a glimpse of how representatively the sample fit the population, this section offers a snapshot of the data.   The characteristics of the dependent variables used in the models, as well as of housing mix, will be explored by reporting such descriptive statistics as measures of central tendency, the normality of the distribution of key variables.  Additionally, the frequency of scores, or levels, of ordinal dependent variables will be 139  examined.  Finally, maps are used to show both the geographic distribution of key variables as well as to highlight issues such as clustering in the housing mix variable.  5.2.1 Housing Mix Table 27 presents some descriptive statistics of the housing mix variable, defined at successively larger network buffers around a respondent’s residential postal code.  A few trends are apparent from casual observation. First, as the search geography increases, the average housing mix score grows.  This aligns with expectations that housing typologies are clustered and segregated, governed by such factors as zoning bylaws and land economics.  We would expect less variation in the diversity and distribution of different housing typologies within a 500m walk of any given residential address than we would within a 5 km walk of that same address.  Eventually, as search geographies increase to cover very large proportions of the region, we might expect there to be little heterogeneity or variance among housing mix scores in the sample.  This is apparent in shrinking standard deviations and rising minimum values of housing mix scores as the search geography is increased to its maximum extent of 5 km.  However, the highest housing mix observed in any of the observations remains relatively fixed around a score of 1.728.  This suggests that there are neighbourhoods in the sample where there is a very diverse representation of housing types, even within a small search area.  Conversely, minimum scores of 0 at 500m and 1 km suggest that there are also communities which are represented by only one type of dwelling.  Housing mix scores are shown to be slightly left, or negatively, skewed, indicative of a slight concentration of observations at higher housing mix                                                  28 Given that the housing mix variable is an entropy measure of 7 different categories, the highest possible score – where each of the 7 categories would be fully represented and exactly equal in size – would be 1.946.   140  scores.  As is expected, increasing search geographies returns higher average housing mix scores, while some of the low scores (particularly in lower density, suburban areas) are retained.  As a result, negative skewness increases. As network buffer sizes increase and housing mix scores cluster around the mean, the kurtosis – or “peakedness” of the distribution – also increases. The decreasing number of observations found as search geography increases in size reflects two factors. As search buffer sizes increase and observations cluster around the mean, the number of outliers that are removed increases, lowering the number of observations that were included in the sample.  Meanwhile, the sharp drop in observations between 3km and 5km measures of housing mix reflects a computational constraint that returned fewer observations because of the size of the calculation that ArcGIS had to undergo.  The number of residential properties that are included in a spatial join calculation around a network buffer increases exponentially with each additional kilometer in search area.  To illustrate this with a simplified example, a circular crow fly buffer of a 1 km radius would encompass an area of 3.14 square kilometers (r2); a circular crow fly buffer of 5 kilometers would encompass an area twenty five times that size.  This put severe constraints on the ability for ArcMap to create spatial joins on network buffers that sometimes consisted of several hundred thousand properties, even after several iterations were run.  Luckily, a retrospective evaluation of the results revealed that the 1,036 observations that were returned were not compromised in their calculation of the housing mix variable.  Nevertheless, this reduced the sample size available for analysis, and the outcomes of  outcomes of models where housing mix is measured at 5 km should be interpreted with some caution.    141  Table 27: Housing mix defined around home postal code, descriptive statistics  Search Geography around Respondent’s Postal Code Statistic 500 m 1 km 2 km 3 km 5km       N  1,183   1,183   1,160   1,135   1,036        Mean 0.826 1.011 1.227 1.346 1.456 Median 0.879 1.105 1.324 1.431 1.506 Standard Deviation 0.397 0.384 0.328 0.250 0.162 Minimum 0.000 0.000 0.318 0.593 0.974 Maximum 1.623 1.688 1.711 1.692 1.674       Kurtosis -0.654 -0.188 0.020 0.155 0.241 Standard Error 0.142 0.142 0.144 0.145 0.152 Skewness -0.334 -0.729 -0.986 -0.984 -0.963 Standard Error 0.071 0.071 0.072 0.073 0.076        The same descriptive statistics on housing mix scores gathered around a respondent’s workplace postal code reveal similar trends (Table 28); mean housing mix scores rise, and standard deviations fall as the search geography increases.   Table 28: Housing mix defined around workplace postal code, descriptive statistics  Search Geography around Respondent’s Workplace Postal code Statistic 1 km 2 km 3 km     N 397 409 411     Mean 0.850 1.074 1.212 Median 0.950 1.242 1.316 Standard Deviation 0.462 0.415 0.359 Minimum 0.000 0.000 0.000 Maximum 1.715 1.687 1.683     Kurtosis -1.140 -0.976 -0.277 Standard Error 0.122 0.121 0.120 Skewness -0.301 -0.585 -0.840 Standard Error 0.244 0.241 0.240      142  The housing mix scores measured at a 2 km network buffer29 around respondent’s home postal codes are mapped in Figure 19 and compared against walkability index values.  Although housing mix scores were only calculated around the postal codes of CLASP survey respondents, and not all postal codes in the region, a trend becomes immediately clear: while the core downtown peninsula contains very low housing mix, the urban, walkable areas immediately off the downtown peninsula contain very high levels of housing mix.  High housing mix extends southeastward toward the suburban city of Surrey, mostly along a rapid transit corridor that has been the target for infill growth and residential intensification by the region since the mid-1980s (Tomalty 2002).  The high housing mix of the inner city areas immediately off the downtown peninsula runs somewhat counter to the claims of local observers that there is not sufficient housing diversity in the City of Vancouver’s neighbourhoods (Villagomez et al. 2012). Bands of low housing mix exist in the outer southeastern suburbs as well as in the wealthy, low density suburban communities on the north shore of Burrard Inlet, opposite the City of Vancouver.                                                        29 A 2 km buffer was chosen for mapping purposes since analysis in the subsequent chapter will reveal that predictive ability peaks at this spatial definition of housing mix. 143  Figure 19: Housing mix scores at 2km, Metro Vancouver  A cross demarcates the postal code in the sample with the lowest observed housing mix score of 0.05.  This postal code, in the outer suburban municipality of Maple Ridge, also contained a very low walkability score of -5.02, putting it squarely in the lowest quartile for walkability.  However, housing mix does not perfectly track walkability: the postal code with the highest housing mix score was located in the southeastern corner of the City of Vancouver in a postal code that did not even score in the top quartile of walkability scores.  Indeed, as Figure 20 reveals, the areas with the highest walkability scores in the region – the very high density neighbourhoods of the downtown Vancouver peninsula – also had some of the lowest housing 144  mix scores, since these areas were overwhelmingly characterized by multifamily highrise condominiums and apartments.   Figure 20: Housing mix at 2 km and walkability, central Metro Vancouver  A cluster analysis was performed in ArcMap using Anselin Local Moran’s I (Anselin 1995) to examine whether the clustering pattern of high and low housing mix values (measured at 2 km) was significantly different from what it might have been if due to random chance (Figure 21).  Given that residential intensification is subject to local land use and zoning bylaws, and the cooperation of established homeowners (Fischel 2004), we might expect many significant clusters of low and high housing mix.  Figure 21 confirms that the downtown peninsula and peripheral suburban areas contain significant clusters of low housing mix.  High 145  housing mix clusters exist in the east end of the City of Vancouver, and in a band extending across the eastern, inner suburbs.  Figure 21: Housing mix at 2km, spatial clustering, Metro Vancouver   The same cluster analysis results are mapped at a finer scale against median household income (by DA) and zoomed in on the City of Vancouver (Figure 22).  While it is important to recognize that the housing mix values are calculated based on properties extending 2 km away from the centroid that is indicated on the map, there still is a clear trend toward high housing mix values being significantly clustered in lower and middle income parts of the city, such as East Vancouver.     146  Figure 22: Housing mix clusters and median DA household income, central Metro Vancouver  At first glance, Figure 22 seems to align with Fischel’s (2004) homevoter hypothesis: higher income communities are successful at thwarting residential intensification efforts and keeping area incomes high.  As such, residential intensification is confined to neighbourhoods where lower income residents have less power over the planning process30.  Pettit (1993) documented the tactics to resist intensification efforts used by affluent homeowners in the single family home neighbourhoods of the city’s West Side in the early 1990s.  She found that homeowners were very organized and had a savvy understanding of urban design processes that                                                  30 Like the proverbial “chicken and egg” question, it is unclear whether household incomes are low in high housing mix areas because housing mix enables lower income residents to sort into these communities (as advocates would suggest), or whether areas that had lower incomes to begin with had less power to fight residential intensification and, therefore, a rise in housing mix. 147  enabled them to resist efforts to add density to these areas by working with local planners to change aesthetics guidelines in overlay districts (ibid).  Several years later, (Punter 2003, 173-181) documented how participants in a Local Area Plan31 for the west side community of Dunbar were successful in working with planners to retain the area’s single family home character against future intensification efforts.  In any case, Figure 22 suggests that areas with equally high walkability levels, both quite close to the region’s CBD, can have significantly different housing mix levels (compare against Figure 16).  These areas are also, apparently, areas with a greater proportion of lower income residents.  Does this mean that housing mix enables lower income households to match into Smart Growth areas?  This relationship will be explored in the next chapter.  5.2.2 Neighbourhood Preferences The degree to which respondents prefer Smart Growth neighbourhoods is measured by the NPS: the results of a PCA on 5 trade-off questions of neighbourhood design from the CLASP survey (see Chapter 4, section 4.4.3.2).  A positive NPS indicates that a respondent prefers Smart Growth neighbourhoods; a negative NPS indicates that a respondent prefers a suburban neighbourhood and an NPS of 0 indicates indifference between Smart Growth and suburbia.  The NPS scores of the CLASP sample and of various population subgroups are reported in Table 29.  These groups were chosen because they represent population subgroups that receive considerable attention with respect to housing concerns in Metro Vancouver, Canada, and abroad.  Briefly, Canadian housing policies have been criticized by observers for favouring ownership over                                                  31Instead of a city-wide official community plan, the City of Vancouver conducts strategic land use planning at the neighbourhood level through Local Area Plans.  A separate Local Area Plan is conducted for each of the 22 official neighbourhoods in the city (boundaries are shown in Figure 15).   148  renting (Hulchanski 2007), and low rental vacancy rates, particularly in the City of Vancouver and other municipalities in the centre of the region, have been an ongoing concern (Canada Mortgage and Housing Corporation 2014).  Seniors have been reported to have higher preferences for homes in Smart Growth communities, and accommodating the large and aging baby boom cohort poses a challenge to planners (Myers and Ryu 2008; Myers and Gearin 2001).  Some observers feel that the lack of suitable housing typologies in Smart Growth areas prohibits families from raising children (Villagomez 2011) and this perceived problem has received attention from the popular press (Egan 2005).  Finally, many media articles in the Vancouver region have opined on the possibility that demand for investment properties from wealthy foreign nationals, particularly from China, Hong Kong, and Taiwan has contributed to the region’s escalating home prices (Young 2015; Marlow and Jang 2014).   Table 29: Neighbourhood preference scores for selected population groups Statistic CLASP Sample Renters People over 60 Families with Children East Asians       N 1,186 478 373 222 125       Mean 7.007 9.217 7.123 3.329 6.310 Median 8.060 10.104 9.013 3.546 5.948 Standard Deviation 10.56 9.81 10.56 10.08 8.26 Minimum -22.38 -22.38 -22.38 -22.38 -16.09 Maximum 22.38 22.38 22.38 22.38 22.38       Kurtosis -.404 -.023 -.350 -.430 -.007 Standard Error .142 .223 .252 .325 .430 Skewness -.503 -.657 -.595 -.119 -.077 Standard Error .071 .112 .126 .163 .217        With mean and median scores all well above 0, Smart Growth is popular not only among all respondents but among the different population subgroups.  Renters and seniors express higher 149  preferences for Smart Growth than the sample average, while families with children and respondents of East Asian ancestry report lower preferences than average. Interestingly, apart from where average scores lie, the preference distributions of the different subgroups may be nearly identical.  With the exception of East Asians, there were respondents who indicated both the maximum (22.38) and minimum (-22.38) preference levels for Smart Growth among all population subgroups, while all subgroups showed similar standard deviations and were all approximately normal with respect to their kurtosis and skewness.  This suggests that – at least in Metro Vancouver - Smart Growth is not a niche product whose preference levels reflect the responses of a small, but intensely loyal, group.  Instead, Smart Growth neighbourhoods seem to be widely accepted by members of all groups, and that the kinds of neighbourhoods that are preferred run along the entirety of the suburban-Smart Growth spectrum. This suggests to policymakers that there is some importance in adequately supplying homes in neighbourhoods at all levels of Smart Growth intensity.  5.2.3 Neighbourhood Match As was described in the Methods chapter, neighbourhood match is, in fact, two separate variables estimated in two parallel models. Subjective neighbourhood match is a respondent’s direct assessment of their own neighbourhood compared to their preferences based on trade-off questions in a survey.  Objective neighbourhood match is a comparison of a respondent’s survey-elicited preference to objectively-measured qualities of the built environment (i.e. the walkability index score of the postal code in which they reside).  The binary outcome of being “matched”, or “mismatched”, is derived from the difference between choice and preference (the NPS), also known as a “neighbourhood match score”. A positive neighbourhood match score indicates that a 150  respondent lives in an environment with fewer Smart Growth attributes than they would prefer, a negative neighbourhood match score indicates that a respondent lives in an environment with more Smart Growth attributes than they would prefer, and a score of zero indicates that a respondent is exactly matched.  This approach mirrors that of Cao (2008), who used differences between perception and preference as evidence of whether access to Smart Growth communities was sufficient or insufficient. Table 30: Neighbourhood match score, descriptive statistics Statistic Subjective Neighbourhood Match Score Objective Neighbourhood Match Score    N  1,108   1,162     Mean .401 -2.510 Median .034 -.290 Standard Deviation 6.169 14.623 Minimum -17.96 -48.40 Maximum 18.79 40.06    Kurtosis .859 .721 Standard Error .147 .143 Skewness -.023 -.755 Standard Error .073 .072     Table 30 profiles the differences between subjective and objective neighbourhood match scores revealing some of the key differences between subjective perceptions and objective measures.  The mean subjective match score of 0.401, and the median subjective match score of 0.034, indicate that, on average, many people feel that they are matched; that the neighbourhood in which they live closely aligns to their preferences for Smart Growth or suburbia.  On the other hand, the mean objective neighbourhood match score of -2.510 suggests that a larger number of people are mismatched into areas with higher walkability (i.e. more Smart Growth attributes) 151  than they would prefer.  Although many of these households may not perceive that they live in the “wrong” kind of community, this type of “mismatch” may have other ramifications in the form of satisfaction and health.  The range of objective neighbourhood match score values is considerably larger than that of subjective neighbourhood match scores.  This reflects the fact that people’s perceptions of their neighbourhood, as they indicate on the survey, do not differ greatly from their perceptions of what they prefer.  However, people’s preferences for Smart Growth attributes were often quite different from the actual, measurable qualities of the neighbourhood in which they resided.  To conclude from Table 30, the typical respondent in the CLASP survey sample perceives that they live in a neighbourhood that closely aligns with their preferences (with a slight perception of being mismatched into a less Smart Growth area than desired), while the typical respondent may actually live in an area with more Smart Growth attributes than they desire. However, these circumstances probably do not apply equally to different groups in the population.  Renters, seniors (people over the age of 60), families with children, and East Asian respondents are analyzed for their match dynamics in Tables 31-34.  Based on subjective match averages, most population subgroups analyzed perceive that they are reasonably well matched.  The median subjective neighbourhood match score of 0 reported for owners, people under 60, families with children, and non-East Asian participants corresponds to people who provided  identical responses to both the preference and current neighbourhood questions and who, therefore, perceive that they are exactly matched.  Only seniors showed notably different responses.  People over 60 had a median subjective neighbourhood match score of .802, suggesting that seniors perceive that they live in neighbourhoods that have fewer Smart Growth attributes than they would desire.  Median objective neighbourhood match scores varied more 152  considerably.  Interestingly, renters’ were identically mismatched – slightly away from Smart Growth - whether measured subjectively or objectively (Table 31).  Owners, people under age 60, and non-East Asian respondents were more likely to be more mismatched toward more Smart Growth than preferred when neighbourhood environment was measured objectively.  Seniors (people over 60) were objectively mismatched in neighbourhoods that had fewer Smart Growth attributes than they preferred, somewhat validating their perceptions.      Table 31: Neighbourhood match score, descriptive statistics, renters vs. owners Tenure Renters Owners Neighbourhood Match Score Subjective Objective Subjective Objective      N 651 651 438 438      Mean 0.648 -1.21 0.187 -4.603 Median 0.216 0.216 .000 -1.819 Standard Deviation 6.273 12.91 5.919 15.235 Minimum -17.63 -48.3 -17.5 -48.4 Maximum 18.79 40.06 18.68 25.11      Kurtosis 0.65 1.754 1.158 0.366 Standard Error 0.191 0.191 0.233 0.233 Skewness 0.031 -0.932 -0.103 -0.849 Standard Error 0.096 0.096 0.117 0.117         153  Table 32: Neighbourhood match score, descriptive statistics, by age of respondent Age of respondent Under 60 Over 60 Neighbourhood Match Score Subjective Objective Subjective Objective      N 743 743 346 346      Mean .081 -3.542 1.281 -.498 Median .000 -1.126 .802 1.388 Standard Deviation 6.007 14.09 6.331 13.546 Minimum -17.50 -48.40 -17.63 -46.08 Maximum 18.79 40.06 18.75 28.23      Kurtosis .867 1.100 .795 1.227 Standard Error .179 .179 .261 .261 Skewness -.043 -.942 .007 -.941 Standard Error .090 .090 .131 .131       Table 33: Neighbourhood match score, descriptive statistics, by family type Family type Families with Children Childless Families Neighbourhood Match Score Subjective Objective Subjective Objective      N 201 201 888 888      Mean -.413 -.247 .660 -3.102 Median .000 .339 .139 -.631 Standard Deviation 6.195 10.79 6.107 14.56 Minimum -15.35 -48.30 -17.63 -48.40 Maximum 18.79 21.30 18.75 40.06      Kurtosis .295 1.214 .974 .902 Standard Error .341 .341 .164 .164 Skewness -.060 -.637 .003 -.911 Standard Error .172 .172 .082 .082         154  Table 34: Neighbourhood match score, descriptive statistics, East Asian vs. non-East Asian respondents Respondent’s ethnicity East Asian Non-East Asian Neighbourhood Match Score Subjective Objective Subjective Objective      N 111 111 978 978      Mean .537 -2.397 -.201 -4.142 Median .255 -.125 .000 -2.480 Standard Deviation 6.135 13.86 6.115 15.03 Minimum -17.63 -48.40 -16.02 -47.57 Maximum 18.79 40.06 18.75 21.88      Kurtosis .842 1.074 1.068 1.358 Standard Error .156 .156 .455 .455 Skewness -.047 -.907 .315 -1.120 Standard Error .078 .078 .229 .229       5.2.3.1 Categories of Neighbourhood Match Neighbourhood “matching” is analyzed not just as a binary outcome, but is also divided into discrete categories that are estimated using multinomial logit models (see Chapter 4, Section 4.4.3.3).  These categories include preferring Smart Growth and being matched, preferring suburbia and being matched, as well as being mismatched into higher Smart Growth environments than one would prefer, and being mismatched into lower Smart Growth environments than one would prefer. Table 35 highlights the frequencies of these observations for subjective and objective neighbourhood match definitions, while these matching categories are mapped in Figures 23 and 24 for subjective and objective neighbourhood match, respectively.      155  Table 35: Neighbourhood match categories, frequency of observations  Subjective Match Objective Match Match Category N % N %      Matched 554 50.0% 539 46.3% Not Matched 554 50.0% 624 53.7%      Prefers Smart Growth and Matched 463 41.8% 408 35.1% Prefers Suburbia and Matched 91 8.2% 131 11.3% Mismatched into higher Smart Growth 269 24.3% 344 29.6% Mismatched into lower Smart Growth 285 25.7% 280 24.1% TOTAL 1,108  1,163        Figure 23: Subjective neighbourhood match categories, Metro Vancouver  156  Figure 24: Objective neighbourhood match categories, Metro Vancouver  While table 35 demonstrates that exactly half of all respondents are matched subjectively, fewer than half are matched objectively.  When neighbourhood matching is defined objectively and broken into categories, a greater proportion of the sample is either matched into suburbia or mismatched into a higher Smart Growth environment than their preference compared to when respondents define match subjectively.  Figures 23 and 24 highlight these differences, while Figure 25 focuses on this discrepancy by mapping only households who were objectively matched, but not subjectively matched, and vice versa. Individuals who are objectively, but not subjectively, matched comprise people who perceive that they are “mismatched” but actually are 157  matched according to the availability of Smart Growth neighbourhoods in the region32.  The spatial distribution of these respondents is largely arranged around the “middle” of the Smart Growth spectrum.  That is, in established suburban communities and some of the dense, highly walkable, neighbourhoods of the City of Vancouver lying outside of the downtown peninsula.  As Figure 25 shows, inner city respondents who were objectively, but not subjectively, matched often perceived they lived in an environment with too many Smart Growth attributes for their tastes, while respondents in the same category in suburban areas felt they lived in areas with too few Smart Growth, or too many suburban, attributes for their tastes. In contrast, people who were subjectively, but not objectively, matched tended to live on the downtown peninsula, with a scattering of respondents living in peripheral suburban and exurban communities to the region’s east.  Not surprisingly, these people found themselves in the opposite circumstances as their subjectively matched, but objectively mismatched, counterparts: those who lived in exurban communities actually lived in communities that had too few Smart Growth attributes compared to their preferences (although they did not perceive this), while those who lived downtown lived in communities with too many Smart Growth attributes compared to their preferences.  This phenomenon, as well as comparisons of mean and median NPS and CNS scores (Table 29), suggest that people perceive their neighbourhood to align closely to their preferences; many downtown residents assessed their surroundings to have fewer Smart Growth attributes than in reality, while quite a few suburbanites with Smart Growth preferences considered their                                                  32 More correctly, if not succinctly, people who are objectively matched have preference scores which correspond within half a standard deviation to the equivalent walkability index score.  In other words, their preference score among the CLASP sample closely mirrors the walkability index score of their neighbourhood among the full distribution of neighbourhoods in Metro Vancouver. 158  neighbourhoods to be more urban in nature than the objective measures would suggest.  As Table 29 demonstrates, while residents indicated a strong preference for Smart Growth-oriented communities (a positive score indicates a preference for Smart Growth), these values are still considerably below the comparable walkability index scores for downtown neighbourhoods, many of which exceed a value of 15.  Not surprisingly, most respondents who were objectively, but not subjectively, matched occupied this middle ground between low density, outer suburban communities and the Manhattan-level densities of the downtown Vancouver peninsula.  In other words, when people evaluate their neighbourhood subjectively, a large number of them believe their area to be more “Smart Growth-oriented” than that neighbourhood might be among the total population of neighbourhoods in the region when measured objectively. Figure 25: Neighbourhood match categories, subjective and objective conflicts   159  5.2.4 Neighbourhood Satisfaction Unlike neighbourhood match, there seems to be no discernible spatial relationship to how satisfied respondents are with their neighbourhood (Figure 26).  In general, most respondents give a favourable impression of their community, with nearly a third of all respondents rating their neighbourhood a 10 out of 10;  less than a tenth of all surveyed rated their neighbourhood a 5 out of 10, or below (Table 36).  Figure 26: Neighbourhood satisfaction responses, Metro Vancouver    160  Table 36: Neighbourhood satisfaction, frequency of scores Satisfaction Score (0=very unsatisfied; 10 = very satisfied) N Percent of Total 0 4 0.3 1 8 0.7 2 5 0.4 3 19 1.6 4 17 1.4 5 57 4.8 6 74 6.2 7 170 14.3 8 264 22.3 9 185 15.6 10 383 32.3 TOTAL 1,186 100.0    Mean 8.15  Median 8  Standard Deviation 1.892  Question posed: “How satisfied are you with your neighbourhood (0 = dislike very much, 10=like very much)  A close-up of the most central part of the region, focused on Downtown Vancouver and the innermost neighbourhoods, shows that this trend is apparent at a finer scale.  Almost all respondents are generally satisfied, although there is a small pocket of dissatisfaction centered on the traditionally low income, inner city areas of the Downtown East Side and Strathcona. While high satisfaction scores predominate, there is a slight adjustment in mean satisfaction scores when the sample is stratified by objective and subjective neighbourhood match scores, and the type of neighbourhood matching or mismatching that respondents fall under.  When neighbourhood match is defined subjectively (Table 37), respondents who prefer Smart Growth and are successfully matched report the highest mean satisfaction levels, and respondents who are mismatched into a neighbourhood with more Smart Growth attributes than their preference report the lowest.  In contrast, when match is defined objectively (Table 38), people who prefer Smart Growth and are successfully matched report the lowest mean 161  Figure 27: Neighbourhood satisfaction responses, central Metro Vancouver  satisfaction levels, while people who are mismatched into lower Smart Growth areas than they prefer actually report the highest.  Still, these differences are slight, and because they are just descriptive results, we cannot infer with any confidence that Smart Growth environments lead to lower satisfaction levels, per se. One observation gleaned from the map of satisfaction in central Vancouver (Figure 27) is that respondents living in low income inner city areas seemingly report lower satisfaction levels than respondents living in higher income areas with a similar built form (for simplicity, only 162  income is mapped, not walkability as in other maps).  This might suggest that poverty, which is a personal characteristic that technically should not have any bearing on neighbourhood satisfaction, nevertheless exerts a strong impact on life satisfaction, which cannot be psychologically divorced from neighbourhood satisfaction (Campbell, Converse, and Rodgers 1976).  In the case of objectively-matched Smart Growth dwellers reporting lower rates of satisfaction, it is possible that many low income people “prefer” Smart Growth areas because the availability of rental stock and a good public transit system (and their financial inability to afford car ownership) compels them to choose and anchor their preferences in urban areas, rather than in suburban areas that they may more freely prefer if they had the means.  These people may, in turn, report lower neighbourhood satisfaction levels.  These challenges will be discussed in considerable detail in the final chapter. Table 37: Neighbourhood satisfaction, descriptive statistics, subjective neighbourhood match categories   Matched Mismatched Satisfaction Score  Prefers SG Prefers Suburbia Into Higher SG Into Lower SG      N 463 91 269 285      0 (very unsatisfied) 0.2% 0.0% 0.0% 1.1% 1 0.4% 1.1% 0.7% 0.7% 2 0.2% 0.0% 0.4% 0.7% 3 1.3% 1.1% 1.9% 1.4% 4 0.9% 0.0% 2.6% 1.4% 5 5.6% 3.3% 5.6% 2.1% 6 5.2% 8.8% 6.3% 5.3% 7 10.8% 14.3% 18.6% 13.0% 8 20.5% 25.3% 24.2% 23.9% 9 18.4% 16.5% 13.8% 14.7% 10 (very satisfied) 36.5% 29.7% 26.0% 35.8%      Mean 8.37 8.21 7.91 8.26 Median 9 8 8 9 Standard Deviation 1.795 1.716 1.872 1.975  163  Table 38: Neighbourhood satisfaction, descriptive statistics, objective neighbourhood match categories  Matched Mismatched Satisfaction Score  Prefers SG Prefers Suburbia Into Higher SG Into Lower SG      N 408 131 344 280      0 (very unsatisfied) 0.7% 0.0% 0.3% 0.0% 1 1.0% 0.8% 0.9% 0.0% 2 0.5% 0.0% 0.3% 0.7% 3 1.7% 2.3% 1.5% 1.4% 4 1.0% 0.8% 2.9% 0.4% 5 5.1% 2.3% 6.4% 2.9% 6 6.9% 6.1% 6.7% 5.0% 7 13.7% 13.7% 16.3% 13.6% 8 22.8% 26.7% 21.8% 20.4% 9 16.2% 14.5% 13.4% 17.5% 10 (very satisfied) 30.4% 32.8% 29.7% 38.2%      Mean 8.06 8.27 7.95 8.49 Median 8 8 8 9 Standard Deviation 1.993 1.75 1.97 1.655  5.2.5 Health Outcome Variables In the final research question, the effect of neighbourhood matching was estimated on two health outcome variables: self-reported health status and body mass index (BMI).    5.2.5.1 Self-Reported Health Status Table 39 profiles the frequency of self-reported health status scores in the sample.  These scores are stratified by subjective neighbourhood match (Table 40) and objective neighbourhood match (Table 41) and disaggregated by the four neighbourhood match categories. In contrast to neighbourhood satisfaction, average self-reported health status scores seem to lie toward the middle of possible choices.  There is no considerable difference in health scores when the results are stratified by objective and subjective neighbourhood match, nor within matching categories.  A map of self-reported health status scores also shows no spatial trends, although a noticeable 164  pocket of poor health scores can be observed in the lower income Whalley neighbourhood of the City of Surrey (Figure 28). Figure 28: Self-reported health scores, Metro Vancouver  A closer view (Figure 29) reveals that these pockets may just be a few isolated incidences surrounded by fairly typical responses, and that one’s self-reported health status does not seem closely related to the median household income of the DA.  Nor does health status seem to be linked to the walkability of an area (compare against the walkability map in Figure 15).These lack of trends may underscore the problem with inferring too much from a self-reported health measure which may be influenced by many other factors, few of which were queried in this residential preference survey.    165  Figure 29: Self-reported health scores and median DA household income  Table 39: Self-reported health status, frequency of scores Self-Reported Health Status  N Percent of Total 1 = “poor” 46 3.9 2 = “fair” 239 20.2 3 = “good” 428 36.1 4 = “very good” 329 27.7 5 = “excellent” 138 11.6 6 = “don’t know/prefer not to say” 6 0.5 TOTAL 1,186 100.0    Mean1 2.77  Median1 3  Standard Deviation1 1.026  Question posed: “In general, compared to people your own age, would you say your health is (select one)” 1. “Don’t know/prefer not to say” responses (Category #6) omitted.  166  Table 40: Self-reported health status, descriptive statistics, subjective neighbourhood match categories  Matched Mismatched Self-Reported Health Status  Prefers SG Prefers Suburbia Into Higher SG Into Lower SG      N 463 91 269 285      1 (poor) 3.0% 1.1% 4.5% 4.9% 2 19.2% 25.3% 22.3% 16.8% 3 35.6% 30.8% 35.7% 40.4% 4 28.7% 30.8% 28.6% 24.6% 5 (excellent) 13.2% 12.1% 8.2% 12.3% Don’t know/prefer not to say 0.2% 0.0% 0.7% 1.1%      Mean1 3.30 3.27 3.14 3.23 Median1 3 3 3 3 Standard Deviation1 1.02 1.012 1.004 1.032 Question posed: “In general, compared to people your own age, would you say your health is…(select one)”  1. “Don’t know/prefer not to say” responses (Category #6) omitted  Table 41: Self-reported health status, descriptive statistics, objective neighbourhood match categories  Matched Mismatched Self-Reported Health Status  Prefers SG Prefers Suburbia Into Higher SG Into Lower SG      N 408 131 344 280      1 (poor) 3.9% 3.8% 2.9% 5.0% 2 17.4% 26.7% 22.4% 17.9% 3 41.4% 32.1% 33.1% 33.3% 4 27.2% 26.7% 27.6% 30.1% 5 (excellent) 9.1% 9.9% 14.0% 13.6% Don’t know/prefer not to say 1.0% 0.8% 0.0% 0.4%      Mean1 3.20 3.12 3.27 3.29 Median1 3.00 3 3 3 Standard Deviation1 0.97 1.042 1.05 1.07 Question posed: “In general, compared to people your own age, would you say your health is…(select one)”  1. “Don’t know/prefer not to say” responses (Category #6) omitted    167  5.2.5.2 BMI measures The mean BMI for the CLASP sample was 25.85 (Table 42).  Among the 1,082 participants who supplied information on their height and weight, 190 individuals had a BMI over 30 kg/m2 and could be classified as ‘obese’.  These figures are almost identical to the 17.4% of Metro Vancouver residents estimated to be obese by the Canadian Community Health Survey (Navaneelan and Janz 2014).  Table 42: BMI variables, descriptive statistics Statistic BMI logBMI    N 1,082 1,082    Mean 25.85 3.24 Median 25.09 3.22 Standard Deviation 4.861 0.182 Minimum 13.67 2.62 Maximum 43.40 3.77    Kurtosis .746 -.013 Standard Error .149 .149 Skewness .830 .308 Standard Error .074 .074      BMI does not appear to be related to the type of neighbourhood matching category.  As Table 43 shows, means, medians and standard deviations are all roughly similar across all groups, regardless of whether people prefer Smart Growth or suburbia and are subjectively matched or mismatched.  Subjective match, of course, is based on an individual’s perception of the Smart Growth or suburban attributes of their community, and not the actual objectively measured built form attributes that have been demonstrated to be linked to BMI outcomes (Frank et al 2006), so it is perhaps not surprising that being subjectively matched or mismatched has seemingly little bearing on BMI scores.  However, these findings are replicated across categories 168  of objective neighbourhood match (Table 44), although individuals who prefer and are matched into suburbia seem to have marginally higher BMI scores than others.  Respondents who are objectively mismatched into areas of higher Smart Growth than their preferences, or areas of lower Smart Growth than their preferences have very comparable BMI scores.  This finding may be due to the fact that people who were objectively mismatched are scattered across the region (See Figure 23), since mismatch is based on one’s initial preference level.  For example, two hypothetical households living next to one another in a reasonably walkable area may fall under opposite mismatch categories: the individual who prefers an area with a lot of Smart Growth features would be mismatched into a lower Smart Growth neighbourhood, while his/her neighbour, who prefers areas with very few Smart Growth features would effectively be mismatched into an area with more Smart Growth attributes.  This echoes the findings of Frank et al. (2006), who show that the built environment one finds oneself in may matter in explaining walking behaviour and BMI more than one’s preferences.  Table 43: BMI, descriptive statistics, subjective neighbourhood match categories  Matched Mismatched Statistic Prefers SG Prefers Suburbia Into Higher SG Into Lower SG      N 428 81 249 256      Mean 25.90 25.86 26.02 25.80 Median 24.96 25.02 25.29 25.18 Standard Deviation 4.919 4.812 5.199 4.648 Minimum 16.46 18.14 16.64 13.67 Maximum 43.40 42.29 41.66 41.60          169  Table 44: BMI, descriptive statistics, objective neighbourhood match categories  Matched Mismatched Statistic Prefers SG Prefers Suburbia Into Higher SG Into Lower SG       N 372 116 316 258      Mean 25.90 26.65 25.79 25.78 Median 25.09 25.79 25.15 25.08 Standard Deviation 5.083 5.305 4.937 4.226 Minimum 13.67 16.64 16.95 17.03 Maximum 42.18 41.00 43.40 41.04       Finally, figure 30 maps BMI against walkability.  While a discernible trend is far from clear, it would appear as if the spatial distribution of overweight and obese individuals increases in suburban areas of low walkability as existing evidence would suggest.   170  Figure 30: BMI observations, Metro Vancouver  5.2.6 Conclusions Descriptive statistics were presented to gain insight into how representative the survey sample was to the population, and also to understand the characteristics and spatial distribution of key variables.  On the whole, the sample was fairly representative of the population in terms of the geographical distribution of respondents, housing characteristics, and household SES.  Some demographic characteristics were atypical: sample respondents were somewhat older, less likely to have children, and more educated than the average Metro Vancouver resident.  Ethnic representation was difficult to compare to the region because of the way in which the 2011 NHS gathered information on ethnicity.   171  In terms of variable characteristics, the descriptive statistics for housing mix was presented at 5 different network buffer definitions, ranging from 500 meters to 5 km.  As expected, mean and median scores increase as the search geography widens to take into account more spatial heterogeneity.  The maximum score remains relatively similar across all 5 geographies, indicating that there are highly mixed neighbourhoods even at small spatial scales.  The distribution of housing mix scores was mapped at a 2 km definition, and showed that the lowest housing mix scores tended to cluster on the downtown peninsula, while the highest housing mix scores were clustered in inner city neighbourhoods of the City of Vancouver immediately off the downtown peninsula, particularly on the City’s eastern side.  Areas of high housing mix can nevertheless be found across the entire region. Neighbourhood preference scores demonstrate that the average household prefers Smart Growth over suburban neighbourhood designs, with Smart Growth preferences extending to different subgroups of the population.  Renters, in particular, show a strong preference for Smart Growth communities.  The average survey respondent was reasonably well matched into the neighbourhood type of their choice when match was measured subjectively.  When measured objectively, there is a slight tendency for households to be mismatched into areas with more Smart Growth attributes than they express preferences for.  When neighbourhood match is divided into four categories, a higher proportion of respondents prefer Smart Growth and are matched when measured subjectively than objectively, and a higher proportion of respondents are mismatched into an environment with more Smart Growth attributes than their preferences when measured objectively rather than subjectively.  Most households express high levels of satisfaction with their neighbourhood, although there are some pockets of dissatisfaction in lower income areas of the central parts of the region.  For self-reported health status, respondents 172  reported fairly typical scores, with no apparent spatial trends or differences in terms of neighbourhood matching ability.  BMI scores in the survey sample are nearly identical to those found in the Canadian Community Health Survey.  Respondents who prefer and are matched into suburban neighbourhood environments report marginally higher BMI scores than others. 173  Chapter 6: Model Results 6.1 Introduction This chapter describes the results of the statistical models that were used to answer the three research questions.  The results are presented in order, beginning with the first research question.  The chapter concludes by tying the results of all three research questions together.  6.2 Statistical Models Answering Research Question 1 The first research question of this dissertation seeks to answer whether housing mix significantly predicts the ability for households to match the neighbourhood type that they live in with the neighbourhood type of their preferences.  Both “subjective” and “objective” definitions of match are predicted in parallel but separate models.   The analysis begins with simple logistic regression models that estimate the effect housing mix, by itself, might have on the ability to match as a binary outcome.  These series of models separately estimate the effect of housing mix at different spatial scales, and separate models are presented for housing mix measured around a respondent’s home postal code and a respondent’s workplace postal code.  From there, analysis progresses to multinomial logistic models estimating how housing mix predicts different matching categories; that is, estimating the probability that respondents with preferences for Smart Growth or Suburbia ended up matched or mismatched into an environment with more Smart Growth preferences than they prefer.  In these models, respondents who were mismatched into areas with fewer Smart Growth attributes were treated as the reference group.  While the odds ratios of belonging to these categories are estimated, housing mix at different spatial scales remain the only independent variables used in these analyses.   174  The most important series of models used to answer research question 1 are the multivariate models that are presented in section 6.2.3.  These multinomial logit models predict whether housing mix, measured by a network buffer of 2 km around a respondent’s home postal code, increases the odds of respondents with Smart Growth preferences matching into Smart Growth neighbourhoods compared to mismatching into neighbourhoods with fewer Smart Growth attributes.  Relevant personal characteristics and attributes of the surrounding built environment are controlled for. The same models are then separately estimated using stratified samples of respondents over 60 years of age versus those under 60, and renters versus owners.  The intent of these separate analyses is to test the effect of housing mix on match in subgroups that are often the intended target of housing mix and affordability policies (see Chapter 5).  6.2.1 Simple Models Predicting Neighbourhood Match To begin to understand whether housing mix is a significant predictor of neighbourhood match, a simple logistic regression model consisting of only one independent variable – housing mix – is run predicting the probability of neighbourhood matching, either subjectively or objectively.   Table 45 displays how one standard deviation increase in housing mix around a respondent’s home postal code predicts the odds associated with being subjectively matched into the neighbourhood type of their choice.  The search geographies represent the size of the network buffer around which housing mix is calculated. For example, a search geography of 500 m indicates that the housing mix score comprises every property within the area encompassed by a 500 m road-based path from the centroid of the postal code in which the respondent resides.   175  Table 45: Binary logistic regression predicting subjective neighbourhood match, housing mix defined around home postal code  Search geography 500 m 1 km 2 km 3 km 5km       Constant 0.998 0.995 1.003 0.993 1.006 Sig. .976 .928 .956 .904 .923       Z_HsgMix 1.001 0.988 0.895 0.985 0.981 Sig. .980 .841 .071 .806 .765       N  1,105   1,105   1,085   1,060   981  R2 (Nagelkerke) .000 .000 .004 .000 .000 * Significant at p. <0.05  ** Significant at p.<0.01  ***Significant at p.<0.001  As Table 45 demonstrates, housing mix around a respondent’s home address is not a significant predictor of subjective match, although defining housing mix within a 2 km network buffer is weakly significant (at p<0.10)  and, with an odds ratio of 0.895, actually associated with lowering the probability of subjectively matching.  The predictive power of these models is also very weak or negligible: Pseudo-R2 values do not exceed 0.004.  Given that the ability to live in the neighbourhood type of your preference may depend on a range of other factors – from price to the amount of financial resources at one’s disposal, to a host of other personal characteristics – the singular effect of housing mix in predicting match may not be expected to be high.  Nevertheless, Table 45 reveals that the data does not seem to fit the models at all.   Does prediction improve if housing mix is defined around a person’s workplace?  There is a possibility that many respondents choose homes close to their workplaces, so the housing mix around one’s place of employment may have more of a bearing on one’s ability to match.  Unfortunately, as Table 46 shows, housing mix within 3 different network buffer distances of a respondent’s workplace does not significantly alter the odds of subjective matching, and the data does not fit the model at all.   176  Table 46: Binary logistic regression predicting subjective neighbourhood match, housing mix defined around workplace postal code Search geography 1 km 2 km 3 km     Constant 0.818 0.822 0.832 Sig. .054 .058 .074     Z_HsgMix 1.023 1.007 .979 Sig. .832 .946 .839     N  369   379   381  R2 (Nagelkerke) .000 .000 .000 * Significant at p. <0.05  ** Significant at p.<0.01  ***Significant at p.<0.001  When match is defined objectively rather than subjectively (Table 47), then housing mix significantly increases the probability that respondents will match into the neighbourhood type that they prefer.  Significant associations, however, are only found when housing mix is measured within 500 meters, 2 km, and 5 km of a respondent’s address.  The predictive power remains low, but peaks at a Pseudo-R2 of 0.060 when housing mix is defined within a 2 km network buffer of a respondent’s home postal code.  In any case, the model fit is considerably better than in the parallel model predicting subjective neighbourhood match.  Meanwhile, housing mix measured around a respondent’s work address remains an insignificant predictor of objective neighbourhood match, much as it was an insignificant predictor of subjective neighbourhood match (Table 48).   177  Table 47: Binary logistic regression predicting objective neighbourhood match, housing mix defined around home postal code  Search geography 500 m 1 km 2 km 3 km 5km       Constant 0.840 0.842 0.835 0.864 0.847 Sig. .003** .003** .003** .014* .008**       Z_HsgMix 1.265 1.116 1.574 1.109 1.135 Sig. .000*** .064 .000*** .086 .045*       N  1,179   1,178   1,155   1,130   1,031  R2 (Nagelkerke) .018 .004 .060 .004 .005 * Significant at p. <0.05  ** Significant at p.<0.01  ***Significant at p.<0.001  Table 48: Binary logistic regression predicting objective neighbourhood match, housing mix defined around workplace postal code  Search geography 1 km 2 km 3 km     Constant 0.824 0.820 0.821 Sig. .056 .047* .048*     Z_HsgMix 0.945 0.952 0.984 Sig. .575 .625 .875     N  394   406   408  R2 (Nagelkerke) .001 .001 .000 * Significant at p. <0.05  ** Significant at p.<0.01  ***Significant at p.<0.001  Why does housing mix significantly predict objective while failing to predict subjective neighbourhood match?  The answer may lie in how respondents perceive the Smart Growth qualities of their neighbourhood, and the distribution of subjectively versus objectively matched respondents.  Many respondents lived on the downtown peninsula, a community with an average 178  density of 17,139/km2 (44,389/square mile)33 - whose postal codes all lie at the extreme end of the region’s walkability index scores. Because the downtown peninsula consists almost overwhelmingly of multifamily apartment units, housing mix scores are among the lowest in the region. If a sufficient number of downtown residents feel that they are matched with their neighbourhood preferences, it is possible that housing mix will not be a significant predictor of subjective neighbourhood match.  Many downtown respondents expressed only moderate preference levels for Smart Growth communities, so they were objectively mismatched into neighbourhoods that contained more Smart Growth attributes than they preferred.  However, when they were asked to evaluate their surroundings, most downtown respondents felt that they lived in communities with modest levels of Smart Growth attributes.  Based on their perceptions, they were subjectively matched.  By way of example, a respondent with a clear preference for suburban neighbourhood attributes subjectively reported that the residential environment surrounding their Yaletown address (a downtown neighbourhood consisting almost exclusively of high-rise condominium towers with a population density of over 13,000/km2) featured those same suburban qualities.  This respondent was therefore subjectively “matched” according to his or her suburban preferences. In contrast, most downtown respondents were objectively “mismatched” – residing in a neighbourhood that had many more Smart Growth attributes than they preferred.  These results highlight the tenuousness of relying solely on self-reported assessments of the built environment.  To test a phenomenon such as an individual’s ability to match into their preferred neighbourhood type, measuring “subjective match” alone may, at first glance, appear                                                  33 Based on 2011 National Household Survey 179  to suffice. After all, if people perceive that they live in the neighbourhood type of their preference, why prod any further?   However, given that Smart Growth built environments have been extensively linked to social and economic benefits (see Chapter 2, section 2.1.2), and that people may be unreliable at assessing the built environment in which they live, it is important to also define match based on actual, measurable attributes of the built environment (i.e. objective neighbourhood match).  If objective neighbourhood match leads to positive health and satisfaction outcomes, and housing mix is a significant and positive predictor of objective neighbourhood match, then we may infer that housing mix is a worthwhile policy for planners to pursue.   Nevertheless, individual behaviour may be more a function of the perceptions people make of their environment.  An area may be objectively walkable, but if individuals perceive that the area presents obstacles to walking, they will curtail the amount of walking they partake in (Gebel et al. 2011).  Subjective match – the perception that one lives in their preferred environment - is still an important consideration that should not be dismissed, and is not dismissed from further analysis in this dissertation.  The outcomes from both subjective and neighbourhood match models will be analyzed in parallel for all questions in this dissertation.  6.2.2 Simple Models Predicting Neighbourhood Matching Category  While housing mix, on its own, may or may not significantly predict whether people match, does housing mix differently affect the ability for people with preferences for different neighbourhood types to match?  Given that housing mix policies in Metro Vancouver are intended to assist households to live in “complete” (i.e. Smart Growth) communities, and given 180  that regional observers - as well as advocates of Smart Growth principles - champion housing mix as an approach to residential matching, there is much more interest in understanding whether housing mix will help people with Smart Growth preferences match into these kinds of neighbourhoods, than whether it will assist people who prefer suburbia to find housing in those kinds of environments.  A simple binary “match” outcome will not distinguish between households who prefer Smart Growth and are successfully matched and households that prefer suburbia and are successfully matched. In the second set of models, a multinomial logit regression is run predicting how a one standard deviation increase in housing mix changes the odds associated with belonging to 3 different categories of match: - Preferring Smart Growth and being matched - Preferring suburbia and being matched - Being mismatched into a higher Smart Growth environment than their preference Because the impetus is on understanding whether housing mix is associated with the ability to match people with preferences for Smart Growth, the reference group with which those three categories are compared are respondents who end up mismatched into a neighbourhood with fewer Smart Growth attributes (i.e. more suburban qualities) than they express preferences for.     181  Table 49: Multinomial logistic regression predicting subjective neighbourhood match, housing mix defined around home postal code  Match Category N Z_HMix_ 500m Z_HMix_ 1km Z_HMix_ 2km Z_HMix_ 3km Z_HMix_ 5km        Prefers Smart Growth and matched 461 1.099 .959 .876 .926 1.029 Sig.  .222 .584 .102 .335 .734        Prefers suburbia and matched 91 .590 .958 .594 .836 .537 Sig.  .000*** .725 .000*** .142 .000***        Mismatched into higher Smart Growth 269 .967 .941 .840 .856 .874 Sig.  .693 .479 .052 .077 .144        Mismatched into lower Smart Growth 284 ref ref ref ref ref        R2 (Nagelkerke)  .029 .001 .020 .004 .036  Significant at p. <0.05 ** Significant at p.<0.01  ***Significant at p.<0.001  Table 49 shows the results of the multinomial logit model predicting how housing mix affects the odds of belonging to these 3 categories if they are defined using the subjective match approach.  Housing mix does not significantly increase the odds that people will belong to the “Prefers Smart Growth and matched” category than the “mismatched into lower Smart Growth” category.  In other words, housing mix is not significantly associated with subjectively matching households to their Smart Growth preferences.  On the other hand, housing mix significantly lowers the odds that respondents will belong to the “prefers suburbia and matched” category, at least if housing mix is measured within 500m, 2km and 3km of a respondent’s address.  This suggests that individuals who prefer and perceive that they live in suburbia are more likely to live in an area with low housing mix.  Model fit is substantially improved over the binary model presented in Table 45, but still remains relatively low. The odds of belonging to the different 182  subjective match categories compared to the reference group is unaffected by housing mix when housing mix is measured around a respondent’s workplace, and model fit remains poor (Table 50).  Table 50: Multinomial logistic regression predicting subjective neighbourhood match categories, housing mix defined around workplace postal code  Search geography N Z_HMix_ 1km_work Z_HMix_ 2km_work Z_HMix_ 3km_work      Prefers Smart Growth and matched 134 1.173 1.065 1.048 Sig.  .226 .634 .719      Prefers suburbia and matched 32 1.006 .938 .919 Sig.  .976 .747 .664      Mismatched into higher Smart Growth 100 1.247 1.065 1.090 Sig.  .121 .658 .545      Mismatched into lower Smart Growth 103 ref ref ref      R2 (Nagelkerke)  .009 .002 .003 * Significant at p. <0.05 ** Significant at p.<0.01  ***Significant at p.<0.001    When match is measured objectively, rather than subjectively, the outcome of the model changes considerably (Table 51).  Increasing the housing mix by one standard deviation improves the odds that a respondent who prefers Smart Growth will be matched by nearly 72% (when housing mix is measured at 2 km).  At the same time, increased housing mix lowers the odds of being matched into suburbia and lowers the odds that a household will be mismatched into an environment with more Smart Growth attributes (i.e. a more walkable neighbourhood than they would prefer).  This last category is likely tied to the considerable number of households who express moderate Smart Growth preferences but live on the downtown 183  peninsula – a very dense, extremely walkable residential environment that is almost exclusively composed of highrise, multifamily towers and some of the lowest housing mix scores in the region.  Respondents who live on the downtown peninsula form the overwhelming majority of respondents who are objectively mismatched into higher Smart Growth environments than their preference.  At the same time, model fit is considerably improved; when housing mix is measured at 2 km, the influence of this variable alone seems to predict nearly 16% of the variance in the model. Table 51: Multinomial logistic regression predicting objective neighbourhood match categories, housing mix defined around home postal code  Match Category N Z_HMix_ 500m Z_HMix_ 1km Z_HMix_ 2km Z_HMix_ 3km Z_HMix_ 5km        Prefers Smart Growth and matched 408 1.520 1.049 1.718 1.118 1.290 Sig.  .000*** .564 .000*** .191 .008**        Prefers suburbia and matched 131 .579 .814 .536 .716 .523 Sig.  .000*** .053 .000*** .001** .000***        Mismatched into higher Smart Growth 344 .936 .855 .601 .728 .798 Sig.  .433 .059 .000*** .859 .011*        Mismatched into lower Smart Growth 277 ref ref ref ref ref        R2 (Nagelkerke)  .083 .010 .156 .023 .084 * Significant at p. <0.05  ** Significant at p.<0.01  ***Significant at p.<0.001  In contrast, measuring housing mix around the workplace remains an insignificant predictor of the odds of belonging to each of the three objective match categories (Table 52).  Not only does housing mix around a workplace not significantly predict the ability to match for 184  any group, but increasing the search geography does not make any discernible difference to the odds of matching nor is predictive power improved.  Measuring housing mix around the workplace will not be considered in any further models. Table 52: Multinomial logistic regression predicting objective neighbourhood match categories, housing mix defined around workplace postal code  Search geography N Z_HMix_ 1km_work Z_HMix_ 2km_work Z_HMix_ 3km_work      Prefers Smart Growth and matched 131 1.036 1.036 1.045 Sig.  .787 .784 .729      Prefers suburbia and matched 47 1.060 1.030 1.065 Sig.  .741 .865 .712      Mismatched into higher Smart Growth 108 1.137 1.120 1.087 Sig.  .359 .407 .535      Mismatched into lower Smart Growth 100 ref ref ref      R2 (Nagelkerke)  .003 .002 .001 * Significant at p. <0.05  ** Significant at p.<0.01  ***Significant at p.<0.001  When measured around a respondent’s home address, increasing housing mix at certain search geographies is significantly associated with raising the odds of being objectively matched into a community with Smart Growth attributes that align with the respondent’s preferences compared to being objectively mismatched into a neighbourhood with fewer Smart Growth attributes than the respondent prefers.  At the same time, increasing housing mix lowers the odds that a respondent with suburban preferences lives in the neighbourhood type of their choice and lowers the odds that a respondent is mismatched into a community with more Smart Growth 185  attributes than they prefer.  This suggests that housing mix may be a useful policy tool to match people into moderate level Smart Growth environments – neither peripheral suburban areas that are characterized overwhelmingly by low density housing options (i.e. single family homes and some ground-oriented attached housing), nor in the very dense downtown core characterized overwhelmingly by high density, multifamily housing types.   6.2.3  Discrepancies in Matching Outcomes at Different Spatial Scales Why is housing mix only a significant predictor of match at certain buffer distances? Moreover, why do significant search geographies exist at the smallest, largest and mid-sized buffers of the geographic spectrum (i.e. at 500m, 2 km, and 5 km)?  For example, the measures of central tendency showed 500 m and 1 km housing mix scores to be very similar to one another  (Table 5 in Chapter 5), even though one measure significantly predicts objective matching ability, while the other does not.  Similarly, measuring housing mix around a 5 km buffer, which contains the least amount of heterogeneity in its observations, is a significant predictor of objective neighbourhood match, while measuring housing mix around a 3 km network buffer is not. Figures 31 and 32 provide a possible explanation for why these discrepancies occur.  As was shown in the previous chapter, many respondents who were not objectively matched lived on the downtown peninsula, where the built environment was too dense for many people’s preferences.  The downtown peninsula was also shown to contain some of the lowest housing mix scores in the region because this area overwhelmingly consisted of multifamily apartments and condominiums.  When housing mix is measured around a 500 m buffer of a postal code centroid on the edge of the downtown peninsula – such as postal code V6Z 2Z5 – it only 186  captures other downtown properties (Figure 31).  On the other hand, when housing mix is measured around a 1 km buffer of the same centroid, the buffer often extends across bridges into neighbourhoods with more housing variety.  For example, a 1 km network buffer drawn around V6Z 2Z5 travels across the Cambie Bridge to encompass parts of the South False Creek area which contain more townhomes and even some single family homes.  This raises the housing mix score between 500 m and 1 km, even though the downtown-dwelling participant belonging to that observation is not objectively matched.  When the search geography is increased again to 2 km, the housing mix score may be raised again but, at the same time, the housing mix scores in locations where respondents are objectively matched – such as in the inner city neighbourhoods of East Vancouver and the eastern inner suburbs – are raised as well.  The inability to predict match at 1 km and 3 km may be a function of how housing mix scores are calculated for downtown respondents who do not objectively match relative to inner city and inner suburban respondents who do.      187  Figure 31: Housing mix at 500 m, central Vancouver  Figure 32: Housing mix at 1 km, central Vancouver  188  While housing mix may be associated with objective neighbourhood matching, does this relationship still hold if other, perhaps more relevant, factors are controlled for?  The ability for households to match into their neighbourhood may depend much more on their income (i.e. their ability to pay), and the importance they assign to other attributes of homes and neighbourhoods other than just Smart Growth characteristics, than on the diversity of housing options that may allow people to sort into these communities.  Additionally, certain population groups may be at a distinct advantage or disadvantage in sorting into the neighbourhood areas of their choice based on the availability of housing that may suit their needs.  In particular, families with children may be constrained by the costs of finding suitably-sized homes, particularly in Smart Growth neighbourhoods which are typically more centrally-located and command a price premium, while seniors – who typically report higher preference levels for Smart Growth areas – may possibly have greater success at downsizing.  Similarly, renters may be restricted from moving into desired areas because the rental stock – or the type of home that is usually rented out – may be confined to certain parts of the region.  6.2.4 Multivariate Models Predicting Neighbourhood Matching Category In the final series of models for the first research question, a multinomial logit model is performed to test whether housing mix remains a significant predictor of objective match - or becomes a significant predictor of subjective match – when relevant personal characteristics are controlled for.  Since housing mix’s predictive power peaked when it was defined around a 2 km network buffer for both subjective and objective neighbourhood match models, only housing mix defined at 2 km will be considered in these models.  For the sake of brevity, and because the ability to match people with Smart Growth preferences into Smart Growth neighbourhoods is 189  under much more scrutiny, results will only be reported predicting the odds of matching respondents into Smart Growth preferences compared to being mismatched into areas of lower Smart Growth (the reference group).  The three series of covariates included the neighbourhood importance factors – factors which proxy for the importance people assign to features of homes and neighbourhoods that are unrelated to Smart Growth – the principal component proxying for objectively measured structural attributes of homes (BuiltForm_2km_Factor), and household income.  Together, these three series of covariates were chosen for the final model for two reasons.  First, these variables were considered to best represent both the attributes of properties that households consider in home-purchasing or rental decisions, as well as their ability to pay (i.e. income).  Secondly, the variables were relatively uncorrelated with one another.  The inclusion of further predictor variables, such as level of education and median property values, would have increased the likelihood of multicollinearity.  Table 53 reports the results of these multinomial logit models predicting the odds of subjectively matching respondents with Smart Growth preferences.  Four models are run, with each model adding a new variable or series of variables.  Model 1 only includes the neighbourhood importance factors (see Chapter 4, section 4.4.4) as predictor variables.  All importance factors appear to be significant, although their effect on increasing or decreasing the odds of residential matching vary considerably.  People who express high preferences for nearby neighbourhood amenities (NeighbAm_factor) – such as stores and services within walking distances – have 42% higher odds of being Smart Growth matched than being mismatched into neighbourhoods with less Smart Growth attributes than their preferences.     190  Table 53: Multinomial logistic regression results predicting odds of subjectively matching neighbourhood to Smart Growth preferences  Model Number1 Predictor (1) (2) (3) (4)      Z_HsgMix_2km    .923 Sig.    .505      NeighbAm_factor 1.419 1.317 1.298 1.301 Sig. .000*** .007** .016* .015* RegAccess_factor 1.185 .994 .984 .999 Sig. .032* .953 .879 .994 JobSchool_factor .784 .907 .922 .919 Sig. .002** .347 .459 .445 Social_factor 1.302 1.349 1.324 1.336 Sig. .001** .002** .006** .005** HomeAtts_factor .846 .959 .981 .985 Sig. .031 .660 .848 .884      BuiltForm_2km_Factor  .702 .723 .753 Sig.  .001** .005** .035*      HHInc_under$40k   ref Ref HHInc_$40_$100k   .698 .695 Sig.   .136 .132 HHInc_over$100k   .986 .973 Sig.   .967 .935 N  1,108  798 701 699 R2 (Nagelkerke) .161 .164 .167 .168 * Significant at p. <0.05  ** Significant at p.<0.01  ***Significant at p.<0.001 1 Reference group is respondents who are “Mismatched into lower Smart Growth”..  This is not surprising, suggesting that people who seek to be in areas that have Smart Growth features  make an effort to live in Smart Growth areas.  People who prize the ability to access regional jobs and services (RegAccess_factor) also are more likely to be Smart Growth matched, as are people who value proximity to cultural offerings, friends and families when they seek new neighbourhoods to live in (Social_factor).  However, people who put greater emphasis on finding large, spacious homes with yards (HomeAtts_factor) and people who seek homes near 191  their work or school (JobSchool_factor) are less likely to be subjectively Smart Growth matched.  Of these latter two findings, the first is relatively intuitive: people who care more about the spaciousness and privacy of their homes are less likely to want to live in Smart Growth areas where these things are almost always compromised to provide better access to stores, services and amenities.   The second finding – that people who value jobs and school access are less likely to be Smart Growth matched – might either suggest that many jobs are increasingly located in suburban areas (cf. Levine – jobs housing balance), that institutions for higher learning are located in suburban areas34, or that respondents with preferences for suburban environments also value easy access to highways to reach their jobs, all of which, in the Vancouver region, are located in suburban areas.   The addition of a principal component measuring a host of objective built environment attributes around the same 2 km network buffer as housing mix (BuiltForm_2km_Factor) renders all but two neighbourhood importance factors insignificant (Model 2).  However, the model fit is not improved dramatically. NeighbAm_factor and Social_factor continue to be significant predictors that are associated with improved odds of subjective Smart Growth matching, while BuiltForm_2km_Factor is significantly associated with lower odds of Smart Growth matching. Not surprisingly, if a respondent lives in an area with larger homes, a greater proportion of homes with yards, and a higher average number of bedrooms, they are less likely to live in a Smart Growth neighbourhood.  The number of observations included in the analysis shrinks dramatically from 1,108 to 798 on account of the large number of outliers that are omitted when the Built Form component is included.  The significance and direction of the variables are                                                  34 In the Vancouver region, the two largest university campuses are quite physically separated from surrounding built up areas and are quite distant from established residential areas that may be considered to be walkable; many community colleges are located in suburban municipalities. 192  retained when the household income of the respondent is controlled for in model 3.  However, household income categories do not significantly predict whether a respondent with Smart Growth preferences will raise their odds of subjectively matching.  Finally, adding housing mix at 2 km (Model 4) to the model does not seem to improve the odds of Smart Growth matching, nor is the predictive power of the model increased.  Housing mix does not appear to make a difference in the ability for people to subjectively match into Smart Growth. Table 54: Multinomial logistic regression results predicting odds of objectively matching neighbourhood to Smart Growth preferences  Model Number1 Predictor (1) (2) (3) (4)      Z_HsgMix_2km    1.877 Sig.    .000***      NeighbAm_factor 1.093 0.807 0.787 0.819 Sig. .265 .068 .061 .131 RegAccess_factor 1.134 .799 .804 .721 Sig. .116 .048* .083 .015* JobSchool_factor .865 .994 1.027 1.019 Sig. .067 .959 .832 .882 Social_factor 1.172 1.138 1.148 1.169 Sig. .045* .220 .233 .189 HomeAtts_factor .885 1.060 1.008 1.029 Sig. .127 .592 .946 .812      BuiltForm_2km_Factor  .287 .239 .220 Sig.  .000*** .000*** .000***      HHInc_under$40k   ref ref HHInc_$40_$100k   1.411 1.599 Sig.   .210 .094 HHInc_over$100k   .829 .868 Sig.   .616 .709 N  1,163  825 723 721 R2 (Nagelkerke) .136 .438 .453 .486 * Significant at p. <0.05  ** Significant at p.<0.01  ***Significant at p.<0.001 1 Reference group is respondents who are “Mismatched into lower Smart Growth”.  193  The parallel model predicting objective Smart Growth matching shows very different results (Table 54).  In the first model, only proximity to family, friends and cultural amenities (Social_Factor) significantly raises the odds of Smart Growth matching.  The odds of objectively matching into a Smart Growth neighbourhood is significantly diminished for every standard deviation increase in the BuiltForm_2km_Factor (model 2), and the addition of this component renders every neighbourhood importance factor insignificant with the exception of the RegAccess_factor, which is associated with lower odds of Smart Growth matching. Additionally, model fit is substantially improved, with over 40% of the variance in the data explained by the variables in the model.  Income remains an insignificant predictor of match (model 3), while housing mix remains significantly and positively associated with raising the odds of objective Smart Growth matching, even after all the other factors have been controlled for (Model 4).  Model fit, however, improves only slightly.  An additional 3% of the variance in the data is explained by the inclusion of housing mix into the model.  This is not surprising, given that housing mix was never expected to be a major motivator (or barrier) for people to sort into the neighbourhood of their choice. This last finding may be the most important finding encountered so far: living in an area with a higher level of housing mix significantly raises the odds that respondents with Smart Growth preferences will objectively match, controlling for a range of personal and environmental characteristics.      194  6.2.4.1 The Insignificance of Household Income Why is household income not a significant predictor that people will match their neighbourhood with their Smart Growth preferences?  To provide an answer for this, a multinomial logit model was run predicting the odds of matching into the 3 match categories with just income categories as predictor variables.   Table 55: Multinomial logistic regression results predicting odds of subjectively matching neighbourhood to Smart Growth preferences, effect of household income only   Subjective Match Predictor N Prefers SG & Matched Prefers suburbia & matched Mis-matched into higher SG      HHInc_under$40k 213 ref ref ref HHInc_$40_$100k 617 .668 2.157 .920 Sig.  .046* .062 .778 HHInc_over$100k 207 .739 2.120 1.125 Sig.  .224 .109 .619 N  403 82 242 R2 (Nagelkerke)  .017 017 017 * Significant at p. <0.05  ** Significant at p.<0.01  ***Significant at p.<0.001 1. Reference group is respondents who are “Mismatched into lower Smart Growth”. (n = 260)  In the subjective match model (Table 55), respondents with household incomes in the “middle” range of between $40,000 to $100,000 a year have significantly lower odds (OR = 0.668) of matching into Smart Growth neighbourhoods rather than mismatching into an area with lower Smart Growth compared to their low income (below $40,000) counterparts.      195  Table 56: Multinomial logistic regression results predicting odds of objectively matching neighbourhood to Smart Growth preferences, effect of household income only   Objective Match Predictor N Prefers SG & Matched Prefers suburbia & matched Mis-matched into higher SG      HHInc_under$40k 213 ref ref ref HHInc_$40_$100k 617 1.025 2.388 1.008 Sig.  .905 .010* .972 HHInc_over$100k 207 .635 2.073 .829 Sig.  .076 .056 .472 N  362 125 298 R2 (Nagelkerke)  .016 .016 .016 * Significant at p. <0.05  ** Significant at p.<0.01  ***Significant at p.<0.001 1. Reference group is respondents who are “Mismatched into lower Smart Growth”. (n = 252).  In the objective match model (Table 56), middle income respondents were twice as likely as their low income counterparts to match into suburbia as opposed to mismatch into an area with fewer Smart Growth attributes.  When these findings are compared against a map of the geographic distribution of respondents by income category (Figure 33), a trend emerges showing that middle income and high income respondents are scattered somewhat uniformly across the region and its walkability quartiles, while lower income respondents appear to be clustered in high walkability (i.e. high Smart Growth) areas, both within and outside the City of Vancouver.  Not surprisingly, these areas are characterized by high net residential densities and, ostensibly, contain a greater proportion of cheaper multifamily dwelling types.  These respondents are also likely to be renters (Figure 34), and renters are disproportionately represented among lower income respondents in the sample (Table 57).     196  Table 57: Tenure and household income, cross-tabulation of frequencies  Tenure Under $40,000 $40,000-$100,000 Over $100,000 Total       Own N 124 413 171 708  % 17.5% 58.3% 24.2% 100.0%       Rent N 221 217 40 478  % 46.2% 45.4% 8.4% 100.0%       Total N 345 630 211 1186  % 29.1% 53.1% 17.8% 100.0%  Figure 33: Income category of respondents, Metro Vancouver     197  Figure 34: Tenure of respondents, Metro Vancouver  While higher income groups may be able to sort themselves into a variety of built environments according to their preferences, lower income groups may be restricted to living in high Smart Growth areas. Low income respondents overwhelmingly report preferences for Smart Growth environments, but this begs the question as to how many of those low income respondents would prefer suburbia if they did not find themselves restricted into those parts of the region where rental housing is available, and where more jobs and services can be accessed without a car.  It is plausible that more low income respondents who “prefer” Smart Growth would willingly trade smaller apartments and higher density neighbourhoods for more spacious homes and suburban neighbourhood qualities if they could afford it.  However, meeting their 198  basic needs for affordable shelter and affordable transportation options may compel them to indicate that they prefer a Smart Growth environment which can provide these amenities when responding to the survey. In other words, for some respondents, a “preference” for Smart Growth environments may not be a “true” preference at all; it may represent a forced choice, and this phenomenon – if it exists – cannot be captured using survey methods. This raises issues with the robustness of stated preference surveys.  Supposedly, one of the benefits of stated preference, compared to revealed preference research, is the ability to estimate the demand for latent, or under provided alternatives.  However, it may be difficult, if not impossible, for respondents to divorce their “wants” – preferences for products with a disregard for personal circumstances that may restrict them from accessing these products – from their “needs”.    6.2.4.2 Stratified Multivariate Models by Population Subgroup Finally, the multivariate model was run on separate subsamples of the population to see if housing mix impacted the matching behaviour of these groups differently.  Separate models were run on renters versus owners (Table 58) and respondents over the age of 60 versus under the age of 60 (Table 59) A decision to stratify the sample by families with children versus childless households, and by East Asian respondents versus non-East Asian respondents was abandoned because the sample size of the two groups was below what was needed to achieve statistical power.  Both subjective and objective definitions of match are run.     199  6.2.4.2.1 Renters versus Owners Table 58 displays the difference in Smart Growth matching outcomes between renters and owners.   Table 58: Multinomial logistic regression results predicting odds of matching neighbourhood to Smart Growth preferences, stratified by tenure   Subjective Match Objective Match Predictor Renters Owners Renters Owners      Z_HsgMix_2km .659 .937 1.120 1.933 Sig. .135 .655 .780 .004**      NeighbAm_factor 1.417 1.294 .780 .846 Sig. .076 .063 .279 .311 RegAccess_factor .848 .977 .641 .745 Sig. .412 .876 .048* .115 JobSchool_factor 1.515 0.750 .996 1.018 Sig. .051 .044* .984 .916 Social_factor .873 1.539 1.261 1.119 Sig. .468 .001** .234 .481 HomeAtts_factor 1.713 0.794 1.004 1.063 Sig. .005** .086 .984 .705      BuiltForm_2km_Factor .849 .803 .420 .170 Sig. .592 .179 .010* .000***      HHInc_under$40k ref ref ref ref HHInc_$40_$100k .510 1.150 2.105 1.160 Sig. .072 .684 .053 .747 HHInc_over$100k 1.194 1.358 2.569 0.532 Sig. .808 .466 .271 .241 N 478 422 478 429 R2 (Nagelkerke) .207 .230 .483 .503 * Significant at p. <0.05 ** Significant at p.<0.01  ***Significant at p.<0.001 1 Reference group is respondents who are “Mismatched into lower Smart Growth”.  Housing mix is not a significant predictor of matching ability for any group other than for owners, when Smart Growth matching is defined objectively.  This infers that only owners are able to take advantage of housing mix to sort themselves into Smart Growth areas.  As the map 200  of tenure distribution in the sample shows (Figure 34), most renters are confined to high walkability/“Smart Growth” areas where housing mix is relatively low, likely owing to the predominance of multifamily rental buildings.  A close up of the central portion of the region, centered around the City of Vancouver shows that renters take advantage of the belt of high housing mix that extends around the inner city neighbourhoods ringing the downtown peninsula (Figure 35).  However, few renters who live in these high housing mix areas seem to be matched with their preferences, at least objectively.  Many renters were objectively mismatched into areas that contained higher Smart Growth attributes than they expressed preferences for.  Interestingly, many of these same respondents were subjectively matched (Figure 36).   Figure 35: Renters by objective neighbourhood match categories, central Metro Vancouver  201  Figure 36: Renters by subjective neighbourhood match categories, central Metro Vancouver  This lends some support to the idea that renters, many of them with lower incomes, might be compelled to “prefer” their current neighbourhood environment in light of their difficulties in accessing other housing options (or difficulty in accessing destinations without a car) in neighbourhoods that align more closely with what their true preference might be.  If this phenomenon can be demonstrated, a meaningful policy prescription might be to improve the availability of rental stock in areas outside of the areas with the highest Smart Growth attributes and improve transportation alternatives in these areas.   202  6.2.4.2.2 Stratification Results by Age Group: Under 60 and Over 60 Table 59 highlights the difference in the odds of Smart Growth matching between people over the age of 60 and those under.   Table 59: Multinomial logistic regression results predicting odds of matching neighbourhood to Smart Growth preferences, stratified by age   Subjective Match Objective Match Predictor Under 60 60 and over Under 60 60 and over      Z_HsgMix_2km 1.089 .616 1.908 1.784 Sig. .584 .027* .010* .052      NeighbAm_factor 1.333 1.314 .696 1.074 Sig. .027* .213 .031* .772 RegAccess_factor .886 1.251 .537 1.275 Sig. .380 .282 .000*** .338 JobSchool_factor .983 .683 .947 1.033 Sig. .905 .058 .738 .890 Social_factor 1.396 1.191 1.261 .988 Sig. .010* .329 .124 .955 HomeAtts_factor 1.053 0.951 0.993 0.974 Sig. .684 .798 .964 .911      BuiltForm_2km_Factor .663 .962 .191 .297 Sig. .020* .868 .000*** .001**      HHInc_under$40k ref ref ref ref HHInc_$40_$100k .593 .970 1.767 1.287 Sig. .083 .945 .102 .625 HHInc_over$100k .893 0.969 .854 .812 Sig. .777 .960 .732 .784 N 495 204 510 211 R2 (Nagelkerke) .170 .285 .501 .519 * Significant at p. <0.05  ** Significant at p.<0.01  ***Significant at p.<0.001 1 Reference group is respondents who are “Mismatched into lower Smart Growth”..  Increasing housing mix lowers the odds that people over 60 will subjectively match into a Smart Growth neighbourhood. Housing mix significantly raises the probability that people under 60 will objectively match into Smart Growth areas, but it is not a significant predictor of the 203  same behaviour among people over 60.  Unlike renters, seniors are dispersed across Metro Vancouver (Figure 37).  When the settlement and objective neighbourhood match patterns of seniors are zoomed in on the central part of the region, it becomes apparent that seniors are less likely than renters to take advantage of high housing mix in high walkability neighbourhoods (Figure 38).  Fewer seniors live in the neighbourhoods to the east and south-east of the downtown peninsula characterized by higher degrees of housing mix.  Even fewer are objectively matched into their preferred environment; most seniors who lived in the inner city were objectively mismatched into a neighbourhood that contained a higher amount of Smart Growth than their preferences, especially when compared with renters.  Seniors were revealed to have lower preference scores for Smart Growth than renters (Table 29, Chapter 5), but still had median preference scores that were above average.  The fact that many seniors lived in the downtown peninsula, with its concentrated availability of multifamily apartment towers (and rental stock), or were dispersed across the metropolitan region, suggests that the best strategies to enable “aging in place” might be to encourage multifamily apartment construction across the metropolitan region.  This is a separate consideration from improving “housing mix”, which entails constructing a variety of different housing options within a small area.  Further research is needed.   204  Figure 37: Respondents over 60 years of age by objective neighbourhood match categories, Metro Vancouver    205  Figure 38: Respondents over 60 years of age by objective neighbourhood match categories, central Metro Vancouver   6.2.5 Summary of Findings Increased neighbourhood housing mix was found to significantly increase the probability of objective neighbourhood matching, and also to increase the probability that a household with preferences for Smart Growth would objectively match, rather than live in a neighbourhood with fewer Smart Growth attributes than they preferred.  These findings apply when other neighbourhood importance factors, the structural attributes of nearby properties, and household income are controlled for.  However, these findings are not replicated when match  is defined 206  subjectively.  Furthermore, housing mix does not significantly predict the odds that seniors (respondents over 60) and renters will objectively match into Smart Growth neighbourhoods.    6.3 Statistical Models answering Research Question 2 Regardless of whether housing mix enables some people to match more than others, are the people who match more satisfied with their neighbourhood?  Although “neighbourhood matching” has never been tested as a predictor of satisfaction in the residential satisfaction studies reviewed in this dissertation (Chapter 3, section 3.3.2), conventional wisdom would lead us to expect that the ability to match would be positively and significantly associated with neighbourhood satisfaction.  The models presented in this section test the relationship between neighbourhood match and neighbourhood satisfaction, controlling for a range of socioeconomic and demographic predictors as well as variables representing the built environment of the area surrounding a respondent’s home and their satisfaction with other aspects of their neighbourhood.  These relationships are tested using ordered logit models that estimate the odds that a respondent is one level more dissatisfied with their neighbourhood; as such, when interpreting odds ratios it is important to recognize that a significant odds ratio above 1 is associated with greater levels of dissatisfaction, so odds ratios below 1 are normatively “better”.  6.3.1 Simple models predicting Neighbourhood Satisfaction The first set of models (Table 60) investigate whether being “matched” – either subjectively (model 1) or objectively (model 2) – is a significant predictor of neighbourhood satisfaction on its own, without any control variables.  207  Table 60: Ordered logit results predicting degree of neighbourhood dissatisfaction, subjective vs. objective neighbourhood match  Model (1) (2) Variable Odds Sig. Odds Sig.      Subjective_match 0.705 .001**        Objective_match   1.060 .573      Pseudo-R2 (Nagelkerke)  .010  .000 Model Fit  .001  .573 * Significant at p. <0.05  ** Significant at p.<0.01  ***Significant at p.<0.001  Table 60 reveals that being subjectively matched significantly lowers the odds that a respondent will be more dissatisfied with their neighbourhood.  However, being objectively matched is not significantly associated with any change in neighbourhood satisfaction, and the inclusion of objective neighbourhood match explains none of the variance in the model.  It seems intuitive that subjective neighbourhood match – which is the respondent’s own assessment of whether their neighbourhood aligns with their Smart Growth preferences – should be a significant predictor of neighbourhood satisfaction while objective neighbourhood match – which is a comparison of a respondent’s Smart Growth preferences to a walkability index score which they are oblivious about – is not.  Nonetheless, it is more important for Smart Growth planners and researchers to understand if being matched or mismatched into different types of neighbourhoods – Smart Growth versus suburbia – have different impacts on neighbourhood satisfaction.  In Table 61, and all subsequent models, match is divided into the four match categories introduced in the multinomial logit models used to predict neighbourhood matching.  As with those models, households who are mismatched into lower Smart Growth areas than their preference are used as a reference group.   208  Table 61: Ordered logit results predicting degree of neighbourhood dissatisfaction, neighbourhood match categories Model (1) (2) Variable Odds Sig. Odds Sig. Subjective Match     Prefers SG and Matched 0.930 .587   Prefers suburbia and Matched 1.195 .406   Mismatched into higher SG 1.560 .003**   Mismatched into lower SG ref    Objective Match     Prefers SG and Matched   1.452 .007** Prefers suburbia and Matched   1.254 .234 Mismatched into higher SG   1.655 .000*** Mismatched into lower SG   ref  Pseudo-R2 (Nagelkerke)  .014  .012 Model Fit  .001  .004 * Significant at p. <0.05  ** Significant at p.<0.01  ***Significant at p.<0.001  Table 61 shows that not all match categories have equal bearing on neighbourhood satisfaction.  In the subjective model (model 1), being matched into a Smart Growth neighbourhood or being matched into suburbia does not significantly raise or lower the odds of dissatisfaction compared to being mismatched into a lower Smart Growth environment than your preferences.  However, being mismatched into an environment that has more Smart Growth attributes than you prefer significantly lowers the odds of being satisfied (i.e. it is associated with increased odds of higher neighbourhood dissatisfaction).  In contrast, people who prefer Smart Growth and are objectively matched into areas with comparable walkability index scores (model 2) report higher odds of dissatisfaction compared to people who are objectively mismatched into areas with lower Smart Growth than they might prefer.  Households who are mismatched into 209  areas that have more Smart Growth attributes than they might prefer have even higher odds of being dissatisfied.  However, the level of neighbourhood satisfaction that respondents report is more likely due to personal characteristics as well as aspects of the environment that respondents perceive positively or negatively, rather than their match status.  The next series of models incorporate these considerations.    210  6.3.2 Multivariate Models of Neighbourhood Satisfaction Table 62 investigates whether the neighbourhood satisfaction trends reported in the simple models continue if important SES and demographic characteristics are controlled for35.    Table 62: Ordered logit results predicting degree of neighbourhood dissatisfaction, neighbourhood match categories, socioeconomic and demographic covariates Model (1) (2) Variable Odds Sig. Odds Sig. Subjective Match     Prefers SG and Matched 0.850 .268   Prefers suburbia and Matched 1.588 .046*   Mismatched into higher SG 1.634 .003**   Mismatched into lower SG ref    Objective Match     Prefers SG and Matched   1.338 .053 Prefers suburbia and Matched   1.503 .044* Mismatched into higher SG   1.522 .007** Mismatched into lower SG   ref       Other Predictors     Age_under60 1.396 .023* 1.322 .049* Children_inHH 1.392 .030* 1.566 .003** Ethnicity_NotMatched 1.033 .815 1.080 .574 HHInc_under$40k ref  ref  HHInc_$40_$100k 0.723 .035* 0.715 .024* HHInc_over$100k 0.543 .002** 0.494 .000*** HHInc_NotMatched 1.158 .272 1.174 .213 Length_of_Res 1.000 .970 0.998 .718 Tenure_Rent 1.356 .021* 1.275 .062 Commute_BelowMedian 0.781 .067 0.817 .124 Pseudo-R2 (Nagelkerke)  .072  .061 Model Fit  .000  .000 * Significant at p. <0.05  ** Significant at p.<0.01  ***Significant at p.<0.001                                                  35 The inclusion of these predictor variables was checked for multicollinearity since it was suspected that many of these socioeconomic and demographic variables might be correlated.  Surprisingly, all variables had a VIF below 1.5. 211  In the subjective model (Model 1), being under the age of 60, having children under 18 in the household, and being a renter were significantly associated with higher levels of dissatisfaction.  As expected, households with higher incomes reported significantly lower levels of neighbourhood dissatisfaction.  Interestingly, living in a neighbourhood where the respondent was not part of the ethnic majority (Ethnicity_NotMatched), having a different household income from your census tract median (HHInc_NotMatched), and having a longer than average commute time than the sample median (Commute_BelowMedian) were not significant predictors of neighbourhood satisfaction, contrary to evidence on residential segregation from the US by Kiel and Zabel (2008) and Ioannides and Zabel (2008) and running somewhat contrary to theories that households back municipal land use regulation that ensures that they segregate by race or income (Ihlanfeldt 2004; Fischel 2004)36.  When these SES and demographic variables are controlled for,  respondents who prefer suburbia and are matched - in both the subjective and objective models - report significantly lower satisfaction levels, as do respondents who are mismatched into higher Smart Growth areas compared to the reference group.  Respondents who prefer Smart Growth, and are matched, are no more likely to be satisfied than people who prefer Smart Growth but are mismatched into more suburban areas than they report preferences for.   If built form covariates are controlled for, instead of SES and demographic factors, satisfaction results differ considerably (Table 63).  In model 1 and model 3, the effect of subjective and objective match type is tested, respectively, on neighbourhood satisfaction, controlling for whether a household lives on a cul de sac, the type of housing they report living in (with single family home dwellers serving as the reference category) and the principal                                                  36 It is possible that multifamily dwellings enable vertical, rather than spatial segregation and that people’s acknowledgment of the income and ethnic makeup of their surroundings is influenced by this.  This phenomenon is not tested and discussion of its plausibility or veracity is beyond the scope of this dissertation. 212  component that proxies for the built environment characteristics of the home within a 2km network buffer of their address (BuiltForm_2km_Factor).   The results for both subjective and objective match models show that living in an ostensibly smaller and denser housing typology is associated with significantly lower levels of neighbourhood satisfaction than living in a single family home.  Interestingly, people who live in communities with a high “Built Form” component score – that is, areas where homes have greater interior space, a higher proportion of yards, higher values and more bedrooms – report higher odds of neighbourhood dissatisfaction. While this is unexpected, it is likely that all the factors that contribute to satisfaction may have been captured in the other variables.  Brown et al (1985) found that households who lived on cul de sac held stronger ties to their neighbours than residents of through-streets, and this is echoed by the finding that respondents who lived on cul de sacs were associated with higher odds of satisfaction, albeit only in the objective match models.  Controlling for all built form factors (models 1 and 2), only households who were subjectively mismatched into higher Smart Growth environments than their preference reported significantly different – in this case lower – satisfaction levels than the reference group.  In the objective models (models 3 and 4), however, all groups reported lower satisfaction levels than people who found themselves mismatched into areas that had fewer Smart Growth attributes than their preferences.  Respondents who were objectively mismatched into higher Smart Growth areas, in particular, were three times more likely to report more dissatisfaction with their neighbourhood compared to the reference group.  While the previous research question suggested that neighbourhood housing mix may be associated with objective neighbourhood matching, greater housing mix was also associated with higher odds of neighbourhood 213  dissatisfaction, in both subjective and objective models where it was added (models 3 and 4) after other built form variables were controlled for.  This finding refutes the original hypothesis described in the conceptual framework (Figure 5, Section 4.1) that housing mix would be an insignificant predictor of neighbourhood satisfaction on its own.  The results from table 63 suggest that households who live in non-Smart Growth surroundings, characterized by lower density, more spacious housing forms and less diversity in available housing options are, in fact, the most satisfied.  This trend seems to apply, regardless of whether respondents prefer to be in Smart Growth, or not. 214  Table 63: Ordered logit results predicting degree of neighbourhood dissatisfaction, neighbourhood match categories, built form covariates Model (1) (2) (3) (4) Variable Odds Sig. Odds Sig. Odds Sig. Odds Sig. Subjective Match         Prefers SG and Matched 0.824 .245 0.845 .312     Prefers suburbia and Matched 1.228 .515 1.282 .437     Mismatched into higher SG 1.449 .046* 1.499 .030*     Mismatched into lower SG Ref  Ref      Objective Match         Prefers SG and Matched     1.802 .002** 1.258 .007** Prefers suburbia and Matched     2.135 .008** 1.688 .006** Mismatched into higher SG     3.280 .000*** 2.330 .004** Mismatched into lower SG     Ref  Ref           Other Predictors         Z_HsgMix_2km   1.175 .047*   1.258 .007** Cul_de_Sac 0.735 .080 0.757 .115 0.649 .019* 0.671 .020* HsgType_MFHhigh 1.993 .004** 2.002 .004** 1.944 .005** 1.977 .004** HsgType_MFHlow 2.183 .000*** 1.997 .000*** 2.430 .000*** 2.234 .000*** HsgType_GroundAttached 1.469 .070 1.499 .030* 1.615 .019* 1.507 .048* HsgType_SingleFam ref  ref  ref  ref  BuiltForm_2km_Factor 1.462 .000*** 1.325 .003** 1.862 .000*** 1.654 .000*** Pseudo-R2 (Nagelkerke)  .059  .064  .087  .098 Model Fit  .000  .000  .000  .000 * Significant at p. <0.05  ** Significant at p.<0.01  ***Significant at p.<0.001  215  Finally, the importance of neighbourhood matching is compared against the importance of deriving satisfaction from other aspects of one’s neighbourhood (Table 64).  In these models, the effect that a respondent’s match status has on neighbourhood satisfaction is tested, while controlling for ten questions inquiring about their satisfaction with other aspects of their community.  Among the satisfaction questions, the level of satisfaction that people have with the crime in their community (Crime_) has the strongest bearing on neighbourhood satisfaction.  Respondents who reported being very dissatisfied with the level of crime in their neighbourhood were twenty times more likely to be dissatisfied with their neighbourhood than people who were very satisfied with crime levels in their area.  Other strong, significant predictors of neighbourhood dissatisfaction include being dissatisfied with access to recreational facilities (AxsRecr_), access to food options (AxsFood_), knowing your neighbours (KnowNeighbours_), and noise (Noise_).   Once these satisfaction questions were included, no category of objective match was associated with improved neighbourhood satisfaction levels compared to people who were mismatched into lower Smart Growth environments.  Households who were subjectively mismatched into higher Smart Growth, or even matched into the suburban environment of their preference, were more likely to be dissatisfied with their neighbourhood than the reference group.  However, households who were subjectively matched into Smart Growth were no more likely to be dissatisfied than the reference group.  These findings suggest that the factors that led respondents in Smart Growth areas to report lower satisfaction levels are probably captured by factors queried in the satisfaction questions (e.g. crime, noise, yard size).  The percent of respondents who reported dissatisfaction with crime in their community, for example, was highest among people who were objectively matched into Smart Growth (17.9%) and people 216  who were objectively mismatched into an area with more Smart Growth attributes than they preferred (21.5%).  In contrast, only 13.5% of respondents who were objectively mismatched into areas with fewer Smart Growth attributes and 13.0% of respondents who were matched into suburbia expressed dissatisfaction with the crime levels in their neighbourhood.  Whether these findings reflect actual crime rates in the area is unknown, but some of the sacrifices that are made to live in Smart Growth areas, whether real or perceived, may disproportionately contribute to the dissatisfaction people have with their communities compared to the sacrifices that are made to live in suburbia.     217  Table 64: Ordered logit results predicting degree of neighbourhood dissatisfaction, neighbourhood match categories,  satisfaction questions   Model  1 2 Variable N Odds Sig. Odds Sig. Subjective  Match Measures      Prefers SG and Matched  463 1.056 .714   Prefers suburbia and Matched 92 1.683 .024*   Mismatched into higher SG 269 1.391 .046*   Mismatched into lower SG 285 ref    Objective  Match Measures      Prefers SG and Matched 408   1.299 .088 Prefers suburbia and Matched 131   1.268 .243 Mismatched into higher SG 344   1.258 .148 Mismatched into lower SG 280   Ref        Satisfaction Questions      Commute_StrongDis 29 0.764 .270 0.733 .397 Commute_SomeDis 106 1.164 -.152 1.164 .452 Commute_SomeSat 486 1.247 -.220 1.181 .207 AxsRecr_StrongDis 9 8.478 .001** 5.141 .010* AxsRecr_SomeDis 42 2.931 .000*** 2.629 .001** AxsRecr_SomeSat 508 2.017 .000*** 2.054 .000*** AxsFood_StrongDis 11 3.584 .029* 3.660 .011* AxsFood_SomeDis 98 2.504 .000*** 2.389 .000*** AxsFood_SomeSat 497 2.058 .000*** 2.089 .000*** ChildcareProx_StrongDis 15 1.297 .656 1.359 .606 ChildcareProx_SomeDis 60 0.589 .068 0.631 .094 ChildcareProx_SomeSat 586 0.960 .787 1.012 .934 SchoolQual_StrongDis 27 2.002 .132 1.896 .143 SchoolQual_SomeDis 75 1.389 .228 1.329 .280 SchoolQual_SomeSat 653 1.197 .274 1.060 .713 HomeSize_StrongDis 39 1.629 .156 2.124 .017* HomeSize_SomeDis 190 1.249 .245 1.203 .315 HomeSize_SomeSat 502 1.181 .251 1.114 .444 YardSize_StrongDis 108 1.215 .422 1.371 .171 YardSize_SomeDis 193 1.738 .004** 1.709 .005** YardSize_SomeSat 462 1.452 .014* 1.359 .038* Crime_StrongDis 37 20.275 .000*** 17.257 .000*** Crime_SomeDis 152 5.318 .000*** 5.678 .000*** Crime_SomeSat 569 2.001 .000*** 2.134 .000*** KnowNeighbours_StrongDis 44 3.673 .000*** 4.701 .000*** KnowNeighbours_SomeDis 203 2.147 .000*** 2.360 .000*** KnowNeighbours_SomeSat 557 1.261 .140 1.345 .054 Noise_StrongDis 63 5.696 .000*** 5.258 .000*** Noise_SomeDis 239 1.766 .001** 1.718 .001** Noise_SomeSat 473 1.081 .608 1.048 .751 Pseudo-R2 (Nagelkerke) Model Fit Test of Parallel Lines (sig)   0.433 0.000 0.399  0.435 0.000 0.560  * Significant at p. <0.05  ** Significant at p.<0.01  ***Significant at p.<0.001  218  6.3.3 Summary of Findings Respondents who were matched with their Smart Growth preferences were not significantly more likely to be satisfied with their neighbourhood environment compared to respondents who were mismatched into a neighbourhood with fewer Smart Growth attributes than their preferences.  These findings applied to both subjective and objective definitions of match.  Respondents who found themselves mismatched into areas with more Smart Growth attributes than their preferences were significantly more dissatisfied than households who were mismatched into areas with fewer Smart Growth attributes than their preferences when controlling for SES, demographic and built form covariates.  This, along with the finding that higher density housing typologies are associated with increased odds of dissatisfaction, suggests that the trade-offs associated with Smart Growth environments – such as increased noise, reduced privacy, and a greater perception of crime – may overwhelm positive contributions to neighbourhood satisfaction such as greater access to recreation and food options.  Accordingly, an increase in neighbourhood housing mix was associated with greater odds of neighbourhood dissatisfaction.   6.4 Statistical Models answering Research Question 3 Finally, models linking the ability to match and measurable outcomes of health are investigated in the third research question.  These include models predicting the effect that neighbourhood matching may have on one’s subjective perception of health (self-reported health status) and a measure of health that is subject to less personal interpretation (BMI). 219  6.4.1 Models predicting self-reported health status Tables 65 and 66 display outcomes from ordered logit models predicting self-reported health status for subjective and objective definitions of match, respectively.  Since the model predicts the odds of a respondent reporting one better level on the self-reported health status scale (e.g. from a level of 3 (“good”) to 4 (“very good”)), odds ratios above 1 can be interpreted as being normatively better (raising the odds of reporting a higher health status), while odds ratios below 1 can be interpreted as being normatively worse.   Model 1 only predicts the effect that the type of residential matching may have on improving self-reported health status.  When defined subjectively or objectively, the type of neighbourhood match or mismatch a respondent encounters is not significantly associated with any change in their self-reported health status on its own.  In model 2, categorical variables are added coding for respondents who are either very satisfied (a satisfaction rating of 10/10) or very unsatisfied (a satisfaction rating of 5/10 or below) with their neighbourhood.  As expected, respondents who are very satisfied with their neighbourhood are significantly more likely to report a better health status.  220  Table 65: Ordered logistic regression predicting self-reported health status, subjective neighbourhood match categories  Model 1 2 3 4 Variable Odds Sig. Odds Sig. Odds Sig. Odds Sig. Subjective Match          Prefers Smart Growth and Matched  1.128 .378 1.132 .366 1.339 .047* 1.220 .183 Prefers Suburbia and Matched 1.066 .770 1.091 .691 0.905 .667 0.981 .935 Mismatched into higher SG 0.865 .348 .930 .640 0.990 .951 0.940 .708 Mismatched into lower SG ref  ref  ref  ref           Other Predictors         WalkIndex       1.022 .000*** Age     1.020 .000*** 1.050 .001** Gender_Male     0.910 .430 0.906 .414 HHInc_under$40k     ref  ref  HHInc_$40_$100k     1.641 .001** 1.682 .001** HHInc_over$100k     2.301 .000*** 2.326 .000*** Ethn_European     0.715 .094 0.695 .071 Ethn_Canadian     0.669 .076 0.656 .063 Ethn_EastAsian     0.737 .022* 0.728 .017* Ethnicity_allothers     ref  ref           NeighbSat_very_sat   2.056 .000***     NeighbSat_very_unsat   .538 .002**     NeighbSat_allothers   ref      N  1,102   1,102   981   981  Pseudo-R2 (Nagelkerke)  .004  .055  .055  .065 Test of Parallel Lines  .218  .375  .437  .348 a. * Significant at p. <0.05  ** Significant at p.<0.01  ***Significant at p.<0.001   221  Table 66: Ordered logistic regression predicting self-reported health status, objective neighbourhood match categories  Model 1 2 3 4 Variable Odds Sig. Odds Sig. Odds Sig. Odds Sig. Objective Match          Prefers Smart Growth and Matched  0.841 .218 0.905 .480 0.919 .574 0.755 .076 Prefers Suburbia and Matched 0.718 .084 0.738 .115 0.622 .018* 0.661 .039* Mismatched into higher SG 0.933 .635 1.029 .848 1.013 .934 0.630 .019* Mismatched into lower SG ref  ref  ref  ref           Other Predictors         WalkIndex       1.085 .000*** Age     1.020 .000*** 1.021 .000*** Gender_Male     0.896 .347 0.900 .370 HHInc_under$40k ref  ref  ref  ref  HHInc_$40_$100k     1.571 .002** 1.656 .001** HHInc_over$100k     2.124 .000*** 2.190 .000*** Ethn_European     0.807 .285 0.814 .303 Ethn_Canadian     0.678 .083 0.704 .119 Ethn_EastAsian     0.741 .020* 0.752 .027* Ethnicity_allothers ref  ref  ref  ref           NeighbSat_very_sat   1.921 .000***     NeighbSat_very_unsat   0.538 .001**     NeighbSat_allothers ref  ref  ref  ref  N  1,157   1,157   1,031   1,031  Pseudo-R2 (Nagelkerke)  .003  .048  .047  .062 Test of Parallel Lines  .096  .225  .284  .268 b. * Significant at p. <0.05  ** Significant at p.<0.01  ***Significant at p.<0.001 222  The relationship between health status and neighbourhood satisfaction may be bidirectional: if people feel they are much healthier than people of comparable age, they may project their overall satisfaction onto their satisfaction with their neighbourhood, even though the two should, in theory, be unrelated.  This aligns with Campbell’s original theory of life satisfaction where satisfaction consisted of interlinked domains, and people who were fulfilled in one aspect of their life often were fulfilled in others (Campbell, Converse, and Rodgers 1976).  It also highlights the difficulty of drawing meaningful conclusions from neighbourhood satisfaction studies.  Was the respondent truly satisfied with their neighbourhood, or were they satisfied with some other, latent, aspect of their life which they then projected onto their satisfaction with their neighbourhood?   In the third model, socioeconomic and demographic predictors are added. When these variables are added, subjectively matched Smart Growth dwellers have higher odds of reporting positive health status scores than people who are mismatched into less Smart Growth (Table 65, model 3).  Meanwhile, objectively matched suburbanites are significantly less likely to report positive health scores (Table 66, model 3).  This finding augurs well for Smart Growth advocacy.  Not surprisingly, the higher a respondent’s income, the higher the odds were of reporting positive health scores.  Each additional year of age was associated with a 2% increase in the odds of reporting one higher health score.  Somewhat interestingly, respondents of East Asian ancestry were significantly less likely to report positive health scores compared to all other ethnic groups.  When the walkability index score is added to the model (Model 4), the 223  significance of smart growth matching disappears37.  Living in a more walkable neighbourhood is significantly associated with reporting a more favourable health status.  6.4.2 Models predicting BMI While a perceived health outcome is interesting, is neighbourhood match associated with another health measure that may suffer from less response bias?  While self reported by study participants, and not objectively measured, BMI is at least based on measurable variables that are not subject to much interpretation: height and weight.    Because BMI was normally distributed in the sample (see Chapter 3), an ordinary least squares regression was used for estimation, and the dependent variable was log transformed to interpret the results as a percent change – as opposed to a unit change – in BMI.  The predictor variables used in each of the 4 models are identical to the ones used to predict health status.  Table 67 describes results for the subjective match model, while table 68 describes results for the objective match model. Table 67 reveals that no subjective match category is significantly associated with any change in BMI. Being objectively matched into a suburban environment was associated with a .037% rise in BMI, although only at a significance level of p<0.10 (Table 68).  The inclusion of neighbourhood satisfaction dummies in model 2 was not significantly associated with any gain in BMI.  Among socioeconomic and demographic predictors, men were associated with significantly lower BMI, an extra year of age added 0.002% to average BMI scores, and respondents of East Asian ancestry had 0.094% lower BMI values than other ethnicities.  This last finding contrasts with the results of self-reported health status where East Asians were more                                                  37 It is important to note that objective neighbourhood match is defined based on the walkability index, so interpreting the addition of walkability as an additional predictor variable in the objective match model (Table 21) should not be interpreted. 224  likely to be dissatisfied with their health compared to other ethnic groups.  For every 1 unit increase in walkability there is a 0.005% decrease in BMI, confirming a multitude of evidence linking the built environment that people reside in to their obesity levels.  When walkability is controlled for, people who are mismatched into higher Smart Growth environments actually report significantly higher BMI levels compared with people who were mismatched into lower Smart Growth environments.  This finding confirms those demonstrated by Frank et al. (2006), which showed that the actual neighbourhood environment had a greater impact on walking behaviour and predicting BMI than whether people were matched into the neighbourhood types of their preferences.225  Table 67: Ordinary least squares regression predicting BMI, subjective neighbourhood match categories  Model 1 2 3 4 Variable Coeff. Sig. Coeff. Sig. Coeff. Sig. Coeff. Sig.          Constant 3.230 .000*** 3.238 .000*** 3.203 .000*** 3.222 .000***          Subjective Match          Prefers Smart Growth and Matched  .007 .591 .008 .539 .007 .592 .017 .208 Prefers Suburbia and Matched .007 .768 .006 .785 -.004 .850 -.013 .548 Mismatched into higher SG .010 .527 .008 .582 .017 .264 .022 .143          Other Predictors         WalkIndex       -.005 .001** Age     .002 .000*** .002 .000*** Gender_Male     -.047 .000*** -.048 .000***          HHInc_$40_$100k     .006 .626 .003 .823 HHInc_over$100k     -.002 .923 -.004 .806 Ethn_European     -.007 .533 -.007 .554 Ethn_Canadian     .019 .382 .019 .364 Ethn_EastAsian     -.094 .000*** -.092 .000***                   NeighbSat_very_sat   -.023 .057     NeighbSat_very_unsat   -.009 .642              N   1,081    1,081    1,081    1,081  R2  .000  .004  .090  .099 Adjusted R2  -.002  -.001  .081  .090 * Significant at p. <0.05  ** Significant at p.<0.01  ***Significant at p.<0.001 Dependent variable is the natural log of BMI. 226  Table 68: Ordinary least squares regression predicting BMI, objective neighbourhood match categories  Model 1 2 3 4 Variable Coeff. Sig. Coeff. Sig. Coeff. Sig. Coeff. Sig.          Constant 3.227 .000*** 3.236 .000*** 3.205 .000*** 3.218 .000***          Objective Match          Prefers Smart Growth and Matched  .010 .507 .008 .585 .014 .305 .021 .130 Prefers Suburbia and Matched .037 .063 † .036 .071 † .028 .142 .018 .350 Mismatched into higher SG .006 .688 .005 .741 .011 .438 .038 .021*          Other Predictors         WalkIndex       -.005 .001** Age     .002 .000*** .002 .000*** Gender_Male     -.048 .000*** -.048 .000***          HHInc_$40_$100k     .005 .708 .002 .890 HHInc_over$100k     -.003 .852 -.005 .773 Ethn_European     -.008 .497 -.010 .423 Ethn_Canadian     .018 .384 .015 .478 Ethn_EastAsian     -.094 .000*** -.092 .000***                   NeighbSat_very_sat   -.022 .062     NeighbSat_very_unsat   -.009 .664              N   1,081    1,081    1,081    1,081  R2  .003  .007  .090  .101 Adjusted R2  .001  .002  .082  .091 † Significant at p. <0.10 * Significant at p. <0.05  ** Significant at p.<0.01  ***Significant at p.<0.001 Dependent variable is the natural log of BMI. 227  6.4.3 Summary of Findings The results of the health models reveal that being matched into the environment you desire – whether toward Smart Growth or toward suburbia – has little bearing on your BMI or your self-reported health status, at least when other, more relevant characteristics are controlled for.  Your age, ethnicity and whether you live in a walkable neighbourhood – regardless of whether it is the neighbourhood type you prefer – all have more significant effects on your self-reported health status and BMI.  6.5 Conclusions The three research questions have been answered separately in a series of models presented in this chapter, but no attempt has been made to tie all the evidence together.  It is easy to lose perspective after being presented with forty pages of regression outputs, so Figures 37 and 38 link the models together in a coherent framework.  These two figures present the accumulated evidence linking increased neighbourhood housing mix to different types of Smart Growth matching and on to satisfaction and health outcomes based on the various models explored in this chapter.  Although it is almost certain that some relationships are bidirectional, the framework is presented in a linear format, with the arrows indicating the direction of the relationship studied in the model.  A green arrow indicates a significant, positive relationship, while a red arrow indicates a significant, negative relationship38; a gray arrow denotes that the relationship is not significantly different from that of the reference group, where respondents are mismatched into areas with less Smart Growth than their preferences. Figure 39 shows the                                                  38 Although the dependent variable for research question 2 was neighbourhood “dissatisfaction”, for simplicity, it is presented as neighbourhood “satisfaction” so that a negative arrow indicates a significant predictor that a respondent will be less satisfied with their neighbourhood. 228  pathway of relationships in the models where no control variables were included, and only the relationships indicated in the figure are accounted for.  Figure 40, by comparison, shows the same pathway in models where SES and demographic variables are controlled for.  The specific table and model references are indicated above the arrows. Figure 39: Outcomes pathway, no control variables  Colour of arrows indicate significance of relationships (green = positive; red = negative; gray = no significance) of respondents who were matched to Smart Growth compared to respondents who were mismatched into a lower Smart Growth environment than their preferences (reference group).     Housing Mix(2 km)Objectively Matched into SGSubjectively Matched into SGNeighb. SatisfactionHealth StatusBMI229  Figure 40: Outcomes pathway, controlling for SES and demographic variables  * Controlling for other built form covariates (Table 63). Colour of arrows indicate significance of relationships (green = positive; red = negative; gray = no significance) of respondents who were matched to Smart Growth compared to respondents who were mismatched into a lower Smart Growth environment than their preferences (reference group).  From figures 3 and 4, a bigger picture emerges: housing mix is associated with the ability for respondents to match into a Smart Growth neighbourhood, when match is measured objectively, but the pathway to greater neighbourhood satisfaction and better health outcomes ends there.  Housing mix is not a significant predictor that respondents will perceive that they are matched to the neighbourhood environment they prefer (i.e. subjectively matched), which means that housing mix does not contribute to the significant, downstream association between people who subjectively match into Smart Growth and higher self-reported health statuses.  When SES and demographic variables are controlled for, increased neighbourhood housing mix is negatively associated with the level of neighbourhood satisfaction that respondents report.  Even the finding that increased housing mix enables households to objectively match into the Smart Growth neighbourhoods of their choice is incomplete: only certain groups of people enjoy the significant relationship between housing mix and objective neighbourhood matching.  People Housing Mix(2 km)Objectively Matched into SGSubjectively Matched into SGNeighb.SatisfactionHealth StatusBMITable 63*, models 2 & 4230  over 60 and renters are among the subgroups for which housing mix has no effect on residential matching.  And, yet, these are precisely the groups that housing mix policies are intended to help.  People under 60 and owner occupiers – who seem to be the beneficiaries of neighbourhood housing mix – were not the intended target of the RGS’ “complete communities” strategy (Metro Vancouver 2011), nor were these groups the targets of local and regional affordable housing strategies (City of Vancouver 2013; Metro Vancouver 2007).  There is a possibility that housing mix helps people who could have accessed the neighbourhood of their choice, anyway.  Another possibility is that neighbourhoods with a diverse range of housing typologies may, in fact, be highly coveted, and therefore attract people with the affluence or ability to sort into their preferred community in any case.  On the surface, then, housing mix does not appear to be a very effective tool for enabling people to settle where they want.  In turn, allowing people to settle where they want does not seem to lead to positive social outcomes such as greater neighbourhood satisfaction or better health outcomes, at least when other factors are taken into account.  Should we dismiss housing mix as a toothless policy, or do local housing affordability circumstances interfere with its mode of action?  Is a neighbourhood with a high housing mix a highly coveted, luxury product? Paradoxically, a community with a diverse range of housing typologies might also have features that attract affluent people, and therefore work against the intent of a policy that was meant to house people of lesser means or abilities.  Do the findings in this dissertation reflect actual behaviour, or are there flaws in the research design that may lead to false answers?  Finally, if we are provided with this new knowledge on housing mix – and we accept the evidence at face value - what should planners and policymakers do?  These issues will be explored in the final chapter of this dissertation. 231  Chapter 7: Conclusions 7.1 Introduction In the final chapter of this dissertation, the evidence on neighbourhood housing mix is reviewed and a judgement is made about the usefulness and viability of housing mix as a strategy to house people in their preferred neighbourhoods.  The discussion chapter begins by acknowledging the many limitations inherent to the design of this research project which may have biased results.  The challenge of using Metro Vancouver as a study context, in light of the region’s unique housing dynamics, is then considered in more detail. Conclusions are drawn about the usefulness of neighbourhood housing mix as a planning strategy, and a variety of policy alternatives are considered.  Finally, future research studies building off the lessons learned from this dissertation are outlined.  7.2 Limitations of Research Design The findings of this dissertation have revealed that housing mix is a significant predictor that households with Smart Growth preferences will objectively match into these types of communities, but not that they will subjectively match, and that there is no further link between the ability to objectively match and one’s health and satisfaction outcomes.  Are these results valid, or are there flaws in the way the research was designed that may have led to these particular outcomes?  Beginning with the sampling strategy and ending with a discussion on certain peculiarities of Metro Vancouver as an urban context, this section explores the possibility that issues with the design of the research project may have influenced results. The list of limitations explored in this section is not meant to be exhaustive, but is meant to highlight some 232  of the largest potential sources of error and bias that should be acknowledged when interpreting the results.  7.2.1 Potential Issues with the Sampling Strategy Most of the data obtained for this dissertation was sourced from the CLASP survey, a stated residential preference study whose intent was to document the demand and satisfaction for Smart Growth communities and, to a lesser extent, to draw a connection between neighbourhood types and individual travel behaviour and health outcomes.  The sampling strategy employed to obtain a representative sample was to stratify by household income tertiles and walkability quartiles.  The aim of this dissertation, on the other hand, was to investigate the link between housing mix and the ability to match into different neighbourhood types.  Had this dissertation employed its own survey, a more suitable sampling strategy would have been to stratify based on the walkability index quartile (to proxy for neighbourhood type) and to incorporate the full variance of neighbourhood housing mix scores in the region.  A retrospective cross-tabulation of the frequencies of observations by walkability quartile and housing mix quartiles (measured around respondents’ addresses at 2km) reveals that lower walkability index scores (e.g. 1st and 2nd walkability quartiles) are underrepresented in the sample (Table 69). Within the highest housing mix quartile, there is a distinct lack of respondents who live in the lowest walkability quartiles.     233  Table 69: Walkability quartile and housing mix quartile, cross-tabulation of frequencies   Walkability Index Quartile1  Housing Mix Quartile2 1st 2nd 3rd 4th Total        1st N 81 46 38 156 321  %3 25.2% 14.3% 11.8% 48.6% 100.0%        2nd N 27 45 70 136 278  % 9.7% 16.2% 25.2% 48.9% 100.0%        3rd N 11 38 88 144 281  % 3.9% 13.5% 31.3% 51.2% 100.0%        4th N 8 31 62 179 280  % 2.9% 11.1% 22.1% 63.9% 100.0%        Total N 127 160 258 615 1,160  % 10.9% 13.8% 22.2% 53.0% 100.0% 1. Measured at a 1 km network buffer around a respondent’s postal code. 2. Measured at a 2 km network buffer around a respondent’s postal code. 3. Percentages reported are row percentages (i.e. percent of housing mix quartile responses living in different walkability index quartiles).  These findings may reflect reality: it is reasonable to assume that the areas with the lowest walkability index scores also contain overwhelmingly low housing mix scores, since these communities almost exclusively comprise low density, suburban and exurban areas that only contain single family detached homes.  Deliberately oversampling respondents who lived in high housing mix but low walkability index communities likely would have provided a more robust comparison of how housing mix affects the ability for individuals to sort into Smart Growth areas versus suburban areas.  Currently, the results suggest that greater housing mix lowers the odds that households with preferences for suburban areas (i.e. areas with a low walkability index) will match, either subjectively or objectively, but this finding may have been based on only a handful of observations.  Luckily, for the purposes of this project, the main findings that were reported involved a comparison of the ability to match into Smart Growth (i.e. higher 234  walkability scores) versus being mismatched into more suburban areas than preferred.  The number of observations across different housing mix scores in higher walkability quartiles is roughly evenly distributed.  The opportunity for purposive sampling in the future should extend to deliberately oversampling population subgroups that are the intended beneficiaries of housing mix policies.  For example, there were not enough families with children represented in the sample, which forced the researcher to abandon any attempt at running multinomial logit models to determine if housing mix had any effect on the settlement patterns of families.  This is a major omission, since housing families with children in pricy, Smart Growth-type communities has been the subject of both local concern over housing mix sufficiency (City of Vancouver 2015b; Villagomez 2011), local policies to attract families with children to Smart Growth neighbourhoods (City of Vancouver 1992) and even articles in the popular media (Benfield 2014; Egan 2005).   7.2.2 Potential Issues with the Design of the Survey Instrument Beyond the sample of respondents queried, another possible source of bias might be the survey instrument that respondents were presented with.  Issues with earlier generations of residential preference surveys have already been raised in Chapter 3, Section 3.2, including the use of single variables to measure residential environment and preference, and the lack of incorporating implicit trade-offs  between environments in the phrasing of questions.  For the most part, these issues have been addressed in the CLASP survey instrument.  Still, other issues remain. The CLASP instrument relies on visual cues and it is possible that some respondents focused on the illustrations rather than reading the descriptions that accompanied the pictures 235  (Malizia and Exline 2000).  The visual cues used in the CLASP survey are identical to ones used in previous residential preference surveys, most notably a major study conducted in 2003 in Atlanta (Levine and Frank 2006).  One of the most important sources of bias may come from the fact that the visual cues and textual descriptions of the trade-offs were not recalibrated to adjust for the substantial differences between the built form (in both Smart Growth and suburban areas) and typical property prices in Metro Vancouver versus Atlanta.  For example, the fifth trade-off question describes a trade off to be made between a Smart Growth and a suburban neighbourhood –where the Smart Growth option entails nearly identical single family homes to the suburban option, albeit on lots less than half the size (Figure 41).  Implicit in this trade-off is that both options are priced roughly equivalent to one another. Figure 41:Trade-off question 5 from CLASP survey    236  In Metro Vancouver, the trade-off may operate the other way around; lot sizes of single family homes may not differ markedly between Smart Growth and suburban areas, but price differences are considerable.  In the City of Vancouver, where the majority of the region’s Smart Growth neighbourhoods are to be found, the median lot size for a single family detached home in 2011 was 4,705 ft2, and the median property value of single family homes was $980,000.  In Maple Ridge, an outer suburban municipality about 30 km to the east, the median lot size for single family homes was 7,200 ft2 (or roughly 53% larger than in the City of Vancouver), while the median property value for single family homes was $437,000 – or less than half of what one might have paid in the City of Vancouver in 2011.  Indeed, the trade-off implied in Figure 39 is misleading because no single family home – and not even most two bedroom condominium apartments – can be purchased in the central-most portions of the region for the price that single family homes command on the region’s periphery.  This particular trade-off question likely biased responses toward preference for Smart Growth neighbourhoods.   In future preference studies, the trade-offs posed should be recalibrated to reflect regional housing contexts.  To do this, a hedonic regression can be applied on a sample of property sales across a broad spectrum of neighbourhood types in Metro Vancouver (Green and Malpezzi 2003).  “Spatial submarkets” (Jones, Leishman, and Watkins 2003; MacLennan and Tu 1996) – geographic areas where the value for a housing service differs significantly from other areas – can then be identified and hypothetical “homes” consisting of a different bundle of attributes can be imputed for different areas of the region roughly corresponding to equivalent price points.  Alternatively, a trade-off like the one depicted in figure 39 can be presented but with an explicit reference to the price differences between the two communities.  The trade-off description would acknowledge, for example, that the Smart Growth neighbourhood can be had as is – but for a 237  half a million dollar price premium. Incorporating price and attribute trade-offs can be expanded upon by asking respondents at which price point they might be willing to make a trade-off, and comparing the dollar value to their household income as a rudimentary way of examining the “willingness to pay” to live in Smart Growth.  While these ideas are conceptual, and likely fraught with problems of their own, at the very least trade-off questions in stated preference surveys should be adjusted to reflect local market realities. Given the high price point of much of Vancouver’s Smart Growth supply, another unique research opportunity that adds both to understanding Smart Growth preferences and housing affordability knowledge would be to investigate whether low income households made personal sacrifices to live in Smart Growth communities.  The evidence from this dissertation has shown that many people, including lower income renters, prefer Smart Growth.  But what other aspects of life did they have to forego to live in these kinds of areas? Households find ways to adapt to expensive housing circumstances by foregoing other necessities of life; when a sufficient number of life’s necessities are sacrificed, individuals begin to suffer from housing affordability problems (Stone 2006).  This “shelter poverty” conception of housing affordability (ibid) contrasts with the much more blunt measurement of shelter expenditure to income that has been roundly criticized for focusing on an arbitrarily-defined ratio that ignores the utility people may have of paying high amounts for a home in certain areas (Hancock 1993; Hulchanski 1995).  The list of North American regions known for their Smart Growth environments also reads like a list of the most expensive cities on the continent –Vancouver, Boston, the Bay Area, New York City – and this has not escaped the eye of critics (Demographia 2008).  If Smart Growth offers demonstrated benefits to its residents, more research is needed to understand what sacrifices the 238  most economically vulnerable groups have made to continue to live in these communities and, hopefully, to design policies to address these needs.  7.2.3 Potential Issues from Survey Responses Some of the sources of bias in survey responses may be expected from the inclination of respondents to respond a certain way, rather than to the actual questions as they are designed.  For example, a theory was proffered in Chapter 5 that renter households were necessarily constrained to certain areas of the region where rental housing was supplied.  A distribution of renters and owners by walkability index quartile shows that renters are much more confined to the topmost walkability quartile (Table 70), and suggests that the region’s rental stock may be confined to areas with the most Smart Growth.  Because renters are confined to these areas, their “preferences” for Smart Growth may have, in fact, been forced choices based on where they were able to settle.  Similarly, households without the financial means of car ownership may have been compelled to prefer Smart Growth not because they disliked driving, but because they had no choice but to situate themselves in areas where car ownership was not as much of a necessity.   Table 70: Walkability quartile and tenure, cross-tabulation of frequencies   Walkability Index Quartile1  Tenure 1st 2nd 3rd 4th Total        Renters N 16 45 94 323 478  % (total) 3.3% 9.4% 19.7% 67.6% 100.0%        Owners N  130 122 164 292 708  % (total) 18.4% 17.2% 23.2% 41.2% 100.0%        Total N 146 167 258 615 1,186 1. Measured at a 1 km network buffer around a respondent’s postal code. 239   Additionally, measures of satisfaction with one’s neighbourhood should be interpreted with caution.  Nearly forty years ago, Campbell already demonstrated that people had a tendency to project their satisfaction with an unrelated aspect of their life onto their self-reported satisfaction with their neighbourhood (Campbell, Converse, and Rodgers 1976).  This was shown when self-reported health status was a strong, significant predictor of neighbourhood satisfaction, even though the two should, technically, not be at all related; a rational actor would ostensibly be able to separate the two conceptions before making a rating.  It is also possible that people adjust their satisfaction upward and shift their neighbourhood preferences toward the circumstance in which they find themselves.  This variation on Tversky and Kahneman’s (1974) status quo bias (viz. Marsh and Gibb (2011)) should be more fully explored in future residential preference studies.  A longitudinal approach that administers the same preference survey to subjects before and after they move homes could theoretically compare responses to questions about neighbourhood preference and account for this bias39.   It may be prudent for future residential preference and Smart Growth satisfaction studies to integrate more lessons from behavioural economics.  7.2.4 Potential Issue with the Construction of Variables The raw data obtained from the survey and other data sources are eventually turned into useful variables for analysis, adding another layer of uncertainty and bias.  Some of the issues surrounding the use of network buffers have already been discussed in the methods chapter.                                                   39 Luckily, at the time of writing, the researcher had just concluded gathering data from exactly this kind of study. 240  Most glaringly, limitations in computing power meant that more than 100 observations had to be dropped when computing the housing mix around a 5 km network buffer of a respondent’s address.  The remaining 1,030 observations were properly calculated, but the models are fitted on a reduced sample.  This means that models of how housing mix at a 5km network buffer predicts neighbourhood match should be interpreted with caution.  Data on the structural attributes of rental apartment buildings was also missing from the parcel data, and could not be incorporated into the construction of the BuiltForm_2km_Factor principal component. The definition of the variable “housing mix”, itself, is based on the diversity and distribution of seven housing types defined for property assessment purposes.  Whether these seven types affect housing behaviour is unknown – indeed, that was the crux of the original research question.  More empirical means of defining housing mix, such as running latent class analyses based on inputs of square footage, property value, and number of bedrooms, were pursued but later dismissed as being too detached from useful policymaking.  After all, homes are approved and built according to local zoning and development regulations, not the unique latent classes identified by a faraway researcher.  While the seven housing types were used to define housing mix because of a lack of other empirical guidance, it is almost certain that the models would have yielded different results if the categorization scheme were arranged differently. The appropriate geographical scale at which to measure housing mix is another limitation.  Ideally, the geography at which to measure housing mix would be analogous to the psychological area in which each household would search for homes.  This psychological area is unique from home seeker to home seeker, and likely depends on a variety of personal factors ranging from the proximity to work, to proximity to relatives and friends, to even intangible aspects of neighbourhood character.  In this dissertation, 5 network buffers of increasing size 241  were used to model this search area, and the network buffer at 2 km was shown to have the greatest predictive ability.  However, the 5 network buffer sizes were arbitrarily defined at set distances from each centroid (e.g. 500m, 1 km, 2 km, 3 km and 5 km, regardless of the location).  An empirical method exists of defining where the average household may search for homes, and this may be a more robust spatial unit over which to define housing mix.  This approach is the spatial submarket approach (Jones, Leishman, and Watkins 2003; MacLennan and Tu 1996), where a cluster analysis identifies the rough geographical outlines of precincts within the urban housing market where home values, controlling for all other structural attributes of homes, are significantly similar.  If a detailed dataset exists of recently-transacted homes exists, hedonic regressions can be applied and cluster analyses performed to reveal the boundaries of distinct spatial submarkets in which housing mix can be measured. The approach to defining the main dependent variable “neighbourhood match” is also a source for uncertainty.  Preference was defined on a small choice set of just five trade-off questions, rather than the original eight, since three questions made explicit references to housing mix and would have thus introduced an element of the independent variable into the dependent variable.  Five questions represents a definite improvement over earlier generations of stated preference surveys which attempted to elicit a preference for a complex, multi-faceted good such as Smart Growth attributes in a residential neighbourhood with a single question.  Still, the five questions do not capture the full complexity of the different amenities and disamenities that Smart Growth offers to different individuals40.                                                      40 Nor would it be possible to capture all the nuances of what a neighbourhood represents in a stated preference survey.  At some point a line must be drawn that suggests that the answers obtained from the survey instrument are “good enough” to portray the kind of community that a respondent prefers and the kind of environment that a respondent imagines they live in.  242  These five questions are then subjected to a principal components analysis to isolate a single component representing the Smart Growth attributes of a neighbourhood.  For “subjective neighbourhood match”, the same factor loadings are used to calculate both the neighbourhood preference score (NPS) and the current neighbourhood score (CNS).  Additionally, it is assumed that the respondent answering the question has the same spatial concept of “neighbourhood” when they answer the questions about preference as when they assess their own surroundings.  Because the NPS and CNS are identical both conceptually and spatially, we can assume that “subjective neighbourhood match” is a relatively robust indicator of how well people are able to match into their preferred environment.  “Objective neighbourhood match”, on the other hand, may represent more of an “apples to oranges” comparison; preference scores are elicited the same way as with subjective neighbourhood match, but the current neighbourhood is represented by a walkability index score calculated around a 1 km network buffer of each respondent’s postal code.  While the walkability index and NPS were standardized to one another (see Chapter 4, section 4.4.3.2), ultimately the concept of “neighbourhood” being compared  to define objective neighbourhood match come from separate sources and are conceived of (both the attributes being compared and the spatial area under evaluation) in different ways.  The walkability index and a respondent’s evaluation of their own neighbourhood (i.e. the “current neighbourhood score”) were validated against each other using a correlation test and a simple logistic regression to ensure that the two measures of neighbourhood environment were roughly similar (Chapter 3, section 3.4.3.2).  They both are significantly and strongly correlated to one another, and they both predict preference scores in similar ways, but they are ultimately not the same.  This means that objective neighbourhood match scores and subjective neighbourhood match scores cannot be directly compared against one another, and that objective neighbourhood match may not 243  represent a “true match” of preference to environment in the same way that subjective neighbourhood match does.  It is difficult to ascertain how this issue may have affected results, but this limitation to the design of the objective neighbourhood match variable should be acknowledged, especially in light of the finding that housing mix only significantly predicts objective, rather than subjective, neighbourhood match.  The meaningfulness of housing mix, as a policy to house people in the neighbourhood types that they prefer, should be considered in light of this result.  7.2.5 Potential Issues with the Statistical Models The statistical models used to test the relationship between the variables leads to another source of uncertainty and potential error. Logit models were chosen to predict the odds that a respondent “matched” into their particular neighbourhood of their preference. As described in Chapter 4, section 4.4.3.3, the original intention of measuring match as a continuous variable was abandoned because a normal distribution of the degree to which individuals are matched into their environment was impossible even after applying transformations. Not surprisingly, this reflects the fact that far more respondents were more “matched” than  “mismatched” into the environment of their choice, so the distribution would have been invariably skewed. The use of logit models necessitated the definition of “match” as either a binary outcome or a categorical variable, even though almost no household exactly matches their residence to their preferences.  As a result, a model was used that predicts the odds of belonging to discrete categorical outcomes, even though “neighbourhood match” hardly fits the description of a discrete outcome.  Finally, the multivariate models of neighbourhood matching may not control for some of the factors behind why people sort into certain neighbourhoods.  Indeed, it is unlikely that Smart 244  Growth – or suburban – neighbourhood character is the most important reason that people choose to reside in the neighbourhoods they do.  To a certain extent these factors are accounted for by running a factor analysis of the importance people assign to 25 considerations of home and neighbourhood attributes that lie outside of the domain of Smart Growth or suburban built character (Question B1 of the CLASP survey; see Appendix).  Additionally, income and a principal component representing the size, cost, yard presence and number of bedrooms of homes within the area are controlled for.  Given the data available, this model attempts to represent all of the important factors that affect household’s decisions on residential location, but it is extremely unlikely that these measures capture all the factors of any household.  7.2.6 Summary of Limitations George Box, a British mathematician once remarked that “all models are wrong” but some were useful (Box 1976).  This truism applies to social science research as well; models that attempt to represent reality are necessarily circumscribed and many of these limitations are difficult, if not impossible, to account for.  The limitations identified in this project touch almost every aspect of its research design.  Many of them are circumstances of the nature of the original study: the sampling strategy was not intended to answer a question about housing mix, so stratifying and oversampling respondents in certain housing mix segments was never considered.  Others should be remedied in future studies: survey questions that present choices of neighbourhood design should incorporate the realities of the marketplace – both in terms of what types of homes are available and what they may be priced at. Other shortcomings reflect the novelty of the research question: since housing mix has never been the subject of intense study, there is little guidance on how to define mix based on the typology of homes available and an 245  arbitrary categorization scheme was applied. Finally, some limitations stem from the nature of the data or the availability of computing power: observations were dropped from housing mix calculations around 5 km network buffers and the non-normality of the data led to the selection of a logit model to predict the odds of whether households matched, or not, rather than an ordinary least squares model that estimated how closely households lived in an environment that aligned with their preferences. Taken together, these limitations would have undoubtedly affected the final results that were presented in profound ways. However, one other aspect of the research design that might have influenced the results deserves further scrutiny and its own section.  Perhaps the most interesting source of bias does not come from the design of the research but from the context in which this phenomenon is studied. As a metropolitan region in North America, Metro Vancouver may be so unique in its planning history and the dynamics of its housing market that it cannot serve as a suitable test bed for understanding the impact that housing mix may play on residential matching, let alone provide results that one can render universal judgments upon. This dilemma is considered in the following section.  7.3 The Unique Context of Metro Vancouver For many, Vancouver has become a planning model for other cities (Grant 2009; Punter 2003).  But does that mean that outcomes from Vancouver-based models should be used to help other cities plan?  As described in Chapter 3, Metro Vancouver has been lauded for its extensive history of regional coordination in growth management (Tomalty 1997), its commitment to securing good urban design outcomes in new developments (Punter 2002), and innovative practices in tailoring residential developments to specific groups (MacDonald 2005).  Unlike many American (and other Canadian) metropolitan regions, Smart Growth living has always 246  been desirable; a mix of housing typologies within neighbourhoods has been a longstanding policy, and these findings have revealed that every municipality in the region seemed to contain some degree of housing mix.  Vancouver seems, on the surface, to be an ideal context for studying the effects of housing mix and Smart Growth preferences. However, Vancouver is also infamous for its housing affordability problems..  The beauty of Vancouver’s physical setting and its clement weather41, its long history of real estate development and speculation, and its more recent ties to the Asia Pacific region and affluent Asian immigration (Ley 2010) has created a region that has attracted wealth, investment and resulting high home prices.  Although Metro Vancouver is neither the largest metropolitan area in Canada, nor the region with the highest incomes, it has historically commanded the highest housing prices in the country (Teranet 2015).  Exacerbating these conditions, Vancouver is also one of the most constrained metropolitan regions for outward development in North America, hemmed in by mountains, the ocean and the US border on three of its sides.  The average single family home in Metro Vancouver sold for $976,700 in 2014 (Metro Vancouver 2015).  Even the price point of housing typologies such as townhomes and rowhouses, midway in size and density between single family homes and multifamily condominium apartments, may be prohibitively expensive.  By 2014, the average sale price of a row or townhouse on the City of Vancouver’s desirable west side reached $738,000; the Metro Vancouver average price for a row or townhouse was $471,200 (ibid).   It is possible that cities and regions may only begin to take housing mix seriously when the intended products of housing mix strategies – such as mid-sized, ground-oriented attached                                                  41 At least by Canadian standards. 247  housing typologies – become luxury products.  At that point, housing mix ceases to be effective policy since the housing products it attempts to create cannot help a broad spectrum of the population sort into their desired neighbourhood.  On the other hand, a low-priced metropolitan region may not feel compelled to pursue a housing mix strategy at all, since most residents can be affordably housed in single family detached homes.  In both scenarios, the relationship between housing mix and the ability to match is difficult to disentangle.  This may be the paradox – or, perhaps more accurately, the “Catch 22” – of studying housing mix. Is it possible that a neighbourhood with a high housing mix is, in itself, a luxury “good”?  Housing mix might be a value-adding amenity not unlike access to a rapid transit station or a good view.  Alternatively – and perhaps more plausibly – builders may attempt to cater to demand in a desirable neighbourhood by shoehorning different, denser housing types while maintaining the ground-oriented character of a neighbourhood.  In that scenario, housing mix is a response to desirability rather than a cause.  In the absence of sales data to perform a hedonic regression of housing mix and other attributes on selling prices, a  correlation test was run on housing mix scores and average property values, both measured using the same 2 km network buffer around respondents’ addresses .  The results showed a negative, but weak correlation (-0.048) between property value and housing mix, but a relationship that is also insignificant even at p<0.10 (.0102).  When housing mix is mapped onto property values for the region (Figure 42), it becomes apparent that housing mix and property values are not correlated.  Some of the highest-valued areas, such as the western side of the City of Vancouver and the wealthy suburb of West Vancouver (a separate municipality) contain both very high and very low housing mix, respectively.  Meanwhile, lower-valued areas such as parts of suburban Surrey and Langley contain both areas of high and low neighbourhood housing mix. 248  Figure 42: Housing mix and property values, Metro Vancouver   The lack of a trend is even clearer in the central-most part of the region (Figure 41).  In the City of Vancouver, high levels of housing mix cross wealth lines, with wealthy neighbourhoods such as Kitsilano having comparable levels of housing mix to neighbourhoods in traditionally working class East Vancouver, where comparable properties are often hundreds of thousands of dollars less expensive.   249  Figure 43: Housing mix and property values, central Metro Vancouver  If these maps are to be trusted, the region’s commitment to residential intensification seems to be working.  A diversity of housing typologies are found in all corners of the region, and across areas with a diverse range of socioeconomic makeup and Smart Growth attributes.  Despite the mix of housing across all neighbourhood types, housing mix does not appear to predict matching abilities.  Nor is a lack of housing mix synonymous with suburbia; the areas with the lowest housing mix appear to be in downtown Vancouver among the densest, most walkable neighbourhoods in the entire region.  Neighbourhood housing mix may be a principle of the Smart Growth movement, but it is not necessarily a feature of Smart Growth neighbourhoods – at least not in Metro Vancouver.  An objection may be raised that a 2 km 250  network buffer encompasses housing typologies that may be far away and even physically segregated from other parts of the same neighbourhood.  Is a stable single family home neighbourhood adjacent to a large-scale apartment complex really the type of community that regional policymakers envision when they advocate “complete communities”?  Or do they imagine neighbourhoods where next door neighbours live in different housing typologies; where the mix is much more fine-scaled and mid-rise apartment blocks stand adjacent to single family homes and jostle with laneway homes?  Maybe neighbourhoods conceived in this way are better able to integrate a broader spectrum of the population.  However, this dissertation has demonstrated that measuring housing mix at a fine scale (i.e. within 500 meters – or an area reached in less than a ten minute walk) is no better at predicting the ability of households to match, and that predicting residential match peaks around a 2 km search geography.  7.4 Is Housing Mix Useful Policy? If the limitations of this dissertation are acknowledged and the findings are accepted, then increasing the diversity of housing typologies at the neighbourhood level is not a useful strategy for enabling people to match into their preferred neighbourhood types.  Housing mix exists across a broad spectrum of neighbourhoods, by design and by price, but still fails to be associated with the ability for individuals to sort into these communities. This conclusion is underscored by findings showing that housing mix is only a significant predictor of one type of neighbourhood match – objective match – and then only for owner-occupiers, and respondents under 60.  In other words, population subgroups like senior citizens and renters, who have long been the intended beneficiaries of housing mix policies, are not impacted by these initiatives at all.  Moreover, objectively matching into Smart Growth communities does not predict any 251  further positive outcomes such as improved neighbourhood satisfaction levels, reporting a better health status, or reduced BMI levels.  Controlling for SES and personal factors, greater housing mix appears to significantly lower a respondent’s satisfaction with their neighbourhood.  Housing mix may even be a solution in search of a problem: half of respondents felt that they had sorted into their preferred neighbourhood type, and the vast majority of respondents were very satisfied with the neighbourhoods they lived in.  While housing mix within a neighbourhood has been championed by numerous policy documents as a means to assist people of limited financial means in sorting into their desired neighbourhoods, the findings from this dissertation suggest that there is no evidence to support that claim.  7.5 Future Policy Directions and Future Research Directions So if housing mix is ineffective, what policy alternatives exist to house people in Smart Growth areas?  And what policies might be useful in regions, such as Vancouver, where land costs are high and homes are often expensive?  The final section of this chapter will ruminate on some of the strategies that exist, and then offer a suggestion as to what sort of policy could be enacted in a region like Metro Vancouver.  Future research trajectories will also be offered.    7.5.1 Policy Alternatives This dissertation has shown that housing mix is ineffective at enabling a greater proportion of the population to reside in the neighbourhood of their preferences. This dissertation has also shown that the majority of respondents – both as a whole, and across different population subgroups - prefer neighbourhoods that have more Smart Growth attributes than the average neighbourhood in the region.  If housing mix does not work, what alternative policies 252  exist to sort people into Smart Growth communities, and would they be effective in Metro Vancouver?  Increasing the housing mix in existing Smart Growth neighbourhoods can loosely be considered to be a “supply side” approach to achieving a goal of greater Smart Growth matching, since it ostensibly involves changes to the residential stock in those areas through new construction or conversion.  On the “demand side”, the most prominent policy to achieve Smart Growth matching into existing properties may be the use of “Location Efficient Mortgages” (LEM) (Center for Neighborhood Technology 2010).  A LEM is effectively a preferential mortgage rate given to home buyers who choose transit-friendly locations often associated with Smart Growth neighbourhoods.  The premise behind LEMs is that households who live in transit-friendly areas spend less on transportation since they can forego car ownership (Brookings Institution, Center for Transit Oriented Development, and Center for Neighborhood Technology 2006).  The transportation-related savings are analogous to an increase in disposable income that can be used to offset higher housing costs.  In this sense, “location efficient” homeowners may be less risky to default than homeowners with similar financial circumstances in areas where car ownership may be a necessity.  Relative to other Smart Growth or housing policies, LEMs initially received widespread media attention because they seemed to promise low levels of government intervention, and LEM programs could effectively be managed by private sector lenders (Blackman and Krupnick 2001).  However, the track record of LEMs has been more mixed.  Blackman and Krupnick (2001) analyzed over 8,000 Chicago-area loans matched to both location efficient and non-location efficient areas and showed that LEM-recipients were just as likely to default as non-recipients, and that making loans available to borrowers in location efficient areas was no less risky than giving such loans to a random sample 253  of borrowers.  The authors suggest that the cost savings of transport did not offset the added costs of inner city housing in Chicago, and that LEMs assume, perhaps incorrectly, that homeowners use the money saved from transportation and put it directly into mortgage payments (ibid).  Similar conclusions were reached in a policy report in Metro Vancouver written by McClanaghan & Associates (2006): the sheer cost difference of a home in an inner city, transit-friendly location in Metro Vancouver compared to a suburban home essentially negates any purported savings in transportation, and properties with desirable transportation access or proximity almost always integrate this added utility into their prices to some extent42.  Moreover, a large proportion of jobs in Metro Vancouver – as in many American metropolitan regions – are located in automobile-oriented suburban areas, necessitating car ownership (Metro Vancouver 2014). In light of the recent defeat of a regional plebiscite to fund public transit improvements in suburban areas, the prospect of being able to access these jobs by public transit seems unlikely to change any time soon.    Another alternative is to demand exactions from developers of new projects in exchange for the permission to build, and to use this money to fund various programs designed to house more people in Smart Growth communities.  Exactions, or impact fees, were the fastest growing source of municipal revenue in US municipalities in the 1990s and 2000s (Pagano 2003) and the manner in which they are levied can take on a variety of forms.  In Metro Vancouver, two impact fees familiar to any developer seeking a rezoning application are development cost levies (DCLs) and community amenity contributions (CACs).  DCLs are exactions charged on                                                  42 In a personal communication, Dale McClanaghan remarks about the difficulty of approaching experienced real estate economists at mortgage lending institutions and proposing a policy that, at its core, assumes that real estate markets are inefficient at valuing something as essential as location.  The familiar real estate agents’ refrain of “location, location, location!” comes to mind. 254  development calculated by square footage according to an area-specific fee schedule.  (City of Vancouver 2015a).  The money raised by DCLs is restricted to funding parks, replacement housing (if units were demolished), childcare facilities, and transportation infrastructure (ibid). CACs, on the other hand, are far more ad hoc in their application, and are in-cash or in-kind43 contributions to help improve neighbourhood-specific amenities or services on a case-by-case basis.  The specific outcomes of CACs are the product of a negotiation process between the developer and the city planning department. In theory, impact fees such as DCLs and CACs are intended to cover the social and economic costs imposed by new development, and to ameliorate burdens on existing infrastructure posed by the arrival of additional residents (Gyourko 1991; Altshuler and Goméz-Ibáñez 1993; Gray 2012).  Exactions have certainly been quite lucrative for the City of Vancouver, which estimates that DCLs and CACs contributed to $133 million in improvements in 2013 alone44 (City of Vancouver 2014).  Could CACs and DCLs be harnessed to pay for Smart Growth initiatives?   Unfortunately, the policy of demanding exactions from new development for revenue purposes runs into both practical limits and ethical challenges.  Altshuler and Goméz-Ibáñez (1993) contend that impact fees on new development are ethically problematic because they shift the responsibility of paying for current municipal costs onto future residents.  Gyourko (1991) has argued that impact fees represent a form of property tax abatement for existing residents, passing the responsibility for funding municipal services to individuals who do not yet reside in the community and cannot participate in local decision-making.  As a form of revenue generation, impact fees seem to align with the concepts advanced by endogenous zoning                                                  43 That is, constructed directly by the developer in lieu of providing funding to the City. 44 Since this value includes both cash revenue and in-kind contributions, the value is difficult to measure. 255  theorists and Fischel’s (2004) homevoter hypothesis where municipal governments regulate land to appease existing homeowners and to keep property taxes low (Calabrese, Epple, and Romano 2007).  In Vancouver, CACs are intended to pay for the capital construction of new amenities within the neighbourhood, but reflect existing demands, rather than the demands of incoming residents.  However, the temptation to use developer-paid CACs to fund existing operating expenses – usually by topping up endowment funds – has proven difficult to resist (Gray 2012).  Impact fees may also raise the cost of new supply as developers recoup the added cost of providing these amenities by increasing the asking price of their new units (Burchell and Mukherji 2004).  In particular, the ad hoc nature of exactions such as CACs, which are often subject to  lengthy and uncertain negotiation processes, impose time costs on new development, and may stifle the motivation of developers to produce new units (viz. Mayer and Somerville 2000).  Finally, the revenues or contributions raised by impact fees can vary greatly from year to year.  In 2013, the City of Vancouver raised $133 million from these exactions, but only raised $68 million in 2012; it nearly set a record with $180 million raised in 2011 (City of Vancouver 2014).  These variations reflect the timetable of when major developments are approved, and also the financial viability of pursuing large scale development projects (ibid).  The lucrativeness of CACs may also diminish if development occurs away from the downtown core, where city planners have less flexibility in granting developers density bonuses in exchange for handsome contributions (Gray 2012).  What is needed is a much more stable source of funding to pursue long-range Smart Growth objectives.  Such a source of funding would have to be applied fairly – on existing residents rather than on potential new ones.  Luckily such a source of funding may already exist. 256  With $3.68 paid in property taxes for every $1,000 in assessed value, Vancouve