UBC Theses and Dissertations

UBC Theses Logo

UBC Theses and Dissertations

A land use regression model for ultrafine particles in Vancouver, Canada Abernethy, Rebecca 2012

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

Item Metadata

Download

Media
24-ubc_2012_spring_abernethy_rebecca.pdf [ 5MB ]
Metadata
JSON: 24-1.0072650.json
JSON-LD: 24-1.0072650-ld.json
RDF/XML (Pretty): 24-1.0072650-rdf.xml
RDF/JSON: 24-1.0072650-rdf.json
Turtle: 24-1.0072650-turtle.txt
N-Triples: 24-1.0072650-rdf-ntriples.txt
Original Record: 24-1.0072650-source.json
Full Text
24-1.0072650-fulltext.txt
Citation
24-1.0072650.ris

Full Text

A LAND USE REGRESSION MODEL FOR ULTRAFINE PARTICLES IN VANCOUVER, CANADA  by Rebecca Abernethy  B.Sc., The University of British Columbia, 2008  A THESIS SUBMITTED IN PARTIAL FULFILLMENT OF THE REQUIREMENTS FOR THE DEGREE OF  MASTER OF SCIENCE in THE FACULTY OF GRADUATE STUDIES (Occupational and Environmental Hygiene)  THE UNIVERSITY OF BRITISH COLUMBIA (Vancouver)  March 2012  © Rebecca Abernethy, 2012  Abstract Background and Aims: Epidemiologic studies have associated adverse health outcomes with exposure to traffic-related air pollutants, principally NO2, at levels below those showing effects in controlled exposure studies. This suggests the importance of related outdoor air contaminants, such as ultrafine particles (UFP) (<0.1µm in diameter). Presently, no UFP monitoring exists in North America and little information is available regarding UFP spatial distributions. We measured particle number concentrations (PNC) in Vancouver to develop a land use regression (LUR) model for use in epidemiologic studies and to identify important sources of UFP. Methods: During a two-week sampling period in spring 2010, PNC were measured with portable condensation particle counters (CPC) for 60-minutes at eighty locations used previously to characterize spatial variability in nitrogen oxides. Continuous PNC measuring occurred at four additional locations to assess temporal variation. LUR modeling was conducted using 135 geographic predictors, including: road length, vehicle density, intersection and bus stop density, land use type, fast food restaurant density, population density and others, following previously developed methods. A novel buffer approach incorporated meteorologic data through wedge-shaped wind roses from measurements made during PNC sampling, in addition to circular buffers.  ii  Results: The range of measured (60-minute median) PNC across locations varied 70-fold (range: 1500 – 105 000 particles/cm3, mean [SD] = 18 200 [15 900] particles/cm3). Correlations of PNC with concurrently measured two-week average NO2, NO and NOX concentrations at the same sites were 0.64, 0.65 and 0.70. A model (R2 = 0.48, leave-one-out cross validation R2 = 0.32) predicting PNC included length of truck routes within 50m, density of fast food locations within 200m and ln-distance to the nearest port. LUR models created with wind rose shaped buffers had lower predictive power than models with circular buffers (R2 = 0.29 – 0.34). Conclusions: Measured PNC was highly variable across the Metro Vancouver region and correlated with nitrogen oxides. Geographic predictors explained a smaller proportion of variability in PNC than found previously for nitrogen oxides, suggesting some common sources and additional unknown factors influencing PNC spatial variability. This represents the first LUR model for UFP in North America.  iii  Preface The Metro Vancouver ultrafine particle sampling campaign, temporal data correction methods as well as summary statistics on measured PNC was presented orally at the 50th meeting of the Pacific Northwest International Section (PNWIS) of the Air and Waste Management Association in Missoula, Montana, November 3 – 5, 2010. Sampling results as well as preliminary LUR models for ultrafine particles UFP were shared on a poster presentation at the British Columbia Particulate Matter Symposium in Vancouver, British Columbia, June 9, 2011. Sampling results and LUR model results, including models incorporating meteorologic variables, were presented in a poster during the student symposium and in an oral presentation during the 23rd meeting of the International Society for Environmental Epidemiology (ISEE) conference in Barcelona, Spain, September 13 – 16, 2011. Measurement data from the sampling campaign (section 3.2) were included in a journal article entitled: ‘Proximity of public elementary schools to major roads in Canadian urban areas.’ This paper was submitted to the International Journal of Health Geographics in September 2011, accepted in December 2011 and is now published online in its final form. My contribution to this paper was the Metro Vancouver ultrafine particle sampling campaign measurement data, including design and execution of the measurements and adjustments (temporal and instrument) to raw data. In addition I reviewed the manuscript and provided comments.  iv  Table of Contents  Abstract.................................................................................................................................... ii
 Preface..................................................................................................................................... iv
 Table of Contents .................................................................................................................... v
 List of Tables .......................................................................................................................... ix
 List of Figures......................................................................................................................... xi
 List of Abbreviations ........................................................................................................... xiii
 Acknowledgements .............................................................................................................. xiv
 Chapter 1: Introduction ........................................................................................................ 1
 1.1
 Outdoor air pollution ................................................................................................... 1
 1.1.1
 Ultrafine particles ................................................................................................. 2
 1.1.1.1
 Health effects of UFP exposure ..................................................................... 2
 1.1.1.2
 Sources of UFP and factors affecting concentrations.................................... 5
 1.1.2
 Methods of modeling PNC: LUR and others........................................................ 8
 1.1.2.1
 Incorporating meteorology in LUR models................................................. 13
 1.2
 Rationale .................................................................................................................... 13
 1.3
 Objectives and hypotheses......................................................................................... 14
 Chapter 2: Methods ............................................................................................................. 16
 2.1
 UFP sampling ............................................................................................................ 16
 2.1.1
 Pilot study ........................................................................................................... 16
 2.1.2
 UFP sampling campaign study design................................................................ 17
 2.1.2.1
 Site selection ................................................................................................ 17
 v  2.1.2.2
 Timing of sampling campaigns.................................................................... 20
 2.1.2.3
 Sampling equipment .................................................................................... 20
 2.1.2.4
 Data collected during sampling ................................................................... 21
 2.2
 Data analysis .............................................................................................................. 24
 2.2.1
 Descriptive statistics ........................................................................................... 24
 2.2.2
 Temporal corrections .......................................................................................... 24
 2.2.3
 Instrument correction.......................................................................................... 26
 2.2.4
 Comparisons with fall 2009 PNC measurements ............................................... 26
 2.3
 LUR modeling ........................................................................................................... 27
 2.3.1
 Geographic predictor variables........................................................................... 28
 2.3.1.1
 Circular buffers ............................................................................................ 29
 2.3.1.2
 Wind rose shaped buffers ............................................................................ 30
 2.3.1.3
 Sources of geographic predictor variables................................................... 32
 2.3.2
 LUR model building procedure .......................................................................... 39
 2.3.3
 Regression maps ................................................................................................. 41
 2.4
 Evaluation of measurements and models................................................................... 41
 2.4.1
 Sixty-minute measurement evaluation................................................................ 41
 2.4.2
 Model evaluation ................................................................................................ 42
 Chapter 3: Results ............................................................................................................... 44
 3.1
 Sampling outcomes.................................................................................................... 44
 3.1.1
 Instrument correction factor................................................................................ 44
 3.1.2
 Temporal correction factors................................................................................ 44
 3.1.3
 Summary statistics .............................................................................................. 50
 vi  3.1.4
 Comparing PNC with nitrogen oxides................................................................ 52
 3.1.5
 Meteorologic data ............................................................................................... 54
 3.1.6
 Univariate correlations with PNC50 and geographic variables ........................... 56
 3.2
 PNC LUR models ...................................................................................................... 57
 3.2.1
 Sensitivity analyses............................................................................................. 60
 3.2.2
 Pollution surface images..................................................................................... 61
 3.3
 Evaluation .................................................................................................................. 66
 3.3.1
 Assessment of 60-minute samples...................................................................... 66
 3.3.2
 Comparing spring 2010 with fall 2009 measurements ....................................... 68
 3.3.3
 Internal validation ............................................................................................... 69
 3.3.4
 External validation.............................................................................................. 70
 Chapter 4: Discussion.......................................................................................................... 72
 4.1
 PNC measurements.................................................................................................... 72
 4.2
 LUR models............................................................................................................... 74
 4.2.1
 Comparison of Vancouver PNC LUR models with other LUR models............. 75
 4.2.2
 Comparison with other LUR models incorporating meteorology ...................... 82
 4.2.3
 Key findings of the Vancouver PNC LUR models............................................. 84
 4.3
 Strengths .................................................................................................................... 84
 4.4
 Limitations ................................................................................................................. 86
 4.5
 Applications ............................................................................................................... 88
 4.6
 Future work................................................................................................................ 88
 4.7
 Conclusion ................................................................................................................. 89
 References.............................................................................................................................. 91
 vii  Appendices............................................................................................................................. 99
 Appendix A Metro Vancouver ultrafine particle database ................................................. 99
 Appendix B Ultrafine particle sampling campaign manual.............................................. 103
 Appendix C Ultrafine particle and nitrogen oxides comparisons..................................... 119
 Appendix D Commands used in GIS variable extraction and R scripts........................... 123
 D.1
 Geospatial modelling environment command examples .................................... 123
 D.2
 R script examples................................................................................................ 124
 Appendix E 2010 Ultrafine particle sampling campaign field notes................................ 128
 Appendix F Predictor variables for PNC LUR models (‘classic buffer’)......................... 129
 Appendix G Predictor variables for PNC LUR model (based on measured wind rose data) .......................................................................................................................................... 130
 Appendix H Predictor variables for PNC LUR model (based on monitored wind rose data from a central monitoring station) .................................................................................... 131
  viii  List of Tables  Table 2.1 Data collected during spring 2010 UFP sampling campaign ................................. 22
 Table 2.2 Potential variables for UFP LUR modeling (adapted from (56)) ........................... 37
 Table 2.3 Summary of UFP LUR models created for Metro Vancouver ............................... 41
 Table 3.1 Within-day variation correction factors applied to raw data .................................. 45
 Table 3.2 Between-day variation correction factors applied to raw data ............................... 47
 Table 3.3 Summary statistics for corrected one-hour PNC at 80 sites: median, 10th and 90th percentile................................................................................................................................. 52
 Table 3.4 Correlation coefficients between PNC and nitrogen oxides................................... 52
 Table 3.5 Examination of three sites with high residual values when comparing NOX and UFP levels............................................................................................................................... 54
 Table 3.6 Summary statistics for site-specific and central station measured meteorology variables .................................................................................................................................. 54
 Table 3.7 Range of wedge buffer lengths applied in measured meteorologic LUR model.... 55
 Table 3.8 Selected univariate correlations with PNC50 .......................................................... 56
 Table 3.9 LUR models developed for PNC in Metro Vancouver .......................................... 58
 Table 3.10 Selected univariate correlations between fast food and port variables and trafficrelated variables ...................................................................................................................... 60
 Table 3.11 Correction factors for comparing fall 2009 to spring 2010 PNC measurements . 68
 Table 3.12 Predicting PNC at fall sites based on road length (PNC50) LUR model from spring sampling.................................................................................................................................. 69
 Table 3.13 Predicting PNC levels at repeated and fixed sites by road length (PNC50) model71
 ix  Table 4.1 Summary of LUR models published by other researchers for comparison with the Vancouver PNC LUR models................................................................................................. 76
  x  List of Figures  Figure 2.1 UFP sampling area in Metro Vancouver, showing fall locations (22 in orange), spring locations (58 in yellow plus 22 in orange, total 80), fixed sites (4), central monitoring site (1) with major road network............................................................................................. 19
 Figure 2.2 PNC sampling campaign equipment set-up, showing the CPC 3007 on the left and Kestrel 4500 on the right ........................................................................................................ 21
 Figure 2.3 Steps to creating a LUR model in Metro Vancouver (reprinted from (56)).......... 28
 Figure 2.4 500m and 1000m circular buffer around a point (sampling site) (adapted from (54)) ........................................................................................................................................ 30
 Figure 2.5 Panel A: An example of a wind rose from a sampling site, used to make wind rose shaped buffers. Panel B: The windrose shapefile, used to extract values from geographic predictor variables. The blue, full extent of the wedges represents a 100% (maximum length) buffer while the yellow smaller wedges is a 50% buffer........................................................ 32
 Figure 3.1 Within-day variation correction factors by hour of day ........................................ 45
 Figure 3.2 Mean PNC by hour at four fixed sites over entire sampling period...................... 46
 Figure 3.3 Between-day variation correction factors by day of sampling.............................. 47
 Figure 3.4 Mean PNC by day at four fixed sites over entire sampling period ....................... 49
 Figure 3.5 Histograms of corrected PNC distributions (pt/cm3). A: PNC50. B: Ln-PNC50. C: PNC10. D: PNC90..................................................................................................................... 51
 Figure 3.6 Residual plot of PNC50 (pt/cm3) compared with average NO (ppb) from the same locations at the same sampling time ....................................................................................... 53
  xi  Figure 3.7 Pollution surface for road length, PNC50 LUR model. Variables in model: length of truck route within 50m, distance to port (ln km), density of fast food restaurants within 200m. ...................................................................................................................................... 62
 Figure 3.8 Pollution surface for vehicle density, PNC50 LUR model. Variables in model: density of trucks within 25m, distance to port (ln km)........................................................... 63
 Figure 3.9 Pollution surface for road length, PNC10 LUR model. Variables in model: length of truck route within 50m, distance to port (ln km). Note: different scale than previous figures. .................................................................................................................................... 64
 Figure 3.10 Pollution surface for road length, PNC90 LUR model. Variables in model: length of truck route within 50m, distance to port (ln km), density of fast food restaurants within 200m. Note: different scale than previous figures.................................................................. 65
 Figure 3.11 Random selection of 60-minute PNC measurements compared with two-week overall PNC at the same location. A, B, C, D: four fixed site locations................................. 67
  xii  List of Abbreviations CPC – condensation particle counter DMTI – Desktop Mapping Technology Incorporated (DMTI Spatial is a Canadian publisher of spatial data) GME – geospatial modeling environment LOO – leave one out LUR – land use regression MV – Metro Vancouver NO, NO2, NOx – nitric oxide, nitrogen dioxide, nitrogen oxides PM2.5 – fine particulate matter (particles with an aerodynamic diameter <2.5µm) PNC – particle number concentration or count, in particles per cubic centimeter pt/cm3 – particles per cubic centimeter RMSE – root mean square error SD – standard deviation SOEH – School of Environmental Health (now the Occupational and Environmental Health theme within the School of Population and Public Health) SFU – Simon Fraser University UBC – University of British Columbia UFP – ultrafine particles (particles with a diameter <0.1µm)  xiii  Acknowledgements First, I wish to thank my thesis committee for their academic guidance and expertise. My supervisor, Michael Brauer, has been available, encouraging and extremely helpful over the past two years through this project. Ryan Allen and Ian McKendry, my committee members, have also provided valuable assistance to this work. I offer thanks to fellow students at the former School of Environmental Health (SOEH) and School of Population and Public Health (SPPH) for technical and peer support. In particular, Cris Barzan, Rongrong Wang and Perry Hystad provided a great deal of help with sampling equipment and computer programs. I am grateful for the work completed by my six research assistants, Rehema Ahmed, Sara Casallas, Mary Choi, Christie Cole, Brian Lee and Sarah St. John, during the field sampling campaign. This study was made possible due to funding from Health Canada. I wish to thank friends near and far for moral support and fun times while completing this degree. Finally, I am grateful to my parents for their unwavering support, always.  xiv  Chapter 1: Introduction This chapter includes an overview of ultrafine particles (UFP), including the relationship with observed health impacts, sources and factors influencing concentrations, as well as information on land-use regression (LUR) modeling. A rationale for development of a UFP LUR in Metro Vancouver as well as study objectives and hypotheses are detailed.  1.1  Outdoor air pollution An increasing base of evidence suggests chronic exposure to outdoor air pollution is  associated with a variety of adverse respiratory, cardiovascular and reproductive health impacts. (1-7) The World Health Organization estimates that 1.4% of the global burden of disease can be attributed to outdoor air pollution. (8) With a significant portion of particulate matter (PM) and other pollutants in ambient air originating from motor vehicles, the direct health impacts of traffic-related air pollution (TRAP) and other sources are substantial. (8) Air pollution epidemiologic studies compare measured or modeled air pollutant exposure levels with health outcomes in a defined population. (9) Routine government-mandated air pollution monitoring typically used for exposure estimates lacks the station density required to capture small-scale spatial variations in concentrations of air contaminants. (9) In Metro Vancouver 22 monitoring stations are irregularly located throughout the region, placed to characterize regional background air quality and typically not located in close proximity to major roadways or other key sources of air pollution. (10) The shortest distance between two monitors in Metro Vancouver is 1.67km, with most monitors located an estimated five to seven kilometers apart. (10) The design of this network does not capture small-scale spatial  1  variability in pollutant concentrations, with major differences occurring over a distance of 300m or less in other studies. (15)  1.1.1  Ultrafine particles UFP are a fraction of PM defined as having a diameter less than 0.1µm. (11) UFP are  often a direct product of combustion, though some secondary atmospheric formation occurs. (11,12) UFP are reactive and tend to coagulate (form accumulation mode particles) or disperse shortly after being emitted (on a scale of minutes to hours). (12) UFP are prevalent in urban areas and can account for over 90% of outdoor airborne particles by number. (11) PM2.5 (diameter less than 2.5µm) are regularly measured in British Columbia to evaluate observed levels relative to the BC Air Quality Objectives. (13) This measurement quantifies the mass of particles, of which UFP only constitute a small portion. UFP are one type of PM linked to adverse health outcomes and could be partially responsible for the effects observed in epidemiologic studies. (14,15) Given the importance of identifying specific components of the air pollution mixture that are harmful to humans, for the design of interventions and evidence-based air quality management strategies, understanding the spatial variability of specific outdoor air pollutants in addition to those routinely monitored is an important policyrelevant objective.  1.1.1.1  Health effects of UFP exposure UFP are of particular interest from a health perspective, as these very small particles  represent a large surface area which can contact cells in the body. UFP are defined by size and can exist in different shapes and may contain an array of contaminants, such as metals. 2  (11,12) Further, the small size of the particles make them respirable, permitting entry to the gas exchange area of the lung, as well as enabling entry into the circulatory system and penetrating cells via nonphagocytic means. (11) UFP interactions within cells generate reactive oxygen species, leading to oxidative stress (an imbalance of destructive reactive oxygen species and the body’s ability to repair damage). (16) To date, a limited number of epidemiologic studies have been conducted for UFP. A summary of toxicology and shortterm epidemiology studies on UFP and as well as expert opinions on health effects are provided below. Animal studies have shown acute UFP toxicity effects and the ability of UFP to translocate within the body. One study using polytetrafluoroethylene (PTFE) particles with an approximate diameter of 18nm induced lethality in rats after fifteen minutes of inhalation at 50 µg/m3 (the author likened concentrations of 10 000 – 40 000 pt/cm3 to be about 2 µg/m3, so 50 µg/m3 then represents 250 000 – 1 000 000 pt/cm3). (17) Rapid translocation of UFP to epithelial, interstitial and endothelial sites was observed. The same study found that aging PTFE fumes agglomerated into larger particles that reduced their toxicity. Another study again exposing rats to UFP in the form of TiO2 particles found significantly more bronchoalveolar inflammation and interstitial and lymph node transfer of UFP compared with fine PM (PM2.5) at 23 mg/m3 (equal doses of the contaminants by mass). (18) Further research on rats has shown that UFP can translocate into other organs in the body and remain at high levels up to seven days after exposure. (17,19) Animals were exposed to labelled 13C at 160 µg/m3 for six hours; afterward, the radioactive UFP were found in the liver and olfactory bulb in the brain. (19) Health effects of this translocation are unknown. The contrast in health effects from different exposure levels mentioned (lethal outcomes at 50 3  µg/m3 for fifteen minutes versus survival at 160 µg/m3 for six hours) illustrates the variability of effects possible from UFP due to differing size and chemistry between particles. It also highlights the complex task of studying mixtures of UFP, in terms of particle size and composition, found in outdoor air. A review paper (20) on human health effects of air pollution concluded that there were relatively few studies on the health impacts of UFP compared with fine and coarse PM, especially long-term epidemiologic studies. The review described some short-term studies linking UFP exposure to adverse cardiovascular outcomes, reduced pulmonary function, acute respiratory symptoms, autonomic nervous system problems, changes in blood pressure and impaired endothelial functions. Notable short-term studies found increases in UFP exposure to be correlated with increases in all-cause, cardiovascular and cardio-respiratory mortality. (21,22) A controlled exposure study in North Carolina found 24-hour standard deviation heart-beat intervals to decrease significantly while D-dimer plasma factor increased significantly (approximately 17%) with increased controlled exposure to UFP, which could suggest an increased tendency toward clotting and cardiac outcomes. (23) A study in Germany found that increased exposure to UFP was linked with increased plasma sCD40L levels, a marker for platelet activation. This can lead to coagulation and inflammation and decreased platelet count, which were hypothesized to be due to activation binding platelets to tissues. (24) Limited-term studies positively associated UFP exposure to all-cause morbidity, cardiovascular, pulmonary and hospital admission outcomes. (25) One paper described an expert European panel (12 individuals including clinicians, toxicologists and epidemiologists) that followed systematic methods to determine the likelihood that UFP exposure contributes to health outcomes. (26) The panel generally 4  agreed that short-term exposure to high UFP concentrations was likely to contribute to allcause mortality, cardiovascular hospital admissions, aggravation of symptoms in asthma patients and decreases in lung function. With long-term exposure, the experts tended to believe UFP are causally linked to all-cause mortality and cardiovascular morbidity. The panel was most confident that a pathway involving respiratory inflammation leading to a cardiac event through processes involving plaque rupture and thrombosis is most likely responsible for the outcome, when assuming UFP are causally related to cardiac events. The panel emphasized that spatially resolved UFP estimates are needed in order to conduct muchneeded long-term epidemiologic studies. While more research in this area is needed, overall, existing work suggests that UFP exposure may lead to a number of adverse health outcomes, especially cardiovascular events.  1.1.1.2  Sources of UFP and factors affecting concentrations The primary source of UFP in outdoor air is motor vehicle exhaust, with more  emissions generated from heavy-duty diesel than spark ignition engine vehicles. (27-31) Low-emission (of carbon dioxide) diesel engines tend to produce higher concentrations of UFP than conventional ones. (27,28) For gasoline-powered automobiles, UFP emissions are higher when the vehicle is older, accelerating or driving at high speeds. (28) Stationary combustion sources, such as industrial parks, may be important sources of UFP emissions, (32) while residential wood combustion is also recognized as a major source, particularly at night and during cold seasons. (33,34)  5  There is indication from work in Amsterdam that marine vessels also significantly contribute to UFP levels onshore near major shipping areas. (35) The PM distribution emitted by ships is dominated by UFP. (36,37) Restaurants, especially fast food and grill outlets that conduct a lot of frying, are also thought to contribute to UFP concentrations. For example, measurements from the California Children’s Health Study reported high UFP levels (similar to those near major freeways) at a monitoring site next to a fast-food restaurant (38) while a study of UFP concentrations in Manchester, England suggested high restaurant density as a predictor of elevated UFP concentrations. (39) Indoor measurements of UFP indicate cooking as important source (32) and it is hypothesized that ventilation of indoor-generated UFP outside could increase outdoor concentrations also. (39) Additional work has specifically suggested meat cooking as a source of UFP. (33,40,41) A study in a single home in Reston, Virginia found particularly high levels (>100 000 pt/cm3) during cooking with gas or electric stoves or during use of electric toaster ovens. (42) Levels inside restaurants were also found to be consistently high during the course of eating a meal, with levels ranging from 50 000 – 200 000 pt/cm3. (42) Concentrations of UFP are highly spatially variable. In a recent study conducted on cyclist exposures in Vancouver, UFP were heterogeneously distributed in space; areas with heavy traffic volumes had the highest concentrations. (30) Typically UFP measured by PNC exhibit much greater spatial variability within urban areas than measures of particle mass concentration. A publication from Augsburg, Germany found UFP to be 1.5 – 1.8 times higher at traffic versus background locations (15), although a study in The Netherlands found PM2.5 to be only 17 – 18% higher at traffic sites compared with background sites. (43) A study in Amsterdam reported that median PNC levels were 80% higher at traffic locations 6  compared with urban background locations, while PM2.5 at the same sites was only 9% higher at traffic sites. (44) Several factors can explain UFP spatial patterns, including proximity to major roadways and meteorologic conditions. First, proximity to motor vehicle traffic, a major source, greatly affects UFP concentration. UFP levels have been shown to decrease exponentially with distance from major roadways. (15,28,45,46) Specifically, a 100m buffer around a traffic source found highest UFP concentrations (15,28) while over 300m away UFP levels were significantly reduced and similar to background concentrations. (46) There may be some temporal variation with UFP concentration, where a peak is reached during morning rush hour, which gradually declines over the course of the day, until a smaller peak is seen in the afternoon. (31,47) One study in the Los Angeles area measured UFP on and off major roadways and concluded that coupled with typical time allocation patterns of area residents, nearly 50% of most people’s UFP exposure occurred on freeways or arterial roadways, underscoring the importance of proximity as well as time-of-day when considering UFP exposure. (48) Second, meteorologic and topographic variables have an influence. Low and high wind velocities tend to lower UFP concentrations; at low speeds more time is provided for the particles to coagulate, while at high speeds rapid atmospheric dilution occurs. (12) Wind direction largely determines particle location. Long-range transport can increase UFP levels downwind of a primary emission site. (45) Downwind and medium wind speed conditions near roadways are associated with higher UFP levels. (48) Other research shows that unstable conditions disperse UFP, reducing concentrations near breathing height. (49) Relatively low temperature and high humidity are associated with slower atmospheric dispersion, suggesting 7  UFP levels are greater in the winter than summer; further, higher temperatures in summer favour increases in accumulation-mode (PM2.5 range) particles. (15) Nocturnal surface temperature inversions pre-sunrise have also been associated with high concentrations of UFP further away from major roadways than observed during peak hours. (15,49) A study in Montreal found UFP concentrations to be inversely associated with wind speed and temperature. (50) Physical geographic factors, such as topography and altitude also dictate dispersion, where low-lying valleys collect PM and high elevations have greater atmospheric dispersion. (46,48,51) Other land-use variables play a role in UFP concentration as well: industry operation, population density (which may simply be a surrogate for more sources in an area) and agricultural activity all increase concentrations principally within 500 – 1000m radii zones of the source. (14,15)  1.1.2  Methods of modeling PNC: LUR and others Given the limited ability of routine monitoring networks to characterize spatial  variability in air pollution, high-resolution exposure estimates are required for epidemiologic and risk assessment studies related to TRAP and other sources of outdoor air pollution. A study on NO2 in Montreal noted that concentrations decreased logarithmically with increasing distance to nearest highway, with the most profound decreases occurring within 200m of the roadway. (52) Another study from New York reported that UFP concentration decreased 15 – 20% within 100m of a major roadway. (53) A study from Los Angles reported PNC decreasing exponentially with distance to a major highway, with levels 300m away indistinguishable from background concentrations and the majority of the decrease in 8  concentrations occurring within the first 100m. (12) These examples highlight the spatial variability of outdoor air pollution, especially UFP, on scales of less than 300m. As the monitoring network typically consists of stations located kilometers apart and not close to major roads, these sources of data are insufficient to reasonably reflect spatial exposure patterns of these types of pollutants. LUR models are a powerful, commonly used method for capturing this variability and estimating exposures for large study populations dispersed throughout large regions. (54,55) LUR models developed following specific procedures have traditionally been based on spatially dense measurements of TRAP indicator pollutant(s) (commonly NO2 as it can be measured easily and affordably relative to other contaminants) and readily available geographic variables (such as land use designations, road types, elevation) surrounding the measurement locations. (55,56) The relationship between these site attributes and pollutant concentrations is determined empirically using standard multiple linear regression methods. (55,56) LUR models have been applied in a number of epidemiologic studies; for example, many have investigated TRAP and childhood respiratory disease outcomes. (4,57,58) While there are a number of other modeling options, outlined below, for demonstrating spatial variability of air contaminant concentrations, LUR has been shown to offer a resource-effective, reliable approach for providing exposure estimates. (44,56,59) Three main types of alternatives are emission inventories, dispersion models and receptor methods. One major limitation of all alternatives, including LUR for UFP, is the lack of measurement data in many regions, including Vancouver. Emission inventories estimate the amount of a pollutant released into the greater environment from specific sources. Data from emissions inventories can be geocoded and 9  used to create a map of emissions, highlighting areas of expected high concentrations. (9,56) However, emissions inventories are subject to large error and are not reported as concentrations, so supplemental field measurements also need to be completed. (56) Further, these models only incorporate emissions, excluding atmospheric dispersion and transport. Proximity models could be considered a type of emission inventory where the distance from the source is used to estimate exposure level. These types of studies are limited by considering the nearest location of a type of source (such as major road) to the exclusion of other source contributors that may also be located nearby and still contribute to pollutant concentrations. (9,56) Dispersion models predict pollutant levels based on emissions and factors that may affect concentrations once released, such as meteorology and chemical and physical processes. (9,56) Dispersion models require detailed, localized input, which may be difficult to obtain and technically challenging to incorporate. Emissions data is needed for all pollutants from all sources and becomes complex when atmospheric chemical reactions occur and secondary pollutants (such as coagulation of UFP) are formed. (9,56) Eulerian Grid models are one type of dispersion model that are computationally complex and created at coarse resolutions (magnitude of kilometers). (56) Gaussian plume models are widely used to predict concentrations any distance from a source, but generally do not consider atmospheric chemistry. (9) LaGrangian or ‘puff’ models incorporate pockets of air pollution being moved by wind and are often regarded as more theoretically accurate than plume models. (56)  10  Receptor models begin with a pollutant measurement at specific sites and work backwards to determine contributions from sources, by incorporating chemical composition data and other information. (9,56) Interpolation models, which could use network monitoring data, employ geospatial methods such as kriging, incorporating spatial dependence in measured levels to create a continuous pattern of pollution in space. This predicts concentrations at unmeasured locations in a study area. This option tends to homogenize exposure levels over an area and is particularly unreliable near edges of study zones. (9,56) Since UFP are not routinely monitored or regulated by governing bodies, (60) this method is particularly limiting for UFP. Of the alternatives listed above, only emissions inventories would be practical for discerning spatial patterns of UFP in Vancouver, as dispersion models would have difficulty incorporating the reactive nature of UFP and virtually no data exist to develop receptor models. LUR then offers a high-resolution exposure estimate solution for UFP. LUR models for Metro Vancouver and southern BC have been developed for NO/NO2, PM2.5 and Black Carbon and woodsmoke (61,62), demonstrating the feasibility of this modeling approach in the region. Most LUR modeling for TRAP exposure assessments have focused on NO and NO2. (55) While exposure estimates from models based on NO2 have been linked to adverse health outcomes, these concentrations are often below those that have demonstrated negative impacts in controlled exposure studies or in epidemiologic studies of indoor gas combustion sources. (21,63,64) It is possible that in studies of within-city air pollution gradients, associations observed between NO2 and health impacts acts as a surrogate for other harmful 11  species in the traffic pollution mixture. (63,64) Correlation between UFP and NO/NO2 levels has been reported in a limited number of studies, but further comparison of measurements is needed. (47,65,66) In a Los Angeles study, moderate correlation between PNC and NO/NO2 levels were found (r = 0.08 – 0.68, depending on site) with high correlations found near traffic sites during morning rush hour (r > 0.9). (65) In Copenhagen, PNC-NOX temporal and spatial correlations were also high (r > 0.84). (47) A LUR model for UFP has been successfully developed for the city of Amsterdam and is at the time of writing the only published LUR model for UFP. (44) Measurements were successfully obtained at 46 sites outside homes, representing high traffic (N = 20) and urban background (N = 26) locations over a one-week period at each location, spread over two years. A single continuously monitoring background location provided data for temporal correction. Significant predictors of UFP concentrations included high traffic intensity (specifically traffic intensity and inverse distance to road squared), address density and industrial land use. In a study conducted in Corpus Christi, Texas, UFP at the census block level were modeled with a modified California Line Source 4 (CALINE4) Gaussian dispersion model incorporating distance-scaling functions (which weight the receptor measurement according to distance from a major source: a major road). In New York, UFP levels were predicted with an autoregressive statistical model based on a dense (N = 8225) measurement campaign incorporating traffic count and distance from two major roadways, but not other land-use variables. (53) This study used smoothing-distance effects to monitor the percent decrease in UFP from two specific roadways and presented an explanatory equation showing how UFP changed in relation to proximity to the roadways. Neither of these models are both spatially 12  highly-resolved and cover large areas, two criteria that are important in epidemiologic studies of UFP.  1.1.2.1  Incorporating meteorology in LUR models A number of studies have found that varying meteorology, especially regarding wind  speed and direction, can substantially influence PNC and other pollutants. Most LUR models use a circular buffer around a sampling point to determine the influence of various geographic predictor variables on pollutant concentrations, under the assumption that over the long term, areas surrounding a point in all directions influence the pollutant in the centre equally. The absence of an approach to incorporate meteorologic phenomena in LUR models has been repeatedly noted as a limitation. (12,15,45,50,67)  1.2  Rationale UFP represent the majority of PNC in urban air and are linked to a number of adverse  health impacts. (14,15,68) UFP are different than more conventionally measured PM mass concentrations (PM2.5 and PM10) in that their physical and chemical properties, largely due to their small size, may result in different toxicological impacts. Independently, measurement data on UFP in Vancouver adds to understanding concentrations and patterns of the pollutant in the region, due to its absence from regulatory monitoring. Comparing UFP with regulated contaminants, such as NOX, provides insight to the degree contaminants share sources. A model that can predict individual exposures of residents based on residential location (56,61) may be useful to epidemiologic and risk assessment studies evaluating health effects of UFP. More directly, a LUR model for UFP has the potential to indicate the relative importance of 13  different sources in predicting pollutant concentrations. Both a predictive model of UFP and better understanding of UFP sources are relevant when considering policy implications, such as potential monitoring, exposure limits and air quality management.  1.3  Objectives and hypotheses  The specific objectives of this project are: 1. To conduct spatial measurements of PNC at 80 varied sampling sites located throughout the Metro Vancouver region. 2. To compare one-hour median PNC values and two-week average measurements of NO and NO2 at the same sampling locations. •  Hypothesis: PNC and NO or NO2 concentrations are predicted to be highly, positively correlated, based on common sources and other studies showing correlation between TRAP components.  3. To assess the relationship between measured PNC and spatial predictors characterizing traffic flow, land use, topography and fast-food restaurant density/proximity in the study region. •  Hypothesis: PNC is expected to be higher near high traffic areas, especially where diesel trucks travel, near fast-food locations and near industrial sources.  4. To use the PNC measurements in combination with the above spatial predictors and meteorologic data to develop a LUR model for PNC in Metro Vancouver.  14  •  Hypothesis: Spatial patterns in PNC can be predicted using geographic predictors and meteorologic data. These models can explain a similar portion of variability as NOX models.  15  Chapter 2: Methods This chapter describes the methods used to develop LUR models for UFP in Vancouver, including the measurement campaign, data preparation after sample collection, creation of models and evaluation.  2.1  UFP sampling Measurement of a pollutant in a spatially dense network is the first step in the  development of a LUR model (and an informative exercise independently, especially for a contaminant such as UFP which is not routinely measured or regulated in North America). (60)  2.1.1  Pilot study Preliminary sampling was conducted at 22 sites in Metro Vancouver in October 2009  using a P-trak 8525 device (TSI®, Shoreview, MN, 2007) to record PNC values over onehour periods. These pilot sites were selected from 116 sites used during the 2003 NO and NO2 measurement period in the region by convenience from those representing a range of NOX concentrations. Following examination of the collected data, including correlation between these UFP measurements and NOX measurements made at the same locations during a 2003 (61) sampling campaign, a decision was made to proceed with a full PNC sampling campaign in conjunction with a 2009/2010 NOX campaign.  16  2.1.2  UFP sampling campaign study design Sampling for PNC was conducted at 80 sites (locations shown in Figure 2.1) across  Metro Vancouver in April and May 2010. The 22 sites sampled in October 2009 are also shown. Four additional sites across the region were monitored continuously during the sampling period to provide data for temporal correction. All details of the sampling campaign are located in Appendix B, the ultrafine particle sampling campaign manual.  2.1.2.1  Site selection The 80 sites where PNC was measured during the spring campaign were a subset of  the 116 selected by the location-allocation method during the 2003 NO and NO2 measurement period in the region. (61,69) Sampling locations were selected to characterize variability and to be geographically representative of the region. The study area was separated into four geographic quadrants, with 16 – 18 sites selected in each quadrant, including sites with low and high NO/NO2 concentrations as measured in 2003 and those near fast food restaurants. The quadrants were not based on hard boundaries, but rather general portions of the region. They were: northwest (Vancouver), southwest (Richmond, Delta), northeast (east Burnaby, Coquitlam) and southeast (Surrey). A fifth grouping represented 12 geographic outlying sites in the region, including locations in the municipalities of North Vancouver, Tsawwassen and Abbotsford, to enhance the applicability of the model for the entire Lower Mainland region. All 22 sites from 2009 pilot sampling were included in spring sampling. Repeat sampling on different days occurred at 20 sites to validate the temporal adjustments applied to the measurements during the spring sampling campaign. 17  Further, four stationary sites measured PNC levels continuously to provide data for temporal (within day and between day) correction factors. The four fixed site locations were determined by attempting to represent different parts of the study area (including west and east parts of the region, those in high-traffic neighbourhoods and low-traffic neighbourhoods). Convenience also played a role in the fixed site selection, as the instruments had to be located at homes for security purposes and since the instruments required changing alcohol wicks on a daily basis. One monitor was located in eastern Vancouver (Kingsway, Site 1), one toward southern Vancouver (Cambie Village Site 2), one in and one in Port Coquitlam (Site 3) and one on the west side of Vancouver (Kitsilano, Site 4), covering three of the quadrants described above (excluding Surrey and outlying areas). One central monitoring station (T18, South Burnaby) was selected to provide regional meteorologic data for the models due to its central location within the study region and is also shown in Figure 2.1.  18  Figure 2.1 UFP sampling area in Metro Vancouver, showing fall locations (22 in orange), spring locations (58 in yellow plus 22 in orange, total 80), fixed sites (4), central monitoring site (1) with major road network  19  2.1.2.2  Timing of sampling campaigns The main sampling campaign occurred in April – May 2010 (April 19 – May 6) to  correspond with NOX sampling. As one of the study objectives was to compare PNC and NOX values at the same locations during the same time period, the sampling campaigns were conducted at the same time of year (NOX sampling in Metro Vancouver occurred in fall 2009 and spring 2010 as well). Measurements were taken Monday to Friday, between 8:00am – 6:00pm. The four continuous monitors recorded values seven days a week, twenty-four hours a day during the entire sampling period, to be sure that data covering all mobile sampling times was collected and to provide base data to determine a two-week period PNC average.  2.1.2.3  Sampling equipment PNC were measured by Condensation Particle Counter (CPC) 3007 devices (TSI®,  Shoreview, MN, 2007) in the spring sampling campaign. P-trak 8525 (TSI®, Shoreview, MN, 2007) were used in the fall pilot. P-trak 8525 and CPC 3007 devices were co-located during spring sampling to develop a correction factor between the two types of instruments, in order to compare the fall 2009 and spring 2010 measurements. Samples at all 80 sites were one-hour in length, with values recorded at one-minute intervals. The four fixed monitors recorded fiveminute average UFP concentrations every fifteen minutes (the greatest frequency possible, given equipment operation and instrument memory limitations). Additionally, at all sites, Kestrel® 4500 Pocket Weather Trackers (Nielsen-Kellerman, Boothwyn, PA, 2009) were used to record meteorologic variables, including wind speed, wind direction, temperature, humidity and pressure, as well as elevation. These variables were  20  recorded at five-minute intervals during the one-hour measurement period, the highest frequency option given the device’s memory capacity. A GPSMAP® 60CSx or Geko 301 (Garmin®, Olathe, KS, 2009) global positioning system (GPS) device was used to record latitude, longitude and elevation at each site. Waypoints were converted to Universal Transverse Mercator (UTM) zone 10 using an online calculator. (57) For sites with repeat sampling, the mean of measured coordinates was used to identify locations.  Figure 2.2 PNC sampling campaign equipment set-up, showing the CPC 3007 on the left and Kestrel 4500 on the right  2.1.2.4  Data collected during sampling Table 2.1 outlines meteorology and field observations recorded at all 80 sampling sites  and the four continuous monitoring locations.  21  Table 2.1 Data collected during spring 2010 UFP sampling campaign  80 Mobile sites  4 Fixed sites  Measurement parameter UFP concentration (PNC, pt/cm3) Wind speed (m/s) Wind direction (°) Temperature (°C) X, Y coordinates (°) Elevation (m)  Method  Heavy-duty diesel vehicle counts Fast food restaurant counts  Visual observation of trucks and buses on road/intersection Visual observation of locations visible when walking 100m in all directions from site Completion of Community Ultrafine Particle Sampling Field Log, Appendix B CPC 3007  Field notes and photographs UFP concentration (PNC, pt/cm3)  CPC 3007  Kestrel 4500 Kestrel 4500 Kestrel 4500 GPS  GPS  N (# of sites) 80 (plus 20 repeats)  Duration  80 (plus 20 repeats) 80 (plus 20 repeats) 80 (plus 20 repeats) 80 (plus 20 repeats)  60 min sample 60 min sample 60 min sample  5 min  2 x 5 min sample  every 30 min  18 days (entire sampling period) 18 days (entire sampling period) 18 days (entire sampling period) 18 days (entire sampling period)  5 min average every 15 min 60 min  80 (plus 20 repeats) 80 (plus 20 repeats)  60 min sample  Averaging time 1 min  5 min 5 min  80 (plus 20 repeats) 80 (plus 20 repeats) 4  Wind speed (m/s)  Kestrel 4500  4  Wind direction (°)  Kestrel 4500  4  Temperature (°C)  Kestrel 4500  4  60 min  60 min  22  Four additional site-specific variables were created and outlined below, for use in the ‘best model.’ 1. Manual heavy-duty vehicle counts The mean of two, five-minute counts taken during each sample was calculated, divided by five (to obtain number of vehicles/minute) and multiplied by the number of minutes in the sample. 2. Wind direction Research assistants estimated the overall wind direction during sampling based on visual observation (such as the direction trees were blowing). It is acknowledged this is a subjective assessment. Under conditions where no wind was discernable, the upwind designation was applied. •  0 - upwind of potential major UFP source (major roads)  •  1 - downwind of potential major UFP source (major roads)  3. Fast food or other grill sites •  0 - no locations visible within 1000m of the sampling site  •  1 - one location within 100m of the sampling site  •  2 – two locations within 100m of the sampling site  •  3 – three or more locations within 100m of the sampling site  •  4 - at least one location more than 100m but less than 1000m from the sampling site (or at least one location visible from site)  4. Construction activities visible within 500m of sampling site that were potential UFP emission sources •  0 – no 23  •  2.2 2.2.1  1 – yes  Data analysis Descriptive statistics As many pollutants follow a lognormal distribution (70), the distributions of both median  and log transformed (ln) median PNC values from each sample are displayed in Figure 3.5. The mean, median, 10th and 90th percentiles were all used for LUR modeling, the latter two assessed spatial variability in background levels of UFP (PNC10) and locations with high peak levels (PNC90).  2.2.2  Temporal corrections Raw PNC measurements were adjusted for temporal variation, both within and between  days, as well as for sampler variation to create a database of comparable values for LUR modeling. Data collected for temporal adjustments and calculations are in Appendix A. Within-day variation adjustments were made using data from the four fixed locations according to methodology utilized previously (71) and the calculation is shown in equation 2.1.  Equation 2.1. Within-day correction factor (for each sampling hour)  24  CF = correction factor h = hour for correction factor n = hour of day (0 – 23) k = fixed site number (1 – 4) d = day for correction factor m = day of sampling (1 – 19)  First, the mean PNC value for each sampling hour (median value at the four sites for that hour averaged) was calculated. Next, the overall mean value (median value of each hour at the four sites averaged) was determined. The overall mean was divided by the hourly mean value to determine the correction factor for a particular hour. Whole hours (i.e. from 09:00 to 10:00) were calculated and the correction factors were applied according to the hour the majority of the sample was in. Between-day variations were calculated in a similar manner, shown in equation 2.2.  Equation 2.2 Between-day correction factor (for each sampling day)  *Symbols in equation 2.2 are the same as in equation 2.1  25  First, the mean PNC value for each sampling day (median value at four sites for that day averaged) was calculated. Next, the overall mean value (median value of each day at each of the four sites averaged) was determined. The overall mean value was divided by the daily mean value to determine the correction factor for a particular day.  2.2.3  Instrument correction Two CPC instruments were used to measure PNC at the 80 spring sites. Based on  collocation testing, one of the two instruments had consistently had lower readings than the other. One device was deemed most ‘correct’ based on its consistency with eight other instruments used during collocation testing (Appendix A), so a correction factor for the other instrument was developed. The correction factor was created by dividing raw measurements from one instrument by the other, taken at the same location at the same sampling minute, and averaged over all collocated minutes in three tests.  2.2.4  Comparisons with fall 2009 PNC measurements As described above (section 2.1) PNC measurements were also made in fall 2009 at 22 of  the 80 sites included in the spring sampling. For comparison with measured and modeled PNC in spring 2010, these measurements were first corrected for instrument differences. A correction factor was applied to convert measurements from the P-trak 8525 (fall) to the CPC 3007 (spring). This factor (1.51) was based on three collocated outdoor samples, described in Appendix A. Next, predictions from the spring model were compared to unadjusted and adjusted fall measurements. In the unadjusted assessment, the spring model was used to directly predict measured PNC from the fall; R2 and RMSE between unadjusted measurements and predictions 26  were calculated. In the adjusted assessment, temporal corrections for time of day (section 2.2.2, developed during spring sampling) were applied to the fall measurements. Two different adjustments for season were applied in separate comparisons: 1) based on the ratio of the overall mean (one-hour median values from all sites) PNC in the fall and in the spring (at the same sampling locations only) and 2) based on NO concentrations measured at all available Metro Vancouver monitoring stations (N = 20) during the two sampling periods. This second adjustment was conducted to determine the feasibility of using a regularly monitored pollutant to correct for seasonal variability. To summarize, a total of three comparisons between fall and spring values were made: a comparison of predicted and unadjusted measurements, a comparison of predicted and adjusted measurements based on the ratio of spring:fall PNC measurements and a comparison of predicted and adjusted measurements based on the ratio of spring:fall NO measurements from Metro Vancouver monitoring stations.  2.3  LUR modeling The Metro Vancouver UFP LUR models were developed using the same procedure as the  original NO/NO2 Vancouver models (61); the schematic in Figure 2.3 outlines the basic procedure in model creation. First, geographic predictor variables that may explain the variability in pollutant concentration, such as road networks, land use classifications, topographic features were calculated using geographic information systems (GIS) software for each measurement location. From this a regression model was created that predicts the level of the contaminant at any site in the region based on the geographic predictors. A pollution surface map, which is a visual representation of the regression model, can then be generated. 27  Figure 2.3 Steps to creating a LUR model in Metro Vancouver (reprinted from (56))  2.3.1  Geographic predictor variables One hundred thirty-five geographic predictor variables were calculated in preparation for  UFP LUR model development. Variable layers were developed using ArcGIS ArcInfo software, version 10.1 (ESRI®, Redlands, CA, 2011). The complete list of geographic predictors is outlined in Table 2.2. Selection of predictors was based on those used to develop previous LUR models in Vancouver, (61) but a number of additional variable types were included for UFP LUR model development, highlighted in yellow. For each variable type, data were obtained in either vector (shapefile) or raster format (from sources outlined in Section 2.3.1.3) and various ArcGIS tools were used to prepare each variable type with the appropriate buffer size or distance. For example, all data sources were 28  clipped by a shapefile created previously (56,61) to limit the geographic extent of the study area (and projected to UTM Zone 10N). Using the select by features tool, new layers featuring only specific road types or locations of intersections were created. Shapefiles were then converted to a raster and the appropriate buffer size (if applicable) was generated. Then, commands using Geospatial Modelling Environment (GME) software (Spatial Ecology, 2011) generated values of a particular variable for each of the 80 sampling locations. Commands used included: intersect points in raster (isectpntrst), which calculated the raster value (for example, population density within 2500m) at each sampling location (point file) and sum line length in polygon (sumlinelengthspolys), which was used to sum the length of lines (such as highways) that intersected wind-rose shaped polygons from each sampling location. Appendix D provides a sample of GME commands used. Distance calculations of each variable at each sampling location were completed using the ‘near’ tool in ArcGIS.  2.3.1.1  Circular buffers Most geographic predictor variables were derived for different buffer sizes, following  two methods. The first is the ‘classic’ circular buffer, shown in Figure 2.4. These circular buffer sizes ranged from 25m to 2500m (buffer sizes for each variable type are listed in Table 2.2) depending on the variable class and likeliness that fine scale changes would impact UFP levels markedly, (e.g. highways, major roads, truck routes) based on previous research. (53) In the example shown in Figure 2.4, a 500m and 1000m buffer are drawn from a radius of that distance from the sampling location. Values for the variables were generated by summing the lengths of lines within the buffer in the case of a polyline variable (such as road length), summing the area  29  within the buffer in the case of a polygon variable (such as land use), or counting the number of points within the buffer in the case of a point variable (such as bus stops).  Figure 2.4 500m and 1000m circular buffer around a point (sampling site) (adapted from (54))  2.3.1.2  Wind rose shaped buffers The second method for buffer generation was a novel approach incorporating measured  wind speed and direction at time of pollutant sampling. In this procedure, a wind rose for each sampling location was created using the software WindRose PRO (Enviroware, Concorezzo IT, 2011); an example is shown in Figure 2.5. A wind rose includes the direction the wind was blowing, the length of time it was blowing in that direction (reflected by the magnitude of the wedge) and the wind speed (reflected by the colour of the wedges in Panel A). The shape of the wind rose for each site was exported from WindRose PRO as a shapefile, where the direction and time were preserved (Panel B). Two types of wind roses were created. The first was developed from data measured at each location during the sampling campaign with the Kestrel 4500. The shapefile was assigned a 30  maximum length equal to the mean wind speed measured multiplied by the duration of the sample (minutes), to represent the maximum source area that could influence the sampling point. Buffers representing 1/8, ¼, ½, ¾ and 1 times the maximum length of the wind rose were also created, with the same angles and shapes as the maximum wind rose. Figure 2.5, Panel B shows an example of a 50% buffer in yellow and a 100% buffer in blue. Seventy-seven geographic predictor variables were derived for the meteorologic LUR models, shown in Table 2.2. Nine sites had a wind speed of <0.5 m/s measured (limit of detection of Kestrel 4500) during sampling. Since no wind rose could be produced for these sites, a circular buffer was used, assuming all directions affected the transport of UFP equally. In these instances, the circular buffer radius selected as half of the limit of measurement for the Kestrel 4500 was multiplied by the number of seconds in the sample. This value (0.25 m/s x ~3600 s = 900 m) represented the maximum distance the pollutant could have traveled in the sampling time. Circular buffers of 1/8, ¼, ½ and ¾ of this maximum distance were offered to models. The second type of wind rose buffer used data from a central (T18 South Burnaby, Figure 2.2) Metro Vancouver monitoring site to create a regional rose that was applied to all sites. This wind rose was based on wind speed and direction data measured from 8:00am to 6:00pm for all days during the sampling campaign. These measurements were collected at 10m above ground, in contrast to the Kestrel 4500, which measured at about 1.5m above the ground, the same height as the CPC 3007 PNC sampler.  31  A  B  Figure 2.5 Panel A: An example of a wind rose from a sampling site, used to make wind rose shaped buffers. Panel B: The windrose shapefile, used to extract values from geographic predictor variables. The blue, full extent of the wedges represents a 100% (maximum length) buffer while the yellow smaller wedges is a 50% buffer.  2.3.1.3  Sources of geographic predictor variables Road length variables were obtained from the CanMap 2010 Streetfile (Desktop Mapping  Technologies (DMTI) Spatial Inc). Highways (RD1) were defined as carto classifications 1 and 2 (freeways and primary highways, respectively) while major roads (RD2) were carto classifications 3 and 4, (secondary highways and major roads, respectively). (72) The resolution of all road length variables was 10m. Intersections of highways with major roads or other highways were determined from the same data source using the script “Calculate Fnode Tnode 2.0” in ArcGIS 9.2 to generate nodes. Each node had a valence assigned to it and those with a  32  value ≥3 were defined as intersections. The point density tool was used to sum the number of intersections in a given buffer. Truck route data was compiled manually by Dr. Sarah Henderson (54), using route data collected from municipalities and the CanMap 2003 Streetfile (DMTI Spatial Inc). Truck routes from the municipalities of Vancouver, West Vancouver, Burnaby, New Westminster, Coquitlam, Port Coquitlam, Surrey, Langley, Delta and White Rock were obtained from open source data on city websites or contact with planning departments. Other cities advise trucks to follow major arterial routes, RD1 and RD2, including North Vancouver, Richmond, Port Moody, Maple Ridge, Pitt Meadows and Abbotsford. Traffic density variables were obtained from the Translink (regional transportation authority for Metro Vancouver) EMME/2 traffic model for 2008 provided by Clark Lim and Ken Tseng. EMME/2 is transportation planning software that estimates different types of traffic densities, speeds and times on road segments. Vehicle densities were based on morning peak hour volumes, from 7:30 – 8:30am. Data were received in a polyline file and auto volume was used for vehicle density and light truck and heavy truck volumes combined for truck density. The feature to point tool in ArcGIS was used to convert the file to points at a resolution of 1m and point density tool used to calculate the density of traffic volume points and associated values in the appropriate buffer size. Bus stop location information was acquired as a point shapefile from Translink Transit GIS data, 2009. Data were accessed from Abascus on the UBC Library website. A search of “BC bus stops” yielded the file. Land use classifications were based on data from the Metro Vancouver regional office (2006), as there were inconsistencies with land use data classifications between different years 33  from DMTI Spatial Inc. Data were available at a resolution of 10m. (12,17) Seventeen classifications were aggregated into five categories as follows: Industrial = Industrial-Extractive + Industrial + Transportation, Communication Utilities; Residential = Single family + Rural + Townhouses + Low-rise apartments + High-rise apartments; Open = Agricultural + Harvesting and research + Open and undeveloped + Regional watershed + Recreation and protected natural areas + Lakes and water bodies; Government = Institutional; Commercial = Commercial + Commercial – Residential/Mixed. The Metro Vancouver data did not include Abbotsford (and several sampling sites were located there) so DMTI Spatial Inc. land use data from 2006 was used outside of the Greater Vancouver Regional District. The normalized difference vegetation index (NDVI), measuring the percentage of vegetation cover, was obtained from NASA Global Inventory Modeling and Mapping Studies (NASA GIMMS) group (2002). Dr. Jason Su provided the layer at a 25m resolution. Population and dwelling density information was taken from Statistics Canada, Census of Canada using dissemination area level (2006). The population or number of dwellings was assigned to the centroid of its polygon and the calculate density (kernel) feature in ArcGIS ran at the buffers listed in Table 2.2. The resolution of these files was 5m. Location variables were derived from their respective variable category (i.e. distance to nearest highway was calculated from the CanMap 2010 Streetfile). Port, airport and shoreline GIS data also came from CanMap 2010 Streetfile. (58) Preparation of a restaurant location shapefile was completed manually. First, fast food chain restaurant locations that conduct cooking, grilling or deep-frying were located and their coordinates obtained using Google Earth. The selection of these restaurants was based on the assumption that these cooking methods could produce significant UFP emissions. These chains 34  included A&W, Burger King, Church’s Chicken, Dairy Queen, Kentucky Fried Chicken, McDonald’s and Wendy’s. Next, we searched the study area in Google Earth for restaurants that contained the word “grill” “Indian” or “Chinese.” For each potential restaurant that was identified, we assessed the name and description to determine if it was actively cooking food and included only those that were in the fast food restaurant database. For example, a location called “Benny’s Grill” would be included in the final restaurant list whereas “Golden Dragon Cooking Supplies” would not. This approach assumed that Indian and Chinese restaurants (which are common in the study areas) employed cooking methods including a substantial amount of frying, which may be a source of UFP emissions. These point locations (a list of X,Y coordinates) were then imported into ArcGIS, converted to a point shapefile and used in further geographic analysis. In addition to the geographic predictor variables considered for LUR modeling outlined in Table 2.2, additional variables combining traffic density and road length measures (such as [traffic density x 1/(distance_to_highway)] or [traffic_density x 1/(distance_to_highway)^2] were explored, as the LUR model developed for Amsterdam used similar compound variables. (44) These variables did not increase the R2 of the models in preliminary analyses and were excluded from further model development in favour of road length, traffic density and distance to features used in the LUR models developed previously in Vancouver (56,61), which also included ln-distances to features. The measured mean temperature during sampling was a variable offered to the road length, measured wind rose buffers (PNC50) model (model 9 in Table 3.9). Temperature was included given its potential to influence PNC levels. After preliminary analysis, temperature was not found to be a significant predictor of PNC. 35  2.3.  Best model Four additional variables were created, based on counts and field notes collected during  the UFP sampling campaign, to add to a ‘best model’ that incorporated site-specific field observations (Appendix E). This specific information, which may help describe determinants of PNC, would not be useful for estimating concentrations at unmonitored locations. The four additional variables were manual heavy-duty vehicle counts, wind direction, fast food sites and construction activities, detailed in section 2.1.2.4.  36  Table 2.2 Potential variables for UFP LUR modeling (adapted from (56))  Variable category (N)  Description  Sub-category  Road length (27)  Total length (km) of 3 road types  RD1 (highways) RD2 (major roadways) TRK (truck routes) AD (automobiles) TD (trucks)  Vehicle density (18)  Density (vehicles/ha) of 2 types during morning peak Bus stop density Density of bus (15) stops Intersection density (15)  Land use (20)  Vegetation index (1) Population density (6)  Density of intersections of highways or major roads Total area (ha) of 5 land-use designations  Normalized difference index Density of population (people/ha)  BUS  Circular buffer radii (m)  Wind rose Source buffer radii (portion of max. length) 25, 50, 75, 100, 0.125, 0.25, CanMap Streetfile, 200, 300, 500, 0.5, 0.75, 1.0 DMTI Spatial Inc. 750, 1000 (2010) 25, 50, 75, 100, Not used Translink EMME 200, 300, 500, Model (2008) 750, 1000 25, 50, 75, 100, 200, 300, 500, 750, 1000 25, 50, 75, 100, 200, 300, 500, 750, 1000  0.125, 0.25, 0.5, 0.75, 1.0  Translink Transit GIS data, 2009  0.125, 0.25, 0.5, 0.75, 1.0  CanMap Streetfile, DMTI Spatial Inc. (2010)  RES (residential) COM (commercial) GOV (government) IND (industrial) OPN (open area) VEG  300, 400, 500, 750  0.125, 0.25, 0.5, 0.75, 1.0  Greater Vancouver Regional District, 2001  N/A  N/A  POP (population)  750, 1000, 1250, 1500, 2000, 2500  Not used  NASA Global Inventory Modeling and Mapping Studies (2002) Statistics Canada, Census of Canada (2006).  INTER  37  Table 2.2 Potential variables for UFP LUR modeling (adapted from (56))  Variable category (N)  Description  Sub-category  Dwelling density (6) Location (19)  Density of dwellings (units/ha) Variables describing attributes of site location (km from site)  DWEL (dwelling)  Fast food Distance to restaurants/grills nearest restaurant (8) (km), restaurant density Temperature Measured (°C)  Circular buffer radii (m) 750, 1000, 1250, 1500, 2000, 2500 N/A  ELEV (elevation) X (latitude) Y (longitude) DIST_HWY DIST_RD DIST_TRK DIST_AIR DIST_SHOR DIST_BUS DIST_INTER DIST_PORT (distance to highway, major road, truck route, airport, shoreline, bus stop, intersection, port) FF (density of 50*, 75, 100, restaurants in buffer) 200, 300, 500, FFD (distance to 750, 1000 nearest restaurant)  Wind rose Source buffer radii (portion of max. length) Statistics Canada, Not used Census of Canada (2006). N/A CanMap Streetfile, DMTI Spatial Inc. (2010)  0.125, 0.25, 0.5, 0.75, 1.0  Developed for this project (2011)  N/A  *the nearest fast food restaurant/grill to any sites was 40m, so the 25m buffer was not used for this variable class  38  The highlighted portions of Table 2.2 indicate variables that were developed for UFP LUR modeling and were not considered in earlier (61) Vancouver LUR models of NOX and PM2.5. Buffers less than 100m are also novel. In total, 135 variables were used during LUR model development with circular buffers. The number of variables offered to the wind rose buffer models was 77.  2.3.2  LUR model building procedure The LUR model development followed the same procedure as described previously and  used in the development of previous Vancouver LUR models. (56,61) The steps in creating a LUR were: 1. Rank all variables by absolute correlation with measured pollutant. 2. Identify the highest-ranking variable in each group (variables of the same type, but with different buffer radii). 3. Exclude variables highly correlated (r>0.6) with the highest-ranking variable in each category. 4. Enter all remaining variables in a backward stepwise regression. 5. Remove variables that are inconsistent with a priori hypotheses (e.g. a negative coefficient for a traffic-related variable, as PNC is expected to increase with increasing length of roads within a buffer). 6. Remove variables that are insignificant (p≥0.05) 7. Repeat steps 4, 5 and 6 to converge on a final model. Remove any variable that contributes less than 1% to the model R2 for a concise final model.  39  The final product is a multiple linear regression model of the form: Pollutant Concentration = α + β1X1 + β2X2 + β3X3 - β4X4 + e where α is the constant intercept, β1 is the coefficient for variable X1 (e.g. length of major roads in a 100m radius buffer) and e is the random error term. Potential predictors (Table 2.2) were systematically evaluated by calculating correlations between each variable and the UFP measurement. PNC measurements used as outcome variables during modeling included PNC50, ln-PNC50, PNC10 and PNC90. One value for each site from the one-hour sample recorded at one-minute intervals was used to create the appropriate distribution. This method sought to achieve a final, parsimonious model with minimal colinearity in predictors. All model building was conducted using R software version 2.10.1 (R Development Core Team, Vienna AT, 2009). Code examples are located in Appendix D. The two distinct types of models were developed: the road length model, which initially included all potential predictors except vehicle density variables, and the vehicle density model, which initially included all potential predictors except road length variables. As the models were created, it was apparent that the models built using road length variables and the median PNC values tended to explain more of the variation in measurements than others. As such, models that were developed subsequently with wind rose shaped buffers used an outcome of PNC50 and road length variables (excluding vehicle density variables). Table 2.3 below displays the models that were created and are shown in Chapter 3. Spreadsheets containing PNC measurements along with extracted geographic variable values at each site, which were used to build the LUR models are contained in the appendices. Appendix F includes data used to build the models based on circular buffers and included variables incorporated into the best model. Appendix G contains data used to build the model 40  with wind rose shaped buffers, based on measured wind speed and direction. Appendix H contains data used to build the model with wind rose shaped buffers using data from a central monitoring location.  Table 2.3 Summary of UFP LUR models created for Metro Vancouver  Model PNC50 (pt/cm3) ln-PNC50 (ln pt/cm3) PNC10 (pt/cm3) PNC90 (pt/cm3) PNC50 (pt/cm3), measured wind rose buffers PNC50 (pt/cm3), central station wind rose buffers PNC50 (pt/cm3), best model 2.3.3  Road Length X X X X X X X  Vehicle Density X X X X  Regression maps Following model development, the weighted sum tool in ArcGIS was used to create a  pollution surface for each model by multiplying variable rasters by variable coefficients and adding a fixed value intercept raster. This yielded a map of the study area with predicted concentrations of the pollutant based on the model to be used for visual identification of potential ‘hotspots’ and other spatial patterns. Maps were truncated at the maximum PNC measured.  2.4 2.4.1  Evaluation of measurements and models Sixty-minute measurement evaluation In order to assess the representativeness of 60-minute UFP sample compared to longer-  term averages, data from the four fixed sites were evaluated. For each of the four fixed sites, the overall mean (mean, median value of all hours) during the two-week sampling campaign was calculated. Then, 14 random one-hour mean PNC measures from each site were selected and 41  compared to the overall mean, to evaluate the representativeness of (the equivalent) one-hour sample measurements compared to the overall two-week period mean. One sample was selected for each sampling day (Monday – Friday) during sampling hours (8:00am – 6:00pm), to match the mobile sampling campaign protocol. All three corrections were applied to the median PNC measurement from each of the 100 samples (80 sites plus 20 repeats), according the time and day of the sample. The effectiveness of the adjustment factors was determined by comparing the absolute difference of the median PNC values as well as the root mean square error (RMSE) for raw and corrected data for the 20 sites that had repeat measurements taken.  2.4.2  Model evaluation Initial model performance was assessed on the adjusted R2 values of the models.  Subsequent evaluations of model stability were based on the R2 and root mean square error (RMSE) from a “leave one out” (LOO) cross-validation approach, in which models were regenerated by systematically leaving out one of the 80 sites and then using the remaining 79 sites to predict the concentration at the site that was left out. (56,61) R code for this evaluation is in Appendix D. The R2 of the simple linear model between predictions from the LOO models and actual measurements and then the RMSE from these comparisons was calculated. A higher LOO-R2 value and lower RMSE value (and standard deviation) were associated with more stable models. This LOO evaluation was completed on all eleven LUR models. Finally, the model with the highest R2 and lowest RMSE was compared to measured PNC at 20 sites where duplicate measurements were collected on separate sampling days and at the four fixed sites. For these four sites, comparisons were made for the measured PNC levels 42  both during the specific sampling hours (8:00am to 6:00pm, Monday to Friday) and during the entire study period (twenty-four hours a day, seven days a week). These comparisons were evaluated using RMSE.  43  Chapter 3: Results This chapter will outline results of the UFP sampling campaign as well as LUR models and evaluations.  3.1 3.1.1  Sampling outcomes Instrument correction factor Two CPC 3007 instruments were used to measure PNC at the 80 locations around Metro  Vancouver. Based on an outdoor collocation pretest and a sampling collocation measurement, an instrument correction factor of 1.29 (standard deviation = 0.19) was developed in order to make measurements from the two CPC devices comparable. (Correction factors of the individual collocation examples were 1.42 and 1.16.) Data used to create this correction are located in Appendix A.  3.1.2  Temporal correction factors Temporal correction factors applied to the raw data to account for variability in time of  day and day of sampling were normalized to the overall mean PNC from the hourly measurements across the four fixed sites. The within-day variation corrections, shown in Table 3.1, range from 0.836 – 1.00, with a mean of 0.93 and standard deviation of 0.051. The trend of hourly adjustment factors throughout the day is shown in Figure 3.1.  44  Table 3.1 Within-day variation correction factors applied to raw data  Hour of day 08:00 09:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00 17:00 18:00  Correction factor 0.95 0.96 0.92 0.96 0.87 0.84 0.90 0.92 1.00 1.00 0.96  Figure 3.1 Within-day variation correction factors by hour of day  The range of within-day correction factors was quite small, indicating there was little temporal variation in PNC from 8:00am to 6:00pm. From examining Figure 3.1, there appeared  45  to be a slight increasing trend in adjustment factors at later hours of the day, meaning that measurements made later in the day more closely reflect the overall mean PNC (twenty-four hours per day, seven days per week) and that measurements made earlier in the day are higher, which could be associated with morning rush hour. Figure 3.2 shows the mean hourly PNC values at the four fixed sites.  20000
  Mean
PNC
(pt/cm3)
  18000
 16000
 14000
 12000
 10000
 8000
 6000
 4000
 2000
 0
 1
 2
 3
 4
 5
 6
 7
 8
 9
 10
 11
 12
 13
 14
 15
16
 17
18
 19
 20
 21
 22
 23
 24
 Hour
of
sampling
(24h)
 Site
1
  Site
2
  Site
3
  Site
4
  Figure 3.2 Mean PNC by hour at four fixed sites over entire sampling period  Adjustment factors to correct for the day of sample (between-day variation) are presented in Table 3.2. The range of correction factors applied was 0.613 – 3.82 with a mean of 1.52 and standard deviation of 1.03. The trend of daily adjustment factors throughout the sampling period is shown in Figure 3.2.  46  Table 3.2 Between-day variation correction factors applied to raw data  Day Monday, 04/19/10 Tuesday, 04/20/10 Wednesday, 04/21/10 Thursday, 04/22/10 Friday, 04/23/10 Monday, 04/26/10 Tuesday, 04/27/10 Wednesday, 04/28/10 Thursday, 04/29/10 Friday, 04/30/10 Monday, 05/03/10 Tuesday, 05/04/10 Wednesday, 05/05/10 Thursday, 05/06/10  Correction factor 0.91 0.68 0.75 0.61 0.96 0.91 1.1 1.3 1.4 0.90 1.9 2.9 3.2 3.8  Figure 3.3 Between-day variation correction factors by day of sampling  47  From examining Figure 3.3, there is an upward trend of daily correction factors as sampling progressed. The last three days in particular (adjustments 2.9 – 3.8) have markedly higher correction factors than the first eleven days, which increases the slope of the line. Separate between-day correction factors were calculated based on only the two sites that were operated during the last three days of the campaign (the other two instruments had to be returned from loan and were no longer used). These new correction factors displayed a similar pattern as above, with correction factors on the last three days 50% higher than the campaign mean correction factor (overall mean of 1.22 versus last three days mean of 1.84), while the correction factors based on four sites were about 100% higher the last three days (overall mean of 1.52 versus last three days mean of 3.3). Thus, the trend of increasing correction factors on the last three days of the campaign (analogous to lower concentrations) appears to reflect atmospheric conditions and is not simply due to missing measurements. Figure 3.4 shows the trend in mean PNC value by day at the four sites over the sampling period. It is clear that there is a general downward trend in PNC toward the end of sampling, which contributes to the elevated correction factors during the last three days of the campaign. Unique samples were collected at only four sites during the last three days of the campaign (the rest of the measurements on these days were repeats), so the number of sites included in LUR modeling that had large correction factors applied to them were minimal.  48  20000
  Mean
PNC
(pt/cm3)
  18000
 16000
 14000
 12000
 10000
 8000
 6000
 4000
 2000
 0
 1
 2
 3
 4
 5
 6
 7
 8
  9
 10
 11
 12
 13
 14
 15
 16
 17
 18
  Day
of
sampling
 Site
1
  Site
2
  Site
3
  Site
4
  Figure 3.4 Mean PNC by day at four fixed sites over entire sampling period  All three corrections (instrument, within-day, between-day) were applied to the median PNC measurement from each of the 80 samples. Hereafter, all PNC measurements described are corrected, unless noted otherwise. For example, site 48, measured on April 19, 2010 from 9:40am to 10:39am had a raw median PNC count from the sample of 55 139 pt/cm3. An instrument factor of 1.29, a within-day adjustment of 0.92 (Table 3.1) and a between-day adjustment of 0.91 (Table 3.2) were applied to generate an adjusted median PNC concentration of 55139 x (1.29) x (0.92) x (0.91) = 58 871 pt/cm3. The root mean square error (RMSE) of the absolute differences between repeat measurements at 20 sites, before and after application of adjustment factors were 10 800 and 13 700 pt/cm3, respectively, an increase of 27%. However, the absolute median difference from uncorrected to corrected measurements decreased 11% (from 7500 to 6700 pt/cm3). Based on  49  these evaluations, it is unclear whether the corrections reduced temporal variability, but does suggest that factors other than time account for differences in PNC at the same locations. Notably, the corrections applied to the raw measurements were mostly relatively small, with the range of the three correction types ranging from 0.61 to 3.8.  3.1.3  Summary statistics Histograms showing the distribution of adjusted measurements from the 80 sites sampled  are displayed in Figure 3.5.  50  A  B  C  D  Figure 3.5 Histograms of corrected PNC distributions (pt/cm3). A: PNC50. B: Ln-PNC50. C: PNC10. D: PNC90.  The distribution of PNC50 appears lognormal in panel A, while the ln-transformed PNC50 dataset in panel B appears to be approximately normally distributed. PNC10 and PNC90 in panel C and D both are skewed to the right, appearing lognormally distributed. Despite ln-PNC50 being the mostly clearly normally distributed dataset, all were used in LUR modeling as it became 51  evident that using PNC50 as the outcome variable yielded models with systematically higher R2 values than using ln-PNC50, (Table 3.9). Summary statistics from the 80 sites sampled are reported as the median, 10th and 90th percentile of the 60 points recorded in a 60-minute sample, shown in Table 3.3.  Table 3.3 Summary statistics for corrected one-hour PNC at 80 sites: median, 10th and 90th percentile  Statistic Minimum 10th percentile First quartile Mean Median Third quartile 90th percentile Maximum Standard deviation 3.1.4  PNC50 (pt/cm3) 1500 4200 7900 18 200 13 600 21 900 30 400 105 000 15 900  Ln-PNC50 (ln-pt/cm3) 7.30 8.47 8.97 9.50 9.54 9.99 10.53 11.6 0.78  PNC10 (pt/cm3) 1100 3900 6100 12 600 10 000 15 400 25 400 72 300 10 100  PNC90 (pt/cm3) 2300 3700 11 900 30 200 20 447 41 500 61 600 169 000 28 100  Comparing PNC with nitrogen oxides Measurements of NOX are compared with PNC in Table 3.4 and Figure 3.6.  Table 3.4 Correlation coefficients between PNC and nitrogen oxides  Correlated pollutants PNC50 and NO PNC50 and NO2 PNC50 and NOX ln-PNC50 and ln-NO ln-PNC50 and ln-NO2 ln-PNC50 and ln-NOX  Pearson correlation coefficient 0.65 0.64 0.70 0.63 0.61 0.60  Spearman correlation coefficient 0.56 0.57 0.58 0.56 0.57 0.58  52  Figure 3.6 Residual plot of PNC50 (pt/cm3) compared with average NO (ppb) from the same locations at the same sampling time  PNC and NOX were highly correlated (r = 0.61 – 0.70), suggesting the pollutants have some of the same sources. The residual plots between regression of PNC50 with NO2 and NOX displayed similar patterns to that of NO (Figure 3.6). Residuals were evenly distributed around zero, with three outlier locations evident (shown in red). Table 3.5 outlines some known geographic features of these sites that might explain some source differences between PNC and NOX at those locations. Notably, all three sites had very high truck counts compared with other sites and were on or near truck routes. The site with the highest residual value, E7, was also very close to the port. This indicates that heavy-duty truck traffic and possibly port proximity may be more specific predictors of PNC than they are of NOX. However, site 35 had a negative residual 53  and high truck counts, so there be other unaccounted for factors explaining the difference between PNC and NOX.  Table 3.5 Examination of three sites with high residual values when comparing NOX and UFP levels  Site E7  Site description Clark Drive and Pender Avenue, Vancouver  Residual (+/-) +  48  Knight Street and East 36 Avenue, Vancouver Copley Street and East 15th Avenue, Vancouver  +  35  3.1.5  -  Site features -high truck count (378 in one hour) -on truck route -bus stop located nearby (within 20m) -high area of industrial land -fast food restaurant outlet nearby -very close to port (30m) -high truck count (366 in one hour) -on truck route -high truck count (292 in one hour) -close to truck route -close to bus stop -close to intersection  Meteorologic data Table 3.6 includes summary statistics for measured meteorology variables at the 80 sites  sampled and those measured at the central Metro Vancouver (T18, Burnaby South) monitoring station data during the same time period.  Table 3.6 Summary statistics for site-specific and central station measured meteorology variables  Statistic Minimum First quartile Mean Median Third quartile Maximum Standard deviation  Wind speed (m/s) 80 sites Central (T18) 0 1.1 0.5 2.1 1.8 3.0 1.2 2.9 2.2 3.7 16.3 6.6 2.3 1.2  Temperature (°C) 80 sites Central (T18) 8 5.3 11.1 9.3 12.9 11.3 12.3 10.8 14.2 12.5 22.7 21.6 2.8 3.0 54  The wind speed measured at the Burnaby South monitoring station was systematically higher (the mean was 1.7 and median 2.4 times greater) than the measured wind speeds from the 80 sampled sites. The central monitoring station had consistently lower temperatures than the 80 sites. These differences in meteorologic variable trends can likely be attributed to the height of the monitors. The central monitor, at 10m above ground, captured a more generalized pattern in the region whereas the Kestrel 4500® at each sampling location made recordings from 1.5m, the height the CPC 3007 devices were measuring (and approximate human breathing height). The range of wedge buffer lengths used in the wind rose buffer methodology models is presented in Table 3.7.  Table 3.7 Range of wedge buffer lengths applied in measured meteorologic LUR model  Buffer size (portion of maximum wind rose length) 0.125 0.25 0.50 0.75 1.0  Minimum (m) 113 225 450 675 900  Maximum (m) 3767 7534 15067 22601 30134  Mean (m) 776 1553 3106 4659 6211  Median (m) 536 1073 2145 3218 4290  Standard deviation (m) 772 1545 3090 4635 6179  Notably, the total range of buffer sizes shown in Table 3.7 is from 113 – 30 134m, which represents a much larger maximum range than used in circular buffer variables (where 2500 m was the largest buffer size).  55  3.1.6  Univariate correlations with PNC50 and geographic variables Selected univariate correlations between geographic predictor variables and PNC50 are  shown in Table 3.8.  Table 3.8 Selected univariate correlations with PNC50  Variable ln-Distance to nearest intersection Mean temperature Mean wind speed Length of highways within 1000m Population within 2500m Number of intersections within 25m Number of fast food sites within 200m Area of industrial land within 750m Number of bus stops within 100m Length of major roads within 100m Truck density within 25m Automobile density within 25m Truck density within 50m  Pearson correlation coefficient -0.049 0.03 0.12 0.12 0.24 0.27 0.36 0.40 0.45 0.49 0.58 0.60 0.63  Univariate correlations between predictor variables and PNC50 measurements showed that the highest overall correlation was for truck routes within 50m (r = 0.63), followed by density of automobiles within 25m and the density of trucks within 25m. The highest distancerelated correlation (ln-distance to the nearest intersection of major roads, r = -0.049) was quite low suggesting that simple road proximity measures are not very useful in the prediction of UFP spatial variability. The highest land use correlation was for industrial land within 750m at r = 0.40. With fast food and grill restaurants in a 250 m buffer, the correlation with PNC50 was r = 0.36.  56  Mean sample temperature and wind speed, measured at the time of sampling, were weakly positively correlated with PNC50 suggesting that these factors were not very important in predicting PNC.  3.2  PNC LUR models A total of eleven LUR models for UFP were developed and are summarized in Table 3.9.  Models described as “road length” had all qualifying variables initially offered into the model, except the vehicle density variables. Models described as “vehicle density” had all qualifying variables initially offered into the model, except the road length variables. The wind rose buffer based models (model 9 and 10 below) were only based on road length variables, as preliminary models had shown higher R2 values for road length versus vehicle density models. A model for PNC50 was also created using all possible variables (road length and vehicle density) and this final model was identical to the road length model. Road length models consistently yielded higher R2 values than their vehicle density counterparts by 0.03 to 0.10. PNC50 models predicted 5 – 8% more variability than ln-PNC50 models. PNC90 models for both road length and vehicle density explained 1 – 5% more variability than PNC50 models for the same model type, which in turn explained 9% more variability than the PNC10 models. Models with wind rose shaped buffers had R2 values 0.14 and 0.19 lower than the PNC50 road length model. The models with the highest R2, 0.53, were a road length model of PNC90 and a model of PNC50 including site observations. The model with the highest R2 for median PNC levels excluding site observations was the road length model (R2 = 0.48).  57  Table 3.9 LUR models developed for PNC in Metro Vancouver  Model type 1. Road length, PNC50 2. Vehicle density, PNC50 3. Road length, ln-PNC50  4. Vehicle density, ln-PNC50 5. Road length, PNC90 6. Vehicle density, PNC90  Variables  Coefficients (SE) (pt/cm3 or ln-pt/cm3) • Truck route length (50m buffer) 150 800 (25 200) • Distance to port (ln km) -3000 (900) • Fast food site density (200m buffer) 3700 (1700) • Intercept 17 700 (2500) • Truck density (25m buffer) 88 (14) • Distance to port (ln km) -3600 (990) • Intercept 20 600 (2500) • Truck route length (50m buffer) 6.6 (1.3) • Distance to fast food site (ln km) -0.29 (0.088) • Distance to port (km) -0.041 (0.014) • Latitude 3.4 x 10-5(1.3 x 10-5) • Area government land (300m buffer) -0.085 • Intercept -7.3 (6.1) • Distance to bus stop (ln km) -0.16 (0.060) • Truck density (25m buffer) 2.2 x 10-3(8.6 x 10-4) • Distance to intersection (ln km) -0.14 (0.058) • Intercept 8.7 (0.16) • Truck route length (50m buffer) 292 500 (40900) • Distance to port (ln km) -4600 (1600) • Fast food site density (100m buffer) 17 700 (6600) • Intercept 27 400 (4100) • Vehicle density (25m buffer) 8.0 (1) • Distance to port (ln km) -5700 (1700) • Intercept 30 800 (4400)  P values <0.001 0.002 0.03 <0.001 <0.001 <0.001 <0.001 0.005 0.009 0.03 0.004 0.009 0.02 <0.001 0.004 0.009 <0.001 0.001  Partial R2 Model R2 0.36 0.10 0.05 Model: 0.48 0.32 0.11 Model: 0.42 0.19 0.08 0.06 0.05 0.03 Model: 0.40 0.16 0.13 0.11 Model: 0.37 0.43 0.07 0.06 Model: 0.53 0.37 0.08 Model: 0.43  LOO R2 RMSE (SD) 0.32 12 800 (23 100) 0.29 13 100 (22 900) 0.30 0.64 (0.78)  0.32 0.63 (0.78) 0.41 20 800 (34 700) 0.31 22 400 (38 800)  58  Table 3.9 LUR models developed for PNC in Metro Vancouver  Model type  Variables  7. Road length, PNC10  • • • • • • • • • • • • • • • • • • •  8. Vehicle density, PNC10 9. Road length, measured wind rose buffers (PNC50) 10. Road length, central site wind rose buffers (PNC50) 11. Best model (PNC50)  P values  Truck route length (50m buffer) Distance to port (ln km) Intercept Truck density (25m buffer) Distance to port (ln km) Intercept Distance to intersection (ln km) Distance to port (ln km) Intercept  Coefficients (SE) (pt/cm3 or ln-pt/cm3) 97 700 (17700) -2000 (700) 13 600 (1809) 57 (10) -2400 (700) 14 900 (1800) -4900 (1000) -3900 (1100) 14 300 (3600)  Distance to port (ln km) Area residential land (100% buffer) Distance to major road (km) Distance to highway (ln km) Intercept Truck route, 50m Distance to port (ln km) Truck count Fast food site density, 200m Intercept  -3400 (1100) -0.49 (0.19) -16 200 (5000) -2500 (1100) 36 300 (42300) 103 100 (28100) -3100 (900) 38 (12) 3400 (1600) 16 800 (2400)  0.003 0.01 0.002 0.02  <0.001 0.005 <0.001 0.001 <0.001 <0.001  <0.001 <0.001 0.002 0.04  Partial R2 Model R2 0.31 0.08 Model: 0.37 0.24 0.11 Model: 0.33 0.23 0.13 Model: 0.34  LOO R2 RMSE (SD) 0.22  0.12 0.10 0.09 0.09 Model: 0.29 0.20 0.19 0.14 0.05 Model: 0.53  0.14  9100 (15600) 0.19 9300 (16600) 0.20 14 100 (24300)  14 700 (24900) 0.38 12 300 (21300)  SE = standard error; SD = standard deviation Model R2 are adjusted R2 values LOO = Leave one out cross validation, R2 between predicted and measured values RMSE = root mean square error 59  3.2.1  Sensitivity analyses Additional univariate correlations between distance to the nearest fast food restaurant,  fast food restaurant density within 200m, distance to the nearest port and traffic-related variables were also conducted, after the first variables appeared in a number of final models. Table 3.10 shows the results.  Table 3.10 Selected univariate correlations between fast food and port variables and traffic-related variables  Variable 1  Variable 2  Distance to nearest restaurant (km) Number of restaurants in 200m Number of restaurants in 200m Number of restaurants in 200m Distance to port (km) Distance to port (km) Distance to port (km) Distance to port (km)  Distance to nearest intersection (km) Length truck route within 50m Length of major roads within 100m Number of bus stops in 100m Length truck route within 50m Density of trucks within 50m Length of major roads within 100m Distance to nearest bus stop (km)  Pearson correlation coefficient 0.37 0.28 0.39 0.55 -0.19 -0.11 0.27 0.68  There is a strong correlation (r = 0.68) between distance to the port and distance to the nearest bus stop and a fairly strong correlation (r = 0.55) between number of restaurants within 200m and number of bus stops within 100m. All other comparisons show weak to moderate correlations, suggesting that restaurant and port variables independently contribute to observed UFP levels, or that other factors not considered are responsible for the PNC results obtained. As a sensitivity analysis, a model was built with 79 sites, excluding the site with the highest median PNC (as it was an outlier). In this model, variables were the same as in the PNC50 road length model and coefficients were comparable: length of truck routes in 50m (coefficient 105 000 pt/cm3), ln-distance to nearest port (coefficient -900 pt/cm3) and density of fast food restaurants in 200m (coefficient 5400 pt/cm3). The model R2 was 0.45 and LOO-R2 was 0.34. 60  The somewhat lower model R2 and higher LOO-R2 indicated that some of the instability of the road length PNC50 model was due to this one site.  3.2.2  Pollution surface images Maps representing the road length and vehicle density LUR models for PNC50, and those  for PNC10 and PNC90 are presented in Figures 3.7 – 3.10. The vehicle density map overall predicted higher PNC than the road length map (Figure 3.8 versus 3.7) as the colour is darker in Figure 3.8 and there is not a gap of medium red over Delta (the region that appears as an island on the left side of the maps about two-thirds of the way down) as there is in Figure 3.7. It is evident that PNC was heterogeneous in this region. The road length model created a more heterogeneous surface than the vehicle density model, as the truck routes and nodes of restaurants were more evidently highlighted. The vehicle density model, with the density of trucks in a 25m buffer as a significant variable, appeared relatively homogeneously dispersed as the density of trucks is not restricted to official truck routes. The PNC10 model might best represent sites with background levels of UFP while the PNC90 model could characterize spikes in concentrations or hotspots. As the two surfaces look similar and the models only differed by fast food locations within 100m variable (in the PNC90 model), it seems plausible that fast food outlets may be an important source of peak UFP levels. In all maps, the predicted level of PNC declined from west to east (generally as the distance to the nearest port increased) and away from the area more densely populated and developed. The truck route pattern was visible in all surfaces (three models have a truck route variable in them and the remaining model has density of trucks in 25m in it).  61  Figure 3.7 Pollution surface for road length, PNC50 LUR model. Variables in model: length of truck route within 50m, distance to port (ln km), density of fast food restaurants within 200m.  62  Figure 3.8 Pollution surface for vehicle density, PNC50 LUR model. Variables in model: density of trucks within 25m, distance to port (ln km)  63  Figure 3.9 Pollution surface for road length, PNC10 LUR model. Variables in model: length of truck route within 50m, distance to port (ln km). Note: different scale than previous figures.  64  Figure 3.10 Pollution surface for road length, PNC90 LUR model. Variables in model: length of truck route within 50m, distance to port (ln km), density of fast food restaurants within 200m. Note: different scale than previous figures.  65  3.3 3.3.1  Evaluation Assessment of 60-minute samples To evaluate the representativeness of 60-minute samples on longer-term PNC, a random  selection of 14, 60-minute PNC50 measurements (one from each day of the sampling campaign, during sampling hours) from each of the four fixed sites was plotted along with a line showing the overall mean value for that site during the entire two-week sampling campaign. Plots for the four sites are shown in Figure 3.11.  66  A  B  C  D  Figure 3.11 Random selection of 60-minute PNC measurements compared with two-week overall PNC at the same location. A, B, C, D: four fixed site locations.  67  The 14 random points are somewhat evenly distributed around the mean, with a tendency to over-estimate the mean (in panel D, this obvious pattern is enhanced by two outliers). This suggests that a 60-minute sample can characterize long-term averages of PNC (two-week averages in this case) reasonably well, though it may overestimate overall means. It should be noted that the 14 samples from each location were selected from sampling hours (8:00am to 6:00pm) while the overall mean value was calculated over 24 hours. As PNC is lower overnight than during the day (as observed in instrument logs) the positively skewed plots in Figure 3.11 are to be expected.  3.3.2  Comparing spring 2010 with fall 2009 measurements The Pearson correlation between adjusted PNC50 measurements at 22 sampling locations  in fall 2009 and spring 2010 was 0.58. The road length (PNC50) model developed with spring PNC measurements was used to predict the unadjusted and adjusted measurements collected in the previous fall campaign. Adjustment factors are displayed in Table 3.11 and an evaluation of the predicted versus measured concentrations is in Table 3.12.  Table 3.11 Correction factors for comparing fall 2009 to spring 2010 PNC measurements  Description P-trak 8525 to CPC 3007 values Fall to spring seasonal adjustment (based on mean PNC50 measures) Spring to fall adjustment Fall to spring seasonal adjustment (based on mean NO measures) Spring to fall adjustment  Correction factor 1.51 0.596 1.68 0.264 3.79  68  Table 3.12 Predicting PNC at fall sites based on road length (PNC50) LUR model from spring sampling  Method of comparison Uncorrected Temporally corrected (hour of day, season) by UFP Temporally corrected (hour of day, season) by NO  RMSE (pt/cm3) 23 500 (31500) 12 100 (14000) 13 900 (15800)  R2 (adjusted) 0.29 0.28 0.26  The R2 of the seasonal validations were slightly lower than the LOO method (0.26 – 0.29 versus 0.32). The RMSE was approximately halved when fall measurements were temporally adjusted. RMSE values for adjusted fall data were large at over 12 000 pt/cm3, representing 7282% of the mean (fall PNC corrected mean: 16 900 pt/cm3). Temporal adjustment by UFP measurements versus by NO measurements did not greatly impact the RMSE of the predicted PNC, as those using the other pollutant to adjust had only 14% greater error.  3.3.3  Internal validation The LOO cross-validation on the eleven LUR models yielded R2 between predicted and  measured values from 0.14 and 0.41 (central site wind rose buffers for PNC50 and road length for PNC90 models respectively) and RMSE between 9100 and 22 400 (road length for PNC10 and vehicle density for PNC90 models). This is shown in Table 3.9. LOO values shared a similar pattern to model R2, with decreases from the model R2 ranging from 14% (ln-PNC50 by vehicle density, model 4) to 52% (PNC50, road length model based on central wind-rose shaped buffers, model 10). The PNC90 road length model had a higher LOO R2 than the PNC50 model with site observations. RMSE values were high, ranging from 9300 to 22 400 pt/cm3. Considering mean values were 12 600, 18 200 and 30 200 pt/cm3 for PNC10, PNC50 and PNC90, RMSE values were 72 – 74% of the mean for the relevant distribution. RMSE values were generally higher for vehicle density models, PNC90 models (and 69  lower than PNC50 for the PNC10 models) and wind-rose shaped buffer models compared with the PNC50 road length model.  3.3.4  External validation Using the road length PNC50 model to predict replicate measurements at 20 locations,  (i.e. measurements not included in building the model, but at sites that were included) yielded a larger validation R2 (0.63) and smaller RMSE (10 600 pt/cm3) than the LOO methods, shown in Table 3.13. The model explained more variability in these repeat measures and they were subject to less error than the LOO approach. This is logical as these same 20 sites were included in the model building (but with different measurements taken at a different time). This outcome increased confidence in the ability of the model (PNC50, road length) to predict the spatial pattern of UFP, however, the absolute error (RMSE 58% of the mean) is still quite large. Similarly, the PNC50 road length model was also used to predict PNC at the four fixed sites, both during sampling hours (8:00am to 6:00pm, Monday to Friday) and during all hours during the two-week sampling period. Results are in Table 3.13. The RMSE values increased to about 15 000 pt/cm3, which is larger than the LOO results and represents about 82% of the mean. This was a high error level, suggesting predictions at these sites were not very accurate. Interestingly, the RMSE for the four sites was nearly the same for sampling hours versus all hours, indicating that temporal variation is not that important when predicting PNC. The R2 was not calculated, as it is not a valid statistic for four sites.  70  Table 3.13 Predicting PNC levels at repeated and fixed sites by road length (PNC50) model  Twenty repeat sites Four fixed sites: sampling hours Four fixed sites: all hours  RMSE (SD) (pt/cm3) 10 600 (13 400) 15 000 (16 600) 15 200 (16 400)  R2 (adjusted) 0.63 N/A N/A  Overall, these external validation results increase confidence in the predictability of the PNC50 road length model, but highlight its limitations due to large RMSE values.  71  Chapter 4: Discussion This chapter compares the Vancouver PNC measurements and LUR models with other LUR models to contextualize this work. The key findings of this study, as well as strengths, limitations, applications, future work and success of the modeling process are described.  4.1  PNC measurements Variation in PNC was dictated more by spatial factors than temporal factors. Fairly small  temporal correction factors (most between 0.6 and 1.5) were applied to raw measurement data. Further, a LUR model developed from spring measurements was able to predict measurements of PNC during the fall nearly as well as in the spring, with similar validation R2 values (0.32 versus 0.28 – 0.29) and comparable RMSE values (12 800 versus 12 100 – 13 900 pt/cm3). Temporal adjustments for the hour of the day and season reduced the RMSE markedly (from 23 500 to 12 100 and 13 900 pt/cm3). Interestingly, the R2 and RMSE values for models with temporal adjustment based on PNC measurements and models adjusted for temporal trends using NO measurements from the Metro Vancouver monitoring sites were quite similar (0.28 and 12 000 pt/cm3 versus 0.26 and 13 900 pt/cm3). This suggests that NO (or likely NO2 or NOx), a routinely monitored, related pollutant in the TRAP mixture, could be used to seasonally adjust PNC data. This could be useful in an epidemiologic application of the UFP LUR model to times of the year when continuous measurements of PNC are not available. Conversely, spatial distribution of PNC varied widely across the region, with a seventyfold increase from minimum to maximum site median measurements observed (1500 – 105 000  72  pt/cm3). There was a three- and seven-fold increase from 25th to 75th percentile and 10th to 90th percentile respectively. The last three days of between-day temporal correction factors (2.9 – 3.8) were about twice as large as the corrections applied (0.61 – 1.9) in earlier days of sampling. This increase could be caused by the lack of measurements from two fixed sites with higher PNC which were stopped a few days earlier as instruments had to be returned. The relatively large daily correction factors in the last three days of sampling compared to the other days might have biased the results by increasing the median values for measurements collected on those days. This would be a positive bias, that could have influenced models to over-predict PNC levels. However, as a number of the sites visited during the last days of sampling were repeat sites (and were not included in model development, only evaluation) the potential impact of this is likely to be minimal. Only four sites sampled in the last three days were first-time samples, with measurements used to make the LUR models. Additionally, review of the trend of PNC over the last three days of sampling showed a real decrease in concentrations, which is further supported by correction factors developed from the two sites that were operated during the last three days showing similar patterns as when all four sites were included in calculation of correction factors. Between-day corrections were greater than within-day corrections (ranges), with the former similar in magnitude to spatial variability. This trend could be due to varying activity patterns by day of the week. It is unlikely that temperature, which varies more within a day than between days usually and was not an important predictor of PNC, can predict this difference. Examination of Figure 3.11 shows that the 14 randomly selected one-hour samples at three sites are approximately evenly distributed around the individual site mean during the two weeks of continuous measurements, while site four’s samples are skewed above the mean. From 73  this, one-hour measurements appear to reflect at least two-week periods reasonably well. Further, a correlation of 0.58 between one-hour measurements in fall 2009 and spring 2010 also suggests that one-hour samples can capture a fair amount of long-term spatial variability in PNC.  4.2  LUR models Eleven LUR models for UFP in Vancouver were produced. The amount of variability in  measured PNC explained by the models (R2) ranged from 0.29 – 0.53. The most representative model from this suite was the road length (PNC50) model, with an R2 of 0.48, as it had the highest R2 for a model predicting a median or ln-median PNC value and lowest RMSE. The LOO cross-validation R2 was 0.32 and RMSE was 12 800 pt/cm3, representing 70% of the mean. The LOO cross-validation approach to assess the R2 and RMSE of predictions versus measurements yielded a range of 0.14 – 0.41 and 9100 – 22 400 (pt/cm3), which suggest modest performance. The R2 between predictions and measurements using the road length (PNC50) model and 20 repeated sites was 0.63, indicating the model performed quite well in predicting PNC at those locations. The RMSE at repeated sites was slightly lower than for the model by the LOO method, at 10 600 (pt/cm3). Using the road length (PNC50) model to predict PNC at the four fixed sites yielded RMSE values around 15 000 pt/cm3, quite high considering fixed site mean values were 8000 – 14 000 pt/cm3. However, there was little difference between the RMSE of the sampling hours and that of all hours of the two-week period (15 000 versus 15 200 pt/cm3), suggesting that this model can predict PNC mean levels over a longer term (at least two-week periods). Creation of regression maps for some of the PNC models (Figures 3.7 – 3.10) showed that the road length models typically predicted lower levels of PNC than vehicle density models, 74  but had prediction values more heterogeneously distributed. In all surfaces, PNC tended to decline from west to east, suggesting that some factors in the western part of the region (ports, increased population density and associated motor vehicle activity) accounted for increased levels of PNC.  4.2.1  Comparison of Vancouver PNC LUR models with other LUR models Table 4.1 lists a selection of previously published LUR models with which to compare  the Vancouver PNC LUR model results. It is important to note that this listing of models is not comprehensive; a review on LUR models in 2008 (55) identified 25 published models and there are even more available now. These comparisons focus on models that examined the same pollutant (PNC) (44), were located in the same region (Vancouver) (61), or provided examples of the greater North American and world context. A discussion comparing the model in Table 4.1 to the PNC models in Vancouver follows.  75  Table 4.1 Summary of LUR models published by other researchers for comparison with the Vancouver PNC LUR models  Location (N) Vancouver (80) Vancouver (114) (114)  Investigator Pollutant (year) Abernethy PNC et al. (2012) Henderson logNO et al. (2007) (61) NO2  (25) Vancouver (39) Amsterdam (46)  Larson et al. (2009) (71) Hoek et al. (2011) (44)  Netherlands Brauer et al. (40) (2003) (74) Munich (40) Stockholm County (42) Amsterdam Briggs et al. (80) (1997) (59)  Variables in LUR models  -TRK.50, LN-DIST.PORT, FFD.200 -removal of site observations -RD.100, RD.1000, RD2.100, POP.2500, ELEV, X, Y; AD.100, TD.1000, ELEV, X, Y -RD.100, RD.1000, RD2.100, POP.2500, COM.750, ELEV, X; AD.100, TD.200, TD.1000, COM.750, POP.2500, ELEV PM2.5 -ELEV, COM.300, RES.750, IND.300, AD.100, ELEV, COM.300, RES.750 Black -RD1.500, RD2.750, ELEV, X, RD2.300, carbon TD.300 PNC -traffic intensity x inverse DIST.RD2, DWELL.300, PORT.300 -removal of site observations PM2.5 -traffic intensity x inverse DIST.RD2, OPEN.300, AD.100 PM2.5 -AD.250, address 300m, traffic site, Absorbance DIST.RD2, region, TRAF.50, TRAF.50250, street canyon, POP.1000-5000 PM2.5 Absorbance PM2.5 Absorbance NO2 -RD1.50, RD2.50, RD1.200, RD2.500, RD3.50, RD3.200, Built_land.100, DIST.RD1  LUR R2 (validation R2) 0.53 (0.38) 0.48 (0.32) 0.57-0.62 (0.36-0.65)  RMSE (SD) 12800 (23100) 1.65-1.9 (12.7-13.1)  Mean or median measurements (SD) 13600 pt/cm3 a  0.56-0.60 (0.31-0.79)  0 (2.612.65)  16.2 (5.6) ppb b  0.52 (0.14)  0 (1.5-1.53)  4.08 (1.98) µg/m3 b  0.51-0.72 (0.40-0.63) 0.65 (0.57)  0.4-5.08  9.72 (4.85) σap  Not published  22400-40400 pt/cm3 a  0.44 0.54 (0.50) 0.73-0.78 0.81-0.90 0.56-0.76 0.67-0.83 0.50-0.63 0.66-0.76 0.62 (0.79 – 10 sites)  31.3 (21.6) ppb b  a  22-24 µg/m3 a 1.1 – 1.6 µg/m3 for PM2.5 0.22 – 0.31 10-5/m for absorbance 4.45 µg/m3  17.5 (µg/m3) b 1.64 (10-5/m) b 13.6 (µg/m3) b 1.84 (10-5/m) b 10.3 (µg/m3) b 1.28 (10-5/m) b 36.4-51.8 (6.4-12.2) µg/m3 b 76  Table 4.1 Summary of LUR models published by other researchers for comparison with the Vancouver PNC LUR models  Location (N) New York (36-62) North RhineWestphalia, Germany (40) Montreal (57)  Investigator (year) Ross et al. (2007) (75) Hochadel et al. (2006) (76)  Pollutant  Variables in LUR models  PM2.5  -traffic counts 500m, POP.1000, IND.300, OPEN.1000 -TD.250-10000, AD.250, DIST.RD2, building count 2500x2500 -TD.100-10000, AD.100, DIST.RD1  Gilbert et al. (2005) (77)  NO2  Toronto (94)  Jerrett et al. (2007) (78)  NO2  Los Angeles (23)  Moore et al. (2006) (79)  PM2.5  a  median  b  mean  PM2.5 PM2.5 Absorbance  -DIST.HWY, traffic counts RD1, RD1.100, RD2.100, RD3.500, OPEN.100, POP.2000 - RD1.200, RD2.50, IND.750, DWEL.2000, X, DWIND.1500, TRAF.500 -TRAF.300, IND.5000, GOV.5000  LUR R2 (validation R2) 0.61-0.64  RMSE (SD)  0.07  2.16  Mean or median measurements (SD) 14.3-15.3 (1.4-1.8) (µg/m3) b 22.4 (µg/m3) b  0.81  0.144  1.71 (10-5/m) b  0.54 (0.52)  Not published  11.6 (3.0) ppb b  0.67  0.163 ln ppb (1.2 ppb)  32.2 (9.2) ppb b  0.69 (0.63)  Not published  18.4 (6.01) µg/m3 b  0.77-0.93  * Note: variables presented in this table are denoted using the same coding as variables used in Vancouver model development, shown in Table 2.2 for those that are common. Other variables are written out fully. 77  Land use regression models are geographically specific, due to different geographic and meteorologic factors in different regions as well as varying emissions and chemical features of pollutants. It is difficult to transfer models developed in one region to another and to compare models created in other parts of the world. (80,81) The transfer of the Vancouver LUR model for NO2 (61) to Victoria and Seattle, locations geographically close and similar to Vancouver, yielded reduced explanation of variation compared with the original model (R2 of 0.33 and 0.51, compared with 0.62). (80) Similar results were obtained when interchanging NO and NO2 models between Edmonton and Winnipeg. (81) As such, it is important to recognize that the exercise of comparing the Vancouver UFP LUR model to other models is more qualitative than quantitative in meaning. The PNC model developed for Amsterdam is similar in several ways to the models developed in Vancouver. In both cases, relatively few variables (usually about three), including one related to ports, were in the final model. This is strong evidence that something linked to ports is contributing to PNC. The PM distribution emitted by ships is known to be dominated by UFP (36,37) and has been shown to be a likely important source of UFP in a separate study in Amsterdam. (35) The Amsterdam model also included a combined traffic intensity and inverse distance to nearest major road variable, while the principal Vancouver model included length of truck route within 50m. Both cases suggest vehicle traffic as an important source of UFP, while the Vancouver model more specifically suggests heavy-duty diesel engine trucks to be the primary traffic-related source. This finding is consistent with other studies that have shown exponential decreases in PNC with increased distance from heavy-duty vehicles. (12,15,28) The third variable in the Amsterdam model was address density within 300m and in Vancouver it was fast food locations in 200m. This difference could indicate distinct sources of UFP in the 78  two locations, although the Amsterdam model did not offer restaurant related variables to the models. Notably, the Vancouver models did have population and dwelling density (address density) variables offered to them, but none of the final models contained any of those variables (in contrast with some of the other LUR models shown in Table 4.1). This could be because population or address density type variables are surrogates for other variables, such as restaurant density or traffic density, which may be the actual drivers of PNC levels. Another possible explanation is that Vancouver uses less fossil fuel (electricity is a major heating source) for residential heating than other areas. (33,34) Vancouver also has a relatively mild climate, so lower fossil fuel use due to less heating in winter could be another reason for this observation. The R2 for the Amsterdam PNC model including site observations was 0.65 (validation: 0.57), which is analogous to the 0.53 (validation: 0.38) found in Vancouver ‘best’ model (model 11, Table 3.9). In this case the Amsterdam model explained more of the variability in PNC than the Vancouver model and the LOO cross-validation R2 remained quite a bit higher. However, once site-specific observations were excluded, the Amsterdam model R2 dropped to 0.44 (no validation R2 reported), while the main Vancouver model was 0.48 (LOO R2 = 0.32). The main site observations that contributed to the more predictive model in Amsterdam were distance measurements to the edge of the nearest road. The authors suggested the discrepancy could be due to GIS measurements going to the middle of a street, small GIS errors having an amplified effect in a compact city and contributions of mopeds and motorcycles traveling at the edge of the road. In Vancouver, manual truck counts increased the R2, providing the most accurate heavyduty traffic data for a specific site at a specific time. As only a model where variable information is available at any site in the area of interest is applicable in an epidemiologic setting, these two models from different parts of the world seem quite similar in their predictive power. 79  The median PNC was higher in Amsterdam than Vancouver, 22 400 – 40 400 pt/cm3 versus 18 200 pt/cm3. This range is due to the stratification of traffic and background sites. The Amsterdam model also included fewer sites than Vancouver (46 versus 80), which may partially account for the higher R2 of the ‘best model’ as LUR models with fewer sites tend to generate higher R2 (they may over-estimate the explanatory ability of the model). (55,56) The Vancouver PNC LUR model explained a bit less variability than the Vancouver NOX models, which had R2 from 0.56 – 0.62. The NOX models contained more variables than the PNC model, including multiple road length or vehicle density variables, a population density variable and a geographic location variable. The Vancouver PM2.5 model, based on only 25 sampling locations, had an R2 of 0.52, which is comparable to the PNC50 and PNC90 road length models of 0.48 and 0.53. The validation R2 for NOX showed similar patterns as that of PNC, ranging from 0.31 – 0.79. This represents a fairly large (up to 0.25) decrease in R2 values from model to validation, which was observed to a generally lesser degree in the PNC models. External validation with repeat and monitoring sites showed high R2 values, up to 0.79, which is similar to the increase to 0.63 shown in validation of 20 repeat sites with PNC. The RMSE values for NO modeling were only about 6% of the mean, whereas they were over 50% for PNC, indicating one has much more confidence in the predicted levels of NO than of PNC. The Vancouver Black Carbon model performed similarly to the NOX models, with a higher model R2 than the PNC model (0.51 – 0.72) and comparatively high validation R2 (0.40 – 0.63) values. However, the RMSE values represented 4 – 52% of the mean, which is closer to the scale of the PNC RMSE values. This higher error (relative to NOX models) may be due to the smaller number of sites (39).  80  Similarly, the Amsterdam PNC LUR model (R2 = 0.44 – 0.65) explained a bit less variability than the LUR model for PM2.5 done at the same time (R2 = 0.54 without site observations), the LUR models for PM2.5 and Absorbance across the Netherlands in a previous study (R2 = 0.73 – 0.90) and the original NO2 model for Amsterdam (R2 = 0.62). (44,59,74) There is insufficient published validation numbers to compare within the Netherlands as some studies reported LOO R2, some sub-sample R2, and some RMSE. Generally, LOO R2 values were slightly lower than model R2 and RMSE numbers were less than 20% of the mean of measurements. As with the Vancouver models, a number of the variables in the Dutch models were similar, pertaining to length of roads (especially major roads within 500m), vehicle density and building density. Also as in Vancouver, the PNC model remained the only one to include a variable related to the port, underscoring the possibility that port-related emissions are an important source of UFP specifically. As with the NOX models in Vancouver discussed earlier, the other models in Table 4.1 tended to have some type of population variable and geographic location variable in addition to traffic or road variables. R2 for the other models ranged from 0.5 – 0.9, placing the Vancouver PNC models at the lower end of this range of explanation offered by models. RMSE values for the other models ranged from 4 – 24% of the reported measurement mean, which is markedly less than the >50% RMSE numbers obtained from the Vancouver PNC models. However, many of the models in Table 4.1 were based on fewer sites than the Vancouver PNC model, which could contribute to those models having higher R2 values. Overall, the development of LUR models for UFP in Vancouver was successful. The amount of variability the models explained (R2 = 0.29 – 0.53) is on the lower end of the spectrum of related published LUR models, but within the range of other studies. 81  As the UFP LUR models highlighted the sharp spatial gradient of that pollutant with inclusion of variables with buffer sizes <100m, one reason these models tend to have lower R2 values could be due to the difficulty in modeling something that varies so much over small distances and the limitation of accurate coordinates. Perhaps the most limiting aspect of the Vancouver PNC LUR models developed are the very high RMSE values associated with the models, often exceeding 50% of the mean measurements for a particular dataset. However, internal and external validation methods showed R2 reasonably close to model R2 values, suggesting models are at least reliable in predicting overall trends in concentration on a fine spatial scale.  4.2.2  Comparison with other LUR models incorporating meteorology Other studies have found that varying meteorology, especially regarding wind speed and  direction, can influence PNC substantially. (12,15,45,50) Several studies have incorporated wind speed and direction into LUR models for other pollutants. One LUR model for NO2 in Portland incorporated wind direction using the wind rose for the entire region during the sampling time (mean wind direction and two standard deviations on either side for the range) and applied this wedge-shaped buffer to the geographic predictor variables (in concert with circular buffers). (51) Inclusion of wind direction improved model predictive power by 0.15. Another study in Hamilton-Toronto on NO2 interpolated observed wind speed and direction from weather stations and found this to improve models slightly (R2 increased 0.01 – 0.04). (82) However, a coarse surface (10km grid) of wind patterns did not improve models in this case. A third study from Vancouver incorporated wind speed, wind 82  direction and cloud cover in a source area LUR model for NO and NO2 and found that due to interpolating meteorology variables, this approach was not more helpful than LUR using circular buffers. (73) A fourth paper modeling PM2.5, elemental carbon and NO2 in Boston incorporated wind direction using meteorology data from one central monitoring location. (67) Variables such as percent of hours a site was downwind of a major road, mean wind speed during sampling time and percent of daytime with still winds were developed. These variables appeared in final models and the authors recommended further studies incorporating more site-specific meteorology variables. In this project, site-specific meteorology data as well as central station monitoring data were used to create two models (model 9 and 10 in Table 3.9) using wind rose shaped buffers instead of circular buffers. Use of these buffers did not improve predictive abilities of the models, as R2 values dropped from 0.48 to 0.34 and 0.29. It is unclear why the meteorology models did not improve R2 or maintain the R2 previously achieved. Perhaps due to the tendency of UFP to be short-lived and reactive (rapidly coagulating in the atmosphere), transport by wind over 60 minutes is not as important in determining concentrations as sources directly around a point (equidistant from the sampling location in a circular buffer). As most of the wind speeds were quite low (interquartile range for measured speeds was 0.5 – 2.2 m/s), perhaps circular buffers were a sufficient approximation for source contribution. One additional way of testing meteorologic impacts on PNC is to conduct measurements in areas where strong wind directions are highly predictable, such as near the ports, where winds are predominantly onshore or offshore. From this experience, the additional effort of measuring meteorology and conducting additional analyses and model development was not worthwhile. 83  4.2.3  Key findings of the Vancouver PNC LUR models Variables that were consistently significant in final models included: those related to  heavy-duty trucks (length of truck route or density of trucks), distance to the port, density or proximity to fast food restaurants or grills, distance to the nearest intersection of major roads. This suggests these variables as important sources of UFP in Metro Vancouver. The heavy-duty diesel truck related variables in the final models were for small buffer sizes (truck route length within 50m and truck density within 25m), suggesting that marked changes in UFP concentration occur on a very small scale from a source. Further, the distance to port variable, which had a negative coefficient, frequently showed up in the form ln-distance to port in the models, as PNC decreased logarithmically from the port. Fast food restaurant variables in final models were for 100m or 200m buffers, underscoring the fine spatial scale of UFP distribution. The smallest buffer investigated for fast food restaurants was 50m, as the closest restaurant to a sampling site was 40m (so a 25m buffer was uninformative).  4.3  Strengths This design of this study benefitted from previous LUR model development in  Vancouver for other pollutants and associated methodology developed therein. The number of sites and length of measurements, presence of four fixed site monitors, quantity and type of geographic predictor variables and inclusion of meteorologic considerations enhance the robustness of the methods employed to develop UFP LUR models for Vancouver. Mobile sampling occurred at 80 sites for 60-minute periods, which is at the high end of number of sites needed to create a LUR, as 40 – 80 locations have been suggested by others (55) as the minimum number needed to provide enough data to resolve spatial patterns in pollutant 84  concentration. Further, 60-minute samples appeared to adequately represent long-term averages of PNC (two-week and seasonal). Overall, the number and length of samples optimized the quantity of data and resource expenditure. Another major strength of this study was the inclusion of four fixed sites, which continually monitored PNC during sampling and were used to temporally adjust data. These four locations were situated to capture suspected high and low PNC locations (high traffic and low traffic areas), to be representative of the region. Other studies used a single central monitoring site, often measuring a different pollutant to correct data, or had no corrections at all. (50,67,71) From the temporal adjustments, it can be concluded that temporal variation (factor of three variability) appears to be less important than spatial variation (factor of seventy variability from minimum to maximum, factor of seven from 10th to 90th percentiles) in determining UFP levels. A total of 135 geographic predictor variables were prepared for consideration in the LUR models. This represents nearly a doubling of the number of variables offered to NOX LUR models created previously in the region (61) and included many new types of variables, highlighted in Table 2.2. Many more factors were considered when attempting to predict PNC levels than had previously been done. Inclusion of new variables was valuable in this study, as some of the new variables were included in the final LUR models and evidently helped to explain sources and factors affecting PNC. They were especially relevant as many of the final models only included a small number of variables (two to five) and contributed substantial partial R2 values: 0.05 – 0.24 for fast food to distance to nearest intersection respectively. This is the first LUR model to include fast food restaurant variables as potential predictors. Notably, these were significant positive predictors of PNC in the two most predictive models (PNC50, road  85  length and PNC90, road length), suggesting that there should be more emphasis on restaurants as potential UFP sources in urban areas. A novel aspect of this LUR model development was the incorporation of meteorologic variables, including wind speed, wind direction and temperature measured at the sampling site. While inclusion of these data did not improve the predictive power of the models in this instance, failure to consider meteorology is often cited as a limitation in other LUR models predicting air pollution species. (45,50,67)  4.4  Limitations There are some limitations associated with this study. First, the LUR models were based  on measurements made in the spring and may not represent the entire year. While a subset of PNC measurements at 22 sites completed in the fall were correlated with spring measurements, values were 1.68 – 3.79 times higher. Using the spring model to predict fall measurements yielded validation results (RMSE) comparable to spring measurements, suggesting this model is as applicable in spring as it is to fall seasons. Further, while fall and spring were selected as periods of the year to best represent the annual average of NOX, the annual pattern of PNC in Metro Vancouver is not well known. A paper from Stockholm, Sweden used a combination of measurements each month of the year over a number of years and dispersion modeling to estimate monthly PNC concentrations in that city. (83) The researchers found a 2.5-fold change between the month with the highest levels (December) and lowest (August). Concentrations tended to be lowest during the summer, June-September, and highest during fall and winter, November to February. The annual average PNC levels are well represented in the months of  86  April and October in Stockholm, the same months when Vancouver sampling was completed, providing support that samples during these times reflects the annual average. The UFP LUR models exhibited similar performance to the Amsterdam UFP LUR model (44) and relatively poor performance compared to LUR models for other pollutants in this region (Table 4.1). Model R2 values were not higher than 0.53, LOO R2 were not higher than 0.41 and RMSE values for all models were quite high, representing 68 – 81% of the mean. These observations all suggest that other unaccounted factors also contribute to UFP concentration variability or that the short-term presence of UFP in air is difficult to model. Application of these models in epidemiologic studies should consider these limitations. However, the most predictive models do account for nearly half of the variability in PNC levels and due to the pollutant’s highly variable spatial distribution, likely can offer a better estimate than a coarsely applied estimate over a large area. Two predominant variables in final LUR models were distance to the nearest port and density of fast food outlets. However, it is unclear what the specific source of UFP is relating to these variables, as emissions may be from the port and fast food sites themselves or those variables may be a surrogate for sources at or near those locations. Contributions to PNC are likely from a combination of both scenarios. Correlations between fast food and port variables with traffic-related variables showed some moderate relationships between these two types of variables. Based on Table 3.10, it seems possible that bus stops, major roads and intersections contribute to PNC near fast food restaurants. Somewhat surprisingly, distance to the nearest bus stop was highly correlated with distance to the port (r = 0.68) while other traffic-related variables were weakly correlated with the port. This suggests that bus stops may play a role in observed high levels of PNC near the port. 87  4.5  Applications The UFP LUR models developed in this study have the capacity to be used in  environmental epidemiology studies in Metro Vancouver, linking levels of pollutants to health outcomes in an exposed population. While the LUR models do not have R2 values as high as other models and RMSE values are large, these models are fairly temporally stable (60-minute samples represent long-term levels fairly well and can be used to predict concentrations in other seasons) and can provide predictions of relatively high or low PNC, capturing trends and contrasts in levels. This represents an important addition to information currently available to assess long-term, population-level health effects of UFP. The understanding of temporal trends of UFP (through temporal correction) as well as spatial variation and distribution of UFP in the Metro Vancouver is useful to know when developing future sampling campaigns which can focus on teasing out specific UFP sources and considering seasons for sampling. The regression maps created from the LUR models inform policymakers where UFP hotspots are located, which could be used to inform air quality management. Finally, the commencement of a database of UFP measurements in Metro Vancouver offers a launching point for future sampling campaigns and for possible future development of a monitoring network and regulatory limits in the area.  4.6  Future work Additional work on refining UFP LUR models for Metro Vancouver and for better  understanding the sources of UFP and factors influencing concentrations can be completed in 88  order to better understand spatial-temporal patterns of UFP in the area and to increase confidence in applying models to epidemiologic studies. Additional sampling around the ports and fast food locations can be completed to determine whether the port itself (ships) and restaurants themselves (cooking processes) contribute to UFP levels or whether related traffic is responsible. Further assessment of the validity of 60-minute samples to accurately reflect long-term concentrations of UFP can be completed, as well as determination of whether shorter time periods can capture the same trends. Sampling could also be conducted over different times of the year, in order to gain a better understanding of annual PNC patterns. This should be a lower priority than other sampling as comparison between fall 2009 and spring 2010 measurements yielded seasonal differences (adjustment factor 1.68 – 3.79) but the temporal variation was not found to be as important as spatial variation.  4.7  Conclusion All project objectives outlined in the introduction have been completed and land use  regression models for ultrafine particles in Metro Vancouver have been successfully developed. Sixty-minute measurements of PNC using CPC 3007 devices at 80 locations, a subset of sites previously used to develop LUR models for other pollutants in Vancouver, measured spatially variable concentrations with median sample values ranging from 1500 – 105 000 pt/cm3. Temporal adjustments to raw data based on four continuous monitors were generally small, ranging from 0.613 – 3.82, suggesting UFP temporal variation over a two-week period is lower than the spatial variability. Comparisons to measurements collected in a different season indicated that seasonally corrected PNC using NO measurements from central monitoring sites is  89  nearly as effective as using PNC data to temporally correct measurements. In addition, one-hour samples seem to represent longer-term (two-week period) PNC averages fairly well. Eleven land use regression models, drawing from 135 potential predictor variables, were created. The most representative model created, based on road length variables predicting PNC50, included the variables: length of truck routes within 50m, density of fast food locations in 200m and ln-distance to the nearest port. This model explained 48% of the variation in PNC; a LOO cross-validation exercise yielded an R2 of 0.32 and RMSE of 12800 pt/cm3 (70% of the mean), suggesting the model is subject to a large degree of error. A new method to incorporate measured and monitored meteorologic variables (wind speed and wind direction) into LUR models was developed and applied but did not improve the predictive power of the models compared to those using circular buffers. The established UFP LUR models can be used in epidemiologic applications in Metro Vancouver. Additional monitoring around ports and fast food sites will help determine the contribution of these features to PNC in the region. This project had laid a foundation for PNC data collection and modeling in Metro Vancouver from which additional studies and perhaps eventually policy-related monitoring protocol can be developed.  90  References 1. Health Effects Institute: Panel on the Health Effects of Traffic-Related Air Pollution. TrafficRelated Air Pollution: A Critical Review of the Literature on Emissions, Exposure, and Health Effects [Internet]. Boston, MA: Health Effects Institute; 2010. Report No.: Special Report 17. [cited 2011 Dec. 13]; Available from: http://pubs.healtheffects.org/getfile.php?u=553 2. Hoffmann B, Moebus S, Dragano N, Möhlenkamp S, Memmesheimer M, Erbel R, et al. Residential traffic exposure and coronary heart disease: results from the Heinz Nixdorf Recall Study. Biomarkers. 2007;14(s1):74–8. 3. Hoffmann B, Moebus S, Stang A, Beck E-M, Dragano N, Mohlenkamp S, et al. Residence close to high traffic and prevalence of coronary heart disease. European Heart Journal. 2006;27(22):2696–702. 4. Brauer M, Hoek G, Smit HA, de Jongste JC, Gerritsen J, Postma DS, et al. Air pollution and development of asthma, allergy and infections in a birth cohort. Eur. Respir. J. 2007;29(5):879–88. 5. Brauer M, Lencar C, Tamburic L, Koehoorn M, Demers P, Karr C. A cohort study of trafficrelated air pollution impacts on birth outcomes. Environ. Health Perspect. 2008;116(5):680– 6. 6. Clark NA, Demers PA, Karr CJ, Koehoorn M, Lencar C, Tamburic L, et al. Effect of Early Life Exposure to Air Pollution on Development of Childhood Asthma. Environ Health Perspect. 2010;118(2):284–90. 7. Karr CJ, Demers PA, Koehoorn MW, Lencar CC, Tamburic L, Brauer M. Influence of Ambient Air Pollutant Sources on Clinical Encounters for Infant Bronchiolitis. Am. J. Respir. Crit. Care Med. 2009;180(10):995–1001. 8. World Health Organization. The world health report 2002: reducing risks, promoting healthy life [Internet]. Geneva, CH: World Health Organization; 2002. [cited 2010 June 1]; Available from: http://www.who.int/whr/2002/en/whr02_en.pdf 9. Jerrett M, Arain A, Kanaroglou P, Beckerman B, Potoglou D, Sahsuvaroglu T, et al. A review and evaluation of intraurban air pollution exposure models. J Expo Anal Environ Epidemiol. 2005;15(2):185–204. 10. Metro Vancouver - Air Quality Index by Station [Internet]. 2010. [cited 2010 Jun 28]; Available from: http://apps2.metrovancouver.org/aqi/ 11. Geiser M, Rothen-Rutishauser B, Kapp N, Schürch S, Kreyling W, Schulz H, et al. Ultrafine Particles Cross Cellular Membranes by Nonphagocytic Mechanisms in Lungs and in Cultured Cells. Environ Health Perspect. 2005;113(11):1555–60. 91  12. Zhu Y, Hinds WC, Kim S, Shen S, Sioutas C. Study of ultrafine particles near a major highway with heavy-duty diesel traffic. Atmospheric Environment. 2002;36(27):4323–35. 13. Government of British Columbia. New Ambient Air Quality Criteria for PM2.5 [Internet]. BC Air Quality. 2009 [cited 2010 Jun 1]; Available from: http://www.env.gov.bc.ca/epd/bcairquality/regulatory/pm25-objective.html 14. Sioutas C, Delfino RJ, Singh M. Exposure Assessment for Atmospheric Ultrafine Particles (UFPs) and Implications in Epidemiologic Research. Environ Health Perspect. 2005;113(8):947–55. 15. Cyrys J, Pitz M, Heinrich J, Wichmann H-E, Peters A. Spatial and temporal variation of particle number concentration in Augsburg, Germany. Sci. Total Environ. 2008;401(13):168–75. 16. Xia T, Li N, Nel AE. Potential Health Impact of Nanoparticles. Annual Review of Public Health. 2009;30(1):137–50. 17. Oberdürster G. Toxicology of ultrafine particles: in vivo studies. Philosophical Transactions of the Royal Society of London. Series A: Mathematical, Physical and Engineering Sciences. 2000;358(1775):2719 –2740. 18. Ferin J, Oberdörster G, Penney J. Pulmonary retention of ultrafine and fine particles in rats. American journal of respiratory cell and molecular biology. 1992;6(5):535. 19. Oberdurster G, Sharp Z, Atudorei V, Elder A, Gelein R, Kreyling W, et al. Translocation of Inhaled Ultrafine Particles to the Brain, Inhalation Toxicology, Informa Healthcare. Inhal. Toxicol. 2004;16(6-7):437–45. 20. Rückerl R, Schneider A, Breitner S, Cyrys J, Peters A. Health effects of particulate air pollution: A review of epidemiological evidence. Inhal Toxicol. 2011;23(10):555–92. 21. Wichmann HE, Spix C, Tuch T, Wolke G, Peters A, Heinrich J, et al. Daily mortality and fine and ultrafine particles in Erfurt, Germany part I: role of particle number and particle mass. Research report (Health Effects Institute). 2000;(98):5–86. 22. Stolzel M, Breitner S, Cyrys J, Pitz M, Wolke G, Kreyling W, et al. Daily mortality and particulate matter in different size classes in Erfurt, Germany. J Expos Sci Environ Epidemiol. 2007;17(5):458–67. 23. Samet J, Graff D, Berntsen J, Ghio A, Huang Y, Devlin R. A Comparison of Studies on the Effects of Controlled Exposure to Fine, Coarse and Ultrafine Ambient Particulate Matter from a Single Location, Inhalation Toxicology, Informa Healthcare. Inhal. Toxicol. 2007;19(Suppl. 1):29–32.  92  24. Rückerl R, Phipps RP, Schneider A, Frampton M, Cyrys J, Oberdörster G, et al. Ultrafine particles and platelet activation in patients with coronary heart disease – results from a prospective panel study. Part Fibre Toxicol. 2007;4:1. 25. Hoek G, Boogaard H, Knol A, de Hartog J, Slottje P, Ayres JG, et al. Concentration Response Functions for Ultrafine Particles and All-Cause Mortality and Hospital Admissions: Results of a European Expert Panel Elicitation. Environ. Sci. Technol. 2010;44(1):476–82. 26. Knol A, de Hartog J, Boogaard H, Slottje P, van der Sluijs J, Lebret E, et al. Expert elicitation on ultrafine particles: likelihood of health effects and causal pathways. Part. Fibre Toxicol. 2009;6(19):1–16. 27. Gertler AW, Gillies JA, Pierson WR, Rogers CF, Sagebiel JC, Abu-Allaban M, et al. Realworld particulate matter and gaseous emissions from motor vehicles in a highway tunnel. Res Rep Health Eff Inst. 2002;(107):5–56; discussion 79–92. 28. Kittelson DB. Engines and nanoparticles. Journal of Aerosol Science. 1998;29(5-6):575–88. 29. Buonocore JJ, Lee HJ, Levy JI. The influence of traffic on air quality in an urban neighborhood: a community-university partnership. Am J Public Health. 2009;99 Suppl 3:S629–635. 30. Thai A, McKendry I, Brauer M. Particulate matter exposure along designated bicycle routes in Vancouver, British Columbia. Science of The Total Environment. 2008;405(1–3):26–35. 31. Shi JP, Evans DE, Khan A., Harrison RM. Sources and concentration of nanoparticles (<10nm diameter) in the urban atmosphere. Atmospheric Environment. 2001;35(7):1193– 202. 32. Lurmann F. Where are people exposed to ultrafine particles? HEI Conference, Portland, Oregon; 2009. 33. Kleeman MJ, Riddle SG, Robert MA, Jakober CA, Fine PM, Hays MD, et al. Source Apportionment of Fine (PM1.8) and Ultrafine (PM0.1) Airborne Particulate Matter during a Severe Winter Pollution Episode. Environ. Sci. Technol. 2009;43(2):272–9. 34. Martin CL, Longley ID, Dorsey JR, Thomas RM, Gallagher MW, Nemitz E. Ultrafine particle fluxes above four major European cities. Atmospheric Environment. 2009;43(31):4714–21. 35. Van der Zee S, Van Der Laan J, Hoek G. The contribution of ships to measured NOx and ultrafine particle concentrations along the waterways in Amsterdam. ISEE Conference, Barcelona, Spain; 2011.  93  36. Saxe H, Larsen T. Air pollution from ships in three Danish ports. Atmospheric Environment. 2004;38(24):4057–67. 37. Healy RM, O’Connor IP, Hellebust S, Allanic A, Sodeau JR, Wenger JC. Characterisation of single particles from in-port ship emissions. Atmospheric Environment. 2009;43(40):6408– 14. 38. Kim S, Shen S, Sioutas C. Size distribution and diurnal and seasonal trends of ultrafine particles in source and receptor sites of the Los Angeles basin. Journal of the Air & Waste Management Association (1995). 2002;52(3):297. 39. Dennekamp M, Howarth S, Dick C, Cherrie J, Donaldson K, Seaton A. Ultrafine particles and nitrogen oxides generated by gas and electric cooking. Occup Environ Med. 2001;58(8):511–6. 40. Cass GR, Hughes LA, Bhave P, Kleeman MJ, Allen JO, Salmon LG. The chemical composition of atmospheric ultrafine particles. Philosophical Transactions of the Royal Society of London. Series A: Mathematical, Physical and Engineering Sciences. 2000;358(1775):2581 –2592. 41. Kleeman M, Robert M, Riddle S, Fine P, Hays M, Schauer J, et al. Size distribution of trace organic species emitted from biomass combustion and meat charbroiling. Atmospheric Environment. 2008;42(13):3059–75. 42. Wallace L, Ott W. Personal exposure to ultrafine particles. J Expos Sci Environ Epidemiol. 2011;21(1):20–30. 43. Hoek G, Meliefste K, Cyrys J, Lewné M, Bellander T, Brauer M, et al. Spatial variability of fine particle concentrations in three European areas. Atmospheric Environment. 2002;36(25):4077–88. 44. Hoek G, Beelen R, Kos G, Dijkema M, Van der Zee SC, Fischer PH, et al. Land Use Regression Model for Ultrafine Particles in Amsterdam. Environmental science & technology. 2011;45(2):622–8. 45. Zhou Y, Levy JI. Factors influencing the spatial extent of mobile source air pollution impacts: a meta-analysis. BMC Public Health. 2007;7(1):89. 46. Hagler GSW, Thoma ED, Baldauf RW. High-Resolution Mobile Monitoring of Carbon Monoxide and Ultrafine Particle Concentrations in a Near-Road Environment. Journal of the Air & Waste Management Association. 2010;60(3):328–36. 47. Ketzel M, Wåhlin P, Berkowicz R, Palmgren F. Particle and trace gas emission factors under urban driving conditions in Copenhagen based on street and roof-level observations. Atmospheric Environment. 2003;37(20):2735–49. 94  48. Fruin S, Westerdahl D, Sax T, Sioutas C, Fine PM. Measurements and predictors of on-road ultrafine particle concentrations and associated pollutants in Los Angeles. Atmospheric Environment. 2008;42(2):207–19. 49. Hanna SR, Britter R, Franzese P. A baseline urban dispersion model evaluated with Salt Lake City and Los Angeles tracer data. Atmospheric Environment. 2003;37(36):5069–82. 50. Weichenthal S, Dufresne A, Infante-Rivard C, Joseph L. Determinants of ultrafine particle exposures in transportation environments: findings of an 8-month survey conducted in Montreal, Canada. J Expos Sci Environ Epidemiol. 2008;18(6):551–63. 51. Mavko ME, Tang B, George LA. A sub-neighborhood scale land use regression model for predicting NO2. Science of The Total Environment. 2008;398(1–3):68–75. 52. Gilbert NL, Woodhouse S, Stieb DM, Brook JR. Ambient nitrogen dioxide and distance from a major highway. Science of The Total Environment. 2003;312(1-3):43–6. 53. Zwack LM, Paciorek CJ, Spengler JD, Levy JI. Modeling Spatial Patterns of Traffic-Related Air Pollutants in Complex Urban Terrain. Environ Health Perspect. 2011;119(6):852–9. 54. Henderson S, Brauer M. Measurement and modeling of traffic-related air pollution in the British Columbia Lower Mainland for use in health risk assessment and epidemiological analysis [Internet]. Vancouver, BC: University of British Columbia; 2005. [cited 2010 Sept. 1]; Available from: http://www.cher.ubc.ca/UBCBAQS/Traffic%20Report/Traffic%20Final%20Report.pdf 55. Hoek G, Beelen R, de Hoogh K, Vienneau D, Gulliver J, Fischer P, et al. A review of landuse regression models to assess spatial variation of outdoor air pollution. Atmospheric Environment. 2008;42(33):7561–78. 56. Brauer M, Henderson S, Marshall J. A land use regression road map for the Burrard Inlet area local air quality study. [Internet]. 2006. [cited 2010 Sept. 15]; Available from: http://www.cher.ubc.ca/PDFs/GVRD_BI_Report.pdf 57. Brauer M, Hoek G, Van Vliet P, Meliefste K, Fischer PH, Wijga A, et al. Air Pollution from Traffic and the Development of Respiratory Infections and Asthmatic and Allergic Symptoms in Children. American Journal of Respiratory and Critical Care Medicine. 2002;166(8):1092 –1098. 58. Morgenstern V, Zutavern A, Cyrys J, Brockow I, Gehring U, Koletzko S, et al. Respiratory health and individual estimated exposure to traffic-related air pollutants in a cohort of young children. Occupational and Environmental Medicine. 2007;64(1):8 –16. 59. Briggs D, Collins S, Elliott P, Fischer P, Kingham S, Lebret E, et al. Mapping urban air pollution using GIS: a regression-based approach. International Journal of Geographical Information Science. 1997;11(7):699–718. 95  60. Government of Canada HC. Regulations Related to Health and Air Quality [Internet]. 2003 [cited 2010 Jun 1]. Available from: http://www.hc-sc.gc.ca/ewh-semt/air/out-ext/reg-eng.php 61. Henderson SB, Beckerman B, Jerrett M, Brauer M. Application of Land Use Regression to Estimate Long-Term Concentrations of Traffic-Related Nitrogen Oxides and Fine Particulate Matter. Environ. Sci. Technol. 2007;41(7):2422–8. 62. Larson T, Su J, Baribeau A-M, Buzzelli M, Setton E, Brauer M. A Spatial Model of Urban Winter Woodsmoke Concentrations. Environ. Sci. Technol. 2007;41(7):2429–36. 63. Seaton A, Dennekamp M. Hypothesis: Ill health associated with low concentrations of nitrogen dioxide--an effect of ultrafine particles? Thorax. 2003;58(12):1012–5. 64. World Health Organization. Air quality guidelines, global update: particulate matter, ozone, nitrogen dioxide and sulfur dioxide [Internet]. Geneva, CH: World Health Organization; 2005. [cited 2011 Feb. 15]; Available from: http://whqlibdoc.who.int/hq/2006/WHO_SDE_PHE_OEH_06.02_eng.pdf 65. Delfino RJ, Sioutas C, Malik S. Potential Role of Ultrafine Particles in Associations between Airborne Particle Mass and Cardiovascular Health. Environ Health Perspect. 2005;113(8):934–46. 66. Sardar SB, Fine PM, Yoon H, Sioutas C. Associations between particle number and gaseous co-pollutant concentrations in the Los Angeles Basin. Journal of the Air & Waste Management Association. 2004;54(8):992–1005. 67. Clougherty JE, Wright RJ, Baxter LK, Levy JI. Land use regression modeling of intra-urban residential variability in multiple traffic-related air pollutants. Environ Health. 2008;7:17. 68. Vinzents PS, Møller P, Sørensen M, Knudsen LE, Hertel O, Jensen FP, et al. Personal Exposure to Ultrafine Particles and Oxidative DNA Damage. Environ Health Perspect. 2005;113(11):1485–90. 69. Kanaroglou PS, Jerrett M, Morrison J, Beckerman B, Arain MA, Gilbert NL, et al. Establishing an air pollution monitoring network for intra-urban population exposure assessment: A location-allocation approach. Atmospheric Environment. 2005;39(13):2399– 409. 70. Hornung RW, Reed LD. Estimation of Average Concentration in the Presence of Nondetectable Values. Applied Occupational and Environmental Hygiene. 1990;5(1):46–51. 71. Larson T, Henderson SB, Brauer M. Mobile Monitoring of Particle Light Absorption Coefficient in an Urban Area as a Basis for Land Use Regression. Environ. Sci. Technol. 2009;43(13):4672–8.  96  72. DMTI Spatial Inc. DMTI Spatial Inc. CanMap® Streetfiles: User manual v2010.3 [Internet]. 2010. [cited 2011 Jan 5]; Available from: http://abacus.library.ubc.ca/jspui/handle/10573/42371 73. Su JG, Brauer M, Ainslie B, Steyn D, Larson T, Buzzelli M. An innovative land use regression model incorporating meteorology for exposure analysis. Science of The Total Environment. 2008;390(2-3):520–9. 74. Brauer M, Hoek G, van Vliet P, Meliefste K, Fischer P, Gehring U, et al. Estimating LongTerm Average Particulate Air Pollution Concentrations: Application of Traffic Indicators and Geographic Information Systems. Epidemiology. 2003;14(2):228–39. 75. Ross Z, Jerrett M, Ito K, Tempalski B, Thurston GD. A land use regression for predicting fine particulate matter concentrations in the New York City region. Atmospheric Environment. 2007;41(11):2255–69. 76. Hochadel M, Heinrich J, Gehring U, Morgenstern V, Kuhlbusch T, Link E, et al. Predicting long-term average concentrations of traffic-related air pollutants using GIS-based information. Atmospheric Environment. 2006;40(3):542–53. 77. Gilbert NL, Goldberg MS, Beckerman B, Brook JR, Jerrett M. Assessing Spatial Variability of Ambient Nitrogen Dioxide in Montreal, Canada, with a Land-use Regression Model. J. Air & Waste Manage. Assoc. 2005;55:1059–63. 78. Jerrett M, Arain MA, Kanaroglou P, Beckerman B, Crouse D, Gilbert NL, et al. Modeling the intraurban variability of ambient traffic pollution in Toronto, Canada. J Toxicol Environ Health A. 2007;70(3-4):200–12. 79. Moore DK, Jerrett M, Mack WJ, Künzli N. A land use regression model for predicting ambient fine particulate matter across Los Angeles, CA. J. Environ. Monit. 2007;9(3):246– 52. 80. Poplawski K, Gould T, Setton E, Allen R, Su J, Larson T, et al. Intercity transferability of land use regression models for estimating ambient concentrations of nitrogen dioxide. J Expos Sci Environ Epidemiol. 2008;19(1):107–17. 81. Allen RW, Amram O, Wheeler AJ, Brauer M. The transferability of NO and NO2 land use regression models between cities and pollutants. Atmospheric Environment. 2011;45(2):369–78. 82. Arain MA, Blair R, Finkelstein N, Brook JR, Sahsuvaroglu T, Beckerman B, et al. The use of wind fields in a land use regression model to predict air pollution concentrations for health exposure studies. Atmospheric Environment. 2007;41(16):3453–64.  97  83. Johansson C, Norman M, Gidhagen L. Spatial & temporal variations of PM10 and particle number concentrations in urban air. Environmental Monitoring and Assessment. 2006;127(13):477–87.  98  Appendices  Note: Appendices A, C and E – H are electronic appendices. These appendices are spreadsheets with project data. A brief written description of each file is provided below. Appendix A Metro Vancouver ultrafine particle database  This appendix is located in the file “Metro_Vancouver_UFP_database_2010.xls” This file contains all of the measurements made during the spring 2010 UFP sampling campaign. The first tab, “1. Sample UFP Data” includes the spatial UFP measurements from the 80 sites (plus 20 repeats). This includes the date, site identification, measured coordinates of the site, description of the site, instrument used, start and end time of the sample, raw measurements during each minute of sampling, trucks in sample, wind direction during sample, fast food locations, construction code, as well as statistics on data from each sample (minimum concentration, 5, 10, 25, 50, 75, 90, 95%ile, mean and standard deviation) in addition to the three correction factors (time of day, day of sample, instrument) individually and compositely applied. The second tab, “2. Sample meteorologic data” includes direction (of wind), wind speed, temperature, relative humidity and altitude measured during sampling. Calculations of mean sample temperature (used as a variable in LUR modeling) as well as maximum wind rose length (and 0.125, 0.25, 0.50, 0.75 fractions) for each site are also included. The third tab, “Fixed site UFP data” includes all raw data from the 4 fixed sites as well as calculation of the within (hour of day) and between day (day of sample) correction factors, used to temporally adjust mobile data. 99  The fourth tab, “Instrument 4 versus 2” contains collocation data for the two CPC 3007 instruments used in mobile monitoring, which was used to develop the instrument correction factor. The fifth tab, “CPC and P-trak comparisons” contains collocation data for a CPC 3007 instrument and P-trak 8525, which was used to develop the between-instrument correction factor. Table A.1 Sampling site locations and site PNC measurements (N=80)  Site ID X Y 2 490416 5463965 3 494066 5463943 4 496650 5463946 6 494505 5462440 7 496478 5461852 8 512888 5460432 10 495601 5459427 11 498971 5459429 12 514925 5459392 13 490002 5458928 15 496980 5458455 16 502058 5458479 17 510886 5458415 18 518562 5458438 19 490985 5457925 20 485015 5457380 22 489495 5457456 23 495003 5457439 25 503489 5457513 27 514011 5457519 28 516571 5460031 29 498027 5457006 31 491990 5456397 32 493505 5456489 33 485899 5455956 35 495871 5456049 37 497530 5455468 38 506961 5455418 39 509533 5455456  PNC50 Ln-PNC50 PNC10 PNC90 3 3 3 (pt/cm ) (pt/cm ) (pt/cm ) (pt/cm3) 19144 9.9 17118 20160 7652 8.9 5239 15888 15886 9.7 13248 18731 12291 9.4 6637 19757 16280 9.7 13932 18481 6066 8.7 4484 7838 13861 9.5 9784 26406 18726 9.8 6407 49653 7772 9.0 6803 9782 9285 9.1 7145 13316 21999 10.0 19120 40573 4510 8.4 3423 6525 7400 8.9 5063 14509 2655 7.9 1770 4226 22599 10.0 8038 31956 20706 9.9 12601 77282 11473 9.3 9180 14981 45636 10.7 28995 78963 10980 9.3 9291 12078 4297 8.4 2937 6140 20098 9.9 16260 24557 7702 8.9 6210 10310 44891 10.7 36031 65570 26278 10.2 20042 44603 7088 8.9 5054 8078 13274 9.5 9890 20347 29661 10.3 21811 58043 12910 9.5 11952 14643 14651 9.6 9410 36423  100  Table A.1 Sampling site locations and site PNC measurements (N=80)  Site ID X Y 40 491940 5454909 41 520994 5454686 42 486515 5454517 44 499443 5454408 46 516003 5454428 47 488544 5453962 48 494428 5453972 49 497414 5453989 50 490439 5453478 52 508537 5453234 55 494471 5452403 56 500404 5452408 59 503042 5451946 60 491594 5451418 61 498071 5451478 62 530425 5451493 63 493482 5450945 68 512555 5450387 69 514584 5449456 70 509839 5448950 71 511994 5447958 72 507525 5447362 74 490437 5446388 76 516532 5445999 79 490056 5445458 80 512051 5445340 82 509839 5448950 83 486897 5443953 85 490374 5442931 86 491999 5442965 87 506984 5442442 88 522552 5442364 89 486452 5441941 90 509425 5441892 92 524491 5439415 93 531368 5437889 96 514954 5431440 98 512486 5430950 99 514970 5429927 100 494177 5428967  PNC50 Ln-PNC50 PNC10 PNC90 3 3 3 (pt/cm ) (pt/cm ) (pt/cm ) (pt/cm3) 7947 9.0 4818 10716 22773 10.0 15609 48105 7873 9.0 5409 11933 9709 9.2 7517 11649 14673 9.6 13328 16773 17518 9.8 14565 25425 58871 11.0 25298 120766 4812 8.5 16043 21843 6858 8.8 2431 29290 42352 10.7 29199 61448 21121 10.0 16089 32338 15767 9.7 13719 22016 14120 9.6 11268 17631 7206 8.9 5516 10731 19167 9.9 10768 26590 16912 9.7 13610 22296 23353 10.1 17794 34305 19504 9.9 16474 21142 21845 10.0 11295 44402 9309 9.1 6595 14559 29607 10.3 26008 35317 13975 9.5 11548 15731 35895 10.5 23700 57387 11311 9.3 10188 15116 24276 10.1 14923 46466 17012 9.7 13968 23955 14931 9.6 11823 21176 6323 8.8 5422 9331 5516 8.6 4553 8000 1473 7.3 1096 2317 12658 9.4 11362 14012 24972 10.1 15297 51353 2310 7.7 1054 4080 10893 9.3 7638 20548 18110 9.8 13710 30417 5727 8.7 4384 7095 11971 9.4 10722 13808 10814 9.3 8647 19206 8048 9.0 7611 10410 7880 9.0 6467 9244 101  Table A.1 Sampling site locations and site PNC measurements (N=80)  PNC50 Ln-PNC50 PNC10 PNC90 Site ID X Y (pt/cm3) (pt/cm3) (pt/cm3) (pt/cm3) 101 519675 5439863 13480 9.5 9438 26356 102 489749 5450709 43150 10.7 30004 62277 103 489930 5456767 43073 10.7 29695 61336 104 491451 5450831 47856 10.8 33632 81058 105 491555 5453533 37137 10.5 18216 67853 106 494406 5451604 22964 10.0 15200 35107 107 494407 5458937 105000 11.6 72288 169032 108 516809 5450585 4602 8.4 3884 6905 110 507518 5449502 8864 9.1 4698 16903 113 547943 5434160 4455 8.4 3729 9190 114 552095 5435382 4297 8.4 3831 12832 Table A.2 Repeat PNC measurements (N=20)  Site ID 3 11 17 22 23 31 31 33 38 40 49 50 59 60 70 79 85 103 103  PNC50 (pt/cm3) 7514.450802 10557.70521 12724.34099 13787.84694 35677.55335 10744.69571 21500.85245 9335.833256 11873.76788 8144.88406 18378.06365 26138.24295 10377.92845 6400.095252 18972.10104 4349.820324 9428.899416 34498.77198 43390.61669  102  Appendix B Ultrafine particle sampling campaign manual  103  Metro
Vancouver
Ultrafine
Particulate
 Sampling
Campaign
Manual
 Spring
2010
 
 Prepared
by:
Rebecca
Abernethy
 School
of
Environmental
Health
 University
of
British
Columbia
 
  
 104  
  
 April
9,
2010
 
 Thank
you
for
agreeing
to
be
a
part
of
the
Spring
2010
Ultrafine
Particulate
Matter
Sampling
 Campaign
across
Metro
Vancouver.
This
is
a
student
project
based
out
of
the
University
of
 British
Columbia.
 
 This
study
will
provide
important
new
research
to
the
region
in
terms
of
the
sources
of
a
 particular
type
of
air
pollution,
ultrafine
particles,
and
their
exposure
to
humans.
Briefly,
the
 objective
of
this
project
is
to
measure
concentrations
of
ultrafine
particles
across
Metro
 Vancouver
and
to
develop
a
predictive
model
to
capture
fine
spatial
variation
in
ambient
levels.
 
 Please
familiarize
yourself
with
all
items
in
this
manual
and
ensure
you
are
confident
with
all
 procedures
and
pieces
of
equipment.
Follow
the
protocol
and
schedule
strictly;
I
must
approve
 ANY
changes.
However,
providing
feedback
throughout
the
sampling
period
on
methods
is
 encouraged.
 There
is
a
timesheet
in
this
package
for
you
to
record
the
hours
worked.
While
the
schedule
has
 been
made
and
it
is
estimated
that
work
hours
will
be
from
9am
–
5pm
each
day,
it
is
important
 that
actual
hours
are
recorded
and
submitted,
as
that
is
what
your
compensation
will
be
based
 on.
You
will
be
provided
with
an
honourarium
once
the
sampling
period
is
complete
to
thank
 you
for
your
contribution.
 
 Your
commitment
to
this
project
is
appreciated.
Please
let
me
know
if
you
have
any
concerns
or
 feedback,
otherwise
enjoy
your
time
in
the
field!
 
 Sincerely,
 
 Rebecca
Abernethy
 MSc
Student
 School
of
Environmental
Health
 University
of
British
Columbia
 
 rebecca.abernethy@gmail.com
 604‐999‐0510
 
 
 
 
 
 
 
 
 105  Overall
Mobile
Sampling
Protocol
 
 Please
complete
all
of
these
steps
at
each
site.
More
detailed
operating
instructions
for
the
TSI
 3007
Condensation
Particle
Counter,
Kestrel
4500
Pocket
Weather
Tracker
and
GPS
devices
 follow.
 
  Equipment
Checklist
 
 Please
ensure
you
have
all
items
before
leaving
for
the
day.
 
 Sampling
Equipment
 • CPC
3007
 o alcohol
cartridge
 o isopropyl
alcohol

in
small
container
 o (there
is
a
possibility
each
team
will
get
a
spare
3007
part
way
through
sampling)
 • 3007
tripod
mounting
case
 • 3007
tripod
 • Zero
filter
for
3007
 • Kestrel
4500
(and
spinning
device
and
weather
vane)

 • 4500
tripod
 
 Miscellaneous
 • GPS

 • Digital
camera
 • Extra
AA
batteries
 • Extra
AAA
batteries
 • Stopwatch
 • 2
traffic
counters

 • Binder
 o sampling
manual
 o log
sheets
 o site
satellite
photos
(Google
Earth
images)
 o schedule
with
sites
 o list
of
site
addresses/intersections
 • Clipboard
 • Pens
 • Sharpie
marker
 • CPC
3007
instrument
manual
 • Kestrel
4500
instrument
manual
 106  • • •  2
High
visibility
vest
 Cell
phone
 Rebecca’s
cell
number:
(604)
999‐0510
 
  Field
Sampling
Procedure
 
  The
primary
data
collection
period
for
this
project
will
occur
between
9:00
am
and
5:00
pm,
 weekdays.

Please
plan
your
days
accordingly.
 Please
double
check
with
Rebecca
BEFORE
making
changes
to
the
timing
of
sampling
or
 anything
else
to
do
with
the
schedule..

 1. Put
on
your
high
visibility
vests.
 2. Identify
the
next
site
you
will
be
visiting
based
on
the
schedule.
Refer
to
the
list
of
sites
 for
reference.
Please
plan
out
your
route
the
night
before,
in
the
order
outlined
on
the
 schedule
(use
Google
Maps,
GPS
navigation
etc
to
help
you).
Drive
to
the
site.
 3. Place
the
alcohol
wick
in
and
turn
the
3007
on
when
you
estimate
you
are
about
10
 minutes
away
from
the
site,
so
that
it
is
warmed
up
and
ready
to
go
upon
arrival.
(See
 3007
details
for
how
to
do
this)
 4. Pull
up
to
the
site.
Look
for
the
Ogawa
sampler
which
will
already
be
in
place
at
some
 sites.
It
will
be
on
the
pole
identified
in
the
site
description
about
2.5m
above
ground
in
 a
little
white
cup
(the
cup
is
about
10cm
high,
so
it’s
fairly
small).
See
photo.
 a. NOTE:
It’s
a
good
idea
to
park
so
that
you
have
a
good
view
of
the
road
on
which
 you
will
be
counting
traffic.
 
  Installed
Ogawa
sampler.
 
  
  5. Record
the
Site
ID
on
the
field
log.
 6. Prepare
for
sampling.
 a. See
operating
instructions
for
the
CPC
3007,
attached.
 b. See
operating
instructions
for
the
Kestrel
4500,
attached.
 107  i. Ensure
the
CPC
3007
and
Kestrel
4500
are
facing
traffic
and
facing
the
 same
direction.
Indicate
the
direction
the
instruments
are
facing
on
the
 Google
Earth
image
for
that
site.
 7. Start
the
3007
logging,
by
selecting
Log
Mode
1
and
shut
the
case.
Ensure
the
4500
is
 also
recording
by
double‐checking
that
Auto
Store
under
Memory
Options
is
on
(see
 4500
instructions).
Record
the
sample
start
time
on
the
field
log.
Allow
the
devices
to
 run
for
one
hour.
 8. Complete
the
log
sheet
for
the
site:
 a. Record
notes
on
any
interesting
activity
or
nearby
buildings/facilities.
Take
 photos
in
all
directions
of
the
surrounding
site
as
well
as
sampling
equipment.
 Label
the
template
roads
and
indicate
where
the
sample
site
is
on
the
sketch.
 b. Record
the
direction
the
wind
is
coming
from
and
whether
the
sampler
is
upwind
 or
downwind
of
the
road.
 i. Using
a
sharpie
marker,
indicate
the
direction
of
the
wind
on
the
zoomed
 in
Google
Earth
image
for
that
site.
 c. Record
the
name,
location
(ie
“intersection
of
Main
&
Maple”),
and
approximate
 distance
to
any
fast
food
restaurants.
 i. Using
a
sharpie
marker,
indicate
the
location
of
any
nearby
fast
food
 restaurants
on
the
zoomed
in
Google
Earth
image
for
that
site.
 d. Record
any
necessary
drawings
on
the
Google
Earth
images.
If
things
are
 significantly
different
from
the
Google
Earth
image,
draw
nearby
building
and
 activities
on
the
sketch
template,
on
the
log
sheet.
 e. Use
the
hand‐held
clickers
to
count
diesel
truck
and
bus
traffic

travelling
in
both
 directions:
 i. Use
the
stopwatch
to
make
sure
you
count
for
5
minutes.
Count
the
 trucks/buses
on
the
busiest
road
adjacent
to
the
samplers
(ie
the
busiest
 road
near
the
site
that
is
not
blocked
by
buildings)
and
mark
this
road
on
 the
sketch.
 ii. Record
the
number
of
trucks/buses
and
the
counting
duration
(this
 should
be
5
minutes
under
normal
circumstances).
 iii. Repeat
the
5‐min
diesel
truck/bus
count
a
second
time.

The
two
5‐ minute
counts
should
be
conducted
at
least
30
minutes
apart
(to
best
 capture
average
conditions
over
the
hour
of
ultrafine
particle
sampling).
 9. After
1
hour
stop
the
3007
logging
session.
 10. Turn
off
the
3007
and
4500.
 a. Remove
the
wick
from
the
3007
and
immerse
it
in
the
alcohol
container
during
 travel.
Place
the
“for
storage
only”
cap
in
the
instrument
for
travel.
 11. Return
all
equipment
to
the
car.
 12. Go
to
the
next
site.
 13. Repeat
above.
 14. At
the
end
of
the
day,
drivers:
download
CPC
3007
data,
4500
data
and
photos
and
 send
to
Rebecca.
 108  a. For
the
3007:

 i. Install
the
software
on
a
PC.
Connect
device
to
the
computer,
turn
it
on
 and
open
the
Aerosol
Instrument
Manager
Program.

 ii. Click
‘New
File’
and
save
it
in
the
following
format:
“mmddyyUBC#‐ siteID#”
For
example,
a
file
collected
April
20,
2010
with
instrument
UBC
 #5
at
site
ID
15
would
be
saved
as
042010UBC5‐15.
 iii. Then
go
to
Run
‐>
Receive
Logged
Data,
Select
All
and
each
file
for
each
 site
will
come
up.
You
can
scroll
along
the

 iv. Go
to
File
‐>
Export
and
select
.xls
and
Time
Stamp
under
data
point
time
 format.
Save
the.xls
files
as
“April
19,
UBC
#5,
Site
18”
as
there
will
be
 one
file
per
site.
Email
all
files
to
Rebecca.
 v. There
is
a
computer
manual
that
should
help
with
any
problems,
or
ask
 Rebecca.
 b. For
the
4500:
download
the
software
from
 http://www.nkhome.com/kestrel/kestrel‐4500/more.php
(under
software).
 Install
the
program
and
follow
the
manual
attached
to
download
data.
 c. Please
label
all
photos
the
same
way
(ie
042010UBC5‐15).
 Please
record
all
file
names
for
the
3007,
4500
and
photos
on
the
log
sheet
for
 the
appropriate
site.
 15. Ensure
batteries
are
charged
for
the
next
day
and
that
all
equipment
is
in
working
 condition.
 16. Arrange
to
meet
with
your
teammate
for
the
next
day
and
if
necessary,
pass
the
 equipment
on
to
the
next
driver.
 a. A
complete
contact
list
is
included
in
this
manual
and
is
also
on
the
schedule.
  
 ***If
community
members
ask
you
questions,
please
try
to
answer
them
if
you
feel
able.
If
you
 do
not
know
an
answer,
please
do
not
make
something
up

‐
just
direct
them
to
call
Rebecca
 (604‐999‐0510).
 ***Please
be
respectful
of
people
in
the
community
you
are
in
and
activities
going
on.
If
you
 are
asked
to
relocate,
please
do
so.
If
you
feel
unsafe
at
any
time
then
you
should
also
relocate.
 The
sampler
can
be
stopped
(make
a
note
of
when
on
the
logsheet)
and
restarted
in
a
different
 area.
Take
another
GPS
waypoint
as
well.
 ***I
would
much
rather
have
too
much
information
and
notes
than
too
little
–
you
will
have
a
 lot
of
time
while
the
instrument
is
sampling.
Please
be
thorough
in
your
site
descriptions,
 photos
and
notes.
 
 
  Operation
of
the
TSI
3007
Condensation
Particle
Counter
 
 *Note:
this
is
a
summary.
For
details
on
instrument
operation
refer
to
the
Operation
and
Service
 Manual
or
call
Rebecca!
 
 109  1. Place
the
alcohol
wick
in
the
3007.
 a. The
3007
should
be
OFF.
 b. Open
the
clear
alcohol
capsule
and
set
the
alcohol
wick
on
a
clean
surface
with
 the
end
standing
up.
 c. Squeeze
alcohol
into
the
fill
capsule
until
the
liquid
level
is
even
with
the
fill
line
 near
the
base.
 d. After
making
sure
the
alcohol
cartridge
is
clean,
insert
the
alcohol
cartridge
into
 the
fill
capsule
and
lock
it
into
place.
 e. Wait
a
few
minutes
while
the
wick
inside
the
cartridge
soaks
up
alcohol.
 f. Remove
the
alcohol
cartridge
from
the
fill
capsule
and
gently
shake
it
to
allow
 excess
alcohol
to
drain
back
into
the
capsule.
 g. Insert
the
cartridge
into
the
3007
and
lock
into
place.
  
  The
3007
with
the
“for
storage
only”
cap
in
place,
to
be
used
during
transit,
and
wick
soaking
up
alcohol.
  
 2. Turn
the
device
on
and
allow
it
to
warm
up
(10
min).
It
will
count
down
from
600s
and
 you
will
not
be
able
to
do
anything
until
this
countdown
is
complete.
 3. Ensure
enough
battery
life
remains
for
sampling.
 a. Ensure
the
spare
set
is
charged;
charge
drained
batteries
in
vehicle
while
 sampling.
 4. Zero
the
instrument.
 a. Attach
the
zero
filter
assembly
to
the
inlet.
 b. The
particle
concentration
should
go
to
zero
in
approximately
5‐10
seconds.

 Leave
the
zero
filter
attached
to
the
instrument
for
30
seconds
to
make
sure
the
 zero
reading
is
stable.
 c. Record
the
30s
zero
reading
on
the
log
sheet
for
the
first
site
of
the
day.
 d. Remove
the
zero
filter.

The
instrument
is
now
ready
for
operation.
 e. If
there
are
problems,
refer
to
the
troubleshooting
page
in
the
manual
(31).
 i. If
the
device
does
not
go
to
zero,
add
an
additional
zero
assembly.
 Readings
of
up
to
10
are
ok,
but
ensure
you
record
the
reading
on
the
 logsheet.
Anything
above
this,
please
call
Rebecca.
 110  3007
with
zero
filter
assembly
attached.
  
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
  
  5. Prepare
for
sampling.
 a. Place
the
3007
in
its
case
on
the
tripod
where
sample
will
be
collected
from,
 pointing
the
aerosol
inlet
toward
the
road.
 b. The
3007
should
be
approximately
4’
off
the
ground
(i.e.
the
tripod
should
be
 fully
extended).
  
  3007
in
case
on
fully
extended
tripod.
 
  c. 3007
should
be
placed
on
the
edge
of
the
road
(on
a
sidewalk/median
strip
if
 possible)
out
of
the
way
of
oncoming
vehicles
and
pedestrians.
 d. Press
down
arrow
to
‘setup.’
 i. Select
log
internal.
Ensure
it
is
at
1:00
minute.

 6. Start
sampling.
 a. Press
down
arrow
to
'log
mode
1'
on
main
menu
and
hit
enter
(bottom
bent
 arrow
key).

 111  
 1” – select to start “Log mode 
 
 enter key. Press logging with 
 enter key 
again to stop logging 
 after 1 hr.
 
 
 
 
 
 
 
 
 
 1. 2. 3.  
  “enter” key – used to select items  
 b. Sample
for
one
hour.
At
the
end
of
sampling,
press
the
enter
key
to
stop.
  Note:

Always
keep
the
3007
level
during
sampling
 You
can
review
the
sample
information
by
selecting
'sample'
in
the
main
menu,
then
'
 statistics'
in
the
next
menu.
Select
'logged
test
stats'
and
hit
enter.
Use
the
up
and
down
 arrow
keys
to
scroll
through
saved
tests.
 Hit
enter
to
stop
logging.
 Press
the
power
button
to
turn
off.

 Remove
alcohol
cartridge
and
replace
with
the
storage
cap
when
moving
to
the
next
 site/ending
for
the
day.
  Operation
of
the
Kestrel
4500
Pocket
Weather
Tracker
 
 *Note:
this
is
a
summary.
For
details
on
instrument
operation
refer
to
the
Operation
and
Service
 Manual
or
call
Rebecca!
 
 1. Turn
the
device
on
(the
circle
button
with
red
on
it).
 
 
 
 
 
 
 
 
 “on” and selection button 112  
  
 
 
 
 
 
 
 
 
 
 
  
 2. Ensure
the
device
has
sufficient
battery
power
(>80%
at
the
start
of
the
day).
 3. Calibrate
the
compass.
 a. Place
the
4500
upright
in
its
box
and
rotate
around
three
times,
slowly
(each
 rotation
should
be
10s).
Do
not
calibrate
on
metal
surfaces
and
ensure
the
 device
is
straight
the
whole
time.
 b. The
device
will
tell
you
once
calibration
is
complete;
if
it
is
unsuccessful,
refer
to
 the
manual
(page
5‐7)
for
troubleshooting.
 c. The
device
must
be
calibrated
every
time
the
battery
compartment
is
opened
or
 batteries
replaced.
Otherwise,
once
at
the
beginning
of
the
day
is
sufficient.
 
  
 
 
 
 Place
device
upright
in
case
on
a
flat
surface
and
rotate
around
three
times
to
calibrate.
 
 4. Verify
that
the
device
is
logging
and
that
the
time
and
date
are
correct.
 a. Press
the
‘on’
key
for
the
menu.

 b. Select
‘Memory
Options’
with
the
centre,
oval
button
and
ensure
‘Auto
Store’
is
 on
and
that
the
‘Store
Rate’
is
5
minutes.
  113  
  (top)
Menu
showing
Memory
Options.
(bottom)
Under
Memory
Options,
ensure
Auto
Store
is
on.
  
 
  
 5. Affix
the
4500
to
the
tripod
casing,
ensuring
the
device
is
fully
upright
and
facing
into
 the
traffic
(same
direction
as
the
inlet
on
the
3007).
 6. After
sampling,
turn
the
device
off.
 a. Note:
the
Kestrel
will
turn
its
screen
off
automatically
after
about
10
minutes.
As
 long
as
it
says
“Auto
Store”
on
at
the
start
and
end
of
sampling,
it
is
recording.
  Operation
of
the
GPS
 
  
  1. Turn
GPS
on.
 2. Ensure
it
has
adequate
battery
power.
 3. At
the
sampling
location,
save
a
way‐point.
Note
the
point
number
and
 latitude/longitude
down
on
the
logsheet.
  114  Community
Ultrafine
Particle
Sampling
Field
Log
 Date:____________________








CPC
3007
instrument
ID:
UBC
#
____



Kestrel
4500
ID:
UBC
#
____
GPS
ID:
______
 Site
ID:
_______




Site
Description____________________________________________________________________
 CPC
3007
file
name:
_______________


Kestrel
4500
file
name:
__________


CPC
3007
zeroing
reading:
___________
 Photo
file
names:
___________________________________________________________________________________
 
  TRAFFIC
COUNTS
 3007
 Start
 Time
  3007
 Stop
 Time
  :
  :
  Road
Name(s)
  Count
1
 Trucks
  Count
1
 Minutes
  
  
  :
  
  
  
  Count
2
 Trucks
  
 
  WIND
 Sampler
upwind
 Direction
wind
is
 or
downwind
of
 Count
2

 coming
from
 Minutes
 road?

 [N,NE,
E,…]
 [U,
DW]
  :
 
  NEARBY
FAST
FOOD
RESTAURANTS
 Name(s)
 

  Location(s)
  GPS
Point
#
 Approx.
 and
Latitude‐ Distance(s)
 Longitude
 (m)
  
  
  
  
  
  
  
  
  
  
  
  
  Codes:

McDonald’s
=
McD;
Burger
King
=
BK;
KFC;
Tim
  
  If
the
sampler
is
located
near
the
intersection
of
two
 roads
you
should
count
the
trucks
on
both
roads.
  Also
record
the
wind
direction
on
 the
Google
Earth
image.
  Horton’s
=
TH;
A&W
=
AW;
Church’s
Chicken
=
CC,
 Dairy
Queen
=
DQ,
Wendy’s
=
WE;
Other
=
O

 Also
record
the
location(s)
of
nearby
fast
food
 restaurant(s)
on
the
Google
Earth
image.
  
 Site
Notes
(and
sketch
if
different
from
Google
Earth
image):
  115  
  Vehicle
Use
Tracking
Log
 Please
keep
accurate
record
of
the
kilometers
driven
per
day
for
ultrafine
sampling
purposes.
 At
the
end
of
the
sampling
period,
submit
this
form
to
be
compensated
for
vehicle
use
 ($0.41/km),
per
UBC
vehicle
use
policy.
 ***Zipcar
use
will
be
reimbursed.
Please
keep
receipts
for
each
day
but
also
indicate
on
this
 sheet
days
used.
 Vehicle
owner/driver
name:
___________________________
Date
submitted:
 __________________
 
 Date
driven
 Odometer
 Odometer
 Kilometers
 Description
of
 Notes
 reading
start
 reading
end
 driven
(km)
 route
 (km)
 (km)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 116  Time
Sheet
 
 Please
keep
accurate
record
of
the
number
of
hours
worked
per
day.
You
will
be
paid
an
 honourarium
at
the
end
of
the
sampling
period
according
to
number
of
hours
worked.
 Note
that
a
standard
sampling
workday
will
be
from
9am
–
5pm.
If
your
day
will
be
different
 than
that,
particularly
if
it
will
be
longer,
please
call
Rebecca
THAT
DAY
(604‐999‐0510)
to
 discuss
why.
 Name:
___________________________
Date
submitted:
__________________
 
 Date
worked
 Start
time
 End
time
 Hours
worked
 Description
of
 Notes
 route
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 117  Team
Members
 
 Name
 Phone
Number
 Email
Address
 Rebecca
 
 
 Christie
 
 
 Rehema
 
 
 Sara
C
 
 
 Brian
 
 
 Sarah
 
 
 Mary
 
 
 *phone
numbers
and
emails
were
shared
amongst
team
members
to
facilitate
sampling
  118  Appendix C Ultrafine particle and nitrogen oxides comparisons  This appendix is located in the file “UFP_NOx_comparisons.xls” This file contains the average NO, NO2 and NO2 values sampled at one hundred fourteen sites during the same time period as UFP sampling was occurring at eighty of the same sites. UFP median values at the same sites are included in the database.  Table C.1 UFP and NOx comparison data  Site ID 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28  Notes  PNC50 (pt/cm3) 19144 7652 15886 12291 16280 6066 13861 18726 7772 9285 21999 4510 7400 2655 22599 20706 11473 45636 10980 4297 20098  NO2 (ppb) NO (ppb) NOx (ppb) 4.60 4.29 8.89 9.00 6.15 15.15 6.95 6.11 13.06 6.37 5.50 11.87 10.03 7.72 17.75 12.41 13.66 26.07 9.43 9.82 19.25 4.46 5.36 9.81 15.10 10.19 25.29 11.47 16.00 27.47 9.23 7.91 17.13 8.01 5.53 13.55 14.03 10.13 24.16 14.24 23.67 37.91 10.70 7.60 18.30 7.07 4.90 11.97 13.00 18.65 31.65 4.62 2.59 7.22 15.55 10.45 25.99 8.84 22.11 30.95 10.09 7.29 17.38 9.70 14.62 24.32 18.96 37.12 56.08 9.36 4.44 13.80 8.65 3.20 11.85 6.40 4.10 10.50 5.38 8.83 14.22 6.87 8.88 15.75 119  Table C.1 UFP and NOx comparison data  Site ID 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68  Notes  PNC50 (pt/cm3) 7702 44891 26278 7088 13274 29661 12910 14651 7947 22773 7873  NO2 (ppb) NO (ppb) NOx (ppb) 14.49 12.65 27.14 7.56 9.24 16.80 13.09 13.04 26.12 12.31 24.58 36.89 7.48 6.39 13.86 6.56 7.75 14.31 14.15 16.76 30.91 6.96 10.66 17.62 12.98 17.10 30.08 10.28 20.64 30.92 8.96 14.93 23.89 7.31 8.35 15.67 5.40 9.15 14.55 6.06 11.84 17.90  missing 9709 14673 17518 58871 4812 6858 42352 21121 15767 14120 7206 19167 16912 23353  19504  11.96 5.67 6.26 6.14 13.61 8.41 8.36 6.69 16.62 7.35 8.04 8.64 5.86 8.77 5.78 10.04 6.10 7.72 4.16 14.10 9.67 8.05 7.47 8.60 5.97  7.62 7.60 8.25 10.03 16.14 6.93 14.20 8.20 33.27 8.98 9.50 12.48 23.75 13.26 7.32 8.05 11.08 5.04 6.21 37.94 9.61 12.46 12.19 8.78 7.51  19.58 13.27 14.51 16.18 29.75 15.34 22.56 14.89 49.89 16.33 17.54 21.13 29.60 22.03 13.10 18.09 17.18 12.77 10.37 52.04 19.28 20.51 19.67 17.38 13.49  120  Table C.1 UFP and NOx comparison data  Site ID 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110  Notes  PNC50 (pt/cm3) 21845 9309 29607 13975 35895  missing  NO2 (ppb) NO (ppb) NOx (ppb) 13.84 24.79 38.63 7.60 7.40 14.99 7.79 14.75 22.54 7.17 6.67 13.84 5.84 7.44 13.28 12.06 18.49 30.54 7.02 7.12 14.15  11311 24276 17012 14931 6323 5516 1473 12658 24972 2310 10893 18110 5727 11971 10814 8048 7880 13480 43150 43073 47856 37137 22964 105000 4602 8864  4.51 3.65 8.42 13.06 7.23 6.62 2.90 5.63 4.26 3.92 4.24 8.46 4.09 5.25 6.97 8.46 3.12 6.21 3.54 4.24 4.59 4.32 2.35 4.26 7.58 8.86 10.31 13.06 7.25 11.03 20.31 5.16 10.06 9.99  6.34 8.12 11.57 25.98 6.54 11.69 11.51 9.01 5.74 7.24 8.08 20.86 7.81 7.44 8.89 20.40 4.17 17.37 5.21 4.70 6.94 9.26 5.28 4.30 18.69 15.23 31.82 31.37 22.89 37.47 43.22 5.35 7.50 17.80  10.84 11.77 19.99 39.04 13.77 18.31 14.42 14.65 10.00 11.16 12.32 29.32 11.89 12.69 15.85 28.86 7.29 23.58 8.75 8.94 11.53 13.58 7.64 8.56 26.27 24.08 42.12 44.43 30.15 48.49 63.53 10.51 17.56 27.79 121  Table C.1 UFP and NOx comparison data  Site ID 111 112 113 114 115 116  Notes  PNC50 (pt/cm3)  NO2 (ppb) NO (ppb) NOx (ppb) 7.40 21.92 29.32 6.10 15.21 21.31 4455 5.12 6.31 11.43 4297 5.15 8.59 13.75 4.65 4.43 9.08 9.89 8.41 18.29  122  Appendix D Commands used in GIS variable extraction and R scripts  D.1  Geospatial modelling environment command examples GME from Spatial Ecology was used to extract values for the geographic predictor  variables (shown in Table 2.2 and 2.3) used in LUR model development. Below are examples of commands used to derive this information. Using the traditional buffer, the ‘isectpntrst’ command was used to intersect the sampling location point shapefile with the raster (with appropriate buffer). The example below is for determinging the length of highways 25m from a site. isectpntrst(in="C:\Users\rabernethy\Desktop\Thesis\3. GIS data\ufp_final_waypoints\80points.shp", raster="C:\Users\rabernethy\Desktop\Thesis\3. GIS data\Road\rd1_25_f", field="rd1_25");  For the wind rose buffers, a polygon shapefile of the wind roses from each site, with site attribute information attached, interacted with polyline files to sum the line length within the polygon, with point files to count the points within the polygon and with polygon files to calculate the overlapping area. Examples of the use of these three commands include the length of highways in a 0.125 maximum wind rose buffer, the number of bus stops in a 0.25 maximum wind rose buffer and the area of commercial land in a 1.0 maximum wind rose buffer. sumlinelengthsinpolys(line="C:\Users\rabernethy\Desktop\Thesis\3. GIS data\Road\RD1.shp", poly="C:\Users\rabernethy\Desktop\Thesis\3. GIS data\Windroses\0125\0125_merge_j80.shp", field="rd1_0125w");  123  countpntsinpolys(poly="C:\Users\rabernethy\Desktop\Thesis\3. GIS data\Windroses\025\025_merge_j80.shp", pnt="C:\Users\rabernethy\Desktop\Thesis\3. GIS data\bus_stops_march09\BusStops_09March2010_Clip.shp", field="bus_wr_025"); isectpolypoly(in="C:\Users\rabernethy\Desktop\Thesis\3. GIS data\Windroses\100\100_joinatt.shp", poly="C:\Users\rabernethy\Desktop\Thesis\3. GIS data\Landuse\landuse_GVRD\comwr.shp", field="CATNUM", prefix="cw100", thematic=TRUE); D.2  R script examples R was used to prepare and validate (through LOO cross validation) the LUR models. The following script is for the development of a LUR model, including ranking the  correlation of the variables with PNC, removing variables correlated >0.6 with the highest ranking variable in their category and creating a linear model.  # Function to return a table of rankings between input Y and the potentially predictive covariates. # Function variable group should be of form "rd1." # Function output is a list of character class, names of variables # y = character value ("pnc50", "lnpnc50", "pnc10" etc.) # ranks = data frame with three columns (variable name, univariate r2 with Y, and variable type) get_ranks_table = function(y, modeltype){ if (modeltype == "length"){ covariates = names(ourdata[7:48]) } if (modeltype == "density"){ covariates = names(cbind(ourdata[7],ourdata[20:60])) } r2 = numeric() for (variable in covariates){ model = lm(get(y)~get(variable), na.rm = T) r2 = c(r2, as.double(summary(model)$r.squared))} ranks = as.data.frame(covariates) ranks$r2 = r2 124  ranks$covariates = as.character(ranks$covariates) ranks$vartype = strsplit(ranks$covariates,"\\.") ranks$vartype = sapply(ranks$vartype, '[[', 1) ranks$vartype = sapply(ranks$vartype, paste, ".", sep='') return(ranks) } # Function correlated # Function # Function  to return the names of variables within the same group by less than 0.6 variable group should be of form "rd1." output is a list of character class, names of variables  include_in_stepwise = function(ranks){ vargroups = as.character(unique(ranks$vartype)) stepwiselist = character() for (group in vargroups){ maxr2 = max(ranks$r2[ranks$vartype == group]) maxvar = ranks$covariates[ranks$r2 == maxr2] subdata = as.data.frame(ourdata[,grep(group, names(ourdata))]) if (dim(subdata)[2] > 1){ varcor = cor(subdata)[maxvar,] valid = c(maxvar, names(varcor)[varcor < 0.6]) } else {valid = c(maxvar)} stepwiselist = c(stepwiselist, valid) } return(stepwiselist) } # Function to variables # pollutant = # modeltype = # stepmodel =  return a step-wise selected model from a set of eligible pollutant type "length" or "density" step-wise selected model of class lm.  get_stepwise_model = function(y, modeltype){ ranks = get_ranks_table(y, modeltype) stepwiselist = include_in_stepwise(ranks) variablelist = c(y, stepwiselist) newdata = ourdata[,variablelist] basemodel = lm(newdata, na.action = na.omit) stepmodel = step(basemodel, trace = FALSE) 125  return(stepmodel) } The stepmodel would be updated to only include variables with coefficients in an a priori direction and with a p value < 0.05. stepmodel = update(stepmodel, .~.-dist.portln) stepmodel = step(basemodel, trace = FALSE) summary(stepmodel)  The LOO cross validation entailed scripting R to re-create the model at each site excluding its information from the site and then predicting the concentration of the pollutant at that site from the new model. The script is shown below.  # Leave-one-out setwd("/Users/rebecca/Desktop/ufp") ufp = read.table("pn_final.csv", header=TRUE, sep=",") store<-matrix(nrow=78,ncol=3) for (i in 1:78){ ourdata=as.matrix(ufp) ourdata=ourdata[-i,] ourdata=as.data.frame(ourdata) attach(ourdata) #stepmodel=lm(pnc.50~dist.port.ln+newff.200+trk.50) # pnc_50 (length) store[i,1]=predict(stepmodel, newdata=ufp[i,]) store[i,2]=ufp[i,"pnc.50"] store[i,3]=store[i,1]-store[i,2] detach(ourdata) } store=as.data.frame(store) 126  names(store)=c('predicts','measurements','diff') write.csv(store, file=paste('LOOmeas.csv',sep='')) mean(store$diff, na.rm=T) sd(store$diff, na.rm=T) cor(store$predicts,store$measurements)  127  Appendix E 2010 Ultrafine particle sampling campaign field notes  This appendix is located in the file “UFP_sampling_campaign_field_notes.xls” This spreadsheet contains all of the notes collected during the spring 2010 UFP sampling campaign. The tab “1. Field notes” simply contains all field notes received from research assistants. The tab “2. New variables” is where data from the notes was compiled to create four new variables (number of trucks counted, upwind or downwind, number of visible fast food sites nearby and presence of construction or abnormal activity), which were used in the best LUR model.  128  Appendix F Predictor variables for PNC LUR models (‘classic buffer’)  This appendix is located in the file “pnc_final.csv” This file contains the extracted geographic predictor variable values for all of the sites, as well as PNC50, ln-PNC50, PNC10 and PNC90 values used to develop LUR models. The four variables added into the ‘best model’ (based on site observations) are included also. Category abbreviations are identified in Table 2.2.  129  Appendix G Predictor variables for PNC LUR model (based on measured wind rose data)  This appendix is located in the file “pnc_measure_wind.csv” This file contains the geographic predictor variable values for all sites (this is a subset of all variables; those included are listed in Table 2.3) extracted by wind rose shaped buffers based on wind data measured at the time of each sample.  130  Appendix H Predictor variables for PNC LUR model (based on monitored wind rose data from a central monitoring station)  This appendix is located in the file “pnc_mv_wind.csv” This file contains the geographic predictor variable values for all sites (this is a subset of all variables; those included are listed in Table 2.3) extracted by wind rose shaped buffers based on average wind data from a central monitoring station (T-18, South Burnaby) during the entire sampling period.  131  

Cite

Citation Scheme:

        

Citations by CSL (citeproc-js)

Usage Statistics

Share

Embed

Customize your widget with the following options, then copy and paste the code below into the HTML of your page to embed this item in your website.
                        
                            <div id="ubcOpenCollectionsWidgetDisplay">
                            <script id="ubcOpenCollectionsWidget"
                            src="{[{embed.src}]}"
                            data-item="{[{embed.item}]}"
                            data-collection="{[{embed.collection}]}"
                            data-metadata="{[{embed.showMetadata}]}"
                            data-width="{[{embed.width}]}"
                            async >
                            </script>
                            </div>
                        
                    
IIIF logo Our image viewer uses the IIIF 2.0 standard. To load this item in other compatible viewers, use this url:
http://iiif.library.ubc.ca/presentation/dsp.24.1-0072650/manifest

Comment

Related Items