{"@context":{"@language":"en","Affiliation":"http:\/\/vivoweb.org\/ontology\/core#departmentOrSchool","AggregatedSourceRepository":"http:\/\/www.europeana.eu\/schemas\/edm\/dataProvider","Campus":"https:\/\/open.library.ubc.ca\/terms#degreeCampus","Creator":"http:\/\/purl.org\/dc\/terms\/creator","DateAvailable":"http:\/\/purl.org\/dc\/terms\/issued","DateIssued":"http:\/\/purl.org\/dc\/terms\/issued","Degree":"http:\/\/vivoweb.org\/ontology\/core#relatedDegree","DegreeGrantor":"https:\/\/open.library.ubc.ca\/terms#degreeGrantor","Description":"http:\/\/purl.org\/dc\/terms\/description","DigitalResourceOriginalRecord":"http:\/\/www.europeana.eu\/schemas\/edm\/aggregatedCHO","FullText":"http:\/\/www.w3.org\/2009\/08\/skos-reference\/skos.html#note","Genre":"http:\/\/www.europeana.eu\/schemas\/edm\/hasType","GraduationDate":"http:\/\/vivoweb.org\/ontology\/core#dateIssued","IsShownAt":"http:\/\/www.europeana.eu\/schemas\/edm\/isShownAt","Language":"http:\/\/purl.org\/dc\/terms\/language","Program":"https:\/\/open.library.ubc.ca\/terms#degreeDiscipline","Provider":"http:\/\/www.europeana.eu\/schemas\/edm\/provider","Publisher":"http:\/\/purl.org\/dc\/terms\/publisher","Rights":"http:\/\/purl.org\/dc\/terms\/rights","RightsURI":"https:\/\/open.library.ubc.ca\/terms#rightsURI","ScholarlyLevel":"https:\/\/open.library.ubc.ca\/terms#scholarLevel","Title":"http:\/\/purl.org\/dc\/terms\/title","Type":"http:\/\/purl.org\/dc\/terms\/type","URI":"https:\/\/open.library.ubc.ca\/terms#identifierURI","SortDate":"http:\/\/purl.org\/dc\/terms\/date"},"Affiliation":[{"@value":"Science, Faculty of","@language":"en"},{"@value":"Resources, Environment and Sustainability (IRES), Institute for","@language":"en"}],"AggregatedSourceRepository":[{"@value":"DSpace","@language":"en"}],"Campus":[{"@value":"UBCV","@language":"en"}],"Creator":[{"@value":"Saraswat, Arvind","@language":"en"}],"DateAvailable":[{"@value":"2016-01-06T02:19:39","@language":"en"}],"DateIssued":[{"@value":"2015","@language":"en"}],"Degree":[{"@value":"Doctor of Philosophy - PhD","@language":"en"}],"DegreeGrantor":[{"@value":"University of British Columbia","@language":"en"}],"Description":[{"@value":"Urban air pollution is a major health and environmental concern worldwide, and the levels are extremely high in New Delhi, India. This research is focused on the spatial and temporal variability of air pollutant concentrations and its implications for population exposure in New Delhi.  Since traffic is considered a significant source of air pollutants in urban environments, robust and multiple linear regression models were used to understand the impact of local traffic flow on ambient concentrations of PM\u2082.\u2085, CO, NO and NO\u2082 at a busy intersection. To elicit the spatiotemporal variability of PM\u2082.\u2085 and its constituents (black carbon and ultrafine particles), land use regression (LUR) models were developed. Separate morning and afternoon models were developed using 136 hours (39 sites), 112 hours (26 sites) and 147 hours (39 sites) of PM\u2082.\u2085, BC and UFPN data, respectively. Finally, to understand how spatiotemporal variations in PM\u2082.\u2085 concentrations impact population exposure, a probabilistic simulation framework was developed to integrate the PM\u2082.\u2085 LUR models with time-activity data obtained from a field survey. Regression models explained about 50\u201380% variability in hourly pollutant concentrations and localized traffic flow explained up to 19% of variability on that scale. Auto-rickshaw and truck flow had a higher influence on NO\u2082 and PM\u2082.\u2085 concentrations, respectively. Independent variables in the LUR models included population density, distance from major roads, and major and minor road lengths in buffers of different radii; measurements from a fixed continuous monitoring site were also used as independent variables in the PM\u2082.\u2085 and BC models. The temporal term explained most of the variability (63\u201377%) in PM\u2082.\u2085 and BC models compared to spatial variables (4\u201316%). Exposure simulations indicate that the estimated annual average PM\u2082.\u2085 exposure (109 \u00b5g m-\u00b3) was high compared to North American or European cities. PM\u2082.\u2085 exposures were highest during the winter months (~200 \u00b5g m-\u00b3) compared to the summer months (~50 \u00b5g m-\u00b3). Ignoring mobility (i.e. exposure during transport or at work\/school locations), as is generally assumed in epidemiologic studies of long-term exposure, underestimated PM\u2082.\u2085 population exposure by about 11%.","@language":"en"}],"DigitalResourceOriginalRecord":[{"@value":"https:\/\/circle.library.ubc.ca\/rest\/handle\/2429\/56224?expand=metadata","@language":"en"}],"FullText":[{"@value":"   AIR POLLUTION IN NEW DELHI, INDIA: SPATIAL AND TEMPORAL PATTERNS OF AMBIENT CONCENTRATIONS AND HUMAN EXPOSURE   by  ARVIND SARASWAT, P.Eng. B. Tech, Indian Institute of Technology Delhi, 2003    A DISSERTATION SUBMITTED IN PARTIAL FULIFILLMENT OF  THE REQUIREMENTS FOR THE DEGREE OF  DOCTOR OF PHILOSOPHY  in  THE FACULTY OF GRADUATE AND POSTDOCTORAL STUDIES (Resource Management and Environmental Studies)    THE UNIVERSITY OF BRITISH COLUMBIA (Vancouver) December 2015   \u00a9 Arvind Saraswat, 2015 ii  Abstract  Urban air pollution is a major health and environmental concern worldwide, and the levels are extremely high in New Delhi, India. This research is focused on the spatial and temporal variability of air pollutant concentrations and its implications for population exposure in New Delhi.  Since traffic is considered a significant source of air pollutants in urban environments, robust and multiple linear regression models were used to understand the impact of local traffic flow on ambient concentrations of PM2.5, CO, NO and NO2 at a busy intersection. To elicit the spatiotemporal variability of PM2.5 and its constituents (black carbon and ultrafine particles), land use regression (LUR) models were developed. Separate morning and afternoon models were developed using 136 hours (39 sites), 112 hours (26 sites) and 147 hours (39 sites) of PM2.5, BC and UFPN data, respectively. Finally, to understand how spatiotemporal variations in PM2.5 concentrations impact population exposure, a probabilistic simulation framework was developed to integrate the PM2.5 LUR models with time-activity data obtained from a field survey.   Regression models explained about 50\u201380% variability in hourly pollutant concentrations and localized traffic flow explained up to 19% of variability on that scale. Auto-rickshaw and truck flow had a higher influence on NO2 and PM2.5 concentrations, respectively. Independent variables in the LUR models included population density, distance from major roads, and major and minor road lengths in buffers of different radii; measurements from a fixed continuous monitoring site were also used as independent variables in the PM2.5 and BC models. The temporal term explained most of the variability (63\u201377%) in PM2.5 and BC models compared to spatial variables (4\u201316%). Exposure simulations indicate that the estimated annual average PM2.5 exposure (109 \u00b5g m-3) was high compared to North American or European cities. PM2.5 exposures were highest during the winter months (~200 \u00b5g m-3) compared to the summer months (~50 \u00b5g m-3). Ignoring mobility (i.e. exposure during transport or at work\/school locations), as is generally assumed in epidemiologic studies of long-term exposure, underestimated PM2.5 population exposure by about 11%.   iii  Preface  \uf0b7 A portion of Chapter 1 was presented at the annual conference of the Air and Waste Management Association and published in the conference proceedings. I conducted the literature review and wrote the manuscript, and Prof. Milind Kandlikar provided overall guidance and critical feedback on the manuscript.   Saraswat, A. and Kandlikar, M. (2010). Air Pollution in India: A Review of Air Quality Monitoring, Trends and Sources. Air & Waste Management Association\u2019s 103rd Annual Conference, Calgary, 2010.  \uf0b7 Chapter 2 was written as an independent manuscript that will be submitted for publication to a peer-reviewed journal. I designed the research, conducted fieldwork in New Delhi, conducted data analysis and wrote the manuscript. Prof. Milind Kandlikar and Prof. Michael Brauer provided guidance and feedback on model development and the manuscript. Prof. Geetam Tiwari provided technical inputs related to the study design and logistical support for traffic data collection.  \uf0b7 Chapter 3 was written as an independent manuscript and has been published in a peer-reviewed journal. The study design was developed in collaboration with co-authors; Josh Apte and Julian Marshall led data collection in New Delhi while working on another project, and I conducted data analysis for LUR model development and wrote the manuscript. All co-authors provided feedback on the LUR models and on the manuscript. Adapted with permission from: Saraswat, A., Apte, J. S., Kandlikar, M., Brauer, M., Henderson, S. B. & Marshall, J. D. 2013. Spatiotemporal Land Use Regression Models of Fine, Ultrafine, and Black Carbon Particulate Matter in New Delhi, India. Environmental Science & Technology, 47, 12903-12911. Copyright (2013) American Chemical Society. \uf0b7 Chapter 4 was written as an independent manuscript that has been submitted for publication to a peer-reviewed journal. I designed the research, conducted fieldwork in New Delhi, conducted data analysis and wrote the manuscript. Prof. Milind Kandlikar and Prof. Michael Brauer provided guidance and feedback on the study design and model development, and on the manuscript. Prof. Arun Srivastava provided logistical support iv  for fieldwork and feedback on the manuscript. Staff at IDS, New Delhi, provided logistical support related to conducting the time-activity survey. The following students at Jawahar Lal Nehru University provided assistance with field sampling: Sumant Srivastava and Himanshu Lal. Ethics approval (H11-00469) was obtained from the Behavioural Research Ethics Board of the University of British Columbia for conducting the time activity-survey. v  Table of Contents  Abstract .......................................................................................................................................... ii Preface ........................................................................................................................................... iii Table of Contents ...........................................................................................................................v List of Tables ................................................................................................................................ ix List of Figures .................................................................................................................................x List of Abbreviations ................................................................................................................. xiv List of Symbols .............................................................................................................................xv Acknowledgements .................................................................................................................... xvi Dedication ................................................................................................................................. xviii Chapter 1: Introduction ................................................................................................................1 1.1 Urban Air Pollution......................................................................................................... 1 1.1.1 Fine Particulate Matter ................................................................................................ 1 1.1.2 Other Important Urban Air Pollutants ........................................................................ 3 1.2 Air Pollution in India ...................................................................................................... 3 1.2.1 Air Pollution in New Delhi, India ............................................................................... 6 1.2.2 Source Contributions ................................................................................................ 11 1.3 Spatiotemporal Variability of Air Pollution & Exposure ............................................. 13 1.4 Research Objectives ...................................................................................................... 15 1.4.1 Objective 1 ................................................................................................................ 15 1.4.2 Objective 2 ................................................................................................................ 15 1.4.3 Role of Traffic (Objective 1) .................................................................................... 16 vi  1.4.3.1 Research Question #1 ....................................................................................... 16 1.4.4 Spatiotemporal Variations (Objective 1) .................................................................. 17 1.4.4.1 Research Question #2 ....................................................................................... 17 1.4.5 PM2.5 Population Exposure (Objective 2) ................................................................. 17 1.4.5.1 Research Question #3 ....................................................................................... 18 1.5 Overview of Dissertation .............................................................................................. 18 Chapter 2: The Effect of Localized Traffic Flow on Ambient Air Quality at a Busy Traffic Intersection in New Delhi ............................................................................................................20 2.1 Introduction ................................................................................................................... 20 2.2 Materials and Methods .................................................................................................. 23 2.2.1 Hourly Pollutant Concentrations............................................................................... 23 2.2.2 Weather Data ............................................................................................................ 24 2.2.3 Traffic Data Collection ............................................................................................. 24 2.2.4 Nighttime Traffic Data .............................................................................................. 26 2.2.5 Traffic Data Coding .................................................................................................. 27 2.2.6 Missing Flow Estimation .......................................................................................... 27 2.2.7 Missing Value Estimation by Random Sampling ..................................................... 28 2.2.8 Regression Models .................................................................................................... 28 2.3 Results and Discussion ................................................................................................. 30 Chapter 3: Spatiotemporal Land Use Regression Models of Fine, Ultrafine, and Black Carbon Particulate Matter in New Delhi, India .......................................................................41 3.1 Introduction ................................................................................................................... 41 3.2 Materials and Methods .................................................................................................. 43 vii  3.2.1 Site Selection ............................................................................................................ 43 3.2.2 Instrumentation and Field Measurements ................................................................. 43 3.2.2.1 GPS and Meteorological Data .......................................................................... 44 3.2.2.2 PM2.5 Data Collection ....................................................................................... 45 3.2.2.3 Black Carbon Data Collection .......................................................................... 45 3.2.2.4 UFPN Data Collection ...................................................................................... 46 3.2.2.5 Data Reduction and Quality Control................................................................. 46 3.2.3 Spatial and Socioeconomic Variables ....................................................................... 47 3.2.4 Model Building ......................................................................................................... 47 3.2.5 Model Evaluation ...................................................................................................... 50 3.2.6 Regression Mapping ................................................................................................. 50 3.3 Results ........................................................................................................................... 50 3.3.1 PM2.5 Models ............................................................................................................. 51 3.3.2 Black Carbon Models ............................................................................................... 52 3.3.3 UFPN Models ........................................................................................................... 52 3.4 Discussion ..................................................................................................................... 55 Chapter 4: Estimating PM2.5 Population Exposure in New Delhi, India, Using a Probabilistic Simulation Framework and Spatiotemporal Land Use Regression Models ....59 4.1 Introduction ................................................................................................................... 59 4.2 Materials and Methods .................................................................................................. 61 4.2.1 Key Input Data and Models ...................................................................................... 63 4.2.1.1 Zonal Planning Data for Home\u2013Work Trip Generation ................................... 63 4.2.1.2 Output From Spatiotemporal LUR Models ...................................................... 64 viii  4.2.1.3 Time-Activity Data ........................................................................................... 66 4.2.1.4 Ratio of In-vehicle Concentration to Near-Vehicle Concentration, \u03bavm ........... 67 4.3 Results ........................................................................................................................... 68 4.4 Discussion ..................................................................................................................... 76 Chapter 5: Conclusion .................................................................................................................80 5.1 Key Findings and Implications ..................................................................................... 80 5.2 Final Reflections ........................................................................................................... 83 Bibliography .................................................................................................................................87 Appendices ....................................................................................................................................99 Appendix A ............................................................................................................................... 99 Appendix B ............................................................................................................................. 102 Appendix C ............................................................................................................................. 104 C.1 Simulation Algorithm ............................................................................................. 106 Appendix D ............................................................................................................................. 109 D.1 Survey Instrument ................................................................................................... 109  ix  List of Tables   Table 2.1: Estimated modal share for the ITO intersection. ......................................................... 30 Table 2.2: PM2.5 multiple linear regression and robust regression models. .................................. 35 Table 2.3: NO2 and NO multiple linear regression and robust regression models. ...................... 36 Table 2.4: CO multiple linear regression and robust regression models. ..................................... 37 Table 3.1: Potentially predictive independent variables. .............................................................. 49 Table 3.2: Descriptive statistics for hourly median concentrations of PM2.5, BC, and UFP at the land use regression sites. ............................................................................................................... 51 Table 3.3: Final spatiotemporal LUR model specifications and results for PM2.5, BC and UFPN........................................................................................................................................................ 53 Table 4.1: Median daily travel time from survey data, measured PM2.5 concentrations on the study route and estimated values for \u03bavm for each mode of travel. ............................................... 69 Table 4.2: Descriptive statistics for average daily PM2.5 exposure for Scenarios 1, 2 and 3 ....... 70  Table B.1: LUR models using robust regression ........................................................................ 103  x  List of Figures   Figure 1.1: PM2.5 trend at the ITO intersection in New Delhi (Sept 2006 \u2013 Sept 2014). Each point represents a monthly mean, the bold line is a fitted smooth trend and the colored band represents the 95% confidence interval of the fit. ............................................................................................ 8 Figure 1.2: NO2 trend at the ITO intersection in New Delhi (Sept 2006 \u2013 Dec 2014). Each point represents a monthly mean, the bold line is a fitted smooth trend and the colored band represents the 95% confidence interval of the fit. ............................................................................................ 9 Figure 1.3: CO trend at the ITO intersection in New Delhi (Sept 2006 \u2013 Dec 2014). Each point represents a monthly mean, the bold line is a fitted smooth trend and the colored band represents the 95% confidence interval of the fit. .......................................................................................... 10 Figure 1.4: O3 trend at the ITO intersection in New Delhi (Sept 2006 \u2013 Dec 2014). Each point represents a monthly mean, the bold line is a fitted smooth trend and the colored band represents the 95% confidence interval of the fit. .......................................................................................... 11 Figure 2.1: Traffic flow diagram for ITO intersection. Video camera positioned at box 1 captured the following flows: 1\u20133 (from 1 to 3), 1\u20134, 4\u20131, 2\u20131 and 3\u20131. For one full day (June 27, 2007) a second camera was mounted on another pedestrian overpass (seen as box 2). The second camera captured the following flows: 4\u20132, 3\u20132, 2\u20134 and 2\u20131. For the purpose of simplicity, from here onwards, we refer to the sum of flow 1\u20133 and flow 1\u20134 as Flow1 and the sum of flow 4\u20131, flow 2\u20131 and flow 3\u20131 as Flow2. Further, we will designate the sum of flow 2\u20134 and flow 2\u20131 as Flow3 and the sum of flow 4\u20132 and flow 3\u20132 as Flow4. Flow1 and Flow2 were captured by the camera positioned at box 1 and Flow3 and Flow4 were captured by the camera positioned at box 2..................................................................................................................................................... 25 xi  Figure 2.2: A picture of Flow 1 and Flow 2 taken from box 1. .................................................... 26 Figure 2.3: Average hourly traffic flow by mode. ........................................................................ 31 Figure 2.4: Diurnal variation of PM2.5 at the ITO intersection (September 2007 to December 2014). The solid lines represent the mean values and shaded area represents the 95% confidence interval. ......................................................................................................................................... 32 Figure 2.5: (a) Expected reduction in PM2.5 concentrations if the initial truck flow is reduced by 50%. (b) Expected reduction in NO2 concentrations if the initial auto-rickshaw flow is reduced by 50%. ......................................................................................................................................... 39 Figure 3.1: Land use regression sites and rooftop site overlaid on major road network. ............. 44 Figure 3.2: Predicted average PM2.5, BC, and UFPN spatial variation for the duration of the study. Transect for Figure 3.3 displayed in part (a). ..................................................................... 54 Figure 3.3: Predicted average PM2.5, BC, and UFPN variation along an arbitrary transect (from point 1 to point 2, Figure 3.2(a)). The red lines indicate results for the afternoon, and the blue lines indicate results for the morning. ........................................................................................... 56 Figure 4.1: A visual representation of the simulation process. *Indicates that the step was repeated 100,000 times. ................................................................................................................ 62 Figure 4.2: (a) Cumulative frequency distribution (CFD) for average daily PM2.5 exposures for Scenarios 1, 2 and 3; (b) Average daily exposure by microenvironment for Scenario 1, Scenario 2 and Scenario 3; (c) CFD for winter (December, January and February); (d) CFD for monsoon season (July, August and September). .......................................................................................... 71 Figure 4.3: Monthly variation of average daily PM2.5 exposure: 95th, 75th, median, 25th and 5th percentile for Scenario 1. .............................................................................................................. 72 xii  Figure 4.4: Average PM2.5 concentrations for morning, afternoon and nighttime hours (Scenario 1). .................................................................................................................................................. 73 Figure 4.5: Map of the modeling domain along with average annual PM2.5 exposure in \u00b5g m-3. 74 Figure 4.6: (a) Annual average daily PM2.5 exposure for the 16 zones along with monthly variation of 95th, 75th, median, 25th and 5th percentiles (b): Difference between average daily PM2.5 exposures for winter and monsoon months for the 16 zones along with monthly variation of 95th, 75th, median, 25th and 5th percentiles. The modeling domain covered a small fraction (<5%) of zones KI and N; hence the estimates of exposure for these two zones are unreliable. . 75 Figure 4.7: Annual average PM2.5 exposure as a function of distance from nearest major road. . 76  Figure A.1: Diurnal variation of NO2 at ITO intersection (September 2006 to December 2014). The solid line represents the mean values and shaded area represents the 95% confidence interval. ......................................................................................................................................... 99 Figure A.2: Diurnal variation of NO at ITO intersection (September 2006 to December 2014). The solid line represents the mean values and shaded area represents the 95% confidence interval. ....................................................................................................................................... 100 Figure A.3: Diurnal variation of CO at ITO intersection (September 2006 to December 2014). The solid line represents the mean values and shaded area represents the 95% confidence interval. ....................................................................................................................................... 101 Figure B.1: Diurnal variation of PM2.5 concentrations at the rooftop site by month. ................. 102 Figure B.2: Diurnal variation of BC concentrations at the rooftop site by month. ..................... 102 Figure C.1: Monthly PM2.5 patterns (mean and 95% confidence interval) at the fixed regulatory monitor (ITO: Sept 2006 to Sept 2014). ..................................................................................... 104 xiii  Figure C.2: Fitted beta distribution compared to distribution of observed travel time fraction. 105 Figure C.3: Study route (~50 km along Ring Road) also shown are major roads in New Delhi...................................................................................................................................................... 106 xiv  List of Abbreviations   BC  black carbon (light absorbing component of PM, measured by light absorption) CO  carbon monoxide CNG  compressed natural gas (mainly methane) IQR  interquartile range (25th \u201375th percentile) LUR  land use regression NO2  nitrogen dioxide NO  nitric oxide NOx  oxides of nitrogen (NO + NO2) O3  ozone PM  particulate matter, also referred to as aerosol when suspended in a gas PM2.5  particulate matter with aerodynamic diameter \u2264 2.5 \u00b5m PM10  particulate matter with aerodynamic diameter \u2264 10 \u00b5m SOA  secondary organic aerosol SO2  sulfur dioxide UFP  ultrafine particles (aerodynamic diameter \u2264 0.1 \u00b5m) UFPN  ultrafine particle number concentration (particle cm-3)  VOCs  volatile organic compounds      xv  List of Symbols   \u00b5g m-3  micrograms per cubic meter \u00b5m  microns or micro meters (10-6 m) nm  nanometer (10-9 m)  xvi  Acknowledgements   I must thank my supervisor, Prof. Milind Kandlikar, for asking critical questions and for providing direction to my research. I am thankful to Prof. Kandlikar for accommodating my requests for countless brainstorming sessions, even on weekends. I must also thank members of my supervisory committee, Prof. Michael Brauer and Prof Madhav Badami, for their guidance, encouragement, critical feedback and support. I must thank Prof. Brauer for his thoughtful comments on my research. I am grateful to all members of my supervisory committee for making my research one of their priorities.     I am grateful for the financial support received from the University of British Columbia in the form of the Four Year Fellowship (FYF) and the Olav Slaymaker Scholarship in Environment. This work was carried out with the aid of a grant (105407-9906075-078) from the International Development Research Centre, Ottawa, Canada. Information on the Centre is available on the web at www.idrc.ca. The Auto-21 NCE also provided partial funding for this work along with ORSIL# F07-0010: Leaders Opportunity Fund grant. I must also thank Anna W\u00e4rje for editing a portion of this dissertation at very short notice.   Most importantly, I would like to thank my wife, Dr. Pooja, for supporting me through this arduous process of writing. This dissertation would not have been possible without her support. I must thank my parents, who worked very hard to support my education until I finished my undergraduate degree, for their enormous love and encouragement. I am grateful to my elder brother Narendra, elder sister Hina and younger brother Manoj, who have always supported my dreams and ambitions.   I must also thank staff and students at the Institute for Resources, Environment and Sustainability, who make it a great space for learning. I am grateful for the support I received from my friends at the Graduate Student Society (GSS) of UBC Vancouver, especially from my dear friend Dr. Mrigank Sharma.  xvii  Last but not least, I am grateful for the support that I received from my colleagues at the Ministry of Environment: Ralph Adams, Robyn Roome, Doug Hill, Matt Lamb-Yorski, Ed Hoffman, Jennifer McGuire and others in the Environmental Protection Division. xviii  Dedication    To my wife and my parents  1  Chapter 1: Introduction   1.1 Urban Air Pollution Urban air pollution is a major environmental and public health concern worldwide. In 2012, 3.7 million (about 6.7% of total) deaths worldwide were attributed to ambient air pollution by the World Health Organization (WHO).1 Urban air pollution is comprised of a mixture of pollutants including particulate matter (PM), oxides of nitrogen (NOx), sulfur dioxide (SO2), ozone (O3), carbon monoxide (CO) and volatile organic compounds (VOCs). Those air pollutants that are of greatest concern from a public health perspective are termed \u201ccriteria pollutants,\u201d whose ambient levels and emissions are subject to regulatory standards, including ambient air quality standards and source-specific discharge limits and standards. Numerous scientific advancements in monitoring, exposure assessment, modeling, toxicology and epidemiology that have resulted in a better understanding of the adverse health effects of urban air pollution also provide the knowledge base according to which air pollutants are regulated. Air quality standards within jurisdictions are based on available epidemiologic and toxicological information for specific pollutants. For example, in 2005, the WHO conducted a thorough risk assessment and developed ambient air quality guidelines for PM, O3 and SO2 that were stringent compared to previous levels; nitrogen dioxide (NO2) standards were kept unchanged.2   In addition to outdoor sources of air pollution, indoor sources of air pollution such as tobacco smoke and the combustion of solid fuels used for cooking and heating also have a significant impact on human health, especially in developing countries.3 The focus of this research, however, is on urban ambient air pollution.  1.1.1 Fine Particulate Matter PM generally consists of both solid and liquid phase matter in different size ranges, and its sources are both natural and anthropogenic in nature. Anthropogenic sources include, but are not limited to, vehicular emissions, power generation and industrial emissions, biomass burning and resuspension of road dust. Adverse health effects have been conclusively linked to PM air 2  pollution,4-12 and hence PM is of key interest in air pollution assessments. There is strong evidence that fine particulate matter or PM2.5 (aerodynamic diameter \uf0a3 2.5 \uf06dm) causes adverse health effects in exposed individuals, primarily to their cardiovascular and respiratory systems.2 PM2.5 is mostly generated due to the combustion of fossil fuels in motor vehicles, industries and power plants, and due to the combustion of biomass. PM2.5 is generated by anthropogenic and natural sources. Anthropogenic sources include combustion of fossil fuels in motor vehicles and industries, as well as non-combustion sources like resuspended road dust. Secondary PM2.5 (sulfate, nitrate and secondary organic aerosol (SOA)) can also be formed due to nucleation and gas-to-particle conversion. Secondary sulfates and nitrates are formed due to reaction of ammonia with SO2 and NOx. Therefore it is important to reduce ammonia emissions from significant sources like agricultural application of fertilizers and agricultural burning.13 SOA formation primarily occurs due to atmospheric oxidation of volatile organic compounds (VOCs) by OH radicals, NO3 radicals, O3 or Cl atoms.14   PM2.5 mass concentration has been used in epidemiologic studies and for regulatory purposes; there are, however, several other characteristics, such as particle surface area, oxidative potential, charge and particle count that may also be significant in terms of health effects.12 Notwithstanding variability in PM characteristics that may influence health risks, the reported health risk estimates per unit PM mass fall within a narrow range of values.5 There have been numerous epidemiologic studies that show that both short-term and long-term exposures to PM2.5 are associated with increased mortality,4, 11 and a causal relationship between PM2.5 exposure and cardiovascular morbidity and mortality11 is widely accepted within the scientific community. Traffic-related air pollution has emerged as an important contributor to cardiovascular risk11, 15 associated with air pollution exposure. Additionally, two constituents of PM2.5\u2014black carbon (BC) and ultrafine particles (UFPs, <0.1 \uf06dm)\u2014have received increasing attention. BC is a product of incomplete combustion, and short-term health studies suggest that BC may be a better indicator of harmful particulate matter from combustion processes, especially traffic, than undifferentiated PM mass.16 The unique physical properties of UFPs, including their potential for translocation into the bloodstream, have led researchers to believe that UFPs may have specific or enhanced toxicity vis-\u00e0-vis other PM fractions. Although exposure studies, epidemiologic 3  studies and toxicological studies of UFPs have not provided consistent findings regarding their effects, independent health effects of UFPs cannot be ruled out.17  1.1.2 Other Important Urban Air Pollutants Carbon monoxide (CO), nitrogen dioxide (NO2), sulfur dioxide (SO2), ozone (O3) and VOCs are the other key urban air pollutants. CO is a product of incomplete combustion of fossil and solid biomass fuels; motor vehicle emissions are a major source of CO in urban settings. There are several species of nitrogen oxides, but NO2 is most important in terms of human health effects. In urban environments, most of the ambient NO2 is emitted as nitric oxide (NO) from power plants and motor vehicles, which then undergoes atmospheric transformation to NO2.2 Ambient O3 is a secondary air pollutant and is formed via complex chemical reactions involving VOCs, NOx and sunlight.18 SO2 is largely produced by combustion of fossil fuels like coal and diesel. These fossil fuels contain varying amounts of sulfur that get converted to SO2 during combustion. Health effects of CO, NO2, SO2 and O3 exposure have been well documented in the literature.2 In this research, I focus primarily on PM2.5, as it has been most consistently associated with adverse health effects. This research is located in New Delhi, India, due to the high levels of ambient air pollution and its potential effects on a population of about 16.7 million. However, the results are informative for managing air quality in other cities in India and elsewhere.    1.2 Air Pollution in India Many researchers have explored the relationship between development and environmental degradation on a national scale through Environmental Kuznets curves (EKC). The EKC hypothesizes a relationship between environmental degradation in general\u2014and air pollution in particular\u2014with per capita income in the shape of an inverted-U, wherein high and low values of income are associated with lower concentrations, while middle income regions show higher levels of air pollutants. Seledon and Song19 used a cross-national panel of data and reported that per capita emissions of CO, suspended particulate matter (SPM), NOx and SO2 exhibit inverted-U relationships with per capita income. In a similar vein, Grossman and Krueger20 examined the relationship between per capita income and environmental indicators and found that economic 4  growth brings an initial phase of deterioration followed by a subsequent phase of improvement, thus supporting the EKC hypothesis. Others have noted that the EKC hypothesis is problematic because of heteroskedasticity and simultaneity in input variables.21 If we assume that the inverted-U relationship is valid, in order to reach a turning point (i.e. the point on the curve where environmental degradation stops and alleviation begins) per capita incomes in developing countries will need to increase over time to a level equivalent to a GDP per capita of $8000 (USD).20 As incomes increase, the expectation is that better emission control technologies and improved regulation resulting from the increased public demand for improved environmental quality will reduce the levels of outdoor air pollution, while the use of clean cooking fuels like liquefied petroleum gas will help alleviate indoor air pollution. India\u2019s GDP per capita in 2014 was $1631.22 By this measure, India has some distance to go before the environmental benefits of economic growth are realized.  The Central Pollution Control Board (CPCB), a unit within the Central Government\u2019s Ministry of Environment and Forests, is responsible for collecting air pollution data and prescribing air quality standards. The functions and powers of CPCB are defined in the Water (Prevention and Control of Pollution) Act, 1974 and the Air (Prevention and Control of Pollution) Act, 1981. CPCB plans and coordinates the nationwide programs and the activities of the State Boards (provincial), and provides technical assistance and guidance to the State Boards. The State Boards are constituted under the two acts mentioned above and are bound by the directions of CPCB. District offices of the State Boards report to the Regional or the Zonal Offices and the Regional offices report to the State Boards. Agencies involved in the air pollution-monitoring program are the State Pollution Control Boards, Pollution Control Committees (in union territories), Zonal offices and the National Environmental Engineering Research Institute, Nagpur. CPCB coordinates with these agencies to ensure the consistency of air quality data, and provides technical and financial support for operating the monitoring stations. CPCB is responsible for executing the nation-wide National Air Quality Monitoring Program (NAMP). The objectives of the program are to determine the status and trends of ambient air quality and identify non-attainment cities.23 The network consists of 523 operating stations covering 215 cities. Under NAMP, three pollutants are monitored at all the stations: PM10 (particulate matter 5  with aerodynamic diameter \uf0a310 \uf06dm), nitrogen dioxide (NO2) and sulfur dioxide (SO2). Currently, PM2.5 is continuously monitored at select locations24 such as New Delhi (8 sites), Mumbai (2 sites), Bengaluru (3 sites), Hyderabad (2 sites) and Chennai (3 sites).   Air quality in Indian cities has steadily deteriorated due to rapid industrial growth and unplanned urbanization in the recent past.25 During 2014\u20132015, India\u2019s GDP grew by 7.4%.26 Industrial growth has accompanied increases in population: over the past decade the population of India increased by 182.7 million, and currently stands at 1.21 billion.27 Migration to urban centers from rural areas has contributed to a dramatic rise in urban populations and population densities in India\u2019s megacities. Centers with high population density are also likely to have higher levels of air pollution due to proximity to traffic and industrial sources, which in turn leads to greater numbers of people being exposed to higher levels of air pollution.  Motor vehicles, power plants, industries and biomass combustion have been identified as major sources of air pollution in urban India.28 In 2012, India had 160 million registered motor vehicles, including 72 million motorized two-wheelers and 13 million cars. The number of motor vehicles grew at a rate of 10.5% per year between 2002 and 2012.29 The growth rate of motorized traffic has been far greater than the growth rate of road networks, resulting in capacity saturation leading to higher traffic emissions and deterioration of air quality. There are about 15 cars per thousand persons in India,29 and this number is about 10 times larger for New Delhi. Rates of vehicle ownership in India are still far lower than in developed countries (~500 per thousand) and the increase in the number of motor vehicles is expected to continue in the near future.30, 31 Thus air pollution due to motor vehicles will grow as a challenge for India in the years to come.  As would be expected, the major sources of air pollution in rural environments are quite different from those in urban environments. An important source of air pollution exposure in rural India is the use of solid fuels for cooking. About 86% of rural households in India use solid fuels for cooking, while only 26% of urban households use solid fuels for cooking.27 Other important rural sources of PM include open field burning of agricultural residue, which is known to impact air 6  quality on a regional scale.32 PM2.5 emissions due to the use of solid fuels for cooking contribute significantly to the burden of disease in India.33, 34 Use of solid fuels for cooking is less prevalent in urban areas, and the magnitude of the problem is greater in rural areas. In the case of New Delhi, only 3% households use solid fuels for cooking27. Although indoor exposure from widespread use of solid fuels for cooking is an important health concern in rural India, the scope of this research is limited to ambient air pollution in urban environments, specifically New Delhi.   1.2.1 Air Pollution in New Delhi, India New Delhi, India\u2019s capital, is a megacity whose population has increased from 13.8 million in 2001 to 16.7 million in 2011. During the same period, the population density went up from 9300 persons per square km to 11,300 persons per square km. This growth in population has come at a time when the Indian economy, especially in its major cities, has boomed. New Delhi had a 16% annual compounded growth rate for gross state domestic product (at current prices) between 2007\u20132008 and 2011\u20132012. Annual per capita income (at current prices) was 212,219 INR in 2013\u20132014 and had an annual growth rate of 14% during 2012\u20132015.35 Rising incomes have been accompanied by degradation in New Delhi\u2019s air quality.   The current levels of air pollution in New Delhi are among the highest in the major cities of the world. In 2014, the annual average PM2.5 concentrations at six regulatory stations maintained by the Delhi Pollution Control Committee (DPCC) ranged from 125 to 191 \u00b5g m-3.35 These levels far exceeded the regulatory standard for PM2.5 (an annual mean of 40 \u00b5g m-3). The average annual concentrations of NO2 in the same year ranged from 51\u2013106 \u00b5g m-3, exceeding the applicable regulatory standard (an annual mean of 40 \u00b5g m-3). In 2014, SO2 ranged from 12\u201320 \u00b5g m-3 and the regulatory standard for SO2 was met at all the six sites (an annual mean of 60 \u00b5g m-3). The SO2 levels in New Delhi have consistently met the regulatory standard since 199735 and will not be discussed further. During 2014, PM10 ranged from 203\u2013583 \u00b5g m-3, more than three times higher than the applicable standard (an annual mean of 60 \u00b5g m-3). Annual averages of CO ranged from 1280\u20132640 \u00b5g m-3, and O3 ranged from 32\u201396 \u00b5g m-3. The applicable regulatory standards for CO (2 and 4 mg m-3 for 98th percentile of 8 h and 1 h means, respectively) and O3 (100 and 180 \u00b5g m-3 for 98th percentile of 8 h and 1 h means, respectively) 7  were met at all but one site.35 Recently, the U.S. Embassy in New Delhi started monitoring ambient PM2.5 as well.36 The Embassy is located in an area that is considered relatively \u201cclean,\u201d yet the reported annual mean for 2014 was 129 \u00b5g m-3. Overall, the PM2.5 levels are alarmingly high compared to the recommended levels by the WHO (an annual mean of 10 \u00b5g m-3 for PM2.5).   Long-term hourly time series data are available only for a single site (the ITO intersection, maintained by CPCB) in New Delhi. Figures 1.1, 1.2, 1.3 and 1.4 were created using the smoothTrend function in the Openair package in R37 and show the trends in PM2.5, NO2, CO and O3 at the ITO intersection. This function uses hourly data to calculate monthly mean and fits a smooth trend line using generalized additive modeling. In Figure 1.1, monthly concentrations show a cyclical pattern, with peaks in winter (November\u2013January) and troughs in monsoon season (July\u2013September). The fitted trend shows a small increase in PM2.5 levels over time, but the increase is not statistically significant. It can be seen that the local PM2.5 standard (an annual mean of 40 \u00b5g m-3) was consistently and grossly exceeded.  Figure 1.2 shows the trend in NO2 levels at the ITO intersection from September 2006 to December 2014. It can be seen that there has been a minor decline in the NO2 levels during this time period, but there is no clear trend. The NO2 levels at the ITO intersection during this time period exceeded the applicable regulatory standard (an annual mean of 40 \u00b5g m-3). The trend in CO levels at the ITO intersection during this time period is shown in Figure 1.3. The applicable regulatory standards for CO (2 and 4 mg m-3 for 98th percentile of 8 h and 1 h means, respectively) were exceeded for all years during this time period at the ITO intersection. The CO levels dropped in 2010 but there has been an increase since 2013. The trend in O3 levels at the ITO intersections during this time period is shown in Figure 1.4; the applicable regulatory standards (100 and 180 \u00b5g m-3 for 98th percentile of 8h and 1 h means, respectively) were exceeded every year from 2009\u20132014.    8   Figure 1.1: PM2.5 trend at the ITO intersection in New Delhi (Sept 2006 \u2013 Sept 2014). Each point represents a monthly mean, the bold line is a fitted smooth trend and the colored band represents the 95% confidence interval of the fit.    9   Figure 1.2: NO2 trend at the ITO intersection in New Delhi (Sept 2006 \u2013 Dec 2014). Each point represents a monthly mean, the bold line is a fitted smooth trend and the colored band represents the 95% confidence interval of the fit.   10   Figure 1.3: CO trend at the ITO intersection in New Delhi (Sept 2006 \u2013 Dec 2014). Each point represents a monthly mean, the bold line is a fitted smooth trend and the colored band represents the 95% confidence interval of the fit.  11   Figure 1.4: O3 trend at the ITO intersection in New Delhi (Sept 2006 \u2013 Dec 2014). Each point represents a monthly mean, the bold line is a fitted smooth trend and the colored band represents the 95% confidence interval of the fit.   1.2.2 Source Contributions  There are 2.8 million cars, 5.7 million motorized two-wheelers, 81,000 motorized three-wheelers, 79,000 taxicabs and 20,000 buses in New Delhi.35 The overall growth rate of motorized vehicles is about 7% per year. Buses, taxicabs and motorized three-wheelers in New Delhi are required by law to run on Compressed Natural Gas (CNG). Introduction of CNG as a mandatory fuel for public transport vehicles in New Delhi has been a major policy measure taken to alleviate air pollution in New Delhi, and as such has provided a \u201cnatural experiment\u201d to test the impact of fuel switching on air quality. Many researchers have studied the impact of CNG use on ambient air quality in New Delhi. Kathuria38 analyzed daily ambient air quality data from June 1999 to September 2003 obtained from the ITO intersection and reported that NOx levels had increased after the conversion, whereas PM10 levels showed only a marginal fall, and CO 12  levels showed a significant decline. Kandlikar39 used singular spectrum analysis to isolate trends and seasonal cycles from daily concentrations of SO2, NOx, CO and PM10. Periods of sharp reductions were reported for both SO2 and CO concentrations in 2001 and 2002, respectively. NOx concentration were reported to show a sustained rise from 2000 to 2004, followed by small decline thereafter; PM10 concentrations were reported to show no trend over the time period.   Emission inventories provide a list of air pollution sources and emitted pollutants for a particular area. Emissions inventories have been developed for New Delhi40, 41 but they are limited due to the lack of availability of necessary data. A recent regional PM2.5 emissions inventory for New Delhi and its neighboring cities put the contribution of the transport sector highest (17%) followed by power plants (16%), brick kilns (15%), industries (14%), domestic sources (12%), waste burning (8%) and others.42 Receptor-modeling studies use chemical analysis to estimate the fraction of ambient concentration of a pollutant contributed by a specific source. Source apportionment studies using chemical mass balance models provide reasonable qualitative understanding of sources, despite the fact that source profiles for Indian cities have not been developed yet. It is important to recognize that when doing source apportionment, except for positive matrix factorization (PMF),43 one has to start with some \u201csignificant\u201d sources. Inclusion of sources is based on information available from past studies, emissions inventories etc. If a source is not considered at all in the model, the results may grossly overestimate the contributions of other sources.   There have been a few particulate matter source apportionment studies in New Delhi.44, 45 A recent study used chemical mass balance (CMB8) modeling to apportion PM2.5 to sources:46 fossil fuel combustion (gasoline, diesel and coal) was found to contribute 25\u201333% of measured PM2.5, and biomass combustion to contribute 7\u201320%. The results indicated that there was no single dominant source and source contributions vary by season, with the highest PM2.5 concentrations in winter and the lowest in summer. Another recent source apportionment study47 in New Delhi found that primary traffic emissions contributed about 16\u201318% of ambient PM2.5; biomass burning was found to contribute about 23% during winter. In the case of New Delhi, 90% of households use LPG, about 5% use kerosene and only about 3% of households use solid 13  fuels,27 suggesting that the use of solid fuels for cooking is not widespread in New Delhi. Srivastava et al. (2009)48 used scanning electron microscopy energy dispersive X-ray analysis to study differences in morphology and elemental composition of particles in five different size ranges from two sites in New Delhi. One of the sites was a major traffic intersection (ITO) and the other was an urban background site (Jawahar Lal Nehru University) with very low local traffic and no major sources of air pollution in the surrounding area. The authors reported that the elemental composition did not change substantially in the fine size ranges across the sites, indicating homogeneous composition of fine particles across the city. Back-trajectory and satellite data analyses have shown that regional sources of air pollution can cause very high levels of PM2.5 in New Delhi during the post-monsoon season.49    1.3 Spatiotemporal Variability of Air Pollution & Exposure The concentrations of air pollutants vary in both time and space, and within-city variability of air pollutants may be larger than between-city variability.50 Using data from fixed site monitors as a proxy for exposure may introduce errors in risk estimates due to exposure misclassification, especially in epidemiologic studies of long-term exposure (i.e. cohort studies).51 Therefore, the development of intra-urban air pollution exposure models emerged as a priority research area.52 There are multiple ways to model the within-city variability of air pollutants including land use regression (LUR) models, dispersion models, hybrid models and others. Jerrett et al. (2005) conducted a systematic review of intra-urban air pollution exposure models. They recognized LUR models as a reliable and cost-effective option, along with the importance of integrating time-activity patterns in exposure assessment. Hoek et al. (2008)53 conducted a review of published LUR models and concluded that their performance was better than or equivalent to geo-statistical methods and dispersion models. The authors also highlighted the importance of including a temporal component in LUR models that could be applied in studies that require exposure estimates at a finer temporal scale.  Studies of within-city variability of air pollutants and exposure have been scarce in India. Guttikunda and Calori (2013)42 used a dispersion model to predict the variability of PM2.5 and 14  PM10 in New Delhi (using 1x1 km grid cells). Generally, a finer spatial resolution is desirable to highlight pollutant variability, which generally is of the order of tens of meters. Recently, the Health Effects Institute funded coordinated studies of short-term exposure to PM10 and all-natural-cause mortality in New Delhi and Chennai, India, as part of the Public Health and Air Pollution in Asia (PAPA) Program.54 Gaseous pollutants were also included for analyses in New Delhi. In the Chennai study, concentrations from the nearest ambient monitor were used to assign exposure to each cell in ten zones to address the spatial heterogeneity of PM10. However, the spatiotemporal variations within each zone were not addressed. A centering technique was used to assign exposures in the New Delhi study and within-city variability was not accounted for.  There are two methods by which to determine the variability of exposure of an urban population: computer modeling (indirect approach) and personal monitoring (direct approach).55 Although personal monitoring can provide reliable data, it tends to be costly. The modeling approach or indirect approach uses empirical distributions of exposure in specific microenvironments and time-activity data to simulate exposure.56, 57 The main advantage of this approach is that it is cost-effective, though validation of the results can be difficult since they must be compared with directly measured exposure levels.56 A weak correlation between ambient measurements and personal monitoring data supports the use of the simulation approach for exposure assessment in epidemiologic studies.57 A number of exposure simulation models have been developed for air pollutants, including for PM2.5, PM10, NO2, CO 58, 59. For example, Burke et al. (2001) developed a probabilistic model to estimate the distribution of PM2.5 exposure by randomly sampling from various input distributions.60 Outputs from air quality models have been also been used as inputs in population exposure models to obtain improved estimates of exposure.61 Setton et al. (2011)62 used simulation approach, including LUR models, to understand the effect of ignoring daily mobility of individuals on exposure to traffic-related air pollution and health effect estimates. The authors reported that ignoring mobility of individuals may cause negative bias in models of exposure to outdoor air pollution and health effects.62   15   1.4 Research Objectives  1.4.1 Objective 1 Concentrations of air pollutants vary with time and space. Pollutant concentrations at a given point are dependent on proximity to sources, topography, meteorology and atmospheric transformations of precursor gases (if applicable).  Fixed site monitors are used to estimate the average levels of air pollutants and to determine attainment of applicable ambient air quality standards and help understand the spatial and temporal variability of urban air pollutants in large cities. Despite the high levels of measured concentrations of PM2.5 in New Delhi, the number of sites that measure PM2.5 concentrations was fairly limited until recently.   Since there is limited information available regarding spatial variation of PM in New Delhi, this research aims to fill that gap. A key objective of this work is to understand the patterns of spatial and temporal variability of pollutants, particularly PM2.5, in New Delhi.   1.4.2 Objective 2  Once we characterize spatial and temporal variability of a pollutant, we can use that information, along with time-activity data, to model the variability in population exposure. Exposure to an air pollutant is defined as contact between an individual and the pollutant at a given concentration over a defined duration of time.63 Exposure assessment is a necessary input into epidemiologic studies that develop empirical relationships between exposure to an air pollutant and its health effects. Direct measurement of personal exposure is cost-prohibitive, and using a small sample size may bias the effect estimates.52 In most epidemiologic studies of long-term air pollution exposure, given that only residential addresses may be available, it is assumed that individuals stay at home only; in other words, mobility of individuals is ignored and ambient levels of the pollutant of interest near an individual\u2019s residence are used to estimate average exposure. This can lead to exposure misclassification and cause bias in epidemiologic analyses.51    16  A second objective of this work is to characterize the spatial and temporal variability in PM2.5 exposure by combining data on where people live, commute and work with spatiotemporal variability of PM2.5 concentration in New Delhi.  In order to meet these research objectives, I have focused my efforts on three major research themes. First, I have sought to understand how local traffic flow impacts ambient air quality on a local scale. This allows us to understand the role of traffic vis-\u00e0-vis background or large-scale variations. Second, I have modeled the citywide spatial and temporal patterns of PM2.5 and its two components (BC and UFPs) using surrogates for traffic, land use and population. These Land Use Regression (LUR) models provide critical insights into spatial and temporal variability of the pollutants, which in turn is an important input in exposure assessment. Third, I have developed a probabilistic simulation framework that integrates spatiotemporal variation in PM2.5 concentrations quantified using LUR models with time-activity patterns to elicit the spatial and temporal variability in PM2.5 population exposure. This framework allows us to translate ambient concentrations into estimates of exposure, and evaluate the effect of time-activity patterns and mobility on exposure.   1.4.3 Role of Traffic (Objective 1) It is recognized that traffic emissions are a significant source of air pollutants in New Delhi. Due to the importance of traffic-related air pollution and the continuing increase in the number of motor vehicles in New Delhi, it is important to understand the influence of local traffic flow on ambient air quality. This will help in understanding the likely outcome of policy measures that may mandate a reduction in traffic flow on certain routes or for a specific mode of traffic.   1.4.3.1 Research Question #1  Can we model the relationship between local traffic and fleet composition, and temporal variations in air pollution at a high traffic location? Will reductions in local traffic flow lead to reductions in pollutant (PM2.5, NO, NO2 and CO) concentrations on a localized scale? I address these questions in Chapter 2 of this dissertation.  17   1.4.4 Spatiotemporal Variations (Objective 1)  Modeling techniques like the LUR approach can highlight intra-urban variability of urban air pollutants.64 LUR models can provide a concentration map for a city, which in turn can be used for epidemiologic studies, risk assessment and prioritization of air quality management.53 Model development includes development of a statistical relationship between land use variables and pollutant concentrations measured at a number (20\u2013100) of representative sites. This relationship is used to predict pollutant concentrations over a bigger domain of interest. LUR has been commonly used to estimate annual concentrations of a pollutant,53 however, spatiotemporal LUR models have also been reported with finer time resolution.65, 66 Land use, road network, population density and traffic flow variables are typically used as inputs to the models, although LUR models have been reported to perform well even in the absence of traffic flow data as an input.67 Several site selection processes have been used by researchers,53, 67-71 but in essence a reasonable number of sites that capture the variability in land use are needed.   1.4.4.1 Research Question #2 What are the predictors of spatial and temporal variations in PM2.5 concentrations in New Delhi? Can we model the spatiotemporal variation in PM2.5 concentrations, with reasonable accuracy, without making a significant financial investment in a fixed monitoring network? Are spatiotemporal patterns for PM2.5 components like black carbon (BC) and ultrafine particles (UFP, <0.1 \u00b5m) similar to that of PM2.5? I address these questions in Chapter 3 of this dissertation.  1.4.5 PM2.5 Population Exposure (Objective 2) We know that pollutant concentrations vary in both space and time, and that people visit several microenvironments in a day. If we want to estimate a population\u2019s exposure to a certain pollutant, we can combine time-activity patterns with the spatiotemporal pollution maps. Modeling can be used to characterize the variability in population exposure.55, 60 Spatiotemporal models, like LUR models, can provide time-varying concentration maps which can be combined 18  with time-activity patterns to estimate population exposure.72-74 Generally, epidemiologic studies of long-term exposure have not accounted for mobility and the role of exposures encountered during transportation.52, 53 Ignoring mobility tends to negatively bias effect estimates due to the fact that air pollution exposure tends to be higher in the transport microenvironment.62  1.4.5.1 Research Question #3 How do spatiotemporal variations in pollutant concentrations affect the variability in population exposure? Can we combine spatiotemporal LUR models with time activity data to estimate the spatial and temporal variability of population exposure? Does ignoring mobility in epidemiologic studies of long-term exposure lead to significant underestimation of population exposure? I answer these questions in Chapter 4 of this dissertation.  1.5 Overview of Dissertation In Chapter 2, I use robust and multiple linear regression models to evaluate the role of local traffic flow (and specific modes) on ambient air quality (measured as PM2.5, NO, NO2 and CO) at hourly time resolution at a busy traffic intersection in New Delhi. These models can provide tools to help estimate the air quality impacts of regulatory restrictions on traffic flow in Delhi.  In Chapter 3, I present LUR models that highlight the spatial and temporal variation of PM2.5, BC and UFP. The PM2.5 and BC models are spatiotemporal, i.e., they can be used to generate citywide pollution maps at hourly time resolution. These models provide critical insights into small-area variations in PM2.5, BC and UFPN concentrations in New Delhi. The LUR models presented in Chapter 3 can be used to inform management strategies and exposure assessment studies.  In Chapter 4, I present a probabilistic simulation framework that integrates PM2.5 LUR models with time-activity data and other factors to estimate the spatial and temporal variability in PM2.5 pollution exposure. This framework can be used to provide improved estimates of exposure while taking mobility of the population into account. The framework is a general one and can be applied to other locations. 19   In Chapter 5, I conclude with an overall summary of my research and a discussion of the implications, limitations and recommendations for future research. I also present some recommendations for policy makers in New Delhi based on conclusions from this research.    20  Chapter 2: The Effect of Localized Traffic Flow on Ambient Air Quality at a Busy Traffic Intersection in New Delhi   2.1 Introduction Rapid growth and urbanization have contributed to the deterioration of ambient air quality in Indian cities, and Indian cities are among the most polluted in the world.75 In 2010 in India, PM2.5 exposure contributed to 600,000 deaths and 17.7 million healthy years of life lost, and was ranked as the fifth highest risk factor for mortality and the seventh highest risk factor for overall disease burden.76, 77 New Delhi, India\u2019s capital, is of special interest not only because of its high levels of air pollution,75 but also because of the exposure levels experienced by a population of 16.7 million.27 Ambient concentrations of fine particulate matter (aerodynamic diameter \u2264 2.5 \u00b5m or PM2.5) ranged from 125\u2013191 \u00b5g m-3 in New Delhi in the year 2014,35 significantly higher than the WHO air quality guideline2 (an annual average of 10 \u00b5g m-3). Emissions from automobiles are widely considered to be a significant driver of increased air pollution in Delhi.42, 47 An estimated 8.3 million motor vehicles ply the streets of New Delhi, a number that has seen an annual growth rate of 7%;35 from 2003 to 2013 there was a 100% increase in the number of motor vehicles in the city.78 During the same period there was a 10% increase in the length of the road network,35 resulting in greater traffic congestion and correspondingly higher levels of vehicular emissions.   Controlling emissions from automobiles has been an important focus of policy measures aimed at improving air quality in urban India. At a national level, the move away from two-strokes engines for use in motorcycles and improved vehicle emissions standards have contributed to improvements in air quality.79 New Delhi has also been the focus of additional regulation aimed at reducing automobile emissions including the conversion of the public transit fleet (buses, taxi cabs and auto-rickshaws) to run on compressed natural gas (CNG)80 and the prohibition on the entry of commercial heavy-duty trucks into the city during the day. These measures have had only a limited impact on the reduction of ambient PM2.5 concentrations.80, 81  21   Though the regulatory focus has been on motor vehicles, there is evidence to show that emissions from other sources may be equally significant, and that automobiles are not the only important source of air pollution in Indian cities. A recent PM2.5 emissions inventory estimated a 17% contribution to primary PM2.5 emissions from traffic (similar to the contribution from other major sources: industries (14%), power plants (16%), domestic (12%) and waste burning (8%).42 The transport sector contributes about 18% of total CO emissions and 53% of total NOx (NO2 & NO) emissions in New Delhi.42 These estimates for PM2.5 emissions from traffic are consistent with source apportionment studies that peg the contribution of primary vehicular emissions at less than a quarter (19\u201324%)46 to less than a fifth (16\u201319%)47 of ambient levels of PM2.5.   Air pollutant concentration at a given location depends on the contributions of local sources in the immediate vicinity as well as more distant sources. Air quality models that integrate local scale processes as well as the regional transport and transformation of pollutants can help identify the drivers of poor air quality on a regional scale. Regional air quality models can also be used to compare the relative effects of the local sources with those of regional sources of air pollutants.82 However, such models require extensive data on the sources of air pollution and surface and upper air meteorological conditions. Consequently, predictions of PM2.5 along roadways have tended to rely upon simplified air quality models that are often used for routine air quality management. These include box models, Gaussian plume models, puff models and statistical\/empirical models.83   Empirical and statistical approaches have often been used to estimate the contribution of traffic to measured gaseous and PM concentrations at near-roadway sites.83-86 These include: tunnel\/roadway studies used to estimate fleet emission factors;87, 88 twin site studies that can isolate the incremental effect of traffic at a given location relative to a background site;89, 90 vehicle specific tracer studies that use chemical analysis;91 and analysis of weekday\u2013weekend differences in gaseous and particulate measurements that might be attributed to traffic. A recent review by Pant et al. (2013)92 concluded that traffic emissions continue to contribute 22  substantially to primary PM emissions in urban areas, although quantitative knowledge of the contribution, especially of non-exhaust emissions to PM concentrations, remains uncertain.   Pollution hotpots at near-roadway locations are a quick way to assess whether regional air quality targets are being exceeded,93 to examine sources and concentration profiles in specific locations,94 to assess the impacts on specific communities95 and to examine the effect of specific mitigation measures on ambient concentrations.96 One particular hotspot in New Delhi, the ITO intersection has been well studied, primarily because it is only location for which India\u2019s Central Pollution Control Board has made daily and hourly data available in the public domain over a long period of time. The ITO intersection is also one of the largest intersections in New Delhi and is thus an ideal location to study the influence of local traffic on ambient air quality, since traffic sources are in the immediate vicinity while other sources of atmospheric pollutants are distant. Data for particulate matter (PM10 and PM2.5), carbon monoxide (CO) and nitrogen oxides (NOx) from the ITO intersection has been used to study the long-term trends and seasonal cycles of air pollutant concentrations,39 and to evaluate the impact of CNG regulation on ambient pollutant concentration.38, 97   There are several contributors to traffic-related PM2.5 at the ITO intersection. These include direct tailpipe emissions of PM2.5, secondary formation of PM2.5 from precursor gases emitted by traffic, non-exhaust PM (brake wear, tire wear and road dust) and the traffic-related urban background concentrations. In addition to accurate estimates of regional and local emission sources, the prediction of PM2.5 and other pollutant concentrations using air quality models also requires an understanding of how primary and secondary components are formed, how they are transported and how they are chemically transformed in the atmosphere. For NO, NO2 and CO the primary proximate sources at the ITO intersection are direct vehicular emissions.   Recent work on the data from ITO has also focused on emission sources and meteorological drivers that explain observed diurnal and seasonal variation in air pollutant concentrations. Chelani (2013)98 used time series analysis and nonparametric wind regression to conclude that local as well as regional and trans-boundary sources influenced PM2.5 concentrations at the ITO 23  intersection. Ghosh et al. (2015)99 studied the variability of synoptic and local airflow patterns that impact aerosol concentrations at the ITO hotspot location. Their work determined that there were substantial regional and sub-regional components in the observed PM2.5 concentrations. Guttikunda and Gurjar (2011)100 used a Lagrangian model to understand the role of seasonality in PM2.5 concentrations in New Delhi. They reported that assuming constant emissions, predicted PM2.5 concentrations were 40\u201380% higher in the winter months and 10\u201360% lower in the summer months compared to the annual mean. This work differs from these previous efforts in that it is the first to quantify the effect of measured local traffic flow on ambient air quality at the ITO location. Estimating the contribution of traffic flows to nearby pollutant concentrations can help us differentiate the effect of localized sources.   2.2 Materials and Methods We measured traffic flow data for multiple vehicular modes (i.e. buses, trucks, cars, auto-rickshaws) that use three different fuels at the ITO intersection. Buses and auto-rickshaws in New Delhi use CNG as fuel, and heavy-duty vehicles (trucks) use diesel. Trucks are only allowed on city roads during nighttime (2200\u20130700). Private cars use both gasoline and diesel, while taxis operate on CNG.  Concentration of gaseous (CO, NO & NO2) & particulate (PM2.5) pollutants at the ITO intersection and meteorological data for New Delhi were obtained from publicly available sources. Multiple linear regression and robust regression approaches were used to model the influence of traffic flow and meteorological variables on ambient concentrations of the pollutants.  2.2.1 Hourly Pollutant Concentrations We used hourly concentrations of NO2, NO, CO, and PM2.5 at the ITO intersection as measured by the Central Pollution Control Board (CPCB) during the duration of the study. Data was accessed through the CPCB website.24 This monitoring station is located by the curbside at the northeast corner of the intersection, with sampling inlet located about 4 meters above the ground.  24  2.2.2 Weather Data Temperature, wind speed and precipitation data for New Delhi for the period of study were obtained from the integrated surface dataset provided by the US National Climatic Data Centre101 for Safdarjung Airport. Weather data were available at a three-hour resolution and were linearly interpolated to estimate the hourly values. We used a proxy for hourly estimates of mixing height \u2013average values for the period of July\u2013September 2011 from the work of Tiwari et al. (2013).102 Consequently, the mixing height profile that we used was not for the sampling year and was assumed to be identical to actual mixing height variation during the sampling period.   2.2.3 Traffic Data Collection  Data on traffic flow was gathered at the ITO intersection by videotaping vehicular movement during night and daylight hours. Daytime traffic data was collected for one full week from 26 June 2007 through 3 July 2007. Traffic was video recorded for 15 min during every hour starting at 8 am and ending at 8 pm. Figure 2.1 provides a schematic of the ITO intersection, including directions of traffic flow relative to the location of video cameras. A video camera was mounted on a pedestrian overpass (in front of Police Headquarters) and focused on the intersection, is marked as box #1 in Figure 2.1. Traffic videos were captured from this spot for all 7 days of the week. This camera captured Flow 1 and Flow 2 throughout the study duration. A second camera was located on another pedestrian overpass for one full day to capture Flow 3 and Flow 4 (Figure 2.1). Figure 2.2 includes a picture of Flow 1 and Flow 2 taken from box 1 at the ITO intersection.  Apart from the one full week of data collection, additional data during the morning and afternoon hours were collected on 16th June (Flow 1 and Flow 2 from 11 am\u20134 pm), 17th June (Flow3 and Flow4 from 11 am\u20133 pm), 20th June (Flow 3 and Flow 4 from 4 pm\u20137 pm) and 22nd June (Flow 3 and Flow 4 from 7 am\u201311 am).   25   Figure 2.1: Traffic flow diagram for ITO intersection. Video camera positioned at box 1 captured the following flows: 1\u20133 (from 1 to 3), 1\u20134, 4\u20131, 2\u20131 and 3\u20131. For one full day (June 27, 2007) a second camera was mounted on another pedestrian overpass (seen as box 2). The second camera captured the following flows: 4\u20132, 3\u20132, 2\u20134 and 2\u20131. For the purpose of simplicity, from here onwards, we refer to the sum of flow 1\u20133 and flow 1\u20134 as Flow1 and the sum of flow 4\u20131, flow 2\u20131 and flow 3\u20131 as Flow2. Further, we will designate the sum of flow 2\u20134 and flow 2\u20131 as Flow3 and the sum of flow 4\u20132 and flow 3\u20132 as Flow4. Flow1 and Flow2 were captured by the camera positioned at box 1 and Flow3 and Flow4 were captured by the camera positioned at box 2.         26   Figure 2.2: A picture of Flow 1 and Flow 2 taken from box 1.    2.2.4 Nighttime Traffic Data  Nighttime traffic data collection was done during the nights of June 26\u201327 (9 pm to 2 am), June 28\u201329 (10 pm\u20133 am) and June 29\u201330 (9 pm\u20132 am). Low traffic volumes and sufficient lighting at the intersection allowed for the collection of nighttime data. Nighttime traffic videos were obtained by mounting a video camera at a corner of the intersection (Figure 2.1). During each hour, Flow 1 and Flow 2 were captured for 15 min and Flow 3 and Flow 4 were captured for 15 min.   27  2.2.5 Traffic Data Coding Traffic videos were played frame by frame in laboratory conditions and traffic was visually classified into five modes: car, motorized two-wheeler, motorized three-wheeler, bus and heavy-duty truck (referred to as truck from here onwards). Individual vehicles were counted for each mode using mechanical tally counters. Classified traffic counts obtained from the traffic videos in 15 min intervals were linearly interpolated to estimate hourly traffic flow.   2.2.6 Missing Flow Estimation We recorded Flow 1 and Flow 2 for one entire week; Flow 3 and Flow 4 were recorded for one full day, and so were missing for 6 days. We estimated Flow 3 and Flow 4 for those days assuming these data were missing completely at random (MCAR).103 We used available data for Flow 1, Flow 2, Flow 3 and Flow 4 to develop linear regression models to estimate the missing values of Flow 3 and Flow 4.  Flow3i = \u03b20i +\u03b21i * Flow1i + \u03b22i * Flow2i + \u03b5i      (1) Flow4i = \u03b20i\u2019 +\u03b21i\u2019 * Flow1i + \u03b22i\u2019 * Flow2i + \u03b5i\u2019     (2) Where subscript \u2018i\u2019 refers to the traffic mode (car, motorized two-wheeler, motorized three-wheeler, bus and truck), and \u03b5i & \u03b5i\u2019 are error terms assumed to be normally distributed. We used Monte-Carlo simulation and determined that parameter estimates obtained from equations (1) and (2) were unbiased.104 We fitted the linear models to estimate missing values of Flow 3 and Flow 4.   For each mode, we aggregated the number of vehicles attributed to Flow 1, Flow 2, Flow 3 and Flow 4 to obtain the count for total vehicles passing through the intersection in an hour. We repeated this for all modes to obtain an aggregated hourly traffic flow through the intersection classified by mode.   28  2.2.7 Missing Value Estimation by Random Sampling During the week of data collection we did not have any information about traffic flow during certain hours. To compensate for this absence we estimated the total traffic and traffic counts by mode through the intersection during these hours by means of random sampling from the collected data at the same hours during other days of the week. These estimated values were incorporated into the data on modal traffic counts.   2.2.8 Regression Models Robust regression models105 and multiple linear regression models were used to understand the relationship between the dependent and predictor variables. Inspecting least squares residuals may not necessarily identify high leverage points.106 Robust regression is used to ensure that high leverage points do not influence parameter estimates.107 In this paper robust models were implemented in the Robustbase package in R,108 this package uses an MM-type regression estimator.109 We chose the natural log of hourly concentrations as the dependent variable; independent variables included natural log of lagged hourly concentrations, hourly traffic flows (by mode and in aggregate), weather variables and proxy for mixing height. We used the logarithm of pollutant since pollutant concentrations tend to follow a lognormal distribution;110 analysis of air pollutant concentrations for all pollutants at the ITO location revealed that they departed from normality and showed a rightward skew. High levels of first-order auto-correlation in hourly pollutant concentrations (0.78, 0.68, 0.6 and 0.60 for PM2.5, CO, NO and NO2 respectively) suggested the use of the lagged term.    To avoid multicollinearity111 we evaluated the correlation between the independent variables. A high degree of correlation was observed between hourly flows of cars, motorized two-wheelers, auto-rickshaws and buses (0.6\u20130.9). Hourly flows for these four modes were therefore summed to create an aggregated independent variable T_Flow. We also checked the correlation between T_Flow and emissions factors weighted flow, for these four modes. The correlation between T_Flow and the emissions factors weighted flow was quite high (0.99\u20130.87) and therefore we just used T_Flow. As expected hourly truck flow did not have a high correlation with other modes and was retained as an independent variable. Final models were further assessed for 29  multicollinearity with the variance inflation factor112 and for autocorrelation in residuals using the Durbin-Watson test.113, 114 Variance explained by each predictor variable in multiple regression models was estimated using the relaimpo package in R.115   The model equations in their most general form for each pollutant were as follows:  log(Conci)= \u03b20 + \u03b21 * log(lagged_Conci ) + \u03b22 * T_Flowi + \u03b23 * Trucki  + \u03b24 * Autoi + \u03b25 *       Tempi + \u03b26 * PCPi + \u03b27 * WSPDi +  \u03b28 * MHTi +    \u03b5i`     (3) where, log(Conci) is log of pollutant concentration in hour \u2018i\u2019 log(lagged_Conci) is log of pollutant concentration in hour \u2018i-1\u2019 T_Flowi is total number of cars, motorized two-wheelers, auto-rickshaws and buses passing through the intersection in hour \u2018i\u2019 Trucki is total number of trucks passing through the intersection in hour \u2018i\u2019 Autoi is total number of auto-rickshaw passing through the intersection in hour \u2018i\u2019 (used only in NO and NO2 models) Tempi is temperature in degrees Celsius for hour \u2018i\u2019 PCPi is precipitation in mm for hour \u2018i\u2019 WSPDi is average wind speed in m\/s for hour \u2018i\u2019  MHTi is a proxy for mixing height for hour \u2018i\u2019  \u03b5i is error term; assumed to be normally distributed.  In addition to the most general model described in equation 3, several alternate models for each pollutant were tested. These were comprised of models that included\/excluded the lagged concentration term and meteorological variables.     30  2.3 Results and Discussion The estimated flows for cars, motorized two-wheelers, autos-rickshaws, buses and trucks for the observed time periods are shown in Figure 2.3. With the exception of trucks, all modal counts were lowest from 2 to 3 am on Sunday. Hourly car flow through the intersection varied between 605 and 6149 per hour, with the highest value observed on a Wednesday between 6 and 7 pm. Hourly motorized two-wheeler flow ranged from 219\u20137487 per hour, with the highest value observed on a Wednesday between 9 and 10 am. Hourly auto-rickshaw flow ranged from 303\u20133076 per hour, with the highest value observed on a Wednesday between 3 and 4 pm. Hourly bus flow ranged from 11\u2013546 per hour, with the highest value observed on a Monday between 9 am\u201310 am. Hourly truck flow ranged from 15\u2013224 per hour, with the highest value observed during Sunday night between 1 and 2 am and the lowest value was observed on Sunday between 11 am\u201312 noon. Modal share for different modes is presented in Table 2.1. Modal share and hourly flow for trucks were higher during the nighttime due to regulatory restriction on truck movement during the day.   Table 2.1: Estimated modal share for the ITO intersection. time car motorized two-wheeler auto-rickshaw bus truck 8am\u20148pm  36% 40% 19.5% 3% 0.5% 8pm\u20148am 43.5% 28% 22% 2% 4%  31   Figure 2.3: Average hourly traffic flow by mode.   Figure 2.4 shows diurnal variation of PM2.5 at the ITO intersection during the period of study; diurnal variation of NO2, NO and CO are shown in Figure A.1, Figure A.2 and Figure A.3, Appendix A. All three pollutants exhibit strong diurnal patterns and the diurnal patterns are different for each pollutant. A visual comparison of the occurrence of traffic peaks with the occurrence of peaks in hourly PM2.5 concentrations (Figure 2.4) shows that the peaks in pollutant concentrations do not coincide with peaks in traffic flow.  32   Figure 2.4: Diurnal variation of PM2.5 at the ITO intersection (September 2007 to December 2014). The solid lines represent the mean values and shaded area represents the 95% confidence interval.   Regression models for PM2.5, NO2 (& NO) and CO are presented in tables 2.2, 2.3 and 2.4 respectively. The proxy variable for mixing height had a counterintuitive sign and its inclusion did not improve model performance; it was therefore dropped. For PM2.5, model 1 (Table 2.2), which did not include the lagged log-concentration term, showed a statistically significant relationship between T_Flow and log-concentration. However, this model, (unlike models that include a lagged log-concentration term) did not pass the Durbin-Watson test113, 114 for autocorrelation, resulting in artificially inflated p-values for model coefficients. Model 2 for PM2.5 included a lagged log-concentration term and had a higher R-squared value (0.65); the lagged concentration term explained 54% of the variability in the PM2.5 concentration. In this model, the variable T_Flow that aggregates all vehicles (with the exception of heavy-duty trucks) and the variable Truck, which is the heavy-duty truck count, and meteorological variables were not statistically significant. Model 3 for PM2.5 is a robust regression version of model 2, and here 33  in addition to the lagged concentration term the variable T_Flow and the variable Truck were statistically significant; model 3 also has a higher R-squared value (0.81) and lower standard error, compared to model 2. Model 4 for PM2.5 excludes meteorological variables that were not significant in all the other models; the results are similar to those from model 3. In each case, the coefficient of the lagged concentration term is about 0.8, which indicates a high level of persistence for the pollutant. We recommend Model 4 as our preferred model for PM2.5 due to its robust nature, lowest standard error and absence of redundant terms. It is important to note that for robust regression models, variability explained by each model term cannot be calculated.  For NO2 we present a number of models (Table 2.3) including some focused on the role of auto-rickshaw emissions. Models 1 and 2 did not include the lagged concentration term, and both models failed the Durbin-Watson test for residual autocorrelation. The addition of the lagged concentration term as an independent variable in model 3 improved model performance by improving the R-squared value, reducing the standard error and removing the residual auto-correlation in the error term. The lagged concentration term was statistically significant at the 95% confidence level and explained 32% of the variability. In model 4 for NO2, the lagged concentration term and the variable Auto (auto-rickshaw counts) had statistically significant coefficients; the lagged concentration term explained 28% of the variability, while auto-rickshaw counts explained 11% of the variability. Models 5 and 6 for NO2 are linear and robust versions of model 4 without the meteorological variables; in both cases, the lagged log-concentration variable was significant, as was the variable Auto. Meteorological variables were not statistically significant in these models. We have focused on NO2 models because of NO2\u2019s adverse health effects.2 Models 7 and 8 are for NO and performed slightly worse than corresponding NO2 models (5 and 6).  Regression models for CO are presented in Table 2.4. Models 1 and 2 for CO did not pass the Durbin-Watson test for residual autocorrelation. In model 3 for CO lagged concentration term and T_Flow were both statistically significant predictors. The lagged concentration term explained 49% of the variability in CO concentrations, and T_Flow explained 7%. Adding truck flow, temperature, precipitation and wind speed did not improve model performance (model 4).  34  Models 5 and 6 for CO are robust regression versions of models 3 and 4. In model 5 the lagged concentration term, T_Flow, temperature and wind speed terms had statistically significant coefficients. Excluding truck flow, temperature, precipitation and wind speed terms did not substantially alter model performance (model 6).  Overall, these results show that the lagged concentration terms explained much of the variability in the hourly ambient concentrations of all pollutants in this study (54% for PM2.5, 43% for NO, 30% for NO2 and 49% for CO). The coefficient of the lagged concentration term was highest across models for PM2.5 and NO (~0.8), followed by CO (~0.7), and NO2 (~0.6). The statistical significance and magnitudes of these coefficients reflect the high degree of first-order auto-correlation in pollutant concentrations or the persistence of the atmospheric concentrations of these pollutants at the ITO intersection. These results are similar to findings of Thomas and Jacko,85 who used regressions and neural network models to predict ambient concentrations of PM2.5 and CO using heavy-duty traffic flow and speed at a freeway location in Indiana, USA. They reported that the lagged concentration term alone could explain about 80% of variability in hourly PM2.5 concentrations and 65% of variability in hourly CO concentrations; the addition of meteorological variables did not improve the performance of their models. The coefficients of the lagged concentration terms in PM2.5 and CO models are also in agreement with values reported by Thomas and Jacko (2007).85   These results show that meteorological variables had little influence on measured concentrations of all three pollutants. This is likely due to the short duration of the study and to the use of rather crude meteorological data. Although mixing height is important in determining pollutant concentrations, failure to include mixing height as a predictor in empirical statistical models should not adversely affect model performance due to high spatial and temporal variability of mixing height and due to inherent limitations of standard equations used to estimate mixing height.116 To a great extent, the lagged concentration term incorporates some of the meteorological effects and the influence of changes in mixing height.    35    Table 2.2: PM2.5 multiple linear regression and robust regression models.  dependent  variable intercept log(lagged_Conci) T_Flowi Trucki Tempi PCPi SPDi R2 a (SE)b  1. Log(PM2.5)  4.6**   4.0\uf0b410-5 * [5%]c 5.5\uf0b410-4 [<1%] -4.0\uf0b410-2  [5%] -3.4\uf0b410-2 [<1%] -3.5\uf0b410-2 [3%] 0.13 (0.48) 2. Log(PM2.5) 1.0 0.82** [54%] -1.5\uf0b410-5 [3%] 7.3\uf0b410-4 [<1%] -1.7\uf0b410-2 [3%] -1.6\uf0b410-2 [<1%] -2.7\uf0b410-2 [2%] 0.65 (0.31) 3. Log(PM2.5) [Robust]d 0.64 0.81** 1.8\uf0b410-5 ** 1.5\uf0b410-3 * -8.0\uf0b410-3 -5.4\uf0b410-3 -1.4\uf0b410-2  0.81 (0.16) 4. Log(PM2.5) [Robust] 0.29 0.82** 1.6\uf0b410-5 ** 1.7\uf0b410-3 **    0.79 (0.16) *Statistically significant at 95% confidence level. **Statistically significant at 99% confidence level. a Model coefficient of determination (R2). b Residual standard error. c Percent of variance explained by the variable. d Robust regression model.            36   Table 2.3: NO2 and NO multiple linear regression and robust regression models.  dependent variable intercept log(lagged_Conci) T_Flowi Trucki Autoi Tempi PCPi SPDi R2 a (SE)b 1. Log(NO2) 3.0**  2.4\uf0b410-5 [2%]c 1.7\uf0b410-3 [1%]  3.6\uf0b410-2 * [4%] 1.4\uf0b410-2 [<1%] -4.4\uf0b410-2 [1%] 0.09 (0.35) 2. Log(NO2) 3.0**   1.9\uf0b410-3  [1%] 3.1\uf0b410-4 * [16%] 2.3\uf0b410-2  [2%] -5.7\uf0b410-3 [<1%] -3.7\uf0b410-2 [1%] 0.21  (0.33) 3. Log(NO2) 1.0 0.60** [32%] 2.0\uf0b410-5 [2%] 1.1\uf0b410-3   1.3\uf0b410-2 [2%] 7.0\uf0b410-3 [<1%] 8.4\uf0b410-3 [<1%] 0.38 (0.29) 4. Log(NO2) 1.3** 0.53** [28%]  1.2\uf0b410-3 [<1%] 2.1\uf0b410-4 * [11%] 7.2\uf0b410-3 [2%] -6.8\uf0b410-3 [<1%] 6.3\uf0b410-3 [<1%] 0.42 (0.28) 5. Log(NO2)  1.48** 0.56** [30%]   2.4\uf0b410-4 ** [19%]    0.49 (0.27) 6. Log(NO2) [Robust]d 1.3 0.58**   2.3\uf0b410-4 **      0.58 (0.21) 7. Log(NO) 0.33 0.82** [43%]   1.4\uf0b410-4 ** [1%]    0.44 (0.41) 8. Log (NO) [Robust] 0.55 0.75**   1.7\uf0b410-4 **     0.48 (0.36) *Statistically significant at 95% confidence level. **Statistically significant at 99% confidence level. a Model coefficient of determination (R2). b Residual standard error. c Percent of variance explained by the variable. d Robust regression model.    37   Table 2.4: CO multiple linear regression and robust regression models.  dependent variable intercept log(lagged_Conci) T_Flowi Trucki Tempi PCPi SPDi R2 a (SE)b 1. Log(CO) 7.6**    3.2\uf0b410-5 ** [7.9%]c 2.3\uf0b410-3 * [4.0%] -2.1\uf0b410-2 [3.6%] 1.6\uf0b410-2 [<1%] 1.6\uf0b410-2 [<1%] 0.16 (0.26) 2. Log(CO) [Robust]d 7.8**  3.6\uf0b410-5 ** 2.1\uf0b410-3 -2.7\uf0b410-2  *  1.3\uf0b410-2  3.5\uf0b410-2  0.30 (0.19) 3. Log(CO) 1.5** 0.76** [49%] 2.2\uf0b410-5 ** [7%]     0.56 (0.19) 4. Log(CO) 2.2** 0.72** [44%] 2.3\uf0b410-5 ** [5.9%] 5.5\uf0b410-4 [2%] -1.6\uf0b410-2  [2.7%] 1.7\uf0b410-2 [<1%] 3.3\uf0b410-2 * [<1%] 0.56 (0.19) 5. Log(CO) [Robust] 2.8** 0.65** 2.4\uf0b410-5 ** 4.4\uf0b410-4   -1.8\uf0b410-2  **  1.6\uf0b410-2  3.6\uf0b410-2 * 0.66 (0.14) 6. Log(CO) [Robust] 2.0** 0.69** 2.4\uf0b410-5 **     0.66 (0.15) *Statistically significant at 95% confidence level. **Statistically significant at 99% confidence level. a Model coefficient of determination (R2). b Residual standard error. c Percent of variance explained by the variable. d Robust regression model.   38   The influence of traffic on PM2.5 concentration is captured by two variables: the variables T_Flow and Truck. Neither variable was statistically significantly in linear regression models, but both were significant at the 95% confidence level or better in the robust regression models (models 3 and 4, Table 2.2) where the effect of outliers is removed. In the robust models the estimated coefficients for Truck are two orders of magnitude greater than the estimated coefficients for T_Flow. This is consistent with known factors for PM2.5 emissions associated with heavy-duty diesel trucks that are significantly greater than those for vehicles that operate on gasoline and compressed natural gas.80 Model 4 (Table 2.2) can be used to estimate the relationship between the variable Truck and ambient PM2.5 concentration. Figure 2.5(a) shows the relationship between the variable Truck and the expected percentage decrease in PM2.5 concentrations if the hourly truck counts were reduced by 50%. For example, given an initial truck count of 120 vehicles per hour, a 50% reduction in truck count, i.e., if we reduce truck count to 60 vehicles per hour, we can expect a 10% reduction in hourly PM2.5 concentrations at the ITO intersection. Such reductions will only be possible during nighttime, since truck flow is restricted during daytime. These findings are relevant for the current policy in Delhi whereby heavy-duty trucks are allowed into the city-limits only during the nighttime.   The influence of the traffic variable T_Flow on CO concentrations was statistically significant at the 95% confidence level across all the models. If we reduce T_Flow from 8000 vehicles per hour to 4000 vehicles per hour we can expect a reduction of 9% in hourly levels of ambient CO. Such a reduction would be achievable during daytime only as T_Flow is quite low during nighttime. The auto-rickshaw count variable Auto was a better predictor of hourly NO2 concentrations than the T_Flow variable. This is consistent with findings that NOx emissions from CNG auto-rickshaws can be high. Authorities in New Delhi had introduced CNG as a mandatory fuel for public transport vehicles in the year 2001. This measure led to improvements in levels of CO and SO2 but had little impact on PM2.5 levels and has been suggested to have increased the levels of NO2.38, 39, 117 We used model 6 (Table 2.3) to estimate reduction in ambient NO2 concentrations if we reduce auto-rickshaw flow by 50% as presented in Figure 2.5 (b). If the initial auto-rickshaw flow is 1500 vehicles per hour, a 50% reduction in auto-rickshaw 39  flow, i.e., if we reduce the from 1500 to 750 vehicles per hour, will result in a 16% reduction in hourly NO2 levels at the ITO intersection. Such a reduction will only be attainable during daytime because auto-rickshaw flow is quite low during nighttime.   Figure 2.5: (a) Expected reduction in PM2.5 concentrations if the initial truck flow is reduced by 50%. (b) Expected reduction in NO2 concentrations if the initial auto-rickshaw flow is reduced by 50%.   Overall, these results show that local traffic flow at the ITO intersection explained limited variability in hourly ambient concentrations of PM2.5. This is consistent with a number of studies including those by Thomas and Jacko (2007)85 for a freeway in Indiana, by Whitlow and co-workers (2011)118 in New York, by Levy and co-workers (2003)84 in Boston, Massachusetts and by Keuken and co-workers (2013)119 in Rotterdam. Street Canyon studies also report small increases in PM2.5 concentrations due to traffic over the urban background.120 Each of these studies reports small effects on PM2.5 measurements from local traffic. It must be noted that these models only explain the variations in hourly pollutant concentrations due to traffic flow at the 40  ITO intersection. These models do not explain the variations in background levels\u2014plausibly captured by the lagged concentration term\u2014which are driven by citywide and regional emissions from a variety of sources including traffic, as well as meteorology. While local traffic flow may not be the only driver of near-road concentrations, it is well established that exposure to traffic-related air pollutants is increasing in developing countries and must be considered in the context of high population density along traffic corridors.15 Reductions in traffic emissions by providing improved and reliable public transit systems along with infrastructure for walking and bicycling, and improved vehicular emission control standards, will reduce the citywide levels of these pollutants.121-123 The results presented here are relevant for understanding the impact of changes in traffic flow at the intersection level, and the results are informative for strategies aimed at reducing PM2.5, NO2 and CO emissions by manipulating traffic flow. The models presented here are applicable for summer season only; separate models can be developed for winter season. We recommend the use of our robust regression models for future analyses. Our results show that reductions in traffic flow on a localized scale are expected to yield limited reductions in ambient concentrations of PM2.5, NO2 and CO on that spatial scale. To achieve significant reductions, management actions related to transport must achieve reductions in traffic volumes (or emissions per vehicle) on a larger spatial scale, i.e., on a citywide or regional scale.           41  Chapter 3: Spatiotemporal Land Use Regression Models of Fine, Ultrafine, and Black Carbon Particulate Matter in New Delhi, India   3.1 Introduction Air pollution has been an environmental and health concern in New Delhi, the Indian capital, for decades. Fine particles (PM2.5 or particulate matter with aerodynamic diameter dp \u2264 2.5 \u00b5m) have been consistently associated with adverse health outcomes.9, 10, 124, 125 Ambient PM2.5 frequently exceeds regulatory standards in New Delhi, with annual average concentrations of 123 \uf0b1 87 \u00b5g m-3 between 2008 and 2011 across seven regulatory monitoring stations.42 Those concentrations are an order of magnitude higher than the guideline value of 10 \u00b5g m-3 (annual mean) set by the World Health Organization.2 For comparison, annual average PM2.5 concentrations in New York City were around 11 \u00b5g m-3 in 2009126 and ~ 100 \u00b5g m-3 or more in Beijing in 2010.127 Black carbon (BC) and ultrafine particles (UFP, < 0.1 \u00b5m diameter) are constituents of PM2.5. BC is an indicator of incomplete combustion, e.g., from biomass burning and motor vehicles.128 Apart from adverse health effects due to inhalation, BC absorbs sunlight and is a climate forcing agent.129 Elevated BC levels have been associated with regional climate effects across India and China.130 Recent research also indicates that BC may be a more specific marker than PM2.5 for health effects.131 Sources of UFP include emissions from internal combustion engines, power plants, incinerators, forest fires, and cooking.132 UFP have also been associated with adverse health effects.132-134   Ambient PM2.5 pollution is understood to be one of the leading risk factors for premature mortality in South Asia, resulting in more than 750,000 premature deaths in 2010.77 Poor air quality in New Delhi is an important public health issue for the city\u2019s 16.7 million people,27, 38, 39, 100, 135-138 yet epidemiologic studies139, 140 have been limited by the quality, duration and spatial coverage of the urban air quality monitoring network. New Delhi has seven continuous monitoring stations for PM2.5. Better coverage may be required to assess the spatial variability of PM2.5 and, hence, the variation in human exposure. In the absence of more detailed monitoring data, land use regression (LUR) models can be an effective tool for assessing within-city 42  variability of air pollution.64 LUR and results are useful for epidemiologic studies, risk assessments, and prioritizing air quality management.53  LUR was developed as an alternative to dispersion models and as a means to assess small-scale spatial variation of air pollutants within urban areas.  LUR models have been developed for many cities in North America and Europe.53, 141 Models may be used to assign population exposure for epidemiologic studies. In brief, a statistical relationship is established between land use characteristics and pollutant concentrations measured by targeted sampling at a limited number of sites, and then, the relationship is used to predict pollutant concentrations at unmeasured locations throughout a given domain. Generally, LUR has been used to estimate mean annual concentrations of a pollutant, using one to two week sampling at 20\u2013100 sites.53 Land use, road network, population density and traffic flow variables are typically used as inputs to the models, though LUR models have been reported to perform well even in the absence of traffic flow data as an input.67 The site selection processes vary67-70 but in general a minimum number of spatially dispersed sites characterizing different land uses that are able to capture the spatial variation of a pollutant for a given domain are required. A small number of sites and a large number of predictor variables can lead to inflated R2 values.142  Air pollution in Indian cities likely has a wider range of sources than air pollution in the European and North American cities where LUR has been previously applied: not just traffic and industry, but also small-scale, distributed sources such as biomass burning for cooking and heat, open burning of solid waste, and diesel generators for backup power.28, 143, 144 In the case of New Delhi, there is also substantial seasonal variation in absolute concentrations of PM2.5 and percentage contributions from different sources.46, 100, 102     There is also substantial variation in PM2.5 concentrations during daylight hours (diurnal variation).42 To capture this trend, we sought to develop separate LUR models for morning and afternoon periods. Specifically, PM2.5 and BC models were developed to predict the spatial distribution of pollutants over time, using data from a fixed continuous monitoring site. The models are thus spatiotemporal rather than simply spatial. These models could be used to obtain 43  estimates for any given time interval within the study period. For example, hourly concentration estimates can be obtained if we input hourly pollutant concentration in a model equation. As no continuous ultrafine particle number concentration (UFPN) measurements were available, the UFPN models only describe spatial variability in the morning and afternoon hours for the duration of the study.     3.2 Materials and Methods  3.2.1 Site Selection We undertook field measurements of air quality in New Delhi during February\u2013May 2010, encompassing the local spring and summer seasons. We employed a site selection approach similar to what Brauer et al.74 employed for the TRAPCA study in three European cities with about 40 sites in each city. We used local knowledge, Google Earth, and city maps (1:5500 to 1:12,500) to classify neighborhoods based on the following criteria: population density, distance to the city center (Connaught Place), residential or commercial type, density of the road network, and green spaces. Sites were allocated to neighborhoods that captured maximum variation in these variables. We made minor adjustments to site locations during the monitoring campaign on the basis of preliminary site visits; adjustments were less than 10 m in magnitude and were made to ensure that a spot was available to leave monitoring devices undisturbed at normal breathing height for the sampling period.   3.2.2 Instrumentation and Field Measurements Data collection occurred in parallel with fieldwork for a companion study of urban and in-vehicle exposure to particulate matter in New Delhi.145 The air quality sampling instruments, protocols, and post-processing techniques have been described previously.145 Continuous monitoring of PM2.5 and BC was conducted at a fixed rooftop site located in a high-income residential neighborhood in southern New Delhi (Figure 3.1). The rooftop provided a background site that was relatively free from the influence of local traffic and point sources, so that data from this site could be used to characterize citywide diurnal and seasonal trends. 44  Sampling at LUR sites was divided into morning and afternoon sampling periods. At each site, measurements were collected for 1\u20133 h during the morning (0800\u20131200) and\/or afternoon (1200\u20131800). Given equipment limitations, only one LUR site (plus the one fixed-location) was sampled at a time. Sampling at LUR sites was conducted close to normal breathing height, in contrast to regulatory monitoring which is often on rooftops.  3.2.2.1 GPS and Meteorological Data We used a GPS device (GPSMap 60CSx, Garmin Inc.) with an accuracy of  \u00b1 3-5 m to record the spatial coordinates of all sites at the time of data collection. We recorded meteorological data (temperature, relative humidity, wind speed, wind direction, and rainfall; all recorded at 5 min intervals) via a weather station (Model PWS\u20131000TD, Zephyr Instruments, East Granby, CT) at the fixed (i.e., central rooftop) location.     Figure 3.1: Land use regression sites and rooftop site overlaid on major road network.   45  3.2.2.2 PM2.5 Data Collection Fine particulate matter was measured using two TSI DustTrak 8250 aerosol monitors (TSI Inc., Shoreview, MN, USA) fitted with PM2.5 impactor inlets. The DustTrak infers particle mass concentrations based on 90\u00ba light scattering measured by a laser photometer. This detection method is subject to error because relative humidity (RH) and particle properties (i.e., density, shape, size, refractive index) influence particle light scattering.145-147 To account for the RH effect, we applied an empirical correction equation146 to the raw DustTrak measurements using 5 min average RH measurements from the rooftop fixed site. The regular sampling program of DustTrak measurements was supplemented with ~35 colocated, time-integrated (~1\u20134 h) gravimetric measurements of PM2.5 collected with a single-stage impactor, from which a nonlinear gravimetric calibration curve was developed145. The final PM2.5 mass determination was obtained by applying this gravimetric calibration to the time-resolved RH-corrected DustTrak observations.148  During each measurement session, PM2.5 data collection was carried out simultaneously at the central rooftop location and at one of the LUR sites. PM2.5 data were collected at 48 LUR sites (44 afternoon sessions, 22 morning sessions; Figure 3.1). Nine afternoon sites and three morning sites were later dropped from the analysis because population and road network data were not available in GIS format at those locations. This could have been avoided if the availability of these data (in GIS format) was verified before the sampling. We employed a time resolution for measurements of 30 s (rooftop site) and 1 s (LUR sites).  3.2.2.3 Black Carbon Data Collection Similar to PM2.5, BC data collection was carried out simultaneously at the central rooftop location and at one of the LUR sites. Concentrations were measured using portable aethalometers (model AE-51 microAeth, Magee Scientific, Berkeley, CA). A previously developed empirical correction factor149 was validated for use in New Delhi145 and then applied to raw BC data to correct for underestimation of BC concentrations with increasing aethalometer filter loadings.149, 150 Five-minute moving averages were used to remove sharp, consecutive negative, and positive concentrations peaks that can reflect measurement artifacts.145, 151 Overall, BC concentrations 46  were measured at 30 LUR sites (29 afternoon sessions and 20 morning sessions).  Four afternoon sessions and three morning sessions were dropped from the analysis because population and road network data were unavailable in GIS format at those locations.  3.2.2.4 UFPN Data Collection Ultrafine particle concentrations were measured using a Condensation Particle Counter (CPC, model CPC 3007, TSI Inc., Shoreview, MN). Although this instrument provides the total number count of all particles in the size range 10 nm < dp < 1 \u00b5m, this result closely approximates UFPN (dp < 100 nm) under the conditions encountered in this study. Concentrations were measured at 1 Hz. Because UFPN concentrations often exceeded the upper measurement limit (105 particles cm-3) of the CPC,145 we employed a custom-built dilutor, which reduced inflow concentrations by a factor of 5.5.  In addition, we applied the empirical correction factor of Westerdahl et al. (2005)152 to account for particle coincidence errors when diluted UFPN concentrations exceeded 105 particles cm-3.  Sampling for UFPN was conducted at LUR sites only, because a second CPC was not available. We measured UFPN concentrations at 48 LUR sites (46 afternoon sessions and 21 morning sessions). UFPN data for nine afternoon sites and two morning sites were dropped from the analysis because population and road network data were unavailable in GIS format.  3.2.2.5 Data Reduction and Quality Control Processed data for all pollutants were used to obtain hourly medians for the LUR sites. Any hour with less than 15 min of data was discarded. The arithmetic mean of these medians for morning or afternoon hours was assigned to an LUR site as its concentration for morning or afternoon hours, respectively.  For the rooftop site, 10th percentile concentrations were computed for each hour, and any hour with less than 15 min of data was discarded. The 10th percentile was selected to approximate the urban background diurnal profile and to be free from the influence of short-duration peaks due to local sources. The arithmetic mean of the hourly 10th percentile concentrations corresponding to 47  the sampling period at an LUR site was computed for each LUR site. The natural logarithm of this value was used as an independent variable in the model building processes for PM2.5 and BC and is hereafter called ln(ROOF).  3.2.3 Spatial and Socioeconomic Variables We used maps in a GIS to generate 14 spatial variables related to land use and demographics (Table 3.1). Shape files were converted into rasters of 5 m cells using ESRI ArcGIS 9.3. Separate rasters were created for major roads, minor roads, and green spaces. The road rasters were used to estimate road length around each LUR site for chosen radii, and to estimate the shortest distance from each LUR site to the nearest major and minor roads. A green-space raster was used to obtain area of green-space around each LUR site for chosen radii.   We used Indian Census data to generate independent socioeconomic variables (Table 3.1) as possible surrogates for sources such as domestic wood and waste burning. Census data were based on the 2001 census by the Government of India, available at the ward level for New Delhi. A ward is an electoral unit for the local municipal government, with a total of 156 wards in New Delhi. Population attributes per ward were assigned to the ward centroids and the Spatial Analyst feature in ArcGIS was used to obtain smoothed density surfaces67 for the total population and for two indicators of low socioeconomic status: illiterate population and Scheduled Caste and Scheduled Tribe population. Scheduled Caste and Scheduled Tribe are defined in Articles 341 and 342, respectively, of the Indian Constitution and are generally considered to be socioeconomically disadvantaged groups.153 In total, 14 variables were considered in LUR model development.   3.2.4 Model Building Air pollutant concentrations are typically log-normally distributed.154 We used the natural logarithm of the mean of measured hourly medians at each LUR site as the dependent variable in each model, following verification with the Kolmogorov\u2013Smirnov test155 in the R statistical software package.156 Eligible independent variables included the land use and socioeconomic variables (Table 3.1) for all models and ln(ROOF) for the PM2.5 and BC models.  48   For PM2.5 and BC models, we assumed that the pollutant concentrations were associated with a multiplicative combination of a background temporal component (from the rooftop site) and the spatial components. We also assumed that (a) the temporal component was spatially invariant and (b) the spatial components were temporally invariant. The first assumption is supported by the existence of a strong temporal correlation between measurements at the LUR sites and the rooftop site. The second assumption is reasonable given that changes in the spatial variables would be small over the modeling period. For example, we do not expect that the patterns of population density changed over the 13 weeks of sampling.     We generated two temporal models per pollutant, each based on log hourly median concentrations: morning models (observations 0800\u20131200), and afternoon models (1200\u20131800). We used a model-building algorithm similar to the one developed by Henderson et al. (2007),67 which was designed to produce a parsimonious model in which the influence of individual variables is interpretable and consistent with a priori assumptions about determinants of spatial variability in air pollution. A statistical relationship was established between the observed pollutant concentrations at the LUR sites, the potentially predictive land use and socioeconomic variables, and the rooftop concentration variables (for PM2.5 and BC). These relationships were used to predict pollutant concentrations throughout the spatial domain. The model-building algorithm is as follows: 1. Each variable was ranked based on its correlation with the log of hourly median concentration. 2. The top-ranking variable in each sub-category (e.g. major roads) was identified. 3. Variables in each sub-category that had a correlation of 0.6 or higher with the top-ranking variable (Step 2) were dropped. 4. Remaining variables were included in robust linear regression models. 5. Variables that were not significant at a 90% confidence level or that had a coefficient with a counter-intuitive sign were dropped.  6. Repeat Steps 4 and 5 to convergence.   49   Table 3.1: Potentially predictive independent variables.  variable description median (IQRa) DIST.J1 shortest distance to the nearest major road from a cell (m) 65 (24\u2013130) DIST.J2 shortest distance to the nearest minor road from a cell (m) 37 (14\u201377) RD1.100 RD1.250 RD1.500 major road length in a buffer of 100, 250 and 500 m radii respectively (m). 200 (0\u2013271) 672 (467\u20131040) 2210 (1349\u20133534) RD2.100 RD2.250 RD2.500 minor road length in a buffer of 100, 250 and 500 m radii respectively (m). 247 (42\u2013459) 1315 (586\u20132745) 5600 (3177\u201311351) GREEN.100 GREEN.250 GREEN.500 area of green space in a buffer of 100, 250 and 500 m radii respectively (m2). 0 (0\u20132119) 2425 (0\u201318437) 31750 (981\u201398200) SC.5000 density of Scheduled Caste and Scheduled Tribe population (persons\/hectare) obtained using density analysis with 5000 m radius. 25 (17\u201344) ILL.5000 density of illiterate population (persons\/hectare) obtained using density analysis with 5000 m radius. 50 (27\u201386) POP.5000 density of population (persons\/hectare) obtained using density analysis with 5000 m radius. 177 (97\u2013289) ln(ROOF) log of mean hourly 10th percentile concentrations from the rooftop site. 4.4 (4\u20135) PM2.5 1.6 (1.1\u20132.7) BC aInterquartile range.   Steps 1 and 2 of the model-building algorithm provide ordered sets of the most predictive variables and step 3 avoids collinearity between independent variables from the same subcategory. Diagnostic plots in step 4 showed a few high leverage points in almost all models. Robust linear regression was used to prevent the undue influence of the high leverage points on coefficient estimates.105 Steps 4 and 5 were repeated to convergence, and the resulting models had those variables only that were statistically significant at 90% confidence level. Here we present the regular linear regression coefficients and statistics for the independent variables 50  selected using this model-building algorithm. Robust linear regression coefficients and statistics are presented in Appendix B (Table B.1). For multiple regression models with more than one independent variable, we also report the variance explained by each predictor variable, estimated using the relaimpo package in R.115  3.2.5 Model Evaluation Models were evaluated with diagnostic plots and leave-one-out cross validation.157 We used ESRI ArcGIS9.3 to obtain Moran\u2019s I statistic to check for presence of spatial autocorrelation158 in the LUR models. For the PM2.5 and BC models, Moran\u2019s I was calculated for every two week period, in the morning and afternoon models. For UFPN models, Moran\u2019s I was calculated for the morning and afternoon models for the complete duration of the study. All models were also checked for temporal autocorrelation using Durbin\u2013Watson test.113   3.2.6 Regression Mapping We rendered the regression models as concentration maps for the study domain using the Spatial Analyst feature in ArcGIS. We used coefficient estimates obtained in the model building process to predict pollutant concentrations for each cell.   3.3 Results  Median  (interquartile range) concentrations for PM2.5, BC, and UFPN across all LUR sites were, respectively, 133 (96\u2013232) \u03bcg m-3 PM2.5, 11 (6\u201321) \u00b5g m-3 BC, and 40 (27\u201372) \u00d7 103 particles cm-3 (Table 3.2). These pollutant concentrations are substantially higher than those reported in other LUR studies.53 The correlation between the hourly median concentrations of PM2.5 and the corresponding BC concentrations at LUR sites was 0.79. At the LUR sites the coefficient of correlation between the hourly median concentrations of PM2.5 and UFPN was 0.42, and for BC and UFPN the correlation was 0.51. The coefficient of correlation between hourly medians at the fixed rooftop site and the corresponding hourly medians at the LUR sites was 0.84 for PM2.5 and 0.85 for BC.  Average temperatures recorded during the sampling campaign were as follows: 20 \u00b0C for February, 28 \u00b0C for March, 36 \u00b0C for April, and 37 \u00b0C for the first week of May. 51   Values of Moran\u2019s I confirmed the presence of spatial autocorrelation, such that sites with higher concentrations were more likely to be closer to other sites with higher concentrations. No evidence of temporal autocorrelation was found using the Durbin-Watson test. Monthly diurnal variation plots for PM2.5 and BC have been included in Appendix B (Figure B.1 and B.2).   3.3.1 PM2.5 Models The morning PM2.5 model (Table 3.3, Figure 3.2(a)) explained 85% of the variability in measured concentrations. The explanatory variables were population density, distance to nearest major road, and ln(ROOF), which accounts for citywide day-to-day and diurnal variation. Coefficient estimates indicate that the predicted concentrations increased with population density and decreased as the distance to nearest major road increased.   Table 3.2: Descriptive statistics for hourly median concentrations of PM2.5, BC, and UFP at the land use regression sites.  species GM (GSD) median min P10 P25 P75 P90 Max N (sites, h) PM2.5   (\u00b5g m-3)  140  (1.8) 133 40 61 96 232 335 680 39 136 BC (\u00b5g m-3)  12  (2.6) 11 2 4 6 21 43 140 26 112 UFPN (103 cm-3)  43 (2.0) 40 7 18 27 72 113 190 39 147   The afternoon PM2.5 model (Table 3.3, Figure 3.2(a)) was more poorly fit than the morning model. The explanatory variables were population density and ln(ROOF). To understand the relative importance of temporal and spatial variation, we dropped the ln(ROOF) term from both the morning and afternoon models but no useful models were obtained. The ln(ROOF) term explained most of the variability (69%) and the spatial variables explained much less variability (16% in morning and 4% in afternoon), suggesting a dominant role of temporal variation in predicting the PM2.5 measurements.  52  3.3.2 Black Carbon Models The morning BC model (Table 3.3, Figure 3.2(b)) explained 86% of the variability in measured BC concentrations.  The explanatory variables were population density and ln(ROOF). The afternoon model (Table 3.3, Figure 3.2 (b)) explained 69% of the variability in the measurements. For both morning and afternoon BC LUR models, distance from the nearest major road was chosen as model variable using the robust linear regression selection algorithm (Table C.1, Appendix C), but its coefficients were not statistically significant in the regular linear regression models. That finding suggests that robust regression can improve the performance of LUR models. The morning BC LUR model was a better fit than the afternoon model. To understand the relative importance of temporal and spatial variation, we dropped the ln(ROOF) term from both BC models; as with PM2.5, no useful models were obtained. The ln(ROOF) term explained most of the variability (63\u201377%) and the spatial variables explained much less variability (6\u20139%), suggesting a dominant role of temporal variation in predicting the BC measurements. For both PM2.5 and BC, predicted spatial variability was higher for morning models compared to afternoon models.   3.3.3 UFPN Models Unlike the PM2.5 and BC models, the UFPN models did not include measurements from the fixed rooftop site as an independent variable. The morning model (Table 3.3, Figure 3.2(c)) had population density as the only statistically significant predictor. The afternoon model (Table 3.3, Figure 3.2(c)) included population density and minor road length in a buffer of 500 m as statistically significant predictor variables.         53  Table 3.3: Final spatiotemporal LUR model specifications and results for PM2.5, BC and UFPN.  model model terms and coefficients: log-median concentration a R2 b LOOE c constant ln(ROOF) POP.5000 DIST.J1 RD2.500   PM2.5 (morning) 1.22** 0.77 ** 2.0 \u00d7 10-3 ** -8.2 \u00d7 10-4 **  0.85 0.25 N = 19d  (69%)e (10%) (6%)    PM2.5 (afternoon)  0.01 0.98 ** 1.5 \u00d7 10-3 **   0.73 0.29 N = 35  (69%) (4%)     BC  (morning) 0.80** 0.73 ** 2.2 \u00d7 10-3 ** -7.2 \u00d7 10-4  0.86 0.34 N = 17  (77%) (7%) (2%)    BC (afternoon) 0.37 0.83 ** 1.5 \u00d7 10-3 * -1.0 \u00d710-3  0.69 0.44 N = 25  (63%) (4%) (2%)    UFPN (morning) 10.35**  3.3 \u00d7 10-3 **   0.28 0.59 N = 18        UFPN (afternoon) 9.79**  2.3 \u00d7 10-3 *  4.7 \u00d7 10-5 * 0.23 0.72 N = 37   (11%)  (12%)   a Spatiotemporal LUR models predict the natural log of the hourly median concentration at each fixed site. b Model coefficient of determination (R2). c Leave-one-out cross-validation error. d N - number of sites. e Percent of variance explained by the variable. **Statistically significant at 95% confidence levels. *Statistically significant at 90% confidence level. 54   Figure 3.2: Predicted average PM2.5, BC, and UFPN spatial variation for the duration of the study. Transect for Figure 3.3 displayed in part (a).   55  3.4 Discussion This study documents the first PM2.5, BC, and UFPN LUR models developed for New Delhi, India, and among the first developed for highly polluted locations in rapidly developing economies.159, 160 The PM2.5 and BC models are spatiotemporal65 in nature, leveraging data from a fixed continuous monitoring site. As such, the PM2.5 and BC models can be used to predict pollutant concentrations for smaller time periods (~1 h periods, which is in contrast to the annual averages predicted by most conventional LUR models). We developed separate models for morning and afternoon hours because of the strong diurnal patterns. For all three pollutants, the spatial patterns were widely different for morning and afternoon hours. Spatial patterns of PM2.5 and BC were similar at both times of day, and both were different from UFPN. The LUR models predicted higher concentrations for, and high variability during, morning hours. Predicted pollutant levels (Figure 3.3) along an arbitrary transect for morning and afternoon models illustrate this point. This observation (concentrations are more uniform during afternoon than during morning) is likely attributable to more rapid meteorological dispersion in the afternoon.42, 148, 161 Concentrations of BC, and spatial autocorrelation in modeled BC concentrations, are higher in morning than in afternoon. The UFPN model also showed greater spatial autocorrelation in morning than in afternoon.  56   Figure 3.3: Predicted average PM2.5, BC, and UFPN variation along an arbitrary transect (from point 1 to point 2, Figure 3.2(a)). The red lines indicate results for the afternoon, and the blue lines indicate results for the morning.   High correlation between measurements at the LUR sites and measurements at the fixed site indicate the important role of temporal variability and urban background concentrations. The high correlation between BC and PM2.5 measurements at LUR sites and the similar spatial patterns suggest that similar factors (sources and\/or meteorology) may be driving both concentrations. Different spatial patterns and weaker correlation between UFPN and the other pollutants indicate that the factors influencing UFPN may be different from those for PM2.5 and BC. The observed correlation between PM2.5 measurements at LUR sites and PM2.5 measurements at the fixed site is higher than the reported correlation in a similar study in Massachusetts that used a similar modeling approach.162   57  Mapping the spatial variability of air pollutants is useful to understand population exposures, and the results can complement information obtained from the regulatory monitoring networks used for air quality management and planning. LUR models typically use one or two weeks of data collected simultaneously at a number of sites.53 Such monitoring is resource intensive, with the cost of a traditional PM2.5 LUR study (including ~ 40 sites) being approximately 30,000 Euros, assuming that the air-monitoring equipment is available to use.53 Purchasing the monitoring equipment would roughly double the cost. These levels of funding are often unavailable in developing countries. The spatiotemporal approach used here is applicable in more resource-constrained situations, and these results demonstrate that a reasonable understanding of spatiotemporal patterns of PM2.5 and BC can be developed. Furthermore, we expect that performance of UFPN LUR models will improve with the availability of continuous monitoring data from a fixed site. This approach is similar to the development of LUR models with mobile monitoring, which has been explored elsewhere.163, 164   There are other challenges to developing LUR models for cities in developing countries. Socioeconomic data related to the independent variables in LUR are often not available, or are available at a coarse spatial resolution. For example, the models for New Delhi show that higher levels of PM2.5, BC, and UFPN are associated with higher population density. However, population data were available at ward level only; the average ward size is ~100,000 people (Delhi: 16.7 million people, 156 wards). Availability of population data on a finer scale may improve model performance, as might data on other sources such as biomass burning. The available geographic predictors explained a much smaller proportion of variability in measured PM2.5, BC and UFPN than observed in other studies,53 indicating the importance of temporal variability and the likelihood of uncharacterized sources and sources that do not correlate with land-use. Performance of the LUR models is in part limited by the influence of uncharacterized sources such as biomass burning (sources that are not necessarily correlated with land use) and the likely impact of a far greater number of distributed sources compared with developed country settings. The presence of spatial autocorrelation is also a limitation of the LUR models. We did not include spatial clustering in the predictions as it decreases interpretability by replacing important but unknown predictors. 58   The LUR models were limited by the availability of geographic data for constructing predictor variables. Other LUR models have considered more than 100 variables to describe spatial contrasts,165 compared with the 14 available for this analysis. It is recognized that a larger number of sampling sites can benefit validation of LUR models and is recommended for future projects.166 Finally, the models are only applicable from February through May, which is when the measurements used here were conducted. Separate models could be developed for other months using the approach employed here.    59  Chapter 4: Estimating PM2.5 Population Exposure in New Delhi, India, Using a Probabilistic Simulation Framework and Spatiotemporal Land Use Regression Models   4.1 Introduction Air pollution is a growing environmental and public health concern in Indian cities, especially in its capital, New Delhi. Annual average concentrations of fine particulate matter (PM2.5 or particulate matter with aerodynamic diameter \uf0a32.5\u00b5m) in New Delhi are considerably higher than the WHO Air Quality Guideline2 of 10 \u00b5g m-3. Annual average concentrations at specific locations ranged from 125\u2013191 \u00b5g m-3 in 2014.35 In 2010, exposure to PM2.5 contributed to 0.6 million deaths and 17.7 million healthy years of life lost in India; exposure to PM2.5 is ranked as the fifth highest risk factor for mortality and seventh highest risk factor for overall disease burden in India.76, 77   While there are multiple sources of air pollution in New Delhi, including biomass burning, power plants, industry and vehicular traffic, there is no single dominant source of PM2.5. Source apportionment studies have shown that primary traffic emissions account for about 20% of ambient PM2.5 mass concentration.46, 47 The number of motor vehicles in New Delhi has been increasing at 7% per year35 for the past decade. In 2012, there were roughly 2.3 million cars, 4.6 million motorized two-wheelers, 88,000 auto-rickshaws and 34,000 buses plying the streets of New Delhi. Transport and land-use policies in New Delhi that have tended to discourage walking and bicycling have exacerbated this trend.167 The increased number of motor vehicles has led to high volumes of traffic and congestion-related air pollution hotspots,39, 168 with significant consequences for public health.169   PM2.5 measurements are only collected at eight fixed locations in New Delhi, and may not capture the spatial heterogeneity of particle concentrations. This can to lead to inaccurate estimates of population exposure, which may impede air quality management efforts and will lead to exposure misclassification, thus potentially biasing epidemiologic analyses.51 Direct 60  measurement of personal exposure is prohibitive for large, representative samples, and the small number of observations in personal monitoring studies can lead to biased estimates.52 Exposure modeling is an inexpensive alternative to personal monitoring for assessing the magnitude of and variability in population exposure.55, 60 Spatiotemporal air quality models such as land use regression (LUR) models that map ambient concentrations of pollutants have been widely developed and applied to population exposure assessment in urban settings. These models can be used to integrate spatial variation in atmospheric concentrations with time-activity based models to estimate exposure of a large number of individuals.72-74  To date, studies of population level exposure to air pollutants, including those using LUR models, have typically not accounted for population mobility and the role of exposures encountered during transportation.52, 53 Commuting is generally a high-exposure activity in areas where the vehicle density is high,170 and can contribute substantially to total daily exposure171 and consequent health effects. Recent studies show that in-vehicle exposure to PM2.5 is associated with cardiovascular effects in healthy adults.172 Ignoring mobility negatively biases the effect estimates since air pollution exposure tends to be higher during commutes.62 The choice of travel mode might therefore be an important determinant of air pollution exposure for commuters,173 especially in places such as New Delhi with high levels of congestion. Additionally, modal choice is highly associated with socioeconomic status (SES), and modes used by low-income commuters might result in greater exposure. A population with a lower SES may also be more susceptible to adverse health effects of environmental pollutants174 including air pollutants.175 However, a simple comparison of exposure to pollutants across different modes of transport can be misleading,176 since the effect of modal choice needs to be understood relative to total exposure.   The objective of this study was to quantify the annual average exposure to PM2.5 and its variability for working and non-working adult sub-populations in New Delhi, with an emphasis on assessing the impact of exposure during commutes relative to exposures at home and work locations. A probabilistic simulation framework was developed to quantify the exposure of individuals in a spatially explicit manner, incorporating exposures at home, at work and for 61  different travel modes during commutes to and from work. We further estimated population exposure as a function of seasonal and diurnal spatiotemporal variations that influence PM2.5 concentrations in Delhi.    4.2 Materials and Methods  Exposure to an air pollutant is determined by the concentration of the pollutant and the time spent in contact with the pollutant.63 Total exposure thus depends on the sum of exposures in different microenvironments as shown in Eq. (1).  E=\u2211fi Ci  i= h,w,t   (1) where, E is the time-weighted average exposure of an individual and fi and Ci are the time fraction spent in and concentration in i different microenvironments (h = home; w= work; t = travel) respectively.   We used a probabilistic simulation framework to estimate the variability in exposure to PM2.5 of outdoor origin averaged across a simulated population. Simulation approaches have been used by a number of researchers to model exposure to air pollutants.56, 58-60, 177 For the purpose of simulation, it was assumed that there are three microenvironments that adequately capture an individual\u2019s daily PM2.5 exposure: home, work and transport. Only PM2.5 of outdoor origin was considered in this study\u2014in other words, the indoor sources of PM2.5 were ignored. The simulation assumes that ambient air quality at a specific location, as predicted by an LUR model, is sufficient to estimate exposure. The common caveat that the simulation outputs are based on a combination of modeled and empirically estimated inputs, and so have not been evaluated, also applies. A visual representation of the simulation process is provided in Figure 4.1. A detailed description of the simulation algorithm is provided in Appendix C.  62   Figure 4.1: A visual representation of the simulation process. *Indicates that the step was repeated 100,000 times.                             *Select home zone; select work zone using Gravity model; and select home and work locations *Extract morning and afternoon concentrations at home and work locations using LUR models *Compute distance between home and work locations and create a distribution for distance Use planning data for zones to build a Gravity model for trip generation [Eq. (2)]  *Assign travel time fraction and generate time spent at work *Use morning and afternoon concentrations to compute nighttime and average daily concentrations at home locations using diurnal patterns from fixed site monitor (LUR) *Compute microenvironmental concentrations for home, work and transport microenvironments for March (use kvm ratio for Scenario 3) LUR models: 1. Create distributions for morning and afternoon concentrations for major and minor roads 2. Create distributions for kvm ratios for each mode  For each month, scale microenvironmental concentrations using monthly patterns from fixed site (ITO); and multiply home and work microenvironmenal concentrations by I\/O ratio  Create distributions for I\/O ratios for summer and non-summer months  Compute average exposure for each month and full year [Eq. 1] 63  We implemented three simulation scenarios:  Scenario 1: Individuals were assumed to spend all their time at home. This scenario may be applicable to the very young and old segments of the populations and is a typical assumption of epidemiologic studies of long-term exposure to air pollution where exposures are estimated based on residential locations. Scenario 2: Individuals were assumed to travel for work or educational purposes; in-vehicle PM2.5 concentrations during commute were assumed to be identical to the near-road PM2.5 concentrations. Scenario 3: Individuals were assumed to travel for work or educational purposes; in-vehicle PM2.5 concentrations during the commute were estimated using in-vehicle to near-vehicle concentration ratios (\u03bavm). This scenario builds on Scenario 2 to incorporate differences in exposure based on mode of transport.  4.2.1 Key Input Data and Models Four datasets, including two primary datasets developed specifically for this simulation and two secondary datasets, were used. The secondary datasets included zonal planning data used to develop a gravity model for trip generation, and output from LUR models developed previously for Delhi.178 The primary data sets included travel-time survey data for New Delhi residents and measurements from an in-vehicle exposure monitoring campaign conducted on a busy road in New Delhi.   4.2.1.1 Zonal Planning Data for Home\u2013Work Trip Generation We used 16 planning divisions or zones defined by Delhi Development Authority179 for trip generation. A smoothed population density domain was created using population data in GIS (ArcGIS10). We selected 100,000 points from the population density domain; the probability of selection of a point was proportional to the population density at that point. This set of points (PTS_POP) was used for selecting home locations for each simulated individual. We also selected another set of 100,000 random points (PTS_RAND) to represent work locations for each simulated individual. In either case, the probability of selection was set to zero for roads and water bodies.  64   A home zone for a simulated individual was selected first, with the probability of selection proportional to the population in that zone. The simulated individual\u2019s home location was then selected by randomly sampling from the subset of PTS_POP lying within the home zone. A work zone was similarly selected by random sampling with probability of selection proportional to the employment and enrollment numbers in that zone. If the work zone was different from the home zone, the work zone was redefined using a Gravity model.180 Probability of selection for the redefined work zone was directly proportional to the sum of employment and enrollment numbers, and inversely proportional to the distance between the centroid of the work zone and the centroid of the home zone as in Eq. (2). Thereafter work locations were selected by randomly sampling from the subset of PTS_RAND that lay within the boundaries of the work zone. We then computed the distance \ud835\udc37\u2032(\ud835\udc56, \ud835\udc57) between the home and work locations for each simulated individual. Each individual in the simulation had a unique combination of home and work locations.  \u03a1(\ud835\udc56, \ud835\udc57) = \u039aEmp(j)(\u03a3Emp(j))\u00d7\ud835\udc37(\ud835\udc56,\ud835\udc57)      (2) where P(i,j) is probability that a simulated individual with home zone i will work in zone j; Emp(j) is the sum of employment and enrollment numbers in zone j; D(i,j) is the distance between centroids of zones i and j; and \u039a is a constant.  4.2.1.2 Output From Spatiotemporal LUR Models We used morning (0800\u20131200) and afternoon (1200\u20131800) PM2.5 concentration maps for the month of March (2010) from recently developed LUR models for Delhi.178 We extracted morning and afternoon PM2.5 concentrations at home locations from the spatiotemporal LUR models. Spatiotemporal LUR models (PM2.5) were developed for morning and afternoon hours using data from 39 sites. Independent variables that included data from a fixed site monitor located in a residential neighborhood, along with distance from major and minor roads and population density, were found to be significantly correlated with PM2.5 concentrations. Other environmental variables such as proportion of green space, and socio-economic variables such as caste composition, were included but were not found to be significant. The models explained 65  more variability in morning (85%) than in the afternoon (73%). Nighttime concentrations for one month (March) were predicted at each home point using the measured diurnal pattern of PM2.5 at the fixed site monitor in the LUR study, located in a residential neighborhood. These concentrations were scaled using data from a regulatory fixed site monitor (at New Delhi\u2019s ITO intersection) to obtain concentrations for months other than March. Monthly patterns at the fixed site (ITO) monitor are shown in Figure C.1, Appendix C.   Similarly, we extracted morning and afternoon concentrations at the work locations from the spatiotemporal LUR models. Indoor concentrations at home and work locations were estimated by multiplying the ambient concentrations at those locations by the indoor\/outdoor PM2.5 concentration (I\/O) ratio. The I\/O ratio provides a general understanding of the steady-state relationship between indoor and outdoor concentrations and is influenced by a number of factors including indoor sources, penetration factor, air exchange rate, outdoor concentration and seasonality.181 In the absence of indoor sources, the I\/O ratio will increase with air exchange rates181 and will be less than or equal to unity. We used PM2.5 I\/O ratio data from a study conducted in urban North India182 which reported PM2.5 I\/O ratios varying from 0.76 to 0.97. For this study I used a PM2.5 I\/O ratio of 0.76\u20130.88 for the winter (November through February) months and 0.88\u20131 for other months. These ranges are consistent with the reported seasonal data for residential locations in north-central India.183 As the focus of this research is on exposure to PM2.5 of outdoor origin, the effect of environmental tobacco smoke (ETS) and other sources of exposure to PM2.5 of indoor origin were not considered.   To estimate near-road concentration of PM2.5, we extracted average PM2.5 concentrations for major and minor roads from morning (0800\u20131200) and afternoon (1200\u20131800) spatiotemporal LUR model outputs. We then fit normal distributions to the extracted values and obtained probability density functions (PDFs) for PM2.5 concentrations on major and minor roads during afternoon and morning periods. We sampled from these distributions to obtain near-road PM2.5 concentrations in order to estimate exposure in the transport microenvironment for scenarios 2 and 3. This approach allowed the assignment of near-road PM2.5 concentrations that captured the spatial variation in exposure expected to be experienced by a commuter.  66   4.2.1.3 Time-Activity Data  In addition to unique home and work locations, each simulated individual was assigned a value for time spent in each microenvironment. Time-activity data were obtained from a travel time survey that was conducted in New Delhi during April 2011. Ethics approval (H11\u201300469) was obtained from the Behavioural Research Ethics Board of the University of British Columbia.  The intent of the survey was to assess the time spent by participants in different microenvironments, viz., home, work and transport. Six interviewers from a local NGO (Institute for Democracy and Sustainability, IDS) were retained to administer a survey instrument (Appendix D) that was translated to Hindi prior to use in the field. A half-day training session was conducted for the interview team prior to the launch of the survey campaign. The survey instrument was administered to a total of 1012 participants in 19 low and middle-income neighborhoods identified in consultation with staff at IDS. In each neighborhood, we surveyed between 20 and 50 participants, depending on the size of the neighborhood and the size of the team of enumerators available that day. We started from a neighborhood landmark and surveyed as many participants as possible, without trying to create a representative sample of the population. Individuals over the age of 18 years were taken as participants in this study. Informed consent was obtained from all participants before the interviewers asked survey questions, and responses were noted.   Estimates for travel times were provided by the survey responses; a probability density function (PDF) was fitted to the fraction of time in a day used for commuting (Figure C.2, Appendix C). We also assumed that total travel time was equally split into morning and afternoon peak hours184 and less than 10% of travel time was spent on minor roads. The fraction of time spent on minor roads was predicted using a uniform distribution with values ranging between 0 to 10%. Travel time fractions were assigned to simulated individuals in proportion to the distance  \ud835\udc37\u2032(\ud835\udc56, \ud835\udc57) travelled between home and work locations, calculated using the Gravity model. This was done by identifying the decile bin of the distance between a pair of home and work locations, followed by assigning a randomly sampled value for travel time from the corresponding decile bin of the travel time fraction PDF. 67   We also assumed that time spent at a work or educational institution followed a normal distribution with mean=8 h and standard deviation=0.5 h. This is consistent with a previous time activity survey in New Delhi.184  4.2.1.4 Ratio of In-vehicle Concentration to Near-Vehicle Concentration, \u03bavm In simple terms, the in-vehicle concentration can be expressed as a multiple of a near-vehicle ambient term with an associated error term,185 assuming that there are no within-vehicle sources like smoking, as shown in Eq. (3).  Civ,vm = \u03bavm . Cnv + \u03b5 vm  (3) where Civ,vm is the in-vehicle (iv) PM2.5 concentration for vehicle mode vm; Cnv  is the near-road PM2.5 concentration; \u03bavm is the ratio of in-vehicle to near-vehicle PM2.5 concentration for vehicle mode vm where is vm=vehicle mode (car, bus, auto-rickshaw (motorized three-wheeler) and motorized two-wheeler) and \u03b5vm is an error term.  Intra-vehicle variability of in-vehicle concentrations is influenced by factors including ventilation system, window position, air-conditioning, vehicle speed, road type and a time of day effect.185, 186 Additionally, inter-vehicle variability exists and is based on differences in these factors across vehicle types within a mode as well as variability across different modes.176 In the simple representation provided in Eq. (3), these variations are captured by a probability distribution for \u03bavm.  In order to estimate \u03bavm, we measured PM2.5 concentrations for four common modes of travel on a busy arterial route (Figure C.3, Appendix C). PM2.5 concentrations were measured using a personal sampler (37 mm PTFE within a PEM sampler, SKC Inc., connected to a personal sampling pump at a flow rate of 4Lmin-1) in four common modes of transport: bus, air-conditioned car, motorized two-wheeler and auto-rickshaw. Filters were conditioned for 24 hours in a desiccator before pre- and post-weighing (Satorious GD603). Flow calibrations were conducted with a TSI\u00ae calibrator before each sampling trip. The sampling equipment was enclosed in a backpack with the sampling inlet at breathing height. Trips were made using each mode of transportation during morning and afternoon peak hours on a fixed route on a busy 68  arterial route (Ring Road, Figure C.3, Appendix C). This route was chosen because it is one of the highest traffic density routes in the city. Each trip was about 50 km, with six trips for each of the four modes, covering approximately 1200 km in total. We extracted average PM2.5 concentrations for the study route from the LUR models. We developed Bayesian priors based on the small sample of six trips per mode to develop distributions for \u03bavm ratios for each mode. Measured concentrations for each mode were divided by the average roadside concentrations obtained from LUR models to obtain \u03bavm ratios. We sampled from these distributions in proportion to modal share of each mode to assign a value for the \u03bavm ratio to each simulated individual.   4.3 Results  Out of 1012 survey participants, 30% belonged to the 18\u201325 years age group, 45% were in the 26\u201340 years age group, 21% were in the 41\u201365 years age group and 4% were above 65 years of age. 56% participants were male and 44% were females. 16% of participants were illiterate, 66% had some level of formal education and 18% had a degree or a diploma. 58% of survey participants were unemployed and the remaining participants were employed part-time or full-time. 73% of participants belonged to lower income households (monthly income< 20 000 INR) and 27% participants belonged to households with a monthly income greater than 20 000 INR. LPG (Liquefied Petroleum Gas) was used as the household cooking fuel for 89% participants, 7% used kerosene and 4% used solid fuels (wood or coal). Median travel times derived from the time-activity survey for the four travel modes ranged from 40 to 75 minutes (Table 4.1). The travel time fraction ft was modeled as a beta distribution that was fit to the survey response data. The proportion of time that an individual spent at home (fh) was estimated as a residual after accounting for travel and work times, and accounted for the bulk of the time spent in all microenvironments in a 24 h period. We assumed that individuals spent 8 h (SD=0.5 h) at work every working day. Table 1 also shows geometric mean and geometric standard deviation of measured PM2.5 concentrations in the four modes, and mean and standard deviation for the estimated \u03bavm ratios for each mode. The histogram of observed travel time fraction along with corresponding fitted beta density function are shown in Figure C.2, Appendix C.  69   Table 4.1: Median daily travel time from survey data, measured PM2.5 concentrations on the study route and estimated values for \u03bavm for each mode of travel. primary mode of travel median daily travel time (h)   (n*) in-vehicle PM2.5 concentration GM (GSD)  estimated \u03bavm ratio mean (SD) Bus 1 (180) 362 (1.2)   2.19 (0.17) Motorized two-wheeler 0.7(146) 230 (1.1) 1.37 (0.06) Car (AC) 0.7(75) 108 (1.0)  0.63 (0.01) Auto-rickshaw 1.2 (6) 433 (1.1) 2.5 (0.08) *Sample size (n) SD: standard deviation, GM: geometric mean, GSD: geometric standard deviation Concentrations in \u00b5g m-3   Descriptive statistics for the annual levels of exposure for the three scenarios are provided in Table 4.2. Figure 4.2(a) shows the cumulative frequency distribution for average daily PM2.5 exposures for Scenarios 1, 2 and 3. Figure 4.2(b) shows micro-environmental contributions to average daily exposure for Scenario 1, Scenario 2 and Scenario 3. The average annual PM2.5 exposure for Scenario 1 was 109 \u00b5gm-3 (IQR: 97\u2013120 \u00b5g m-3). This scenario is representative of individuals staying at home, i.e., it provides an estimate of the exposure for the non-working population.   Scenario 2 assumed that all simulated individuals commute to work or educational institutions and in-vehicle exposure is same as near vehicle exposure: all commuters at a given point in space and location, irrespective of mode, had the same PM2.5 exposure. In other words, this scenario assumed that PM2.5 exposure of pedestrians and bicyclists was same as those of car, auto, bus or motorized-two-wheeler users on a given road segment at a given time.  The average annual PM2.5 exposure for Scenario 2 was 121 \u00b5g m-3 (IQR: 110\u2013131 \u00b5g m-3). PM2.5 exposure at home constituted about 60% (IQR: 55\u201365%) of total daily exposure and PM2.5 exposure while commuting constituted about 5% (IQR: 2\u20137%) of total daily exposure.    70   Table 4.2: Descriptive statistics for average daily PM2.5 exposure for Scenarios 1, 2 and 3 scenario 25th percentile average daily exposure (\u00b5g m-3) average daily exposure (\u00b5g m-3) 75th percentile average daily exposure (\u00b5g m-3) % contribution of transport microenvironment to average daily exposure (IQR) % contribution of home microenvironment to average daily exposure (IQR) Scenario1 97 109 120 0 100 Scenario2 110 121 131 5 (2\u20137) 60(55\u201365) Scenario3 114 125 136 8(4\u201311) 58(53\u201363)   For Scenario 3, where we assumed that all simulated individuals commute to work or educational institutions and in-vehicle exposure was estimated using Eq. (3), average annual PM2.5 exposure was 125 \u00b5g m-3 (IQR: 114\u2013136 \u00b5g m-3). PM2.5 exposure at home constituted about 58% (IQR: 53\u201363%) of total daily exposure, and PM2.5 exposure while commuting constituted about 8% (IQR: 4\u201311%) of total daily exposure. Under Scenario 3, commuters were on average exposed to PM2.5 concentrations that were twice as high as the average concentration they were exposed to at home. However, because commute times were small relative to the time spent at home and at work, and because PM2.5 concentrations at home and work locations were also high, the contribution of commuting to total exposure was low.    To assess temporal and spatial variation in exposures, we focused on Scenario 1 as it captures the broad patterns of exposure in the city; accounting for exposure during commutes makes little difference to the basic inferences about temporal (seasonal and diurnal scale) and spatial variation in exposures drawn from Scenario 1.    71   Figure 4.2: (a) Cumulative frequency distribution (CFD) for average daily PM2.5 exposures for Scenarios 1, 2 and 3; (b) Average daily exposure by microenvironment for Scenario 1, Scenario 2 and Scenario 3; (c) CFD for winter (December, January and February); (d) CFD for monsoon season (July, August and September).   There was a strong seasonal component to PM2.5 concentrations in New Delhi,39 with winter peaks and summer lows, and this is reflected in daily population exposure levels averaged on a monthly basis (Figure 4.3). Figures 4.2(c) and 4.2(d) show cumulative frequency distributions for average daily PM2.5 exposures in winter (November through January) and the monsoon season (July through September) respectively. For Scenario 1, average population exposure levels were lowest in July (59, IQR: 49\u201371 \u00b5g m-3), August (53, IQR: 44\u201363 \u00b5g m-3) and September (56, IQR: 47\u201367 \u00b5g m-3), which are also the months for precipitation from the annual 72  monsoon. Population exposures in the winter months (November (195, IQR: 163\u2013232 \u00b5g m-3), December (182, IQR: 152\u2013216 \u00b5g m-3) and January (157, IQR: 131\u2013187 \u00b5g m-3)) were much higher, with daily average exposures about four times higher compared to the monsoon season. As seen from Figure 4.2(d), differences in seasonal averages for daily exposure between scenarios 1 and 3 can be as high as 25 \u00b5g m-3. During the winter months, mean daily population exposure under Scenario 3 was 207 \u00b5g m-3, and over 25% of the simulated population experienced exposure levels greater than 189 \u00b5g m-3.   Figure 4.3: Monthly variation of average daily PM2.5 exposure: 95th, 75th, median, 25th and 5th percentile for Scenario 1.   In addition to seasonal effects, annual average PM2.5 population exposures also reflect significant diurnal variation in PM2.5 concentrations in New Delhi. Average hourly exposure levels were lowest in the afternoon (1200\u20131800): ~80 \u00b5g m-3, and highest during morning hours (800\u20131200): ~120 \u00b5g m-3 as shown in Figure 4.4. During the morning hours, shallower mixing heights result 73  in traffic-related pollution hotspots, while higher mixing heights in the afternoon102 result in spatially more uniform and lower levels of exposure.178  Figure 4.4: Average PM2.5 concentrations for morning, afternoon and nighttime hours (Scenario 1).   74   Figure 4.5: Map of the modeling domain along with average annual PM2.5 exposure in \u00b5g m-3.  75   Figure 4.6: (a) Annual average daily PM2.5 exposure for the 16 zones along with monthly variation of 95th, 75th, median, 25th and 5th percentiles (b): Difference between average daily PM2.5 exposures for winter and monsoon months for the 16 zones along with monthly variation of 95th, 75th, median, 25th and 5th percentiles. The modeling domain covered a small fraction (<5%) of zones KI and N; hence the estimates of exposure for these two zones are unreliable.  Modeling results were used to generate zonal PM2.5 exposure profiles (Figure 4.5). Annual average daily exposures for residents of each of the 16 zones (A through PII) are shown in Figure 4.6(a) and range from 89 \u00b5g m-3 (Zone PI) to 128 \u00b5g m-3  (Zone A). Variation of annual averages across zones is smaller than seasonal differences within a zone. Median winter \u2013monsoon difference (Figure 4.6 (b)) in average daily exposures can vary dramatically across the 16 zones and range from 101\u2013144 \u00b5g m-3. Such large spatial variations in exposure result from a confluence meteorology and PM2.5 emissions from a variety of sources.100 These results elucidate 76  the variation of average annual exposure as a function of distance from the nearest major road (Figure 4.7). Near road residents (living within 100 m of a major road) face annual average exposures that are 20 \u00b5g m-3 (16%) greater compared to those who live \u2265 725m away; 95% of the simulated individuals lived within 725 m from a major road. The effect of proximity to roads and other emission sources interacts with the seasonal meteorological effects, viz., the effects of mixing height and monsoon-related wet deposition of atmospheric aerosols187 in producing large spatial differences across seasons.    Figure 4.7: Annual average PM2.5 exposure as a function of distance from nearest major road.   4.4 Discussion  The estimated PM2.5 exposures for all three scenarios were high (>109 \u00b5g m-3) compared to studies in the United States and Europe (~20 \u00b5g m-3).58, 60 PM2.5 exposures were lowest if we assumed that individuals stay at home (Scenario 1). Exposure at home constituted more than 58% of average daily exposure (Scenario 2 and 3). Commuters were exposed to higher levels of 77  PM2.5 during travel, about two times more compared with exposures at home. These findings are similar to those by Apte et al. (2011),145 who found that PM2.5 exposure of auto-rickshaw commuters in New Delhi were comparable to daily exposures in high income countries. However, the overall contribution of the transport microenvironment to average daily exposure is small (5\u20138%) because exposures at work and home are high as well. Consequently, the influence of travel mode on annual average PM2.5 exposure is limited. Ignoring the mobility of the working population resulted in a small underestimation of annual average PM2.5 exposure: 11% compared to Scenario 2 and 15% compared to Scenario 3.    These results also show that the lowest PM2.5 exposures can be expected in the month of August, and the highest exposures can be expected during November. Very high levels of exposure during winter are of obvious concern due to adverse health effects of acute exposure.10 Source apportionment studies show that wood burning contributes to about 7\u201323% of ambient PM2.5 in Delhi.46, 47 Analyses of satellite data and back-trajectory analyses have shown that high levels of PM2.5 during wintertime episodes are associated with regional smog.49, 188 It follows therefore that the focus of air pollution improvements in New Delhi should have a regional basis, in contrast to the current focus on local sources alone. Overall, management actions taken to reduce PM2.5 levels in the winter are likely to yield maximum reductions in the average annual PM2.5 population exposure. Though our results show that the spatial variation in annual average exposures is low, there is some heterogeneity in annual exposure levels. Zones A, B, E, G and H had the highest levels of average annual PM2.5 exposure (128\u2013116 \u00b5g m-3), and these zones also experienced high levels of exposure (211\u2013192 \u00b5g m-3) during winter. These levels are plausibly driven by the significant explanatory variables in LUR models: population density and distance from major roads. Zonal variations provide limited additional information about where reductions in PM2.5 concentrations might result in the largest public health benefits.  In summary, a key finding of this study is that annual average levels of PM2.5 exposure across the entire city are high, and are particularly severe in the winter months. Targeted exposure reduction measures during winter can be expected to yield a significant reduction in annual PM2.5 population exposure. Another important finding is that ignoring mobility can be expected to 78  underestimate exposure by about 10% in the case of New Delhi. This in turn can lead to bias health effect estimates in epidemiologic analyses. It is recommended that mobility be accounted for in future epidemiologic studies in New Delhi and elsewhere.   The modeling framework developed here has several limitations, including those related to: the limitations of LUR models used; the time-activity surveys; and in-vehicle measurements. Outputs from the LUR models that underlie the results are driven largely by spatial variation in two variables\u2014population density and proximity to the road network. There are a number of other particulate sources that contribute to poor air quality in New Delhi, including small industry sources, informal burning of foliage in the winters as well as the effect of trash burning,42 that are not used as independent variables in the LUR models. Thus, LUR models used to assign concentrations are unable to capture the impacts of these sources on exposure. Additionally, the outputs from LUR models used in the simulation are based on a PM2.5 measurement campaign carried out in Delhi from February through May. In the absence of LUR models developed for the entire year, LUR output was scaled to obtain data for other months using available monitoring data from a regulatory monitoring station. The high degree of temporal correlation (0.84) between PM2.5 concentrations measured at multiple locations and a fixed site in New Delhi178 support the use of scaling.    A number of limitations also result from the use of modeling (as opposed to direct measurement using personal monitoring) as a way to assess population exposure. The allocation of time-activity patterns for simulated individuals to three activity categories was a necessary modeling simplification. Individuals do more than stay at home and go to work or schools. However, the goal was to simulate exposures at the population levels for which such coarse allocations might be adequate. Additionally, in estimating exposure in home and work microenvironments, we assumed that for a simulated individual, indoor concentrations at work place and residence can be estimated by multiplying ambient outdoor concentrations with an I\/O ratio derived from work done at a North Indian city close to New Delhi.182, 183 This assumption has limitations due to inherent uncertainties in the I\/O ratio,181 and exposure estimates can be refined with an improved understanding of the variability in I\/O ratios. We also note that the distributions for \u03bavm ratio 79  used in this work are based on a small sample of measurements, and measured in-vehicle concentrations are higher compared to another recent study.145    Better LUR models, i.e., with a larger number of predictor variables that capture PM2.5 sources or their drivers, coupled with a greater number of PM2.5 sampling sites, might be needed to improve the performance of the LUR models at finer spatial scales. Though a number of regulatory stations have been added recently, the public availability of data from such stations remains poor, and data are hard to access. Widespread availability of such data could help improve the accuracy of LUR based simulation results and help reveal more targeted management options. 80  Chapter 5: Conclusion   This chapter includes a discussion of how the research objectives presented in Chapter 1 were achieved and the specific research questions answered. I present a summary of the findings put forward in Chapters 2, 3 and 4, and how these findings complement each other. It is recognized that Chapter 4 already integrates the spatiotemporal LUR models with time-activity patterns and other factors that determine exposure. I also present in this chapter a discussion of the implications and limitations of this research, and recommendations for future research.   5.1 Key Findings and Implications The research presented in Chapter 2 addresses my first research question regarding the influence of local traffic flow on ambient air quality on that scale. Chapter 2 documents the relationship between local traffic flow and ambient air quality (hourly concentrations of PM2.5, NO, NO2 and CO) at a busy intersection (ITO) in New Delhi using robust and multiple linear regression models. The results show that the relationship between traffic flow and ambient air quality on a localized scale is a rather complicated one. In linear regression models, local traffic flow and lagged concentration terms explained up to 19% and 54% of the variability in hourly pollutant concentrations, respectively. The lagged concentration term can be interpreted as capturing the influence of local and background emissions in the previous hour as well as the influence of meteorology. Despite the generally weak associations with local traffic flow, certain modes of traffic such as auto-rickshaws and trucks were found to have a higher influence on the hourly concentrations of NO2 and PM2.5. It is recommended that robust regression models be used for future analyses as they provide robust coefficient estimates. It must be noted that the variability explained by each model term cannot be calculated for robust regression models.  Additionally, the results show that policy measures focused on traffic flow on a localized scale will have limited impact on ambient air quality on that spatial scale. For example, a 50% reduction in truck flow (from 120 to 60 trucks per hour, during nighttime) at the ITO intersection 81  is expected to result in about 10% reduction in ambient PM2.5 levels measured at that intersection. Since truck flow is restricted during the daytime, reductions of this magnitude of reduction can only be achieved during nighttime. A 50% reduction in auto-rickshaw flow (from 1500 to 750 per hour) during daytime is expected to result in a 16% reduction in hourly NO2 concentrations at the ITO intersection. These model estimates are for summer season; estimated decrease for winter months can be obtained by developing similar models for those months. Reductions in traffic flow on a localized scale are expected to yield limited reductions in ambient concentrations of PM2.5, NO2 and CO on that spatial scale. Consequently, significant reductions can only be accomplished if management actions related to transport can achieve reductions in traffic flow (or emissions per vehicle) on a larger spatial scale, i.e., on a citywide or regional scale.   Chapter 3 provides insights into the spatial and temporal variations in PM2.5 concentrations and those of its components\u2014black carbon (BC) and ultrafine particles (UFP)\u2014on a citywide scale using land use regression (LUR) models. The LUR models for PM2.5 and BC are spatiotemporal, i.e., they can be used to predict citywide concentrations at hourly timescales. Population density and proximity to major roads were important predictors of PM2.5, BC and ultrafine particle number concentrations (UFPN) in New Delhi. There was a high level of correlation in the measured levels of PM2.5 and BC at sites distributed across the city with a fixed \u201cbackground\u201d monitoring site. Models also showed greater temporal variability (~70%) than spatial variability (<20%) in PM2.5 and BC levels. These results are consistent with the finding presented in Chapter 2 that pollutant concentrations at a given location are influenced to a great degree by large-scale temporal variations. The LUR models showed higher concentrations and higher variability in PM2.5 and BC concentrations during morning hours, with differing spatial patterns for these pollutants during morning and afternoon hours. Different spatial patterns and weaker correlation between UFPN and other pollutants also indicate that the factors driving UFPN variability are different from the factors driving PM2.5 and BC variability. Overall, the approach presented is a cost-effective approach compared to the traditional approach of simultaneous sampling at multiple sites. The LUR models presented in this research can be used for informing 82  management strategies and monitoring network configuration, and can also be used as an input into exposure assessment studies, as demonstrated in Chapter 4.   PM2.5 exposure assessment is a challenge for long-term epidemiologic studies due to small-area variations in concentrations. LUR has been widely used as a tool to model the small-area variations in concentrations of common air pollutants, including PM2.5. LUR models have also been increasingly used to assign estimates of exposure in epidemiologic studies.126, 189 For example, Dadvand et al. applied a spatiotemporal exposure assessment framework using LUR models to study the association between preeclampsia and air pollution.190 The research presented in Chapter 4 focused on the development of a probabilistic simulation framework to integrate the spatiotemporal LUR models with time-activity data and other locational factors that determine exposure to PM2.5 of ambient origin. Dons et al. (2014)191 used a similar modeling approach to estimate personal exposures to BC using hourly BC LUR models and time-activity data. The use of the probabilistic simulation framework allowed for better characterization of within-city variability (spatial and temporal) of PM2.5 exposure. It is an improvement over the use of ambient concentrations at home locations to estimate exposure in epidemiologic studies of long-term exposure because it provides the ability to account for mobility patterns of the population. This framework can be used to build on the work that has been already been done to understand the relationship between short-term exposure to air pollution and daily mortality in India cities,54 as well as in future epidemiologic studies of acute or chronic exposure. The results demonstrate that ignoring mobility of the simulated population resulted in underestimation of exposure, but by a relatively small amount (~11%) despite the fact that the simulation assumed more than 8 hours away from home. The results also show that there is tremendous seasonal variability in PM2.5 exposure. Average population exposure in November (~200 \u00b5g m-3) was about five times higher compared to August (~ 40 \u00b5g m-3). The results presented in Chapter 4 are specific to New Delhi, but the simulation framework can be applied to any other city with simple modifications. This methodology is expected to improve estimates of PM2.5 exposure compared to the use fixed site monitoring data or ignoring mobility patterns of individuals.   83   5.2 Final Reflections  The research presented in Chapter 2 is applicable to impacts of local traffic flow on ambient air quality on that scale. In other words, citywide improvements in air quality can only be accomplished if traffic volumes are reduced on a citywide and regional scale. In their review, Hoek et al.53 discussed a number of limitations related to the use of LUR models for exposure assessment. In this research, I have addressed two of those limitations for New Delhi: (a) limited temporal coverage of LUR models; and (b) the effect of ignoring time-activity patterns and assuming that indoor concentrations are same as outdoor concentrations in exposure assessment. Spatiotemporal LUR (PM2.5 and BC) models presented in Chapter 3 partially resolve the issue of limited temporal coverage for New Delhi by providing pollution maps on an hourly scale for the duration of the LUR study. The probabilistic simulation framework presented in Chapter 4 allows for the integration of time-activity patterns, the relationship between indoor and outdoor concentrations and elevated concentrations in transport microenvironment to obtain more reliable estimates of population exposure. However some of the limitations that generally apply to LUR models53 remain. First, one or more predictor variables (such as population density) used for developing LUR models may be related to health risk factors for a specific disease under consideration and may cause confounding in epidemiologic studies.192 However, this must be resolved on a case-by-case basis by proper treatment of predictor variables. Second, like other methods of exposure assessment, LUR models have a limited ability to separate the influence of highly correlated pollutants.53, 193 For example, high levels of PM2.5 in the transport microenvironment may be highly correlated with the levels of NO2. In such a scenario, independent health effects of the correlated pollutants cannot be evaluated.193 Third, the LUR models do not provide source-specific contributions to exposure.53 That limitation, however, does not impact the usability of results for epidemiologic studies of total exposure. The LUR models performed worse than those reported in most other locations. This is plausibly due to the predominance of temporal variability, as indicated by the percentage of variance explained by the temporal term. The presence of spatial autocorrelation and the inability to account for the 84  influence of uncharacterized sources like biomass burning are other limitations of the LUR models.  The spatiotemporal LUR models were developed and validated for the months of February\u2013May only, and the model outputs have been scaled for other months. Although scaling modeled concentrations to extrapolate findings over time is a method often used,53 it is a limitation of the exposure simulation work presented in Chapter 4. The implicit assumption for such extrapolation is that spatial patterns are seasonally invariant. However, spatial patterns of pollutant concentrations may change according to seasons, so simple scaling might lead to inaccuracies during extrapolation. The lack of spatially fine-grained data on variables such as population density, employment and enrollment in educational institutions may also have reduced the ability to make accurate spatial predictions of pollutant concentrations. Perhaps the most important limitation of the exposure simulation work is that the results have not been validated by field-based measurements of exposure through personal monitoring. Personal monitoring for validation will provide important insights into the performance of the simulation framework.   Overall, the performance of the LUR models and the probabilistic simulation framework can be improved by additional research in the following areas: (a) development of LUR models for the months of June\u2013January, i.e., for the months that were not covered by the LUR study; (b) development of representative PM2.5 concentrations, due to ambient levels, for indoor microenvironments; (c) development of a detailed database of time-activity patterns; (d) better characterization of the transport microenvironment by development of an empirical relationship between in-vehicle and near-vehicle concentrations of PM2.5; (e) development of an empirical relationship between indoor and outdoor PM2.5 concentrations for representative households and workplaces; and most importantly (f) the validation of simulation results by personal monitoring.   The results presented in this dissertation also highlight the important role of large-scale temporal variations in determining PM2.5 population exposure. Regional influences on PM2.5 levels in New Delhi, especially during winter, must be considered to develop effective air quality management strategies. Analyses of satellite data and back trajectory analyses have provided critical insights 85  into the regional scale of the air pollution problem that affects New Delhi. Jethva et al. (2005)194 studied the seasonal variation of aerosol over the Indo-Gangetic plains (New Delhi is part of these plains) using Moderate Resolution Imaging Spectroradiometer (MODIS), Aerosol Robotic Network (AERONET) and Total Ozone Mapping Spectroradiometer (TOMS) data. They found that spatial and temporal variations in aerosol optical depth (AOD) were quite low in winter compared to summer. Gautam et al. (2007)195 conducted analysis of MODIS data over the Indo-Gangetic plains and concluded that fine-mode aerosol constitute most of the haze during winter months. Badarinath et al. (2009)32 studied long-range transport of aerosol due to agricultural crop burning in the Indo-Gangetic plains using LIDAR and satellite data. They concluded that agricultural crop residue burning in the Indo-Gangetic plains affects air quality on a regional scale. These findings, along with satellite images196 and ambient air monitoring data, support the hypothesis that the high levels of PM2.5 in New Delhi during winters are strongly  influenced by regional air pollution. The regression models presented in Chapter 2 and the LUR models presented in Chapter 3 also point to predominant role of large-scale temporal variations. Given the role of regional sources, air quality management strategies on a regional scale are imperative for reducing PM2.5 population exposure in New Delhi.   In summary, this research provides critical insights into spatial and temporal variability of PM2.5 concentrations and exposure in New Delhi. Through this research, I have developed a simple probabilistic simulation framework that uses the spatiotemporal LUR models (presented in Chapter 3) to predict the variability in PM2.5 population exposure in New Delhi. The estimated annual average PM2.5 exposures in New Delhi were high (more than \uf0b410) compared to North American cities. Although localized sources such as traffic were important, regional scale influences seemed to have a much larger impact on PM2.5 population exposure in New Delhi. Seasonal variations in PM2.5 exposure during the time period of the study were large\u2014exposure levels in the winter months were about five times higher than exposure levels in the monsoon months\u2014and were plausibly driven by local as well as regional sources, and by meteorology. Targeted management strategies for the winter months are likely to result in maximum reductions in population exposure in New Delhi. Further, given the role of citywide and regional sources that contribute to high levels of PM2.5 in New Delhi during the winter months, 86  management actions need to focus on a full range of sources on a regional scale, and not just traffic sources, in order to achieve any significant reductions in ambient levels of PM2.5 in New Delhi.   Lastly, this research provides estimates for the underestimation of PM2.5 exposure in New Delhi if mobility is ignored, i.e. if we assume that individuals stay at home only\u2013 a common assumption in epidemiologic studies.  This result has important implications for the current practice of ignoring mobility in epidemiologic studies of PM2.5 exposure. Ignoring mobility is expected to bias the health effects estimates. The methodology presented in this dissertation may be applied to future epidemiologic studies to adjust for mobility.  87  Bibliography  1. Burden of disease from Ambient Air Pollution for 2012. http:\/\/www.who.int\/phe\/health_topics\/outdoorair\/databases\/AAP_BoD_results_March2014.pdf?ua=1 (March 1, 2015). 2. WHO. Air Quality Guidelines, Global Update 2005; World Health Organization: Copenhagen, 2006; http:\/\/www.euro.who.int\/__data\/assets\/pdf_file\/0005\/78638\/E90038.pdf?ua=1 (May 25, 2008). 3. Bruce, N.; Perez-Padilla, R.; Albalak, R. Indoor air pollution in developing countries: a major environmental and public health challenge. Bulletin of the World Health Organization 2000, 78, 1078-1092. 4. Brunekreef, B.; Holgate, S. T. Air pollution and health. The Lancet 2002, 360, 1233-1242. 5. Davidson, C. I.; Phalen, R. F.; Solomon, P. A. Airborne Particulate Matter and Human Health: A Review. Aerosol Sci. Technol. 2005, 39, 737-749. 6. Cohen, A. J.; Anderson, H. R.; Ostro, B.; Pandey, K. D.; Krzyzanowski, M.; K\u00fcnzli, N.; Gutschmidt, K.; Pope III, C. A.; Romieu, I.; Samet, J. M., Urban air pollution. World Health Organization: Geneva, Switzerland: 2004; Vol. 2, p 1353-1433. 7. Englert, N. Fine particles and human health\u2014a review of epidemiological studies. Toxicology Letters 2004, 149, 235-242. 8. HEI Outdoor Air Pollution Among Top Global Health Risks in 2010, Risks Especially High in India and Other Developing Countries of Asia. http:\/\/www.healtheffects.org\/International\/HEI-India-GBD-PressRelease021313.pdf (October 29, 2014). 9. Krewski, D.; Jerrett, M.; Burnett, R. T.; Ma, R.; Hughes, E.; Shi, Y.; Turner, M. C.; Pope, C. A.; Thurston, G.; Calle, E. E.; Thun, M. J.; Beckerman, B.; DeLuca, P.; Finkelstein, N.; Ito, K.; Moore, D. K.; Newbold, K. B.; Ramsay, T.; Ross, Z.; Shin, H.; Tempalski, B. Extended follow-up and spatial analysis of the American Cancer Society study linking particulate air pollution and mortality; Research report 140; Health Effects Institute: Boston, MA, Jun, 2009; http:\/\/pubs.healtheffects.org\/view.php?id=315 (March 21, 2011). 10. Pope, C. A.; Dockery, D. W. Health effects of fine particulate air pollution: Lines that connect. J. Air Waste Manage. Assoc. 2006, 56, 709-742. 11. Brook, R. D.; Rajagopalan, S.; Pope, C. A.; Brook, J. R.; Bhatnagar, A.; Diez-Roux, A. V.; Holguin, F.; Hong, Y.; Luepker, R. V.; Mittleman, M. A. Particulate matter air pollution and cardiovascular disease an update to the scientific statement from the American Heart Association. Circulation 2010, 121, 2331-2378. 12. EPA, U. S. Final Report: Integrated Science Assessment for Particulate Matter; U.S. Environmental Protection Agency: Washington, DC, 2009; http:\/\/cfpub.epa.gov\/ncea\/isa\/recordisplay.cfm?deid=216546 (January 4, 2011). 13. Anderson, N.; Strader, R.; Davidson, C. Airborne reduced nitrogen: ammonia emissions from agriculture and other sources. Environment International 2003, 29, 277-286. 14. Ziemann, P. J.; Atkinson, R. Kinetics, products, and mechanisms of secondary organic aerosol formation. Chemical Society Reviews 2012, 41, 6582-6605. 88  15. HEI. Traffic-Related Air Pollution: A Critical Review of the Literature on Emissions, Exposure, and Health Effects; Health Effects Institute: Boston, MA, 2010; http:\/\/pubs.healtheffects.org\/view.php?id=334 (January 5, 2011). 16. Janssen, N. A.; Gerlofs-Nijland, M. E.; Lanki, T.; Salonen, R. O.; Cassee, F.; Hoek, G.; Fischer, P.; Brunekreef, B.; Krzyzanowski, M., Health effects of black carbon. WHO Regional Office for Europe Copenhagen: 2012. 17. HEI. Understanding the Health Effects of Ambient Ultrafine Particles; Health Effects Institute: Boston, Massachusetts, 2013; http:\/\/pubs.healtheffects.org\/getfile.php?u=893 (September 25, 2014). 18. Seinfeld, J. H. Air pollution: A half century of progress. AIChE Journal 2004, 50, 1096-1108. 19. Selden, T. M.; Song, D. Environmental Quality and Development: Is There a Kuznets Curve for Air Pollution Emissions? Journal of Environmental Economics and Management 1994, 27, 147-162. 20. Grossman, G. M.; Krueger, A. B. Economic growth and the environment; National Bureau of Economic Research: 1994;  21. Stern, D. I.; Common, M. S.; Barbier, E. B. Economic growth and environmental degradation: the environmental Kuznets curve and sustainable development. World development 1996, 24, 1151-1160. 22. World Development Indicators. http:\/\/databank.worldbank.org\/data\/reports.aspx?source=world-development-indicators (March 1, 2015). 23. CPCB. Annual Report 2011-12; Central Pollution Control Board: New Delhi, India, 2013; http:\/\/www.cpcb.nic.in\/upload\/AnnualReports\/AnnualReport_43_AR_2011-12_English.pdf (August 2, 2014). 24. CPCB Real Time Air Quality Data. http:\/\/www.cpcb.nic.in (August 1, 2015). 25. MEFCC. State of Environment Report; Ministry of Environment, Forest and Climate Change, Government of India: New Delhi, India, 2009; http:\/\/envfor.nic.in\/sites\/default\/files\/SoE Report_2009_2.pdf (July 16, 2015). 26. MOF. Economic Survey 2014-15; Ministry of Finance, Government of India: New Delhi, 2015; http:\/\/indiabudget.nic.in\/es2014-15\/echapter-vol2.pdf (July 16, 2015). 27. Census of India, 2011. http:\/\/www.censusindia.gov.in\/2011census\/population_enumeration.html (July 16, 2015). 28. Kandlikar, M.; Ramachandran, G. The causes and consequences of particulate air pollution in urban India: a synthesis of the science. Annu. Rev. Energy Environ. 2000, 25, 629-684. 29. GOI. Road Transport Year Book; Ministry of Road Transport & Highways , Government of India: New Delhi, India, 2013; http:\/\/morth.nic.in\/index2.asp?slid=291&sublinkid=137&lang=1 (June 6, 2014). 30. KPMG. The Indian Automotive Industry; India, 2010; https:\/\/http:\/\/www.kpmg.de\/docs\/Auto_survey.pdf (January 30, 2014). 31. HEI. Health Effects of Outdoor Air Pollution in Developing Countries of Asia: A Literature Review; Health Effects Institute: Boston, MA, 2004; http:\/\/pubs.healtheffects.org\/getfile.php?u=13 (September 13, 2007). 89  32. Badarinath, K. V. S.; Kumar Kharol, S.; Rani Sharma, A. Long-range transport of aerosols from agriculture crop residue burning in Indo-Gangetic Plains\u2014A study using LIDAR, ground measurements and satellite data. Journal of Atmospheric and Solar-Terrestrial Physics 2009, 71, 112-120. 33. Chafe, Z. A.; Brauer, M.; Klimont, Z.; Van Dingenen, R.; Mehta, S.; Rao, S.; Riahi, K.; Dentener, F.; Smith, K. R. Household Cooking with Solid Fuels Contributes to Ambient PM(2.5) Air Pollution and the Burden of Disease. Environ. Health Perspect. 2014, 122, 1314-1320. 34. Smith, K. R. National burden of disease in India from indoor air pollution. Proceedings of the National Academy of Sciences 2000, 97, 13286-13293. 35. GNCTD. Economic Survey of Delhi 2014-2015; Government of National Capital Territory of Delhi: 2015; http:\/\/delhi.gov.in\/wps\/wcm\/connect\/DoIT_Planning\/planning\/whats+new\/economic+survey+of+delhi+2014+-+2015 (July 17, 2015). 36. U.S. Mission India NowCast Air Quality Index. http:\/\/newdelhi.usembassy.gov\/airqualitydataemb.html (September 1, 2015). 37. Carslaw, D. C.; Ropkins, K. Openair\u2014an R package for air quality data analysis. Environmental Modelling & Software 2012, 27, 52-61. 38. Kathuria, V. Impact of CNG on vehicular pollution in Delhi: a note. Transp. Res. D. 2004, 9, 409-417. 39. Kandlikar, M. Air pollution at a hotspot location in Delhi: Detecting trends, Seasonal cycles and oscillations. Atmos. Environ. 2007, 41, 5934-5947. 40. Mohan, M.; Dagar, L.; Gurjar, B. Preparation and validation of gridded emission inventory of criteria air pollutants and identification of emission hotspots for megacity Delhi. Environ. Monit. Assess. 2007, 130, 323-339. 41. Gurjar, B. R.; van Aardenne, J. A.; Lelieveld, J.; Mohan, M. Emission estimates and trends (1990\u20132000) for megacity Delhi and implications. Atmos. Environ. 2004, 38, 5663-5681. 42. Guttikunda, S. K.; Calori, G. A GIS based emissions inventory at 1 km \u00d7 1 km spatial resolution for air pollution analysis in Delhi, India. Atmos. Environ. 2013, 67, 101-111. 43. Reff, A.; Eberly, S. I.; Bhave, P. V. Receptor Modeling of Ambient Particulate Matter Data Using Positive Matrix Factorization: Review of Existing Methods. J. Air Waste Manage. Assoc. 2007, 57, 146-154. 44. Srivastava, A.; Gupta, S.; Jain, V. Source apportionment of total suspended particulate matter in coarse and fine size ranges over Delhi. Aerosol Air Qual. Res 2008, 8, 188-200. 45. Srivastava, A.; Jain, V. K. Source apportionment of suspended particulate matters in a clean area of Delhi: A note. Transp. Res. D. 2008, 13, 59-63. 46. Chowdhury, Z.; Zheng, M.; Schauer, J. J.; Sheesley, R. J.; Salmon, L. G.; Cass, G. R.; Russell, A. G. Speciation of ambient fine organic carbon particles and source apportionment of PM2.5 in Indian cities. J. Geophys. Res., Atmos. 2007, 112. 47. Pant, P.; Shukla, A.; Kohl, S. D.; Chow, J. C.; Watson, J. G.; Harrison, R. M. Characterization of ambient PM 2.5 at a pollution hotspot in New Delhi, India and inference of sources. Atmos. Environ. 2015, 109, 178-189. 48. Srivastava, A.; Jain, V.; Srivastava, A. SEM-EDX analysis of various sizes aerosols in Delhi India. Environ. Monit. Assess. 2009, 150, 405-416. 90  49. Sati, A. P.; Mohan, M. Analysis of air pollution during a severe smog episode of November 2012 and the Diwali Festival over Delhi, India. International Journal of Remote Sensing 2014, 35, 6940-6954. 50. Miller, K. A.; Siscovick, D. S.; Sheppard, L.; Shepherd, K.; Sullivan, J. H.; Anderson, G. L.; Kaufman, J. D. Long-Term Exposure to Air Pollution and Incidence of Cardiovascular Events in Women. N. Engl. J. Med. 2007, 356, 447-458. 51. Wilson, J. G.; Kingham, S.; Pearce, J.; Sturman, A. P. A review of intraurban variations in particulate air pollution: Implications for epidemiological research. Atmos. Environ. 2005, 39, 6444-6462. 52. Jerrett, M.; Arain, A.; Kanaroglou, P.; Beckerman, B.; Potoglou, D.; Sahsuvaroglu, T.; Morrison, J.; Giovis, C. A review and evaluation of intraurban air pollution exposure models. J Expo Anal Environ Epidemiol 2005, 15, 185-204. 53. Hoek, G.; Beelen, R.; de Hoogh, K.; Vienneau, D.; Gulliver, J.; Fischer, P.; Briggs, D. A review of land-use regression models to assess spatial variation of outdoor air pollution. Atmos. Environ. 2008, 42, 7561-7578. 54. HEI. Public Health and Air Pollution in Asia (PAPA): Coordinated Studies of Short-Term Exposure to Air Pollution and Daily Mortality in Two Indian Cities.; Health Effects Institute: Boston, MA., 2011; http:\/\/pubs.healtheffects.org\/view.php?id=357 (March 21, 2012). 55. Ott, W. R. Exposure Estimates Based on Computer Generated Activity Patterns. Clinical Toxicology 1983, 21, 97-128. 56. Klepeis, N. E. An introduction to the indirect exposure assessment approach: modeling human exposure using microenvironmental measurements and the recent National Human Activity Pattern Survey. Environ. Health Perspect. 1999, 107, 365-374. 57. McCreddin, A.; Alam, M. S.; McNabola, A. Modelling personal exposure to particulate air pollution: An assessment of time-integrated activity modelling, Monte Carlo simulation &amp; artificial neural network approaches. International Journal of Hygiene and Environmental Health 2015, 218, 107-116. 58. Hanninen, O.; Kruize, H.; Lebret, E.; Jantunen, M. EXPOLIS simulation model: PM2.5 application and comparison with measurements in Helsinki. J Expo Anal Environ Epidemiol 2003, 13, 74-85. 59. Setton, E. M.; Keller, C. P.; Cloutier-Fisher, D.; Hystad, P. W. Spatial variations in estimated chronic exposure to traffic-related air pollution in working populations: A simulation. International Journal of Health Geographics 2008, 7, 39-39. 60. Burke, J. M.; Zufall, M. J.; Ozkaynak, H. A population exposure model for particulate matter: case study results for PM(2.5) in Philadelphia, PA. J Expo Anal Environ Epidemiol 2001, 11, 470-89. 61. Isakov, V.; Touma, J. S.; Burke, J.; Lobdell, D. T.; Palma, T.; Rosenbaum, A.; k\u00d6zkaynak, H. Combining Regional- and Local-Scale Air Quality Models with Exposure Models for Use in Environmental Health Studies. J. Air Waste Manage. Assoc. 2009, 59, 461-472. 62. Setton, E.; Marshall, J. D.; Brauer, M.; Lundquist, K. R.; Hystad, P.; Keller, P.; Cloutier-Fisher, D. The impact of daily mobility on exposure to traffic-related air pollution and health effect estimates. J Expos Sci Environ Epidemiol 2011, 21, 42-48. 63. Ott, W. R. Concepts of human exposure to air pollution. Environment International 1982, 7, 179-196. 91  64. Marshall, J. D.; Nethery, E.; Brauer, M. Within-urban variability in ambient air pollution: Comparison of estimation methods. Atmos. Environ. 2008, 42, 1359-1369. 65. Rose, N.; Cowie, C.; Gillett, R.; Marks, G. B. Validation of a Spatiotemporal Land Use Regression Model Incorporating Fixed Site Monitors. Environ. Sci. Technol. 2011, 45, 294-299. 66. Smargiassi, A.; Brand, A.; Fournier, M.; Tessier, F.; Goudreau, S.; Rousseau, J.; Benjamin, M. A spatiotemporal land-use regression model of winter fine particulate levels in residential neighbourhoods. Journal of Exposure Science and Environmental Epidemiology 2012, 22, 331-338. 67. Henderson, S. B.; 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, 2422-2428. 68. Gilbert, N. L.; Goldberg, M. S.; Beckerman, B.; Brook, J. R.; 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, 1059-1063. 69. Jerrett, M.; Arain, M. A.; Kanaroglou, P.; Beckerman, B.; Crouse, D.; Gilbert, N. L.; Brook, J. R.; Finkelstein, N.; Finkelstein, M. M. Modeling the intraurban variability of ambient traffic pollution in Toronto, Canada. J. Toxicol. Env. Health Part A 2007, 70, 200-212. 70. Kanaroglou, P. S.; Jerrett, M.; Morrison, J.; Beckerman, B.; Arain, M. A.; Gilbert, N. L.; Brook, J. R. Establishing an air pollution monitoring network for intra-urban population exposure assessment: A location-allocation approach. Atmos. Environ. 2005, 39, 2399-2409. 71. Ryan, P. H.; LeMasters, G. K. A review of land use regression models for characterizing intraurban air pollution exposure. Inhal. Toxicol. 2007, 19, 127-133. 72. Zou, B.; Wilson, J. G.; Zhan, F. B.; Zeng, Y. Air pollution exposure assessment methods utilized in epidemiological studies. J. Environ. Monit. 2009, 11, 475-490. 73. Briggs, D. The role of GIS: coping with space (and time) in air pollution exposure assessment. Journal of Toxicology and Environmental Health, Part A 2005, 68, 1243-1261. 74. Brauer, M.; Hoek, G.; van Vliet, P.; Meliefste, K.; Fischer, P.; Gehring, U.; Heinrich, J.; Cyrys, J.; Bellander, T.; Lewne, M.; Brunekreef, B. Estimating Long-Term Average Particulate Air Pollution Concentrations: Application of Traffic Indicators and Geographic Information Systems. Epidemiology 2003, 14, 228-239. 75. Ambient (outdoor) air pollution in cities database 2014. http:\/\/www.who.int\/phe\/health_topics\/outdoorair\/databases\/cities\/en\/ (September 1, 2015). 76. GBD Profile: India, The Global Burden of Disease Study 2010 (GBD 2010). http:\/\/www.healthdata.org\/gbd\/data (July 20, 2015). 77. Lim, S. S.; Vos, T.; Flaxman, A. D.; Danaei, G.; Shibuya, K.; Adair-Rohani, H.; AlMazroa, M. A.; Amann, M.; Anderson, H. R.; Andrews, K. G. A comparative risk assessment of burden of disease and injury attributable to 67 risk factors and risk factor clusters in 21 regions, 1990\u20132010: a systematic analysis for the Global Burden of Disease Study 2010. The lancet 2012, 380, 2224-2260. 78. GNCTD. Statistical Abstract of Delhi; Government of National Capital Territory of Delhi: New Delhi, 2014; http:\/\/www.delhi.gov.in\/wps\/wcm\/connect\/f508bc8046667b0e9cf6bcf5a4ed47e7\/Stattistical+Abstract+of+Delhi+2014.pdf?MOD=AJPERES&lmod=66436406&CACHEID=f508bc8046667b0e9cf6bcf5a4ed47e7 (May 6, 2015). 92  79. Goel, R.; Guttikunda, S. K. Evolution of on-road vehicle exhaust emissions in Delhi. Atmos. Environ. 2015, 105, 78-90. 80. Reynolds, C. C. O.; Kandlikar, M. Climate Impacts of Air Quality Policy: Switching to a Natural Gas-Fueled Public Transportation System in New Delhi. Environ. Sci. Technol. 2008, 42, 5860-5865. 81. Reynolds, C. C. O.; Grieshop, A. P.; Kandlikar, M. Climate and Health Relevant Emissions from in-Use Indian Three-Wheelers Fueled by Natural Gas and Gasoline. Environ. Sci. Technol. 2011, 45, 2406-2412. 82. Vallero, D., Chapter 27 - Modeling Applications. In Fundamentals of Air Pollution (Fifth Edition), Vallero, D., Ed. Academic Press: Boston, 2014; pp 683-753. 83. Pierce, T. E., Emission and air quality modeling tools for near-roadway applications. US Environmental Protection Agency, Office of Research and Development: 2008. 84. Levy, J. I.; Bennett, D. H.; Melly, S. J.; Spengler, J. D. Influence of traffic patterns on particulate matter and polycyclic aromatic hydrocarbon concentrations in Roxbury, Massachusetts. Journal of Exposure Science and Environmental Epidemiology 2003, 13, 364-371. 85. Thomas, S.; Jacko, R. B. Model for forecasting expressway fine particulate matter and carbon monoxide concentration: application of regression and neural network models. J. Air Waste Manage. Assoc. 2007, 57, 480-488. 86. Aldrin, M.; Haff, I. H. Generalised additive modelling of air pollution, traffic volume and meteorology. Atmos. Environ. 2005, 39, 2145-2155. 87. Rogak, S. N.; Pott, U.; Dann, T.; Wang, D. Gaseous Emissions from Vehicles in a Traffic Tunnel in Vancouver, British Columbia. J. Air Waste Manage. Assoc. 1998, 48, 604-615. 88. Grieshop, A. P.; Lipsky, E. M.; Pekney, N. J.; Takahama, S.; Robinson, A. L. Fine particle emission factors from vehicles in a highway tunnel: Effects of fleet composition and season. Atmos. Environ. 2006, 40, Supplement 2, 287-298. 89. Yan, B.; Zheng, M.; Hu, Y.; Ding, X.; Sullivan, A. P.; Weber, R. J.; Baek, J.; Edgerton, E. S.; Russell, A. G. Roadside, Urban, and Rural Comparison of Primary and Secondary Organic Molecular Markers in Ambient PM2.5. Environ. Sci. Technol. 2009, 43, 4287-4293. 90. Oliveira, C.; Pio, C.; Caseiro, A.; Santos, P.; Nunes, T.; Mao, H.; Luahana, L.; Sokhi, R. Road traffic impact on urban atmospheric aerosol loading at Oporto, Portugal. Atmos. Environ. 2010, 44, 3147-3158. 91. Harrison, R. M.; Deacon, A. R.; Jones, M. R.; Appleby, R. S. Sources and processes affecting concentrations of PM10 and PM2.5 particulate matter in Birmingham (U.K.). Atmos. Environ. 1997, 31, 4103-4117. 92. Pant, P.; Harrison, R. M. Estimation of the contribution of road traffic emissions to particulate matter concentrations from field measurements: A review. Atmos. Environ. 2013, 77, 78-97. 93. Crabbe, H.; Beaumont, R.; Norton, D. Assessment of Air Quality, Emissions and Management in a Local Urban Environment. Environ. Monit. Assess. 2000, 65, 435-442. 94. Vardoulakis, S.; Valiantis, M.; Milner, J.; ApSimon, H. Operational air pollution modelling in the UK\u2014Street canyon applications and challenges. Atmos. Environ. 2007, 41, 4622-4637. 95. Wu, X. M.; Fan, Z. T.; Ohman-Strickland, P. Time-location patterns of a population living in an air pollution hotspot. Journal of Environmental and Public Health 2010, 2010. 93  96. Vos, P. E. J.; Maiheu, B.; Vankerkom, J.; Janssen, S. Improving local air quality in cities: To tree or not to tree? Environmental Pollution 2013, 183, 113-122. 97. Narain, U.; Krupnick, A. The impact of Delhi's CNG program on air quality. Available at SSRN 969727 2007. 98. Chelani, A. B. Statistical Characteristics of Ambient PM2. 5 Concentration at a Traffic Site in Delhi: Source Identification Using Persistence Analysis and Nonparametric Wind Regression. Aerosol and Air Quality Research 2013, 13, 1768-1778. 99. Ghosh, S.; Biswas, J.; Guttikunda, S.; Roychowdhury, S.; Nayak, M. An investigation of potential regional and local source regions affecting fine particulate matter concentrations in Delhi, India. J. Air Waste Manage. Assoc. 2015, 65, 218-231. 100. Guttikunda, S. K.; Gurjar, B. Role of meteorology in seasonality of air pollution in megacity Delhi, India. Environ. Monit. Assess. 2011, 184, 3199-3211. 101. Integrated Surface Database,NOAA. https:\/\/http:\/\/www.ncdc.noaa.gov\/isd\/data-access (December 12, 2007). 102. Tiwari, S.; Srivastava, A. K.; Bisht, D. S.; Parmita, P.; Srivastava, M. K.; Attri, S. D. Diurnal and seasonal variations of black carbon and PM2.5 over New Delhi, India: Influence of meteorology. Atmospheric Research 2013, 125\u2013126, 50-62. 103. Little, R. J.; Rubin, D. B., Statistical analysis with missing data. John Wiley & Sons: 2014. 104. Gujarati, D. N., Basic Econometrics. 4th. In New York: McGraw-Hill: 2003. 105. Huber, P. J., Regression. In Robust Statistics, John Wiley & Sons, Inc.: 2005; pp 153-198. 106. Rousseeuw, P. J.; Leroy, A. M., Robust regression and outlier detection. John Wiley & Sons: 2005; Vol. 589. 107. Western, B. Concepts and Suggestions for Robust Regression Analysis. American Journal of Political Science 1995, 39, 786-817. 108. Rousseeuw, P.; Croux, C.; Todorov, V.; Ruckstuhl, A.; Salibian-Barrera, M.; Verbeke, T.; Koller, M.; Maechler, M. robustbase: Basic Robust Statistics. https:\/\/cran.r-project.org\/web\/packages\/robustbase\/ (August 15, 2015). 109. Yohai, V. J. High Breakdown-Point and High Efficiency Robust Estimates for Regression. The Annals of Statistics 1987, 15, 642-656. 110. Ott, W. R. A Physical Explanation of the Lognormality of Pollutant Concentrations. J. Air Waste Manage. Assoc. 1990, 40, 1378-1383. 111. Farrar, D. E.; Glauber, R. R. Multicollinearity in regression analysis: the problem revisited. The Review of Economic and Statistics 1967, 92-107. 112. Mansfield, E. R.; Helms, B. P. Detecting Multicollinearity. The American Statistician 1982, 36, 158-160. 113. Durbin, J.; Watson, G. S. Testing for serial correlation in least squares regression. II. Biometrika 1951, 38, 159-177. 114. Tillman, J. A. The Power of the Durbin-Watson Test. Econometrica 1975, 43, 959-974. 115. Gr\u00f6mping, U. Relative importance for linear regression in R: the package relaimpo. Journal of statistical software 2006, 17, 1-27. 116. Aron, R. Mixing height\u2014an inconsistent indicator of potential air pollution concentrations. Atmos. Environ. 1983, 17, 2193-2197. 94  117. Ravindra, K.; Wauters, E.; Tyagi, S. K.; Mor, S.; Van Grieken, R. Assessment of air quality after the implementation of compressed natural gas (CNG) as fuel in public transport in Delhi, India. Environ Monit Assess 2006, 115, 405-17. 118. Whitlow, T. H.; Hall, A.; Zhang, K. M.; Anguita, J. Impact of local traffic exclusion on near-road air quality: Findings from the New York City \u201cSummer Streets\u201d campaign. Environmental Pollution 2011, 159, 2016-2027. 119. Keuken, M. P.; Moerman, M.; Voogt, M.; Blom, M.; Weijers, E. P.; R\u00f6ckmann, T.; Dusek, U. Source contributions to PM2.5 and PM10 at an urban background and a street location. Atmos. Environ. 2013, 71, 26-35. 120. Zwack, L. M.; Paciorek, C. J.; Spengler, J. D.; Levy, J. I. Characterizing local traffic contributions to particulate air pollution in street canyons using mobile monitoring techniques. Atmos. Environ. 2011, 45, 2507-2514. 121. Badami, M. Transport and Urban Air Pollution in India. Environmental Management 2005, 36, 195-204. 122. Badami, M. G.; Haider, M. An analysis of public bus transit performance in Indian cities. Transportation Research Part A: Policy and Practice 2007, 41, 961-981. 123. Guttikunda, S. K.; Mohan, D. Re-fueling road transport for better air quality in India. Energy Policy 2014, 68, 556-561. 124. Dockery, D. W.; Pope, C. A.; Xu, X. P.; Spengler, J. D.; Ware, J. H.; Fay, M. E.; Ferris, B. G.; Speizer, F. E. An association between air pollution and mortality in 6 United States cities. N. Engl. J. Med. 1993, 329, 1753-1759. 125. Pope, C. A.; Ezzati, M.; Dockery, D. W. Fine-particulate air pollution and life expectancy in the United States. N. Engl. J. Med. 2009, 360, 376-386. 126. Ross, Z.; Ito, K.; Johnson, S.; Yee, M.; Pezeshki, G.; Clougherty, J. E.; Savitz, D.; Matte, T. Spatial and temporal estimation of air pollutants in New York City: exposure assignment for use in a birth outcomes study. Environ. Health 2013, 12, 51. 127. Wang, J.-F.; Hu, M.-G.; Xu, C.-D.; Christakos, G.; Zhao, Y. Estimation of Citywide Air Pollution in Beijing. PLoS ONE 2013, 8, e53400. 128. Novakov, T.; Andreae, M. O.; Gabriel, R.; Kirchstetter, T. W.; Mayol-Bracero, O. L.; Ramanathan, V. Origin of carbonaceous aerosols over the tropical Indian Ocean: Biomass burning or fossil fuels? Geophys. Res. Lett. 2000, 27, 4061-4064. 129. Grieshop, A. P.; Reynolds, C. C. O.; Kandlikar, M.; Dowlatabadi, H. A black-carbon mitigation wedge. Nature Geosci. 2009, 2, 533-534. 130. Menon, S.; Hansen, J.; Nazarenko, L.; Luo, Y. Climate effects of black carbon aerosols in China and India. Science 2002, 297, 2250-2253. 131. Janssen, N.; Hoek, G.; Simic-Lawson, M.; Fischer, P.; van Bree, L.; ten Brink, H.; Keuken, M.; Atkinson, R. W.; Anderson, H. R.; Brunekreef, B. Black carbon as an additional indicator of the adverse health effects of airborne particles compared with PM10 and PM2. 5. Environ. Health Perspect. 2011, 119, 1691-1699. 132. Oberd\u00f6rster, G.; Oberd\u00f6rster, E.; Oberd\u00f6rster, J. Nanotoxicology: An emerging discipline evolving from studies of ultrafine particles. Environ. Health Perspect. 2005, 113, 823-839. 133. Seaton, A.; Macnee, W.; Donaldson, K.; Godden, D. Particulate air pollution and acute health effects. Lancet 1995, 345, 176-178. 95  134. Peters, A.; Wichmann, H.; Tuch, T.; Heinrich, J.; Heyder, J. Respiratory effects are associated with the number of ultrafine particles. Am. J. Respir. Crit. Care Med. 1997, 155, 1376-1383. 135. Jayaraman, G.; Nidhi. Air pollution and associated respiratory morbidity in Delhi. Health Care Manage. Sci. 2008, 11, 132-138. 136. Saksena, S.; Joshi, V.; Patil, R. S. Cluster analysis of Delhi's ambient air quality data. J. Environ. Monit. 2003, 5, 491-499. 137. Shandilya, K. K.; Khare, M.; Gupta, A. B. Suspended particulate matter distribution in rural-industrial satna and in urban-industrial South Delhi. Environ. Monit. Assess. 2007, 128, 431-445. 138. Srivastava, A.; Jain, V. K. Seasonal trends in coarse and fine particle sources in Delhi by the chemical mass balance receptor model. J. Hazard. Mater. 2007, 144, 283-291. 139. Cropper, M. L.; Simon, N. B.; Alberini, A.; Arora, S.; Sharma, P. K. The health benefits of air pollution control in Delhi. Am. J. Agric. Econ. 1997, 79, 1625-1629. 140. Rajarathnam, U.; Sehgal, M.; Nairy, S.; Patnayak, R. C.; Chhabra, S. K.; Kilnani; Ragavan, K. V. Time-series study on air pollution and mortality in Delhi; Research Report 157; Health Effects Institute: Boston, MA, Mar, 2011;  141. Novotny, E. V.; Bechle, M. J.; Millet, D. B.; Marshall, J. D. National satellite-based land-use regression: NO2 in the United States. Environ. Sci. Technol. 2011, 45, 4407. 142. Basaga\u00f1a, X.; Rivera, M.; Aguilera, I.; Agis, D.; Bouso, L.; Elosua, R.; Foraster, M.; de Nazelle, A.; Nieuwenhuijsen, M.; Vila, J.; K\u00fcnzli, N. Effect of the number of measurement sites on land use regression models in estimating local air pollution. Atmos. Environ. 2012, 54, 634-642. 143. Guttikunda, S. Air pollution in Delhi. Econ. Polit. Wkly. 2012, 47, 24-27. 144. Guttikunda, S. K.; Jawahar, P. Application of SIM-air modeling tools to assess air quality in Indian cities. Atmos. Environ. 2012, 62, 551-561. 145. Apte, J. S.; Kirchstetter, T. W.; Reich, A. H.; Deshpande, S. J.; Kaushik, G.; Chel, A.; Marshall, J. D.; Nazaroff, W. W. Concentrations of fine, ultrafine, and black carbon particles in auto-rickshaws in New Delhi, India. Atmos. Environ. 2011, 45, 4470-4480. 146. Ramachandran, G.; Adgate, J. L.; Pratt, G. C.; Sexton, K. Characterizing indoor and outdoor 15 Minute average PM2.5 concentrations in urban neighborhoods. Aerosol Sci. Technol. 2003, 37, 33-45. 147. Arku, R. E.; Vallarino, J.; Dionisio, K. L.; Willis, R.; Choi, H.; Wilson, J. G.; Hemphill, C.; Agyei-Mensah, S.; Spengler, J. D.; Ezzati, M. Characterizing air pollution in two low-income neighborhoods in Accra, Ghana. Sci. Tot. Environ. 2008, 402, 217-231. 148. Both, A. F.; Balakrishnan, A.; Joseph, B.; Marshall, J. D. Spatiotemporal Aspects of Real-Time PM2. 5: Low-and Middle-Income Neighborhoods in Bangalore, India. Environ. Sci. Technol. 2011, 45, 5629-5636. 149. Kirchstetter, T. W.; Novakov, T. Controlled generation of black carbon particles from a diffusion flame and applications in evaluating black carbon measurement methods. Atmos. Environ. 2007, 41, 1874-1888. 150. Jimenez, J.; Claiborn, C.; Larson, T.; Gould, T.; Kirchstetter, T. W.; Gundel, L. Loading effect correction for real-time aethalometer measurements of fresh diesel soot. J. Air Waste Manage. Assoc. 2007, 57, 868-873. 96  151. Hansen, A. D. A.; Rosen, H.; Novakov, T. The aethalometer -- An instrument for the real-time measurement of optical absorption by aerosol particles. Sci. Tot. Environ. 1984, 36, 191-196. 152. Westerdahl, D.; Fruin, S.; Sax, T.; Fine, P. M.; Sioutas, C. Mobile platform measurements of ultrafine particles and associated pollutant concentrations on freeways and residential streets in Los Angeles. Atmos. Environ. 2005, 39, 3597-3610. 153. Dushkin, L. Scheduled caste policy in India: history, problems, prospects. Asian Surv. 1967, 7, 626-636. 154. Limpert, E.; Stahel, W. A.; Abbt, M. Log-normal distributions across the sciences: Keys and clues. Bioscience 2001, 51, 341-352. 155. Massey, F. J. The Kolmogorov-Smirnov test for goodness of fit. J. Am. Stat. Assoc. 1951, 46, 68-78. 156. R: A Language and Environment for Statistical Computing, R Foundation for Statistical Computing: Vienna, Austria, 2011. 157. Picard, R. R.; Cook, R. D. Cross-validation of regression models. J. Am. Stat. Assoc. 1984, 79, 575-583. 158. Getis, A. Reflections on spatial autocorrelation. Reg. Sci. Urban Econ. 2007, 37, 491-496. 159. Chen, L.; Bai, Z.; Kong, S.; Han, B.; You, Y.; Ding, X.; Du, S.; Liu, A. A land use regression for predicting NO2 and PM10 concentrations in different seasons in Tianjin region, China. Journal of Environmental Sciences 2010, 22, 1364-1373. 160. Allen, R. W.; Gombojav, E.; Barkhasragchaa, B.; Byambaa, T.; Lkhasuren, O.; Amram, O.; Takaro, T. K.; Janes, C. R. An assessment of air pollution and its attributable mortality in Ulaanbaatar, Mongolia. Air Quality, Atmosphere & Health 2011, 1-14. 161. Tripathi, S. N.; Dey, S.; Tare, V.; Satheesh, S. K. Aerosol black carbon radiative forcing at an industrial city in northern India. Geophys. Res. Lett. 2005, 32, L08802. 162. Clougherty, J.; Wright, R.; Baxter, L.; Levy, J. Land use regression modeling of intra-urban residential variability in multiple traffic-related air pollutants. Environ. Health 2008, 7, 17. 163. 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, 2429-2436. 164. Larson, T.; Henderson, S. B.; 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, 4672-4678. 165. Wang, R.; Henderson, S. B.; Sbihi, H.; Allen, R. W.; Brauer, M. Temporal stability of land use regression models for traffic-related air pollution. Atmos. Environ. 2013, 64, 312-319. 166. Johnson, M.; Isakov, V.; Touma, J. S.; Mukerjee, S.; \u00d6zkaynak, H. Evaluation of land-use regression models used to predict air quality concentrations in an urban area. Atmos. Environ. 2010, 44, 3660-3668. 167. Tiwari, G. Transport and land-use policies in Delhi. Bulletin of the World Health Organization 2003, 81, 444-450. 168. Mishra, D.; Goyal, P. Estimation of vehicular emissions using dynamic emission factors: A case study of Delhi, India. Atmos. Environ. 2014, 98, 1-7. 169. Levy, J. I.; Buonocore, J. J.; Von Stackelberg, K. Evaluation of the public health impacts of traffic congestion: a health risk assessment. Environ. Health 2010, 9, 1-12. 97  170. Karanasiou, A.; Viana, M.; Querol, X.; Moreno, T.; de Leeuw, F. Assessment of personal exposure to particulate air pollution during commuting in European cities\u2014Recommendations and policy implications. Sci. Tot. Environ. 2014, 490, 785-797. 171. K\u00fcnzli, N.; Kaiser, R.; Medina, S.; Studnicka, M.; Chanel, O.; Filliger, P.; Herry, M.; Horak Jr, F.; Puybonnieux-Texier, V.; Qu\u00e9nel, P.; Schneider, J.; Seethaler, R.; Vergnaud, J. C.; Sommer, H. Public-health impact of outdoor and traffic-related air pollution: a European assessment. The Lancet 2000, 356, 795-801. 172. Riediker, M.; Cascio, W. E.; Griggs, T. R.; Herbst, M. C.; Bromberg, P. A.; Neas, L.; Williams, R. W.; Devlin, R. B. Particulate Matter Exposure in Cars Is Associated with Cardiovascular Effects in Healthy Young Men. American Journal of Respiratory and Critical Care Medicine 2004, 169, 934-940. 173. Jerrett, M.; Burnett, R. T.; Kanaroglou, P.; Eyles, J.; Finkelstein, N.; Giovis, C.; Brook, J. R. A GIS - environmental justice analysis of particulate air pollution in Hamilton, Canada. Environment and Planning A 2001, 33, 955-973. 174. Evans, G. W.; Kantrowitz, E. SOCIOECONOMIC STATUS AND HEALTH: The Potential Role of Environmental Risk Exposure. Annual Review of Public Health 2002, 23, 303-331. 175. Forastiere, F.; Stafoggia, M.; Tasco, C.; Picciotto, S.; Agabiti, N.; Cesaroni, G.; Perucci, C. A. Socioeconomic status, particulate air pollution, and daily mortality: differential exposure or differential susceptibility. American journal of industrial medicine 2007, 50, 208-216. 176. Krzyzanowski, M.; Kuna-Dibbert, B.; Schneider, J., Health effects of transport-related air pollution. WHO Regional Office Europe: 2005. 177. Kruize, H.; Hanninen, O.; Breugelmans, O.; Lebret, E.; Jantunen, M. Description and demonstration of the EXPOLIS simulation model: two examples of modeling population exposure to particulate matter. J Expo Anal Environ Epidemiol 2003, 13, 87-99. 178. Saraswat, A.; Apte, J. S.; Kandlikar, M.; Brauer, M.; Henderson, S. B.; Marshall, J. D. Spatiotemporal Land Use Regression Models of Fine, Ultrafine, and Black Carbon Particulate Matter in New Delhi, India. Environ. Sci. Technol. 2013, 47, 12903-12911. 179. RITES. Transport Demand Forecast Study and Development of an Integrated Road Cum Multi-Modal Public Transport Network for NCT of Delhi; Government of National Capital Territory of Delhi: New Delhi, 2010; http:\/\/delhi.gov.in\/wps\/wcm\/connect\/d8707180436ea1acbf5affeaf072cdcf\/Final_TDFS_DIMTS_RITES+Study.pdf?MOD=AJPERES&lmod=825067854&CACHEID=d8707180436ea1acbf5affeaf072cdcf (March 21, 2015). 180. Wilson, A. G. A statistical theory of spatial distribution models. Transportation Research 1967, 1, 253-269. 181. Chen, C.; Zhao, B. Review of relationship between indoor and outdoor particles: I\/O ratio, infiltration factor and penetration factor. Atmos. Environ. 2011, 45, 275-288. 182. Massey, D.; Masih, J.; Kulshrestha, A.; Habil, M.; Taneja, A. Indoor\/outdoor relationship of fine particles less than 2.5\u03bcm (PM2.5) in residential homes locations in central Indian region. Building and Environment 2009, 44, 2037-2045. 183. Massey, D.; Kulshrestha, A.; Masih, J.; Taneja, A. Seasonal trends of PM10, PM5.0, PM2.5 & PM1.0 in indoor and outdoor environments of residential homes located in North-Central India. Building and Environment 2012, 47, 223-231. 98  184. Saksena, S.; Prasad, R. K.; Shankar, V. R. Daily Exposure to Air Pollutants in Indoor, Outdoor and In-vehicle Micro-environments: A Pilot Study in Delhi. Indoor and Built Environment 2007, 16, 39-46. 185. Jiao, W.; Frey, H. Method for Measuring the Ratio of In-Vehicle to Near-Vehicle Exposure Concentrations of Airborne Fine Particles. Transportation Research Record: Journal of the Transportation Research Board 2013, 2341, 34-42. 186. Ott, W.; Klepeis, N.; Switzer, P. Air change rates of motor vehicles and in-vehicle pollutant concentrations from secondhand smoke. J Expos Sci Environ Epidemiol 2007, 18, 312-325. 187. Tiwari, S.; Chate, D.; Pragya, P.; Ali, K.; Bisht, D. S. Variations in Mass of the PM10, PM2. 5 and PM1 during the Monsoon and the Winter at New Delhi. Aerosol and Air Quality Research 2012, 12, 20-29. 188. Bisht, D. S.; Dumka, U. C.; Kaskaoutis, D. G.; Pipal, A. S.; Srivastava, A. K.; Soni, V. K.; Attri, S. D.; Sateesh, M.; Tiwari, S. Carbonaceous aerosols and pollutants over Delhi urban environment: Temporal evolution, source apportionment and radiative forcing. Sci. Tot. Environ. 2015, 521-522, 431-45. 189. Jerrett, M.; Burnett, R. T.; Beckerman, B. S.; Turner, M. C.; Krewski, D.; Thurston, G.; Martin, R. V.; van Donkelaar, A.; Hughes, E.; Shi, Y.; Gapstur, S. M.; Thun, M. J.; Pope, C. A. Spatial Analysis of Air Pollution and Mortality in California. American Journal of Respiratory and Critical Care Medicine 2013, 188, 593-599. 190. Dadvand, P.; Figueras, F.; Basaga\u00f1a, X.; Beelen, R.; Martinez, D.; Cirach, M.; Schembari, A.; Hoek, G.; Brunekreef, B.; Nieuwenhuijsen, M. J. Ambient air pollution and preeclampsia: a spatiotemporal analysis. Environ. Health Perspect. 2013, 121, 1365-1371. 191. Dons, E.; Van Poppel, M.; Kochan, B.; Wets, G.; Int Panis, L. Implementation and validation of a modeling framework to assess personal exposure to black carbon. Environment International 2014, 62, 64-71. 192. Moore, D. K.; Jerrett, M.; Mack, W. J.; Kunzli, N. A land use regression model for predicting ambient fine particulate matter across Los Angeles, CA. J. Environ. Monit. 2007, 9, 246-252. 193. Brauer, M.; Hoek, G.; Van Vliet, P.; Meliefste, K.; Fischer, P. H.; Wijga, A.; Koopman, L. P.; Neijens, H. J.; Gerritsen, J.; Kerkhof, M.; Heinrich, J.; Bellander, T.; Brunekreef, B. 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, 1092-1098. 194. Jethva, H.; Satheesh, S.; Srinivasan, J. Seasonal variability of aerosols over the Indo\u2010Gangetic basin. Journal of Geophysical Research: Atmospheres (1984\u20132012) 2005, 110. 195. Gautam, R.; Hsu, N. C.; Kafatos, M.; Tsay, S. C. Influences of winter haze on fog\/low cloud over the Indo\u2010Gangetic plains. Journal of Geophysical Research: Atmospheres (1984\u20132012) 2007, 112. 196. Seasons of Indian Air Quality. http:\/\/earthobservatory.nasa.gov\/IOTD\/view.php?id=84731&eocn=image&eoci=related_image (September 1, 2015). 99  Appendices  Appendix A      Figure A.1: Diurnal variation of NO2 at ITO intersection (September 2006 to December 2014). The solid line represents the mean values and shaded area represents the 95% confidence interval.   100   Figure A.2: Diurnal variation of NO at ITO intersection (September 2006 to December 2014). The solid line represents the mean values and shaded area represents the 95% confidence interval.  101   Figure A.3: Diurnal variation of CO at ITO intersection (September 2006 to December 2014). The solid line represents the mean values and shaded area represents the 95% confidence interval.    102   Appendix B     Figure B.1: Diurnal variation of PM2.5 concentrations at the rooftop site by month.   Figure B.2: Diurnal variation of BC concentrations at the rooftop site by month. 103   Table B.1: LUR models using robust regression model model terms and coefficients: log-median concentration a LOOE b constant ln(ROOF) POP.5000 DIST.J1 RD2.500  PM2.5 (morning) 1.07 0.80 ** 2.0 \u00d7 10-3 ** -7.5 \u00d7 10-4 **  0.25 N = 19c       PM2.5 (afternoon)  -0.29 1.05 ** 1.4 \u00d7 10-3 **   0.29 N = 35       BC  (morning) 0.64* 0.75 ** 2.5 \u00d7 10-3 ** -6.7 \u00d7 10-4 *  0.34 N = 17       BC (afternoon) -0.03 0.97 ** 2.2 \u00d7 10-3 ** -1.1 \u00d710-3 **  0.45 N = 25       PN (morning) 10.38**  3.1 \u00d7 10-3 **   0.59 N = 18       PN (afternoon) 9.76**  2.3 \u00d7 10-3 **  4.9 \u00d7 10-5 ** 0.72 N = 37       aSpatiotemporal LUR models predict the natural log of the hourly median concentration at each fixed site. bLeave-one-out cross-validation error. cN \u2013 number of sites. **Statistically significant at 95% confidence levels. *Statistically significant at 90% confidence level.               104  Appendix C     Figure C.1: Monthly PM2.5 patterns (mean and 95% confidence interval) at the fixed regulatory monitor (ITO: Sept 2006 to Sept 2014).   105   Figure C.2: Fitted beta distribution compared to distribution of observed travel time fraction.   106   Figure C.3: Study route (~50 km along Ring Road) also shown are major roads in New Delhi.   C.1 Simulation Algorithm We selected a set of 100,000 points based on population density in our domain, referred to as PTS_POP and another set of 100,000 points randomly from our domain, referred to as PTS_RAND. Fit a beta distribution to the fraction of time spent travelling in a day (obtained from survey data), probability density function referred to as TT_PDF. Fit normal distributions to extracted morning and afternoon PM2.5 concentrations (from spatiotemporal LUR models for the month of March) for major and minor roads, PDFs referred to as MOR_MJ_PDF, MOR_MIN_PDF, AFT_MJ_PDF and AFT_MIN_PDF, respectively.  Following steps were involved in simulating an individual\u2019s daily exposure: Repeat 100,000 times Steps 1 through 13  1. Select home zone: probability selection was proportional to the population of a zone. 107  2. Select work zone: probability of selection was proportional to the sum of employment and enrollment in educational institutions in a zone. If work zone is different from home zone: select work zone using the Gravity model. The probability of selection of a zone was proportional to the sum of employment and enrollment in educational institutions in a zone and inversely proportional to distance of its centroid from the centroid of the home zone. \u03a1(\ud835\udc56, \ud835\udc57) = \u039aEmp(j)(\u03a3Emp(j))\u00d7\ud835\udc37(\ud835\udc56,\ud835\udc57) 3. Randomly sample, with replacement, one point from PTS_POP that lies within the home zone (Step 1), referred to as PT_Hi (home location)  4. Randomly sample, with replacement, one point from PTS_RAND that lies in work zone (Step 2), referred to as PT_Wi (work location) 5. Extract ambient morning and afternoon concentrations from spatiotemporal LUR models at PT_Hi and PT_Wi, referred to as c_h_mori, c_h_afti, c_w_mori and c_w_afti respectively 6. Calculate distance between PT_Hi and PT_Wi, referred to as DISTi  Create a distribution for distance between home and work locations (DIST) 7. For Scenario 2, \u03bavmi =1 For Scenario 3, generate \u03bavmi  ratio by sampling from subjective probability distributions in proportion to their modal share. 8. For each DISTi compute the decile bin it belongs to and randomly sample, with replacement, from the corresponding decile bin of TT_PDF to obtain TTi, which is the fraction of time spent commuting every day. 9. Generate fraction of time spent at work (TWi) using a normal distribution with mean= 8\/24 and standard deviation=0.5\/24 10. Estimate average daily ambient concentration (cdaily_i) at home using afternoon concentration at home and data from the fixed LUR site (for March). This is average daily exposure assuming no commute (Scenario1). 11. Compute exposure during morning commute for Scenario2 and Scenario 3 (using Civ = \u03bavmi \uf0b4 Cnv): a. Randomly sample, with replacement from MOR_MJ_PDF and MOR_MIN_PDF to obtain cmmj_i and cmmin_i  108  b. Use uniform random sampling to pick a value between 0 and 0.1, this is the fraction of time spent on minor roads to obtain fmor_i c. Exposure during morning commute= (TTi\/2) \u00d7 \u03bai \u00d7 [cmmj \u00d7(1-fmor_i) cmmin_i \u00d7 fmor_i] Compute exposure during afternoon commute for Scenario2 and Scenario 3: d. Randomly sample, with replacement from AFT_MJ_PDF and AFT_MIN_PDF to obtain camj_i and camin_i  e. Use uniform random sampling to pick a value between 0 and 0.1, this is the fraction of time spent on minor roads to obtain faft_i f. Exposure during afternoon commute= (TTi\/2) \u00d7 \u03bai \u00d7 [camj_i \u00d7(1-faft_i) camin_i \u00d7 faft_i] 12. Compute initial estimate of exposure at work: (TWi\/24)*[c_w_mori + c_w_afti]\/2 13. Compute average nighttime concentration at PT_Hi using know diurnal pattern from a fixed LUR site (for March), referred to as cni. Compute initial estimate of nighttime exposure at home: (1- TWi\/24 - TTi )* cni 14. For all months, scale the microenvironmental components from Steps 10 through 13 using data from the fixed regulatory monitoring site (ITO).  15. Generate I\/O ratio for each month by sampling from the I\/O distributions. 16. For each month, multiply values obtained in Steps 10 through 13 by I\/O ratios for respective month and calculate average exposure (E=\u2211fi Ci).     109   Appendix D    D.1 Survey Instrument  1. What is your age? a) 18-25 b) 26-40 c) 41- 65 d) 66- and above  2. You identify yourself as: a) Male b) Female c) Other  3. What is your current marital status? 1) Never Married 2) Currently married 3) Separated\/divorced\/widowed  4. Do you belong to a Scheduled Caste (SC) or a Scheduled Tribe? 1) SC 2) ST 3) No  5. What is your highest level of education?  1) None 2) Literate 3) High School 4) Intermediate 5) Diploma 6) Degree  6. Where do you live? 1) Colony\/Village\/Slum________ 2) Ward number  ________  7.  Do you live in a pucca house? [A pucca house is made from brick and mortar.] 1) Yes 2) No  110  8. What is the total number of individuals in your household? 1) Two or less 2) Three to five 3) Six or more  9. What is the number of dwelling rooms in your household? 1) One 2) Two 3) Three 4) Four or more  10. What is the primary source of lighting in your household? (Choose one) 1) Electricity 2) Kerosene 3) Candles 4) Other, please specify_____  11. What is the secondary source of lighting in your household? 1) Electricity 2) Kerosene 3) Candles 4) None 5) Other, please specify_____  12. Availability of kitchen in your household: 1) Cooking inside- has separate kitchen 2) Cooking inside- NO separate kitchen 3) Cooking outside- has kitchen 4) Cooking outside- no kitchen  13. What is the primary fuel used for cooking in your household? (Choose one) 1) Firewood \/ crop residue \/ cow-dung cakes 2) Coal\/lignite \/ charcoal 3) Kerosene 4) LPG \/ PNG 5) Electricity 6) Other, please specify______ 7) No cooking  14. What is the secondary fuel used for cooking in your household? 1) Firewood or crop residue or cow-dung cakes 2) Coal\/lignite\/charcoal 3) Kerosene 4) LPG\/PNG 5) Electricity 111  6) Other, please specify______ 7) Not Applicable   15. Are you the person who prepares meals in your household? 1) Yes    2) No  16. At what times are meals prepared in your household? 1) Morning : from _________ to________ 2) Afternoon : from _________ to________ 3) Evening : from _________ to________   17.  On a normal work day, what time do you leave for work from your home and what time do you return home? 1) Leave at __________ 2) Return at __________ 3) No fixed time 4) Not applicable  18. How much time do you spend outdoors on a typical day for recreational purposes, eg in garden or park? 1) Less than 1 hour 2) Between 1 and 2 hours 3) Between 2 and 3 hours 4) More than 3 hours 5) Not applicable  19 On a normal day, what time do you go to sleep and what time do you wake up? 1) Sleep at __________ 2) Wake up at ________  20. How many windows and ventilators do you have in your house? 1) Windows     ________ 2) Ventilators ________ 3) Others, please specify ________  21. Do you have an air conditioning unit in your house? 1) Yes    2) No  22. Are you currently doing paid work? 1) Full time  2) Part-time 3) Non-worker  23. (If response to question 22 was \u20181\u2019 or \u20192\u2019) Please select your category of work: 112  1) Office type 2) Outdoors, close to a busy road 3) Outdoors, not close to a busy road 4) Household work 5) Other, please specify______  24. (If response to question 22 was \u20183\u2019) Please select your non-economic activity: 1) Student 2) Household duties 3) Dependant 4) Pensioner 5) Other  25. What is the one way distance to your place of work from your residence?  1) Less than 5 km 2) Between 5 and 10 km 3) Between 10 and 15 km 4) Between 15 and 20 km 5) More than 20 km 6) Not Applicable  26. What is the average time that you spend commuting to your place of work from your residence (one way only)? 1) __________Hours and ___________ minutes 2) Not applicable  27. What is (are) your primary mode(s) of travel and the average distance traveled by that mode and the amount of time spent using that mode of travel? 1) Bus   : distance_________km; time  ____ hours ____minutes 2) Bicycle  : distance_________km; time  ____ hours ____minutes 3) Motorbike  : distance_________km; time  ____ hours ____minutes 4) Car \/ taxi cab : distance_________km; time  ____ hours ____minutes 5) Auto-rickshaw : distance_________km; time  ____ hours ____minutes  6) Walking  : distance_________km; time  ____ hours ____minutes 7) Metro  : distance_________km; time  ____ hours ____minutes 8) Cycle-rickshaw : distance_________km; time  ____ hours ____minutes 9) Other, please specify________  28. (If you selected option 4 in question 27) Do you have an air conditioning unit in your car? 1) Yes    2) No  29. How far is the nearest bus stop from your residence? 1) ____________minutes walk   30. How far is the nearest bus stop from your work place? 113  1) ____________ minutes walk 2) Not applicable   31. On an average working day, how much time do you spend commuting \/in traffic? 1) __________Hours and __________minutes  32. On an average working day, how much time do you spend in close proximity to traffic when you are not commuting? 1) __________Hours and __________minutes  33. What is your average household income level? 1) Less than Rs 5,000 per month 2) Rs 5,000- 10,000 per month 3) Rs 10,000-20,000 per month 4) 20,000-40,000 per month 5) More than 40,000 per month    34. What do you think of air quality in your neighborhood? 1) Very good 2) Good 3) Average 4) Bad 5) Very bad  35.  What do you think are major sources of air pollution in Delhi? (You can select more than one) 1) Motor Vehicles 2) Solid waste burning 3) Industries 4) Waste dumps 5) Other, please specify______  36. Do you suffer from asthma or any other breathing disorder? 1) Yes    2) No  Thanks  TO BE FILLED IN BY THE INTERVIEWER  Name of interviewer :______________________________ Place   :______________________________ Date   :______________________________ Time   :______________________________  ","@language":"en"}],"Genre":[{"@value":"Thesis\/Dissertation","@language":"en"}],"GraduationDate":[{"@value":"2016-02","@language":"en"}],"IsShownAt":[{"@value":"10.14288\/1.0223110","@language":"en"}],"Language":[{"@value":"eng","@language":"en"}],"Program":[{"@value":"Resource Management and Environmental Studies","@language":"en"}],"Provider":[{"@value":"Vancouver : University of British Columbia Library","@language":"en"}],"Publisher":[{"@value":"University of British Columbia","@language":"en"}],"Rights":[{"@value":"Attribution-NonCommercial-NoDerivs 2.5 Canada","@language":"*"}],"RightsURI":[{"@value":"http:\/\/creativecommons.org\/licenses\/by-nc-nd\/2.5\/ca\/","@language":"*"}],"ScholarlyLevel":[{"@value":"Graduate","@language":"en"}],"Title":[{"@value":"Air pollution in New Delhi, India : spatial and temporal patterns of ambient concentrations and human exposure","@language":"en"}],"Type":[{"@value":"Text","@language":"en"}],"URI":[{"@value":"http:\/\/hdl.handle.net\/2429\/56224","@language":"en"}],"SortDate":[{"@value":"2015-12-31 AD","@language":"en"}],"@id":"doi:10.14288\/1.0223110"}