UBC Theses and Dissertations

UBC Theses Logo

UBC Theses and Dissertations

Bayesian spatial and ecological modeling of suicide rates Lin, Yi 2009

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

Notice for Google Chrome users:
If you are having trouble viewing or searching the PDF with Google Chrome, please download it here instead.

Item Metadata

Download

Media
24-ubc_2009_fall_lin_yi.pdf [ 2.3MB ]
Metadata
JSON: 24-1.0067638.json
JSON-LD: 24-1.0067638-ld.json
RDF/XML (Pretty): 24-1.0067638-rdf.xml
RDF/JSON: 24-1.0067638-rdf.json
Turtle: 24-1.0067638-turtle.txt
N-Triples: 24-1.0067638-rdf-ntriples.txt
Original Record: 24-1.0067638-source.json
Full Text
24-1.0067638-fulltext.txt
Citation
24-1.0067638.ris

Full Text

Bayesian Spatial and Ecological Modeling of Suicide Rates by Yi Lin B.Sc., Nankai University, 2000 M.Sc., Nankai University, 2003 A THESIS SUBMITTED IN PARTIAL FULFILMENT OF THE REQUIREMENTS FOR THE DEGREE OF MASTER OF SCIENCE in The Faculty of Graduate Studies (Health Care and Epidemiology) THE UNIVERSITY OF BRITISH COLUMBIA (Vancouver) August 2009 c© Yi Lin 2009 Abstract Suicide and suicide attempts constitute major public and mental health prob- lems in many countries. The risk factors of suicide include not only psycho- logical and other individual features but also the characteristics of the com- munity in which the people live. Therefore, in order to better understand the potential impacts of community characteristics on suicide, the regional level effects of suicide need to be thoroughly examined. For this thesis, an ecological analysis was incorporated into a Bayesian disease mapping study in order to estimate suicide rates, explore regional risk factors, and discern spatial patterns in suicide risks. Fully Bayesian dis- ease mapping and ecological regression methods were used to estimate area- specific suicide risks, investigate spatial variations, and explore and quan- tify the associations between regional characteristics and suicide occurrences. The fact that spatially smoothed estimates of suicide rates highlight the high risk regions can act as stable health outcome indicators at the regional level. Furthermore, regional characteristics explored as potential risk factors of sui- cide rates can provide a better understanding of regional variations of suicide rates. Both can help in planning future public health prevention programs. In order to avoid multicollinearity among risk factors and reduce the dimen- ii Abstract sionality of the risk indicators, Principal Component Analysis and Empirical Bayes method (via Penalized Quasi-Likelihood) were applied in variable se- lection and highlighting risk patterns. Using 10-year aggregated data for all age groups and both genders, this study conducted a comprehensive analysis of suicide hospitalization and mor- tality rates in eighty-four Local Health Areas in British Columbia (Canada). A broad range of regional characteristics was investigated and different asso- ciations with suicide rates were observed in different demographic and gender groups. The major regional risk patterns related to suicide rates across age groups were social and economic characteristics, which include unemploy- ment rates, income, education attainment, marital status, family structure, and dwellings. Some age groups also showed a relation to aboriginal popula- tion, immigrants, and language. The results of this study may inform policy initiatives and programs for suicide prevention. iii Table of Contents Abstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ii Table of Contents . . . . . . . . . . . . . . . . . . . . . . . . . . . . iv List of Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vi List of Figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . viii Acknowledgements . . . . . . . . . . . . . . . . . . . . . . . . . . . x 1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1 1.1 Introduction to the problem . . . . . . . . . . . . . . . . . . . 1 1.2 Rationales for the study . . . . . . . . . . . . . . . . . . . . . 4 1.3 Research objectives . . . . . . . . . . . . . . . . . . . . . . . 7 1.4 Organization of thesis . . . . . . . . . . . . . . . . . . . . . . 8 2 Literature review . . . . . . . . . . . . . . . . . . . . . . . . . . 10 2.1 Issues associated with suicide . . . . . . . . . . . . . . . . . . 10 2.1.1 Definition . . . . . . . . . . . . . . . . . . . . . . . . . 10 2.1.2 Risk factors . . . . . . . . . . . . . . . . . . . . . . . . 13 2.1.3 Suicide prevention . . . . . . . . . . . . . . . . . . . . 32 2.2 Statistical methodology . . . . . . . . . . . . . . . . . . . . . 33 2.2.1 Bayesian disease mapping methodology . . . . . . . . 34 2.2.2 Small-area studies . . . . . . . . . . . . . . . . . . . . 37 2.2.3 Principal component analysis . . . . . . . . . . . . . . 40 3 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43 3.1 Study population and Data sources . . . . . . . . . . . . . . . 43 3.1.1 Study region . . . . . . . . . . . . . . . . . . . . . . . 43 3.1.2 Study population . . . . . . . . . . . . . . . . . . . . . 45 3.1.3 Data sources . . . . . . . . . . . . . . . . . . . . . . . 47 iv Table of Contents 3.1.4 Data extraction and aggregation . . . . . . . . . . . . 49 3.2 Descriptive summary of the suicide data . . . . . . . . . . . . 50 3.3 Descriptive summary of regional characteristics . . . . . . . . 57 4 Statistical methodology . . . . . . . . . . . . . . . . . . . . . . 59 4.1 Spatial-ecological models . . . . . . . . . . . . . . . . . . . . 59 4.2 Exploratory analysis . . . . . . . . . . . . . . . . . . . . . . . 63 4.3 Ecological regression . . . . . . . . . . . . . . . . . . . . . . . 66 4.4 Illustrative example . . . . . . . . . . . . . . . . . . . . . . . 68 5 Regional variation of suicide rates and associated ecological characteristics . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73 5.1 Suicide risks among youth (aged 15-24) . . . . . . . . . . . . 74 5.2 Suicide risks among adults (aged 25-64) . . . . . . . . . . . . 78 5.3 Suicide risks among the elderly (aged 65 and over) . . . . . . 90 6 Discussion and future work . . . . . . . . . . . . . . . . . . . . 94 6.1 Findings overview . . . . . . . . . . . . . . . . . . . . . . . . 95 6.1.1 Suicide rates . . . . . . . . . . . . . . . . . . . . . . . 95 6.1.2 Geographical patterns and high risk areas . . . . . . . 95 6.1.3 Risk patterns for different outcomes . . . . . . . . . . 98 6.1.4 Risk patterns in demographic groups . . . . . . . . . . 100 6.1.5 Significant risk patterns . . . . . . . . . . . . . . . . . 102 6.2 Consideration and implication . . . . . . . . . . . . . . . . . 107 6.3 Potential limitations and implications for future work . . . . 113 Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119 Appendices A Regional characteristics profiles . . . . . . . . . . . . . . . . . 132 B Annual HA and HSDA suicide rates . . . . . . . . . . . . . . 139 C Sensitivity analysis for priors and hyperpriors . . . . . . . 146 D Top 10 LHAs with significant RRs . . . . . . . . . . . . . . . 149 v List of Tables 3.1 Crude LHA-specific suicide hospitalization rates, males and females, 1991-2000. (per 100,000 population) . . . . . . . . . . 54 3.2 Crude LHA-specific suicide mortality rates, males and females, 1991-2000. (per 100,000 population) . . . . . . . . . . . . . . . 55 4.1 Estimates for ecological models, male aged 15-24 suicide hos- pitalization . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69 5.1 Summary of univariate analysis and principal component anal- ysis, youth aged 15-24 . . . . . . . . . . . . . . . . . . . . . . 76 5.2 Risk patterns derived from ecological regression, youth aged 15-24 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77 5.3 Summary of univariate analysis and principal component anal- ysis, adults aged 25-64 . . . . . . . . . . . . . . . . . . . . . . 83 5.4 Risk patterns derived from ecological regression, adults aged 25-44 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84 5.5 Risk patterns derived from ecological regression, adults aged 45-64 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85 5.6 Summary of univariate analysis and principal component anal- ysis, elderly aged 65+ . . . . . . . . . . . . . . . . . . . . . . . 91 5.7 Risk patterns derived from ecological regression, elderly aged 65+ . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92 A.1 Regional characteristics and summary statistics across 84 LHAs, B.C., 1991–2000. . . . . . . . . . . . . . . . . . . . . . . . . . 133 C.1 Sensitivity analysis for the prior and hyperprior specifications (Mean (95%CI)**) . . . . . . . . . . . . . . . . . . . . . . . . 147 D.1 Top 10 LHAs with significant RRs, youth aged 15-24 . . . . . 150 D.2 Top 10 LHAs with significant RRs, adult aged 25-34 . . . . . . 151 vi List of Tables D.3 Top 10 LHAs with significant RRs, adult aged 35-44 . . . . . . 152 D.4 Top 10 LHAs with significant RRs, adult aged 45-54 . . . . . . 153 D.5 Top 10 LHAs with significant RRs, adult aged 55-64 . . . . . . 154 D.6 Top 10 LHAs with significant RRs, elderly aged 65-74 . . . . . 154 D.7 Top 10 LHAs with significant RRs, elderly aged 75+ . . . . . 155 vii List of Figures 3.1 Map of Health Authorities and Health Service Delivery Areas, B.C. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45 3.2 Map of Local Health Areas, B.C. . . . . . . . . . . . . . . . . 46 3.3 Population size of B.C. in 1996, by Local Health Areas. . . . . 47 3.4 Crude age-specific rates of suicide hospitalization and death, males and females, 1991-2000 . . . . . . . . . . . . . . . . . . 51 3.5 Annual B.C. crude rates of suicide hospitalization and mortal- ity, males and females, 1991-2000 . . . . . . . . . . . . . . . . 53 4.1 Loadings of risk factors in the significant principal compo- nents, suicide hospitalization of males aged 15-24 . . . . . . . 70 4.2 Maps of unadjusted and risk-adjusted RRs for male aged 15- 24 suicide hospitalization. (High LHAs in black (Pr(RR > 1) > 0.975); Low LHAs in gray ( Pr(RR < 1) > 0.975); No statistically significance in blank) . . . . . . . . . . . . . . . . 72 5.1 Significance of unadjusted RRs, youth aged 15-24 suicide hos- pitalization and mortality. (High RRs in black (Pr(RR > 1) > 0.975); Low RRs in gray (Pr(RR < 1) > 0.975); No statistically significance in blank) . . . . . . . . . . . . . . . . 75 5.2 Significance of risk-adjusted RRs, youth aged 15-24 suicide hospitalization. (High RRs in black (Pr(RR > 1) > 0.975); Low RRs in gray (Pr(RR < 1) > 0.975); No statistically significance in blank) . . . . . . . . . . . . . . . . . . . . . . . 78 5.3 Significance of unadjusted RRs, adults aged 25-64 suicide hos- pitalization. (High RRs in black (Pr(RR > 1) > 0.975); Low RRs in gray (Pr(RR < 1) > 0.975); No statistically signifi- cance in blank) . . . . . . . . . . . . . . . . . . . . . . . . . . 80 viii List of Figures 5.4 Significance of unadjusted RRs, aged 25-64 suicide mortality. (High RRs in black (Pr(RR > 1) > 0.975); Low RRs in gray (Pr(RR < 1) > 0.975); No statistically significance in blank) . 81 5.5 Significance of risk-adjusted RRs, adults aged 25-64 suicide hospitalization.(High RRs in black (Pr(RR > 1) > 0.975); Low RRs in gray (Pr(RR < 1) > 0.975); No statistically significance in blank) . . . . . . . . . . . . . . . . . . . . . . . 89 B.1 Annual HA rates of suicide hospitalization, males and females, 1991-2000 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 140 B.2 Annual HA rates of suicide mortality, males and females, 1991- 2000 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141 B.3 Annual HSDA rates of suicide hospitalization, males and fe- males, 1991-2000 . . . . . . . . . . . . . . . . . . . . . . . . . 142 B.4 Annual HSDA rates of suicide hospitalization, males and fe- males, 1991-2000 (continued) . . . . . . . . . . . . . . . . . . . 143 B.5 Annual HSDA rates of suicide mortality, males and females, 1991-2000 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 144 B.6 Annual HSDArates of suicide mortality, males and females, 1991-2000 (continued) . . . . . . . . . . . . . . . . . . . . . . 145 ix Acknowledgements I would like to express my gratitude to all those who gave me the possibility to complete this thesis. I am extremely thankful to my supervisor, Dr. Ying C. MacNab with continuous support, patience and guidance. I feel lucky to meet her and appreciate her for helping me get through my hard time. It is my honor to have a chance to work with her. I sincerely thank to my other committee members: Dr. John Spinelli and Dr. Sam Sheps. Special thanks to the external examiner. I would like to thank all staff and faculty members at School of Population and Public health for their helps through the past years. Thanks to the BC Vital Statistics, BC Stats, and the Centre for Health Services & Policy Research for providing the data for my thesis. Thanks to all my friends in School of Population and Public Health and Department of Statistics. I really appreciate your helps and supports. Thanks to my parents for their consideration, love and support. x Chapter 1 Introduction 1.1 Introduction to the problem Suicide is not only a personal tragedy, but also constitutes a major public and mental health problem in many countries (Bertolote et al, 2006). According to the latest available data from WHO reports, globally, about one million people die from suicide every year, and 10 to 20 times more people attempt suicide (WHO, 2008a). This represents one death every 40 seconds and one attempt every 3 seconds, on average. Worldwide, suicide rates have increased by 60% over the last 50 years (WHO, 2008b). Suicide is now among the three leading causes of death among those aged 15-44 years (both sexes) (WHO, 2008b). Although suicide used to be predominant among the elderly, recent suicide rates among young people have been increasing to such an extent that they are now the group at the third highest risk in both developed and developing countries (WHO, 2008a). In Canada, about 3,700 deaths are caused by suicide every year and 14% of them occur in the province of British Columbia (B.C.) (Statistics Canada, 2007). However, suicide rates have remained fairly stable during the last decade. The age-standardized suicide mortality rate has been around 11 1 Chapter 1. Introduction per 100,000 people in total, with the rate for males 3 times higher than for females (Statistics Canada, 2006). Suicide is the second leading cause of death among youth in Canada (Statistics Canada, 2008). Suicide rates are quite high among the elderly, especially in elderly men. The 2003 suicide rate for Canadian males older than 65 (19.7 per 100,000) was nearly twice that of the whole population (11.3 per 100,000) (Statistics Canada, 2006). In the province of B.C., the suicide rates have remained fairly stable over time – about 10 per 100,000 persons per year (450 cases) (B.C. Vital Stats, 2004). Most suicides occur among the young, the elderly, and other vulnera- ble members of society. The suicide mortality rates of B.C. are relatively low compared with other provinces (Statistics Canada, 2006). However, suicide continues to be the second leading cause of death among youth aged 15-24 in B.C. (B.C. Vital Stats, 2002). Geographical variations in suicide rates have been reported in a number of studies (Agbayewa et al, 1998; Exeter and Boyle, 2007; Saunderson and Langford, 1996). In Canada, the suicide mortality rates were high in the Northwest Territories and the Yukon, and low in Nova Scotia, Newfoundland and Labrador (Statistics Canada, 2006). Large geographical variations were observed across the country. In BC, higher suicide rates were found on the west coast of the province, as well as geographical patterns and variations in the suicide rates. Finding explanations for those variations among regions is the goal of this study. The variations can be explained by many sources. One is from the esti- 2 Chapter 1. Introduction mate of suicide rates. The traditional way of estimating suicide rates usually produces large variations because of the small size of a population for a certain region or the rare disease, and it is hard to distinguish the true differ- ences from random errors for the areas. Disease mapping approach enables us to borrow strength from neighboring regions to obtain more conservative smoothed estimates of suicide rates, particularly in terms of signaling high risk areas. By applying disease mapping approach, the current study has bee able to identify high risk areas by smoothing the estimates of suicide rates. Another possible explanation for the variation in suicide rates is the pop- ulation level factors. In ecological studies, population level factors are usually aggregated from individual characteristics as grouped factors. Those grouped factors may present the characteristics of a certain area that can be associ- ated with suicide rates. They may consistently relate to suicide rates as the factors at the individual level, or may be related to suicide rates in a different way. It would be useful to study how risk factors explain the geographical variation in suicide rates at the ecological level. The purpose of this study is to conduct an in-depth analysis of suicide rates among regions to identify the high risk regions and explore how the population level factors explain the variation in suicide rates. If the popu- lation level factors can be identified to account for the variation in suicide rates, then the information can be used to initialize and/or improve any fu- ture prevention programs and further inform health policy makers to help reduce suicide. 3 Chapter 1. Introduction 1.2 Rationales for the study Most suicide studies focus on mental illness or the psychological perspec- tive. Since one often reads that 90% of suicides are associated with mental illness, most clinical studies of suicide have narrowly focused on identify- ing individual risk factors rather than thinking about population mental health in complex social and ecological settings (Knox et al, 2004). Studying population-level risk factors for suicide may help us to better understand how the inequable distributions of the population level factors are related to the differences among suicide rates across regions. Since large, significant differences in suicide rates among regions may indicate that one region needs particular preventative attention, such knowledge can aid efforts to identify regions at a particularly high risk of suicide more efficiently and therefore help to initialize effective programs for suicide prevention. Studies show that aggregated factors have also been found to be associ- ated with suicide rates, although the effects were reduced after accounting for individual factors (Cubbin et al, 2000; Agerbo et al, 2007b). Therefore, it is worth examining whether and how the aggregated population level fac- tors relate to suicide. The aggregated regional factors may pose either an ecological risk (or an ecological protection), even after controlling for the individual-level effects (Agerbo et al, 2007b). If the variation in the suicide rates can be explained by the population-level risk factors, especially demo- graphical, social and economic factors, a study of these factors would be 4 Chapter 1. Introduction helpful in initializing a practical suicide prevention strategy. Ecological studies have been applied to investigate the association be- tween observed incidences of disease and regional risk factors rather than in- dividual risk factors (Bailey, 2001). Numerous ecological studies of regional- level (or population-level or community-level) effects on suicide have been done (Rehkopf and Buka, 2006; Faria et al, 2006; Hasselback et al, 1991; Taylor et al, 2005), but most of them to date have focused on only a few specific socioeconomic factors, such as socioeconomic status, income, and unemployment (Qin et al, 2003; Taylor et al, 2005). However, studies that thoroughly evaluate the impact of multi-dimensional population-level risk factors on suicide have been limited. In the studies and analyses of variations in suicide rates with respect to the ecological risk factors, the estimates of suicide rates are often unstable because they produce large variations due to the fact that the population sizes of certain regions are small or suicide is rare in those areas. Moreover, the suicide rates among neighboring regions are often close and typically exhibit spatial dependency. The measurements of risk factors that are defined based on region are also spatially correlated because people with similar characteristics tend to live close to each other. For instance, since the air pollution from an area’s neighboring regions may affect the disease occurrence in the area, these risk factors in neighboring regions should be taken into account in the estimates of the disease rates for the area. Therefore, both outcome and risk factor data can be seen as spatially correlated and no 5 Chapter 1. Introduction longer independent. The violation of the assumption of independence among observations makes traditional statistical methods inappropriate for making inferences. Disease mapping approach has proven to be a useful component of research on geographical effects. Recent studies on suicide have found that systematic changes in rates depend on the characteristics of the residences in that region; therefore, the variations in suicide rates may be explained by the regional characteristics and the characteristics of neighboring regions (Exeter and Boyle, 2007; Saunderson and Langford, 1996). Using Bayesian disease mapping statistics can provide smoothed estimates of disease risks by accounting for the correlations in the data and detect areas at high risk. Therefore, the present study is designed to identify the regions with high or low risks of suicide occurrences and investigate how the diverse regional characteristics explain the variation in suicide rates, through an innovative use of Bayesian disease mapping method. Through a thoroughly exploration of population-level risk factors on suicide, which link to a broad range of re- gional characteristics, from marital status, language, immigration, aboriginal population, mobility status, through educational attainment, unemployment rates, family structure, and transportation, to income and dwellings, the thesis presents a comprehensive study of the associations between suicide and regional characteristics. The statistical methods involved not only dis- cover the geographical patterns of suicide risks across residential areas, but also examine the extent to which the population-level risk factors explain the variations in suicide rates, in order to provide suggestions and informa- 6 Chapter 1. Introduction tion to public health institutions for improving on current suicide prevention programs or initializing new and better ones. 1.3 Research objectives The primary goal of this thesis is to identify the regions with high or low risks of suicide occurrences, discover the geographical patterns, and investigate the associations of various regional (population-level) demographic, social, economic, and environmental characteristics with suicide rates, through an innovative use of Bayesian disease mapping method. This method not only tends to discover the geographical patterns of suicide risks across residential areas, but also examines the extent to which regional characteristics explain the variation in suicide rates in order to provide information for prevention programs. The secondary objective is to present and implement a Bayesian disease mapping approach for estimating suicide risks and their geographical vari- ations across regions. Poisson regression model with random effects was used to estimate suicide rates and explore the relationships between suicide rates and regional risk factors. The random effects were specified to reflect the potential tendency of “neighboring” regions to have similar risks for the disease. Bayesian approaches were used to give estimates and make infer- ences. In order to avoid multicollinearity among risk factors and reduce the dimensionality of the data, Principal Component Analysis (PCA) was used to highlight the risk patterns for each group and the Empirical Bayes 7 Chapter 1. Introduction method (via Penalized Quasi-Likelihood) (EB-PQL) was applied in the selec- tion of variables. The fact that spatially smoothed estimates of suicide risks highlight the high risk regions means that these estimates can act as stable health outcome indicators at the regional level. However, exploring regional characteristics as potential risk factors of suicide rates provides a better un- derstanding of the regional variations among suicide risks. Nonetheless, both can help provide information for the initialization for public health suicide prevention programs. 1.4 Organization of thesis In the present study, an ecological analysis of suicide rates is incorporated into Bayesian disease mapping study to estimate suicide rates, explore re- gional characteristics, and discern spatial patterns in suicide risks. After the relationship between suicide and regional risk factors, such as age, gender, family structure, and socioeconomic status of people residing in those regions is examined, a more reliable map of the relative risks of suicide, controlling for the spatial correlations is produced. Also, an analytic strategy is introduced for doing variable selection, risk pattern highlighting, and model fitting. The organization of the following parts of the document is as follows: Chapter 2 presents a literature review on both suicide issues and statistical methods. Chapter 3 describes the study population and data sources used in the study. Some primary statistical summaries are presented as well. Chap- ter 4 introduces the statistical methodology in detail, including the model 8 Chapter 1. Introduction structures, estimate and computing issues, variable selection, risk pattern de- velopment, and the practical analytic strategy; finally, illustrative examples of the analysis are provided. Chapter 5 shows the application of the practi- cal analytic approach to estimate suicide hospitalization and mortality rates among 84 Local Health Areas in B.C. during 1991–2000. Along with the conclusion of the study, Chapter 6 discusses the study’s possible limitations and implications for future work. 9 Chapter 2 Literature review In this chapter, we review the pertinent literature about suicide and the statistical methodology used to the study it. 2.1 Issues associated with suicide 2.1.1 Definition While researchers and clinicians often vary in their definitions of suicide, suicide is generally defined as the act of deliberately taking one’s own life. Suicide falls on a continuum from ideation to attempt to completion (Mann, 2002; Lewinsohn et al, 1996). Suicidal ideation can be defined as thoughts or wishes to be dead or to kill oneself; suicide attempt is defined as self-inflicted behaviors that is intended to result in death; and suicide completion is self- inflicted death. According to the International Classification of Diseases, Ninth Revision, Clinical Modification (ICD-9-CM), suicide is coded by E950- 959. The above three types of suicide behavior can either be distinct or over- lapping. Suicidal ideation is thought without action, which is relatively com- mon in suicide behaviors. Suicide attempts vary, depending on the degree of medical lethality or damage. A high lethal attempt means that survival is 10 Chapter 2. Literature review the result of good fortune. A low lethal attempt is an appeal for help. More males tend to make high lethal attempts than females. Suicide completion means that the person dies. Numerous studies have been done to explore suicide at different stages and in different scenarios, from ideation to attempt to completion. While the potential risk factors vary at the different stages, they may be correlated. For example, people who attempt suicide may have had suicide ideation before and may eventually complete suicide. It is worth studying the common and different risk factors in order to develop more effective prevention programs. The four most commonly used methods for committing suicide are firearm, hanging, suffocation, and poisoning (Suicide.org, 2001). There is a distinct gender difference in the manner by which people commit suicide. Males use firearms more often than their female counterparts. The most common method of suicide for all females is poisoning. In our study, the measurements of suicide focus on hospitalization for suicide attempt and suicide completion, since both are relatively severe sui- cide behaviors in which the person is at least injured. Suicide attempts and completions have been investigated over time in previous studies (Beautrais et al, 1996a; Lewinsohn et al, 1996, 2001; Mann and Malone, 1997; Aleck et al, 2006; Iribarren et al, 2000; Spirito and Esposito-Smythers, 2006). The risk factors for suicide attempts and completions often overlap. The study and prevention of suicide attempts are probably the most relevant issues to explore completed suicides. Suicide attempts represent a strong, known risk 11 Chapter 2. Literature review for future attempts or completions (Mann, 2002). Studies about suicide at- tempts provide an important way for understanding suicide completion since suicide completion is extremely rare. About one third of suicide completions have had previous attempts and 10% to 15% of people who attempt suicide will eventually commit suicide (Suominen et al, 2004). In the study of Safer (1997), he found that about 20-25% of adolescent suicide attempts were taken into hospital for medical care (i.e., hospitalized) and of those, about 12-15% wound up in an emergency room. Although many people may not seek medical help for an attempt or may deny that their injuries are the result of a suicide attempt, other studies have found that 30% to 50% of all attempts receive medical care (Kumar et al, 2006; Cantor and Neulinger, 2000). People who attempt suicide that result in hospital admissions usually make severe attempts. They may have similar clinical and psychosocial profiles to those who complete suicide. Since study- ing severe attempts can help us better understand completions and prevent these deaths, severe attempts can be considered as proxies for complete sui- cides. Hospitalizations due to suicide attempt tends to measure non-fatal and severe suicide attempts, the statistics of which are relatively easy to ac- cess in the clinical settings. It has been discovered that hospital data about suicide admissions tend to report more self-poisoning and lacerations cases than others kinds (Cantor and Neulinger, 2000). Cases of hospitalization are usually identified from hospital discharge data because admission data may not include in-hospital deaths. If a person who 12 Chapter 2. Literature review is admitted to hospital due to a suicide attempt dies before being discharged, it should be counted as a suicide death. However, in the admission data, this case may be defined merely as a hospitalized case. In our study, such events have been extremely rare; thus suicide attempts leading to hospitalization can be counted by looking at admission data. In the rest of the thesis, ‘suicide hospitalization’ refers to ‘hospitalization due to a severe suicide attempt’, and ‘suicide death’ or ‘suicide mortality’ refers to ‘suicide completion’. 2.1.2 Risk factors Numerous studies have been done to identify the risk factors for suicide, which suggests that suicide is rarely the response to a single stress. Instead, suicide is the health outcome of a complex interaction among biological, psychological, social, economic, and environmental risk factors. No one factor can be attributed independently to the outcome of suicide. Therefore, a study of the diverse risk factors can provide valuable information to improve our understanding of suicide. Suicides often occur for a number of reasons, including mental illness, serious physical illness, alcohol or other drug dependence, social isolation, death of a loved one, emotional trauma, and unemployment or financial problems (Agbayewa et al, 1998; Beautrais et al, 1996a; Faria et al, 2006; Maris, 2002; McIntosh, 1995; Qin et al, 2003). Among those factors, studies have suggested that mental illness, such as depression, mood disorder, sub- 13 Chapter 2. Literature review stance abuse disorder, and antisocial disorder, are associated with high rates of suicide (Beautrais et al, 1996a). About 90% of suicidal people have psychi- atric disorders and the most common symptoms are depression or psychosis (Mann, 2002). On the other hand, the risk factors for suicide involve not only psycholog- ical and other individual characteristics but also include the regional char- acteristics of the areas in which people live. Since suicide is an individual outcome that takes place in a social context, the regional factors may pose either an ecological risk (or an ecological protection), even after controlling for the individual-level effects (Cubbin et al, 2000; Agerbo et al, 2007b). The regional characteristics represent the nature of the region and are not the simple sum of the independent individuals who live there, even though they are aggregated from individual data. When aggregated to the population- level, those factors may influence suicide rates in the same way as they do at the individual level, or they may be associated with suicide in a different way. The current study examines the associations between suicide rates and regional characteristics at the ecological level. Next, literature regarding risk factors for both suicide and suicidal at- tempts were discussed. Since the goal of our study is to assess the population- level risk factors to explain the variation in suicide rates among regions, the risk factors identified in the individual studies, such as mental illness, family history and interpersonal issues, are briefly summarized in the following part. The demographic, social, and economic risk factors relevant to our study are 14 Chapter 2. Literature review reviewed at both the individual and population levels. Gender When examining how suicide risk varies across major demographic char- acteristics at both the individual and population levels, the data generally show that males have higher mortality rates than females (Agbayewa et al, 1998; Maki and Martikainen, 2007; Taylor et al, 2005). Beautrais (2003) compared serious suicide attempts and suicide among young people (under 25 years of age) and found a gender difference between complete suicides and suicide attempts, that is, females have much higher rates of attempt than males and the rates of suicide deaths are much higher in males. Simi- lar results were also found in other studies (Iribarren et al, 2000; Maki and Martikainen, 2007; Taylor et al, 2005). The differences by gender are often explained by the different methods of suicide. Males’ attempts at suicide tended to be more lethal than females’, a finding that was easily explained by the fact that males use more fast-acting and lethal methods than females (Kumar et al, 2006). Males are more likely to end their lives through effective violent means (firearms, knives, hanging, etc.), while women primarily use more failure-prone methods, such as overdosing on medications (Suominen et al, 2004). Risk factors associated with higher suicide rates also differ in males and females. While employment patterns are associated with elderly male suicide rates, suicide rates in elderly females have been found to be related to such factors as education level, income, and migration levels (Agbayewa et al, 1998). 15 Chapter 2. Literature review Suicide studies according to different age groups Whether at the individual or ecological level, there are well established relationships between age and many diseases including suicide. Suicide attempts and completions are uncommon in childhood (<15 years). Suicide rates increase markedly in the late teens and continue to rise until the early twenties, reaching a level that is maintained through adulthood until the sixth decade, when a small rise in senior populations. Different risk factors were contributed to different age populations (Burrows et al, 2007; Spirito and Esposito-Smythers, 2006; Åhs and Westerling, 2006). Youth suicides and suicide attempts usually stem from the influence of family, peers, school, and the community. Since adolescents and youth usu- ally live with or depend on their parents or other family members, the social, economic, and demographic status of family members, especially parents, has played a large role in youth suicide (Agerbo et al, 2002). Children with single parents or experiencing parental discord and/or impaired parent-child relationships may be at an increased risk of developing a psychiatric disease, which also precipitates their chances of suicide or suicide attempt (Gould et al, 2003; Steele and Doey, 2007; Beautrais, 2000). Having married parents can not only provide social and emotional stability and community integra- tion but also reduce social isolation to protect young people from suicide (Cantor and Neulinger, 2000). Adults are supposed to behave with maturity and responsibility. Because adults are supposed to be able to control their own behaviors, their marriage, 16 Chapter 2. Literature review family, employment, income, and education status may affect their behaviors in a great deal (Agerbo et al, 2007a; Åhs and Westerling, 2006; Cubbin et al, 2000; Hasselback et al, 1991; Rehkopf and Buka, 2006). For example, marriage may offer health and other advantages that divorced people lack and that can potentially reduce the risk for suicide (Faria et al, 2006; Kposowa, 2000). Low socioeconomic status, high mobility, and family instability may increase feelings of social isolation, a feeling that is seen as contributing to considering suicide (Cubbin et al, 2000; Middleton et al, 2004; Qin et al, 2003; Saunderson and Langford, 1996). Finally, low education levels may contribute a certain vulnerability to suicide (Faria et al, 2006; Kaneko and Motohashi, 2007). Elderly people kill themselves at a high rate compared with other gen- erations (Heisel, 2006; McIntosh, 1995). The risks seem to be higher in populations of those who have undergone either the loss of a spouse or who are unmarried (Conwell and Duberstein, 2001; Heisel, 2006); who are liv- ing alone or socially isolated (McIntosh, 1995; Rubenowitz et al, 2001); who are afflicted with identified illnesses, or severe chronic or intractable pain (Voaklander et al, 2008); who are experiencing a decrease in independence (McIntosh, 1995); or who have undergone recent financial difficulties and em- ployment and/or housing changes (Agbayewa et al, 1998; Rubenowitz et al, 2001; Shah et al, 2008). Therefore, it is worth exploring the potential risks from the demographical, social, and economic standpoints for this generation. 17 Chapter 2. Literature review Urban/rural disparity Studies support that rural areas tended to have higher suicide rates than urban areas. Some reasons given for this are a low population density; a high proportion of old adults; high unemployment rates; lack of social support, economic activity and access to medical service; and firearm ownership (Judd et al, 2006; Hirsch, 2006). While Pearce et al (2007) found that both male and female suicide rates were significantly higher in urban than rural areas in earlier times in New Zealand, current urban/rural differences in suicide rates do not appear to be significant. The narrowing of urban/rural differences was, to some extent, a result of the growth in suicide rates in more isolated rural communities and small rural service centers. Mental health Having a psychiatric disorder is the most significant risk factor for suicide across all age groups. Studies have shown that 80% to 90% of people who try to commit suicide have psychiatric disorders (Cheng, 1995; Maris, 2002; Beautrais et al, 1996a). These patients usually have one or more diagnosable psychiatric conditions, particularly depression, hopelessness, mood disorder, schizophrenia, and/or substance abuse. In his review (Mann, 2002), Mann pointed out that approximately 60% of all suicides occurred in persons with a mood disorder. Although suicide is generally a complication of a psychiatric disorder, most persons with a psychiatric disorder never attempt suicide. Having a substance abuse disorder is another risk for suicide, and it is usually related to alcohol-abuse or alcohol-dependency, and/or drug-abuse (Maris, 18 Chapter 2. Literature review 2002; Steele and Doey, 2007). Most people likely to commit suicide require an aggressive catalyst to do so. Such individuals are typically angrier, willing to be aggressive, irritable, or impulsive than non-suicidal control subjects (Maris, 2002). Family history issue Individuals who have family histories of attempted or completed suicide are themselves at higher risk of suicidal behavior. The effects of a family history of mental illness and/or suicide contribute to a great proportion of the risk of suicide. In a study by Iribarren et al (2000), they not only found that the differences in suicide outcomes were related to age, gender and race, but also reported that a prior suicide attempt was the strongest risk predictor for suicide completion. Familial risks may also exert an influence regarding suicide via genetic and biological factors (Brent and Melhem, 2008). Empirical evidence suggested that the tryptophan hydroxylase gene (TPH) may increase the risk of suicide (Kunugi et al, 1999; Rujescu et al, 2003). Neurobiological research has also clearly shown serotonergic function in suicidal individuals as compared to others (Mann and Malone, 1997). Interpersonal issues A prior suicide attempt has been found to be the strongest risk predictor for suicide completion. Hospitalization for suicide attempt was associated with a more than 20-fold greater risk of complete suicide, which suggested the need for psychiatric aftercare following suicide attempts (Iribarren et al, 19 Chapter 2. Literature review 2000). Children experiencing impaired parent-child relations are at an increased risk of suicide in youth category (Agerbo et al, 2002; Qin et al, 2003). Physical and sexual abuse during childhood can result in a high incidence of suicide. Beautrais (2000) concluded that the rates of suicide and suicide attempt ap- pear to increase among young people with histories of childhood physical and sexual abuse and suggested that the most common consequences of physical and sexual abuse were poor mental health outcomes, which may eventually result in suicide. Socio-demographic and socioeconomic factors Socio-demographic and socioeconomic factors, such as marital status, family structure, income, education, and employment, were usually inves- tigated in the suicide studies at both the individual and population levels. As individual risk factors, these factors may present people’s characteristics that potentially affect their suicidal behaviors. At the aggregated popula- tion level, these factors may represent the same associations as those at the individual level, or provide some regional characteristics that reflect social integration, which can assist in explaining the regional variations in suicide rates. The following discussion will clarify the rationale for the connection that these risk factors can be associated with suicide rates at both the indi- vidual and population levels. Marital status At the individual level, marital status is strongly related to suicide, as 20 Chapter 2. Literature review suicide has been shown to be higher for those who are divorced or widowed than for single people. The breakdown of a marriage can increase the suicide risk in general (Agerbo et al, 2007a; Iribarren et al, 2000; Qin et al, 2003). However, the study of gender differences in relation to the effects of marital status suggested that the protective effect of marriage for women was largely the result of being a parent (Qin et al, 2003), and that increased risks of suicide were observed mainly among divorced and separated men rather than their female counterparts (Kposowa, 2000; Qin et al, 2003). A large number of studies have reported that young people are more likely to commit suicide if their families have histories of parental separation or divorce (Agerbo et al, 2002). The impaired relationships of parents may suggest poor attachment and bonding between parents and child, which, in turn, may affect youth behaviors. Elderly suicides were significantly more likely to be among those who were married or widowed (Carney et al, 1994). The strong association between divorce rates and suicide rates can also be observed at the population level. Communities with high marriage rates may indicate that more individuals live with someone, have a greater connection to others, and are less socially isolated. Divorce increases the level of vulner- ability and stress within the family unit, resulting in poor family integration and reducing overall social integration. High divorce rates in a population may cause an increase in the suicide rate for that population due to reduc- tion in the level of social integration. Overall, findings support the notion that a loss of marital status may increase the risk of suicide. Higher rates 21 Chapter 2. Literature review of being unmarried, widowed, or divorced/separated are associated with a higher suicide rate (Masocco et al, 2008), even after adjustment of individual marital status (Agerbo et al, 2007b). However, in the study of Faria et al (2006), the association was found only in males. There were no associations were found in the elderly study by Agbayewa et al (1998), nor in the study by Middleton et al (2004). Divorce rates, at the regional level, especially of females, are highly correlated with youth suicide rates (Cutler et al, 2000). Family structure Family structure, which indicates the size of the family, the number of children in the household, and so on, is highly correlated with marital status. Social support from family may make people feel less socially isolated than those who live alone. Family cohesion help prevent suicide behaviors in children, especially for whom living with parents not only provides social and emotional stability but also reduces the sense of social isolation, which protects young people from suicide (Cantor and Neulinger, 2000). At the individual level, in the exploration of the contribution of parent- hood to protecting against the risk of suicide, a traditional family structure, such as having children in the family or being a parent of a young child, may be associated with lower suicide risks (Qin et al, 2003). In fact, a prospective study in Finland found that suicide rates in females decreased in proportion to the number of children they had (Cantor and Slater, 1995). On the other hand, an increased risk of suicide was associated with people living alone (Agerbo et al, 2007b). 22 Chapter 2. Literature review At the population level, family integration may increase social integration and thus decrease suicide rates. Family cohesion has been reported as a protective factor against suicide (Faria et al, 2006). The proportion of people living alone in a population was found to be associated with an increased risk for suicide even after accounting for individual risk factors (Agerbo et al, 2007b). Furthermore, a high percentage of single parents in an area, which may indicate a family breakdown or lack of family cohesion, was related to high suicide rates (Middleton et al, 2004). Indigenous population Interestingly, there is an abundance of literature on suicide among in- digenous people and most of them were done at the population level (Hunter and Harvey, 2002; Hutchinson, 2005; Hasselback et al, 1991). As it turns out, having cultural values, lifestyles, and social status that are different from the general population may contribute to the high rates of suicide. In their review, Hunter and Harvey (2002) pointed out that the suicide rates among indigenous youth in the United States were three times higher than for the nation as a whole; suicide among indigenous people in Australia were more concentrated among adolescent and young adult males and less common among older populations than in the mainstream; and Maori suicide rates were 1.2 times higher than non-Maori rates. High rates of suicide among Canadian aboriginal peoples have sparked a series of investigations in recent years (Hasselback et al, 1991; Lester, 1996). In Canada, the suicide rates in Aboriginal populations are usually higher than those in the general popula- 23 Chapter 2. Literature review tion. Many of these studies have attempted to understand Aboriginal suicide within a specific regional and cultural context. In their study, Hasselback et al (1991) identified high suicide rates among different cultural groups within Canada. High suicide rates were related to having a large proportion of Na- tive population in the census division. High suicide rates were also found in British Columbia, Canada (Lester, 1996). Mobility High levels of mobility in an area may make it harder for people to build up long-term families, friendships, and/or social connections, therefore in- creasing the degree of isolation, which may result in a high rate of suicide. Having lower mobility levels in a community may increase the likelihood that its members know each other and build a cohesive population, thereby reduc- ing the level of social integration (Schieman, 2005). Since social integration is related to lower rates of suicide; it is logical to conclude that high mobil- ity levels are therefore related to higher rates of suicide. At the population level, studies have found that the mobility of the population may reflect the population growth, and with a high mobility level, the social and family ties may be disrupted, thereby increasing suicide rates (Hasselback et al, 1991). Researches also drew connections between high rates of mobility and suicide rates in elderly women (Agbayewa et al, 1998). Immigrant Large numbers of immigrants from Asia, Africa, and the Caribbean have landed in Canada since the 1960s, which has resulted in a great ethnic and 24 Chapter 2. Literature review cultural mix in the population. Since the immigrant population has filled many occupational roles in the country, it represents an important part of the country’s economic and social growth. However, the unequal social struc- tures and dominant cultural values may contribute additional health risks for immigrant population. The existing literature on the suicide risks of immigration is mixed. Some studies suggest that immigration is positively related to suicide and other re- searches indicate the opposite at both the individual and population levels. At the individual level, researchers have reported that immigration increases the risk of suicide with some immigrant groups in Europe having higher sui- cide rates than the native population, while their divorce, unemployment, and poor social integration contribute to the risk of mental health prob- lems and suicide among immigrants (Qin et al, 2003; Westman et al, 2006; Kposowa et al, 2008). Having a high percentage of immigrants in a region has also been found to contribute to a high rate of suicide in ecological studies (Hasselback et al, 1991; Stack, 1981). Other studies, however, have found no significant association between immigrants and suicide (Åhs and Westerling, 2006; Agbayewa et al, 1998). An individual study conducted by Åhs and Westerling (2006) examined whether one’s country of birth was related to suicide risk, but no significant associations were found in that regard. An ecological study by Agbayewa et al (1998) also found that a high immigrant rate was not related to high suicide rates in elderly, but communities that attract immigrants might have certain attributes that contribute to suicide. 25 Chapter 2. Literature review For instance, speaking a first language other than official one might create barriers to accessing social services, health care services, and networking, especially for immigrants. Socioeconomic factors There is a large body of literature studying the associations between sui- cide and socioeconomic factors at both the individual and population levels. The most commonly used measures for socioeconomic factors are unemploy- ment, income, poverty/deprivation, and educational level. Other measures include: occupation-based social class, income inequality, and dwellings. These factors often correlate to each other and influence suicide simulta- neously. Unemployment Because unemployment usually has a negative impact on an individual’s life, it may increase the risk of subsequently developing a common mental disorder, thus increasing the risk of suicide. At the individual level, most studies have found a positive association between suicide and unemployed status of the individuals (Åhs and Westerling, 2006; Agerbo et al, 2007a,b; Blakely et al, 2003). Although unemployment was associated with increases in risks of suicidal behaviors, much of this association was explained after accounting for mental illness and psychiatric diagnosis (Fergusson et al, 2007; Kumar et al, 2006). The association may be mediated by such common fac- tors as childhood and family antecedents or psychiatric comorbidity (Agerbo et al, 2002). 26 Chapter 2. Literature review Since work provides an important source of social attachment, a commu- nity with a high rate of unemployment is assumed to be less integrated than a community in which most people are employed. Unemployment has received considerable attention as a population level factor in the risk of suicide. At the population level, ecological studies have found a positive relationship be- tween suicide rates and unemployment rates (Agbayewa et al, 1998; Rehkopf and Buka, 2006; Ferrada-Noli, 1997; Chotai, 2005). Since a high percentage of the population from the lower social classes tended to experience higher unemployment, this could result in higher suicide rates (Saunderson and Langford, 1996). While most studies have consistently found that the high unemployment rates contributes to an increased risk of suicide rates, in a Canadian ecological study, the suicide rates in the census divisions were neg- atively associated with unemployment rates (Hasselback et al, 1991). Income Unemployment can often result in a low income, which can make peo- ple feel insecurity and hopeless, and may lead to thought of suicide. At the individual level, most studies have found that suicide risk was strongly asso- ciated with low income (Agerbo et al, 2002, 2007a; Qin et al, 2003; Kposowa, 2000). In particularly, a low income increased the risk of suicide more in male than in female subjects, which may suggest that men respond more strongly to poor economic conditions than women do (Agerbo et al, 2002; Qin et al, 2003; Kposowa, 2000). High poverty rates tend to reduce the level of economic integration experi- 27 Chapter 2. Literature review enced by a group, which can also increase the likelihood of suicide. Increases in family income offered more material resources, educational opportunities and health services, which tended to reduce the risk of suicide (Mathur and Freeman, 2002). At the population level, most researchers who have ex- amined the relationship between income and suicide have found a negative association (Agbayewa et al, 1998; Ferrada-Noli, 1997; Hasselback et al, 1991; Hutchinson, 2005; Abel and Kruger, 2005). Agerbo et al (2007b) and Cubbin et al (2000) showed that the association was significant even after account- ing for individual level income factors. However, low income often interacted with other socioeconomic factors that affect suicide rates, since the associ- ation is often reduced after the adjustment for schooling or unemployment rates (Faria et al, 2006; Newman and Stuart, 2005). Education attainment People with high education levels tend to understand their social environ- ment better and adjust to their surroundings better. Being Well educated enables people to manage well in life. In the individual studies, the suicide rates were elevated by a lack of educational qualifications among both male and female subjects (Åhs and Westerling, 2006; Kposowa, 2000). This is also true among children and youth (Beautrais, 2003) and parents with low educational attainment (Iribarren et al, 2000; Kaneko and Motohashi, 2007). Difficulties in school, dropping out of high school, or not going to college pose significant risks for complete suicide and serious suicide attempts (Gould et al, 2003; Steele and Doey, 2007). These effects may decrease with adjustment 28 Chapter 2. Literature review for personal psychiatric diagnoses (Agerbo et al, 2002; Kumar et al, 2006). At the population level, the generally accepted relationship between sui- cide rates and education is that, as the level of education in a place increases, the suicide rate for that area may decrease, largely due to the increased ac- cess to goods and services resulting from increased education. Educational attainment has often been identified as a protective factor against suicide at the population level (Abel and Kruger, 2005; Faria et al, 2006; Ferrada-Noli, 1997; Kalediene et al, 2006). People having high education levels tend to have lower suicide rates (Hasselback et al, 1991). Poor education in a community is associated with a high rate of suicide (Agbayewa et al, 1998). Dwellings The dwelling status of a community reflects the neighborhood charac- teristics and social environment. The value of the social environment may influence the health of individuals and the population as a whole. The qual- ity of dwellings is also related to people’s income, employment, and family structure. In fact, the interactions among those factors may be associated with suicide rates. Ferrada-Noli (1997) found that an area with a higher proportion of suicides had an increased proportion of the population without home ownership. Indices for socioeconomic factors Besides the aggregated factors used in the studies of suicide, researchers have tended to summarize the social and economic characteristics about the region by developing indices of socioeconomic factors. The most common 29 Chapter 2. Literature review used indices are socioeconomic status (SES), social deprivation, and social fragmentation. Socioeconomic status (SES) SES is a combined total measure of a person’s economic and sociological situation. SES has been proposed as an important risk factor for suicide rates and is usually based on income, education, and occupation. At the individual level, SES may influence whether or not a person can use health services and health institutions and maintain his/her health status. Most studies that have examined association between socioeconomic factors and suicide have reported an increased risk of suicide among individuals from low SES groups (Beautrais et al, 1996b). There are also some studies that showed either small or no effects from SES (Gould et al, 2003). Specially, in children and adolescents, the effects of SES decreased when the study was adjusted for family history of mental illness(Agerbo et al, 2002). At the population level, SES may reflect the overall level of social inte- gration in the community and the health services and structures that can be accessed by the community. People with low socioeconomic status often lack the financial, social, and educational supports that people with high SES en- joy. A variety of studies, coming from North American (Cubbin et al, 2000; Hasselback et al, 1991), Europe (Maki and Martikainen, 2007; Middleton et al, 2004; Rezaeian et al, 2005), and Australia (Taylor et al, 2005), showed how differences in population-level socioeconomic status (SES) contribute to the differences in suicide rates. Most studies have found an inverse relation 30 Chapter 2. Literature review between suicide rates and SES, i.e., the higher the SES of the region, the lower the risk of suicide. Taylor et al (2005) found that SES had a signif- icant association with suicide at the population level, after controlling for the prevalence of mental disorders and other psychiatric symptomatologies. Moreover, SES differentials in suicide have widened in recent years. Maki and Martikainen (2007) conducted a study in Finland, in which they discovered that large socioeconomic differences were related to suicide mortality rates, noting especially high suicide rates among men from the lower socioeconomic classes. Lower SES was related to high suicide risks in census divisions (Has- selback et al, 1991), district level studies (Saunderson and Langford, 1996), and other area-level ecological studies (Rehkopf and Buka, 2006). Social deprivation and social fragmentation Various studies have constructed the measurements of social deprivation and fragmentation at the population level. Many such studies of social de- privation are used to describe the causes of poverty, which is concerned with people’s material circumstances, and how this impacts upon the condition of their lives. Townsend score was derived from census data on unemployment, car ownership, overcrowded housing, and housing tenure as composite indices of deprivation (Gunnell et al, 2005; Congdon, 1996b; Middleton et al, 2004). Developed by Carstairs and Morris (1991), the Carstairs index is an alterna- tive to the Townsend index of deprivation to avoid the use of households as denominators. The Carstairs index is based on four census indicators: low social class, lack of car ownership, overcrowding, and male unemployment. 31 Chapter 2. Literature review The geographical clusters and variations in suicide rates have been explained by the indices of the social deprivation of the region being studied (Congdon, 1996a; Exeter and Boyle, 2007; Rezaeian et al, 2005). Social fragmentation can be the disconnection between groups or com- munities, which may result in a decline in social cohesion. Congdon (1996b) proposed a census-based index of social fragmentation that predicted suicide rates, which he derived from census data on private renting, single person households (aged < 65), unmarried persons, and mobility in the previous year. Further studies often found that suicide rates are more strongly asso- ciated with measures of social fragmentation than with social deprivation at the population level (Whitley et al, 1999; Middleton et al, 2004; Abel and Kruger, 2005). 2.1.3 Suicide prevention From a public health perspective, suicide prevention programs focus on iden- tifying patterns of suicide and suicidal behavior throughout a group or pop- ulation. There are three levels of suicide prevention: primary, secondary (“intervention”), and tertiary (“postvention”) (Cantor and Baume, 1999). With the intention of influencing everyone in the population, primary prevention aims to reduce suicide risks in a population. Primary prevention can be done by public education, crisis hot lines, improving media reporting of warning signs, trends in rates, and treatment in advance of suicide, incor- porating screening programs into primary care practice, or reducing access 32 Chapter 2. Literature review to firearms, drugs, and other common means of suicide (Gould et al, 2003). Secondary prevention usually targets members of the population who are at particular risk for suicide. Such intervention focuses on early detection of individuals at risk, which can be done by improving the skills of local profes- sionals, such as teachers, health workers, or police officers to better recognize and respond to those at risk for suicide, or including suicide prevention pro- grams in school-base curriculums (Maris, 2002). Tertiary prevention – also called “postvention” – targets the particular individuals who have been affected by suicidal behavior. Postvention efforts can include close supervision and treatment of individuals who exhibit suici- dal behaviors or training family members to identify and respond to suicidal behavior. 2.2 Statistical methodology The goal of the study is to identify the areas at high risk for suicide and examine how the population-level risk factors explain the variations in suicide rates. Considerable geographical patterns and variations have been observed in the studies of suicide rates (Beautrais, 2000; Exeter and Boyle, 2007), which is due to the fact that suicide rates among neighboring regions are similar and typically present spatial dependency. The measurements of risk factors that are defined based on the regions are also spatially correlated, because people with similar characteristics tend to live close to each other. The assumption of independence does not hold in the spatial correlated data. 33 Chapter 2. Literature review However, ecological studies often use simple statistical methods to analyze such aggregated data in order to obtain the estimates of certain rates, for example, linear regression and correlation (Faria et al, 2006; Kalediene et al, 2006; Newman and Stuart, 2005; Shah et al, 2008). The interpretations based on these kinds of analysis may potentially contain biases and misleading results, since the methods may not be suitable to the nature of the data and/or fail to control for some important spatial patterns. Moreover, to our knowledge, there are only a limited number of studies that have investigated the simultaneous impacts of the various regional risk factors inherent in the demographic, social, economic, and environmental contexts on suicide rates at the population level (Newman and Stuart, 2005; Shah et al, 2008). If a broad range of population-level risk factors of interest are ready for examining the extent to which and how they can explain the variations in suicide rates and the factors involved are highly correlated, an ecological analysis requires us to find a way to avoid multicollinearity among covariates and reduce the dimensionality of data. Therefore, more sophisticated statistical methods are needed for the ecological and spatial studies to handle the spatially correlated data and multicollinearity among the covariates. 2.2.1 Bayesian disease mapping methodology As discussed in the previous section, the ecological and spatial studies need more sophisticated statistical methods to obtain reliable results by reducing 34 Chapter 2. Literature review random error due to spatial correlations, as well as analyzing and visualizing the risks over space. While the spatial clustering of suicide rates is common among areas, ignoring such spatial correlations of suicide rates may result in biased or misleading conclusions. In fact, researchers whose intention is to model the spatial clusters and variation in their disease mapping studies, have noticed this problem (Congdon, 1996a, 2000; Exeter and Boyle, 2007). Disease mapping refers to estimating and presenting the geographical dis- tribution of a disease within a population. The construction of disease maps has been a central part of descriptive epidemiology throughout its history. From the famous use of mapping in the time of the cholera epidemics in London in the middle of the nineteenth century (Snow, 1936), disease map- ping at the small-area level has grown at a rapid pace (Elliott and Wakefield, 2000; Lawson and Williams, 2001). The variations and clusters of rates in the maps, which are barely discernable in the tabular summary, can play an essential part in the interpretation of the geographical distribution of a disease. Disease maps of the crude mortality/morbidity rates or Standard- ized Mortality Ratio (SMR) have been studied over time (Breslow and Day, 1987). It is well known that maps usually lack reliability when the size of a population for a certain region is small or the disease is rare (Breslow and Clayton, 1993). Because of the dramatic variability in crude rates or SMR, it is hard to distinguish the true differences from random errors for the areas. In order to overcome the drawbacks of crude rates and the SMR, re- searchers have begun to use more sophisticated methods to model disease 35 Chapter 2. Literature review rates. Because of the clustering patterns displayed in the rates of diseases across areas, they account for the geographical variations and clusters in the risks of the diseases by smoothing the estimates of the rates, which they do via “borrowing strength” from neighboring areas. A variety of disease map- ping models are proposed to address the problem, including nonparametric smoothing models, linear Bayes methods, and Bayesian models (Lawson et al, 2000). The generalized linear mixed effects model (GLMM) with Bayesian in- ference facilitates the studies of geographical variations and spatial patterns by estimating the area-specific smoothed disease rates as well as the effects of the potential risk factors related to the disease. This kind of model-based approach can be Bayesian in nature and provide suitable estimates of rela- tive risks for each small area. Clayton and Kaldor (1987) first proposed a GLMM model with a multivariate normal distribution for the random effects that explains spatial correlation and overdispersion. Other possible models for specifying the random effects have been discussed by Besag et al (1991); Besag and Kooperberg (1995); and Ainsworth and Dean (2006). Leroux et al (1999) proposed a new model for studying spatial dependence. Their model is essentially a Gaussian Markov random field (GMRF) model, which is imple- mented by Besag (1974) to specify the conditional distribution in neighboring regions. The model is a so-called Conditional Autoregressive model (CAR). Using the CAR model to specify the random effects, the smoothed estimates are obtained by considering the potential tendency of neighboring regions to 36 Chapter 2. Literature review have similar disease risks. Different CAR models that specify the random effects used in the disease mapping have been further researched in many other studies (Besag and Kooperberg, 1995; MacNab and Dean, 2000, 2001; MacNab, 2003b; Ainsworth and Dean, 2006). The methodology of analyzing GLMM has been expanded from its early developments in the Empirical Bayes methods of estimation via moment methods or penalized quasi-likelihood estimates (EB-PQL) (Besag et al, 1991; Breslow and Clayton, 1993; Clayton and Kaldor, 1987; Clayton and Bernardinelli, 1993; Leroux et al, 1999; MacNab and Dean, 2000) to the Fully Bayesian methods of inference via Markov chain Monte Carlo method (FB-MCMC) (Besag et al, 1991; Bernardinelli and Montomoli, 1992; Elliott and Wakefield, 2000; MacNab and Dean, 2000; MacNab, 2004). While the EB-PQL method estimates parameters effectively, the FB-MCMC approach enables us to quantify uncertainties associated with the estimation for all model parameters, particularly random effects (Bernardinelli and Montomoli, 1992). However, both the EB-PQL and the FB-MCMC methods have weak- nesses. The EB-PQL method underestimates the uncertainties associated with random components and the FB-MCMC method is time-consuming in computation. 2.2.2 Small-area studies The fact that people tend to cluster in areas in systematic ways may reflect the considerable geographical patterns of disease rates among the populations 37 Chapter 2. Literature review of those areas means that the regional characteristics, such as demographical, social, and economic factors that specifically increase the risk of disease need to be examined. Small area studies based on disease mapping methodology may help to tackle how these regional characteristics are related to disease and identify populations at a particularly high risk of disease, controlling for the geographical patterns of the outcome. A number of researches have incorporated ecological studies into disease mapping analysis to explore the geographical variations in disease risks and investigate the associated risks. These studies use different statistical methods to smooth the disease rates and quantify the potential risk factors. We review some of them in the following section. Using Empirical Bayes estimates, Saunderson and Langford (1996) stud- ied the geographical distribution of suicide rates at the district level in Eng- land and Wales. They suggested that the effects of economic hardship, un- employment, and social disintegration also relate to suicide rates, especially for those suffering from psychiatric vulnerability or mental illness. Using a fully Bayesian approach, Congdon (2000) modeled the differential changes in suicide by borough and time over a 16-year period showing special con- cern for the spatial clustering and temporal correlations among relative risks. The results showed that social deprivation may be related to the variations in suicide rates in different areas. Similarly, MacNab (2004) applied Bayesian spatial and ecological models to study the small-area variations in hospitalization rates of accident injuries 38 Chapter 2. Literature review among male children and youth. Her methods not only considered the spa- tial patterns of the disease but also assessed some regional characteristics, including socioeconomic disadvantage, speed, crime, service availability, and utilization. The results showed considerable geographical variations in acci- dent injuries and some associated risk factors. Using one approach may not be sufficient to handle large data with a broad spectrum of risk factors. To highlight the strengths of both the Em- pirical Bayes method and the Fully Bayesian method and minimize their weaknesses, MacNab et al (2006) introduced an analytic strategy for the complementary use of EB and FB inferential techniques for risk assessment and model selection in a small area study of adverse medical events in British Columbia, Canada. The model selection was done using the EB method be- cause of its ability to produce nearly unbiased point estimates and the final ecological model was fitted using the FB method, as it could quantify the uncertainties associated with the estimates of all parameters. Moreover, other statistical methods were also used in small area studies. Exeter and Boyle (2007) determined geographical clusters of suicide rates in Scottish young adults using a spatial scan statistics test. The significant geographical clusters of suicides among young adults that were found in east Glasgow could potentially be explained by the concentration of social depra- vation in that part of Scotland. More recently, researchers have extended the models to allow the inves- tigation of more than one disease simultaneously. Shared Component Model 39 Chapter 2. Literature review can be used to identify the common and disease-specific risks for two or more diseases. This method may provide more convincing evidence of real clus- tering in the underlying risk surface (Knorr-Held and Best, 2001). Shared component models have been used to study the rates and relationships among two or more diseases simultaneously (Knorr-Held and Best, 2001; Knorr-Held et al, 2005; MacNab, 2007; MacNab and Lin, 2008). The smoothed estimates account for the variations and correlations between diseases as well as the variations across diseases, which is a novel element of the shared component approach. 2.2.3 Principal component analysis In observational studies, researchers often need to explore the large num- bers of explanatory factors, which are found to potentially correlate with each other. For example, people with a high income tend to live in areas with better housing and schooling than poor people, so the measurements of income, dwellings, and education may be highly correlated. With limited ob- servations, we are concerned about whether or not adding all the explanatory factors into the analysis is helpful in exploring the relationships of interest. Principal Component Analysis (PCA) can help us deal with such a situation. The idea of PCA is to reduce the dimensionality of a data set and avoid mul- ticollinearity among a large number of interrelated variables, while retaining as much as possible of the variation present in the data set. PCA can help to identify patterns in the data and express the data in such a way as to 40 Chapter 2. Literature review highlight their similarities and differences. PCA involves a mathematical procedure that transforms a number of possibly correlated variables into a smaller number of uncorrelated variables, called principal components (PCs). Depending on the field of application, this is also called the discrete Karhunen-Loève transformation, the Hotelling transformation, or proper orthogonal decomposition. First introduced by Pearson (1901), PCA is a classical statistical method mostly used as a tool in exploratory data analysis and for making predictive models. PCA has been used in a wide range of science fields, such as face recog- nition and image compression in engineering, or searching for genes in mi- croarray data to fix biomedical problems. PCA has also been used in a study of suicide mortality and regional variation to reduce the number of variables while still preserving as much of the original information as possible (Marusic, 1998). The author grouped variables with strong correlations into different risk patterns to examine the associations with suicide rates. Besides reduc- ing the dimensionality of the data and avoiding correlation among variables, PCA has also been used for conducting socioeconomic status indices (Vyas and Kumaranayake, 2006), which can be used in ecological studies. There are alternatives to PCA that can reduce the dimensionality of the data and derive indices, such as correspondence analysis or factor analy- sis. Cortinovis et al (1993) used correspondence analysis to derive an SES measure. The analysis can only perform among categorical variables, so con- tinuous variables were needed to be categorized instead. Lester (1988) used 41 Chapter 2. Literature review factor analysis to develop the indicators for a number of socioeconomic vari- ables in order to explore the relationship between suicide and homicide rates and socioeconomic variables. The advantage of PCA is that it can perform a linear transformation without making any assumptions. In order to evaluate the relationship between suicide rates and various potential risk factors in our study, PCA was used to reduce the number of risk factors into a small number of risk patterns, which were used as covariates to add into the Bayesian disease mapping and ecological regression. In our study, PCA was used mainly to avoid multicollinearity, reduce the dimensionality of the data, and indicate the risk patterns associated with suicide rates in different age and sex groups. Choosing PCs as covariates enabled us to avoid multicollinearity among the variables and reduce the dimensions of covariates, while retaining as much information as possible regarding the risk factors. The results can represent patterns of aggregated population characteristics that are acting as complex regional risk factors for suicide rates. More details about PCA will be discussed in Chapter 4. 42 Chapter 3 Data This chapter describes the data used in the study and presents summary statistics of B.C.’s suicide hospitalization and death data. 3.1 Study population and Data sources 3.1.1 Study region In order to describe the geographical patterns and variations in suicide rates across regions, health service administrative areas were introduced to create appropriate small-area disease mapping models in our study. Provided by the B.C. Ministry of Health, the health care service in B.C. is managed and delivered by five Health Authorities (HAs) in various geographic regions of B.C., in order to ensure B.C. residents have access to a high-quality, patient- centered, and sustainable health care system, regardless of where they live in B.C. During the period of study, 5 HAs coordinated service among 16 Health Service Delivery Areas (HSDAs) and among 84 Local Health Areas (LHAs). According to the statements from the B.C. Ministry of Health (B.C. Health Authorities, 2007), the 5 HAs cooperate with one Provincial Health Services Authority (PHSA), which coordinates and/or provides provincial programs and specialized services. The PHSA is responsible for identify- 43 Chapter 3. Data ing the population’s health needs; planning appropriate programs and ser- vices; ensuring programs and services are properly funded and managed; and meeting performance objectives. The five geographical health authori- ties are: Northern Health Authority, Interior Health Authority, Vancouver Island Health Authority, Vancouver Coastal Health Authority, and Fraser Health Authority. The HAs are further divided into 16 HSDAs. The HSDAs work with the HAs to plan and coordinate the delivery of provincial programs and highly specialized services, such as transplants and cardiac care, as well as govern and manage the organizations that provide health services on a province- wide basis (B.C. Health Authorities, 2007). A map of the HAs and HSDAs is shown in Figure 3.1. Local health areas are the next layer in the provincial health adminis- tration system. B.C. residents in LHAs can evenly access the health care system. A local health area is a unit small enough to balance the representa- tions for homogeneity and variability. The local health area used in the study provides homogeneous health service environments for patients in different locations of B.C. The entire province was divided into specific small areas whose characteristics are similar, in order to better describe their features. Eighty-four local health areas are used as the units to measure outcomes in the current study. The LHAs map is shown in Figure 3.2. Note that the results and conclusions in the thesis have been interpreted as the impacts on populations at the LHA level. 44 Chapter 3. Data Figure 3.1: Map of Health Authorities and Health Service Delivery Areas, B.C. 3.1.2 Study population The present study explored and quantified geographical patterns and varia- tions in suicide hospitalization and mortality rates and associated risk factors in the province of B.C., Canada. The study focused on the occurrences of sui- cide hospitalization due to a severe attempt and suicide death among B.C.’s population during the period 1991–2000. Data were collected based on cal- 45 Chapter 3. Data Figure 3.2: Map of Local Health Areas, B.C. 46 Chapter 3. Data endar years. During the study period, the mid-year population of B.C. in 1996 was about 3.83 million. The map of population size in 1996 is shown in Figure 3.3, which illustrates the distribution of population at the LHA level in B.C. The highest density of population is located in the Greater Vancouver and Interior Health Authority areas. B.C. Population 1996 low med high Figure 3.3: Population size of B.C. in 1996, by Local Health Areas. 3.1.3 Data sources In order to estimate suicide rates and identify the risk factors in various re- gions, administrative data were used to identify suicide events (including hos- pitalization for severe suicide attempts and suicide death), the corresponding population estimates, and the potential risk factors. Hospitalization Data. Data on all hospitalized individuals in B.C. be- tween 1991 and 2000 were available for the study. The records obtained were provided by the B.C. Ministry of Health and based on the hospital admis- 47 Chapter 3. Data sions/separation data in B.C. between 1991 and 2000. The hospitalized data contained patients’ hospital charts from all hospitals in B.C., with the level of care cited as either acute care or day surgery. The majority of individuals hospitalized during the study period had their principal diagnoses coded with an ICD9-CM E-code (0.3% in all hospitalized cases were missing diagnosis codes) and were B.C. residents (the percentage of B.C. residents was 99.9%). Patients who were hospitalized for less than 6 hours, due to sustaining only a minor injury, were not included in the database. The patient’s age, gender, and region of residence (LHA) were coded in the database. Suicide Death Data. The suicide mortality database was developed and provided by the B.C. Vital Statistics Agency. The data included any person who died during the period of 1991–2000 and whose cause of death was coded with any of the ICD9-CM E-codes. The individual’s age, gender, and region of residence were included in the database. Our study focused on B.C. residents who committed suicide in the period from 1991–2000. Population estimates. In order to estimate the suicide rates in all regions, we needed to know the population at risk in order to set it as the denominator. Mid-year estimates of 5-year-group populations at the LHA level were obtained from B.C. STATs, a central statistical agency in B.C. The population at risk during 1991-2000 was estimated by the sum of the 10 mid-year population estimates at the LHA level. Regional characteristic profiles. The regional characteristics were se- lected from the 1996 Census in Canada, which was the mid-year of our study 48 Chapter 3. Data period. The variables in the census were collected from a 20% sample of the entire population. The domains were defined based on topics from the census data. In each domain, several variables were included to represent the major information on the topic. The domains presented the regional characteristics of the population, ranging from marital status, language, immigration, abo- riginal population, mobility status, through educational attainment, unem- ployment rates, family, transportation, to income and dwellings. Appendix A lists all the risk factors of interest from different domains in our study. At the LHA level, most of the variables were summarized as the percentages of the population with certain characteristics. A high percentage in a region in- dicated the concentration of the risk factor in the area. In total, 87 variables from 11 domains were derived from the census. 3.1.4 Data extraction and aggregation In the International Classification of Disease, Ninth Revision, Clinical Modi- fication, External coding scheme, i.e. ICD9-CM E-codes, E950-E959 indicate suicide and self-inflicted injuries. In the hospital separation data, patients with the first occurrence of a cause of injury (usually in the principal diagnosis, coded as E950-E959), and who were admitted to hospital during the study period were recorded as suicide hospitalizations. If a patient’s principal diagnosis was missing and the second diagnosis was available, we used the second diagnosis in the case identification. If the patient was transferred between hospitals, the first 49 Chapter 3. Data admission record was used. Suicide deaths were extracted by the code E950- E959 in the mortality database. Both suicide hospitalization and mortality data were aggregated by gen- der and LHA. The counts were further grouped into 5-year age units, (i.e., <1, 1-4, . . ., 80-84, 85+), 19 groups in total. The crude rates in each 5-year age group and the trend of the rates in all groups are shown in Figure 3.4. The analysis looked at suicide hospitalization and mortality rates for both genders separately. Since we didn’t have sufficient cases in suicide hospital- ization and mortality data for certain age groups, larger age subgroups were created by aggregating 5-year age groups. In order to provide enough cases in each subgroup after checking the trend of the rates, the subgroups were created as: 0-14 as children, 15-24 as youths, 25-34 as young adults, 35-44 as adult working group I, 45-54 as adult working group II, 55-64 as adult working group III, 65-74 as seniors, and 75+ as elderly. 3.2 Descriptive summary of the suicide data During the study period between 1991 and 2000, 38,870 suicide hospitaliza- tions and 4,883 suicide deaths were identified in B.C. The annual rate was the number of suicide hospitalizations or suicide deaths over the aggregated mid-year estimates of population size over all LHAs and all age groups for each year. The B.C. annual suicide hospitalization and mortality rates for all ages are shown in Figure 3.5. The crude suicide hospitalization rates re- mained stable over time, decreased a bit in 1999 and returned to usual in 50 Chapter 3. Data 0 1−4 10−14 20−24 30−34 40−44 50−54 60−64 70−74 80−85 Male Suicide Hospitalization Female Suicide Hospitalization Crude Rates 0 50 10 0 15 0 20 0 25 0 30 0 0 1−4 10−14 20−24 30−34 40−44 50−54 60−64 70−74 80−85 Male Suicide Death Female Suicide Death Crude Rates 0 20 40 60 80 Figure 3.4: Crude age-specific rates of suicide hospitalization and death, males and females, 1991-2000 51 Chapter 3. Data 2000. Roughly overall, about 110 per 100,000 persons were hospitalized due to severe suicide attempts. Females had much higher rates than males (130 vs. 80 per 100,000), which was the opposite of suicide mortality rates. Sui- cide mortality rates continued to be steady over the 10 years and tended to decrease near the end of the period. Male suicide mortality rates were 3 to 4 times more frequent than female, which ranged from 16 to 22 per 100,000 persons. Annual rates were also calculated for the HA and HSDA levels, which showed variations among the health regions (see Appendix B). Besides the gender differences in suicide hospitalization and mortality rates, age differences were found within the male and female categories. Ac- cording to Figure 3.4, age effects presented different patterns in suicide hos- pitalization and mortality rates. The suicide hospitalization rates increased to a peak in adolescence and began to decrease slowly as ages increased, although they increased slightly in seniors. Higher rates occurred among youth and adult working groups, especially in female subjects. Extremely high suicide hospitalization rates occurred in female youth. Suicide mortality rates were more stable than suicide hospitalization rates for both males and females across all age groups. Female deaths were rare across all age groups. While a high density of male suicide mortality was observed in the middle age range, the elderly showed the highest rates. Children under 14 years old rarely completed suicide. The crude suicide hospitalization and mortality rates were also calculated for each group and for males and females separately. Besides the overall 52 Chapter 3. Data 1992 1994 1996 1998 2000 80 90 10 0 11 0 12 0 13 0 14 0 Annual BC crude suicide hospalization rates 1991−2000 Year R at e pe r 1 00 ,0 00  p er so n BC Male Female 1992 1994 1996 1998 2000 5 10 15 20 Annual BC crude suicide death rates 1991−2000 Year R at e pe r 1 00 ,0 00  p er so n BC Male Female Figure 3.5: Annual B.C. crude rates of suicide hospitalization and mortality, males and females, 1991-2000 53 Chapter 3. Data Table 3.1: Crude LHA-specific suicide hospitalization rates, males and fe- males, 1991-2000. (per 100,000 population) Overall LHA-specific rates Male Subgroups Rate Mean Sd Q1 Q3 Children (0–14) 7.80 7.52 8.18 0.00 10.62 Youth (15–24) 130.55 177.30 145.36 99.86 192.20 Young adult(25–34) 129.91 183.80 189.04 104.00 199.80 Working group I (35–44) 116.25 143.80 119.34 84.02 185.60 Working group II(45–54) 74.12 77.79 45.73 45.56 108.60 Working group III(55–64) 44.61 44.30 39.27 18.29 66.75 Seniors (65–74) 33.55 30.80 33.05 0.00 42.87 Elderly (75+) 48.58 43.14 54.93 0.00 61.12 Overall LHA-specific rates Female Subgroups Rate Mean Sd Q1 Q3 Children (0–14) 34.85 46.85 44.01 24.19 61.39 Youth (15–24) 283.47 395.60 439.59 234.10 408.40 Young adult(25–34) 190.47 286.10 277.63 145.60 287.20 Working group I (35–44) 178.23 234.10 161.12 134.60 296.90 Working group II(45–54) 112.72 127.50 81.31 80.28 159.80 Working group III(55–64) 53.06 62.21 57.68 28.38 79.55 Seniors (65–74) 33.26 35.22 59.36 0.00 43.90 Elderly (75+) 32.13 34.35 78.16 0.00 43.15 crude rates shown in Table 3.1 and 3.2, we calculated crude LHA-specific rates for both genders. Crude LHA-specific suicide hospitalization rates were consistently higher than mortality rates among the entire province. Based on the summary statistics of crude LHA rates for suicide hospitalization and mortality, considerable variations were observed among LHAs for all age groups. For adolescents, the male suicide hospitalization rate was 131 per 100,000 54 Chapter 3. Data Table 3.2: Crude LHA-specific suicide mortality rates, males and females, 1991-2000. (per 100,000 population) Overall LHA-specific rates Male Subgroups Rate Mean Sd Q1 Q3 Children(0–14) 0.77 1.40 4.65 0.00 0.90 Youth (15–24) 19.85 28.09 35.91 14.48 32.99 Young adult(25–34) 25.84 35.23 24.38 18.23 46.71 Working group I (35–44) 27.48 31.62 20.70 20.23 36.30 Working group II(45–54) 26.30 33.08 32.78 16.81 36.14 Working group III(55–64) 23.09 26.07 24.97 9.81 33.69 Seniors (65–74) 21.67 29.57 37.51 9.53 37.53 Elderly (75+) 31.85 30.82 35.48 0.00 45.82 Overall LHA-specific rates Female Subgroups Rate Mean Sd Q1 Q3 Children (0–14) 0.37 0.39 1.23 0.00 0.00 Youth (15–24) 5.18 6.90 12.23 0.00 7.63 Young adult(25–34) 6.07 6.52 10.31 0.00 8.48 Working group I (35–44) 7.65 8.12 11.63 0.00 12.57 Working group II(45–54) 8.33 8.61 9.17 0.00 13.17 Working group III(55–64) 7.72 7.04 11.03 0.00 11.03 Seniors (65–74) 7.66 7.10 11.24 0.00 9.01 Elderly (75+) 7.26 4.94 9.41 0.00 7.52 persons, which was about 6.5 times the male suicide mortality rate. The suicide hospitalization rates of female youth were even higher. About 283 per 100,000 females were hospitalized for serious attempts, which was about 57 times the suicide mortality rate. The suicide hospitalization rates for female attempt were more than twice the male rates (283 vs. 131 per 100,000 persons). However, male suicide mortality rates were notably higher than those for females (20 vs. 5 per 100,000 persons), although the number of cases 55 Chapter 3. Data that ended in death was quite low in the youth population. In this study, we observed that young males tended to have more severe suicidal behaviors than females. This seemed to indicate that if a young male attempts suicide, he is more likely to succeed than his female counterpart. Among adults aged 25-64, four 10-year subgroups were used for the anal- ysis. The male suicide hospitalization rate for young adults (aged 25-34) was still high (130 per 100,000 persons) and then decreased as they got older. The suicide hospitalization rate for young males was above 3 times the rate for the senior working group (aged 55-64) (130 vs. 44 per 100,000 persons). Young males tended to attempt more suicides than mature males. This seemed to indicate that, with increased experience people can handle problems in a less negative or dramatic way. Meanwhile, the suicide mortality rates were stable over time and decreased slightly as age increased. While suicide hospitaliza- tion rates for females gradually dropped down from those in the young adult group to those in the middle aged working group, the rate for the middle aged was about half the rates for the senior working group (116 vs. 53 per 100,000 persons). Suicide mortality was rare in females across all age groups. Among adults, suicide hospitalization rates were still higher in females than in males, but the gender gap tended to close as people got older. Female suicide mortality rates were consistently lower than males. Only about 7 per 100,000 persons were found to kill themselves in B.C., according to the average rates in the female population aged 25-64. Male suicide mortality rates were about 3-4 times the female rates. 56 Chapter 3. Data Two 10-year subgroups of the elderly were studied. Suicide hospitaliza- tion rates for those aged 65 and over were quite close in both males and females (Male and Female 65-74: 33 vs. 33; 75+: 48 vs. 32 per 100,000 per- sons). Meanwhile, male suicide mortality rates were 3-4 times higher than female ones (Male and Female 65-74: 22 vs. 8; 75+: 32 vs. 7 per 100,000 persons). Older male groups tended to have higher rates for both suicide hospitalization and mortality. In general, both suicide hospitalization and mortality rates showed large gender differences in B.C. For each gender, the trends of rates were differ- ent with increasing age. Male suicide hospitalization rates showed a small U-shape; female suicide hospitalization rates tended to decrease; male sui- cide mortality rates increased slightly with age; and female suicide mortality rates kept stable across all age groups. Another important feature in suicide hospitalization and mortality rates was large geographical variation, which showed contrasts between genders and age groups. 3.3 Descriptive summary of regional characteristics The statistics for the regional characteristics across all LHAs are summarized in Appendix A. The concentrations of regional characteristics varied among LHAs, which may be related to the variation of suicide rates among areas. The variables chosen to describe and represent the characteristics for each domain potentially correlated, because they delineated the main character- 57 Chapter 3. Data istics from the same perspective. Since the variables between domains were probably correlated as well, they also had the potential to influence each other. For example, variables related to high income were positively associ- ated with those from the high education level domain and negatively related to unemployment rates, because people with higher education levels tended to have better jobs with higher pay. 58 Chapter 4 Statistical methodology In this chapter, a Bayesian disease mapping framework is presented for conducting the spatial and ecological analysis of suicide risks. The Fully Bayesian method is used to estimate the suicide risks and their correlated risk factors. Both the Empirical Bayes method and Principal Component Analysis are used in the exploratory analysis to do variable selection and re- duce the number of covariates and avoid multicollinearity respectively. Note that the present study considers the inequalities in suicide rates by local health areas, by gender, and by different age groups. The following model was fitted for age- and sex-specific outcomes. 4.1 Spatial-ecological models One of the main interests of our study is regional suicide rates. Suicide rates among neighboring regions are close and typically present spatial dependency. The exploratory variables were the potential risk factors for suicide from cen- sus data. The measurements of risk factors that are defined based on the regions are also spatially correlated because people with similar character- istics tend to live close to each other. Therefore, models of regional suicide rates need to account for the spatial dependency. In our study, Poisson re- 59 Chapter 4. Statistical methodology gression model, along with random effects, was used to model the regional suicide rates and tackle the spatial dependency. Suppose Yj is the count of deaths or other outcomes that have occurred in the jth region (j = 1, . . . , J) andNj is the corresponding “at-risk” population size for a certain age and gender group. Yj follows Poisson distribution with mean µj. Poisson regression model expresses that the logarithm of the outcome rate is linked to a linear function such that log(µj) = log(Nj) + β0 + bj , (4.1) where log(µj)− log(Nj) = log(µj/Nj) is the logarithm of the outcome rate; β0 is the logarithm of the overall rate across the regions that accounts for the population average; and bj is the random effect that accounts for both unmeasured confounding and the spatial dependency in the residual of suicide rates. Here, we call it the disease mapping model. When the potential risk factors are accounted for by explaining the vari- ation in suicide rates, Poisson regression model (i.e., ecological regression) can be extended as log(µj) = log(Nj) + β0 + K∑ k=1 βkXkj + bj , (4.2) where the logarithm of the outcome rate log(µj/Nj) is estimated by the ex- ploratory variables ∑K k=1 βkXkj and residual random effects b. β1, . . . , βK represent the coefficients of the associated risk factors, and X1, . . . , XK rep- resent K risk factors; β0 is the logarithm of the overall rate across regions; 60 Chapter 4. Statistical methodology and bj quantifies the residual and unmeasured covariate/confounding effects. RRj = exp(bj) measures the risk-adjusted relative risk (RR) compared to the reference rate (Here the reference rate is the overall rate), with RRj > 1 (RRj < 1) reflecting a high (low) risk of the disease or health outcome oc- currence for the jth region. Therefore, in the Poisson regression model for the regional suicide rates, there are two sources of variation: ∑K k=1 βkXkj, the risk factors related to suicide rates as fixed effects, and bjs, as random effects, which represent the residual and unmeasured confounding. The spatial correlation is accounted for in the residual term. If suicide rates among neighboring regions present spatial dependency, the random effects are assumed to follow a specific multivariate normal dis- tribution to express that neighboring regions have similar rates and risks of suicide. Here, a non-intrinsic Conditional Autoregressive model (CAR) (Leroux et al, 1999) was used for the random effects b = (b1, ..., bJ) ⊤. b ∼ MVN(0,Σ(σ2, λ)), (4.3) Σ(σ2, λ) = σ2D−1, (4.4) D = λQ+ (1− λ)IJ , (4.5) where λ (0 < λ < 1) is a spatial autocorrelation parameter that quantifies the extent to which neighboring regions have similar rates and risks; σ2 is the dispersion parameter that quantifies the relative risk variance; IJ is an identity matrix with the dimension J ; the neighborhood matrix Q is J × J , 61 Chapter 4. Statistical methodology whose jth diagonal element is the number of its neighbors and its off-diagonal elements in each row are −1 if the corresponding areas are neighbors and 0 otherwise. Here, the neighborhood was defined by the areas that share a common boundary. In the ecological model (4.2), the risk factors were typically spatially correlated as well. When the majority of spatial variations are explained by the risk factors, an independent/exchangeable prior (IID prior) is assumed for the random effects b ∼ N(0, σ2IJ), (4.6) which indicates “global” smoothing on the estimates of the rates. In the present study, the random effects b brought a “spatial” dimension into the model formulation to accommodate spatially correlated small area disease rates/risks. In Bayesian inference, the spatial structure for the ran- dom effect is considered as a prior specification, which typically facilitates “borrowing strength” to stabilize area-specific RR estimates. Compared with the crude rates, the resulting inference of relative risks is often more conser- vative, particularly in terms of signaling high risk areas, and thereby results in fewer false positive alarms. In our study, the Empirical Bayes via Penalized Quasi-Likelihood (EB- PQL) method and the Fully Bayesian method via the MCMC technique (FB-MCMC) were used to estimate and make inferences from the Poisson regression model with random effects. The EB-PQL method can efficiently obtain adequate point estimates for fixed and random effects and can also 62 Chapter 4. Statistical methodology produce reasonable confidence intervals for the fixed effects. Its main weak- ness is its underestimation of the variance of the random effects and random components (MacNab et al, 2004). The FB-MCMC method can overcome the drawbacks of the EB-PQL method. The point estimates and uncertainty for all parameters can be quantified directly from posterior distributions. However, this method can be time-consuming and may fail to converge as the models get more complicated. In order to highlight the strengths, both methods are utilized for the inferences from Bayesian disease mapping and the ecological models. The FB-MCMC method was used to estimate area- specific suicide relative risks, investigate spatial variations, and explore and quantify the associations between regional characteristics and suicide occur- rences. The EB-PQL method was applied in the selection of variables. When conducting a Fully Bayesian analysis, it is important to consider the choice of priors. Because the inference from the Bayesian approach tries to combine the information from the data and the knowledge prior to the data, a prior may dominate the estimates of the parameters. A Bayesian sen- sitivity/robustness analysis for a variety of alternative prior and hyperprior specifications can be carried out to examine the impact of priors on posterior inference, see Appendix C. 4.2 Exploratory analysis In order to examine the spatial dependency in suicide risks, the Poisson regression model (4.1) was first fitted using the FB-MCMC method, with a 63 Chapter 4. Statistical methodology CAR prior (4.3) - (4.5) on random effects b. When the posterior estimate of the spatial effect λ is not statistically significant, it usually indicates weak “spatial correlation” in the suicide rates and thus the Poisson regression model (4.1) are fitted with the IID prior (4.6) for the random effects. Since a large number of exploratory variables obtained from the census data were ready to be evaluated, univariate regressions were conducted to select the significant risk factors for suicide rates. A Poisson regression model with one covariate was fitted for each census variable by the EB-PQL method. Generally, many of them were shown to be significantly related to suicide rates. Thus, after the univariate analysis, a large number of significant risk fac- tors, usually more than 20, needed to be assessed in the ecological regression model and they were found to be highly correlated. With the limited num- ber of observations, the capacity of the model can not fit all significant risk factors, as the covariates and the coefficient estimates may attenuate due to collinearity, or change erratically in response to small changes in the model or the data. Therefore, to fit the final ecological model with a reasonable num- ber of covariates, some new methods may be needed to reduce the number of risk factors and handle the collinearity. Conducting Principal Component Analysis (PCA) can provide one way of tackling these problems. Within the significant risk factors in the univariate analysis, PCA was performed to reduce the number of risk factors and avoid multicollinearity by summarizing all significant factors within a few principal components. 64 Chapter 4. Statistical methodology PCA is a multivariate statistical technique used to reduce the dimen- sionality of a data set by retaining those characteristics of the data set that contribute most to its variance. Mathematically, PCA performs a linear transformation, which transforms the original correlated variables into un- correlated components. Each component is a linear weighted combination of the original variables. Given the set of variables X1, . . . , Xp (which are the significant risk factors from the univariate analysis), the principal com- ponents (PCs) are: PC1 = a11X1 + a12X2 + . . .+ a1pXp, · · · , PCk = ak1X1 + ak2X2 + . . .+ akpXp, where akp represents the weight (so-called loading) for the pth variable on the kth principal component. The loadings for the components are given by the eigenvectors of the correlation matrix or the co-variance matrix. The prin- cipal component (PC) usually represents the characteristics of the variables with large loadings. The uncorrelated properties of the components are high- lighted by the fact that they are orthogonal. The subsequent PC accounts for the additional proportion of the total variation in the original variables after the previous component, and the amount of variation explained get smaller and smaller. If the original variables are highly correlated, only a few PCs are required to capture the majority of the information in the original variables. In order to reduce the number of risk factors and avoid multicollinearity 65 Chapter 4. Statistical methodology among the variables, PCA was conducted for the set of significant risk factors identified by the univariate analysis. A few PCs were created to represent the major characteristics of all the significant risk factors and put into a final regression model as covariates. In general, PCs that cumulatively explained about 70% of the total covariates variance in the original variables were selected. Each component is dominated by several regional characteristics, which contribute high loadings to the component. Thus, the component becomes the indicator of these regional characteristics. To visualize the loadings and effects of the risk factors, the loadings with significantly large values can be highlighted in the figures, so that the characteristics of the component can be explored for each group (see the figure in the following example). 4.3 Ecological regression The final ecological models (4.2) were fitted using the FB-MCMC method, with selected PCs as covariates. IID priors (4.6) were applied on the random effects in the ecological models because the selected PCs, which summarized risk factors, had explained the majority of the spatial effects (examined in the following illustrative example). In addition, the Bayesian estimates of the unadjusted and risk-adjusted RRs were obtained for each LHA, which highlighted the high and low risk areas by accounting for those observed risk factors (i.e., covariates in the ecological models). For each age- and sex-specific outcome, the first PC was strongly sig- 66 Chapter 4. Statistical methodology nificant in the ecological model, which contributed to the majority of the reductions in spatial effects and relative risk variance. The second or third PC was significant in some groups. Even though the PCs were not all signif- icant in the final ecological models, we kept them to avoid losing too much information from the original covariates. According to the loadings for each PC, different risk factors were related to suicide rates in different age groups (See application in the next chapter). The risk-adjusted RRs, derived from the ecological models, reflected the residual variation that can be the unmeasured confounders, after adjustments for spatial dependency and known risk factors. These RRs reflected the smoothed estimates of the relative risks for suicide by borrowing information from neighboring regions and removing the known risk factors. In the maps of RRs, the dark areas represent high risk areas, while the light areas represent low risk areas. Generally, while high and low risk patterns of suicide risks differ among age groups, the geographical clusters decreased or disappeared when the regional characteristics were accounted for. The diminishment of high and low risk areas from the unadjusted RRs, as compared to the risk- adjusted area-specific RRs, indicates that variation in suicide risk rates can partly be explained by the adjusted risk factors and the involved spatial effects. This means that the risk-adjusted area-specific RRs can provide more stable and precise estimates of suicide risks after removing some known effects. 67 Chapter 4. Statistical methodology 4.4 Illustrative example To better understand the application of exploratory analysis and ecologi- cal regression, an illustrative example of the disease mapping and ecological analysis of the suicide hospitalization rates of males aged 15-24 is presented here. During the study period, out of a population of 2,569,175 males aged 15-24, there were 3,354 suicide hospitalizations. The crude suicide hospital- ization rate was about 130 per 100,000 persons and varied from 40 to 1,200 per 100,000 persons among LHAs. The corresponding crude RRs ranged from 0.3 to 9.3 across LHAs. Large variations were shown in the crude rates and RRs. Using the FB-MCMC method, the Poisson regression model without co- variates was first fitted on the data. Both significant spatial effects and relative risk variance were observed (λ̂ = 0.76, 95%CI: (0.42, 0.99); σ̂ = 0.65, 95%CI: (0.49, 0.84)). Next, using the EB-PQL method, a univariate analysis was carried out to fit the Poisson regression model with one census variable as a covariate each time. In total, 46 regional characteristics were found to be significant in the univariate analysis. Thirdly, a PCA was conducted for the 46 significant risk factors. Since they were highly correlated, the first 4 PCs can represent the majority of the information in the data, which explained 74% of the total covariates variance. Finally, using 4 PCs as covariates, the ecological model was fitted with a CAR prior and then with an IID prior by 68 Chapter 4. Statistical methodology Table 4.1: Estimates for ecological models, male aged 15-24 suicide hospital- ization CAR prior IID prior Estimate 95%CI Estimate 95%CI Comp.1 -0.35 ( -0.44, -0.26) -0.36 ( -0.44, -0.28) Comp.2 -0.08 ( -0.19, 0.03) -0.10 ( -0.17, -0.02) Comp.3 -0.18 ( -0.29, -0.07) -0.20 ( -0.29, -0.12) Comp.4 0.05 ( -0.04, 0.14) 0.05 ( -0.04, 0.14) Rate per 105 156 ( 130, 188) 154 ( 142, 166) λ 0.66 ( 0.20, 0.98) – σ 0.44 ( 0.30, 0.61) 0.29 (0.22, 0.37) DIC 558 562 Boldface: significant coefficient. –: λ is not available. the FB-MCMC method separately. The posterior means and 95% credible intervals of spatial effect, relative risk variance, and regression coefficients (here, these were principal components) are shown in Table 4.1. The posterior distribution of λ from the model with the CAR prior was quite flat, which indicated that the spatial effect may not be significant. The estimates of the coefficients were close between the model with CAR prior and the model with IID prior, except that the credible intervals were nar- rower in the ecological model with the IID prior. More significant coefficients were identified in the ecological model with the IID prior, since smaller stan- dard deviations were observed. The covariates seemed to have explained the majority of the spatial effects. Therefore, the IID prior was used for the ecological model and the following interpretations were based on the results of the ecological model with the IID prior. 69 Chapter 4. Statistical methodology The first, second, and third components were significant in the ecologi- cal model. The relative risk variance (measured by σ) decreased by 32.5%, compared with the estimates in the disease mapping model. The risk factors included in the ecological model have explained considerable geographical variation. The corresponding loadings for the significant components are shown in Figure 4.1. The variables with high loadings were drawn as dots. Dash lines separated the 11 domains and the names of the domains are la- beled. −0.2 −0.1 0.0 0.1 0.2 Loading for Comp.1 11. Dwellings 10. Family Status 9. Income 8. Transportation 7. Unemployment Rates 6. Education 5. Mobility Status 4. Aboriginal Population 3. Immigration 2. Language 1. Marital Status −0.3 −0.2 −0.1 0.0 0.1 0.2 0.3 Loading for Comp.2 11. Dwellings 10. Family Status 9. Income 8. Transportation 7. Unemployment Rates 6. Education 5. Mobility Status 4. Aboriginal Population 3. Immigration 2. Language 1. Marital Status −0.2 −0.1 0.0 0.1 0.2 0.3 Loading for Comp.3 11. Dwellings 10. Family Status 9. Income 8. Transportation 7. Unemployment Rates 6. Education 5. Mobility Status 4. Aboriginal Population 3. Immigration 2. Language 1. Marital Status Figure 4.1: Loadings of risk factors in the significant principal components, suicide hospitalization of males aged 15-24 The loadings in the top-left panel of Figure 4.1 show that the first com- 70 Chapter 4. Statistical methodology ponent was dominated by regional unemployment rates. The loadings in the top-middle panel show that the second component was dominated by the domains of family structure, education, dwellings, and marital status. The loadings in the top-right panel show that the third component represented risks from the domain of income. Note that the principal components rep- resent patterns of characteristics that are acting as complex regional risk factors for the higher suicide hospitalization rates for male youths. The risk patterns are related to high unemployment rates, a high percentage of popu- lation with low incomes, a high percentage of population with low education levels, a high percentage of population being divorced, a high percentage of population with less children in the household or being single parents, or a high percentage of the population renting their dwellings. The RRs for the 84 LHAs in B.C. can be observed in Figure 4.2. The maps display significance for both unadjusted RRs and risk-adjusted RRs. Comparing the patterns between the unadjusted and the risk-adjusted RRs, some significant patterns in the unadjusted RRs disappeared in the map of the risk-adjusted RRs, which indicates that considerable suicide risk vari- ations may be attributable to the risk factors considered in the ecological analysis. After risk adjustment, some LHAs still had high suicide hospital- ization rates. These were observed in the Vancouver Coastal region and the northern part of Vancouver Island. 71 Chapter 4. Statistical methodology Unadjusted Risk-adjusted Figure 4.2: Maps of unadjusted and risk-adjusted RRs for male aged 15-24 suicide hospitalization. (High LHAs in black (Pr(RR > 1) > 0.975); Low LHAs in gray ( Pr(RR < 1) > 0.975); No statisti- cally significance in blank) 72 Chapter 5 Regional variation of suicide rates and associated ecological characteristics This chapter reports on an in-depth analysis of suicide hospitalization and mortality data. We first report on regional variation and spatial patterns in suicide hospitalization and mortality rates. We then present a spatial and ecological analysis that explores and discusses the regional characteristics related to suicide occurrences. The following sections are organized accord- ing to three demographic groups, each reflecting gender-specific suicide risks among youth (aged 15-24), adults (aged 25-64) and the elderly (aged 65 or older). The analyses of suicide hospitalization and mortality rates were done for age- (aggregated as 10-year groups) and gender-specific groups, respec- tively. Overall, inclusion of the PCs explained the majority of the spatial effects that were observed in the Bayesian estimates of the unadjusted age- and gender-specific RRs and the IID prior was assumed for the risk-adjusted log RRs (i.e. residual random effects) in all ecological models. 73 Chapter 5. Regional variation of suicide rates and associated ecological characteristics 5.1 Suicide risks among youth (aged 15-24) In general, large geographical variation was observed in suicide hospitaliza- tion rates among youth, with the male groups showing large variations in suicide mortality rates. An investigation of regional characteristics was done to explain the variation in terms of area-specific suicide rates. Figure 5.1 shows the Bayesian estimates of the (unadjusted) RRs of sui- cide hospitalization and mortality for males and females respectively. Ge- ographical patterns were observed in both suicide hospitalization and mor- tality rates. For the male population, LHAs with a high risk for suicide hospitalization clustered in the Vancouver Coastal region, Northwest region, and Vancouver Island, while LHAs with a high risk for suicide mortality were observed in the western parts of the province and a few areas of the Interior region. For the female population, LHAs with a high risk for suicide hospital- ization appeared in the Northern Interior region, Vancouver Island, and the central parts of Interior region, whereas LHAs which a high risk for suicide mortality appeared in the Northern Interior region, the Interior region, and islands of the Northern region, and the Vancouver Coastal regions. Table 5.1 shows the summary from the exploratory analysis of youth suicide hospitalization and mortality rates. The numbers of significant risk factors from the univariate analysis and results of the PCA based on those risk factors are presented. For each outcome, the first 4 PCs explained 70% or more of the total covariates variance are included in the ecological regression 74 Chapter 5. Regional variation of suicide rates and associated ecological characteristics Male Hospitalization Female Hospitalization Male Death Female Death Figure 5.1: Significance of unadjusted RRs, youth aged 15-24 suicide hospi- talization and mortality. (High RRs in black (Pr(RR > 1) > 0.975); Low RRs in gray (Pr(RR < 1) > 0.975); No statistically significance in blank) 75 Chapter 5. Regional variation of suicide rates and associated ecological characteristics Table 5.1: Summary of univariate analysis and principal component analysis, youth aged 15-24 Subgroup Univariate analysis PCA Significant PCs (N) Explained factors (N) variance (%) Hospitalization Male 46 4 73.5 Female 49 4 72.6 Mortality Male 39 4 71.0 Female 20 4 76.5 of the suicide hospitalization and mortality rates. For each outcome, the significant PCs in the ecological regressions with large loadings provide the suicide risk patterns. The results are shown in Table 5.2. For suicide hospitalization rates, three and two significant PCs were iden- tified in the ecological regression for males and females respectively. They shared common risk factors in the domain of unemployment rates, marital status, income, and family structure. Risk factors related to education and dwellings were important for male suicide hospitalization rates, whereas risk factors from the domain of aboriginal population exerted the most significant relationship with female suicide hospitalization rates. For suicide mortality rates, fewer cases were observed in this age group. The first PCs were significant in the ecological regression for both male and female subjects. The common risk patterns contained factors from the do- mains of unemployment rates, aboriginal populations, and dwellings for both genders. Risk factors from the domains of education and income were ob- 76 Chapter 5. Regional variation of suicide rates and associated ecological characteristics Table 5.2: Risk patterns derived from ecological regression, youth aged 15-24 Hospitalization Mortality Male Female Male Female Marital status x x Language Immigrant Aboriginal population x x x Mobility Education attainment x x Unemployment rates x x x x Transportation Income x x x Family structure x x Dwellings x x x served to be significantly associated with male suicide mortality rates. The geographical patterns consistently disappeared after accounting for the regional risk factors for suicide hospitalization and mortality rates for both genders. Since none of LHAs showed significantly high RRs for suicide mortality, only the risk-adjusted RRs for suicide hospitalization are presented in Figure 5.2. A few LHAs around the Vancouver Coast and Vancouver Island still indicated high risks for male suicide hospitalization. Other LHAs that were observed to have high risks for female suicide hospitalization were in the western parts of the Northern Interior and the central parts of the Interior region. For the gender-specific suicide hospitalization and mortality outcomes, 77 Chapter 5. Regional variation of suicide rates and associated ecological characteristics the top 10 LHAs with the significantly highest relative risks (if any), based on ecological models with the IID priors, are presented in Appendix D. Male Hospitalization Female Hospitalization Figure 5.2: Significance of risk-adjusted RRs, youth aged 15-24 suicide hos- pitalization. (High RRs in black (Pr(RR > 1) > 0.975); Low RRs in gray (Pr(RR < 1) > 0.975); No statistically significance in blank) 5.2 Suicide risks among adults (aged 25-64) Figure 5.3 presents the Bayesian estimates of the unadjusted RRs of suicide hospitalization among populations aged 25-64. As age increased, fewer high and low risk LHAs were identified for male and female suicide hospitalization separately. For both male and female hospitalization, spatial clusters were observed in adult groups aged 25-34 and 35-44 separately, and variations were shown in working groups aged 45-54 and 55-64. Generally, high RRs 78 Chapter 5. Regional variation of suicide rates and associated ecological characteristics clustered in the western side of the Northern region, the Vancouver Coastal region, and the central areas of the Interior. More high risk areas were observed for females than for males. These appeared in the southwestern parts of the Northwest, the Vancouver Coastal region, Vancouver Island, and the central parts of the Interior region. Figure 5.4 shows the Bayesian estimates of the unadjusted RRs for suicide mortality among adults aged 25-64. While the spatial patterns for both males and females were different from those for suicide hospitalization, large distinctions between males and females were also observed. For the male population, noteworthy spatial patterns were observed in the 35-44 age group and considerable spatial variations were shown in the 45-54 and 55-64 age groups. Some high RRs were observed in the Northwest and central parts of the Interior region. For the female population, spatial patterns were observed only for females aged 25-34 and 35-44. Fewer high risk regions were identified for female suicide death than for males. High risk areas for female mortality were found throughout the Vancouver Downtown East Side and City Centre, and in the southwest of the Interior region. 79 C h a p ter 5 . R eg io n a l v a ria tio n o f su icid e ra tes a n d a sso cia ted eco lo g ica l ch a ra cteristics Male Hosp. 25-34 Male Hosp. 45-54 Female Hosp. 25-34 Female Hosp. 45-54 Male Hosp. 35-44 Female Hosp. 35-44 Male Hosp. 55-64 Female Hosp. 55-64 Figure 5.3: Significance of unadjusted RRs, adults aged 25-64 suicide hospitalization. (High RRs in black (Pr(RR > 1) > 0.975); Low RRs in gray (Pr(RR < 1) > 0.975); No statistically significance in blank) 80 C h a p ter 5 . R eg io n a l v a ria tio n o f su icid e ra tes a n d a sso cia ted eco lo g ica l ch a ra cteristics Male Death 25-34 Male Death 45-54 Female Death 25-34 Female Death 45-54 Male Death 35-44 Female Death 35-44 Male Death 55-64 Female Death 55-64 Figure 5.4: Significance of unadjusted RRs, aged 25-64 suicide mortality. (High RRs in black (Pr(RR > 1) > 0.975); Low RRs in gray (Pr(RR < 1) > 0.975); No statistically significance in blank) 81 Chapter 5. Regional variation of suicide rates and associated ecological characteristics Table 5.3 shows the summary from the exploratory analysis of adult sui- cide hospitalization and mortality rates, including the numbers of significant risk factors from the univariate analysis and the results of the PCA based on those risk factors. Among adults aged 25-64, more regional risk factors were identified for younger groups for both genders and 3-4 PCs explained more than 70% of the total covariates variance for all age- and gender-specific groups. The identified PCs were included in the ecological regressions of the sui- cide hospitalization and mortality rates. The significant PCs in the ecological regressions with large loadings provided the risk patterns for each outcome. The results are shown in Table 5.4 and Table 5.5. Adults aged 25-34 For suicide hospitalization rates, three and two PCs were significant in the ecological regressions for males and females separately. Both groups shared common risk characteristics from the domains of regional unemploy- ment rates, income, education attainment, and language. The domains of family structure and marital status were mainly related to male suicide hos- pitalization rates. For suicide mortality rates, more risk patterns were observed for males than females. Three and one PCs were statistically significant in the ecologi- cal regressions for males and females separately. Marital status, income, and dwellings were the common risk characteristics for both genders. Risk factors related to education attainment, unemployment rates, and immigrants were 82 Chapter 5. Regional variation of suicide rates and associated ecological characteristics Table 5.3: Summary of univariate analysis and principal component analysis, adults aged 25-64 Subgroup Univariate analysis PCA Significant PCs (N) Explained factors (N) variance (%) Young adult 25-34 Hospitalization Male 48 4 72.4 Female 47 4 71.2 Mortality Male 47 4 73.3 Female 5 2 74.2 Working group 35-44 Hospitalization Male 46 4 71.6 Female 50 4 72.0 Mortality Male 45 4 73.4 Female 29 4 76.1 Working group 45-54 Hospitalization Male 30 4 76.7 Female 36 4 77.1 Mortality Male 37 4 73.6 Female 18 3 73.0 Working group 55-64 Hospitalization Male 43 4 73.4 Female 30 4 73.9 Mortality Male 29 4 73.6 Female 33 4 74.6 more associated with male mortality, whereas family structure was associated with female suicide mortality rates. Adults aged 35-44 For both male and female suicide hospitalization, regional unemployment rates, income, and dwellings were the common risk patterns. Marital status and family structure were related to male suicide hospitalization rates and 83 Chapter 5. Regional variation of suicide rates and associated ecological characteristics Table 5.4: Risk patterns derived from ecological regression, adults aged 25-44 Adults aged 25-34 Hospitalization Mortality Male Female Male Female Marital status x x x Language x Immigrant x Aboriginal population x Mobility Education attainment x x x Unemployment rates x x x Transportation Income x x x x Family structure x x x Dwellings x x x x Adults aged 35-44 Hospitalization Mortality Male Female Male Female Marital status x x x Language Immigrant Aboriginal population x Mobility x Education attainment x x x Unemployment rates x x x x Transportation Income x x x x Family structure x x x Dwellings x x x 84 Chapter 5. Regional variation of suicide rates and associated ecological characteristics Table 5.5: Risk patterns derived from ecological regression, adults aged 45-64 Adults aged 45-54 Hospitalization Mortality Male Female Male Female Marital status x x x Language x x x Immigrant Aboriginal population x Mobility x Education attainment x x Unemployment rates x x x Transportation x Income x x x Family structure x x x Dwellings x x Adults aged 55-64 Hospitalization Mortality Male Female Male Female Marital status x x x Language Immigrant Aboriginal population x x x Mobility Education attainment x x x Unemployment rates x x x Transportation x Income x x x Family structure x x Dwellings x x 85 Chapter 5. Regional variation of suicide rates and associated ecological characteristics education attainment was significantly related to female rates. Two and one significant PCs were observed to provide the risk patterns for males and females respectively. For male and female suicide mortality, socioeconomic characteristics from the domains of marital status, education attainment, income, unemployment rates, and family structure were shared by both genders. While aboriginal population was related to suicide mortality rates, mobility status was a high- lighted risk domain for females. People tended to be more mobile at this age than others and the instability of this may raise the risks for suicide. Three significant PCs were revealed for suicide mortality risk patterns for each gender. Adults aged 45-54 The domains of education attainment, unemployment rates, income, and language were observed to be associated with suicide hospitalization for both genders. Between male and female hospitalization, fewer risk patterns were observed to be associated with female suicide hospitalization rates than with those of males. Transportation was one of the most distinct domains, which was only related to female populations. Risk factors from the domains of marital status, family structure, and mobility status often exerted a strong relationship with male suicide hospitalization rates. Three and one significant PCs were identified to explain the risk patterns for male and female subjects respectively. Marital status and family structure were the common domains that were 86 Chapter 5. Regional variation of suicide rates and associated ecological characteristics related to the suicide mortality rates of both genders in this age group. Risk factors from regional unemployment rates, income, and aboriginal population tended to be associated with male suicide mortality rates rather than female ones. Adults aged 55-64 Consistent risk patterns were found for both male and female suicide hospitalization. These were from the domains of aboriginal population, ed- ucation attainment, unemployment rates, and income. Marital status was also related to female suicide hospitalization. In both groups, only the first PCs were statistically significant in the ecological regressions. Family structure, marital status, and dwellings were the common pat- terns for male and female suicide mortality. Regional risk factors from the domains of aboriginal population, unemployment rates, and education at- tainment were only associated with male suicide mortality rates, while risk factors from the domain of transportation and income were related to female suicide mortality rates. Figure 5.5 presents the risk-adjusted suicide hospitalization RRs for males and females. Considerable spatial patterns were still found in males aged 25- 34, and females 25-34 and 35-44. The ecological risk factors for the adult groups only explained part of suicide hospitalization risk variation. Further studies need to be done to investigate other ecological risk factors associ- ated with suicide hospitalization risk variation. The majority of the LHAs in which RRs of male hospitalization became nonsignificant after risk adjust- 87 Chapter 5. Regional variation of suicide rates and associated ecological characteristics ment were in the Vancouver Coastal region, Vancouver Island, and Interior areas. Central Coast, Greater Victoria, and Terrace were usually the LHAs with a high risk after risk adjustment. The geographic patterns of the sui- cide hospitalization rates for females aged 25-64 remained more substantial than those for males after risk adjustments. The Western and Southern parts of the province retained relatively more LHAs with a high risk, which were not explained by the ecological risk factors. The high RRs spread around the Interior, Lower Mainland, and Victoria. Since all the significantly high rates disappeared after accounting for risks, the risk-adjusted RRs for suicide mortality rates were not presented. For each gender-specific suicide hospitalization and mortality outcome, the corresponding top 10 LHAs with the significantly highest relative risks, based on ecological models with the IID priors, are presented in Appendix D. 88 C h a p ter 5 . R eg io n a l v a ria tio n o f su icid e ra tes a n d a sso cia ted eco lo g ica l ch a ra cteristics Male Hosp. 25-34 Male Hosp. 45-54 Female Hosp. 25-34 Female Hosp. 45-54 Male Hosp. 35-44 Female Hosp. 35-44 Male Hosp. 55-64 Female Hosp. 55-64 Figure 5.5: Significance of risk-adjusted RRs, adults aged 25-64 suicide hospitalization.(High RRs in black (Pr(RR > 1) > 0.975); Low RRs in gray (Pr(RR < 1) > 0.975); No statistically significance in blank) 89 Chapter 5. Regional variation of suicide rates and associated ecological characteristics 5.3 Suicide risks among the elderly (aged 65 and over) Suicide hospitalization is an extremely rare event among the elderly, and moderate regional variation was only observed from the rates among those aged 65-74 (maps not shown). The LHAs with a high risk were found in and around the Lower Mainland. The northwest area of the province and the islands of the Northern region showed high suicide mortality rates for males aged 65-74, whereas high RRs were observed in the Southern Okanagan for males aged 75+. Some areas of the Greater Vancouver showed high suicide mortality rates for females aged 65-74. Table 5.6 shows the results from the exploratory analysis, which includes the number of significant risk factors from the univariate analysis and the summary of the PCA on those significant risk factors. Relatively small ge- ographical patterns and variations were observed in the disease mapping models for the elderly. There were fewer significant risk factors related to the elderly than to any other age groups. However, the number of risk factors identified for suicide mortality rates was greater than that for suicide hos- pitalization rates, a finding that was different from the other groups. There were no significant risk factors identified in the suicide hospitalization rates for females aged 75+. Based on the significant PCs in the ecological models, the risk patterns for suicide hospitalization and mortality are shown in Table 5.7. 90 Chapter 5. Regional variation of suicide rates and associated ecological characteristics Table 5.6: Summary of univariate analysis and principal component analysis, elderly aged 65+ Subgroup Univariate analysis PCA Significant PCs (N) Explained factors (N) variance (%) Elderly aged 65-74 Hospitalization Male 17 3 76.0 Female 5 2 79.4 Mortality Male 25 4 74.1 Female 25 3 73.0 Elderly aged 75+ Hospitalization Male 4 1 98.1 Female* 0 - – Mortality Male 7 1 75.2 Female 18 3 70.4 * : No significant risk factors were found for the subgroup. Elderly aged 65-74 For both male and female suicide hospitalization of elderly aged 65-74, family structure and marital status were common risk patterns in both gen- ders. Immigrants, mobility status, and income were related to male hospi- talization rates, whereas education attainment was related to females. As regards suicide mortality for both genders, income was the common domain. Regional unemployment rates, marital status, and dwellings were related to male suicide mortality rates, whereas transportation status, immi- grant status, and language were related to female rates. Elderly aged 75+ For the males, the first PC was significant for the suicide hospitalization 91 Chapter 5. Regional variation of suicide rates and associated ecological characteristics Table 5.7: Risk patterns derived from ecological regression, elderly aged 65+ Elderly aged 65-74 Hospitalization Mortality Male Female Male Female Marital status x x x Language x Immigrant x x Aboriginal population Mobility x Education attainment x Unemployment rates x Transportation x Income x x x Family structure x x x Dwellings x x Elderly aged 75+ Hospitalization Mortality Male Female Male Female Marital status x x Language Immigrant Aboriginal population x Mobility Education attainment x Unemployment rates Transportation x Income Family structure x x x Dwellings x 92 Chapter 5. Regional variation of suicide rates and associated ecological characteristics rates. For this group, it seemed that the only domain related to suicide hospitalization rates was family structure. For the female population aged 75+, as mentioned before, no significant risk factors were identified for the suicide hospitalization rates. Suicide mortality rates for both males and females were related to marital status and family structure. More domains were found to be associated with female mortality, such as aboriginal population, education attainment, transportation, and dwellings. Because the spatial patterns were completely removed after risk adjust- ment, the risk-adjusted RRs for both suicide hospitalization and mortality are not shown. Fewer geographical patterns were observed in the unadjusted RRs for elderly suicide hospitalization and mortality rates, which confirmed that an IID prior should be used in the disease mapping models. The top 10 LHAs with highest RRs for each group are presented in Appendix D. 93 Chapter 6 Discussion and future work The main project for this thesis was a comprehensive analysis of the associa- tion between suicide occurrences and the regional level characteristics, which links to a broad range of demographical, social, economic, and environmen- tal risk factors at the regional level. The statistical method used in the study provides a practical strategy for the labor intensive variable selection, identification of risk patterns, and statistical inference. The goal of this thesis is to explore the extent to which the population- level risk factors explain the variation of suicide rates and to highlight the high risk areas by using recently developed statistical methodologies, which incorporate an ecological analysis into Bayesian disease mapping study. The smoothed estimates of suicide hospitalization and mortality rates were achieved by borrowing information from neighboring regions. The study was based on administrative suicide data collected by the province of B.C., in order to provide an overall picture of the regional risk factors and their associations with suicide rates among B.C. populations at both the provincial and Local Health Area levels. Examining the suicide rates at provincial and regional levels enabled us to track important geographical clusters and highlight re- gional differences in the risk of suicide. 94 Chapter 6. Discussion and future work 6.1 Findings overview 6.1.1 Suicide rates In this study, suicide hospitalization rates were much higher than suicide mortality rates (usually 1.5 – 6.5 times in males, and 4.4 – 55 times in fe- males among different age groups), illustrating that most suicide behaviors ended up in hospital as severe attempts rather than as deaths. Moreover, large differences were found between genders in both suicide hospitaliza- tion and mortality rates. Female suicide hospitalization rates were 1.4 – 2.7 times those of males, which confirmed the gender differences in suicide be- haviors found in other studies (Iribarren et al, 2000; Suominen et al, 2004). The male suicide mortality rates were 3 – 4.3 times those of females, which demonstrated that males tend to engage in more serious suicidal behaviors (Kumar et al, 2006). In our study, the suicide hospitalization rates for pop- ulations aged 65+ were close between males and females, unlike those in the younger age groups, in which females had higher suicide hospitalization rates than males. Suicide mortality rates were higher among the elderly, compared with other groups. This was consistent with Heisel (2006)’s study, which found that elderly people killed themselves at a higher rate than other generations. 6.1.2 Geographical patterns and high risk areas Notable spatial variations were shown in the suicide hospitalization rates of males aged 15-24 and 25-34 respectively, females aged 15-24, 25-34, and 35- 95 Chapter 6. Discussion and future work 44, and also in the suicide mortality rates of males aged 15-24 and 25-34, and females aged 75+. The considerable RRs variations that were observed in the suicide hospitalization and mortality rates were explained by the regional characteristics in the ecological models. Generally, for suicide hospitalization, there were more LHAs with sig- nificantly high RRs than for suicide mortality. Remarkably, after adjusting for the risk factors, the number of significantly high risk LHAs decreased for suicide hospitalization. The disappearance of most of the geographical patterns not only suggests that the ecological risk factors were related to suicide hospitalization rates, but also explains the considerable variations in suicide rates. As age increased, the remaining significant patterns tended to get fewer and disappear. High suicide risk-adjusted RRs were observed in the Vancouver Coastal area, the central region of the Vancouver Island, and the southwest of the Northern region for males; for females, they were the Vancouver Coastal region, the southern and southwestern parts of the Northern Interior, the central area of the Interior, and the northern areas of Vancouver Island. The Center Coast and Greater Victoria were LHAs with a consistently high risk for male and female suicide hospitalization rates. The regional characteristics observed in those LHAs were usually high divorce rates, high percentage of aboriginal population, high percentage of population with low level education, high unemployment rates, or high percentage of population with a dwelling under current construction, all of which may be related to 96 Chapter 6. Discussion and future work the high suicide risks in these areas. Although Greater Victoria had a low percentage of aboriginal population, a low percentage of the population with low education levels, and low regional unemployment rates, it had high risks on suicide rates as well. The high suicide risks there may be associated with the high percentages of persons living alone in the household or of families without children. The low risk areas for suicide hospitalization were the southern and western parts of the Vancouver area, which may have to do with the facts that these two parts of the city share low rates of unemployment, high percentages of the population with high incomes, and high education levels. For suicide mortality, without risk adjustment, the study also showed sig- nificantly high area-specific RRs. Apart from males aged 25-34, the patterns for male suicide mortality rates consistently decreased with age. Most of the LHAs with high RRs were in the northwestern parts of the Northern region and a few areas in the middle part of the Interior. The high RRs disappeared when we adjusted for the regional risk factors. The high risk patterns were rarer in females, since fewer suicide deaths were observed among females. Some LHAs were low risk for female suicide mortality rates, such as the northeastern part of the Northern region and the Vancouver Coastal region. High risk areas clustered in the Vancouver Coastal region and the central and southeastern parts of the Interior region. The Vancouver Downtown East Side, the City Centre, the Stikine, and the Queen Charlotte were high risk LHAs for suicide mortality. The Vancouver 97 Chapter 6. Discussion and future work Downtown East Side and the City Centre shared many regional character- istics, such as a high percentage of the population being divorced, a high percentage of the population having immigrated from Asia, a high percent- age of the population living alone, and a high percentage of the population having lower incomes, and/or spending more money on living places in the ar- eas, all of which may be associated with high rates of suicide mortality. Even though low-level educated populations gathered at the Vancouver Downtown East Side while high-level educated ones gathered at the City Centre, both education levels were related to suicide mortality rates in those areas. Areas with a high percentage of the population being divorced or aboriginal, along with a low percentage of the population having immigrated from Asia or having immigrated between 1961 and 1970 were associated with high suicide rates in the Stikine, whereas areas with a high percentage of the population being divorced or aboriginal, along with a low percentage of the population paying rent, were the regional risk characteristics for the Queen Charlotte. 6.1.3 Risk patterns for different outcomes While various risk factors were shown to be associated with gender-specific suicide hospitalization and mortality rates of diverse age groups, their asso- ciations were mostly consistent across the age groups. More risk factors were identified with respect to suicide hospitalization than mortality rates, espe- cially for the female population. Across age groups, more risk factors were observed in younger groups than older ones. For both males and females, 98 Chapter 6. Discussion and future work fewer risk factors were identified for the senior and elderly groups than for the younger groups. For male suicide hospitalization rates, risk factors varied by age group. The common risk factors across age groups were from the domains of unem- ployment rates, income, education attainment, marital status, family struc- ture, and dwellings. However, for the elderly, unemployment rates and educa- tion attainment were not significantly associated with suicide hospitalization. Besides the common risk patterns for adults, risk factors related to language may be associated with suicide hospitalization in adult groups. For female suicide hospitalization rates, fewer shared risk patterns were identified. Only regional unemployment rates and income were observed in most age groups, except for the elderly. Regional family structure and marital status were important for youth and the elderly. Education attainment and dwellings were the most significant risk patterns for mature groups (both adults and the elderly). Aboriginal population was related to youth suicide hospitalization. For male suicide mortality rates, regional marital status, unemployment rates, income, education attainment, and dwellings were the important com- mon risk domains across most age groups. Marital status related to suicide rates in all age groups, except for the youth (15-24). Aboriginal population and education achievements were more relevant for the youth and adults rather than for the elderly. Family structure tended to be associated with adults rather than the youth. Regional characteristics related to immigrant 99 Chapter 6. Discussion and future work status were associated with suicide mortality rates in some adult groups. For female suicide mortality rates, marital status, dwellings, and family structure were associated with suicide mortality in most age groups. How- ever, aboriginal population only related to female mortality rates of the youth and the elderly. Regional transportation situations related to suicide mortal- ity rates for adults and the elderly; and risks from the domains of income and unemployment rates were associated with suicide mortality rates for youth and adults. 6.1.4 Risk patterns in demographic groups For both male and female youth, social and economic characteristics related to unemployment rates, income, education attainment, marital status, family structure, and dwellings were highlighted as risk patterns for both suicide hos- pitalization and mortality. It has been suggested that this could be because children and youth are largely supported and influenced by their families and surrounding social environments (Agerbo et al, 2002; Beautrais, 2000; Steele and Doey, 2007). In addition, suicide was observed to be more of a risk in the aboriginal population, for which risk patterns were also significant, particularly in youth groups (Hunter and Harvey, 2002; Hutchinson, 2005). For adult groups, the risk patterns were different between males and females for both suicide hospitalization and mortality. Social and economic characteristics related to unemployment rates, income, education attainment, marital status, family structure, and dwellings were the identified risk pat- 100 Chapter 6. Discussion and future work terns for male suicide hospitalization and mortality. Those risk factors have also been found in previous individual and ecological studies (Faria et al, 2006; Middleton et al, 2004; Qin et al, 2003; Cubbin et al, 2000; Hassel- back et al, 1991). Moreover, language showed another risk pattern in male hospitalization rates, while a high percentage of the population being immi- grants and being aboriginal in an area were two more risk characteristics for male mortality. Female suicide hospitalization was largely related to socioe- conomic characteristics, such as unemployment rates, income, and education attainment, while suicide mortality for the same gender was more related to marital status, family structure, dwellings, and transportation status. In our study, the suicide hospitalization rates for populations aged 65+ were close between males and females, unlike those in the younger subgroups, in which females had higher suicide hospitalization rates than males. The main risk patterns were associated with marital status, family structure, and dwellings for both suicide hospitalization and mortality (Conwell and Duberstein, 2001; McIntosh, 1995; Rubenowitz et al, 2001). Socioeconomic factors, such as income, unemployment rates, and education attainment, were not as important for the elderly as for other age groups. There are other factors related to suicide in the elderly, for example, identified illness, depression, or social isolation (Heisel, 2006; Rubenowitz et al, 2001), but our data didn’t cover such information. 101 Chapter 6. Discussion and future work 6.1.5 Significant risk patterns A variety of regional characteristics were observed to be associated with suicide hospitalization and mortality rates for both males and females in dif- ferent age groups. Most ecological associations were found to be consistent with the associations at the individual level. The major regional risk patterns were social and economic characteristics, including unemployment rates, in- come, education attainment, marital status, family structure, and dwellings. The regional risk patterns for some groups were also connected to aboriginal population, immigrants, and language. Unemployment Having a high unemployment rate in a region may decrease the level of social integration in that area, and thus contributed to high suicide rates (Ag- bayewa et al, 1998; Ferrada-Noli, 1997). The positive relationship between suicide rates and unemployment rates indicated by many previous individ- ual and ecological studies was confirmed by our study (Åhs and Westerling, 2006; Agerbo et al, 2007a; Saunderson and Langford, 1996; Chotai, 2005). Our study also showed that regions with high unemployment rates tended to have high suicide hospitalization and mortality rates, with the exception of female suicide mortality, in which that the relationships were not quite as strong. While high regional unemployment rates were significantly asso- ciated with high suicide hospitalization rates for young adults and working groups, this characteristic had less association in the elderly. The association between suicide and unemployment rates may be attenuated by accounting 102 Chapter 6. Discussion and future work for mental illness (Agerbo et al, 2002; Kumar et al, 2006; Fergusson et al, 2007), but this did not fall under the purview of the present study, as the focus is more on regional and ecological data. Income The factors related to regional income levels were highly associated with suicide hospitalization and mortality rates for both genders in most age groups. Regions with a higher percentage of the population with low incomes tended to have higher suicide rates, and regions with a high percentage of the population with high incomes tended to have low suicide rates, findings that are consistent with previous ecological studies (Abel and Kruger, 2005; Hutchinson, 2005). In particular, female mortality is less associated with low income than male, which may suggest that men react more strongly to poor economic conditions than women do – again, this is consistent with previ- ous individual studies (Agerbo et al, 2002; Qin et al, 2003; Kposowa, 2000). Moreover, low income was highly correlated with unemployment rates, and they often interacted to affect suicide rates at both the individual and pop- ulation level (Faria et al, 2006; Newman and Stuart, 2005). Education Previous studies showed that having difficulties in school or having par- ents with low educational attainment posed significant risks for complete suicide and serious suicide attempts (Iribarren et al, 2000; Kaneko and Mo- tohashi, 2007; Gould et al, 2003; Beautrais, 2003), a finding similar to what was observed in our study; that is, regions with high percentages of the pop- 103 Chapter 6. Discussion and future work ulation with high education levels were shown to have low suicide rates. High level educational attainment has been identified as a protective factor against suicide in many ecological studies (Abel and Kruger, 2005; Faria et al, 2006; Ferrada-Noli, 1997; Kalediene et al, 2006). In our study, the associations were confirmed, except that the relationships were not quite as strong in the elderly and female suicide mortality. Marital status In our study, regions with high percentages of the population separated or divorced were shown to have high suicide hospitalization and mortality rates for both genders, except in female youth mortality. Higher rates of being unmarried, widowed, or divorced/separated were associated with a higher suicide rate for males rather than for females in our study, a finding that was consistent with literature at the individual level; that is, the increased risks of suicide were observed mainly among divorced and separated men rather than women (Kposowa, 2000; Qin et al, 2003), and at the population level, the association between suicide rates and divorce rates was found only in males Faria et al (2006). Elderly people who committed suicide were being significantly more likely to be divorced or widowed in our study, which suggested that reducing social isolation may be a useful intervention to reduce the risk of suicide among the elderly (McIntosh, 1995; Rubenowitz et al, 2001). Family structure On the population level, both suicide hospitalization and mortality rates 104 Chapter 6. Discussion and future work were related to family structure for male and female subjects. In our study, regions with high percentages of the population being single parents or with fewer children in the households tended to have high rates of suicide for the youth and adult groups. This association indicated that high levels of family cohesion in an area may reduce its suicide risks, a finding that agreed with both individual (Qin et al, 2003) and population studies (Faria et al, 2006). Regions with a high percentage of single parents may indicate a high percentage of families’ breakdowns or a lack of family cohesion in the region, which may be related to high suicide rates (Middleton et al, 2004). Moreover, regions with a high the proportion of people living alone, especially elderly peopel in our study, were found to have high suicide rates. Similar results were also found in previous studies (Agerbo et al, 2007b). Dwelling status With the exception of female hospitalization, dwelling status was related to suicide hospitalization and mortality rates for most groups. Regions with a high percentage of the population renting their dwellings tended to have a high suicide rate in that area, while regions with a high percentage of the population owning their own houses or dwellings tended to have a low suicide mortality rate in that area. These findings confirmed the results of Ferrada-Noli (1997)’s study, which found that areas with a higher proportion of suicides had an increased proportion of the population who did not own their own home. 105 Chapter 6. Discussion and future work Aboriginal population In our study, regions with a high percentage of the population being aboriginal were more related to high suicide mortality rates than hospitaliza- tion rates for males, and high suicide hospitalization and mortality rates for youth than for older groups. Our study found that regions with a high per- centage of aboriginal population tended to have high suicide mortality rates among male youths and high hospitalization and mortality rates among fe- male youths. Previous ecological studies also found that a large proportion of the population being aboriginal was positively associated with high sui- cide rates, especially in youth (Hasselback et al, 1991; Lester, 1996). In the studies of Chandler and Lalonde (1998) and Chandler et al (2004), among British Columbia’s aboriginal population, they found that there was an ex- treme variability in the suicide rates among various aboriginal communities. In our study, the North American Indian was the major part of the abo- riginal population, so the association with suicide rates was mainly related to North American Indian population. We did not gather enough data to identify whether a high percentage of the population being other aboriginals (e.g., Métis and Inuit) was related to suicide rates. Therefore, further studies need to be done to examine whether or not different aboriginal communities share the same high risk for suicide equally. Immigrants and language The relationships between immigrants and suicide have bee studied in the past, the results are often mix. Some ecological studies have found that 106 Chapter 6. Discussion and future work region with high percentages of immigrants tended to have a high rate of suicide (Hasselback et al, 1991; Stack, 1981), and the others found no as- sociation between suicide and immigrants (Agbayewa et al, 1998; Åhs and Westerling, 2006). In our study, factors related to immigrant population were associated with suicide mortality rates among male adults, particularly, region with a high percentage of population being early immigrants tended to have high suicide mortality rates for males. Although a longer duration of residence may make it easier for them to integrate into society in terms of habit, culture, and language, most immigrants still face cultural contra- dictions between their native and adopted countries of residence and mature adults with experience and good adaptability skills may be able to adjust to the changes (Stack, 1981). Population speak non-official language may have barrier to access social services, health care services, and networking, especially in immigrants. Our study found that factors about language were related to suicide hospitaliza- tion of male adults. 6.2 Consideration and implication A major goal of an epidemiological study is to identify subgroups in the pop- ulation who are at high risk for a disease, so that we can direct preventive efforts to the members of population most likely to benefit from changes in programs and continue surveillance of the disease in the populations (Gordis, 2000). In order to prevent suicide, it is important to get the most up-to-date 107 Chapter 6. Discussion and future work and relevant information to reflect the reality of the situation and illustrate the demographics, associated geographical and environmental features of sui- cide. Our study is carried out to identify the high risk areas of suicide oc- currences and study the associated population level risks of suicide, so that we can use the observed results to inform policy initiatives and programs for suicide prevention. Our hypothesis that suicide rates are not the same throughout the study region is a first step to understanding the geographic nature of suicide. The specific demographic, social, or economic factors that are related to increase the risk of suicide are explored next. The present study attempted to ex- amine the hypothesis and achieve a goal of epidemiology, which identified subgroups in the population who were at high risk for suicide, with a novel statistical methodology in disease mapping and ecological study. Such eco- logical analysis incorporated into disease mapping study can also occur in many other researches besides suicide. For example, the severity of a certain kind of air pollution may affect people in neighboring regions similarly, be- cause the pollutant materials spread in the air over a large area that may cover several nearby regions. Researchers not only need to explore the risk factors related to the disease, but they also need to pay closer attention to the adjustment of spatial correlations that arise in ecological studies. Disease mapping is highly suitable for analyzing epidemiological data, as it provides a useful visual summary of complex geographic information and may identify subtle patterns in the data that may be missed in the tabular 108 Chapter 6. Discussion and future work presentations (Elliott and Wartenberg, 2004). In addition, disease mapping allows us to visualize problems in relation to existing health and social ser- vices and the environment. Disease mapping also functions effectively of bringing epidemiologists’ attention to the geographical patterns of those dis- eases (Lawson and Williams, 2001). When epidemiological studies of disease rates are pursued on a global basis, dramatic variation in the incidence of many diseases are observed. For example, Hodgkin’s disease in Sardinia dur- ing 1983-1985 showed about a 9-fold difference in mortality between two areas (Bernardinelli and Montomoli, 1992). The results supported the hypothesis that specific environmental, social, and economic risk factors may cause the variation in disease rates. The variation in suicide rates was also observed in many studies (Exeter and Boyle, 2007; Saunderson and Langford, 1996). Linking knowledge about the geographic characteristics of suicide is a nec- essary component of effective prevention programs. Knowing where risk is high allows the intervention and prevention to be directed to the right places. The risk patterns of suicide hospitalization and mortality rates varied across age and gender. Although several common risk patterns were observed for most groups, specific risk patterns demonstrated the population charac- teristics with high suicide rates for each age- and gender-specific group. For example, in young groups, economic characteristics such as income and un- employment were highly associated with suicide rates, but in elderly groups, family structure and marital status may be more related to the high rates of suicide. Such information can be used to initialize the focused prevention 109 Chapter 6. Discussion and future work programs to target the different situations in the populations. Regional characteristics described the distribution of the risk factor for the population that lived in that area. In our study, those regional risk factors were derived by aggregating individual data, and the ecological associations between the aggregated variables and suicide were observed consistent with the associations of those factors at the individual level. Thus, based on the present ecological study, the identified ecological risk patterns may provide information for further targeted individual studies. If suicide rate in a certain area was high when the regional characteristics were significant, population with such characteristics may potentially at high risk of suicide. For example, if unemployment rates were found to be associated with high suicide rate in a certain area, unemployed population in that area may have high risks of suicide. It doesn’t necessarily mean unemployed individuals tended to suicide, although sometimes they do. Further study can be done to examine the hypothesis. This is why studies at the individual level is essential in order to establish the actual associations between suicide and risk factors. Nevertheless, the results in our ecological studies can potentially point out the target populations at the most risk, which may help in finding the suicide- prone members of a population and conducting effective prevention programs to help them. If a risk factor is identified in the high risk area, people who work in suicide intervention and prevention should give more attention to the high risk populations. Even if it is difficult to truly improve the conditions that 110 Chapter 6. Discussion and future work put one community at a higher risk than another, the study can at least make people more aware of the regional characteristics that heighten the chances of suicide and address them. For example, a high percentage of the population with low education levels was the risk factor for suicide rates among younger population, a prevention program can focus on improving the education levels of young adults. Although it might take time and effort to make progress in this area, people will at least be aware that they need to pay more attention to educating the younger members of their population. The highlighted high risk areas with their associated risk factors from regional characteristics support our hypothesis that suicide rates are not the same throughout the study region and the information revealed may have implications for future prevention programs. By identifying high risk areas, community-specific programs may be carried out to link to the education and prevention strategies. It is particularly beneficial for individual studies to identify regions with high suicide rates. Looking at the high risk areas and overall picture of suicides found in our study may help the health authorities to find the appropriate place to start target studies. Information can also be used to learn from areas with lower incidence while directing increased attention to areas with higher incidence. We believe that these considerations are important components of any suicide prevention strategy, and argue that most injury prevention could benefit directly from the mapping of risks. In general, this kind of ecological study not only helps to determine the extent of the suicide problem and find regional factors associated with suicide, 111 Chapter 6. Discussion and future work but the results also inform public health policy-making and raise awareness among politicians, policy makers, and the public at large. Since the study was conducted at the population level, the findings regarding the burden and impact of suicide may initialize prevention programs, inform future decisions regarding the allocation of resources for suicide prevention, and guide the planning of future individual studies. Implication of models and analytic strategy In order to identify the regional characteristics associated with the risk patterns of suicide hospitalization and mortality rates for different genders, ages, and areas, a Poisson regression with random effects was proposed and inferences were made via Bayesian approaches. The random effects quantified the residual and unmeasured confounding, and the spatial dependency in the data. Bayesian approaches enable us to make inference on the spatial correlated data and the advantages of both the EB and the FB methods were highlighted in the process of analysis because the weaknesses of one were made up for by the strengths of the other, and vice versa. The process of exploratory analysis provided a strategy to explore the risk factors among numerous candidate factors and tackle the collinearity among the variables. By accounting for regional risks and spatial correlations in the small-area data, we were able to observe the more reliable and smoothed estimates of the suicide rates. The highlighted high and low risk areas may help to inform the development of more efficient and effective future prevention programs. Incorporating the PCA was especially effective for dealing with various 112 Chapter 6. Discussion and future work ecological effects. Since there is a potential for multicollinearity among the effects and high dimension of data involved, applying PCA can avoid the multicollinearity among the risk factors and reduce the dimension of the data. The PCA also helped us explore the risk patterns of the outcome by summarizing the characteristics of the principal components. The significant component usually represents risk patterns from multiple domains and the domains were different among age- and gender-specific groups, which sug- gested that suicide can not be explained by any single risk factor and the associations between suicide and ecological effects were varied from age and gender. Although it is hard to draw causal conclusion on suicide, the identify- ing possible risk patterns provide a better direction for further understanding suicide risks, especially at the ecological level. The univariate analysis for most subgroups showed that more than twenty risk factors were significant, but the first up to four components from the PCA usually explained at least 70% of the total covariates variance. The data were compressed to provide risk patterns of suicide rates without losing much information. 6.3 Potential limitations and implications for future work The findings in this study demonstrated some potential relationships be- tween suicide and regional characteristics. The regional risk factors tended to vary across age, gender, and area. More focused studies need to be done to gain a more in-depth understanding of how certain risk factors affect suicide 113 Chapter 6. Discussion and future work hospitalization or mortality rate for some particular groups. For example, the domain of unemployment included several variables, and each variable expressed different characteristic relevant to unemployment rate. Although unemployment rates were identified to be associated with suicide rates in most groups, different variable may be significantly related to suicide rates in different groups. Those variables may delineate the compositional effect on suicide, or represent the contextual effect on suicide, which need further ex- ploration. Other focus studies could be done, as well, such as studying what kind of family structure is positively or negatively associated with suicide, which regional characteristics of immigrants are contributed to high male suicide mortality rates, what characteristics of the aboriginal population are relate to suicide mortality rates, and so on. Each study could achieve differ- ent goals in terms of finding the effects of regional characteristics on suicide rates. The findings in the study also showed that suicide hospitalization and mortality rates had some similar patterns in risk factors, thus they may po- tentially share some common risk factors and geographical patterns. The shared component model may be used to conduct a joint analysis of suicide hospitalization and mortality rates. This model enabled us to explore the common patterns by facilitating “borrowing strength” between two disease outcomes and to achieve greater efficiency in the disease-specific risk pre- dictions (Knorr-Held and Best, 2001). In particular, for suicide mortality (i.e., extreme rare event) shared component model can borrow information 114 Chapter 6. Discussion and future work from the other outcome, (i.e., suicide hospitalization) in order to potentially improve the suicide mortality risk estimation. Future study can focus on employing the shared component model to explore the geographical patterns and associated risk factors with respected to suicide attempts and comple- tions simultaneously. The present study was an ecological study. Our results have clearly il- lustrated the importance of regional risk factors in explaining variation in suicide hospitalization and mortality rates, after accounting for the spatial dependency. We emphasized the impacts on a population of the risk fac- tors that can precipitate suicide so that a public health agency can use the findings to initialize future suicide prevention programs. While studies of the associations between suicide and various demographical, social, and economic factors at aggregate-levels do not necessarily provide proxies for individual studies, they can potentially provide important information regarding to con- textual effects to suicide and may reflect individual associations. Moreover, other studies also found that regional characteristics may still be associated with suicide rates after accounting for individual factors (Cubbin et al, 2000; Agerbo et al, 2007b). Therefore, it is possible that the differences in pop- ulation composition or other contextual factors operating at an area level, such as SES, or other macro-level features, jointly produce health or social inequalities across space which are related to suicide rates. To better under- stand suicide rates in the population, a multilevel study could be conducted in the future to explore the risk factors that contribute to suicide at both the 115 Chapter 6. Discussion and future work individual and regional levels. To perform a multilevel study, we need data of all risk factors available at both individual and regional levels and for both cases and population at risk, but our current data only covered part of infor- mation at the population level. If multilevel studies need to be conducted, more information needed to be provided. The fact that we collected the number of suicide hospitalization cases due to suicide attempts from data sets covering the entire population in the province during the study period meant that our study reflected the popula- tion characteristics rather than sample estimates. Therefore, the results may be more applicable than those based on non-representative samples. The study period lasted 10 years so that the aggregated data provide a sufficient number of events for us to study the ecological relationship between suicide and regional characteristics in different age groups. However, suicide is a rare event in the general population and we still couldn’t obtain enough in- formation to explore the risk factors for suicide hospitalization and mortality at the regional level in some age groups such as children groups aged 0-14. On the other hand, since the study period was across 10 years, if more data were available, it would be better to concern the temporal components in the spatial and ecological analysis to eliminate some potential known variation. Many people who attempt suicide never come to medical attention or even deny that their injuries are the result of a suicide attempt, and thus it is hard to collect data for measuring suicide attempts completely. In our study, suicide hospitalization was used to measure severe suicide attempts. 116 Chapter 6. Discussion and future work As we known, people who attempt suicide and result in hospital admissions usually make severe attempts. Moreover, studies showed that 30% to 50% of all attempts receive medical care (Kumar et al, 2006; Cantor and Neulinger, 2000). Therefore, suicide hospitalization for severe attempts may reflect the situation of suicide attempt in some degree. Although the data can not cap- ture all suicide attempts, our findings obtained from hospitalization setting may help to identify some problem related to severe suicide attempts. The regional characteristics were collected from census data, which were based on 20% samples of population. The variables in the ecological study may involve measurement errors in representing the regional characteristics for the entire population. Since our ecological study is to explore the associ- ations between suicide rates and regional characteristics in order to highlight high risk areas and identify potential risk factors, there was no causal associ- ations identified. With the findings in our study, further focused studies with well measured data can be done to examine the causal associations between suicide and regional risk factors. Currently, ecological studies use administrative data is an increasingly common practice, because the data carry the advantage of covering large populations over long periods of time and the collection is cheaper than that for primary data. However, the fact that these data were collected for purposes other than research, not only means that many cases might have been under-reported but also creates some limitations with respect to measurement error. The reliability of the data also relies upon the disease 117 Chapter 6. Discussion and future work coding. One concern about the administrative data is the reliability and/or validity of the ICD9 E-codes in the hospital records and death certificates in B.C. The quality of the data can directly influence the inference regarding suicide rates in B.C. In some areas that reported only a few cases, the absence or addition of even one case may change the entire inference about that region. Therefore, the quality of the data needs to be further explored. One solution for this could be that additional data from other sources could be used to compare with the administrative data, for example, selected medical charts could be reviewed and compared with the hospital records. 118 Bibliography Abel EL, Kruger ML. Educational attainment and suicide rates in the United States. Psychol Rep. 2005; 97(1): 25-8. Agbayewa MO, Marion SA, Wiggins S. Socioeconomic factors associated with suicide in elderly populations in British Columbia: an 11-year review. Can J Psychiatry. 1998; 43(8): 829-36. Agerbo E, Nordentoft M, Mortensen PB. Familial, psychiatric and socioe- conomic risk factors for suicide in young people. BMJ. 2002; 325(7355): 74-77. Agerbo E, Gunnell D, Bonde JP, Bo Mortensen P, Nordentoft M. Suicide and occupation: the impact of socio-economic, demographic and psychiatric differences. Psychol Med. 2007; 37(8): 1131-40. Agerbo E, Sterne JA, Gunnell DJ . Combining individual and ecological data to determine compositional and contextual socio-economic risk factors for suicide. Soc Sci Med. 2007; 64(2): 451-61. Åhs AM, Westerling R. Mortality in relation to employment status during different levels of unemployment. Scand J Public Health. 2006; 34(2): 159- 67. Ainsworth LM, Dean CB. Approximate inference for disease mapping. Com- putational Statistics & Data Analysis. 2006; 50: 2552-2570. Aleck O, Stefania M, James T, James D, Ruth H, Lisa C, Amber L, Clyde H. The impact of fathers’ physical and psychosocial work conditions on 119 Bibliography attempted and completed suicide among their children. BMC Public Health. 2006; 6: 77. Bailey TC. Spatial statistical methods in health. Cad Saude Publica. 2001; 17(5): 1083-98. Beautrais AL, Joyce PR, Mulder RT, Fergusson DM, Deavoll BJ, Nigetingale SK. Prevalence and comorbidity of mental disorders in persons making serious suicide attempts: a case-control study. Am. J. Psychiatry. 1996; 153: 1009-14. Beautrais AL, Joyce PR, Mulder RT. Risk factors for serious suicide at- tempts among youths aged 13 through 24 years. J Am Acad Child Adolesc Psychiatry. 1996; 35(9): 1174-82. Beautrais AL. Risk factors for suicide and attempted suicide among young people. Aust N Z J Psychiatry. 2000; 34(3): 420-36. Beautrais AL. Suicide and serious suicide attempts in youth: a multiple- group comparison study. Am J Psychiatry. 2003; 160(6): 1093-9. Bernardinelli L, Montomoli C. Empirical Bayes versus fully Bayesian analy- sis of geographical variation in disease risk. Stat Med. 1992; 11(8): 983-1007. Bertolote JM , Fleischmann A, Butchart A,Besbelli N. Suicide, suicide at- tempts and pesticides: a major hidden public health problem. Bulletin of the World Health Organization. 2006; 84(4): 260-1. Besag J. Spatial interaction and the statistical analysis of lattice systems (with discussions). Journal of the Royal Statistical Society, Series B. 1974; 36: 192-236. Besag J, York J, Mollié A. Bayesian image restoration, with two applications in spatial statistics. Annals of the Institute of Statistical Mathematics. 1991; 43: 1-21. 120 Bibliography Besag J, Kooperberg C. On conditional and intrinsic autoregression. Biometrika. 1995; 82: 733-476. Blakely TA, Collings SC, Atkinson J. Unemployment and suicide. Evidence for a causal association? J Epidemiol Community Health. 2003; 57(8): 594- 600. Brent DA, Melhem N. Familial transmission of suicidal behavior. Psychiatr Clin North Am. 2008; 31(2): 157-77. Breslow NE, Day NE. Statistical methods in cancer research. Volume II–The design and analysis of cohort studies. Oxford University Press; 1987. Breslow NE, Clayton DG. Approximate inference in generalized linear mixed models. JASA. 1993; 88: 9-25. British Columbia Health Authorities. 2007. http://www.healthservices.gov. bc.ca/socsec/roles.html British Columbia Vital Statistics. Annual reports. Victoria, BC: Ministry of Health Planning. 2002. British Columbia Vital Statistics. Annual reports. Victoria, BC: Ministry of Health Planning. 2004. Burrows S, Vaez M, Laflamme L. Sex-specific suicide mortality in the South African urban context: the role of age, race, and geographical location. Scand J Public Health. 2007; 35(2): 133-9. Cantor CH, Slater PJ. Marital Breakdown, Parenthood, and Suicide. Jour- nal of Family Studies. 1995; 1: 91-102. Cantor CH, Baume PJ. Suicide prevention: a public health approach.Aust N Z J Ment Health Nurs. 1999; 8(2): 45-50. 121 Bibliography Cantor C, Neulinger K. The epidemiology of suicide and attempted suicide among young Australians. Aust N Z J Psychiatry. 2000; 34(3): 370-87. Carney S, Rich C, Burke P, Fowler RC. Suicide over 60: the San Diego Study. J Am Geriatr Soc. 1994; 42: 174-180. Carstairs V, Morris R. Deprivation and health in Scotland. Aberdeen Uni- versity Press; 1991. Chandler, MJ, Lalonde C. Cultural continuity as a hedge against suicide in Canada’s First Nations. Transcultural Psychiatry. 1998; 35(2): 191-219. Chandler, Michael J, Lalonde, Christopher E. Transferring whose knowl- edge? Exchanging whose Best Practice? On knowing about indigenous knowledge and aboriginal suicide. In white J, Maxim P and D Beavon eds. Aboriginal Policy Resaerch: Setting the agenda for change. Toronto: Thompson Educational Publishing. 2004; 2. Cheng ATA. Mental illness and suicide. Arch Gen Psychiatry. 1995; 52(7): 594-603 Chotai J. Suicide aggregation in relation to socio-demographic variables and the suicide method in a general population: assortative susceptibility. Nord J Psychiatry. 2005; 59(5): 325-30. Clayton D, Kaldor J. Empirical bayes estimates of age-standardized relative risks for use in disease mapping. Biometrics. 1987; 43(3): 671-81. Clayton DG, Bernardinelli L, Montomoli C. Spatial correlation in ecological analysis. Int J Epidemiol. 1993; 22: 1193-202 Congdon P. The epidemiology of suicide in London. J. Royal. Stat.Soc, Ser A. 1996; 159: 515-533. Congdon P. Suicide and Parasuicide in London: A Small-area Study. Urban Studies. 1996; 33(1): 137-158. 122 Bibliography Congdon, P. Monitoring suicide mortality: a Bayesian approach. European Journal of Population. 2000; 16: 251C284. Conwell Y, Duberstein PR. Suicide in Elders. Annals of the New York Academy of Sciences. 2001; 932: 132-150. Cortinovis I, Vela V, Ndiku J. Construction of a socio-economic index to facilitate analysis of health in data in developing countries. Social Science and Medicine. 1993; 36: 1087C97. Cubbin C, LeClere FB, Smith GS. Socioeconomic status and injury mortal- ity: individual and neighbourhood determinants. J Epidemiol Community Health. 2000; 54(7): 517-24. Cutler DM, Glaser EL, Norberg KE. Explaining the rise in youth suicide. National Bureau of Economic Research Working Paper 7713. 2000. Elliott P, Wakefield JC, Best NG and Briggs DJ. Spatial Epidemiology: Methods and Applications. Oxford University Press, Oxford; 2000. Elliott P, Wartenberg D. Spatial epidemiology: current approaches and fu- ture challenges. Environ Health Perspect. 2004; 112: 998-1006. Exeter DJ, Boyle PJ. Does young adult suicide cluster geographically in Scotland? J Epidemiol Community Health. 2007; 61(8): 731-6. Faria NM, Victora CG, Meneghel SN, de Carvalho LA, Falk JW. Suicide rates in the State of Rio Grande do Sul, Brazil: association with socioeco- nomic, cultural, and agricultural factors. Cad Saude Publica. 2006; 22(12): 2611-21. Fergusson DM, Boden JM, Horwood LJ. Unemployment and suicidal be- haviour in a New Zealand birth cohort: A fixed effects regression analysis. CRISIS. 2007; 28(2): 95-101. 123 Bibliography Ferrada-Noli M. Health and socioeconomic indicators in psychiatric catch- ment areas with divergent suicide rates. Psychol Rep. 1997; 81(2): 611-9. Gordis L. Epidemiology. 2nd ed. Philadelphia: W.B. Saunders; 2000. Gould M, Greenberg T, Velting D, Shaffer D. Youth suicide risk and preven- tive interventions: A review of the past 10 years. Journal of the American Academy of Child and Adolescent Psychiatry. 2003; 42(4): 386-405. Gunnell, DJ, Peters, TJ, Kammerling, RM, Brooks, J. Relation between parasuicide, suicide, psychiatric admissions, and socioeconomic deprivation. BMJ. 1995; 311: 226-230. Hasselback P, Lee KI, Mao Y, Nichol R, Wigle DT. The relationship of suicide rates to sociodemographic factors in Canadian census divisions. Can J Psychiatry. 1991; 36(9): 655-9. Heisel MJ. Suicide and its prevention among older adults. Can J Psychiatry. 2006; 51(3): 143-54. Hirsch JK. A review of the literature on rural suicide: risk and protective factors, incidence, and prevention. Crisis. 2006; 27(4): 189-99. Hunter E, Harvey D. Indigenous suicide in Australia, New Zealand, Canada, and the United States. Emerg Med (Fremantle). 2002; 14(1): 14-23. Hutchinson G. Variation of homicidal and suicidal behaviour within Trinidad and Tobago and the associated ecological risk factors West Indian Med J. 2005; 54(5): 319-24. Iribarren C, Sidney S, Jacobs DR Jr, Weisner C. Hospitalization for suicide attempt and completed suicide: epidemiological features in a managed care population. Soc Psychiatry Psychiatr Epidemiol. 2000; 35(7): 288-96. Judd F, Cooper AM, Fraser C, Davis J. Rural suicide-people or place effects? Aust N Z J Psychiatry. 2006; 40(3): 208-16. 124 Bibliography Kalediene R, Starkuviene S, Petrauskiene J. Social dimensions of mortality from external causes in Lithuania: do education and place of residence matter? Soz Praventivmed. 2006; 51(4): 232-9. Kaneko Y, Motohashi Y. Male gender and low education with poor mental health literacy: a population-based study. J Epidemiol. 2007; 17(4): 114-9. Knorr-Held L, Best N. A shared component model for detecting joint and selective cluster of two diseases. Journal of the Royal Statistical Society: Series A 2001; 164: 73-85. Knorr-Held L, Natário I, Penton S, Rue H, Becker N. Towards joint disease mapping. Statistical Methods in Medical Research 2005; 14: 61-82. Knox KL, Conwell Y, Caine ED. If suicide is a public health problem, what are we doing to prevent it? American Journal of Public Health. 2004; 94(1): 37-45. Kposowa A. Marital status and suicide in the national longitudinal mortality study. J. epidemiol. Community health. 2000; 54: 254-261. Kposowa AJ, McElvain JP, Breault KD. Immigration and suicide: the role of marital status, duration of residence, and social integration. Arch Suicide Res. 2008; 12(1): 82-92. Kumar CT, Mohan R, Ranjith G, Chandrasekaran R. Characteristics of high intent suicide attempters admitted to a general hospital. J Affect Disord. 2006; 91(1): 77-81. Kunugi H, Ishida S, Kato T, Sakai T, Tatsumi M, et al. No evidence for an association of polymorphisms of the tryptophan hydroxylase gene with affective disorders or attempted suicide among Japanese patients. Am. J. Psychiatry. 1999; 156: 774-76. 125 Bibliography Lawson AB, Biggeri AB, Boehning D, Lesaffre E, Viel JF, Clark A, Schlattmann P, Divino F. Disease mapping models: an empirical evalu- ation. Disease Mapping Collaborative Group. Stat Med. 2000; 19(17-18): 2217-41. Lawson AB, Williams FR. An introduction of disease mapping. Wiley; 2001. Leroux B, Lei X, Breslow N. Estimation of disease rates in small areas: a new mixed model for spatial dependence. In: Halloran M, Berry D eds. Statistical Models in Epidemiology, the Environment and Clinical Trials. New York: Springer, 1999; 135-178. Lester D. A regional analysis of suicide and homicide rates in the USA: search for broad cultural patterns. Soc Psychiatry Psychiatr Epidemiol. 1988; 23(3): 202-5. Lester D. Aboriginal suicide in British Columbia. Percept Mot Skills. 1996; 83(3 Pt 2): 1202. Lewinsohn PM, Rohde P, Seeley JR. Adolescent Suicidal Ideation and At- tempts: Prevalence, Risk Factors, and Clinical Implications. Clin Psychor Sci Prac. 1996; 3: 25-46. Lewinsohn PM, Rohde P, Seeley JR, Baldwin CL. Gender differences in suicide attempts from adolescence to young adulthood. J Am Acad Child Adolesc Psychiatry. 2001; 40(4): 427-434. MacNab YC, Dean CB. Parametric boostrap and penalized quasi-likelihood inference in conditional autoregressive models. Statistics in Medicine. 2000; 19: 2421-2435. MacNab YC, Dean CB. Autoregressive spatial smoothing and temporal spline smoothing for mapping rates. Biometrics. 2001; 57(3): 949-56. 126 Bibliography MacNab YC. Hierarchical Bayesian spatial modelling of small-area rates of non-rare disease. Statistics in Medicine. 2003; 22(10): 1761-73. MacNab, YC. Bayesian spatial and ecological models for small-area accident and injury analysis. Accident Analysis and Prevention. 2004; 36(6): 1019- 28. MacNab YC, Farrell PJ, Gustafson P, Wen S. Estimation in bayesian disease mapping. Biometrics. 2004; 60(4): 865-873. MacNab YC, Kmetic A, Gustafson P, Shaps S. An innovative application of Bayesian disease mapping methods to patient safety research: the Canadian iatrogenic injury study. Statistics in Medicine. 2006; 35(23): 3960-3980. MacNab YC. Mapping Disability-adjusted Life Years: A Bayesian hierarchi- cal model framework for burden of disease and injury assessment. Statistics in Medicine. 2007; 26(26): 4746-69. MacNab, YC, Lin, Y. On empirical Bayes penalized quasi-likelihood inference in GLMMs and in Bayesian disease mapping and eco- logical modeling. Computational Statistics and Data Analysis 2008;doi:10.1016/j.csda.2008.07.032 Maki NE, Martikainen PT. Socioeconomic differences in suicide mortality by sex in Finland in 1971-2000: a register-based study of trends, levels, and life expectancy differences. Scand J Public Health. 2007; 35(4): 387-95. Mann JJ, Malone KM. Cerebrospinal fluid amines and higher-lethality sui- cide attempts in depressed inpatients. Biol. Psychiatry. 1997; 41: 162-71. Mann JJ. A current perspective of suicide and attempted suicide. Ann Intern Med. 2002; 136(4): 302-11. Maris RW. Suicide. Lancet. 2002; 360(9329): 319-26. 127 Bibliography Marusic A. Regional variation in suicide mortality in Republic of Slovenia. Crisis. 1998; 19: 159-166. Masocco M, Pompili M, Vichi M, Vanacore N, Lester D, Tatarelli R. Suicide and marital status in Italy. Psychiatr Q. 2008; 79(4): 275-85. Mathur VK, Freeman DG. A Theoretical Model of Adolescent Suicide and Some Evidence from US Data. Health Economics. 2002; 11(8): 695-708. McIntosh, JL. Suicide Prevention in the Elderly (Age 65-99). Suicide and Life-Threatening Behavior. 1995; 25(1): 180-192. Middleton N, Whitley E, Frankel S, Dorling D, Sterne J, Gunnell D. Sui- cide risk in small areas in England and Wales, 1991-1993. Soc Psychiatry Psychiatr Epidemiol 2004; 39: 45-62. Newman SC, Stuart H. An ecologic study of parasuicide in Edmonton and Calgary. Can J Psychiatry. 2005; 50(5): 275-80. Pearson K. On Lines and Planes of Closest Fit to Systems of Points in Space. Philosophical Magazine. 1901; 2(6): 559-572. Pearce J, Barnett R, Jones I. Have urban/rural inequalities in suicide in New Zealand grown during the period 1980-2001? Soc Sci Med. 2007; 65(8): 1807-19. Qin P, Agerbo E, Mortensen PB. Suicide risk in relation to socioeconomic, demographic, psychiatric, and familial factors: a national register-based study of all suicides in Denmark 1981-1997. Am. J. Psychiatry. 2003; 160: 765-72. Rehkopf DH, Buka SL. The association between suicide and the socio- economic characteristics of geographical areas: a systematic review. Psychol Med. 2006; 36(2): 145-57. 128 Bibliography Rezaeian M, Dunn G, St Leger S, Appleby L. The ecological association between suicide rates and indices of deprivation in English local authorities. Soc Psychiatry Psychiatr Epidemiol. 2005; 40(10): 785-91. Rubenowitz E, Waern M, Wilhelmson K, Allebeck P. Life events and psy- chosocial factors in elderly suicides-a case-control study. Psychol Med. 2001; 31(7): 1193-202. Rujescu D, Giegling I, Sato T, Hartmann AM, Moller H-J. Genetic vari- ations in tryptophan hydroxylase in suicidal behavior: analysis and meta- analysis. Biol. Psychiatry. 2003; 54: 465-73. Safer DJ. Self-reported suicide attempts by adolescents. Ann Clin Psychia- try. 1997; 9(4): 263-9. Saunderson TR, Langford IH. A study of the geographical distribution of suicide rates in england and wales 1989-92 using empirical bayes estimates. Soc Sci Med. 1996; 43(4): 489-502. Shah A, Bhat R, Mackenzie S, Koen C. A cross-national study of the re- lationship between elderly suicide rates and life expectancy and markers of socioeconomic status and health care. Int Psychogeriatr. 2008; 20(2): 347-60. Schieman S. Residential Stability and the Social Impact of Neighborhood Disadvantage: Gender- and Race-Contingent Patterns Social Forces. 2005; 83(3): 1031-1064. Snow J. Snow on cholera: being a reprint of two papers. London: The Commonwealth Fund; 1936. Spirito A, Esposito-Smythers C. Attempted and Completed Suicide in Ado- lescence. Annual Review of Clical Psychology. 2006; 2: 237-66. 129 Bibliography Stack S. The effect of immigration on suicide: A cross-national analysis. Basic and Applied Social Psychology. 1981; 2: 205C218. Statistics Canada. Mortality, Summary List of Causes 2000, 2001, 2002, 2003. Ottawa, ON, Canada; 2006. Statistics Canada. Suicide and suicide rates, by sex and by age group. Retrieved May 2007 from http://www40.statcan.ca/l01/cst01/hlth66a. htm?sdi=suicide Statistics Canada. Leading causes of death, 2000 to 2004. Retrieved Dec 2008 from http://www.statcan.gc.ca/daily-quotidien/081204/dq081204c- eng.htm Steele MM, Doey T. Suicidal behavior in children and adolescents. part 1: etiology and risk factors. Can J Psychiatry. 2007; 52(6 Suppl 1): 21S-33S. Suicide.org. U.S. Suicide Statistics. Available at: http://www.suicide.org/ suicide-statistics.html#2001. 2001. Suominen K, Isometsa E, Suokas J, Haukka J, Achte K, Lonnqvist J. Com- pleted suicide after a suicide attempt: a 37-year follow-up study. Am J Psychiatry. 2004; 161(3): 562-3. Taylor R, Page A, Morrell S, Harrison J, Carter G. Mental health and socio-economic variations in Australian suicide. Soc Sci Med. 2005; 61(7): 1551-9. Voaklander DC, Rowe BH, Dryden DM, Pahal J, Saar P, Kelly KD. Medi- cal illness, medication use and suicide in seniors: a population-based case- control study. J Epidemiol Community Health. 2008; 62(2): 138-46. Vyas S, Kumaranayake L. Constructing socio-economic status indices: how to use principal components analysis. Health Policy Plan. 2006; 21(6): 459- 68. 130 Bibliography Westman J, Sundquist J, Johansson LM, Johansson SE, Sundquist K. Coun- try of birth and suicide: A followup study of a national cohort in Sweden. Archives of Suicide Research. 2006; 10: 239C248. Whitley E, Gunnell D, Dorling D, Smith GD. Ecological study of social fragmentation, poverty, and suicide. BMJ. 1999; 319(7216): 1034-7. World Suicide Prevention Day, 10 September 2008: Think Globally, Plan Nationally, Act Locally. World Health Organization. 2008 Suicide prevention. World Health Organization. Available at: http:// www. who.int/mental health/prevention/suicide/suicideprevent/en/index.html, 2008. MRC Biostatistics Unit, Cambridge, UK. WinBUGS (bayesian in- ference using gibbs sampling) software, version 1.4. http://www.mrc- bsu.cam.ac.uk/bugs/winbugs /contents.shtml, 2003. 131 Appendix A Regional characteristics profiles 132 A p p en d ix A . R eg io n a l ch a ra cteristics p ro fi les Table A.1: Regional characteristics and summary statistics across 84 LHAs, B.C., 1991–2000. Domain Derived census variables Mean Sd Range 1. Marital Status • % Population aged 15+ never married 29.76 6.90 18.31 – 54.40 • % Population aged 15+ married 52.96 6.81 24.58 – 63.62 • % Population aged 15+ separated 3.71 0.67 2.36 – 5.87 • % Population aged 15+ divorced 8.18 1.45 5.34 – 12.31 • % Population aged 15+ widowed 5.36 1.59 0 – 8.76 2. Language • % Population with English as 1st official 97.17 2.86 83.72 – 99.75 language spoken • % Population with French as 1st official 1.36 0.61 0 – 3.58 language spoken • % Population with both English and 0.21 0.22 0 – 0.87 French as 1st official language spoken • % Population with neither English nor 1.26 2.65 0 – 13.57 French as 1st official language spoken 3. Immigration • % Immigrants from Europe 58.28 16.48 14.23 – 79.49 • % Immigrants from Africa 1.77 1.63 0 – 8.66 • % Immigrants from Asia 19.69 18.62 0 – 77.10 • % Immigrants from Oceania and other 2.32 1.46 0 – 7.22 • % Immigrants before 1961 31.98 12.47 0 – 56.07 • % Immigrants during 1961 - 1970 19.76 6.51 7.14 – 55.00 • % Immigrants during 1971 - 1980 21.09 6.35 0 – 39.29 • % Immigrants during 1981 - 1990 14.16 6.61 0 – 32.81 Continued on next page133 A p p en d ix A . R eg io n a l ch a ra cteristics p ro fi les Table A.1 – continued from previous page Domain Derived census variables Mean Sd Range 3. Immigration • % Immigrants during 1991 - 1996 12.36 8.95 0 – 42.67 • % Immigrants at age 0 - 4 years 11.86 3.97 0 – 33.33 • % Immigrants at age 5 - 19 years 28.06 5.02 19.39 – 52.38 • % Immigrants at age over 20 years 60.62 6.32 41.67 – 83.33 4.Aboriginal • % Population with Aboriginal origins 10.14 16.30 0.64 – 90.93 Population • % Population with mix-Aboriginal 0.04 0.06 0 – 0.27 origins • % North American Indians and 8.86 15.79 0.37 – 87.66 non-Indians • % Métis Indians and non-Indians 0.88 0.63 0 – 3.94 • % Inuit Indians and non-Indians 0.03 0.05 0 – 0.25 5. Mobility Status • % Population aged 1+ in last year are 19.14 3.31 13.04 – 32.97 movers • % Population aged 5+ in last year are 51.13 6.31 36.03 – 73.98 movers 6. Education • % Population aged 15+ with less than 8.63 3.58 1.95 – 18.82 grade 9 education • % Population aged 15+ with secondary 12.99 2.15 5.07 – 21.57 school graduation certificates • % Population aged 20+ with university 19.78 8.80 8.50 – 58.64 or higher education • % Population with full-time school 49.60 6.40 33.97 – 71.76 Continued on next page134 A p p en d ix A . R eg io n a l ch a ra cteristics p ro fi les Table A.1 – continued from previous page Domain Derived census variables Mean Sd Range 7. Unemployment • Unemployment rate for population aged 11.55 4.39 4.70 – 31.20 Rates 15+ • Unemployment rate for population aged 18.01 5.88 8.90 – 42.90 15-24 • Unemployment rate for population aged 10.33 4.33 3.40 – 29.30 25+ • Unemployment rate for population aged 11.50 4.34 4.71 – 30.86 15+ in private households • Unemployment rate for population aged 13.11 5.08 6.52 – 40.00 15+ without children in private households • Unemployment rate for population aged 9.78 4.23 2.21 – 26.67 15+ with children in private households • Unemployment rate for population aged 12.46 5.49 0– 31.25 15+ with children aged 6- in private households • Unemployment rate for population aged 8.86 4.39 1.88 – 29.27 15+ with children aged 6+ in private households • % Population with full-time work 43.30 6.14 23.43 – 54.82 8. Transportation • % Labor force aged 15+ using public 4.07 6.99 0– 29.73 transit Continued on next page135 A p p en d ix A . R eg io n a l ch a ra cteristics p ro fi les Table A.1 – continued from previous page Domain Derived census variables Mean Sd Range 8. Transportation • % Male labor force aged 15+ using 3.19 5.37 0– 24.04 public transit • % Female labor force aged 15+ using 5.08 8.91 0– 38.66 public transit 9 . Income • Average employment income for 26,079 4,582 16,149 – 48,390 population aged 15+ • % Income from government transfer 74.71 8.40 53.90 – 91.50 • % Employment income 15.02 5.11 6.20 – 31.70 • Average income for males aged 15+ 31,216 6,575 18,072 – 65,140 • Median income for males aged 15+ 26,685 6,512 12,347 – 48,454 • Average income for females aged 15+ 18,109 2,955 13,539 – 30,034 • Median income for females aged 15+ 13,959 2,516 10,062 – 21,842 • % Population with low income 13.84 5.36 4.20 – 40.70 • % Income from Other Sources 10.28 5.34 1.60 – 28.10 • % Total one-family households paying 19.85 7.06 0 – 38.73 30% income on spending • Average census family income 52,692 10,780 33,587 – 110,421 • Median census family income 46,954 9,096 27,520 – 79,443 • Average single-lone male parent family 39,501 12,401 0 – 72,955 income • Average single-lone female parent family 26,438 5,089 16,949 – 44,961 income • Incidence of low income 14.06 5.53 4.20 – 40.70 Continued on next page136 A p p en d ix A . R eg io n a l ch a ra cteristics p ro fi les Table A.1 – continued from previous page Domain Derived census variables Mean Sd Range 9 . Income • Incidence of low income for population 36.36 7.85 8.30 – 67.00 aged 15+ unattached individuals • Incidence of low income for people in 17.46 6.69 7.50 – 51.80 private households 10. Family Status • % Persons living alone in households 9.41 4.59 3.27 – 37.42 • % Seniors (65+) living alone 27.67 6.77 11.76 – 59.44 • % Children aged 6- at home 24.66 2.85 16.84 – 31.14 • % Children aged 6-14 at home 41.59 4.41 25.73 – 49.09 • % Children aged 15-24 at home 27.61 3.29 20.63 – 38.33 • % children aged 25+ at home 6.07 3.38 2.61 – 17.82 • % Census families with 4- persons 67.84 6.52 50.00 – 92.97 • % Census families with 4+ persons 32.10 6.49 7.03 – 50.00 • Average number of persons per family 3.00 0.20 2.30 – 3.70 • % Census families without children 40.09 8.55 17.71 – 66.77 • % Census families with 1 child 23.60 3.32 15.38 – 32.53 • % Census families with 2 or more children 36.17 6.74 9.15 – 53.13 • % Population with lone-parent 13.40 2.72 7.02 – 25.00 • % Babies with low birth weight 4.57 2.12 0 – 12.43 • % Population being teenage mothers 1.25 0.68 0.05 – 3.82 11. Dwellings • % Dwelling rented 69.14 11.91 20.60 – 82.94 • % Dwelling owned 29.23 11.34 16.91 – 79.38 • % Dwelling needing major repairs 11.06 5.85 4.18 – 36.36 • % Dwelling constructed before 1946 10.07 8.05 1.16 – 40.20 Continued on next page 137 A p p en d ix A . R eg io n a l ch a ra cteristics p ro fi les Table A.1 – continued from previous page Domain Derived census variables Mean Sd Range 11. Dwellings • % Dwelling constructed during 12.57 6.31 2.90 – 37.02 1946 - 1960 • % Dwelling constructed during 15.72 4.30 7.25 – 25.32 1961 - 1970 • % Dwelling constructed during 27.66 7.46 11.95 – 46.57 1971 - 1980 • % Dwelling constructed during 19.95 6.74 7.72 – 42.06 1981 - 1990 • % Dwelling constructed during 14.02 5.23 3.29– 24.60 1991 - 1996 • Average gross rent 110,619 53,534 27,884 – 392,925 • % Spending 30%+ on shelter cost 36.49 9.83 14.94 – 62.22 138 Appendix B Annual HA and HSDA suicide rates 139 Appendix B. Annual HA and HSDA suicide rates 1992 1994 1996 1998 2000 50 10 0 15 0 20 0 Interior Year R at e pe r 1 00 ,0 00  p er so n 1992 1994 1996 1998 2000 50 10 0 15 0 20 0 Fraser Year R at e pe r 1 00 ,0 00  p er so n 1992 1994 1996 1998 2000 50 10 0 15 0 20 0 Vancouver Coastal Year R at e pe r 1 00 ,0 00  p er so n 1992 1994 1996 1998 2000 50 10 0 15 0 20 0 Vancouver island Year R at e pe r 1 00 ,0 00  p er so n 1992 1994 1996 1998 2000 50 10 0 15 0 20 0 Northern Year R at e pe r 1 00 ,0 00  p er so n BC Male Female Figure B.1: Annual HA rates of suicide hospitalization, males and females, 1991-2000 140 Appendix B. Annual HA and HSDA suicide rates 1992 1994 1996 1998 2000 0 5 10 15 20 25 30 Interior Year R at e pe r 1 00 ,0 00  p er so n 1992 1994 1996 1998 2000 0 5 10 15 20 25 30 Fraser Year R at e pe r 1 00 ,0 00  p er so n 1992 1994 1996 1998 2000 0 5 10 15 20 25 30 Vancouver Coastal Year R at e pe r 1 00 ,0 00  p er so n 1992 1994 1996 1998 2000 0 5 10 15 20 25 30 Vancouver island Year R at e pe r 1 00 ,0 00  p er so n 1992 1994 1996 1998 2000 0 5 10 15 20 25 30 Northern Year R at e pe r 1 00 ,0 00  p er so n BC Male Female Figure B.2: Annual HA rates of suicide mortality, males and females, 1991- 2000 141 Appendix B. Annual HA and HSDA suicide rates 1992 1994 1996 1998 2000 50 10 0 15 0 20 0 25 0 30 0 35 0 East Kootenay Year R at e pe r 1 00 ,0 00 1992 1994 1996 1998 2000 50 10 0 15 0 20 0 25 0 30 0 35 0 Kootenay Boundary Year R at e pe r 1 00 ,0 00 1992 1994 1996 1998 2000 50 10 0 15 0 20 0 25 0 30 0 35 0 Okanagan Year R at e pe r 1 00 ,0 00 1992 1994 1996 1998 2000 50 10 0 15 0 20 0 25 0 30 0 35 0 Thompson Cariboo Shuswap Year R at e pe r 1 00 ,0 00 1992 1994 1996 1998 2000 50 10 0 15 0 20 0 25 0 30 0 35 0 Fraser East Year R at e pe r 1 00 ,0 00 1992 1994 1996 1998 2000 50 10 0 15 0 20 0 25 0 30 0 35 0 Fraser North Year R at e pe r 1 00 ,0 00 1992 1994 1996 1998 2000 50 10 0 15 0 20 0 25 0 30 0 35 0 Fraser South Year R at e pe r 1 00 ,0 00 1992 1994 1996 1998 2000 50 10 0 15 0 20 0 25 0 30 0 35 0 Richmond Year R at e pe r 1 00 ,0 00 BC Male Female Figure B.3: Annual HSDA rates of suicide hospitalization, males and fe- males, 1991-2000 142 Appendix B. Annual HA and HSDA suicide rates 1992 1994 1996 1998 2000 50 10 0 15 0 20 0 25 0 30 0 35 0 Vancouver Year R at e pe r 1 00 ,0 00 1992 1994 1996 1998 2000 50 10 0 15 0 20 0 25 0 30 0 35 0 North Shore/Coast Garibaldi Year R at e pe r 1 00 ,0 00 1992 1994 1996 1998 2000 50 10 0 15 0 20 0 25 0 30 0 35 0 South Vancouver Island Year R at e pe r 1 00 ,0 00 1992 1994 1996 1998 2000 50 10 0 15 0 20 0 25 0 30 0 35 0 Central Vancouver Island Year R at e pe r 1 00 ,0 00 1992 1994 1996 1998 2000 50 10 0 15 0 20 0 25 0 30 0 35 0 North Vancouver Island Year R at e pe r 1 00 ,0 00 1992 1994 1996 1998 2000 50 10 0 15 0 20 0 25 0 30 0 35 0 Northwest Year R at e pe r 1 00 ,0 00 1992 1994 1996 1998 2000 50 10 0 15 0 20 0 25 0 30 0 35 0 Northern Interior Year R at e pe r 1 00 ,0 00 1992 1994 1996 1998 2000 50 10 0 15 0 20 0 25 0 30 0 35 0 Northeast Year R at e pe r 1 00 ,0 00 BC Male Female Figure B.4: Annual HSDA rates of suicide hospitalization, males and fe- males, 1991-2000 (continued) 143 Appendix B. Annual HA and HSDA suicide rates 1992 1994 1996 1998 2000 0 10 20 30 East Kootenay Year R at e pe r 1 00 ,0 00 1992 1994 1996 1998 2000 0 10 20 30 Kootenay Boundary Year R at e pe r 1 00 ,0 00 1992 1994 1996 1998 2000 0 10 20 30 Okanagan Year R at e pe r 1 00 ,0 00 1992 1994 1996 1998 2000 0 10 20 30 Thompson Cariboo Shuswap Year R at e pe r 1 00 ,0 00 1992 1994 1996 1998 2000 0 10 20 30 Fraser East Year R at e pe r 1 00 ,0 00 1992 1994 1996 1998 2000 0 10 20 30 Fraser North Year R at e pe r 1 00 ,0 00 1992 1994 1996 1998 2000 0 10 20 30 Fraser South Year R at e pe r 1 00 ,0 00 1992 1994 1996 1998 2000 0 10 20 30 Richmond Year R at e pe r 1 00 ,0 00 BC Male Female Figure B.5: Annual HSDA rates of suicide mortality, males and females, 1991-2000 144 Appendix B. Annual HA and HSDA suicide rates 1992 1994 1996 1998 2000 0 10 20 30 Vancouver Year R at e pe r 1 00 ,0 00 1992 1994 1996 1998 2000 0 10 20 30 North Shore/Coast Garibaldi Year R at e pe r 1 00 ,0 00 1992 1994 1996 1998 2000 0 10 20 30 South Vancouver Island Year R at e pe r 1 00 ,0 00 1992 1994 1996 1998 2000 0 10 20 30 Central Vancouver Island Year R at e pe r 1 00 ,0 00 1992 1994 1996 1998 2000 0 10 20 30 North Vancouver Island Year R at e pe r 1 00 ,0 00 1992 1994 1996 1998 2000 0 10 20 30 Northwest Year R at e pe r 1 00 ,0 00 1992 1994 1996 1998 2000 0 10 20 30 Northern Interior Year R at e pe r 1 00 ,0 00 1992 1994 1996 1998 2000 0 10 20 30 Northeast Year R at e pe r 1 00 ,0 00 BC Male Female Figure B.6: Annual HSDArates of suicide mortality, males and females, 1991- 2000 (continued) 145 Appendix C Sensitivity analysis for priors and hyperpriors The priors and hyperpriors were set up for the parameters of Poisson regres- sion model (4.2) with random effects (4.3–4.5) as follows: inverse gamma priors InvGamma(0.001, 0.001) for σ2, uniform priors Unif(0, 1) for λ and β0, normal priors N(0, 100 2) for b and all risk factors βs. Sensitivity was assessed by running the MCMC simulations for various alternative hyperpri- ors with other priors unchanged, such as: Unif(-100, 100) for the regression coefficients βs; Unif(0, 10) for β0; σ 2 ∼ InvGamma(0.1, 0.1), InvGamma(0.5, 0.005) and Unif(0, 100); and λ ∼ Beta(2, 4), Beta(4, 2). The sensitivity analyses were done based on the spatial-ecological model for the suicide hospitalization rates of males aged 15-24. The first 4 PCs of the significant regional characteristics were considered as covariates in the spatial-ecological models. The βs were the corresponding regression coeffi- cients. The posterior means and 95% credible intervals of parameters and DIC from each model are presented in Table C.1. Overall, the resulting inference for all parameters remained relatively con- sistent under the different priors and hyperpriors. The DICs also remained stable across the models. Different priors and hyperpriors had little effect on the fixed effects, i.e., βs. While small changes were observed for the alter- native hyperpriors for σ2 in the posterior estimates of σ2 and λ, the credible intervals overlapped. While the different priors for λ make moderate changes in the posterior estimates of λ, the changes can be ignored in the posterior estimates of the random effects. The priors for the fixed effects βs had little impacts on the posterior estimates of all parameters. The priors and hyper- 146 Appendix C. Sensitivity analysis for priors and hyperpriors Table C.1: Sensitivity analysis for the prior and hyperprior specifications (Mean (95%CI)**) Original model* β ∼ Unif(−100, 100) DIC 508 508 λ 0.66 (0.20,0.98) 0.66 (0.21, 0.98) σ 0.44 (0.30,0.61) 0.44 (0.31, 0.62) m = exp(β0) 0.002 (0.001,0.002) 0.002 (0.001, 0.002) β1 -0.35 (-0.44,-0.26) -0.35 (-0.45, -0.26) β2 -0.08 (-0.19, 0.03) -0.08 (-0.19, 0.02) β3 -0.18 (-0.29,-0.07) -0.18 (-0.29, -0.07) β4 0.05 (-0.05, 0.14) 0.05 (-0.04, 0.14) β0 ∼ Unif(0, 10) σ 2 ∼ InvGamma(0.1, 0.1) DIC 508 507 λ 0.66 (0.22, 0.99) 0.67 (0.24, 0.98) σ 0.44 (0.30, 0.61) 0.46 (0.32, 0.63) m = exp(β0) 0.002 (0.001, 0.002) 0.002 (0.001, 0.002) β1 -0.35 (-0.44, -0.26) -0.35 (-0.45, -0.26) β2 -0.08 (-0.19, 0.03) -0.08 (-0.19, 0.03) β3 -0.18 (-0.29, -0.07) -0.18 (-0.29, -0.07) β4 0.05 (-0.04, 0.14) 0.05 (-0.04, 0.14) σ2 ∼ InvGamma(0.5, 0.005) σ2 ∼ Unif(0, 100) DIC 508 506 λ 0.64 (0.19, 0.98) 0.70 (0.25, 0.99) σ 0.43 (0.29, 0.60) 0.47 (0.32, 0.65) m = exp(β0) 0.002 (0.001, 0.002) 0.002 (0.001, 0.002) β1 -0.35 (-0.44, -0.26) -0.35 (-0.45, -0.25) β2 -0.08 (-0.19, 0.02) -0.08 (-0.20, 0.04) β3 -0.18 (-0.29, -0.08) -0.18 (-0.29, -0.06) β4 0.05 (-0.04, 0.14) 0.05 (-0.04, 0.14) *: Priors are β0 and λ ∼ Unif(0, 1); σ 2 ∼ InvGamma(0.001, 0.001); β ∼ N(0, 1002). **: Posterior mean and 95% credible interval. 147 Appendix C. Sensitivity analysis for priors and hyperpriors Table 3.2(Continued): Sensitivity analysis for the prior and hyperprior specifications (Mean (95%CI)**) Original model* λ ∼ Beta(2, 4) DIC 508 507 λ 0.66 (0.20,0.98) 0.46 (0.16, 0.78) σ 0.44 (0.30,0.61) 0.40 (0.28, 0.55) m = exp(β0) 0.002 (0.001,0.002) 0.002 (0.001, 0.002) β1 -0.35 (-0.44,-0.26) -0.36 (-0.44, -0.27) β2 -0.08 (-0.19, 0.03) -0.09 (-0.18, 0.01) β3 -0.18 (-0.29,-0.07) -0.19 (-0.29, -0.09) β4 0.05 (-0.05, 0.14) 0.05 (-0.04, 0.14) λ ∼ Beta(4, 2) DIC 508 λ 0.70 (0.34, 0.96) σ 0.45 (0.32, 0.61) m = exp(β0) 0.002 (0.001, 0.002) β1 -0.35 (-0.45, -0.26) β2 -0.08 (-0.19, 0.03) β3 -0.18 (-0.29, -0.07) β4 0.05 (-0.04, 0.14) *: Priors are β0 and λ ∼ Unif(0, 1); σ 2 ∼ InvGamma(0.001, 0.001); β ∼ N(0, 1002). **: Posterior mean and 95% credible interval. priors chosen for the analysis seemed reasonable and tended not to influence the resulting inference for the data. 148 Appendix D Top 10 LHAs with significant RRs LHAs were ranked based on the unadjusted RRs for suicide hospitalization and mortality respectively. The boldface LHAs indicated that the risk- adjusted RRs were statistically significant with accounting for risk factors. 149 Appendix D. Top 10 LHAs with significant RRs Table D.1: Top 10 LHAs with significant RRs, youth aged 15-24 Rank Male suicide hospitalization Male suicide death 1 Central Coast Central Coast 2 Upper Skeena Nisga’a 3 Nisga’a Snow Country 4 Prince Rupert Upper Skeena 5 Bella Coola Valley Stikine/Telegraph Creek 6 Queen Charlotte Bella Coola Valley 7 Snow Country Nechako 8 North Thompson Queen Charlotte 9 Stikine/Telegraph Creek Prince Rupert 10 Vancouver Island North Lilooet Rank Female suicide hospitalization Female suicide death 1 Central Coast Central Coast 2 Nisga’a Upper Skeena 3 Upper Skeena Queen Charlotte 4 Prince Rupert Lilooet 5 Smithers Merritt 6 Lilooet Nechako 7 100 Mile House 8 Burns Lake 9 Alberni 10 Bella Coola Valley LHA in boldface has significantly high RRs after risk adjustment. 150 Appendix D. Top 10 LHAs with significant RRs Table D.2: Top 10 LHAs with significant RRs, adult aged 25-34 Rank Male suicide hospitalization Male suicide death 1 Central Coast Vancouver Downtown East Side 2 Nisga’a 3 Upper Skeena 4 Hope 5 Burns Lake 6 Alberni 7 Prince Rupert 8 North Thompson 9 Bella Coola Valley 10 South Cariboo Rank Female suicide hospitalization Female suicide death 1 Central Coast Vancouver Downtown East Side 2 Nisga’a Central Coast 3 Upper Skeena New Westminster 4 Agassiz-Harrison 5 Burns Lake 6 Prince Rupert 7 Lilooet 8 Bella Coola Valley 9 Enderby 10 100 Mile House LHA in boldface has significantly high RRs after risk adjustment. 151 Appendix D. Top 10 LHAs with significant RRs Table D.3: Top 10 LHAs with significant RRs, adult aged 35-44 Rank Male suicide hospitalization Male suicide death 1 Central Coast Vancouver Downtown East Side 2 Vancouver Downtown East Side Vancouver City Centre 3 Nisga’a Stikine/Telegraph Creek 4 Chilliwack Central Coast 5 Hope Queen Charlotte 6 Bella Coola Valley Keremeos 7 Upper Skeena New Westminster 8 Terrace Lilooet 9 Merritt Arrow Lakes 10 Powell River Merritt Rank Female suicide hospitalization Female suicide death 1 Central Coast Vancouver Downtown East Side 2 Nisga’a Vancouver City Centre 3 Upper Skeena Keremeos 4 Hope Southern Okanagan 5 Prince Rupert Penticton 6 Smithers New Westminster 7 100 Mile House Arrow Lakes 8 Vancouver Island North Greater Victoria 9 Vancouver Downtown East Side Sunshine Coast 10 Agassiz-Harrison Nanaimo LHA in boldface has significantly high RRs after risk adjustment. 152 Appendix D. Top 10 LHAs with significant RRs Table D.4: Top 10 LHAs with significant RRs, adult aged 45-54 Rank Male suicide hospitalization Male suicide death 1 Vancouver Downtown East Side Vancouver Downtown East Side 2 Terrace Stikine/Telegraph Creek 3 New Westminster Vancouver City Centre 4 Misson Queen Charlotte 5 Prince Rupert Lilooet 6 Vancouver City Centre New Westminster 7 Chilliwack Courtenay 8 Greater Victoria 9 Nanaimo Rank Female suicide hospitalization Female suicide death 1 Howe Sound Vancouver City Centre 2 Upper Skeena Vancouver Downtown East Side 3 South Cariboo New Westminster 4 Alberni 5 Ladysmith 6 Prince Rupert 7 Chilliwack 8 Nanaimo 9 Greater Victoria LHA in boldface has significantly high RRs after risk adjustment. 153 Appendix D. Top 10 LHAs with significant RRs Table D.5: Top 10 LHAs with significant RRs, adult aged 55-64 Rank Male suicide hospitalization Male suicide death 1 Vancouver Downtown East Side Vancouver Downtown East Side 2 Vancouver City Centre Vancouver City Centre 3 Greater Victoria Stikine/Telegraph Creek 4 Nanaimo Queen Charlotte 5 Lilooet 6 New Westminster Rank Female suicide hospitalization Female suicide death 1 Ladysmith Vancouver Downtown East Side 2 Prince Rupert Vancouver City Centre 3 Maple Ridge New Westminster 4 Chilliwack Greater Victoria 5 Greater Victoria Kootenay Lake/Nelson 6 Nanaimo LHA in boldface has significantly high RRs after risk adjustment. Table D.6: Top 10 LHAs with significant RRs, elderly aged 65-74 Rank Male suicide hospitalization Male suicide death 1 Vancouver Downtown East Side Queen Charlotte 2 Vancouver City Centre Stikine/Telegraph Creek 3 New Westminster Vancouver Downtown East Side 4 Burnaby 5 North Vancouver Rank Female suicide hospitalization Female suicide death 1 Sunshine Coast Vancouver West Side 2 Vancouver City Centre 3 West Vancouver-Bowen Island 4 North Vancouver 5 Richmond 6 Vancouver South 7 Delta 8 Coquitlam 154 Appendix D. Top 10 LHAs with significant RRs Table D.7: Top 10 LHAs with significant RRs, elderly aged 75+ Rank Male suicide hospitalization Male suicide death 1 Vancouver City Centre Vancouver City Centre 2 Greater Victoria Southern Okanagan Rank Female suicide hospitalization Female suicide death 1 Vancouver Downtown East Side 2 Not Applicable Vancouver City Centre 3 Greater Victoria 4 Vancouver West Side 5 New Westminster 6 North Vancouver 155

Cite

Citation Scheme:

        

Citations by CSL (citeproc-js)

Usage Statistics

Share

Embed

Customize your widget with the following options, then copy and paste the code below into the HTML of your page to embed this item in your website.
                        
                            <div id="ubcOpenCollectionsWidgetDisplay">
                            <script id="ubcOpenCollectionsWidget"
                            src="{[{embed.src}]}"
                            data-item="{[{embed.item}]}"
                            data-collection="{[{embed.collection}]}"
                            data-metadata="{[{embed.showMetadata}]}"
                            data-width="{[{embed.width}]}"
                            data-media="{[{embed.selectedMedia}]}"
                            async >
                            </script>
                            </div>
                        
                    
IIIF logo Our image viewer uses the IIIF 2.0 standard. To load this item in other compatible viewers, use this url:
https://iiif.library.ubc.ca/presentation/dsp.24.1-0067638/manifest

Comment

Related Items