A Study of the Spatial Distribution of Suicide Rates

A Study of the Spatial Distribution of Suicide Rates Ferdinand DiFurio, Tennessee Tech University Willis Lewis, Winthrop University With acknowledgements to Kendall Knight, GA, Tennessee Tech University Introduction Suicide is a very personal and sensitive issue -Tragedy, both direct and indirect -Personal Conduct research that may be helpful with prevention This research mixes economics with sociology Suicide Determinants The assumption taken in this paper -micro-level characteristics matter, but macroindirect level Risk factors (there are many) and the pathways -mental health  suicide -other factors directlysuicide *bypass mental health? A mentally healthy individual could experience significant trauma and commit suicide -other factors indirectlymental healthsuicide Suicide Determinants Other factors? Trauma, substance abuse, genetics?, age, previous attempts, family size, family dynamics, social dynamics (i.e change in societal status), trauma (divorce, death of family/friends), religious involvement, length of time from trauma, physical illness, business cycle Suicide Determinants Business cycle? As Snipes et al (2012) point out, Durkheim’s theory “asserted that the suicide rate varies inversely with the stability and durability of social relationships and that any economic change, positive or negative, disrupts that status quo and contributes to an increased incidence of suicide” What our paper does Objective -to identify relationships in aggregated suicide rates From here, we can gain more information on how external factors influence these micro-level characteristics, and ultimately improve prevention Our basic question -Does a relationship exist between countylevel suicide rates and socioeconomic characteristics? How is our paper different? Rich literature in regional variation of suicide rates -Snipes et al (2012), Lester (1985) look at cities -Churchill (1999) look at VT counties -Gruenewald et al (1995) found higher suicides in regions with higher alcohol sales, and higher in states with higher relative rates of divorce, interstate migration, crime, and lower levels of church attendance (Lester 1988) -Lester (1998) Kaplan and Geling, Molina and Duarte (2006) find higher suicides in regions with looser handgun control and higher incidence of gun ownership -Snipes also goes to cite Cebula and Zelenskaya (2006) that find a link bw violent rates of crime Few to none mention spatial dependence (we’ll explain later) Some things to consider Aggregation -the best we can say Causation -classic errors -cannot be implied without theory Age -our data does not measure -risk vary with age What does the literature say about aggregation and causality? Lester and Yang (2003), Preti (2003) -individual level -“Does unemployment increase the risk of of suicidal behavior, or, alternatively, are those with psychiatric problems more likely to become unemployed and also more likely to engage in suicidal behavior?” -time-induced problem E Agerbo (2003) -Karl Pearson, who said that “only correlation and not causation can be estimated from observational data.” Blakely (2003) -mental illness is the significant factor rather than other factors such as financial inadequacies General lit review Snipes (2011), Viren (2005) and Ruhm (2000), among others: the business cycle -Snipes et al suggest that higher unemployment rates increase the chance of suicide for a one month lag -They also report suicide is greater with males -unemployment appears to play a significant role in female suicide This finding contrasts the accepted belief that females are less prone to suicide due to workplace downturns because of their traditional role in the household Viren: is a statistical relationship between aggregate suicides and business cycles Ruhm: “suicide rates are predicted to rise with unemployment rates” Molina and Duarte identify depression along with other key factors : adolescents who have suffered educational failure, who possess a gun, or who are often distressed by their physical appearance are more sensitive to the possibility of attempting suicide -at the aggregate level, this is hard to verify, but we will obtain data on number of mental health facilities per county per capita Our paper’s implicit function Macro model at the county-level: suicide rateit = f(unemploymentit [higher unemployment consistent with an economic downturn, lower unemployment consistent with an economic upswing], unemploymentit-1, crimeit, (county or adjacent county), divorce rateit, religious establishmentsit, mental health accessit, foreclosure rateit (foreclosure rates typically lag unemployment)….with most of these variables, we will investigate lags….particularly with unemployment, divorce, and crime As opposed to… • Micro model at the individual level: suicide risk = f(mental health, mental health access, quality of care for mental health condition, substance abuse, genetics?, trauma (death or illness), religion, age, family size, family structure, family strife, social life, demographics, gun inside home?) What is Spatial Autocorrelation? -An important component of the data analysis in our paper is controlling for spatial dependence -Anselin & Bera: Because counties in the state are contiguous to one another, it is necessary to control for spillover effects that may bias the empirical results and cause spatial measurement errors -the problem of serial correlation in cross-sectional data, but applied to space -widely used technique that’s expanding into different fields How does it apply to this paper? -Intuitively, there is the possibility that what’s going on in surrounding counties (contiguous to) could be impacting suicides in local counties -For instance, employment shocks in a group of metropolitan counties could be impacting suicides in a contiguous rural county -If we don’t identify this, we could miss a large part of the story -If we ignored the issue, what would happen? Bias in the estimated coefficients This would impact the interpretation and understanding of what the data is saying Preliminary results To test for the presence of spatial dependence, and thus, for the appropriate model, a series of steps is required -To avoid intolerable boredom, we skip the details here -Standard Ordinary Least Squares Regression (OLS) is the “go to” model in most studies (or some variation of) -Our tests with OLS indicate unemployment is significant, and this is consistent with most “regional” studies Preliminary results -But the significance disappears in the spatial fixed effects model (the appropriate one for our data) -This raises the question on whether past estimates could have benefitted from spatial analysis: parsing out local effects from spillover is crucial *Caution: we are not saying that job loss is insignificant to the suicide decision -Why care? Resource allocation: In a world of scarcity, putting resources into A involves an opportunity cost: they were taken away from B When policy is developed to funnel resources into “unemployment” programs based on biased results, something else could’ve suffered Preliminary results -Additional findings: our results find no relationship between the local economy and suicide rates in Tennessee (no spatial dependence present so far) -But, there is a relationship between local crime rates and local divorce rates -That is, suicide increases as crime rates goes up and decreases as divorce rates goes up -Since the variables are contemporaneous, no interpretation will be made regarding the crime rate as we are not sure if suicide is reported as a crime, we have to dig deeper OLS Space FE Time FE FE Spatial lag Intercept 21.2 (5.04)*** UR 0.63 (1.88)* 0.28 (0.57) 0.53 (1.56) 0.003 (0.006) INC -0.0001 (-1.08) 0.0003 (1.43) -0.0002 (-2.20)** 0.006 (0.13) REL -0.82 (-0.07) -4.57 (-1.01) 0.37 (0.33) -6.02 (-0.13) Crime -0.06 (-3.92)*** 0.09 (1.89)* -0.06 (3.82)*** 0.09 (1.85)* Divorce -0.13 (-1.007) -0.70 (-2.24)** -0.09 (-0.67) -0.80 (-2.38)** Notes: t-values in parentheses; Significance levels * = 0.1 ** = 0.05, *** = 0.01 What’s next? -more work on the model -foreclosures -mental health access -rural vs urban -lags Thank you OLS Space FE Time FE FE Spatial lag Intercept 21.2 (5.04)*** UR 0.63 (1.88)* 0.28 (0.57) 0.53 (1.56) 0.003 (0.006) INC -0.0001 (-1.08) 0.0003 (1.43) -0.0002 (-2.20)** 0.006 (0.13) REL -0.82 (-0.07) -4.57 (-1.01) 0.37 (0.33) -6.02 (-0.13) Crime -0.06 (-3.92)*** 0.09 (1.89)* -0.06 (3.82)*** 0.09 (1.85)* Divorce -0.13 (-1.007) -0.70 (-2.24)** -0.09 (-0.67) -0.80 (-2.38)** W*suicide 0.02 (0.34) R2 0.0703 0.3988 0.0933 0.3956 LogL -2087 -1962.8 -2079.9 -1968.8 LM spatial lag 0.297 0.004 0.003 0.01 LM spatial error 0.000 0.019 0.553 1.06 Robust LM spatial lag 8.565 4.196 10.08 87.84 Robust LM spatial error 8.268 4.211 10.63 88.89 237.9*** 3.68 261.6*** LR-test Fixed Effects Notes: t-values in parentheses; Significance levels * = 0.1 ** = 0.05, *** = 0.01 ... E Agerbo (2003) -Karl Pearson, who said that “only correlation and not causation can be estimated from observational data.” Blakely (2003) -mental illness is the significant factor rather than... Determinants The assumption taken in this paper -micro-level characteristics matter, but macroindirect level Risk factors (there are many) and the pathways -mental health  suicide -other factors... increase the chance of suicide for a one month lag -They also report suicide is greater with males -unemployment appears to play a significant role in female suicide This finding contrasts the accepted