On predictions, measurement, and causal inference : Essays on statistical correlations and statistical applications in labor market economics and criminology

Sammanfattning: This thesis consists of five papers covering a broad range of statistical themes, from prediction via correlation measures to causal inference.Paper I uses a nationally representative survey of separated Swedish women to examine whether men who strive to control their partners during their relationships are more likely to stalk or assault their ex-partners after separation. The empirical analysis shows that basic measures of control behaviors explain 18% of the variance in stalking victimization and 8% of the assault victimization. The predictive value of the constructed measures of control is shown to by far surpass those of other common risk markers included in the analysis.Paper II presents asymptotic properties of Spearman’s rank correlation for variables with finite support. For variables with finite support, the population version of Spearman’s rank correlation has been derived. Using this result, convergence to a normal distribution irrespectively of dependence is shown, and the asymptotic variance is derived. A small simulation study indicates that the asymptotic properties are of practical importance.Paper III proposes a robust version of the polychoric correlation, a parametric measure of the correlation between two unobserved continuous variables when the observed variables are discrete. Robust polychoric correlation is shown to be consistent and asymptotically normal. Results from a systematic Monte Carlo simulation suggest that the new estimator has better robustness properties than normality based maximum likelihood estimation of the polychoric correlation, to negligible efficiency costs.Paper IV reports results from a large-scale randomized experiment evaluating whether a supported employment rehabilitation intervention strategy can improve labor market opportunities for young adults on disability pension better than regular vocational rehabilitation. In total, 1,062 individuals were randomly assigned between interventions. The main result shows that 18 months after the start of the project, participants offered supported employment have work rates that are approximately 10 percentage points higher than participants who received regular rehabilitation.Paper V estimates consequences of interpersonal violence on victims. Using rich administrative population data for Sweden, individuals who visited a hospital in the years 1998 to 2002 due to assault are compared to individuals who did not have a hospital visit due to assault at that time, but who were statistically indistinguishable from the individuals of interest in the four years prior to the incident. The results suggest that violent crime has large and persistent effects on mortality, suicide, earnings, work status, disposable income, as well as on the number of days on sick leave. Specifically, an assault leading to a hospital visit is estimated to convey losses amounting to 1.4 million SEK per victimized woman and 1.5 million SEK per victimized man, whereof more than 80 percent result from excess mortality. Estimates on socio-economic outcomes show robustness against selection on unobserved characteristics. Estimates on mortality and suicides are very robust.