In my first chapter, I extend current literature on neural networks for estimation of proportional hazards model by allowing for time-varying covariates, estimating the baseline hazard, and adding a parameter for unobserved heterogeneity, which I estimate using fixed effectes. I detail two neural network architectures for estimating this model. I also introduce an estimation technique for a flexible quantile accelerated failure time model using neural networks. Lastly, I demonstrate the flexibility of these methods on both Monte Carlo simulations and empirical data. \In my second chapter, I extend current literature about the effects of gender identity on behavior by estimating the effect of “breadwinner” stigma on unemployment duration. I leverage heterogeneity across the United States in beliefs about women's roles to estimate the impact of gender identity beliefs on the length of unemployment. For men living in states more prejudiced towards women working outside of the household, unemployment durations are negatively correlated with their wife's earnings. This result contradicts what we would expect from the wealth effects literature but is in line with a model that accounts for the behavioral effects of gender identity.
\In my third chapter, I develop a novel method, using google trends, to estimate the salience of police killing stories in different cities. I then use calls for service data from seven cities across the United States to estimate the relationship between police killing news stories and citizens’ reliance on law enforcement. I find strong evidence that news about out-of-policy killings reduces total calls for service. The evidence for in-policy stories is less straightforward, but results suggests a possible positive relationship between in-policy stories and calls for service. Demographic analysis suggests that race is an important factor in determining effect size; neighborhoods with a more significant black population experience greater changes in calls for service.