Punishment and Reoffending: A Probationary Selection Model with Judge Randomization (Job Market Paper) Draft
Abstract: Understanding the impact of punishment on repeat offenses is vital, particularly for assessing its effects on first-time offenders and how their subsequent criminal behavior evolves. Despite extensive research on incarceration, most crimes qualify for probation, placing significant importance on judges' probationary decisions. This study develops a model to analyze judicial decisions, including probation decisions, probation lengths, and sentence lengths, using data from North Carolina new offenders (1995-2010). It reveals that overlooking the probation selection process leads to substantial estimation biases. The findings suggest that longer probation increases reoffending rates, whereas active and suspended sentences have a marginal deterrent effect. The research highlights the influence of judicial discretion in sentencing and shows that more skilled judges, often more lenient, are likely to reduce reoffending rates, in contrast to adherence to strict sentencing guidelines.
Abstract: The sparsity assumption is crucial for estimation and inference in the presence of high-dimensional instruments and controls. However, the sparsity assumptions for both the instrumental and control variables are unverifiable and often violated in economic analysis. In this work, we propose a novel ridge regularization-based treatment effect inference method in the presence of many instrumental variables and control variables without sparsity assumption. To overcome the many instrument biases, we suggest a refitted cross-validation strategy in the construction of the optimal instrument. The proposed estimator outperforms existing sparsity-based approaches in a variety of settings and is applicable to both low- and high-dimensional models, with and without sparsity assumptions. The asymptotic superiority and the limited sample performance are validated with theoretical proof and numerical analysis. This work also contributes to the literature on many instrumental variables by allowing high-dimensional controls. The implementation of the proposed method is available in the R package "HDRRTreat".
Abstract: Pre-exposure prophylaxis, or PrEP, is a daily medicine for people at high risk of contracting HIV to lower their chance of getting infected. Clinical studies show that PrEP can be highly effective, reducing the risk of contracting HIV through sex by more than 99%. Despite the great effectiveness of PrEP in preventing HIV infection, there is concern that PrEP may lead some people to more risky behaviors, such as reducing their use of condoms and increasing their number of sexual partners. Using public data from the Multicenter AIDS Cohort Study, this paper empirically examines such concerns. Exploiting the panel nature of the data, we use a difference-in-difference approach to study the effect of PrEP on risky sexual behaviors and other sexually transmitted diseases (STDs). Our analysis shows that PrEP makes individuals more likely to engage in unprotected sex, have more sexual partners, and more likely to contract syphilis.
Social Networks with Unobserved Links in Panel Data Draft
Abstract: I consider the problem of quantifying the social network effects in a panel data setting where the network links are unobserved. The unobserved social network links are a major difficulty in many empirical works. I propose a methodology to identify and estimate social network models using panel data. I consider two cases, including the network links are fixed or changed over time. This method is useful when network link formation is unknown and may change over time. This method also helps to control the unobserved individual fixed effects. Then, I apply the estimator to the peer effect in the Student/Teacher Achievement Ratio (STAR) Project. Without observing the latent network, I identify and estimate peer effects on students' achievement, showing that peer effects tend to be larger in smaller classes.