Essays on Maternity Leave Policy and Educational Experiments

This dissertation contains two essays on the impact of maternity leave policy and one chapter illustrating the advantage of a rarely used but powerful experimental design. In chapter one, my co-authors and I exploit a sudden expansion of paid leave from 6 to 12 weeks in the United States military to assess impacts on maternal health. Using detailed medical records covering all medical care by mothers in the year following birth, I assess the impact of additional maternity leave on maternal health under multiple designs. I find that the additional leave decreases the likelihood of postpartum depression diagnosis. I provide some evidence that expanding leave reduced mothers' pain and health care utilization but these results are more sensitive to specification choice. These results provide some of the best and most comprehensive evidence of the impacts of paid maternity leave on maternal health.

Chapter two utilizes the same policy variation as chapter one to study impacts of additional paid leave on leave taking, continued employment, and promotion of mothers. Administrative records enable estimation under multiple designs. I find that the policy increases leave taking by 5 weeks, has minimal impacts on continued employment, and significant negative impacts on promotion. Features of the military mean that results are absent return to work and income effects. As a result, my estimates provide evidence which can help reconcile estimates of positive short run and negative long run impacts found in past research.

Chapter 3 builds on my existing published work regarding the impacts of the flipped classroom and the design of educational experiments. When educational interventions are studied using randomization of treatment by classroom and semester, power requires many classrooms and rigorous evaluation is costly. I use placebo treatments and Monte Carlo simulations to evaluate precision under randomization at the classroom by lesson, an alternative discussed in (Wozny et al., 2018). I find that standard errors under randomization by classroom and lesson are one-fourth the size and have less bias than those under randomization by classroom and semester. Improvements are robust to changes in key parameters and of violations of identifying assumptions.