We use a regression discontinuity design that compares narrowly-elected Democrats and Republicans across congressional districts from the 51st to 116th Congress to identify the causal effect of party on congressional action on immigration. We measure immigration positions through two channels: roll call votes on legislation that concerns immigration and the sentiment tone of immigration-related floor speeches from Card et al. 2022. Our results reveal substantial and growing partisan polarization on immigration policy. Democrats are 12.5 to 15 percentage points more likely than Republicans to vote in favor of pro-immigration legislation, with this gap widening significantly over time.

Existing policy encourages married couples to engage in joint-decision making when allocating time among income-earning, household, and leisure activities. Often, it is more efficient for one partner to specialize in household tasks, while the other specializes in earning. Disproportionately, women in heterosexual couples fall into the former camp, taking on the role of secondary earners and engaging in part-time work more often than their male counterparts. This secondary earner status disadvantages divorced women. A feature of the Social Security system tries to insure against this: the spousal benefit. How effective is this policy? We use a regression discontinuity design that exploits the discontinuous feature of the spousal benefit: eligibility is contingent on the marriage lasting must be at least 10 years. Previous research, confirmed by our own preliminary analyses, shows no evidence of manipulation behavior around the cutoff. We thus can compare women around the cutoff to measure the impact of the spousal benefit on financial and mental well-being.

We present a workflow to digitize hundreds of thousands of pages of dense, loosely structured documents using large language models (LLMs). Using three case studies, we show that LLMs can approach "gold-standard" human digitization accuracy at a fraction of the cost and compare well against other alternatives. But using LLMs comes with a significant learning curve, budget uncertainty, and the risk of producing incorrect and difficult-to-check output. We provide a crash course in production-scale LLMs, targeted to a tech-savvy social scientist reader who has used ChatGPT or other web-based AI platforms but who has not yet explored the capabilities of AI APIs. We discuss how to limit risk, manage API budgets, and incorporate human review into an AI workflow.