RESEARCH

Selected papers

A Matter of Timing: Effects of benefit distribution timing on consumption and downstream health outcomes.

It is well documented that individuals or households with little savings do not smooth their consumption between paychecks (or social security checks). Rather, consumption tends to peak when the check arrives and then falls until the arrival of the next check. This cyclical pattern can be particularly damaging for low-income households more likely to be food insecure. This study examines the connection between the timing Social Security and SNAP benefits and the occurrence of emergency room (ER) visits and hospital admissions for hypoglycemia, a condition that is highly sensitive to short-term changes in nutritional intake. I compare the incidence of these hypoglycemic episodes in dual-eligible recipients who receive both of their benefit checks around the same time in the last week of their benefit-month, a time where they are likely to be low on available cash and potentially run out of food, to the healthcare consumption of those who received their benefits weeks apart. This unique setup enables a novel and convincing analysis on the causal impact benefit timing on the exhaustion of food budgets and subsequent diet-related health outcomes. I use 11 years of Medicare data and programmatic details on the distribution schedule for Social Security and SNAP for 10 U.S. states. I create a continuous treatment group based on the gap between the two payment cycles based on 31 birthday groups. Due to the lack of overlap in the treatment within state, I use cardinality matching, a multivariate matching method for covariate adjustment in observational studies that finds the largest possible self-weighted samples across multiple treatments groups that are balanced relative to a covariate profile. I specify a fixed effects Poisson event-study model that evaluates the effects of this payment timing gap during the pay cycle. The regression compares utilization between beneficiaries with no gap in-between benefits and those who recieve them two weeks apart. Future work can build off of this novel design of combining the timing of these two streams of income and investigating their effects on other relevant outcomes. These findings can have further implications in the policy space of optimal benefit design in that a previously overlooked policy parameter, benefit pay frequency, may be an important design feature of social programs.

A Wrinkle in Time: Time Aggregation in Difference-in-Differences Studies

Difference in difference (DID) studies have become a popular method for evaluating the causal effect of a treatment or policy change on an outcome variable. While the "canonical" DID model contains just two time periods, “pre” and “post”, and two groups, “treatment” and “control”, most DID applications in the real-world applications, however, use multiple time periods and exploit variation across groups of units that receive treatment at different times. Researchers choose how to aggregate data from a fine time scale (e.g., daily) to a coarser time scale (e.g., yearly) before applying quasi-experimental analysis techniques. However, these choices are usually ad hoc and not guided by statistical criteria. We consider two aims: 1) to understand the evolution of the outcome over time (possibly as a prerequisite to causal inference) and 2) to estimate and perform statistical inference on a target causal quantity. We consider the optimal level of time aggregation for each aim. We consider a simple case of balanced panel data, a single treatment date, and no time-varying confounding. Then we complicate the data-generating model to include unbalanced panels, staggered treatment adoption, and time-varying treatment. We simulate panel data using flexible parametric data-generating models that produce realistic complexity and specify a variety of realistic parametric models for analysis. Across data-generating models, we vary how we aggregate the data and how we incorporate time evolution in the estimation of and inference on the treatment effect.