Working Papers

Demand for Quality in the Presence of Information Frictions: Evidence from the Nursing Home Market (Reject and Resubmit, AER)

This paper studies consumers' demand for quality in the nursing home market, where information frictions are a potential concern. Using administrative data on the universe of nursing home residents, I estimate substantial variation in quality of nursing homes in California. Yet, structural demand estimates reveal that average demand for quality is very low, with residents who were younger, highly educated, free from dementia, and who made their choices after the introduction of the star rating system being more responsive to quality. Counterfactual simulations suggest that eliminating information frictions may reduce deaths by at least 8 to 28 percent.

Why Are Minorities Disproportionately Concentrated in Low-Quality Nursing Homes?

This paper studies the underlying mechanisms giving rise to high levels of racial segregation and disparities across nursing homes in the US. Descriptively, I find that while residential segregation is an important explanation for racial segregation across nursing homes, it struggles to explain disparities. I provide reduced form evidence for several other explanations for disparities: nursing homes seem to discriminate against minorities in their admission practices, individuals tend to choose nursing homes with a higher share of residents of their own race, and minorities seem less sensitive to changes in nursing home quality. To disentangle and quantify the effects of these proposed mechanisms, I then estimate a structural model. Counterfactual simulations confirm that residential segregation is indeed the main explanation for high levels of segregation, but it is not always the main driving force behind disparities -- in some states, heterogeneity in sensitivity to quality and discrimination by nursing homes play a larger role in explaining disparities.

Estimation of Regression Discontinuity and Kink Designs with Multiple Running Variables [R Code]

In regression discontinuity designs with multiple running variables (MRD), units are assigned different treatments based on whether their values on several observed running variables exceed known thresholds. In such designs, applied work commonly analyzes each running variable separately, estimating a single-dimensional RD design in the first running variable after limiting the sample to the set of individuals qualifying on the second threshold, and vice versa. In this paper, I propose a new estimator for MRD designs using thin plate splines that improves upon the applied practice in two ways. First, the estimator can be used to estimate the conditional average treatment effect at every point on the boundary separating treated and untreated units, and second, it provides efficiency gains by using the entire sample. I also develop analogous estimators for multidimensional regression kink (MRK) and multidimensional regression discontinuity/kink designs (MRDK). I show consistency of these estimators and construct asymptotically valid confidence intervals (CIs), before presenting simulation results showing that they produce estimates and CIs that perform well in finite samples. Finally, I demonstrate the performance of my MRD estimator with two empirical applications: Londoño-Vélez, Rodríguez, and Sánchez (2020) on the effect of financial aid on college enrollment, and Keele and Titiunik (2015) on the effect of political ads on election turnout. R code for estimation and inference will soon be available.

Selection on Unobservables in Discrete Choice Models

Selection on unobservables is an important concern for causal inference in observational studies, and accordingly, previous papers have developed methods for sensitivity analysis for OLS, binary choice models, instrumental variables, and movers designs. In this paper, I develop methods for sensitivity analysis for a setting that has not been previously studied — discrete choice models. In particular, I derive bounds for the omitted variables bias under an assumption about how much the consumer values the omitted variable(s) relative to the included control variables, and about the relationship between the omitted variable and the variable of interest. I provide theoretical results for my bounding procedure, and demonstrate its performance in simulations. Finally, I show in several empirical applications that my procedure produces economically meaningful bounds.

Research in Progress

Peer Effects in the Workplace: Evidence from the Illinois Workplace Wellness Study (with Damon Jones, David Molitor, and Julian Reif)

This paper studies peer effects in health behavior in a workplace setting. To overcome well-known identification challenges faced by studies on peer effects, we design and implement a workplace wellness program for nearly 5,000 employees of the University of Illinois. We collect information on employees’ coworker networks, and randomize eligibility and financial incentives for the wellness program at the individual level. We find that the probability an individual participates in health screenings increases by 3.9 percentage points for each additional coworker randomized into treatment. By contrast, estimates from OLS regressions on individual participation on the leave-out mean of peer participation are much larger, highlighting the importance of using experimental variation to obtain credible estimates of peer effects. Our results are robust to state-of-the-art inference procedures for network settings.

Does Your Roommate Matter for your Health? Evidence from the Nursing Home Market (with Martin Hackmann)

Challenges in Measuring Mental Health Trends

Public awareness of mental health issues has grown in recent years, and there is a common perception that mental health in the population is worsening, concerns that are supported by descriptive evidence. However, changes in public attitudes towards mental health make the interpretation of these trends challenging: low response rates to mental health surveys make their results sensitive to changes in sample selection bias, and even diagnosis rates in comprehensive data sets such as the Medicare data are a function of individuals' willingness to seek professional help. In this paper, I address these measurement issues using a comprehensive data set on all nursing home residents, so it does not suffer from sample selection bias. Similar to Medicare data, it contains information on whether each resident was diagnosed with depression in the recent past, but in addition, it contains a rich set of psychosocial measures, which provides us with a detailed picture of the resident's underlying mental health. I find that while depression diagnoses at admission increased from 19 to 24 percent for residents admitted to a nursing home in California between 2000 and 2010, underlying mental health of these residents (based on observed psychosocial behavior as well as machine-learning predictions using hundreds of covariates) was roughly constant over the same period. These results illustrate the perils of inferring mental health trends from survey evidence or diagnosis trends alone. 

Past Work

Cigarette Consumption and Tax Salience
This paper studies how cigarette consumption responds over time to changes in tax rates. Using a panel of state data, I estimate that the cumulative effect of an excise tax rise on consumption is larger than the cumulative effect of an increase in sales tax, in line with a theory of tax salience. In addition, I find that consumption falls in advance of an excise tax hike, whereas it only falls in the year after a sales tax increase. The pattern of consumption response to sales taxes is also consistent with consumer learning over time.

Does Fake News Affect Voting Behavior? An Instrumental Variable Approach Using Big College Football Games

The issue of fake news has been hotly debated in recent years, with some commentators claiming that it played a role in US presidential elections and the Brexit vote. Despite these claims, there has been limited evidence to date linking fake news directly to voting behavior. In this project, I seek to provide credible evidence on this question by using big college football games as an instrument for fake news consumption. I find that search volumes for pro-Trump fake news terms were lower in counties close to college football teams that played a big game shortly before the election, and also that these counties were less likely to vote for Trump. The magnitude of these estimates suggest that a one-standard deviation increase in search volume for pro-Trump fake news terms increased Trump’s vote share by about 4.5–9 percentage points. Finally, I do not find evidence that fake news affected overall turnout rates, or that fake news resulted in down-ballot effects.