Hi! I'm Dylan. I'm on the 2026-27 job market.
I study labor and econometrics.
Job Market Paper
-
[pdf]Economists disagree about the aggregate effects of the minimum wage and what they mean for basic economic theory. But, there is a mismatch between empirical evidence from policies that apply to entire markets and the textbook monopsony model that describes a single firm (Robinson, 1933). In this paper, we use new firm-level micro data on employment, wages, sales, and prices to study how low-wage employers adjust to minimum wage hikes. We find that firms paying just below the new minimum wage increase employment and lower prices relative to competitors, while firms already paying the old minimum do the opposite. This is new evidence that minimum wages increase employment and lower prices up to a threshold, above which the effects flip signs. Firms that are not directly affected also adjust because they compete with directly affected firms for both workers and customers. We interpret these results in a model of heterogeneous firms with both labor and product market power. We develop a general method to recover the model parameters from the difference-in-differences variation and find that unconstrained markups and markdowns are large. The market-level effects of the minimum wage depend on both the nature of firm-level heterogeneity and market-level spillovers. Spillovers, not reductions in firm-level monopsony markdowns, drive the bulk of employment and wage changes under the minimum wage policies we study.
Publications
- The Review of Economics and Statistics, 2026. [pdf] [package] [github] [ReStat]Standard estimators in information provision experiments place more weight on individuals who update their beliefs more in response to new information. This paper shows that, in practice, these individuals who update the most have the weakest causal effects of beliefs on outcomes. Standard estimators therefore understate these causal effects. I propose an alternative local least squares (LLS) estimator that recovers a representative unweighted average effect in a broad class of learning rate models that generalize Bayesian updating. In five of six recent studies, estimates of the effects of beliefs on outcomes increase. In two, they more than double.
-
Journal of Policy Analysis and Management, 2022
[pdf] [JPAM]The COVID-19 pandemic led to dramatic economic disruptions, including government-imposed restrictions that temporarily shuttered millions of American businesses. We use a nationwide survey of thousands of small business owners to establish three main facts about business owners' decisions to reopen at the end of the lockdowns. First, roughly 60 percent of firms planned to reopen within days of the end of legal restrictions, suggesting that the lockdowns were generally binding for businesses—although nearly 30 percent expected to delay their reopening by at least a month. Second, decisions to delay reopenings did not seem to be driven by concerns about employee or customer health; even businesses in high-proximity sectors with the highest health risks generally reported intentions to reopen as soon as regulations allowed. Third, pessimistic demand projections primarily explain delays among firms that could legally reopen. Owners expected demand to be one-third lower than before the crisis throughout the pandemic. Using experimentally induced shocks to perceived demand, we find that a 10 percent decline in expected demand results in a 1.5 percentage point (8 percent) increase in the likelihood that firms expected to remain closed for at least one month after being legally able to open. We use follow-up surveys to cross-validate expectations with realized outcomes. Overall, our results suggest that governments set more stringent guidelines for reopening than what many businesses would have selected, suggesting that governments may have internalized costs of contagion that businesses did not.
Working Papers
-
Updated October 2026. Under Review. [pdf]Using point-of-sale and payroll data, we estimate the direct tax savings, behavioral tipping response, and incidence of the 2025 No Tax on Tips (NTOT) provision. The average tipped worker directly saves $531, but these savings are concentrated: 26% earn too little to benefit. Tip revenue rose 3.1% post-enactment as firms newly enabled tipping. Workers kept 72 cents of each new tip dollar, while firms lowered base pay by the remaining 28 cents. Overall, NTOT raised the average tipped worker’s after-tax earnings by approximately 3.3%. The new tips are a raise for workers and, partly, a wage subsidy to firms.
-
[pdf]Tipping is large and growing, but we know little about its consequences for workers and firms. We link point-of-sale and payroll records at 20,000 U.S. firms, which lets us observe tips from customer payment to worker earnings. We find that tips are a nearly constant share of revenue, which means hourly pay moves with revenue. Firms that begin collecting tips claw back between 23 and 33 cents of each dollar of tips through lower base wages. Tips reward effort: workers at tipping firms earn more by serving more customers. We show this with demand shocks constructed from customer spending at other businesses. A demand shock raises revenue nearly twice as much at tipping firms as at non-tipping firms, while base wages and prices do not respond. Output per hour rises more despite a larger increase in hours. Effort drives the gap because each worker is more productive, so the firm expands and recovers about a third of the tips. Yet not every firm adopts tipping: the gain is smaller where customers tip less. A model calibrated to the demand-shock estimates implies that tipping benefits workers and firms, raising job value by about half and firm profit by about 30 percent in firms where customers are prone to tip. These gains may explain why tipping has recently grown in many industries.
Work in Progress
-
[draft coming soon]We use novel administrative data covering thousands of small businesses to show that labor market frictions allow firms to impose an hours markdown and assign workers fewer hours than they would prefer at their current wage. When full-time workers leave a firm, their part-time coworkers immediately increase their hours at the same wage rate. Exploiting variation in schedule substitution patterns, we estimate that additional hours significantly reduce worker separations, revealing that workers prefer longer hours at their current wage. We interpret these findings in a simple model where search frictions allow (and incentivize) firms to maintain larger workforces as insurance against future separations, assigning fewer hours per worker. Our estimates imply that the median firm employs 25% more workers than the competitive benchmark, with an hours markdown of 20%.
- [draft in preparation]A large literature develops efficient estimators that place more weight on observations with greater identifying variation (i.e., those that are more "exposed"). However, when treatment effects are heterogeneous, these estimators generally do not recover representative average effects. This paper characterizes the tradeoff between precision and representativeness and shows how to construct instruments that minimize the mean squared error (MSE) of a representative average effect. An important application is to formula instruments, such as shift-share instruments, where treatment effects may be correlated with exposure to treatment. I derive optimal weights for formula instruments with heterogeneous treatment effects and construct a feasible shrinkage estimator. In an application to Chinese high-speed rail, the efficient formula-IV estimate is 0.09, compared with a direct average effect estimate of 0.62. The MSE-optimal instrument recovers an estimate of 0.56 with a standard error 30 percent smaller than the direct average estimator.
-
September 2025. [pdf]Heterogeneity robust difference-in-differences methods typically require control units that remain untreated throughout the entire post-treatment window. This prevents the identification of long-run effects when researchers observe fewer pre-treatment periods than post-treatment periods. We show that cohort-stacked estimators identify long-run effects by chaining together successive not-yet-treated controls. This approach uses overlapping cohorts to extend identification under standard common trends assumptions. We apply this to the earnings effects of parenthood. In a setting where direct methods identify effects only four years post-birth, chaining extends identification to eight years.