Empirical Methods in Labor Economics

Core Idea

Labor economics has been at the frontier of the "credibility revolution" in empirical economics — the shift from structural estimation and OLS regressions toward quasi-experimental methods that provide more credible causal identification. The key challenge is that most labor market questions involve endogeneity: does education cause higher wages, or do more able people both earn more and get more education? Do minimum wages reduce employment, or do states that raise minimum wages differ systematically from those that do not? Four identification strategies dominate modern labor economics: difference-in-differences (comparing changes in outcomes between a treated and control group, as in Card and Krueger's 1994 minimum wage study), regression discontinuity (exploiting sharp eligibility thresholds, as in studies using age cutoffs for policy eligibility), instrumental variables (using exogenous variation to isolate causal effects, as in Angrist and Krueger's 1991 use of quarter of birth to instrument for schooling), and natural experiments (exploiting policy changes, institutional features, or historical accidents as sources of quasi-random variation). These methods transformed the field from one where empirical claims were weakly identified to one with a high standard of causal evidence.

Explainer

The central challenge of empirical labor economics is the same challenge that pervades all empirical social science: establishing causation rather than mere correlation. Does an additional year of schooling cause wages to rise by 8%, or do people who would have earned high wages anyway also happen to get more education? Does the minimum wage reduce employment, or do unobserved economic conditions confound the relationship? These questions cannot be answered by running a regression of wages on schooling, because the regression coefficient conflates the causal effect of schooling with the selection effect of who gets more schooling. The credibility revolution was the field's collective response to this identification problem.

Difference-in-differences (DiD) is perhaps the most widely used quasi-experimental method in labor economics. The core idea is simple: compare the change in an outcome for a group affected by a policy change (treatment group) to the change for a group not affected (control group). The difference in these differences removes time-invariant unobserved factors (which are differenced out within each group) and common time trends (which cancel in the between-group comparison). Card and Krueger's 1994 minimum wage study is the canonical example: they compared employment changes in New Jersey fast-food restaurants before and after a minimum wage increase to changes in neighboring Pennsylvania restaurants over the same period. The critical assumption is parallel trends — that absent the treatment, the two groups would have experienced similar changes. DiD has been extended to staggered adoption settings (where different states adopt policies at different times), to synthetic control methods (which construct a weighted comparison group), and to event-study designs that allow visual assessment of pre-treatment trends.

Regression discontinuity (RD) exploits sharp eligibility thresholds to identify causal effects. When a policy assigns treatment based on whether a running variable crosses a cutoff — financial aid eligibility at a GPA threshold, retirement benefits at age 62, unemployment insurance extension at a duration cutoff — units just above and just below the cutoff are nearly identical except for treatment status, creating a local quasi-experiment. RD designs are compelling because the identification is visual: if the outcome shows a discrete jump at the cutoff, the effect is apparent in a simple graph. In labor economics, RD has been used to study the effects of unemployment insurance duration on job search (using duration cutoffs), the impact of disability insurance on labor force participation (using age-based eligibility rules), and the returns to attending elite universities (using admission score thresholds).

Instrumental variables (IV) address endogeneity by finding a source of variation in the endogenous variable that is uncorrelated with the error term. The instrument must be relevant (correlated with the endogenous variable) and valid (affecting the outcome only through the endogenous variable — the exclusion restriction). Angrist and Krueger's (1991) quarter-of-birth instrument for schooling illustrates the logic: compulsory schooling laws interact with school-entry cutoff dates to create exogenous variation in how much schooling people get. Those born in Q1 can drop out with slightly less schooling than those born in Q4, and this birth-timing variation is plausibly unrelated to ability or other earnings determinants. IV estimates identify a local average treatment effect (LATE) — the causal effect for "compliers" whose behavior is changed by the instrument — which may differ from the average treatment effect in the population.

Natural experiments is the broader category encompassing any situation where institutional features, policy changes, or historical accidents create quasi-random variation that can be exploited for causal inference. The Vietnam draft lottery (random draft numbers used to study the effect of military service on earnings), German reunification (an exogenous shock to labor markets used to study convergence), and immigration shocks from political events (the Mariel boatlift used by Card to study the effect of immigration on native wages) are all natural experiments that provided credible identification for questions that seemed intractable with observational data alone. The common thread is opportunistic identification: rather than designing an experiment, the researcher recognizes that history or institutions have created one.