Development programs rarely assign benefits randomly. Budgets constrain coverage, political pressures shape targeting, and administrative rules determine eligibility. Yet within these constraints lies a methodological opportunity that development economists increasingly exploit: the regression discontinuity design. When eligibility depends on crossing an arbitrary threshold—a test score, an income level, a geographical boundary—individuals just above and below that cutoff become remarkably similar in all respects except their treatment status.

This similarity creates what methodologists call a local randomized experiment. The arbitrariness of where precisely a cutoff falls means that landing slightly above versus slightly below reflects essentially random variation. A student scoring 59 versus 61 on an exam differs by two points of measurement noise, not by fundamentally different ability or motivation. When 60 determines scholarship eligibility, we obtain sharp causal identification without ever conducting an actual randomized trial.

For development economists working in contexts where randomization proves ethically problematic, politically infeasible, or practically impossible, regression discontinuity offers rigorous causal inference from observational data. The method has revolutionized our understanding of interventions ranging from conditional cash transfers to rural electrification to educational subsidies. Understanding its logic, validity requirements, and limitations has become essential knowledge for anyone designing or evaluating development programs.

The fundamental insight underlying regression discontinuity is deceptively simple: discontinuous treatment assignment creates continuity in potential outcomes. Consider a poverty-targeting program that provides cash transfers to households below a wealth index score of 50. Households scoring 49 receive transfers; those scoring 51 do not. If we believe that underlying household characteristics—education, health, economic potential—vary smoothly with wealth scores, then households at 49 and 51 are essentially identical in all respects except their treatment status.

This logic distinguishes regression discontinuity from naive threshold comparisons. Comparing all households below 50 to all those above would confound treatment effects with systematic differences between poorer and wealthier populations. But focusing narrowly on observations near the cutoff eliminates this selection problem. The bandwidth around the threshold determines our effective comparison group, trading off bias reduction against precision loss as we narrow our focus.

Graphically, regression discontinuity manifests as a vertical jump in outcomes at the cutoff point. We model the relationship between the running variable—the score determining eligibility—and the outcome of interest on both sides of the threshold. Any discontinuous jump at the cutoff that cannot be explained by the smooth relationship between running variable and outcome represents the causal treatment effect.

The design comes in two variants with distinct identifying power. Sharp regression discontinuity applies when treatment status changes deterministically at the cutoff—everyone below receives treatment, everyone above does not. Fuzzy regression discontinuity handles partial compliance, where crossing the threshold changes the probability of treatment without determining it completely. The fuzzy design requires instrumental variables reasoning, using the threshold as an instrument for actual treatment receipt.

Critically, regression discontinuity identifies local average treatment effects—impacts for individuals near the eligibility threshold. This population may differ systematically from those far from the cutoff. A scholarship program's effect on students barely qualifying tells us little about effects for high-achieving students who would have qualified regardless. This locality limitation constrains external validity but reflects the honest scope of what the design actually identifies.

Takeaway

Regression discontinuity exploits the arbitrariness of precisely where cutoffs fall—individuals just above and below thresholds differ only by random variation in their running variable scores, creating local randomized experiments from administrative rules.

Regression discontinuity's credibility rests on assumptions that require explicit verification rather than mere assertion. The most critical is no manipulation: individuals cannot precisely control their running variable to sort above or below the cutoff. If scholarship applicants can adjust their test scores, or if bureaucrats can manipulate wealth assessments, the fundamental comparability of units near the threshold collapses. Treatment and control groups would differ systematically in their ability or willingness to game the system.

Testing for manipulation typically involves examining the density of the running variable around the cutoff. Under no manipulation, we expect the distribution to be smooth through the threshold. Bunching—excess mass of observations just above or below the cutoff—suggests strategic sorting. The McCrary density test formalizes this inspection, though visual examination of histograms often proves equally informative. Development contexts with manually recorded data face particular manipulation risks that careful researchers must address.

The second crucial assumption involves continuity of potential outcomes at the threshold. All determinants of the outcome except treatment status must evolve smoothly through the cutoff. If other policies or characteristics change discontinuously at the same threshold, we cannot isolate treatment effects from these confounding discontinuities. When poverty programs use the same wealth cutoff as other benefit eligibility, disentangling effects becomes impossible without additional identifying variation.

Specification testing helps assess this continuity assumption. Researchers examine whether predetermined covariates—characteristics determined before treatment assignment—show discontinuities at the threshold. If age, education, or baseline health jump at the cutoff, something other than random variation determines threshold position. Smooth covariate profiles across the cutoff support the design's validity; discontinuities suggest manipulation or coincident policy changes.

Finally, functional form choices influence results in ways that demand transparency. The relationship between running variable and outcome need not be linear, and misspecifying this relationship biases discontinuity estimates. Modern practice recommends local polynomial regression within optimally chosen bandwidths, along with sensitivity analyses demonstrating robustness to alternative specifications. Presenting results across multiple bandwidths and polynomial orders has become standard practice for establishing credibility.

Takeaway

Valid regression discontinuity requires demonstrating that individuals cannot manipulate their position relative to the threshold and that no other relevant factors change discontinuously at the cutoff—assumptions that demand explicit testing, not mere assertion.

Regression discontinuity has generated some of development economics' most credible impact estimates, particularly for educational interventions. Duflo, Dupas, and Kremer's evaluation of secondary school scholarships in Ghana exploited test score cutoffs determining eligibility, finding substantial effects on enrollment and completion rates. The design's credibility—verified through covariate balance tests and manipulation checks—provided more convincing evidence than observational comparisons between scholarship recipients and non-recipients could have delivered.

Rural electrification studies have particularly benefited from geographic regression discontinuity designs. When infrastructure investments follow administrative boundaries or distance thresholds from substations, households just inside and outside coverage areas become natural comparators. Dinkelman's influential study of South African electrification exploited distance-based installation rules, finding employment effects concentrated among women—a finding that would have been confounded by unobserved location characteristics under simpler identification strategies.

Conditional cash transfer evaluations have repeatedly employed regression discontinuity where poverty scores determine eligibility. Mexico's Progresa program, Colombia's Familias en Acción, and Brazil's Bolsa Família all used targeting scores with defined cutoffs. Researchers exploiting these thresholds have estimated effects on school attendance, health care utilization, and labor supply with credibility that complements evidence from the programs' original experimental evaluations.

The method has also illuminated political economy questions in development contexts. Lee's foundational work on US elections established regression discontinuity for studying incumbency effects, and development economists have extended this logic to examine how narrow electoral victories affect public goods provision, corruption, and economic outcomes across democratic contexts. The approach identifies effects of political control distinct from the characteristics that make candidates electable.

Recent methodological advances have expanded applicability to settings development economists frequently encounter. Geographic regression discontinuity handles two-dimensional boundaries rather than single running variables. Regression discontinuity with multiple cutoffs pools information across different thresholds determining the same treatment. Regression kink designs exploit discontinuities in slopes rather than levels when treatment intensity rather than treatment receipt changes at thresholds. Each extension maintains the core logic while addressing specific empirical challenges common in development policy evaluation.

Takeaway

From scholarship impacts to electrification effects to cash transfer evaluations, regression discontinuity has provided credible causal evidence for development interventions in contexts where ethical or practical constraints precluded randomized assignment.

Regression discontinuity occupies a privileged position in the identification hierarchy—below randomized experiments in internal validity but far above naive observational comparisons. Its credibility derives from transparent assumptions that researchers can directly test rather than merely invoke. When manipulation checks pass, covariates balance, and results prove robust to specification choices, the design delivers causal evidence that approaches experimental standards.

For development practitioners, regression discontinuity offers both evaluation opportunities and program design insights. Existing administrative thresholds become evaluation assets; thoughtfully placed cutoffs in new programs can build in rigorous assessment. The tension between optimal targeting and causal identification need not paralyze design—slightly arbitrary thresholds near optimal levels sacrifice little efficiency while enabling credible learning.

The method's locality limitation actually provides useful information. Knowing that a scholarship helps marginal qualifiers rather than assuming uniform effects across the ability distribution enables more sophisticated cost-effectiveness analysis. Development economics increasingly recognizes that average treatment effects obscure policy-relevant heterogeneity. Regression discontinuity's honest admission of local identification represents methodological maturity, not weakness.