Redesigning Policies for Lasting Impact with Causal Graphs: Evidence from Uganda's Youth Opportunities Program

Problem definition. Anti-poverty programs-cash transfers, vocational training, asset grants-aim to produce durable gains in income, health, and education for poor and underemployed populations. These gains take years, often a decade, to materialize, while the operational decisions that shape them are made on much shorter cycles. To iterate effectively, policymakers need methods that estimate long-term impact from short-term data and identify which short-term changes carry that impact, so the program can be redesigned around the mechanisms that drive durable gains. Methodology/results. We develop a framework that combines short-term experimental data with long-term observational data to estimate the long-term effect of a program and rank design modifications by their predicted long-term gain. We learn the causal mechanisms linking the treatment, intermediate (surrogate) outcomes, and long-term outcomes. These mechanisms are encoded in a causal graph, recovered using our COMB-PC algorithm. Using the learned graph, we identify a valid surrogate set, yielding a consistent estimator for the long-term effect. We then decompose the longterm effect into directed pathways using a graph-constrained structural model in an interpretable way, and rank design modifications by their predicted long-term gain. Applied to the Uganda Youth Opportunities Program (YOP), our method identifies the welfare effects that persist and those that fade, recovering the nine-year impact from only two-year surrogates which are based on two years of post-disbursement data. We find that most of the lasting effect operates through a single mechanism: enrollment in vocational training. Counterfactual analysis on the learned graph shows that strengthening this channel is expected to produce the largest long-term gains. Managerial implications. For policymakers designing the next pilot or scaleup, our framework converts short-horizon experimental evidence into actionable guidance on both whether a program's gains will persist and what design changes improve the next iteration. For YOP, the decomposition isolates a single high-leverage channel-early enrollment in vocational training-and points to a redesign built around raising enrollment: bundling the cash grant with enrollment vouchers, transportation support, or direct connections to training providers.