🧭 What this method does:
New methods are presented to estimate causal effects retrospectively from microdata by harnessing a machine learning ensemble. The approach targets a clearly defined "retrospective intervention effect" founded on hypothetical population interventions, so the causal comparisons are specified rather than ambiguous.
🔑 Problems this approach fixes:
- Conventional regression or matching can leave the relevant retrospective counterfactual unclear; this method pins down the counterfactual as a hypothetical policy intervention.
- Standard models risk misspecification, overfitting, and inefficient or biased use of many identifying covariates; the ensemble lets the data guide how to use a large covariate set in a controlled way.
📊 How the machine learning ensemble is used:
- Uses an ensemble of algorithms to combine predictive information without relying on a single parametric specification.
- Controls model complexity and guards against overfitting while exploiting rich microdata and many covariates.
🔍 Illustration and empirical application:
- The method is illustrated with an analysis of policy options aimed at reducing ex-combatant recidivism in Colombia.
- The illustration shows how retrospective intervention effects can be estimated from observational microdata when guided by an ensemble approach.
⚖️ Why it matters:
- Provides a transparent, well-defined target for retrospective causal claims about policy effects.
- Offers a practical way to leverage large covariate sets without relying solely on potentially misspecified parametric models, improving observational policy evaluation in contexts like post-conflict reintegration.
