The integration of machine learning into development evaluation represents one of the most significant methodological shifts since the credibility revolution introduced randomized controlled trials to the field. Practitioners now routinely encounter causal forests, double machine learning, and predictive targeting algorithms in evaluation reports that would have seemed exotic a decade ago.

The appeal is understandable. Traditional RCTs deliver average treatment effects, but policymakers increasingly demand answers to sharper questions: which households will benefit most from cash transfers, which children respond to remedial tutoring, which farmers actually adopt improved seeds. Machine learning promises to extract these patterns from data that would overwhelm conventional subgroup analysis.

Yet the enthusiasm often outpaces the epistemology. Development evaluation operates under constraints that Silicon Valley applications rarely confront: modest sample sizes, noisy administrative data, high-dimensional heterogeneity across contexts, and consequential decisions affecting vulnerable populations. Algorithms optimized for prediction accuracy on Netflix recommendations behave differently when deployed to identify beneficiaries of maternal health programs. The question is not whether machine learning belongs in development evaluation, but rather how to deploy it without sacrificing the rigor that made experimental methods credible in the first place.

Causal forests, developed by Susan Athey and Stefan Wager, extend random forest algorithms to estimate conditional average treatment effects across covariate space. Unlike traditional interaction terms that require pre-specifying which subgroups might respond differently, these methods let the data reveal heterogeneity patterns that researchers might never have hypothesized.

The methodological appeal is substantial. Pre-registration protocols, while essential for maintaining inferential validity, often force evaluators to commit to subgroup analyses before seeing the data. Causal forests offer a principled alternative: honest sample splitting separates the discovery of heterogeneity from its estimation, preserving valid confidence intervals for the effects the algorithm identifies.

Consider a recent application to Kenya's GiveDirectly cash transfer program. Standard analysis reported robust average effects on consumption and psychological wellbeing. A causal forest analysis, however, revealed that treatment effects varied substantially with baseline food security, gender of household head, and distance to markets—dimensions that interact in ways no pre-specified model would have captured.

The technique shines when heterogeneity is genuinely multidimensional. Traditional approaches struggle beyond two or three interaction terms; causal forests handle dozens of covariates while adjusting for multiple testing through their ensemble structure. For interventions like agricultural extension, where effects depend on soil quality, market access, credit constraints, and household composition simultaneously, the method offers genuine analytical leverage.

But heterogeneity discovered is not heterogeneity explained. Causal forests identify where effects vary, not why. Without theoretical grounding, discovered patterns risk being artifacts of sampling noise, and the substantive interpretation that gives evaluation policy relevance requires domain expertise the algorithm cannot supply.

Takeaway

Machine learning can find patterns humans miss, but pattern detection is not causal understanding—the algorithm tells you where to look, not what you are seeing.

The targeting problem in development is fundamentally one of resource allocation under scarcity. When budgets cover only a fraction of eligible populations, evaluators must identify who benefits most. Machine learning approaches to targeting have proliferated: proxy means tests optimized through gradient boosting, satellite imagery classifying poverty at village level, and mobile phone metadata predicting creditworthiness for microfinance.

Recent evidence suggests these methods can improve targeting accuracy substantially. Work by Aiken and colleagues on emergency cash transfers in Togo demonstrated that phone-based algorithms outperformed geographic targeting and self-declared income measures during COVID-19, reaching poorer households more consistently than administrative rolls permitted.

Yet predictive accuracy and welfare improvement are not synonymous. A targeting algorithm optimized to predict poverty may systematically exclude populations whose deprivation manifests differently—informal workers, recent migrants, or households in surveillance-dark communities. The very features that make individuals predictable often correlate with existing state visibility, reproducing rather than correcting historical exclusions.

More subtly, predicting who will benefit most from an intervention differs from predicting who needs it most. A tutoring program might yield largest test score gains among students already close to proficiency thresholds, but concentrating resources there could widen educational inequality relative to reaching struggling learners. The choice of prediction target embeds normative commitments that algorithmic framing tends to obscure.

Rigorous evaluation of targeting mechanisms therefore requires more than out-of-sample accuracy metrics. It demands explicit welfare analysis comparing algorithmic assignment to counterfactual allocation rules, attention to who is systematically missed, and recognition that the objective function embedded in any predictive model represents a policy choice deserving democratic scrutiny.

Takeaway

Every predictive targeting algorithm contains a hidden answer to a political question: whom do we owe help, and on what grounds do we ration it?

The statistical machinery that enables machine learning's flexibility also creates its distinctive failure modes. Models with millions of parameters can memorize training data patterns that fail to generalize, and cross-validation procedures designed for prediction do not automatically protect causal estimates. In development contexts where sample sizes are modest and treatment effects small, these risks compound dangerously.

External validity poses particular challenges. A model trained on data from one district may embed environmental, cultural, or administrative particularities that fail to transfer. The literature on algorithmic fairness has documented how models trained on non-representative samples produce systematically biased predictions for underrepresented groups—a pattern likely to manifest whenever development algorithms are transported across contexts.

The black-box problem intensifies these concerns. When a linear regression produces implausible coefficients, analysts can diagnose the source. When an ensemble of gradient-boosted trees produces implausible predictions, the diagnostic path is often obscured by design. Interpretability methods like SHAP values offer partial illumination, but they describe model behavior rather than reveal underlying causal structure.

Perhaps most troubling is the temptation to substitute algorithmic sophistication for identification strategy. Machine learning cannot manufacture exogenous variation where none exists. Applying causal forests to observational data preserves whatever confounding was present in the original design; complex nonlinear adjustments for observables cannot resolve selection on unobservables that undermines the analysis at its foundation.

Responsible practice therefore requires explicit pre-analysis plans specifying which algorithms will be applied, systematic robustness checks across model specifications, and honest reporting of prediction intervals rather than point estimates. Machine learning is a complement to careful research design, never a substitute for it.

Takeaway

Algorithmic complexity can launder methodological weakness into apparent sophistication—a bad identification strategy processed through a neural network remains a bad identification strategy.

Machine learning offers development evaluation genuine new capabilities: discovering heterogeneity without pre-specification, improving targeting under budget constraints, and extracting signal from high-dimensional administrative data. These are not marginal improvements but substantive extensions of the evaluator's toolkit.

Yet the methods amplify rather than resolve the field's persistent challenges. Identification requires design; interpretation requires theory; welfare analysis requires normative commitments no algorithm supplies. Practitioners who treat machine learning as a solution to these challenges will produce sophisticated-looking evidence that fails to guide better decisions.

The productive path forward integrates algorithmic flexibility with experimental rigor and domain expertise. Pre-registration, honest sample splitting, and explicit welfare frameworks preserve the credibility that makes development evaluation worth doing. The goal is not to make our methods more impressive, but to make our programs more effective for the people they claim to serve.