Causal Random Forests Model Using Instrumental Variable Quantile Regression

GAICIL Insights

An average treatment effect is a compromise that may describe no one.¹ Causal Random Forests Model Using Instrumental Variable Quantile Regression, by Jau-er Chen and Chen-Wei Hsiang, aims at what averages conceal: who gains, who loses and by how much. The authors propose an econometric procedure estimating quantile treatment effects nonparametrically with machine learning, building on the generalized random forests of Athey and coauthors and combining them with instrumental variable quantile regression.

The procedure does two jobs at once. It estimates the quantile treatment effect without restrictive functional forms, and it yields a measure of variable importance in terms of heterogeneity among control variables. Researchers learn how a treatment affects different parts of the outcome distribution and which characteristics drive the variation. The authors supply the practical algorithm and associated R code, lowering the barrier to adoption. A method with code attached travels much faster than a method without.

Two applications show it working. The first reinvestigates the distributional effect of 401(k) participation on net financial assets, with income, age, education and family size identified as the four most important variables. The second studies the quantile effect on earnings of participating in a job training program, where the previous year's work status and the job training service strategies emerge as important controls. The forest estimates, and it also points to where the action is.

This sits at the frontier where artificial intelligence meets causal inference. Random forests are a staple of machine learning prediction. Generalized random forests adapt them to heterogeneous causal effects. This paper extends the approach to endogenous treatment and distributional outcomes, using instruments to address selection and quantiles to capture heterogeneity. Prediction tools, retrained, become instruments of evaluation.

Variable importance is the policymaker's favorite output. Officials want to know which characteristics determine whether a program works. The procedure flags the controls that matter most for heterogeneity, guiding targeting. For 401(k)s, income, age, education and family size shape the effect on assets. For job training, prior work status and service strategies matter. A training program that knows prior work status matters can redesign its intake; a savings incentive that knows family size matters can tailor its outreach.

Open tools deserve credit. Methods matter only if researchers can use them. Supplying algorithm and code makes it easier for applied economists to adopt machine learning-based causal inference, accelerating diffusion of methodological advances across the profession.

For the economics of AI, this is machine learning reshaping policy evaluation. Flexible algorithms let researchers explore heterogeneity without specifying in advance how effects vary. Combined with instrumental variables, they produce credible causal estimates in observational settings. As administrative and digital data grow richer, such methods will matter more for understanding policies and programs.

Trade-offs remain. Nonparametric methods require substantial data, and variable importance measures need careful interpretation. Validation through simulations and replication across settings helps establish reliability. Revisiting well-studied questions, as these applications do, provides a check against established findings. A new method proves itself by reproducing what is known before claiming what is new.

On policy, the payoff is better information. Retirement saving incentives and job training are evaluated to decide whether to expand, reform or end them. Knowing how effects vary across the distribution and which characteristics drive them supports more precise decisions, and better allocation of public resources.

Pulled together, AI tools can deepen understanding of causal effects. The procedure estimates distributional effects flexibly, accounts for endogeneity and highlights the variables driving heterogeneity. Applied to retirement saving and job training, it yields insights relevant to program design. As such methods spread, empirical economics will be better equipped to answer the questions that matter most: what works, for whom and why.

One question to put to program evaluators: an evaluation that reports a single average should now be treated as incomplete. With tools like this, a ministry can ask which participants a program fails and expect an answer backed by evidence.