Counterfactual Mean-variance Optimization
share
We study a new class of estimands in causal inference, which are the solutions to a stochastic nonlinear optimization problem that in general cannot be obtained in closed form. The optimization problem describes the counterfactual state of a system after an intervention, and the solutions represent the optimal decisions in that counterfactual state. In particular, we develop a counterfactual mean-variance optimization approach, which can be used for optimal allocation of resources after an intervention. We propose a doubly-robust nonparametric estimator for the optimal solution of the counterfactual mean-variance program. We go on to analyze rates of convergence and provide a closed-form expression for the asymptotic distribution of our estimator. Our analysis shows that the proposed estimator is robust against nuisance model misspecification, and can attain fast √(n) rates with tractable inference even when using nonparametric methods. This result is applicable to general nonlinear optimization problems subject to linear constraints whose coefficients are unknown and must be estimated. In this way, our findings contribute to the literature in optimization as well as causal inference. We further discuss the problem of calibrating our counterfactual covariance estimator to improve the finite-sample properties of our proposed optimal solution estimators. Finally, we evaluate our methods via simulation, and apply them to problems in healthcare policy and portfolio construction.
READ FULL TEXT