Estimating Expected Individual Treatment Outcomes by Using Nonparametric Evaluation Methods

Consider the situation faced by an individual which has to choose among different treatments available to him. For selecting the optimal treatment he needs to conjecture for each treatment how his situation after treatment would likely to be. This article proposes a semiparametric method to estimate heterogeneous potential treatment outcomes with respect to individual characteristics. These are important in at least two respects. They can be used ex-ante to improve treatment choices and targeting of programmes to heterogeneous individuals. They also allow an ex-post examination whether each individual received the treatment most adequate to him, therewith allowing an assessment of the efficiency of the selection process. While the evaluation literature has mostly concentrated on average treatment outcomes and has promoted nonparametric techniques for their estimation, the systems proposed for targeting programmes have predominantly relied on fully parametric models. The model proposed in this article combines nonparametric and parametric elements into a GMM estimator to attain more robust estimates. Root-n consistency and asymptotic normality are shown. Furthermore, since usually more than one outcome variable is necessary to describe the after-treatment situation, different ways to summarize the numerous estimates, whose number grows quickly with the number of outcome variables and the number of treatments, are considered. Based on the asymptotic distribution of the estimated expected potential outcomes, stochastic dominance probabilities that certain treatments dominate others in one or all outcome variables are derived. Finally the proposed estimator is applied to Swedish rehabilitation programmes.