Programme evaluation and treatment choice (Q1865754)

Programme evaluation attempts to asses how far a programme/policy has achieved its intended aims. While treatment choice approaches policy evaluation from a different perspective. Namely, it is primarily concerned with the question of which of the available treatments a particular individual should receive. The book begins with an overview of fundamental concepts in programme evaluation and optimal treatment choice in the framework of multiple treatment evaluation and conflicting policy goals. In particular, the concepts of potential outcomes and average treatment effects are introduced. Selection bias is discussed and various non-experimental strategies to identify treatment effects are presented, which include controlling-for-confounding-variables, difference-in-difference and instrumental-variable identification. The nonparametric estimation of mean counterfactual outcomes is considered in detail and generalized matching and re-weighting estimators are discussed, including an examination of propensity score matching and semiparametric efficiency bounds. Then the results of an examination of how individuals should be allocated to programmes if outcome variables are multidimensional are presented. The existing profiling and targeting systems are reviewed and their weakness are illustrated. An alternative two-step process to derive optimal programme choices is suggested. Also, a review of nonparametric identification and the estimation of mean potential outcomes and an examination of nonparametric regression technique is given. The author continues with an investigation of the asymptotic and finite-sample properties of various evaluation estimators. The mean squared error of kernel matching, local linear matching, Seifert\&Gasser matching, re-weighting and regression-based matching are compared to the conventional pair-matching estimator when matching on an observed covariate. Since the local polynomial matching estimators require the choice of a bandwidth value, their precision gains are simulated at different values and their sensitivity to the bandwidth value is explored. As a first step towards developing a data-driven bandwidth selector, an asymptotic approximation to their mean squared error is derived and its approximation precision is evaluated. Furthermore, the applicability of cross-validation bandwidth selection is examined. The finite sample properties are also investigated when matching proceeds on an estimated variable, for example, the estimated propensity score. Finally, a semiparametric GMM estimator of conditional expected potential outcomes is developed. Asymptotic normality is shown and its finite sample properties are investigated by Monte Carlo simulation. The proposed GMM estimator is applied to analyze optimal choices among rehabilitation programmes in Sweden. In particular, the potential re-employment prospects for people with long-term illnesses are estimated and their expected employment outcomes are simulated under the optimal allocation. It is found that a substantial increase in the re-employment rate from 46\$\ improved allocation of participants to programmes.