Zimmert, MichaelMichaelZimmert2023-04-132023-04-132021-09-20https://www.alexandria.unisg.ch/handle/20.500.14171/109968The thesis comprises three essays in the field of causal econometrics. Each of the single essays is dedicated to the question how supervised machine learning algorithms can be integrated in standard causal designs. A first, introductory chapter puts the three following main chapters into perspective. The second chapter studies the link between semiparametric efficiency theory and causal machine learning in semiparametric difference-in-differences estimation. The third chapter investigates nonparametric and machine learning methods for estimating heterogeneous treatment effects. The last chapter analyses the implications of a parental leave reform in Germany using semiparametric difference-in-differences. A major finding from Chapter 2 is that the variance lower bound that is asymptotically attainable for semiparametric difference-in-differences estimators crucially depends on the assumptions imposed on the statistical model. It turns out that lower efficiency bounds can be achieved under more restrictive assumptions. This implies a robustness-efficiency trade-off. Further, the derived efficiency results are useful for integrating machine learning methods in semiparametric difference-in-differences estimation. In particular, it is shown that moment conditions implied by the efficient influence functions lead to plug-in estimators that maintain desirable statistical properties like asymptotic normality and efficiency even when machine learning algorithms are used to estimate first stage nuisance parameters. Chapter 3 studies nonparametric heterogeneous treatment effect estimation when the researcher is interested in effect heterogeneity with respect to a limited number of covariates. However, it turns out that estimation of first stage nuisance parameters with machine learning methods remains feasible even if the dimension of the covariates needed to account for confounding is high. In particular, coupled convergence conditions between the machine learning first stage and the nonparametric second stage are derived. Additionally, it is shown that averaging over the heterogeneous treatment effects estimators leads to an estimator for the average treatment effect that attains the semiparametric efficiency bound. The finite sample performance of the different estimators is compared in various Empirical Monte Carlo experiments. Chapter finds that a parental leave reform in Germany encourages mothers to return to the labour market earlier. As intended by the reform, mothers are mostly encouraged to return in part-time. The semiparametric difference-in-differences approach used allows to estimate first stage nuisance parameters with an ensemble learner. The empirical results are robust against the inclusion of a large number of covariates that are available from administrative data. Additionally, an identification result for heterogeneous treatment effects in semiparametric difference-in-differences is derived. It turns out that mothers with higher opportunity costs of working part-time have lower incentives to return to the labour market earlier.enÖkonometrisches ModellMaschinelles LernenÖkonometrieEDIS-5162doctoral thesis