Estimating Casual Effects from Non-Experimental Data: an Application to Returns to Education
Michael Gebel, Mannheim Zentrum für Europäische Sozialforschung, Universität Mannheim
Estimating causal effects is a central aim of most quantitative empirical analysis. The lack of experimental research designs leads researchers to draw causal inferences from observational data. However, with observational data at hand, one has often to deal with the problem of non-random assignment of individual to the independent variable of interest. For example, when one tries to identify labour market returns to education there is the problem of self-selection of individuals into education based on observed as well as unobserved factors like preferences or ability. Furthermore, it is usually assumed that causal effects are constant across individuals, though observed and unobserved factors can lead to causal effect heterogeneity, i.e. there is a whole distribution of individual causal effects. Conventional least squares and instrumental variable estimators are biased in case of causal effect heterogeneity and non-random assignment of individuals due to self-selection.
To cope with these problems, econometricians and statisticians have developed methods that have found increasing acceptance in sociology. In this context, this paper tries to contribute from a theoretical and
empirical perspective to this evolving field of causal analysis in sociology by presenting treatment effect versions of selection models. First, the paper gives an overview of how to unite the recent treatment
effect literature with classical selection bias literature. Different model variants as well as their estimation procedures are discussed and compared to conventional methods and the alternative method of propensity score matching. Second, several versions of the treatment effect selection model are applied in an empirical study on returns to educational degrees in West-Germany based on data from the German Socio-Economic Panel (GSOEP).
Michael Gebel, Mannheim Zentrum für Europäische Sozialforschung, Universität Mannheim
Estimating causal effects is a central aim of most quantitative empirical analysis. The lack of experimental research designs leads researchers to draw causal inferences from observational data. However, with observational data at hand, one has often to deal with the problem of non-random assignment of individual to the independent variable of interest. For example, when one tries to identify labour market returns to education there is the problem of self-selection of individuals into education based on observed as well as unobserved factors like preferences or ability. Furthermore, it is usually assumed that causal effects are constant across individuals, though observed and unobserved factors can lead to causal effect heterogeneity, i.e. there is a whole distribution of individual causal effects. Conventional least squares and instrumental variable estimators are biased in case of causal effect heterogeneity and non-random assignment of individuals due to self-selection.
To cope with these problems, econometricians and statisticians have developed methods that have found increasing acceptance in sociology. In this context, this paper tries to contribute from a theoretical and
empirical perspective to this evolving field of causal analysis in sociology by presenting treatment effect versions of selection models. First, the paper gives an overview of how to unite the recent treatment
effect literature with classical selection bias literature. Different model variants as well as their estimation procedures are discussed and compared to conventional methods and the alternative method of propensity score matching. Second, several versions of the treatment effect selection model are applied in an empirical study on returns to educational degrees in West-Germany based on data from the German Socio-Economic Panel (GSOEP).
A sandwich will be provided if you register before Tuesday, 15:00h at aias@uva.nl.
Look at our website www.uva-aias.net for more information on the lunch seminar series.
