UE641 - Econometrics 3: Discrete Models and Panel

Intervenant·e·s

• Luc Behaghel [référent·e]   directeur de recherche, INRAE / Paris School of Economics (PJSE)
• Philipp Ketz   assistant professor/maître assistant, Paris-Jourdan sciences économiques

The course covers four broad topics.

After a summary of the traditional approach to causality in cross-sectional linear models (lecture 1), lectures 2-6 present the "treatment effect" or "program evaluation" approach to causality. In lecture 2, we present the treatment effect model, also known as Rubin’s model, that is the common framework used in this approach, and apply it to the analysis of randomized controlled experiments. In lecture 3, we cover advanced issues with instrumental variables, and their use to analyze quasi-experiments. In lecture 4, we analyze regression discontinuity designs. In lecture 5, we cover matching estimators, and in lecture 6, synthetic controls.

Lectures 7 and 8 deals with panel data. We consider them from two perspectives: endogeneity and dynamics. One advantage of panel data over cross-sections is indeed to offer new ways to deal with endogeneity. We present simple models that account for the presence of permanent differences across units (individual effects, lecture 7). We then discuss how instrumental variables can be used in that context. To that end, we introduce a general class of estimator that uses the "generalized method of moments" (GMM) (lecture 8).

Lectures 9-12 cover Maximum Likelihood (ML) estimation and its main applications in applied economics. First, the concept of ML is introduced together with its large sample justification [lecture 9]. Then, we discuss several models which are frequently used in economics and estimated by means of ML [lecture 10-11]. A broad class of models is given by limited dependent variable models. A prominent example is the binary choice model. In this context, we contrast ML estimation with linear regression models that ignore the nature of the binary choice variable. Other examples of limited dependent variable models are (multivariate) discrete choice, censored regression, and duration models. We discuss estimation of these models along with several testing problems of interest, such as model specification. Furthermore, we discuss how ML estimation can be used in the context of sample selection issues, that is when the estimation sample is not representative of the population of interest [lecture 11]. In addition, we discuss alternative, less “parametric” solutions to the problem of sample selection. Last, we discuss an empirical application to illustrate the usage of some of the newly introduced estimation methods used in practice [lecture 12].