Panel Data Analysis

The methods of analysing pooled cross-sectional time series are applicable to a large selection of research questions and travel across disciplines. The pooled cross-sectional time series data structure has the advantage of allowing for testing of highly general theories. We will work with time series aggregated across units such as countries, firms or individuals. These data structures render data analysis more complicated since one has to consider both the time-series aspects such as dynamic effects and spatial aspects such as unit heterogeneity at the same time. The course examines the problems arising from these complex data structures and provides techniques to control and account for specific complications. We will start out by discussing characteristics and types of pooled data and underlying assumptions of basic statistical models for panel data. We then address specification problems such as complex error structures, different kinds of heterogeneity, dynamic specification issues (lag structures), spatial heterogeneity and dependency, time invariant and rarely changing variables in panel data analysis. The course will address these topics by gradually introducing more demanding concepts and techniques. The participants will be trained to select the most appropriate statistical method given the particular research question and a specific data set, to implement statistical techniques using suitable software, and to assess the empirical evidence for the argument.

Day 1

The first lecture approaches the problem of panel data from the linear regression perspective. The regression assumptions are discussed in detail while the accent is on the assumption violations with respect to unit heterogeneity and autocorrelation. The remaining time of the session will address feasible generalized least squares, clustered Huber/White/sandwich estimator and panel-corrected standard errors.

Day 2

The second session is initiated by discussing distinction between the within and between effects. Firstly, we will discuss estimation of between effects model. However, the focus of the session will be on addressing the theory of fixed effects models; particularly, on estimation of fixed effects using dummy variables and demeaning; and estimation of fixed effect in Stata, Furthermore, Hurwicz/Nickell bias is discussed, while the final segment of the class dedicated to discussion of two way fixed effects model.

Day 3

The third session will firstly address one way random effects model. The distinction between random effects and fixed effects models is elaborated in detail. In addition, we will discuss the Hausman test and the selection between random effects and fixed effects models. Furthermore, we will discuss the relevant diagnostics tests. Finally, we will progress to fixed- and random-effects linear models with an AR(1) disturbance and discuss modelling dynamics in panel data.

Day 4

In the fourth session the discussion on unit heterogeneity and dynamics is further extended by addressing random coefficients model. In this section we will address the problem of spatial autocorrelation and methodological alternatives for addressing this type of problems. Lastly, the extensions of panel data analysis to binary, multinomial and censored dependent variables are presented.

Day 5

The final session is dedicated to the more advance topics in panel data analysis. In this session we will discuss several topics including: problems and solutions for unbalanced data sets; time invariant and slowly moving variables; instrumental variable approaches and simultaneous equation models.

The participants need to be familiar with basic statistical concepts, have knowledge of linear regression model, have understanding of the basics of inferential statistics, and some understanding of  generalized least squares and generalized linear models.

The students need to have basic knowledge of Stata.