Panel Data Analysis

While there should be a large overlap between this course and the ones offered by various universities in the world on panel data methods, this course places an emphasis on the connections between these methods and causal inference, which is the primary goal of social science research. Thus, the course begins with an overview of causal inference in the context of observational studies. We will quickly survey the idea of causality and concepts such as average treatment effect (ATE). Importantly, we will draw the connections between these concepts and regression analysis.

In the second half of Day 1, we will look at simple panel data methods. Specifically, we make a distinction between “independently pooled across section” data and “panel” data, with the latter being the focus of this course. We shall see that panel data differs in some important respects from an independently pooled cross section in that a panel consists of the same entities (individuals, firms, countries or whatever) across time. We will then study two basic panel data methods, namely two-period panel data analysis and first differencing.

On Day 2, we focus on slightly more advanced methods for estimating the unobserved effects in the context of panel data analysis. we introduce the fixed effects estimator, which, like first differencing, uses a transformation to remove the unobserved effect prior to estimation. As a result, any time-constant explanatory variables are removed in the process. In contrast to the fixed effect estimator, we also introduce the random effect estimator, which looks attractive when we think the unobserved effect is uncorrelated with all the explanatory variables. For example, when we have good knowledge about the factors affecting the dependent variable and have controlled for these factors in the equation, random effect estimator can sometimes be the preferred strategy. With these foundations, we will then look into a relatively new correlated random effects approach, which provides a synthesis of fixed effects and random effects methods and has been shown to be practically very useful.

Day 3 begins with a lab session, whereby we put into practice the methods that has been introduced in the previous two days. We shall carry out these analyses in Stata and learn about the interpretation of the results. This is also a good opportunity for students who have brought their own data to carry out the analysis for their current research projects. In the second session of the day, we shall come back to the classroom to learn more about the research designs and methods for which panel data can be applied to. Depending on the time available, we will learn about match-pairs samples, cluster samples and the combination of panel data methods with instrumental variable (IV) method. Again, we emphasize the connection between these methods and causal inference.

The first session on Day 4 continues with the discussion on applying panel data methods to other data structures and model specifications. Specifically, we study panel-corrected standard error (PCSE) estimates and dynamic panel methods (Arellano-Bond estimator). For these more advanced methods, we shall not be looking into the technical details but rely more on an intuitive understanding of the challenges these methods address. In the second session, we return to the lab and put these more advanced methods into practice. We will learn how to carry out panel IV analysis, obtain the PCSE and Arellano-Bond estimators and calculate cluster robust standard errors. Again, this is a good opportunity for students to carry out the analyses using their own datasets and check robustness of the results across various model specifications.

Day 5 consists of two seminar sessions in which students will present on their research or research proposals that employ any panel data methods and receive feedback from the instructor and their peers.

Note from the Academic Convenors to prospective participants: by registering to this course, you certify that you possess the prerequisite knowledge that is requested to be able to follow this course. The instructor will not teach again these prerequisite items. If you doubt whether you possess that knowledge to a sufficient extent, we suggest you contact the instructor before you proceed to your registration.

This course builds on ordinary least square (OLS) regression and extends it to data with a time-series-cross-sectional structure. Participants should be familiar with basic statistical concepts such as sample mean and sample variance as well as their properties. In addition, participants should possess basic knowledge of regression analysis (up to multiple regression). Basic skills in Stata are also necessary. Participants who do not meet the above criteria are strongly encouraged to take other relevant courses offered by the Summer School, either before or concurrently with this course.