Panel data analysis: An introductory overview

Abstract

Our ability to manipulate large datasets rapidly on personal computers, together with the increasing availability of panel data on individuals and firms, has revolutionized empirical microeconomics. It should also revolutionize our understanding of the microeconomy and allow us to evaluate received wisdom gained from aggregate time series and micro cross-section data alike. The list of important new applied papers using panel data is too long to mention and so we provide a brief overview of panel data analysis, noting some of the advantages while warning of some potential pitfalls. Panel data refers to cross-section data that have been pooled over time where the same individual agents are followed through time. A common feature of such datasets is that, whereas the number of agents ‘N’ (firms or households) is large, the number of time periods ‘T’ is relatively short. For example, we may have in excess of 400 firms in each company cross-section but only 10 or fewer annual observations on each. A number of attractive features of such data immediately suggest themselves. Time series or temporal analysis can be explored without suffering aggregation bias common in ‘macroeconometric’ studies. Moreover, the consistency of the estimated parameters can often be proved for fixed T provided the cross-section dimension is ‘large’. Indeed, the aim of many microeconometric studies has been to estimate dynamic or temporal effects from relatively few time-series observations exploiting large N asymptotics. However, this result relies fairly heavily on the idiosyncratic nature of micro error terms and considerable care is needed in assessing the appropriateness of large N asymptotics on short panels. Nevertheless, it is clear that features common to individuals across time but which differ across them in any time period can be relatively easily eliminated, removing much of the importance of modelling permanent individual specific unobservable effects which makes analysis on single cross-sections so difficult. This is particularly true where individual specific unobservables are correlated with included regressors. Most importantly perhaps, panel data allow the analyst to exploit the large variation in circumstance of different individuals in any cross-section while still recovering temporal effects in behaviour. As a panel data illustration we can imagine investigating the relationship between taxation and investment behaviour. Clearly