Time-Series Cross-Section Data and Methods

The use of pooled time-series cross-section (PTSCS) data in quantitative political analysis has increased immensely over the last two decades. Pooled data analysis has become the standard especially in subdisciplines such as international relations, comparative politics, and comparative political economy. However, fields that use microdata, such as political behavior or American politics, also increasingly use PTSCS data due to the greater availability of survey data over time. Panel data pool cross-sectional information (number of units N) with information over time (number of time points T), for example, data on individuals or firms at different points in time, information on countries and regions over time, and so on. Thus, panel data consist of repeated observations on a number of units. We can distinguish between cross-sectional dominant data (cross-section time-series, CSTS), time-series dominant data (time-series cross section, TSCS), and pooled data with a fixed number of units and time points. The data structure has implications for the model choice since asymptotic properties of estimators for pooled data are either derived for N → ∞ or T → ∞. In addition, violations of full ideal conditions and specification issues have more or less severe effects for bias and efficiency depending on whether the number of units exceeds the number of observations over time or vice versa. In what follows, we discuss the respective strengths and weaknesses of this method and various ways by which we can cope with some of the inherent problems.

Some have argued that TSCS and CSTS data consist of observations at different points in time for fixed units of theoretical interest, such as countries or dyads, whereas in panel data, the units, mostly individuals in surveys, are of no specific interest and are randomly sampled from an underlying population with all inferences dedicated to uncovering the relationships in the population. Textbooks and articles, however, use these terms quite loosely. This entry follows this trend and discusses general estimation procedures and specification issues with respect to different kinds of data pooling cross-sectional and time-series information.

Advantages and Disadvantages of PTSCS Data Analysis

Panel data pool observations for units (i) and time periods (t). The typical data-generating process can be characterized as

with k independent variables x, which have observations for N units (i) and T periods (t). The dependent variable y is continuous (though in principle, it can be limited dependent, which requires nonlinear estimation procedures) and also observed for i and t. εit describes the error term for observations i and t and we can assume an NT × NT variance-covariance matrix Ω of the error term with the typical element E(εit, εjs). In case all Gauss-Markov assumptions are met (the error term is iid), this model can be straightforwardly estimated by ordinary least squares (OLS). Because PTSCS data combine time-series and cross-section [Page 2616]information, this is rarely the case. However, the analysis of PTSCS data offers significant advantages over the analysis of pure time-series or pure cross-sectional data. First, using pooled data increases the number of observations and, therefore, the degrees of freedom, which allow us to test more complex arguments by employing more complex estimation procedures. More important, most theories in the social sciences generate predictions over space and time, and it seems, therefore, indispensable to test these hypotheses by using data providing repeated information for theoretically interesting units. PTSCS data analysis allows the modeling of dynamics, which is impossible when pure cross sections are examined, which may lead to spurious regression results. Finally, analyzing pooled data allows controlling for unit heterogeneity beyond the inclusion of additional right-hand-side (RHS) variables. Accordingly, pooled data can be used to get rid of some kinds of omitted-variable bias, make the best of the available information, test theories that predict changes, and test theories that predict parameter heterogeneity.

...