Recent developments in the analysis of panel data 1
|Convenor||Dr Klaus Pforr (GESIS – Leibniz-Institute for the Social Sciences )|
|Coordinator 1||Professor Josef Brüderl (Department of Sociology, University of Munich)|
|Coordinator 2||Dr Jette Schröder (GESIS – Leibniz-Institute for the Social Sciences)|
This presentation compares the technical and substantial differences of logisitic panel regression and discrete event history analysis in the social sciences. From these differences, we derive certain strategies for empirical practioners to make accurate and functional decions between these two methods in different research situations. We illustrate our suggestions by a research example based on the Socio-Economic Panel Study (SOEP), analyzing the impact of employment status on cohabiting in early adulthood. We perform this analysis in both analytical frameworks and discuss the surprising differences in results.
Many social scientists turn to panel data for resolving questions of causal ordering. But conventional panel models rely on the assumption of strict exogeneity, which rules out reverse causality. While this assumption can be relaxed in some circumstances, alternative model may also yield biased estimates if the lags between panel waves do not match the actual causal lags of the process under study. Using a substantive example with real data, I compare results obtained by different panel models and discuss the plausibility of the underlying assumptions of these models.
Estimation of the effect of parenthood on real-life outcomes, such a life satisfaction, is challenging due to (1) selection, (2) heterogeneity of the effects, and (3) the outcomes changing with the ages of children. This paper uses 14 waves of the Swiss Household Panel and fixed-effect models to estimate the effect of parenthood on life satisfaction. We discuss the consequences of methodological choices involved, including the choice of the sample, coding of independent variable, choice of the reference category, and functional form of relationship.
This work develops a hidden Markov latent variable model for analyzing multivariate longitudinal data. To reveal the dynamic patterns and possible heterogeneity of the associations and interrelationships among longitudinal observed and latent variables, a mixed hidden Markov model is introduced to model the transition probabilities across different latent states. We develop sound statistical methods to perform parameter estimation, hypothesis testing, and model selection. We conduct simulation studies to assess the empirical performance of the proposed methodologies and apply the model to a longitudinal study of cocaine use. Important insights into effective prevention of cocaine use are obtained.