ESRA logo
Tuesday 14th July      Wednesday 15th July      Thursday 16th July      Friday 17th July     

Thursday 16th July, 09:00 - 10:30 Room: L-101

Recent developments in the analysis of panel data 2

Convenor Dr Klaus Pforr (GESIS – Leibniz-Institute for the Social Sciences )
Coordinator 1Professor Josef Brüderl (Department of Sociology, University of Munich)
Coordinator 2Dr Jette Schröder (GESIS – Leibniz-Institute for the Social Sciences)

Session Details

Panel data offer two major advantages compared to cross-sectional data:
1) They allow to identify causal effects under weaker assumptions (within estimation)
2) They allow to estimate individual trajectories over time (growth curve modeling)
Not all model classes that are available for panel data analysis, exploit these advantages fully.
There is much uncertainty amongst users, which kind of models to use. On the other side, there
are new model classes, for which it is quite unclear what the assumptions are that they need to
identify a causal effect (e.g. structural equation models for panel data, and multi-level models).
Therefore, we especially welcome papers that
1) compare different model classes and their usefulness for panel data analyses, or that
2) apply recently developed model classes and explicate their assumptions.

Paper Details

1. Applying Causal Mediation Analysis Technique in a Study of Positive Youth Development
Dr Youngjo Im (University of Chicago)

This study aims to contribute to an understanding of how development programs can be improved to better support disconnected youth ages 16 to 24, who are neither enrolled in school nor participating in the U.S. labor market. To achieve this goal, the study analyzes 14,327 individuals nationwide who participated in the Job Corps program—the nation’s largest, most comprehensive education and job training program for the disadvantaged. To uncover the underlying links that explain the program effect on youth outcomes, the study uses an innovative analytic approach—the ratio of mediator probability weighting (RMPW) method.

2. The Fixed-Effects model with Individual-Specific Slopes (FEIS)
Dr Volker Ludwig (LMU Munich)

Fixed-effects (FE) models for panel data are widely seen as the best available solution to the problem of unobserved heterogeneity. However, conventional FE estimation relies on the assumption of strict exogeneity. This assumption is violated if the causal variable of interest is related to unobserved heterogeneity regarding the outcome trajectories. Notwithstanding, the FE approach can be extended to allow for heterogeneous growth that is systematically related to the causal variable. The Fixed-Effects model with Individual-specific Slopes (FEIS) relies on a weaker form of the strict exogeneity assumption than FE, making it a useful tool for causal analysis.

3. Multinomial logistic regression with fixed effects
Dr Klaus Pforr (GESIS - Leibniz-Institute for the Social Sciences)

Fixed-effects models, which allow to control for temporally constant unobserved heterogeneity, have been implemented for many statistical software packages for continuous, dichotomous, and count-data dependent variables. We present an implementation of the multinomial logistic regression with fixed effects. Possible applications would be analyses of effects on employment status, with special consideration of part-time or irregular employment, and analyses of effects on voting behavior that implicitly control for long-time party identification rather than measuring it directly. We show an exemplaroty application on the effect of maternal smoking during pregnancy on timeliness of birth.

4. Controlling for time-varying omitted variables in panel data models: Evidence from a Monte-Carlo simulation
Mr Yusep Suparman (Universitas Padjajdaran)
Professor Henk Folmer (Groningen University)

This paper presents evidence from a Monte-Carlo study to control for time varying omitted variables by latent fixed effects regression, demeaning regression, first order differencing regression, autoregression and constrained autoregression model. The data are generated from a standard regression model with three explanatory variables and three time periods. From the regression of the log absolute bias on the log regression parameters, the individual bias and mean squared error as well as the total of the mean squared errors of the estimators of the regression coefficients of the included variables, we find that constrained autoregression is superior to the alternatives.