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

Wednesday 15th July, 16:00 - 17:30 Room: HT-101

Methods for Improving Causal Inference in the Social Sciences
Convenor	Professor Jochen Mayerl (University of Kaiserslautern, Germany )
Coordinator 1	Professor Volker Stocké (University of Kassel, GErmany)
Coordinator 2	Professor Levente Littvay (Central European University, Hungary)

Session Details

The central aim of social sciences is to discover the causal effect of explanatory variables on various outcomes. However, we are often restricted to quasi-experimental, observational and in particular cross sectional survey data. Under these conditions, the causal status of observed associations remains unclear, because of unobserved heterogeneity and reverse causation. At present the methodological tools for improving causal inference are not common knowledge.
These tools are firstly different kinds of fixed-effects estimation (e.g. school fixed-effects) which eliminates the possibly endogenous between-variance. When longitudinal panel data is available, different methods like individual fixed effects panel regression, cross-lagged autoregressive models, latent growth curve models, and autoregressive latent trajectory models are used to decide on causality. Differences-in-differences estimators can be used on the aggregate level. Secondly, matching-techniques (e.g. propensity score matching) comparing the difference of similar members of the treatment and control group are feasible as well. Thirdly, instrumental variable approaches, utilizing exogenous determinants of the treatment condition to estimate causal effects. Fourthly, regression discontinuity techniques are used in order to exploit variance at the edge between strata of the explanatory variables.
All these methods of causal analysis make certain assumptions, for instance no systematic missing data, the SUTVA-condition and unconfoundedness. In contrast to more common methods of analysis, the consequences of violations of these assumptions are much less analyzed. The same is true for decisions which have to be made when causal analyses are applied. For instance, which criterion should be used in the case of matching-techniques? How to judge the exogeneity and strength of an instrumental variable?

This session invites methodological and empirical contributions which apply methods for improving causal inference in survey research, compares them to “naive” methods or presents progress in methodological issues.

Paper Details

1. Methodological Issues in the Analysis of the Scar-Effects of Unemployment
Mr Sebastian Beil (Ruhr University Bochum)

The scar-effects of unemployment on wages or other job-related outcomes cannot be estimated using within-estimators alone: each individual that is affected by unemployment is also affected by job loss. To estimate these effects one has to rely on a difference-in-differences (did) approach, comparing the changes in the outcome for persons that lost their job but did not become unemployed to persons with a subsequent unemployment spell. This kind of comparison faces similar challenges for causal inference as a cross-sectional data analysis. The study proposes a did-matching approach to cope with this identification problem.

2. Better a Living Donkey than a Dead Lion: Identification and Estimation of the Causal Effect at the Aggregate Level, if assumptions at Individual Level are Violated
Dr Ulf Kroehne (German Institute for International Educational Research (DIPF), Frankfurt am Main (Germany))
Professor Johannes Hartig (German Institute for International Educational Research (DIPF), Frankfurt am Main (Germany))
Professor Eckhard Klieme (German Institute for International Educational Research (DIPF), Frankfurt am Main (Germany))

A re-analysis of data from the large-scale study DESI (German English Student Assessment) is presented that aims to falsify the hypothesis that bilingual instruction has a causal effect on English performance. Due to missing covariates and SUTVA violations, a (re-)formulation of the problem at school level was necessary prior to the application of propensity score methods and outcome modelling based adjustment methods. Different equally reasonable treatment effect estimates are presented as “region of uncertainty” that included zero for all considered outcomes. The interpretation of causal effects at aggregate levels is discussed for the perspective of policy makers.

3. The interrelation of immigrants’ interethnic ties and socioeconomic status in Germany. An autoregressive panel analysis.
Mr Sascha Riedel (University of Cologne)

This paper analyses the causality between interethnic ties and socioeconomic status for Italian, Turkish, and former Yugoslavian immigrants in Germany. Referring to social capital theory and its inherent problem of homophily the interrelation between these two constructs remains ambiguous. The data are made up by the German Socio-economic panel study (GSOEP). After demonstrating the drawbacks of existing empirical studies on this issue, results of fixed effects panel regressions and autoregressive cross-lagged panel models (ARM) with latent variables are presented. The final results indicate that interethnic ties positively influence the respondents’ socioeconomic status and therefore support social capital theory.