Symposium on Causality 2012 Jena University
Program & Videos Material related to the contributions About Jena Sponsors Links Imprint / Contact

Courses of our department Symposium on Causality 2006 Symposium on Causality 2008 Symposium on Causality 2010 SMABS/EAM 2012
SMABS 2004 EAM Causal Effects

Department of methodology and evaluation research

Jena University

Home

Symposium "Conditional and Direct Causal Effects"

Schloss Dornburg near Jena (Photograph: Scheere)

Place: Altes Schloss Dornburg near Jena, Großer Kaisersaal

Time: July 24 to 26, 2012

Language: english

Aim of the symposium

The symposium is intended to sharpen our ideas and instruments for the analysis of conditional and direct treatment effects. Why does this aim seem to be warranted? In the last decades, analyses of causal effects have mainly focussed on average total effects of treatments. This is appropriate if we are interested in the overall or total effectiveness of treatments in a population of subjects. However, conditional and direct treatment (intervention, exposition) effects are more informative and scientifically more interesting (MacKinnon, 2008).

With conditional total effects we consider the effects of a treatment on an outcome variable given a value of one or more covariates. Such a covariate can be sex, diagnostic group, severity of symptoms before treatment, motivation for treatment, or a multivariate variable consisting of several of such one-dimensional variables. Such conditional total effects can differ from each other between different values of the covariates. Hence, with conditional total effects we can quantify the total treatment effects for different (groups of) subjects. How big are the total effects for which kind of subjects? Note that total treatment effects may be positive for some and negative for others. Hence, in clinical research, conditional total effects deal with differential indication of a treatment.

Average and conditional total effects deal with the total effects of a treatment. In contrast, with direct effects we ask for the effects of a treatment that are not transmitted by intermediate variables that, in the process considered, occur up to a specified time point. Controlling for all pre-treatment variables and all variables in between treatment and a specified time point of the process, is there still an effect of the treatment? How big are these direct effects on average? How do they differ for different values of the variables controlled? Hence, now we ask: How big are the direct effects for which kinds of subjects? In educational research we may ask if the effects of a teaching method are transmitted to the outcome variable by raising the motivation to learn and by increasing time spent on learning, or if the total effect observed is not mediated through these and other variables occurring in the process up to the time point to which time spent on learning refers.

Asking for direct causal effects of a treatment has gained increasing interest in the last years because it has been shown that the well-known path analysis procedures for the analysis of direct and indirect effects often lead to systematically wrong results even in the randomized experiment. This is due to the fact that independence of treatment and covariates induced by randomization is cancelled as soon as we condition on an intermediate variable that is affected by treatment and is correlated with a pre-treatment variable. In this case, the partial correlation between the treatment variable and the covariate, controlling for the intermediate variable, is not zero. For example, if the intermediate motivation is affected by treatment and it depends on pre-treatment motivation, then a high score on intermediate motivation goes along with a high score on treatment (e. g., 1) and a high score on pre-treatment motivation, whereas a low score on intermediate motivation goes along with a low score on treatment (e. g., 0) and a low score on pre-treatment motivation. Independence of the treatment variable and all pre-treatment variables (induced by randomization) implies unbiasedness of mean differences of the outcome variable between treatment groups, but it does not imply unbiasedness of direct effects if we only include treatment, intermediate, and outcome variables in our analysis.

References:

  • Baron, R. M., & Kenny, D. A. (1986). The moderator-mediator variable distinction in social psychological research: Conceptual, strategic and statistical considerations. Journal of Personality and Social Psychology, 51(6), 1173-1182.
  • MacKinnon, D. P. (2008). Introduction to statistical mediation analysis. New York: Lawrence Erlbaum.

Organizational structure

The structure of the conference was closely followed the "Symposium on Causality 2010". There were 5 focus presentations by leading proponents in different fields of causality research. Each focus presentation was discussed and supplemented by two invited discussants, followed by an open discussion among all participants. There was also be room for participants to present their own research in short presentations.


Focus presentations

Philip Dawid (University of Cambridge, Cambridge, UK)
"A probabilistic approach to direct and indirect effects"
Kosuke Imai (University of Princeton, NJ, USA)
"The potential outcome approach to direct effects: Definition, identification, and sensitivity analysis"
Andreas Klein (Universität Frankfurt, Germany)
"Conceptual difficulties of modeling mediation: Logical and statistical problems intertwined"
David MacKinnon (Arizona State University, AZ, USA)
"Introduction to Mediation Analysis: Importance, Applications, and Examples"
Axel Mayer (Friedrich-Schiller-Universität Jena, Germany)
"Identification and estimation of direct and indirect effects"
Susanne Rässler (Universität Bamberg, Germany)
"Conditional effects: How to use them in practice"
Rolf Steyer (Friedrich-Schiller-Universität Jena, Germany)
"The definition of total, direct and indirect effects and their identification"
Johannes Textor (Utrecht University, Utrecht, The Netherlands)
"Graphical Approaches to Covariate Selection for Direct and Partial Causal Effects"
Felix Thoemmes (Cornell University, NY, USA)
"Covariate selection for total and direct effects: A comparative view"

Discussants

George Marcoulides (University of California at Riverside, CA, USA)
David Rindskopf (City University of New York, NY, USA)
Peter Steiner (University of Wisconsin at Madison, WI, USA)
Stephen West (Arizona State University, AZ, USA)

Previous conferences on Causality