MeMa - Methodology Matters
Prof. Leonce Röth (University of Cologne)
Modelling Causal Pathways - Structural Equations and Mediation of Causal Effects
24 March 2017,
Computer Lab, h. 9.00-12.15 and 13.30-18.00
NASP Graduate School in Social and Political Sciences
Via Pace, 10 - Milan
The history of theories on causality is a history of increasing contextualization of causal claims. This is reflected in the refined counterfactual theory (Lewis 2000) or the structural equation approach (Pearl 2009). Both converge to the folk intuition of causality – basically, distinguishing between operative background conditions and deviating events applying both counterfactual and productive truth makers in particular causal structures.
Accordingly, causal effects have only a limited validity. In typical econometric regression models, effect heterogeneity remains a widely neglected issue. Although many researchers employ or assume interacting or mediating relations in their theories and their models. Controls variables are widely applied as if they would only have direct causal links to the outcome.
The perspective of contextualized causality raise the awareness of background or scope conditions. Participants are encouraged to specify the often implicit background conditions of their own research and learn to use scope conditions to delineate their own ambition to the generalizability and specify the population under scrutiny.
Starting from this exercise we discuss directed graphs as an intuitive way to illustrate causal assumptions. Based on directed graphs, we distinguish unnecessary from necessary controls in multivariate regressions in order to achieve parsimonious models. The mayor criteria to make this distinction are "closed of back-door paths".
Only with an idea of structural causal relations we start to assess potential heterogeneity of causal effects. We clarify that under heterogeneous effects, regression estimates are biased average treatment effects. The heterogeneity of effects over subgroups can be explicitly modelled using mediating graphs in structural equations. We contrast this technic with the conventional way of modelling interactions in multi-variate regressions. Mediation of causal effects will be discussed and applied in close detail in order to raise awareness of heterogeneous effects and its implications for causal identification.
Pearl, J. (2009). Causality. Cambridge University press.
Lewis, P. (2000). Realism, causality and the problem of social structure. Journal for the theory of Social Behaviour, 30(3), 249-268.
Basic knowledge of regression analysis.
We will use a computer lab. Stata and an open source software on causal graphs will be applied, but their mastering is not required. Datasets will be provided by the instructor.
Part I – Causality in Directed (Acyclic) Graphs
Time 9:00 to 10:30
Pearl, Judea. (1995). "Causal Diagrams for Empirical Research." Biometrika 82 (4): 669- 710.
An important founding document for causal diagrams. However, it is very technically written.
Morgan, Stephen L. and Christopher Winship. (2007). Counterfactuals and Causal Inference: Methods and Principles of Social Research. Cambridge: Cambridge University Press.
Chapter 3 (p. 77-95) is a more applied discussion of directed graphs embedded in examples from sociology.
Robins, James M. (2001). "Data, Design, and Background Knowledge in Etiologic Inference," Epidemiology 11 (3): 313-320.
Another less technical description of directed graphs.
Part II – Necessary and Unnecessary Controls or Closing the Backdoor
Morgan, Stephen L. and Christopher Winship (2007). Counterfactuals and Causal Inference: Methods and Principles of Social Research. Cambridge: Cambridge University Press.
Chapter 3 (p. 95-130) is an applied discussion of the back-door criteria.
Pearl, J., Glymour, M., & Jewell, N. P. (2016). Causal inference in statistics: a primer. John Wiley & Sons. Chapter 3 (p. 53-75).
Lunch Break 12:15-13:30
Part III – Effect Heterogeneity and the Average Treatment Effect
Elwert, F., & Winship, C. (2010). Effect heterogeneity and bias in main-effects-only regression models. Heuristics, probability and causality: A tribute to Judea Pearl, 327-36.
Falleti, T. G., & Lynch, J. F. (2009). Context and causal mechanisms in political analysis. Comparative political studies, 9(42), 1143-1166.
Part IV – Interactions versus Structural Equations with Mediation Effects
Pearl, J., Glymour, M., & Jewell, N. P. (2016). Causal inference in statistics: a primer. John Wiley & Sons. Chapter 3 (p. 75-87).
Brambor, T., Clark, W. R., & Golder, M. (2006). Understanding interaction models: Improving empirical analyses. Political analysis, 14(1), 63-82.
Imai, K., Keele, L., & Yamamoto, T. (2010). Identification, inference and sensitivity analysis for causal mediation effects. Statistical Science, 1(25), 51-71.
Part V – Discussion
Background readings will be circulated in advance only to participants.
This seminar is part of MeMa - Methodology Matters Seminar Cycle.