Michel MOUCHART
CORE and ISBA, Universit_e catholique de Louvain
August 14, 2015
Abstract
Background. In demography, and in other social sciences, determining the variables to be controlled for is usually a major problem when analyzing possible causal relations. Clear guidelines in the literature are scarce and actual practice is often questionable. Objectives. A structural modeling approach is presented as a consistent framework for de- termining a coherent set of guidelines for deciding what variables should be controlled.
Methods. The method is based on a recursive decomposition of the multivariate distribution, represented by a directed acyclic graph and reecting the causal mechanism and sub-mechanisms of the data generating process.
Results. Two rules are developed for determining control variables when studying respec- tively the direct and the total e_ects of a cause on an outcome. The rules can easily be applied in the framework of a causal model based on background knowledge and invariant to changes of the environment.
Conclusions. Our approach for determining control variables is simpler and more consistent than the alternative ones based on Pearl's back-door criterion. It takes into account both confounders and other immediate causes leading to variations in the distribution of the outcome variable and possibly being in interaction with the causal variable.
Keywords: Causality, Control, Causal Modelling, Structural Modelling, Recursive Decomposi-tion, Total E_ect, Direct E_ect, Directed Acyclic Graph.
Joint work with: Guillaume WUNSCH and Federica RUSSO
Contact person
Daniela Cocchi