@article{2018:otter:causal_inf,
title = {Causal Inference Using Mediation Analysis or Instrumental Variables - Full Mediation in the Absence of Conditional Independence},
year = {2018},
note = {This paper re-emphasizes the topic of model specification in the
context of mediation analysis and highlights the ambiguous nature of
results that are consistent with partial mediation, both with respect
to the existence of a direct causal effect, and with respect to
correct inference about the indirect effect. While none of our
individual results are genuinely new, they are often only discussed on
the side or as “special topics” in application oriented
discussions, or in the newer, more technical literature on causal
inference in mediation analysis. With this paper, we hope to
contribute to the acute awareness of model specification issues in the
wider community of researchers that rely on mediation analysis for
their substantive research. We clarify the differences between
mediation analysis and IV-estimation as a starting point for the
discussion of model specification in the context of mediation
analysis. Specifically, we highlight the connection between full
mediation and conditional independence and show that conditional
independence may be violated even though full mediation holds at a
causal theory level. When this is the case, standard mediation
analysis yields biased inference both with respect to the direct and
the indirect effect. We emphasize that this is a problem separate
from, and prior to the problem of reliably assessing the statistical
significance of estimates suggestive of an indirect effect. We discuss
how Bayes factors usefully improve on p-values in the context of
testing for the absence of effects and suggest the development of
additional testable hypotheses and of the corresponding testing
routines. Finally, we suggest that researcher relying on mediation
analysis should investigate alternative model specifications, notably
a specification that allows for unobserved common causes of the
mediator and the dependent variable, and a specification that allows
for measurement error in the mediator. This can be easily accomplished
using covariance algebra or structural equation modeling software.
Finally, our results imply that full mediation at a causal theory
level could be more common than currently believed.},
journal = {Marketing ZFP},
pages = {41--57},
author = {Otter, Thomas and Pachali, Max J. and Mayer, Stefan and Landwehr, Jan R.},
volume = {40},
number = {2}
}