[gov3009-l] Robins on "Pure (natural) direct and indirect effects"
mblackwell at iq.harvard.edu
Mon Sep 27 10:32:35 EDT 2010
We hope you can join us for the Applied Statistics Workshop this
Wednesday, September 29th, when we are excited to have Jamie Robins
speaking on direct and indirect effects, based on joint work with
Thomas Richardson. Details of his talk, an abstract, and a link to the
paper are below. The workshop begins at 12 noon and we will serve a
light lunch. Hope to see you there!
"Pure (natural) direct and indirect effects, alternative
counterfactual models, and the philosophy of science"
Harvard School of Public Health
September 29th, 2010, 12 noon
K354 CGIS Knafel (1737 Cambridge St)
We consider four classes of graphical causal models: the Finest Fully
Randomized Causally Interpretable Structured Tree Graph (FFRCISTG) of
Robins (1986), the agnostic causal model of Spirtes et al. (1993), the
Non-Parametric Structural Equation Model (NPSEM) of Pearl (2000), and
the Minimal Counterfactual Model (MCM) which we introduce. The latter
is referred to as ‘minimal’ because it imposes the minimal
counterfactual independence assumptions required to identify those
causal contrasts representing the effect of an ideal intervention on
any subset of the variables in the graph. The causal contrasts
identified by an MCM are, in general, a strict subset of those
identified by a NPSEM associated with the same graph. We analyze
various measures of the ‘direct’ causal effect, focusing on the pure
direct effect (PDE), also called the ‘natural direct effect’. We show
the PDE is a parameter that may be identified in a DAG viewed as a
NPSEM, but not as an MCM. In spite of this, Pearl has given a scenario
in which the PDE corresponds to the intent-to-treat parameter of a
randomized experiment. We resolve this apparent paradox by showing
that implicit within Pearl’s account is an extended causal DAG with
additional variables in which there is a causal contrast that equals
the pure direct effect of Pearl’s original NPSEM. Further, this
contrast is identified from observational data on the original
variables. Finally we relate our results to the work of Avin et al.
(2005) on path-specific causal effects.
Institute for Quantitative Social Science
Department of Government
More information about the gov3009-l