|
Research Currents
Causal
Inference
The Seattle Epidemiologic Research and Information
Center will be publishing a bi-monthly series of
short articles addressing research methodologies
that may be of interest to the VA research community.
This month, we give a brief overview of current
thinking on inferring causality. Please visit our
web site for more details and links to articles
related to causality: http://www.eric.seattle.med.va.gov/research_currents.html.
For many of us in the health sciences, the ultimate
goal of scientific inquiry is to describe the underlying
causal pathways that link genetics, biology, environment,
and human systems to physical and mental health
and to use this understanding to prevent or cure
disease. Determining causality in experimental
and non-experimental observational research can
be challenging and demands rigorous methodologies.
Randomized controlled trials are experiments and
the gold-standard for inferring causality. Although
useful in many medical settings, this approach
is often limited by feasibility issues or ethical
concerns. In the absence of randomization, other
approaches have been conceptualized. In 1965 Sir
Austin Bradford Hill proposed a seminal set of
guidelines for assessing causality that are still
widely used. The guidelines include: strength of
the association, where stronger associations are
more supportive of causality; consistency of the
association, where the association is observed
repeatedly across time, locations, and populations;
temporality, where cause precedes effect; biologic
gradient or dose-response, where more exposure
is associated with more effect; biologic plausibility,
where observed association fits biologic models
from the laboratory; and experimental evidence,
where manipulation of the exposure changes the
outcome.
More contemporary efforts to clarify thinking about
causal inference have relied upon the use of factuals
and counterfactuals, also known as “potential
outcomes.” Most simply this concept argues
that for a person who is exposed to A and then
experiences outcome B, A is considered causal if
and only if, had that same person not been exposed
to A, she/he would not have experienced outcome
B. At issue is that a person cannot simultaneously
be exposed and not exposed to
A. Hence the observed association of exposure A
and outcome B, the factual,
must be compared to an unobservable situation,
the counterfactual, where exposure A does
not occur yet all other factors are identical.
Although it
is not practical to compare factuals with counterfactuals,
the approach can be informative when planning a
study and readily understood when thinking about
the role of randomization in the experimental process:
the creation of 2 or more groups that differ only
in terms of an exposure.
References
1. Hill AB. The environment and disease: association
and causation. Proceedings of the Royal Society
of Medicine 1965;58:295-300.
2. Holland PW. Statistics and causal inference.
J Am Stat Assoc 1989;81:945-66 (includes commentaries
and discussion).
3. Doll R. Controlled trials: the 1948 watershed.
BMJ. 1998;317(7167):1217-20.
http://bmj.bmjjournals.com/cgi/content/full/317/7167/1217
4. Maldonado G, Greenland S. Estimating causal
effects. Int J Epidemiol 2002;31:422-38 (includes
commentaries and discussions).
http://ije.oxfordjournals.org/cgi/content/full/31/2/422
http://ije.oxfordjournals.org/cgi/content/full/31/2/429
http://ije.oxfordjournals.org/cgi/content/full/31/2/431
http://ije.oxfordjournals.org/cgi/content/full/31/2/432
http://ije.oxfordjournals.org/cgi/content/full/31/2/434
http://ije.oxfordjournals.org/cgi/content/full/31/2/435
5. Hofler M. Causal inference based on counterfactuals.
BMC Med Res Methodol 2005;5:28.
http://www.biomedcentral.com/1471-2288/5/28
|
