“The Future
of Longitudinal Studies:
What we know; What we don’t know; What we need to
know”
Inferring Causality
from Longitudinal Studies
Chaired By: Elizabeth Owens, University Of California, Berkeley
Friday, March 21, 2003
Helena
Chmura Kraemer
"Randomized Clinical Trial (RCT) Methodology and Causality"
Stanford University
Inference
for causation lies in study design, not in the analysis of the data.
Unfortunately, untoward inferences are often drawn from correlational
or observational longitudinal studies (e.g., the biomedical “finding”
that hormone replacement prevents heart disease, which current indications
suggest is false). To show that A causes B, one needs to show that:
(1) A precedes B (hence, longitudinal studies), (2) A is correlated
with B (i.e., A is a "risk factor" for B), and (3) that
there is no other explanation for the association other than causality
(i.e., other explanations have been eliminated.). If we know only
(2): A can be described as “a correlate of” B, and if
we know both (1) and (2): A can be described as a “risk factor”
for B. To show (1) requires a longitudinal study and there are many
approaches to showing (2), but how do we show (3)?
One answer is that the causal effect of A [some therapy or intervention]
on B [some outcome measure] for subject i is: E(B/A)-E(B/not-A).
That is, what B would be if exposed to causal factor A, compared
to what it would be if not exposed. Although we can’t make
this calculation for an individual (because exposure at one time
might affect later response), it may be estimated for a population.
This definition of the "causal effect" leads directly
to: the necessity of a representative sample from the population,
the definition of a control/comparison group (not-A), randomization
to the two conditions (A and not-A) to control for selection bias,
"blinded" assessment of outcome to remove the possibility
of measurement bias, analysis by intention to treat, estimation
of the causal effect of A on B in that population, the demonstration
of statistical significance, and the indication of clinical or practical
significance. In short, this leads to RCT methodology.
While cross-sectional studies allow one to identify correlations
only, observational longitudinal studies (OLS) allow one to identify
risk factors and can aid in the design of RCTs. However, only RCT
designs allow one to claim causality with assurance that the proper
criteria have been met, and any causal inference from an RCT, of
course, requires replication, as do all scientific results. However,
this is not to say that OLSs are not of great importance and use.
In an OLS, one might identify 200-300 risk factors, and can begin
to remove the “silt” (proxies, and overlapping risk
factors) to get down to the important constructs and potential moderators
and mediators to test in an RCT. OLSs point us in the right direction,
help us design RCTs (e.g., help us estimate effect sizes), and make
inferences of causation in RCTs possible.
Helena Chmura Kraemer's presentation "Inferring
Causality from Longitudinal Studies" can be viewed in PDF
format, using Adobe®
Acrobat® Reader®.