> I have a question. I am trying to wrap my head around the calculation
> and use of Population Attributable Risk based on a cross sectional
> study.
> From my basic epidemiology courses, I have come to understand that
> such an estimate is only reasonable when one can make plausible
> assumptions of causality. Given that cross sectional studies are
> hardly able to provide a basis for causal inferences, is such a
> measure reasonable to calculate?
While I can see how the idea of population attributable risk is
appealing in terms of translating results, it is still only a
numerical calculation based on study data. As such there are several
issues.
Cross sectional studies, as you indicate are inadequate for causal
inference because of temporality. So I would not use one to calculate
the atributable fraction. But a longitudinal cohort study, is usually
insufficient on it's own to prove causality also. Recall that you
usually need associations, temporal relationships, dose response,
plausibility and consistency.
Even if you had all of that, and pooled several studies together to
get attributable risks, you may still be dealing with confouding or
effect modification that was not considered.
The calculations are only as good as the data that went into them. In
most epidemiological designs, dealing with complex biological
processes and diseases, there are many, many factors that could be
effect modifiers and/or confounders that just aren't evaluated.
Attributable fractions should be evaluated cautiously.
Just my 2 cents.....
Marc
|
