Effect modification and the limits of biological inference from epidemiologic data.