Drawing causal inferences in epidemiologic studies of early life influences.

Observational studies can describe associations between early life exposures and subsequent outcomes in human populations. It is challenging to draw causal inferences from these associations because exposures often occur many years before the outcome and are related to other early life exposures. An approach is required that combines traditional epidemiologic and statistical principles with the use of novel and sophisticated analytic methods. To minimize the bias in longitudinal studies of early origins, researchers need to do all they can to reduce losses to follow-up and to describe individuals who are lost to follow-up. To reduce the role of chance, researchers should concentrate on effect sizes and the strength of the evidence to support these effect sizes, and they should be cautious in their interpretation of subgroup analyses. More complex analytic approaches can and should be used to handle missing data and repeated measurements. Addressing the issue of confounding is not straightforward. Statistical adjustment for the confounders measured in a study may help, but a lack of attenuation does not guarantee that the association is not confounded. Ecologic studies, observational studies in populations with different confounding structures, and the follow-up of randomized trials (where these exist) can be informative. Genetic and nongenetic instrumental variable approaches (eg, Mendelian randomization) may also provide causal insights. These approaches to confounding often require the comparison of data from different populations or a combination of studies to ensure adequate power to provide robust estimates of the causal effect.