When experiments and intervention studies cannot be done for ethical and practical reasons common causes must be excluded by indirect means. The logic of the approach is conceptually straightforward: we see whether A and B covary when possible common causes are statistically "controlled for".
One approach to this goal is to control potentially confounding variables in the study design. For example, we could rule out the hypothesis that any relationship between non-steroidal anti-inflammatory drug (NSAID) use and coronary heart disease was a consequence of concurrent medicine use by: (i) excluding individuals with the disease who used other medicines from a cohort study; (ii) by matching cases and controls on concurrent medicine use; or (iii) by stratifying potential participants in a cohort study on confounding variables and matching on those variables [3,16].
Another commonly used approach to dealing with confounding in epidemiology is covariate adjustment. In this approach, all study participants are measured on potentially confounding variables (covariates) and statistical methods are used to estimate the association between A and B while controlling for the covariates [16-19].
Propensity score analysis can be used to assess the plausibility of selection bias as an explanation of relationships in observational studies where patients select their own treatment [13,16]. In this approach, covariates are used to predict the exposure condition that each individual had the greatest propensity to receive. The resulting "propensity score" can be used either as a matching variable or as a covariate in regression analyses [13,20].
Sensitivity analyses can be used when we do not have measures of potential confounders for covariate adjustment or propensity score analysis. Such analyses explore the plausibility of confounding as an explanation of observed outcomes [13,16,21]. These analyses involve modelling the relationship between medicine use and the health outcome under various scenarios in which a confounding variable is related in varying degrees to both medicine use and the outcome. If the relationship between the two persists when allowance is made for plausible degrees of confounding then we can be more confident that the relationship is likely to be causal ([22], pp 193–196).
The major limitation of all these strategies is that they can only rule out specified alternative hypotheses. That is, we have to identify a candidate common cause that we can then match on, measure and adjust for using covariate adjustment or propensity scores, or model in sensitivity analyses. Randomisation is superior to all these strategies because it rules out all possible common causes, including ones that have not been measured or thought of [12].
Answers to all your Biology Questions
