In carrying out drug safety studies on observational data, one of the major difficulties in assessing causality is that the allocation of treatment is non-random. In observational studies, the indication or reasons for prescribing a drug may be associated with the outcome, which is termed confounding by indication. Other sources of bias include confounding by severity,the protopathic bias and selection bias. The traditional methods of dealing with confounders include matching, stratification and covariate-adjusted regression, and more recently, the use of propensity score methods. One study design proposed by Maclure, the case-crossover design, can in certain circumstances, eliminate control selection bias and confounding by constant within-subject characteristics.