Influence of selection bias on the test decision. A simulation study.

BACKGROUND

Selection bias arises in clinical trials by reason of selective assignment of patients to treatment groups. Even in randomized clinical trials with allocation concealment this phenomenon can occur if future assignments can be predicted due to knowledge of former allocations.

OBJECTIVES

Considering unmasked randomized clinical trials with allocation concealment the impact of selection bias on type I error rate under permuted block randomization is investigated. We aimed to extend the existing research into this topic by including practical assumptions concerning misclassification of patient characteristics to get an estimate of type I error close to clinical routine. To establish an upper bound for the type I error rate different biasing strategies of the investigator are compared first. In addition, the aspect of patient availability is considered.

METHODS

To evaluate the influence of selection bias on type I error rate under several practical situations, different block sizes, selection effects, biasing strategies and success rates of patient classification were simulated using SAS.

RESULTS

Type I error rate exceeds 5 percent significance level; it reaches values up to 21 percent. More cautious biasing strategies and misclassification of patient characteristics may diminish but cannot eliminate selection bias. The number of screened patients is about three times larger than the needed number for the trial.

CONCLUSIONS

Even in unmasked randomized clinical trials using permuted block randomization with allocation concealment the influence of selection bias must not be disregarded evaluating the test decision. It should be incorporated when designing and reporting a clinical trial.