Balancing treatment allocations by clinician or center in randomized trials allows unacceptable levels of treatment prediction.
Hills RK., Gray R., Wheatley K.
OBJECTIVE: Randomized controlled trials are the standard method for comparing treatments because they avoid the selection bias that might arise if clinicians were free to choose which treatment a patient would receive. In practice, allocation of treatments in randomized controlled trials is often not wholly random with various 'pseudo-randomization' methods, such as minimization or balanced blocks, used to ensure good balance between treatments within potentially important prognostic or predictive subgroups. These methods avoid selection bias so long as full concealment of the next treatment allocation is maintained. There is concern, however, that pseudo-random methods may allow clinicians to predict future treatment allocations from previous allocation history, particularly if allocations are balanced by clinician or center. We investigate here to what extent treatment prediction is possible. METHODS: Using computer simulations of minimization and balanced block randomizations, the success rates of various prediction strategies were investigated for varying numbers of stratification variables, including the patient's clinician. RESULTS: Prediction rates for minimization and balanced block randomization typically exceed 60% when clinician is included as a stratification variable and, under certain circumstances, can exceed 80%. Increasing the number of clinicians and other stratification variables did not greatly reduce the prediction rates. Without clinician as a stratification variable, prediction rates are poor unless few clinicians participate. CONCLUSION: Prediction rates are unacceptably high when allocations are balanced by clinician or by center. This could easily lead to selection bias that might suggest spurious, or mask real, treatment effects. Unless treatment is blinded, randomization should not be balanced by clinician (or by center), and clinician-center effects should be allowed for instead by retrospectively stratified analyses.