STATISTICAL-ANALYSIS OF POSSIBLE BIAS OF CLINICAL JUDGMENTS DUE TO OBSERVING AN ON-THERAPY MARKER VARIABLE

In certain double-blind clinical trials there is the possibility that certain 'marker variables' observable during the trial may in part unb lind the trial, even at a subliminal level. At issue is whether or not this potential unblinding biases the investigators' clinical efficacy assessments. This issue arose after the completion of three clinical trials that compared tretinoin emollient cream (TEC) 0.05 per cent to its vehicle in patients with photodamaged skin. The question raised wa s whether or not possible 'subliminal unblinding' of the investigators and patients, due to the cutaneous irritation associated with topical tretinoin, might have caused a treatment bias in the study. To addres s this issue, we undertook a reanalysis of these three clinical trials . In doing so, we develop in this paper a statistical modelling approa ch to address issues of possible bias introduced by the ability to obs erve such marker variables. The approach utilizes a linear discriminan t analysis to introduce an auxiliary categorical variable for the effi cacy analysis, A suitable categorical data model permits the estimatio n of relevant bias effects. We illustrate this approach with data from the three TEC 0.05 per cent trials.