Bias due to observation of different parts of a nonconstant hazard curve over time.

摘要

OBJECTIVES. Uncontrolled data often have to be used in clinical decision marking and in the planning of clinical trials. When such data are used as a basis for comparing different treatment strategies, they tend to generate sources of bias such as inconsistent patient selection, misrepresentation, and measurement errors. A rational usage of uncontrolled data requires identification and handling of different systematic errors when comparing different treatment strategies. Therefore we sought to define a systematic error that deserves more attention in the clinical literature. METHODS. Using hypothetical examples, we illustrate the bias introduced by variations in "lead time," proportions of so-called prevalent/incident cases, and lengths of follow-up between reference and treatment groups. RESULTS. We found these three situations conceptually identical. Bias is introduced because different parts of a nonconstant hazard curve over time are observed. CONCLUSIONS. The systematic error described is potentially important in uncontrolled data provided the hazard rate over time is nonconstant. Thus the seemingly worse outcome after external beam radiation therapy as compared to the outcomes of deferred treatment of radical prostatectomy in compiled patient series of localized prostate cancer may in some part be explained by bias due to observation of different parts of a nonconstant hazard curve over time.