Calibrated Propensity Scores for Comparative Effectiveness Estimates with Disease Subject to Misclassification
Keywords: comparative effectiveness, propensity score, misclassification error, prostate cancer
Generally, medical decisions are performed in a highly uncertain environment. For example, this is particularly true in the management of prostate cancer due to two issues. First, the instrument by which diagnosis (i.e. biopsy) is made is prone to measurement error. Second, depending on the risk stratum determined by the biopsy, the individual has two or more potential competing treatment strategies from which to choose. However, only one of these treatment strategies (radical prostatectomy) provides an almost error-free severity measure appropriate for prognosis and decision-making. This means that estimates of comparative effectiveness in this setting are done under measurement error which is usually disregarded. Depending on the type of adjustment for the comparison, regression model or propensity scoring, the problem is one of misclassification in covariates or response. Specimens from radical prostatectomy provide the perfect scenario in which methods for handling misclassification can be evaluated empirically. Based on disease severity measured from more than 3,500 radical prostatectomy specimens, we assess the characteristics of the misclassification in disease severity as per biopsy, we build a calibration model to account for this error, and finally estimate the influence of the misclassification error and calibration models on estimates of biochemical recurrence, a marker of disease progression and treatment success.