BACKGROUND: Quantitative bias analysis methods can use treatment group–specific positive predictive values (PPV) to correct observed effect-measure estimates biased by differential outcome misclassification. However, if treatment-specific PPVs are unknown, a range of differential misclassification scenarios may be evaluated simultaneously to estimate the potential extent of bias.
OBJECTIVES: To assess and display the potential impacts of all possible combinations of differential misclassification combinations on an observed incidence rate ratio (IRR).
METHODS: A cohort study comparing the rate of an outcome between two treatment groups was used as an example. An overall outcome PPV was available, but treatment-specific PPVs were unavailable due to low event counts in treatment groups. Using the observed cohort IRR, we used every possible combination of differential misclassification, ranging from perfect classification in the treatment group (PPV = 100%) and complete misclassification in the comparator (PPV = 0%) and vice versa to estimate a corrected IRR at every PPV combination.
RESULTS: In the resulting graphical display, a matrix of all corrected IRR values for each combination of possible treatment and comparator PPVs was created; corrected IRR values were shaded with a gradient of darker colors indicating higher IRR values. Thus all possible differential misclassification scenarios were viewed on a single plot. In the cohort example, the observed IRR was 0.44 (13 treated cases, 61 comparator cases), and the overall outcome PPV was 63%. Analysis of the figure and underlying matrix indicated that for a true increased risk greater than IRR = 1.5 to be masked by differential misclassification, treatment PPVs of 50% to 70% must be approximately 33% to 46% higher than comparator PPVs. To estimate the worst-case differential outcome misclassification scenario possible given the observed number of outcome cases by treatment group, we assumed that if the treatment PPV was 100% (all 13 cases confirmed true cases), then the comparator would have a PPV of approximately 55% (34 of 61 confirmed true cases) for the observed overall PPV to be 63% (47 of 74 total cases confirmed), resulting in a hypothetical worst-case corrected IRR of 0.96. PPVs higher in the comparator than the treatment group would result in decreased IRR estimates.
CONCLUSIONS: Simultaneously displaying all possible differential misclassification scenarios provided an efficient method to display and interpret large amounts of data relating to the extent of differential misclassification that would be required to change the interpretation of the primary study results.
