(A-184) Potential for Extreme Bias Due to Outcome Misclassification in Relative Measures of Effect for Rare Time-to-Event Outcomes

Background: Studies of medical product safety often involve comparative analyses of rare time-to-event outcomes with misclassified outcome data. For binary outcomes, a risk ratio is unbiased under non-differential misclassification when specificity is 100%, but if specificity < 100% then it is biased toward one and the magnitude of the bias increases with decreasing outcome incidence. A potential advantage of the risk difference is that it depends on the sensitivity and specificity and not on the outcome incidence.

Objectives: Given the importance of time-to-event outcomes in clinical and epidemiological research, we compared the relative bias of absolute and relative measures of effect using Monte Carlo simulation under varying degrees of outcome misclassification and rare outcome incidences.

Methods: In the simulation, misclassification conditions included: 100% sensitivity paired with two levels of specificity (99%, 99.9%); 100% specificity paired with two levels of sensitivity (90%, 50%); four levels of less than perfect sensitivity and specificity (90%, 50% sensitivity x 99%, 99.9% specificity). Each combination of sensitivity and specificity was evaluated under three cumulative incidences (0.1%, 0.5%, 1.0%). In total there were 24 simulation scenarios. Within each condition, we compared each estimate under misclassification to no misclassification to calculate percent relative bias.

Results: Relative measures of effect were susceptible to considerable bias, particularly when outcome incidence was very low. Absolute measures of effect had less bias than relative measures when specificity and sensitivity were both < 100%, most notable when sensitivity was 90% and in some cases when the sensitivity was as low as 50%.

Conclusions: The results suggest when the outcome incidence is rare, specificity is less than perfect, and sensitivity is at least modest, absolute measures may be preferable to relative measures of effect.