(P-548) Graphical Representation of Multiple Quantitative Bias Analysis Scenarios for Unmeasured Confounding

Background: Quantitative bias analyses can correct effect-measure estimates for unmeasured confounding using assumptions of the strength and prevalence of a hypothetical unmeasured confounder. When these parameters are unknown or vary, evaluating a wide spectrum of assumptions may be of interest.

Objectives: To evaluate and graphically present the potential impact of an array of unmeasured confounding scenarios on an observed incidence rate ratio (IRR).

Methods: An IRR comparing the risk of an outcome in users of a specific medication (exposure group) and a comparator group was estimated from a cohort study. Three scenarios of the association of a hypothetical unmeasured confounder with the outcome were evaluated—risk ratio (RR) = 1.5 (moderate association), RR = 3.0 (strong), and RR = 4.5 (very strong). A factor must be imbalanced between treatment groups to be a confounder. For each hypothetical confounder strength scenario, we calculated IRRs corrected for the unmeasured confounder at every possible imbalance level, ranging from present in all comparator patients and absent in all exposure patients to absent in all comparator patients and present in all exposure patients. For each confounder strength scenario, all possible corrected IRRs were plotted as a function of the confounder imbalance on a single graph using SAS.

Results: In the resulting graphical display, there was a range of possible corrected IRRs at each imbalance level for each strength scenario, as a given covariate imbalance level may be achieved through numerous combinations of confounder prevalences. In the cohort example, the observed IRR was 0.70. In a worst-case scenario of having a hypothetical moderate confounder (RR = 1.5) with 0% prevalence in the exposure group and 100% prevalence in the comparator group, the maximum corrected IRR would be 1.05; any less extreme imbalance would result in IRRs lower than 1.05. A hypothetical moderate confounder (RR = 1.5) would require at least 90% higher prevalence in the comparator group to mask a true IRR greater than 1.00. If the hypothetical unmeasured confounder had a stronger independent relationship with the outcome, RR = 3.0 or 4.5, then a smaller imbalance would be required.

Conclusions: An array of assumptions was evaluated and displayed on a single graph to assess the maximum possible impact of a single unmeasured confounder. Summary plots of multiple confounding scenarios provided an efficient method of displaying and evaluating the potential impact of unmeasured confounding on the results of an observed risk estimate.