Assessing non-inferiority for binary matched-pairs data with missing values: a powerful and flexible GEE approach based on the risk difference.

Abstract

Background: Clinical studies often aim to test the non-inferiority of a treatment compared to an alternative intervention with binary matched-pairs data. These studies are often planned with methods for completely observed pairs only. However, if missingness is more frequent than expected or is anticipated in the planning phase, methods are needed that allow the inclusion of partially observed pairs to improve statistical power.

Methods: We propose a flexible generalized estimating equations (GEE) approach to estimate confidence intervals for the risk difference, which accommodates partially observed pairs. Using simulated data, we compare this approach to alternative methods for completely observed pairs only and to those that also include pairs with missing observations. Additionally, we reconsider the study sample size calculation by applying these methods to a study with binary matched-pairs setting.

Results: In moderate to large sample sizes, the proposed GEE approach performs similarly to alternative methods for completely observed pairs only. It even results in a higher power and narrower interval widths in scenarios with missing data and where missingness follows a missing (completely) at random (MCAR / MAR) mechanism. The GEE approach is also non-inferior to alternative methods, such as multiple imputation or confidence intervals explicitly developed for missing data settings. Reconsidering the sample size calculation for an observational study, our proposed approach leads to a considerably smaller sample size than the alternative methods.

Conclusion: Our results indicate that the proposed GEE approach is a powerful alternative to existing methods and can be used for testing non-inferiority, even if the initial sample size calculation was based on a different statistical method. Furthermore, it increases the analytical flexibility by allowing the inclusion of additional covariates, in contrast to other methods.