Many-to-one comparisons with general linear model in the presence of missing observations

This study proposes a new parametric Dunnett-type procedure for many-to-one comparisons with general linear models (GLM) in the presence of missing observations. The procedure pools GLM results from multiple imputed datasets, accounting for the multiplicity inherent in many-to-one comparisons. The single-step version of the procedure generates simultaneous confidence intervals and corresponding p-values while the multiple-steps (step-down) version of the procedure yields p-values. Simulation studies demonstrate its strong control of FWER and superiority in test power over the traditional methods for multiplicity adjustment (Bonferroni method and Holm-Bonferroni method after combining GLM results from multiple imputed datasets). Specifically, as the missing proportion increases, either version of the proposed procedure enhances disjunctive power compared to Bonferroni method or Holm-Bonferroni method; the step-down version further improves conjunctive power compared to Holm-Bonferroni method. Given the widespread adoption of many-to-one designs and multiple imputation analyses in clinical trials nowadays, the proposed procedure offers a more powerful parametric alternative for multiplicity-adjusted inference based on multiple imputed datasets, applicable whether the goal is to detect any true treatment difference, a specific difference, or all differences simultaneously, thereby boosting trial success rates.