More realistic power estimation for new user, active comparator studies: an empirical example.

Area of Science:

• Pharmacoepidemiology
• Drug safety research
• Biostatistics

Background:

• Pharmacoepidemiologic studies often aim to detect rare outcomes.
• Study design choices to minimize bias can sequentially reduce statistical power.
• This study examines power loss in a new-user active-comparator study of DPP-4 inhibitors and pancreatic cancer.

Purpose of the Study:

• To illustrate the impact of bias-minimizing steps on statistical power in pharmacoepidemiologic studies.
• To compare the power of a study investigating pancreatic cancer incidence after initiating dipeptidyl-peptidase-4 inhibitors (DPP-4i) versus comparators, before and after applying design refinements.
• To inform the design of future new-user active-comparator studies.

Main Methods:

• Identified Medicare beneficiaries initiating DPP-4 inhibitors or comparators (thiazolidinediones, sulfonylureas) between 2007-2009.
• Applied sequential exclusion criteria: prevalent users, single prescription, prevalent cancers, age <66 years, and as-treated censoring.
• Estimated statistical power to detect hazard ratios ≥2.0 before and after applying these bias-minimizing steps.

Main Results:

• The study included 19,388 DPP-4i and 28,846 thiazolidinedione initiators.
• Significant reductions in sample size, outcome events, and person-time occurred with bias-minimizing steps.
• Naïve power (>99%) was substantially higher than the final estimated power (~75%) after applying all design refinements.

Conclusions:

• Steps taken to minimize bias in new-user active-comparator studies significantly affect sample size, outcome counts, and person-time.
• Awareness of these losses is crucial for accurate power estimation in study design.
• Using generic loss percentages can improve power estimates compared to naive approaches that ignore validity-enhancing steps.