Alternative Analytic and Matching Approaches for the Prevalent New-user Design: A Simulation Study

Background: The prevalent new user (PNU) design allows researchers to answer research questions that cannot be addressed via an active comparator new user design (e.g. effects of switching or augmenting vs continuing on treatment). While time-conditional propensity score (TCPS) matching was the initial analytic strategy used for PNU studies, other strategies may be capable of addressing confounding in PNU cohorts but have not been assessed for bias and efficiency.

Objectives: To evaluate relative bias of the risk ratio (RR) estimates from various analytic approaches in a simulated PNU cohort.

Methods: We simulated a population the size of the original PNU study (86,000 total new users of the treatment of interest and comparator), with 4,900 of the comparator new users becoming PNUs. Individuals initiated the comparator with the option to stop, restart, or switch to the treatment of interest over time, with probabilities of each determined by four confounders; the outcome was binary and ended follow-up (similar to a mortality outcome). We evaluated standardized morbidity ratio weighting (SMRW), disease risk scores (DRS), TCPS matching with replacement (TCPS-R), TCPS matching without replacement that matched participants in ascending (TCPS-Ascending) or descending (TCPS-Descending) order of time on treatment, PS matching to a comparator sample created without replacement in ascending (Match-Ascending) or descending (Match-Descending) order of time on treatment, and PS matching when randomly selecting one comparator observation per continuing individual (Match-Random) as analytic options. We compared average three-year RRs from each analytic method across 2,000 replicates to the truth (estimated with a Monte Carlo approach) for both a null effect and a protective effect.

Results: When estimating a null effect, SMRW, DRS, TCPS-R, TCPS-Ascending, and Match-Ascending were all unbiased. TCPS-Descending (RR=0.983), Match-Descending (0.903), and Match-Random (0.802) exhibited progressively more bias, despite balanced covariate distributions. Similar patterns in bias were present when the treatment of interest was protective vs. the comparator. SMRW and DRS were the most precise methods.

Conclusions: Multiple analytic methods can estimate effects without bias in PNU cohorts including SMRW and DRS. If matching without replacement is used (whether in TCPS or other matching strategies) matching in ascending order of time on treatment helps prevent selection bias. Analytic approaches in empirical applications of PNU designs should be unbiased and balance computational complexity, practicality, and interpretability.