Graduate Student Department of Epidemiology, Gillings School of Global Public Health, University of North Carolina at Chapel Hill CARRBORO, United States
Background: Robust evidence on the safety and effectiveness of discontinuing medication use (irrespective of cause) via general discontinuation (i.e., stopping use of a drug by patient choice or provider advice) or prescriber-led deprescribing (i.e., a planned reduction or stopping of medication) is scarce. Uncertainty about the safety and effectiveness of discontinuation and deprescribing could be addressed with randomized trials, but such trials are difficult to conduct and patients in these trials may not be generalizable to real-world populations. Commonly used analytic methods like marginal structural models and time-varying hazard ratios estimate the average treatment effect in everyone, rather than the more clinically relevant average treatment effect in the discontinuers (ATD).
Objectives: Demonstrate application of the prevalent new-user design and time-stratified SMR weighting to estimate the ATD in a simulated population.
Methods: We simulated a cohort of 5,000 patients initiating a chronic treatment at time 0 and followed for 5 units of time (study period) or until outcome (e.g., death). After time 0, whether patients discontinued, continued, or restarted treatment (treatment status) was a function of three covariates (two baseline binary confounders and one continuous time-varying confounder affected by prior treatment status). Outcomes were generated from a log-linear model comprising the treatment status and covariates at each time unit. We created a prevalent new-user cohort, treating “discontinuation” like switching to a new treatment and using “continuation” as the comparator. For the discontinuers, we simulated counterfactual outcomes using the true log-linear model to obtain the risk ratio (RR). We estimated crude and time-stratified SMR weighted RRs. SMR weights were created from a logistic regression model stratified by time since treatment initiation. We obtained the median values of the risk ratio (RR) from 1,000 simulations and obtained the 2.5 and 97.5 percentiles as estimates of the 95% confidence limits.
Results: Of the 5,000 patients, 1,421 (28.4%) discontinued treatment for one or more time units. The true RR in the discontinuers, the ATD, was 2.00. The crude RR for discontinuation in the prevalent new-user cohort was 3.45 (95% CI: [2.60, 4.42]), with a time-stratified SMR weighted RR of 2.01 (95% CI: [1.50, 2.67]).
Conclusions: By creating a prevalent new-user cohort in a population of patients who initiated and discontinued, continued, or restarted a treatment, we were able to recover a nearly unbiased estimate of the treatment effect specifically in the discontinuers, the ATD. The prevalent new-user study design can be used to evaluate the effect of discontinuation on health outcome assuming no unmeasured confounding.
