Methods/Statistics
Adapting the prevalent new user study design to evaluate the effects of drug discontinuation: an example in statins and mortality Qoua Her* Qoua Her Alan Ellis Emilie Duchesneau Til Stürmer Jennifer Lund Viriginia Pate Michael Webster-Clark
Drug discontinuation studies often use causal contrasts to ask questions like: “what if everyone discontinued a drug immediately after some date” vs “what if no one ever discontinued drug?” Combining the prevalent new user (PNU) design and time-stratified odds weighting allows us to ask a different question: “what if the discontinuers had continued the drug rather than discontinued?”
We demonstrated this in a cohort of new statin users (i.e., no statin prescription in the previous year) with at least 5 years of subsequent continuous statin use (i.e., without a gap in statin use of >90 days) from a 20% random sample of Medicare fee-for-service beneficiaries. We adapted the PNU design to estimate the effect of statin discontinuation versus continuation on mortality among discontinuers, treating discontinuation like a switch to new treatment in the PNU design with continuing as the comparator. For each month after 5 years of use, we created exposure sets of discontinuers and continuers with equal duration of previous statin treatment and estimated a month-stratified propensity score for discontinuation (Figure 1). We calculated all-cause mortality rates, the rate difference (IRD), and the rate ratio (IRR) for discontinuers and the continuers with and without odds weights for the continuers. We estimated 95% CIs with the 2.5th and 97.5th percentiles from 150 bootstrap iterations.
We identified 113,109 statin initiators who continued for at least 5 years. Mortality rates per 100 person-years were 12.16 among discontinuers and 6.20 among continuers before applying odds weights. After weighing, the IRD was 3.06 (95% CI: 2.82, 3.43) per 100 person-years, and the IRR was 1.34 (1.30, 1.37).
While additional work (e.g., restriction to less-frail individuals) is necessary to avoid bias from the sick stopper effect, adapting the PNU design to drug discontinuation allows estimation of the treatment effect in a well-defined target population of clinical relevance.
