Presentation (CTI)

OBJECTIVES: Target trial emulation (TTE) supports causal inference from observational data but collapses treatment effects into a single time-constant contrast, obscuring how benefits accrue over exposure duration or attenuate after discontinuation. The goal of this study was to develop a TTE framework that supports the evaluation of duration response and the recovery of individual intention-to-treat effects (ITT).
METHODS: We developed duration-conscious TTE using g-estimation of structural nested mean models (SNMM), clone-censor-weighting, and a centered grace-period encouragement indicator as an instrument. Using SNMM blip functions, the SNMM decomposes the intention-to-treat (ITT) risk difference into on-treatment and washout components that vary with exposure duration. We compare SNMM with a constant-effect TTE in a Monte Carlo simulation and apply both to a claims study of tirzepatide versus sodium-glucose cotransporter-2 (SGLT2) initiation to reduce opioid overdose risk in patients with comorbid type 2 diabetes and opioid use disorder.
RESULTS: In simulations, SNMM reduced ITT risk-difference bias by threefold compared with constant-effect TTE (−0.010 vs. −0.034) and recovered duration-specific response curves that matched the data-generating process. Constant-effect TTE bias had two components: post-grace treatment-uptake dynamics that biased the estimate under a constant per-period effect and additional bias resulting from duration heterogeneity in the treatment effect. In the tirzepatide application, constant-effect TTE showed an 18-month overdose risk reduction of −0.0062 (95% CI: −0.0096, −0.0029), and SNMM showed an 18-month risk difference of −0.0077 (95% CI: −0.0094, −0.0063), relative to SGLT-2s. The SNMM effect grew with each treated month and persisted after stopping, highlighting the lasting benefit of tirzepatide in reducing overdose risk relative to SGLT-2s.
CONCLUSIONS: Duration-conscious TTE via SNMM g-estimation characterizes duration-specific effects that standard TTE frameworks collapse into a single summary. It is most useful when effects depend on exposure duration or when post-grace treatment-uptake dynamics make a constant hazard-ratio assumption untenable.