Instrumental variable estimation of the marginal structural Cox model for time-varying treatments

Summary Robins (1998) introduced marginal structural models, a general class of counterfactual models for the joint effects of time-varying treatments in complex longitudinal studies subject to time-varying confounding. Robins (1998) established the identification of marginal structural model parameters under a sequential randomization assumption, which rules out unmeasured confounding of treatment assignment over time. The marginal structural Cox model is one of the most popular marginal structural models to evaluate the causal effect of time-varying treatments on a censored failure time outcome. In this paper, we establish sufficient conditions for identification of marginal structural Cox model parameters with the aid of a time-varying instrumental variable, when sequential randomization fails to hold due to unmeasured confounding. Our instrumental variable identification condition rules out any interaction between an unmeasured confounder and the instrumental variable in its additive effects on the treatment process, the longitudinal generalization of the identifying condition of Wang & Tchetgen Tchetgen (2018). We describe a large class of weighted estimating equations that give rise to consistent and asymptotically normal estimators of the marginal structural Cox model, thereby extending the standard inverse probability of treatment weighted estimation of marginal structural models to the instrumental variable setting. Our approach is illustrated via extensive simulation studies and an application to estimate the effect of community antiretroviral therapy coverage on HIV incidence.

Association of Blood Pressure with Cardiovascular Outcome and Mortality: Results from the KNOW-CKD Study

Background Optimal BP control is a major therapeutic strategy to reduce adverse cardiovascular events and mortality in patients with CKD. We studied the association of BP with adverse cardiovascular outcome and all-cause death in patients with CKD. Methods Among 2,238 participants from the KoreaN cohort study for Outcome in patients With CKD, 2,226 patients with baseline BP measurements were enrolled. Main predictor was SBP categorized by 5 levels: <110, 110-119, 120-129, 130-139, and ≥140 mmHg. Primary endpoint was a composite outcome of all-cause death or incident cardiovascular events. We primarily used marginal structural models using averaged and the most recent time-updated SBPs. Results During a median follow-up of 10233.79 person-years (median 4.60 years), the primary composite outcome occurred in 240 (10.8%) participants, with a corresponding incidence rate of 23.5 (95% CI, 20.7–26.6) per 1,000 patient-years. Marginal structural models with averaged SBP showed a U-shaped relationship with the primary outcome. Compared to time-updated SBP of 110–119 mmHg, hazard ratios (95% CI) for <110, 120–129, 130–139, and ≥140 mmHg were 2.47 (1.48–4.11), 1.29 (0.80–2.08), 2.15 (1.26–3.69), and 2.19 (1.19–4.01), respectively. Marginal structural models with the most recent SBP also showed similar findings. Conclusions In Korean patients with CKD, there was a U-shaped association of SBP with the risk of adverse clinical outcome. Our findings highlight the importance of BP control and suggest a potential hazard of SBP <110 mmHg.

742Analyzing the statistical pitfalls of treating mediators as confounders

Background Numerous studies indicated that infants born small-for-gestational-age (SGA) are at higher risk of overweight. However, the association between SGA and overweight may be due to overcontrolling for body size. This study aimed to analyze the effect of controlling for child’s weight and height in the association between SGA and overweight in children born preterm. Methods Data were obtained from the Preterm Infant Multicenter Growth Study (n = 1089). The association between SGA and overweight at 36 months corrected age (CA) was analyzed using logistic regression models: 1) crude, 2) adjusted for baseline covariates, 3) adjusted for baselines covariates with additional adjustments separately for child’s weight and height at 21 months CA. Marginal structural models (MSM) with stabilized inverse probability weights were used to estimate the direct effect of SGA on overweight. Results The crude and adjusted models yielded a null association (OR, 95% CI: 0.88, 0.26-2.96; 0.95, 0.28-3.29). Adjusting for later height reversed the effect (OR, 95% CI: 2.31, 0.52-10.26), and adjusting for later weight reversed the effect and provided a significant association (OR, 95% CI: 6.60, 1.10-37.14). The MSMs with height and weight considered as mediators indicated no direct effect of SGA on overweight (OR, 95% CI: 0.83, 0.14-5.01; 0.71, 0.18-2.81). Conclusions Overcontrolling for body size can falsely induce an association between SGA and overweight. Key messages Mediators should not be treated as confounders.

932Using marginal structural models to account for selection bias in the analysis of legacy effect

Background Analyses of legacy effects are mainly based on post-trial follow-up study after initial randomized controlled trials (RCT). However, the differential event rates between arms may cause a violation of randomization balance and induce selection bias in the post-trial analysis. We conducted a simulation to illustrate the bias and explore if marginal structural model (MSM) can address it. Methods Our simulation combined an RCT and an extended follow-up study. The scenarios investigated include different settings of direct treatment effect, legacy effect and underlying event rate. To fit the MSM, we used the inverse probability weighting method. The performance of MSM was compared to the standard model with and without adjustment of baseline covariate. Results Post-trial analysis without making adjustment resulted in biased estimates, and the bias increased with the underlying event rate and treatment effect. Both MSM and standard baseline covariate adjustment equally corrected for the bias if no patients took treatment after the trial. We are currently undertaking analysis for scenarios where some people continue treatment post-trial, and there is treatment confounder feedback; these results will be presented at the Congress. Conclusions Estimation of legacy effects using post-trial data without adjusting for differential survival between randomised treatment arms results in biased estimates. Although both standard covariate adjustment and MSM correct the bias if no patients take treatment in the post-trial period, MSM is expected to be the best method in the more realistic scenario where some patients continue to take treatment, and there is treatment confounder feedback. Key messages Post-trial analysis without making adjustment results in biased estimation of legacy effect. Marginal structural models may be used to address the selection bias.