Tompsett Daniel, Vansteelandt Stijn, Grieve Richard, Dixon Will, Gomes Manuel
Department of Primary Care and Population Health, UCL, London, UK.
Department of Applied Mathematics, Computer Science, and Statistics, University of Ghent, Ghent, Belgium.
BMC Med Res Methodol. 2025 Sep 2;25(1):208. doi: 10.1186/s12874-025-02625-y.
The increased availability of large-scale longitudinal data offers important opportunities to assess the causal effects of health interventions. In this setting, Instrumental Variable (IV) approaches have the potential to reduce the risk of bias from confounding due to unmeasured variables. However, there has been a lack of attention given to the development of IV approaches in settings when both the instrument and the potential confounders vary over time. In this paper we critically evaluate two instrumental variable approaches in time-varying settings.
The paper extends an existing g-estimation method that incorporates time-fixed IVs and compares it to an inverse probability weighting approach that incorporates time-varying IVs. A simulation study investigates the relative performance of these two approaches under varying scenarios. These methods are applied to a retrospective cohort from the US National Databank for Rheumatic Diseases, evaluating the sustained use of Adalimumab (Humira) versus other biologics on the health-related quality of life of patients with Rheumatoid Arthritis. Our case study considers physicians preference for Adalimumab as an instrument.
The g-estimation approach provided unbiased, precise estimates of treatment effects, across a wide range of scenarios, including weak IVs and complex time-varying confounding mechanisms. The performance of the weighting approach was reasonable in scenarios with a moderate or strong time-varying IV, but deteriorated with weak IV strength. Both methods suggest that sustained treatment with Adalimumab does not improve the health-related quality of life of rheumatoid patients, compared to other biologics, but the g-estimation approach led to narrower confidence intervals.
The proposed IV-based g-estimation approach can be reliably used in the estimation of time-varying treatments if a valid time-varying IV is available. The weighting approach offers an accessible alternative but was found to work well only when the IVs are strongly associated with treatment assignment, which is relatively unlikely in real-world applications.
大规模纵向数据可用性的提高为评估健康干预措施的因果效应提供了重要机遇。在这种情况下，工具变量（IV）方法有潜力降低因未测量变量导致的混杂偏倚风险。然而，在工具变量和潜在混杂因素均随时间变化的情况下，对IV方法的开发缺乏关注。在本文中，我们批判性地评估了时变情况下的两种工具变量方法。
本文扩展了一种现有的包含时间固定IV的g估计方法，并将其与一种包含时变IV的逆概率加权方法进行比较。一项模拟研究调查了这两种方法在不同情况下的相对性能。这些方法应用于美国国家风湿病数据库的回顾性队列，评估阿达木单抗（修美乐）与其他生物制剂的持续使用对类风湿关节炎患者健康相关生活质量的影响。我们的案例研究将医生对阿达木单抗的偏好作为一种工具。
在包括弱IV和复杂时变混杂机制在内的广泛情况下，g估计方法提供了无偏且精确的治疗效果估计。加权方法在具有中度或强时变IV的情况下表现合理，但在IV强度较弱时性能会下降。两种方法均表明，与其他生物制剂相比，阿达木单抗的持续治疗并不能改善类风湿患者的健康相关生活质量，但g估计方法导致的置信区间更窄。
如果有有效的时变IV，所提出的基于IV的g估计方法可可靠地用于时变治疗的估计。加权方法提供了一种易于使用的替代方法，但发现仅当IV与治疗分配密切相关时才有效，而这在实际应用中相对不太可能。
