Conference Agenda

Background: Lack of adherence can reduce the effectiveness of beneficial treatments. However, the extent to which adherence affects outcomes is unclear in many settings. For example, is it enough to adhere to treatment for 80% or 90% of the prescribed days? Does it matter whether adherence is higher earlier or later in the treatment schedule? And is it particularly important not to miss consecutive days of treatment? In this work, we explore a methodology to emulate and compare different adherence strategies. Specifically, we use a framework that incorporates explicit grace periods and regimes that depend on natural treatment patterns. Our work is motivated by a clinical question concerning women with early-stage hormone receptor-positive breast cancer, for whom daily endocrine therapy is prescribed for five to ten years. In these patients, young age is associated with both an increased risk of cancer recurrence and suboptimal adherence to endocrine therapy.

Methods: Using French nationwide claims data, we applied the proposed methods to compare the survival benefits achievable in patients with breast cancer by implementing different adherence strategies to endocrine therapy for each age group. We emulated three different adherence strategies that allowed for gaps in treatment of no more than one, three, or six consecutive months.

Results: A total of 121,601 patients were included in the analyses. In patients aged 34 years or younger, strict ET adherence (≤1-month gaps) improved 5-year DFS by 4.3 percentage-points, (95% confidence interval (CI): 2.6-7.2) compared to observed adherence. In this age group, ET adherence strategies allowing for ≤3-month and ≤6-month gaps reduced the 5-year DFS benefit to 1.3 (95% CI:0.2-3.7) and 1.0 (95%CI : -0.2-3.4) percentage-points, respectively. In contrast, DFS benefits of improved ET persistence in patients after 50 years old did not exceed 1.8 percentage-points, compared to observed persistence, regardless of the length of gaps allowed.

Conclusion: Our results show that young women would benefit substantially from stricter adherence to endocrine therapy, with treatment breaks never exceeding one month, highlighting the need for tailored strategies to improve treatment adherence in this population.

Many datasets involve staggered entries, where individuals join the study at different points in time. For example, randomized controlled trials (RCTs) in medicine usually recruit patients over time, and electronic health records contain information from the time a patient enters the healthcare system. Seminal works by Hernán, Brumback, and Robins (2000) and Hernan et al. (2008) on target trials used such datasets, treating the time an individual entered the study as a covariate in regression models. In this talk, I will describe a subtle - but frequently overlooked - positivity violation that appears in the analysis of staggered entry data. Because of this positivity violation, a frequently used class of Inverse Probability Weighting Censoring procedures leads to biased results. I will then propose new adjustment methods to circumvent such bias, and elaborate on how these fit with the target trial emulation framework. Finally, I will outline and compare these two approaches in settings wherein the bias varies in severity, thereby clarifying when one method is preferred over the other.

The target trial framework is a powerful methodology to estimate causal effects in observational studies by emulating randomized controlled trials. It addresses common biases in observational data, such as confounding and selection bias, by conceptualizing a sequence of hypothetical trials initiated at different time points. To increase precision, information is pooled across trials, typically under the assumption that the treatment effect is constant over time and across individuals. However, this can create ambiguity about the causal question when the effects vary or, as frequently happens, the population observed in some emulated trials systematically differs from the target population. Further challenges come when effects are parametrized using non-linear regression models (e.g. logistic regression) where non-collapsibility can create issues of misspecification when pooling different populations.
To address these challenges, this project introduces an assumption-lean strategy for target trial analysis, focussing on the choice of the estimand, rather than the choice of a model. This ensures that the analysis' aim is unequivocal regardless of model misspecification, and that uncertainty assessments reflect only information available in the data. Our proposal consists of several estimands, each related to different data structures and addressing different aspects of the patient population that may be of interest to researchers or decision-makers. For these estimands, corresponding estimators have been developed by the use of G-computation and inverse probability weighting. Applications on simulations and real data on antimicrobial de-escalation in an intensive care unit setting demonstrate the advantages of the proposed methodology over traditional techniques, offering greater clarity and reliability in causal effect estimation.

Background: A key challenge in medical statistics is that there remain many questions about the effects of treatments that are unlikely to be addressed in randomised trials. For example, in cystic fibrosis, long-term treatment effectiveness is a priority research area, but is difficult to address in trials due to feasibility and cost. When a randomised trial is not feasible, an alternative is to use observational data to ‘emulate’ a target trial. Target trial emulation (TTE) helps to avoid commonly occurring biases in observational studies, but it is unclear how best to apply TTE to data from existing cystic fibrosis registries.

The UK Cystic Fibrosis (CF) Registry, managed by the CF Trust, collects data on over 99% of the UK CF population. It includes longitudinal data on clinical variables collected at annual review visits and data on treatment prescription start and stop dates. Our aim is to establish best practices for the application of TTE using UK CF Registry data to estimate the effects of treatments in CF.

Methods and Results: We describe different methodological approaches to applying TTE to these data; a key consideration is how to define time 0. Previous studies applying TTE using these data have defined time 0 as the date of the annual review in which individuals meet the eligibility criteria for the emulated trial. They assume that treated individuals began treatment on the date of the annual review. In reality, individuals may have initiated treatment at any point between two consecutive annual review visits. An alternative approach is to make use of the data on prescription dates and there are different ways that time 0 could be defined for treatment initiators and non-initiators using these data. Another challenge is that the outcomes are often measured at different times (relative to time 0) for different individuals. The proposed methodological approaches are compared in a series of trial emulations of published trials in cystic fibrosis.

Conclusions: TTE is a commonly used approach to investigating treatment effectiveness using observational data, when a randomised trial is not feasible. However, it is not always clear how best to implement TTE using the available data. A key outcome of our study will be practical guidance for the application of TTE using UK CF Registry data. Our findings may be useful for other disease areas that benefit from patient registries.

Target trial emulation is a popular method for estimating effects of treatment regimes from observational data. In the emulation, new trials, indexed by time, are initiated at fixed intervals. A subject participates in every trial for which eligibility criteria are met. Current methods treat each time-specific trial separately. For instance, for a trial comparing the regimes “always” versus “never” treat from initiation at t onwards, it is common to fit a hazard or risk ratio (RR) model that includes a treatment indicator and its potential confounders. Subjects are censored if they later change treatment, with inverse probability weighting to adjust for the censoring. If most subjects change treatment, the estimates will be inefficient. In this paper we propose more efficient estimators by introducing regime-specific structural nested target trial emulation models (SNTTEM). Given a regime, a SNTTEM imposes parametric models for all time-specific blip functions of the eligible subjects and leaves those for the ineligible unrestricted. A time-specific blip function quantifies on a mean scale the effect of initiating the regime at a time t versus one period later, as a function of past history. The intersection of all the earlier time-specific RR models constitutes a single SNTTEM with regime “always take the treatment that one took last time”. We show that SNTTEM can be fitted using g-estimation, a method that censors less and is more efficient than current methods.