Conference Agenda

Background

Studies have suggested that optimal midwifery staffing is associated with spontaneous vaginal births. The gold standard is a midwife-to-woman ratio of one-to-one during established labour and birth. In a clinical site, where the gold standard is established, it is not ethical to conduct a randomised controlled trial with less than standard care. In the real world though, it is not always possible to adhere to the gold standard due to high variation in care demand and ineffective measures to address high demand. Observational data will thus consist of gold standard and non-gold standard cases. To estimate causal effects, a target hypothetical pragmatic randomised trial was conceptualised with the routinely collected hospital data. The aim was to determine the causal effect of gold standard midwifery staffing compared to less than gold standard midwifery staffing on the occurrence of spontaneous vaginal births.

Methods

The target trial was emulated with routine hospital data from one tertiary hospital from 01 January 2019 to 31 December 2022. The exposure was defined as the proportion of time with one-to-one care during active labour, with 100% midwifery staffing as gold standard. All women in active labour were eligible to be included. Women with a planned caesarean section, breech birth, multiple pregnancy and stillbirth were excluded. Women assigned to each staffing exposure were assumed to be comparable conditional on baseline covariates. We considered hypertensive disorders, diabetes, gestational age, parity, maternal age, country of birth, birth weight and labour induction as confounding variables and adjusted for them with inverse probability weighting. We used marginal structural models to calculate the total effect based on the per-protocol policy. Ethical approval to access the routine hospital data has been granted by the local ethics committee.

Results

In total 6,602 cases were included in the analysis and 16.2% (n=1,072) were exposed to gold standard midwifery staffing. Overall, 61% (n=4031) had a spontaneous vaginal birth. Women receiving gold standard midwifery staffing are 5.7% (CI 2.3% – 8.9) more likely to have a spontaneous vaginal birth than women not receiving gold standard midwifery staffing.

Conclusion

The target trial emulation confirmed what cross-sectional studies already have indicated. Women receiving gold standard midwifery staffing are more likely to have a spontaneous vaginal birth. The target trial emulation can be used as a blueprint for further research of causal links between nurse and midwifery staffing and outcomes.

We describe an alternative encoding of time-varying treatment strategies. The motivation for this encoding is to simplify the exposition of identifiability assumptions in many settings of practical interest. As a clinically relevant running example, we consider sequential per-protocol effects, including sequential separable effects. The study of these effects would complement the traditional intention-to-treat analyses of an RCT. We derive new identification results for these per-protocol parameters, which can be described as conditional independencies in causal Directed Acyclic Graphs (DAGs) and Single World Intervention Graphs (SWIGs). Furthermore, we propose several estimators, including one that is semi-parametrically efficient and double-robust. These results are illustrated through an analysis of the Systolic Blood Pressure Intervention Trial (SPRINT), where we estimate a new type of time-varying separable effects, with a clear clinical interpretation as the per-protocol separable effect of taking a modified blood pressure treatment on acute kidney injury.

Consider an optimal dynamic treatment regime, correctly identified from a perfectly executed sequentially randomised experiment. Even when the experimental results are generalisable to a future target population, there is no guarantee that the optimal regime outperforms human decision-makers; human experts can do better than the optimal regime whenever they have access to relevant information beyond the covariates recorded in the experiment. Motivated by this fact, we derive results on a new class of regimes called initiation regimes. These regimes follow human decision-makers until it is more beneficial to initiate a sequential optimal regime from that point onward, and are guaranteed to outperform both entirely human and entirely algorithmic decision-makers, e.g., based on reinforcement learning algorithms. Furthermore, we present modified experimental designs that identify the best initiation regimes, show how the best initiation regime can be identified from classical observational data with commonly invoked assumptions, and give estimation and statistical inference methodology for these regimes. To illustrate the practical utility of our methods, we consider initiation regimes in a case study on back pain treatment.

Lung transplantation was a critical intervention for extending the lifespan of individuals with cystic fibrosis. Since transplant assignment cannot be randomized, evaluating treatment effectiveness relies on observational data. Such data—such as those provided by the United Network for Organ Sharing (UNOS)—offer a valuable opportunity to emulate a target trial. This methodology is widely used to investigate causal relationships using observational data, which inherently contain biases. One major source of bias is confounding due to non-random treatment assignment. Additionally, improper implementation of the emulated trial framework can introduce further biases, such as immortal time bias, which complicates the estimation of treatment effects. Correcting for this bias, however, can induce informative censoring bias, adding another layer of complexity to the analysis.

In this work, we use UNOS data as a case study to develop a methodological framework for emulating target trials in the context of lung transplantation. We address the challenges associated with different types of bias, leading to a sequence of target trials that incorporate time-dependent matching based on the lung allocation score (LAS), the primary known confounder affecting treatment assignment. Specifically, our design involves setting a sequence of time landmarks (e.g., weekly follow-ups) to minimize the time between follow-up initiation and transplantation for the treated group. Each treated individual is matched with a control - who is on the waiting list at the current time - ensuring similarity based on LAS. If a control later receives a transplant, they are artificially censored and subsequently included in the treated population in a later trial. To adjust for this informative censoring, we apply the Inverse Probability of Censoring Weighting (IPCW) method with time-varying weights. We present our results as survival curves aggregated from Kaplan-Meier estimators fitted on each weekly trial.

Finally, we explore a theoretical question regarding the convergence of such an aggregated estimator when based on dependent data, as individuals on the waiting list can participate in multiple trials. We propose to investigate the level of dependency as a function of the total number of individuals available for the study through simulations, analyzing the effects of sequentialization, matching, and censoring.

Background The causal effects of exposures often vary across an individual's lifetime, with certain periods exerting a greater influence on health outcomes or revealing the long-term consequences of risk factors. Capturing these time-varying causal effects provides valuable insights into underlying mechanisms and supports the development of effective health interventions. Multivariable Mendelian Randomization (MVMR) offers a robust framework to estimate the direct effect of any exposure on an outcome, accounting for the influence of other concurrent exposures. However, methods for accounting for time-varying exposures remain limited in the literature.

Methods We propose a Multivariable Functional Mendelian Randomization (MVFMR) approach to estimate the direct effects of multiple time-varying exposures on an outcome. Our method builds on the framework introduced by [1]. First, we apply functional principal component analysis (FPCA) to reduce dimensionality by extracting low-dimensional factors from exposure trajectories. Second, we introduce a data-driven feature selection step to determine the optimal number of functional principal components (FPCs), eliminating the need for predefined selection criteria. Finally, we develop the MVFMR model to assess the direct causal associations of multiple time-varying exposures with the outcome, providing a comprehensive evaluation of their effects across an individual's lifetime.

Results Through simulations, we demonstrate that our feature selection approach effectively identifies the most accurate functional form of the association between time-varying exposures and outcomes. The strength of this approach is that it selects the optimal number of principal components without relying on a priori definitions based on the variability explained by them, criteria that our simulations demonstrate to be unreliable. Furthermore, we demonstrate that our method outperforms separate models in accurately estimating the time-varying effects of multiple exposures when all are associated with the outcome. We applied the proposed methodology to investigate the impact of time-varying genetically predicted systolic blood pressure and LDL cholesterol on the risk of coronary artery disease, using data from the UK Biobank.

Conclusions This study highlights the importance of integrating functional data analysis within the Mendelian Randomization framework to understand how risk factors evolve over a lifetime and estimate their causal effects. This approach enables the identification of critical exposure periods, providing valuable clinical insights that can inform targeted healthcare strategies.

[1] Tian, H., Patel, A., & Burgess, S. (2024). Estimating Time-Varying Exposure Effects Through Continuous-Time Modelling in Mendelian Randomization. Statistics in medicine, 43(27), 5166–5181. https://doi.org/10.1002/sim.10222