ADDRESSING CONFOUNDING IN MIDWIFERY RESEARCH: FROM OBSERVATIONAL STUDIES TO RANDOMIZED CONTROLLED TRIALS

This doctoral thesis investigates the use of causal inference methods in midwifery and maternity care research, with the goal of enhancing causal interpretation validity in observational and randomised study designs. In contemporary maternity care research, the estimation of causal effects is often hampered by problems of confounding, selection mechanisms and the complexity of clinical and organisational pathways. Here, causal inference provides a unified conceptual and methodological framework to overcome these limitations by explicitly formalising assumptions about data-generating processes and by distinguishing between association and causation. The thesis is organised around the central concept of exchangeability, the fundamental condition required for valid causal inference. In observational studies, exposure is not randomly assigned, so exchangeability is usually not guaranteed. This may lead to confounding by baseline prognostic characteristics. In randomised controlled trials, exchangeability is anticipated by design through random allocation, but finite sample variability may still result in residual imbalance in important prognostic factors. In both study designs, causal assumptions are explicitly represented using Directed Acyclic Graphs (DAGs) and counterfactual reasoning to guide covariate selection and identify potential sources of bias. The methodological approaches for confounding control in observational research including regression adjustment, standardisation, inverse probability of treatment weighting (IPTW) and related g-methods are presented and discussed. These approaches are situated within the target trial emulation paradigm, which aims to reconstruct as closely as possible the conditions of a hypothetical randomised experiment from observational data. The strengths and limitations of each technique are demonstrated through simulated clinical scenarios from midwifery practice. The empirical part of the thesis is split into two main applications. The first study examines the association between the organisational level of maternity units and maternal birth satisfaction in an observational multicenter cohort of low-risk women. The non-random allocation of women to different hospital settings dictates the use of IPTW based on the propensity scores to address baseline imbalances notably concerning parity and other obstetric characteristics, to improve covariate balance and bolster causal interpretation. The second is a randomised controlled trial of the effectiveness of maternal postural techniques and rebozo manoeuvres to reduce persistent occiput posterior position in labour. In this experiment, stratified randomisation is conducted to improve baseline comparability for parity, which is an important prognostic factor for intrapartum outcomes. In general, the results of this thesis suggest the importance of explicitly addressing causal assumptions in both observational and experimental research in midwifery. While randomisation remains the gold standard design for causal inference, observational studies require carefully specified analytic strategies to approximate exchangeability. Thoughtful design choices are also helpful for randomised trials when dealing with strong prognostic factors. In conclusion this work provides a consistent framework to incorporate statistical methodology, study design and clinical reasoning in maternity care research through the perspective of causal inference. In doing so it helps strengthen the validity and interpretability of evidence in midwifery research by bridging observational and randomised approaches within a common conceptual framework which in turn supports more robust and clinically meaningful decision-making in maternity care.