Methods/Statistics
Causal Inference in Real-World Dementia Research: A Systematic Review Yi Yang* Yi Yang Yang Yang Yang Yang Department of Psychiatry, University of Oxford, Oxford, UK
Background: Dementia presents major challenges for causal inference due to its multifactorial aetiology, long preclinical phase, and heterogeneous progression. Randomized controlled trials are often impractical or unethical for many dementia-related exposures, leading to heavy reliance on observational studies. However, traditional observational designs, particularly prevalent exposure designs, are highly susceptible to bias, including survivor bias, misaligned time zero, time-varying confounding, and poorly defined interventions. To address these limitations, dementia research has increasingly adopted causal inference frameworks that aim to emulate hypothetical interventions and improve causal interpretability. This systematic review critically evaluates how causal inference methods are applied in observational dementia research, assesses their methodological rigor, and identifies emerging trends and gaps.
Methods: To address the urgent need for a field-level evaluation of how causal inference methods are applied in dementia research, an issue not amenable to single empirical studies, we conducted a systematic review. Following PRISMA guidelines, we searched MEDLINE, EMBASE, Web of Science, PsycINFO, Scopus, and the Cochrane Library for studies published between 1960 and 2024. Eligible studies were observational in design and explicitly applied causal inference methods to investigate dementia-related outcomes, including cognitive decline, disease onset or progression, mortality, and quality of life. Data extraction captured study characteristics, causal frameworks, assumptions, analytical methods, and reporting practices. Risk of bias was assessed using ROBINS-I. Findings were synthesized narratively, with meta-analysis conducted where methodological homogeneity permitted.
Results: From 11,009 records identified, 345 studies were included. Mendelian randomization was the most used causal method, followed by marginal structural models and instrumental variable analyses, with increasing use of hybrid approaches integrating machine learning. Substantial heterogeneity was observed in the definition of causal estimands, handling of time-varying confounding, and transparency of assumptions. Risk of bias varied widely, with frequent concerns related to residual confounding, weak instruments, selection bias, and limited sensitivity analyses. Despite these limitations, several studies demonstrated the potential of causal methods to address questions not amenable to randomized trials.
Conclusions: Causal inference methods offer essential tools for strengthening causal interpretation in dementia research where experimental evidence is limited. However, their application remains inconsistent. Greater alignment with explicit causal frameworks, improved reporting standards, and more rigorous sensitivity analyses are needed to enhance validity and reproducibility. This review highlights both the promise and current limitations of causal inference in dementia research and informs future methodological development.
Systematic review registration PROSPERO (CRD42024619228).