Statistical Methods Used to Predict Progression in Parkinson’s Disease: A Critical Review

Abstract

This thesis critically reviews the statistical methods used to predict progression in Parkinson’s disease, with emphasis on how different model families match the mathematical structure of the outcomes they are de- signed to predict. The review is organised as a structured narrative synthesis of a curated corpus of 50 references spanning Parkinson’s disease prognostic studies, methodological reviews, and reporting or ap- praisal guidance. Across the literature, progression is represented through continuous symptom trajectories, fixed-horizon binary outcomes, milestone-based time-to-event endpoints, multidomain deterioration, latent severity processes, and multimodal subtype or risk-classification tasks. These outcome structures motivate different statistical approaches, including regression and mixed-effects models, Cox and time-dependent sur- vival models, joint longitudinal-survival models, multivariate and latent-variable frameworks, and machine- learning systems. The central conclusion is that no single modelling family is universally best. Rather, methodological appro- priateness depends on the structure of the response variable, the role of repeated measures, the presence of censoring, dropout, or missingness, the required level of interpretability, and the intended clinical or research use of the prediction. The review also shows that many published prognostic models remain limited by weak handling of missing data, incomplete calibration assessment, inconsistent external validation, and inadequate reporting transparency. Its main contribution is therefore not simply to catalogue existing methods, but to clarify how model choice in Parkinson’s disease progression prediction should be driven by statistical structure and methodological maturity rather than by algorithmic novelty alone. The thesis therefore con- tributes a decision-oriented statistical framework for matching Parkinson’s disease progression outcomes to appropriate model families, with explicit attention to assumptions, repeated-measures structure, censoring, missingness, validation, calibration, interpretability, and intended use.