UPDRS Scores and Parkinson’s Disease Milestones

Why Link Motor Scores to a Progression Milestone?

In Parkinson’s disease, repeated motor assessments and time to a clinical milestone describe related but different aspects of a patient’s course. The Unified Parkinson’s Disease Rating Scale motor examination, or the motor examination of the MDS-UPDRS, records observed motor impairment at visits. A progression milestone might instead be the first independently adjudicated transition to a specified disability state. The score varies over time, while the milestone time is observed once, or not at all during follow-up.

Medication makes this pairing especially important to interpret carefully. A patient’s observed motor score can improve after dopaminergic medication even while underlying disease severity is worsening. If that score is treated as a direct, unchanging measurement of progression, the association between motor impairment and milestone risk can be distorted. A joint model can relate the event hazard to a latent motor trajectory while representing medication state in the score model.

For this case study, suppose the marker is a repeated motor score, with higher values indicating worse motor impairment. Define the event as the first independently adjudicated transition to a specified clinical disability milestone, such as Hoehn–Yahr stage 3. The endpoint should be defined consistently and not merely be a threshold crossing in the same score being modeled; otherwise the marker and outcome may be mechanically coupled.

A Joint Model for UPDRS and Milestone Time

Let \(Y_{ij}\) be patient \(i\)’s motor score at visit time \(t_{ij}\), measured in years from a common origin. Let \(A_{ij}\) indicate that the patient was assessed in the medication ON state, with \(A_{ij}=1\) for ON and \(0\) for the reference state. A simple linear mixed-effects trajectory with a medication-state term is

Here \(m_i(t)\) is the patient’s latent motor severity on the chosen reference medication state; \(X_i\) contains baseline covariates; and \(b_{0i}\) and \(b_{1i}\) are patient-specific random intercept and slope. The residuals \(\epsilon_{ij}\) are commonly assumed independent normal errors with mean zero and variance \(\sigma^2\), conditional on the random effects and modeled covariates. The coefficient \(\delta\) represents an average score difference associated with assessment in the ON state, after the modeled trajectory and covariates are accounted for. It is not automatically a causal medication effect.

Let \(T_i\) be time to the milestone or censoring time, and let \(\Delta_i=1\) if the milestone occurs during follow-up and \(0\) if follow-up is right-censored. A shared-parameter event model can be written

The baseline hazard \(h_0(t)\) is the event hazard for the reference covariate and motor-severity values. The reference severity \(m_{\mathrm{ref}}\) is a convenient centering value, and \(A_i(t)\) may represent the medication state or exposure history specified in the analysis. The coefficient \(\alpha\) measures the conditional association between current latent motor severity and milestone hazard; \(\theta\) accounts for the chosen medication term in the event model. A positive \(\alpha\) means higher latent motor severity is associated with a higher instantaneous milestone rate, conditional on the modeled covariates and random effects. This is an association, not a claim that changing a score or medication would cause the event.

Assumptions and Likelihood Justification

The interpretation depends on the model being a reasonable description of the data. The linear trajectory assumes that the latent severity trend is adequately represented by a straight line over the time range studied; nonlinear time terms or splines may be needed otherwise. Random effects are commonly modeled as jointly normal, and the residual distribution and variance structure must fit the score data adequately. Since a motor score is bounded and assembled from ordered item ratings, a normal linear model is an approximation; a model for ordinal items or an appropriate non-normal response may be preferable when the approximation is poor.

The event model assumes the specified hazard form is adequate, including the chosen baseline hazard, the current-value link, and any medication term. The shared random effects must capture the relevant dependence between scores and event times. Conditional on these effects and covariates, the model typically assumes the longitudinal measurement process and event process are independent. Visit timing should be non-informative given the modeled history, or modeled explicitly if patients are assessed more often when worsening. Censoring is assumed conditionally independent of the unobserved event time given the included information. These assumptions are substantive, not guaranteed by fitting software.

The joint likelihood makes the link explicit. Write \(b_i=(b_{0i},b_{1i})^\top\), and let \(f(b_i)\) be its random-effects density. Given \(b_i\), the longitudinal contribution is the product of score densities, while the event contribution is the hazard at the observed event time multiplied by the survival probability through the observed follow-up time. Integrating over the unobserved random effects gives the patient-level contribution:

The event term follows from the survival density: an observed event contributes \(h_i(T_i\mid b_i)\exp[-\int_0^{T_i}h_i(u\mid b_i)\,du]\); a censored observation contributes only the survival probability. Multiplying the two conditional contributions and averaging over the random-effects distribution produces the joint contribution. This integration is why the procedure does not simply treat an estimated score trajectory as known without error.

Worked Interpretation and Medication Examples

Diagnostics, Alternatives, and Limitations

A plausible fit requires checking both parts of the model and their connection. Plot observed score trajectories by patient and compare them with fitted trajectories; look for curvature, changing variability, ceiling or floor effects, and systematic residual patterns. Check whether ON/OFF assessments, dose, and visit timing are recorded consistently. Compare predicted and observed event patterns, including calibration over clinically relevant horizons, and examine whether the fitted hazard adequately represents time dependence. Sensitivity analyses can assess whether conclusions change under nonlinear trajectories, alternative baseline hazards, different medication specifications, or a different association structure, such as linking the hazard to the latent slope as well as the current value.

A common alternative is a two-stage analysis: first estimate each patient’s trajectory, then insert the estimated score or slope into a survival model. This is simpler, but the second stage often treats uncertain estimated trajectories as if they were measured without error. Patients with sparse visits may have especially uncertain estimates, and event-related dropout can make the available trajectories selective. A joint model handles this dependence and uncertainty within one likelihood, but it does not automatically repair a misspecified trajectory, poorly measured medication exposure, informative visit schedules, or unmeasured confounding.

Other limitations matter clinically. A linear mixed model may not reflect heterogeneous or nonlinear progression, and a single additive medication coefficient may be inadequate when dose, timing, and response vary. If medication is initiated or intensified because a clinician observes worsening, adjustment for medication can be confounded by indication. The chosen milestone may also be affected by clinical practice or assessment frequency. A joint model provides a structured association, not a guarantee that the latent trajectory equals biological progression or that the fitted risk transports to a different clinical setting.