Cognitive Scores and Alzheimer’s Disease Progression

Why Link Cognitive Change to Clinical Progression?

In an Alzheimer’s disease cohort, cognitive scores may be collected repeatedly, while time to loss of independent living or death is recorded as a time-to-event outcome. These data are connected: worsening cognition may be associated with a higher rate of reaching the clinical endpoint, and people who become more impaired may be less likely to complete later cognitive assessments. An analysis of scores alone can therefore miss their relationship with clinical progression, while a survival analysis that uses only a baseline score discards subsequent measurements.

A joint model represents the cognitive trajectory and the event process together. The example here is hypothetical: it illustrates how to analyze repeated cognitive scores and a composite endpoint defined as the first occurrence of loss of independent living or death. A composite endpoint does not distinguish its components; the interpretation concerns time to the first event in that definition.

Model and Assumptions

Let \(Y_{ij}\) be the cognitive score for participant \(i\) at time \(t_{ij}\), measured in years since enrollment. Suppose higher scores indicate better cognitive function, and scores are measured in points. Let \(T_i\) be the observed event or censoring time and let \(\delta_i=1\) if the endpoint occurred at \(T_i\), and \(\delta_i=0\) if the participant was right-censored. Let \(b_i=(b_{0i},b_{1i})^\mathsf{T}\) denote participant-specific random effects.

A simple longitudinal submodel is a linear mixed model with a random intercept and slope:

Here \(m_i(t)\) is participant \(i\)’s underlying, or latent, mean score at time \(t\). A common assumption is \(b_i\sim N(0,D)\), where \(D\) is a \(2\times2\) covariance matrix, and \(\epsilon_{ij}\sim N(0,\sigma^2)\), independently across measurements and independently of \(b_i\). Covariates can be added to the mean if they are part of the scientific question; the examples below focus on the trajectory and its link to the event.

For the event submodel, use a proportional-hazards model linked to the current latent score:

In this expression, \(h_i(t\mid b_i)\) is the instantaneous hazard of the composite endpoint at time \(t\), conditional on the participant’s random effects and covariates \(w_i\); \(h_0(t)\) is an unspecified baseline hazard; and \(\alpha\) is the current-value association parameter. Because higher scores represent better function, a negative \(\alpha\) means that higher current scores are associated with a lower hazard, after conditioning on the included covariates and model. This is an association, not a causal effect.

The model relies on substantive and statistical assumptions. The chosen mean trajectory must be adequate over the observed follow-up; the random-effects and residual distributions must be reasonable; and the hazard must be proportional with the specified current-value link. Conditional on the shared random effects and included covariates, the longitudinal measurements and event process are modeled as independent. Censoring is assumed non-informative conditional on the modeled information: after conditioning, censoring should not carry additional information about the participant’s unobserved event time. These assumptions are not guaranteed by fitting the model; they require scientific justification and diagnostic checking.

How the Likelihood Connects the Two Outcomes

Conditional on \(b_i\), the longitudinal contribution is the product of the measurement densities. The event contribution must reflect whether the endpoint was observed or the participant was censored. Write \(f_i(t\mid b_i)\) for the event-time density and \(S_i(t\mid b_i)\) for the event-free survival probability. The observed event-data contribution is

Thus an observed event contributes its event-time density, whereas right-censoring contributes survival through the censoring time. Combining the longitudinal and event contributions conditional on \(b_i\), then integrating over the random-effects distribution, gives the participant’s likelihood contribution:

This integration matters: the random effects are not observed, and the longitudinal measurements inform their distribution for each participant. Through the shared random effects and latent trajectory, the score history and event outcome contribute jointly to estimation.

Worked Interpretation for a Cognitive Trajectory

A current-value link concerns the latent score at a given time, not the score’s rate of change by itself. The distinction is important in interpreting a trajectory: a person’s slope can help determine their current latent score, but the hazard model above contains no separate slope term.

Event and Censoring Contributions in Practice

The survival contribution is often easiest to understand through the cumulative hazard \(H_i(t\mid b_i)=\int_0^t h_i(u\mid b_i)\,du\). The survival function is \(S_i(t\mid b_i)=\exp\{-H_i(t\mid b_i)\}\), and the event-time density is \(f_i(t\mid b_i)=h_i(t\mid b_i)S_i(t\mid b_i)\). These identities show why an observed event and a censored observation contribute different expressions to the likelihood.

Diagnostics, Alternatives, and Limitations

A fitted association is useful only if the model captures the important features of the data. Plot individual and population-average score trajectories against time, checking for curvature, floor or ceiling effects, and changes in measurement patterns. A linear trajectory may be a poor approximation if decline accelerates; residual patterns or systematic departures by follow-up time can reveal this. Check whether the assumed random-effects structure is plausible, whether numerical warnings or unstable estimates occur, and whether estimated hazards and survival patterns are compatible with the observed event experience. The earlier material on diagnosing numerical warnings and checking model output provides general fitting checks; these do not replace disease-specific assessment of the cognitive scale and endpoint.

The assessment process itself deserves attention. Participants with worsening cognition may miss visits, become unable to complete a test, or reach the endpoint before another score is collected. A joint model can account for the association between the modeled trajectory and event time, but it does not automatically correct every form of missingness. In particular, missing scores caused by processes not captured by the shared random effects can still bias inference. The cognitive instrument’s scoring direction, validity over disease stage, and any practice or measurement effects should also be considered.

The main alternatives answer different questions. A baseline-score survival model is simpler but ignores later cognitive measurements. Separate mixed-effects and survival models can describe both outcomes, but the survival analysis may not account for the dependence between the evolving score and event time. A two-stage procedure that estimates a trajectory first and then inserts fitted scores into a survival model can understate uncertainty and can treat estimated scores as if they were observed without error. Joint modeling propagates uncertainty through a shared likelihood, at the cost of stronger modeling assumptions and more demanding estimation.

The current-value model also has limits. It assumes the hazard is linked to the present latent score through the specified function, and it summarizes the composite endpoint rather than distinguishing loss of independent living from death. Its coefficients describe conditional associations within the model, not treatment effects or causal mechanisms. Conclusions may be sensitive to the assumed trajectory, hazard form, random-effects distribution, and censoring assumptions. When any of these are doubtful, report that uncertainty and examine scientifically reasonable alternatives rather than treating one fitted model as definitive.

Common Mistake