CD4 Counts and AIDS-Free Survival in HIV

Why Model CD4 Counts and AIDS-Free Survival Together?

In an HIV cohort, CD4 counts measured repeatedly can describe immune status over time, while AIDS or death marks a clinically important event. These processes are connected: a participant with declining immune status may be more likely to experience the event, and the event itself stops subsequent CD4 measurements. An analysis that treats the last observed CD4 count as an ordinary complete record can therefore miss important dependence between the biomarker history and the event process.

A joint model represents the underlying, partly unobserved CD4 trajectory and relates it to the event hazard. The case study below uses a current-value association: at each time, the event hazard depends on the participant’s latent CD4 level at that time. The resulting association is conditional and model-based. It does not establish that changing CD4 by intervention would cause the estimated change in event risk.

Model and Assumptions

Let \(t=0\) be cohort entry, with time measured in years. For participant \(i\), \(y_{ij}\) is the \(j\)th observed CD4 count, in cells per cubic millimeter, at time \(t_{ij}\). Let \(T_i\) be the time to the first of AIDS or death, and let \(\delta_i=1\) if that event is observed and \(\delta_i=0\) if follow-up is right-censored. This composite endpoint treats AIDS and death as a single event; it does not estimate separate cause-specific associations.

One illustrative longitudinal submodel is a linear mixed model:

where \(\mathbf{X}_i\) contains baseline covariates, \(b_{0i}\) and \(b_{1i}\) are participant-specific random intercept and slope, and the measurement errors \(\epsilon_{ij}\) are assumed independent of the random effects and normally distributed with mean zero and variance \(\sigma^2\). A common assumption is that \((b_{0i},b_{1i})^\mathsf{T}\) has a mean-zero multivariate normal distribution with covariance matrix \(\mathbf{D}\). A linear trajectory is a simplifying choice, not a biological law; plots and residual checks should support it over the follow-up period being analyzed.

For a current-value association, use the event hazard

where \(\lambda_0(t)\) is the baseline hazard, \(\mathbf{W}_i\) contains baseline event-model covariates, and \(\alpha\) is the association parameter. Centering at 500 makes the baseline hazard correspond to a participant whose latent CD4 count is 500 cells per cubic millimeter and whose covariates \(\mathbf{W}_i\) are at their reference values. The division by 100 makes \(\alpha\) the log hazard ratio for a 100-cell higher latent CD4 count, holding other modeled quantities fixed.

This specification relies on assumptions that need to be considered in context. Conditional on the random effects and included covariates, the longitudinal measurement errors and event process are modeled as independent; the link through \(m_i(t)\) captures their modeled dependence. The event model must describe the baseline hazard adequately, and the longitudinal mean and random-effects structure must be plausible. Censoring is assumed non-informative conditional on the modeled information: after conditioning on the relevant covariates and random effects, censoring should not carry additional information about the event time. CD4 measurement times and missingness also require scrutiny; a joint event model does not automatically correct informative visit timing or every form of dropout.

Reading the Estimated CD4 Trajectory

The model distinguishes observed counts from the latent trajectory. An observed count includes measurement variation, whereas \(m_i(t)\) describes the model’s estimated underlying mean at time \(t\). The random intercept and slope allow participants to begin at different levels and change at different rates. In practice, a participant-specific trajectory estimate has uncertainty, especially when there are few measurements or measurements are far from the time being interpreted.

Interpreting the Current-Value Association

Because the hazard model uses \([m_i(t)-500]/100\), the hazard ratio for two otherwise identical participants whose latent CD4 counts differ by 100 cells per cubic millimeter at time \(t\) is \(\exp(\alpha)\) for the higher-count participant compared with the lower-count participant. If \(\alpha<0\), higher latent CD4 is associated with a lower conditional hazard of AIDS or death. Reversing the comparison, the hazard ratio for a 100-cell lower count is \(\exp(-\alpha)\).

The interpretation follows directly by comparing the two hazards at the same time and covariate values. If the second participant’s latent CD4 count is 100 cells per cubic millimeter higher, the difference between their scaled latent values is one:

where \(x(t)=[m(t)-500]/100\). The shared baseline hazard and equal covariate terms cancel. This is a hazard comparison at a given time, not a comparison of cumulative event probabilities over follow-up.

From a Hazard Association to an Illustrative Event Probability

A hazard ratio does not directly give the ratio of event probabilities. To obtain a conditional event probability from a hazard model, one needs the cumulative hazard over the time interval. If the hazard is \(\lambda_i(u)\), the probability of the event by time \(t\), conditional on the modeled quantities, is \(1-\exp\{-\int_0^t\lambda_i(u)\,du\}\). When CD4 changes over time, the hazard changes with it, so the trajectory must be incorporated across the interval.

Model Checks, Alternatives, and Limitations

Diagnostics should address both parts of the model and the link between them. Plot observed CD4 values and fitted trajectories over time, inspect residual patterns and the distributional adequacy of the longitudinal model, and check whether a straight-line trajectory is plausible over the observed follow-up. Examine the event-model fit, including whether the chosen baseline hazard is sufficiently flexible. Sparse measurements, extreme counts, or a few influential participants can make the association estimate unstable. Also check that the fitted model matches the intended endpoint definition, time origin, covariates, and censoring scheme.

An especially important failure mode is to treat the estimated CD4 trajectory as known without uncertainty. In a joint model, uncertainty about the latent trajectory and its association with the event process contributes to inference. A plug-in analysis that first estimates CD4 trajectories and then treats them as exact can understate uncertainty. Conversely, a joint model is not automatically correct merely because it is joint: an implausible trajectory, inadequate event model, or informative visit process can still distort results.

The current-value model is one scientifically motivated choice, not the only possible association structure. A slope-based association asks whether the rate of CD4 change is associated with the hazard; a cumulative association asks about accumulated biomarker history. Those choices answer different questions and require their own clinical justification. A current-value association can also be sensitive to the time scale, functional form, and the extent to which the latent CD4 level captures the relevant biology.

Finally, AIDS and death may have different mechanisms, but this composite endpoint treats them alike for the event analysis. The estimated association is an average model-based relationship for the defined endpoint and cohort, conditional on included covariates and assumptions. It does not imply that CD4 alone explains event risk, and it should not be interpreted as evidence that an intervention that raises CD4 would reduce risk by the same amount.

Common Pitfall