Hemoglobin and Organ Complications in Sickle Cell Disease

Why Hemoglobin and Complication Time Belong Together

In sickle cell disease, hemoglobin is measured repeatedly, while an organ complication may occur at any time during follow-up. The measurements can show a changing pattern, and that pattern may be related to the timing of the event. For example, participants with more severe disease may have lower hemoglobin, different treatment histories, and earlier complications. An analysis that treats every hemoglobin value as an independent observation, or that uses only one measurement, can miss this structure.

The aim here is to interpret a joint model for a repeated hemoglobin trajectory and time to a defined first organ complication. The endpoint must be specified in advance—for example, first adjudicated renal complication under stated diagnostic criteria. It is not enough to say “organ damage,” since different organs and definitions produce different event outcomes. All values and model results in the examples below are hypothetical.

A Joint Model for the Case

A common longitudinal submodel uses a linear time trend and participant-specific random intercept and slope:

Here \(\beta_0\) and \(\beta_1\) describe the population trajectory, \(b_{0i}\) and \(b_{1i}\) allow a participant to differ in level and slope, and \(\mathbf{A}_i(t)\) represents treatment information at that time. The random effects are commonly modeled jointly as a mean-zero normal vector, and residual errors are assumed independent conditional on the random effects and included covariates. In this model, the latent trajectory is \(m_i(t)=\beta_0+\beta_1t+b_{0i}+b_{1i}t+\boldsymbol{\gamma}^{\mathsf T}\mathbf{A}_i(t)\).

One standard event submodel relates the instantaneous complication rate to the current latent hemoglobin, treatment, and disease severity:

The hazard \(h_i(t\mid b_i)\) is the instantaneous event rate at time \(t\), conditional on remaining event-free just before \(t\). The baseline hazard \(h_0(t)\) is left unspecified in a Cox-type model. \(\mathbf{Z}_i\) can include baseline disease-severity measures, such as a prespecified clinical severity score. The parameter \(\alpha\) describes the association between the latent current hemoglobin level and the event hazard, conditional on the other modeled covariates. Because hemoglobin is measured in g/dL, \(\alpha\) is on a per-g/dL scale.

The model’s central feature is that the event submodel uses \(m_i(t)\), not the most recent observed \(Y_{ij}\). The latent value accounts for measurement error and for the fact that measurements are only collected at discrete visits. A different scientific question could motivate an association with slope, such as \(\alpha_1m_i(t)+\alpha_2m_i'(t)\), but that is a distinct association structure and should not be added merely because the data are repeated.

Assumptions and the Joint Likelihood

Interpretation depends on assumptions. The longitudinal mean and variability must be adequately represented; the event hazard must have the specified covariate form; and participants’ event outcomes must be conditionally independent given their modeled random effects and covariate histories. Censoring should be independent of the unobserved event time conditional on the modeled information. If death prevents observation of the organ complication and is related to its risk, treating death as ordinary noninformative censoring may be inappropriate; the endpoint and analysis must address that competing event.

The usual joint likelihood makes the shared structure explicit. Conditional on the random effects, the repeated measurements contribute their longitudinal densities, and the observed event time contributes a hazard-survival term. Integrating over the random-effects distribution gives:

Here \(S_i(t\mid b_i)=\exp\{-\int_0^t h_i(u\mid b_i)\,du\}\) is the conditional event-free survival probability. The term \(h_i(T_i\mid b_i)^{\delta_i}\) supplies an event density when \(\delta_i=1\), and equals one when the observation is censored; the survival term contributes in either case. The integral averages over uncertainty about the participant-specific trajectory rather than pretending the random effects are observed.

Reading the Trajectory Before the Hazard

A useful interpretation begins with the hemoglobin trajectory: its scale, direction, and variation across participants. The fitted population trend is not necessarily the trajectory of every participant. Random intercepts distinguish participants’ typical levels, while random slopes capture differing rates of change. Treatment can also shift or alter the modeled trajectory. The event association should then be interpreted as conditional on the modeled trajectory and covariates, not as an isolated comparison of raw hemoglobin readings.

Interpreting the Marker–Event Association

In the current-value model, two participants with the same modeled treatment and severity covariates but latent hemoglobin values differing by one g/dL have a conditional hazard ratio of \(\exp(\alpha)\), comparing the higher value with the lower value. This is a relative instantaneous event rate among participants still event-free at that time. It is not a ratio of cumulative probabilities, and it does not mean that changing hemoglobin by one unit would cause the hazard to change by that amount.

Treatment and Disease Severity in Context

Treatment is especially important to interpret carefully. In sickle cell disease, treatment may change over follow-up, and changes can depend on symptoms, prior hemoglobin, or clinical severity. A model can include treatment in the longitudinal and event submodels, but ordinary covariate adjustment does not automatically estimate a causal treatment effect. If treatment is initiated because a participant is worsening, treatment status can be associated with both the marker and event risk for reasons that the model has not fully captured. The timing and definition of treatment variables should be explicit.

Disease severity also matters. A severity covariate can help compare participants at similar measured severity, but it cannot remove confounding by unmeasured or poorly measured factors. The estimate \(\alpha\) is therefore a conditional association under the chosen model, not a biological mechanism established by the model.

Diagnostics, Alternatives, and Limitations

Diagnostics should examine both parts of the model. For hemoglobin, assess whether the fitted mean pattern is plausible over time, whether residual variability is adequately represented, and whether random-effect distributions or influential participants are driving the result. For the event model, assess proportional-hazards adequacy and whether the assumed current-value association is supported; a systematically changing association may suggest that the chosen form is inadequate. Compare observed and model-predicted longitudinal patterns and event-free survival, and check whether conclusions change under reasonable alternative trajectory or association specifications.

The observation schedule also matters. Standard formulations generally assume that visit times are noninformative given the modeled information. If participants are measured more often because their condition is worsening, the visit process can itself carry information; ignoring it may distort trajectory estimates. Sparse measurements, few complications, or limited variation in hemoglobin can make the association estimate unstable. A complicated random-effects structure may also be weakly identified in a small data set.

Separate longitudinal and survival analyses are simpler alternatives, but a survival model using the last observed hemoglobin can treat a noisy, irregularly timed value as if it were the true current level. A two-stage approach that first estimates individual slopes and then uses them in a survival model can understate uncertainty from the first stage. Joint modeling addresses the shared latent trajectory and event process in one likelihood, but depends on its modeling assumptions and can be computationally demanding. It does not automatically resolve confounding, informative visit timing, competing risks, or misspecification.

A common mistake is to interpret a negative \(\hat{\alpha}\) as proof that raising hemoglobin prevents organ complications. The model estimates a conditional association between the latent marker and hazard under its assumptions. Treatment selection, disease severity, and other factors may explain part of that association. Causal claims require a design and assumptions suited to causal inference, not just a joint model.