DAS28 and Structural Damage in Rheumatoid Arthritis

Why Link DAS28 to Structural Damage?

In rheumatoid arthritis, disease activity and structural damage are related but are not interchangeable. DAS28 summarizes disease activity using measures such as tender and swollen joint counts, an inflammatory marker, and a patient global assessment. Radiographs, by contrast, assess accumulated joint damage. A patient’s repeated DAS28 measurements can help describe changing inflammatory activity, while the time to a prespecified radiographic progression event can summarize structural outcome.

The measurements are repeated and imperfect: a DAS28 score at one visit is not the patient’s underlying activity at every moment. Further, damage is usually assessed at scheduled imaging visits, so its onset is often known only to have occurred between two radiographs. A joint model represents the underlying activity trajectory and relates it to the damage-event process in one model. The earlier material on joint models supplies the general framework; here the focus is how to specify and interpret it for DAS28 and radiographic damage.

The event criterion must be specified before analysis. For example, investigators might define progression using a validated radiographic scoring system and a prespecified increase from baseline. The exact scoring system, threshold, imaging schedule, and adjudication rules determine what “time to damage” means. The analysis should use one DAS28 variant consistently, such as DAS28-ESR or DAS28-CRP; their values and interpretation should not be treated as automatically interchangeable.

A Joint Model for DAS28 and Damage

A basic longitudinal model describes the observed scores around a patient-specific underlying trajectory. A linear trajectory is useful when its fit is reasonable over the study period:

Here \(m_i(t)\) is the latent DAS28 level for patient \(i\) at time \(t\); \(\beta_0\) and \(\beta_1\) are the population intercept and yearly change; and \(b_{0i}\) and \(b_{1i}\) are patient-specific deviations. A common specification assumes \((b_{0i},b_{1i})\) follows a mean-zero bivariate normal distribution with covariance matrix \(D\), and that the measurement errors \(\epsilon_{ij}\) are independent normal errors with mean zero and variance \(\sigma^2\), conditional on the patient effects. The normality and linearity assumptions are modeling choices to assess, not facts guaranteed by the scale.

One standard event submodel links the instantaneous damage hazard to the current underlying DAS28:

In this expression, \(\lambda_0(t)\) is the baseline hazard, \(Z_i\) is a vector of specified covariates, and \(\gamma\) contains their log-hazard coefficients. The parameter \(\alpha\) is the association between current underlying DAS28 and the damage hazard, conditional on the covariates and the modeled trajectory. For example, \(\exp(\alpha)\) is the hazard ratio for a one-point higher current latent DAS28, comparing patients with the same \(Z_i\) under this model. It is not a causal effect merely because it is expressed as a hazard ratio.

How Interval-Censored Damage Enters the Likelihood

Given a patient’s random effects \(b_i\), the conditional survival function is determined by the cumulative hazard \(H_i(t\mid b_i)=\int_0^t\lambda_i(u\mid b_i)\,du\):

If imaging shows no progression at \(L_i\) and progression at \(R_i\), the conditional probability of an event in that interval is \(S_i(L_i\mid b_i)-S_i(R_i\mid b_i)\). This is the probability of surviving through \(L_i\), minus the probability of surviving through \(R_i\). For a right-censored patient, the contribution is \(S_i(C_i\mid b_i)\). These contributions are conditional on the random effects; the joint likelihood also includes the repeated-score density and integrates over the random-effects distribution.

For a right-censored patient, the expression in braces for the event contribution is replaced by \(S_i(C_i\mid b_i)\). This integration matters: the unobserved patient trajectory affects both the expected DAS28 measurements and the damage hazard. A joint fit estimates the association while allowing for that shared, partly unobserved patient-level variation.

Worked Examples

Diagnostics, Alternatives, and Limitations

A numerical estimate is interpretable only if the model gives a credible account of both processes. For DAS28, inspect observed trajectories and residual patterns over time, including whether a straight-line mean is plausible and whether variability changes with the score or follow-up time. Check whether the assumed random-effects structure is supported and whether individual patterns show curvature that a linear model misses. For the event component, examine the fitted baseline hazard, event-time predictions, and calibration at clinically relevant follow-up times. Check convergence, sensitivity to starting values, and whether the association estimate is unusually sensitive to a small number of patients.

The current-value link is one choice, not a universal biological law. Damage may relate to recent activity, cumulative exposure, or rate of change; those are distinct association structures and should be motivated and compared carefully. A proportional-hazards form also assumes that the covariate effects in \(Z_i\) act proportionally over time as specified. If patients die before radiographic damage can be assessed, treating death as ordinary censoring requires an appropriate independent-censoring assumption. If that assumption is implausible, the scientific target and competing-event analysis need to be reconsidered.

A separate longitudinal model followed by a survival model using the last observed DAS28 is simpler, but it treats that observed score as if it measured current activity without error and may ignore the timing and pattern of earlier measurements. A joint model instead connects the event process to the latent trajectory. This can be valuable when repeated scores are noisy and their shared patient-level features are associated with event risk. It does not automatically remove bias from treatment changes, unmeasured confounding, informative visit schedules, or a misspecified event definition. Those design and modeling issues still require attention.

Radiographs are obtained at visits, so the true onset of progression is generally not observed exactly. Replacing an interval by its midpoint can be a rough approximation, but it discards the actual event-time uncertainty and may be poor when imaging intervals are long or irregular. Prefer an interval-censored event likelihood when the data and software support it. Finally, DAS28 is a clinical activity measure, not a direct measure of accumulated structural injury. A fitted association should be understood as a population-level relationship under the model, not as a guarantee that a particular patient with a high score will progress.