FEV1 and Exacerbations in Chronic Obstructive Pulmonary Disease

Why Separate Chronic Decline From Acute Events?

FEV1, the forced expiratory volume in one second, is measured repeatedly in people with chronic obstructive pulmonary disease (COPD). A gradual decline can reflect chronic lung-function change, while an acute exacerbation can cause a temporary fall in measured FEV1. Those patterns are not interchangeable. If a temporary dip is treated as a lasting change in underlying lung function, an analysis may describe the chronic trajectory incorrectly. If an acute episode is ignored, it may also help explain why measurements or subsequent severe exacerbations occur when they do.

The case-study question is whether a person's underlying FEV1 level, or its rate of change, is associated with time to a first severe COPD exacerbation. The target is an association, not necessarily a causal effect. The analysis must define the longitudinal quantity being related to the event and distinguish that quantity from short-term changes around acute episodes.

A Joint Model With Chronic and Acute Components

Let \(Y_{ij}\) be the FEV1 measured for person \(i\) at time \(t_{ij}\), in years from a specified study origin. Let \(T_i\) be time from that origin to the first severe exacerbation, and let \(\delta_i=1\) if that event is observed and \(\delta_i=0\) if follow-up is right-censored. Let \(b_{0i}\) and \(b_{1i}\) be person-specific random intercept and slope terms. A simple chronic trajectory is

Here \(m_i(t)\) is the latent chronic FEV1 level, measured in liters, and \(m_i'(t)=\beta_1+b_{1i}\) is its rate of change, measured in liters per year. It is not the same as any one observed measurement. Let \(A_i(t)\) describe a recorded recent non-severe acute exacerbation history. For example, it could indicate whether a non-severe episode occurred during a specified recent window. A longitudinal observation model can separate an acute component from the chronic trajectory:

The function \(q\) represents the short-term shift associated with acute history; it could instead be specified using time since the episode, allowing a temporary effect to diminish during recovery. The residual \(\epsilon_{ij}\) represents measurement variation and other short-term variation not captured by the model.

A corresponding event model can relate the hazard of a first severe exacerbation to chronic FEV1, its slope, and recent acute history:

In this expression, \(h_0(t)\) is the baseline hazard, \(w_i\) is a vector of baseline covariates, and \(b_i=(b_{0i},b_{1i})\). The coefficient \(\alpha\) describes the association of current latent FEV1 level with the event hazard, conditional on the other terms. The coefficient \(\eta\) describes the association with the chronic rate of change, and \(\theta\) describes the association with recent acute history. For instance, \(\alpha\) has units of inverse liters, while \(\eta\) has units of years per liter. A negative \(\alpha\) means that higher chronic FEV1 is associated with a lower hazard, holding the other modeled terms fixed.

The model requires substantive and statistical assumptions. The mean trajectory must be adequate for the observed pattern; a straight line is a simplification, not a COPD law. Conditional on the random effects and modeled acute component, residual errors are commonly assumed to have mean zero, constant variance, and no remaining dependence across measurements. A common specification assumes \(b_i\) follows a bivariate normal distribution and is independent of the residual errors. The event model assumes proportional hazards given its covariates and latent trajectory. Censoring must be noninformative conditional on the modeled information. The recorded acute history must be defined consistently and measured well enough for its role in the model to be meaningful.

The association parameters are estimated jointly because the same random effects help describe both the repeated FEV1 measurements and the event risk. Conditional on \(b_i\) and the acute-history path, an individual's likelihood contribution has the form

Here \(S_i(t\mid b_i,A_i)\) is the probability of remaining free of the severe event through time \(t\), conditional on the model terms, and \(f(b_i)\) is the random-effects density. The event contribution is a hazard when the event is observed and a survival probability when it is censored. Integrating over the random effects accounts for their uncertainty; it also links the two data processes. This likelihood form conditions on the recorded acute-history path. If acute episodes are themselves informative and not adequately handled as observed covariates, their process may need to be modeled as well.

Worked Examples

Justification for the Hazard-Ratio Interpretation

At a fixed time, compare two covariate profiles with the same baseline hazard and the same other model terms. Subtracting their log hazards cancels those shared terms. If only latent FEV1 differs, then

This establishes why the coefficient is interpreted per unit difference in latent FEV1, not per unit difference in a single noisy measurement. If slope and acute-history terms also differ, their contributions must be added to the log hazard ratio; attributing the entire comparison to FEV1 level would then be incorrect.

Diagnostics, Alternatives, and Limitations

Diagnostics should examine whether the model's distinctions are supported by the data. Plot observed FEV1 values against time, marking acute episodes and severe events, and compare them with fitted chronic trajectories and acute-adjusted predictions. Systematic curvature suggests that a linear chronic trajectory may be inadequate. Residual patterns by time, patient, or time since an episode can reveal unmodeled structure. Check whether a small number of patients or events dominate the estimated association, and assess whether the proportional-hazards assumption is plausible. Compare fitted and observed event patterns over follow-up. Estimates can be unstable when there are few severe events, sparse FEV1 measurements, or little variation in slopes.

The main alternatives answer somewhat different questions. Separate longitudinal and event analyses do not directly account for the dependence created by shared patient characteristics; they may be useful descriptively, but their event association can be distorted when measurement and event processes are linked. Using the most recent observed FEV1 as a time-varying covariate is simpler, but it mixes measurement error, acute dips, and chronic level unless these are handled explicitly. A landmark analysis relates information available at a prespecified time to later risk, which can be easier to communicate but uses only people still under observation at that landmark and does not represent the full joint trajectory.

The joint model also has limits. Its result depends on the trajectory, random-effects distribution, acute-effect specification, and event model. An unrecorded episode or recovery process may be mistaken for chronic change. The association parameters are conditional model quantities, not automatically causal effects. If the clinical question is about the effect of an intervention, confounding and treatment processes require additional design and modeling considerations. The introductory material on joint modeling gives the broader framework; in this COPD application, careful definition of the chronic marker and the acute component is essential to interpreting the shared association.