Why Sparse Causes Create an Estimation Problem
In a competing-risks joint model, the longitudinal process can be linked to more than one event cause, with a distinct association parameter for each cause. A cause-specific association parameter describes how the modeled longitudinal feature—such as the current value of the underlying marker—is related to the hazard of that cause, conditional on the model. When one cause occurs infrequently, its association may be estimated imprecisely or behave erratically across fits.
The issue is not simply that the full study has a small sample. Information about a particular cause-specific association depends especially on how many events of that cause occur, which participants remain at risk, how much the longitudinal feature varies among those participants, and how well that feature is measured. Many observations or many events of another cause do not automatically provide strong information about the rare cause’s distinct association parameter.
For a current-value association, a representative cause-specific hazard is
where \(i\) indexes participants, \(k\) indexes event causes, \(h_{0k}(t)\) is the baseline hazard for cause \(k\), \(x_i\) is a vector of covariates, \(b_i\) denotes the participant-specific random effects, and \(m_i(t)\) is the modeled longitudinal value. The parameter \(\alpha_k\) is the cause-specific association. For example, a one-unit increase in \(m_i(t)\) multiplies the cause-\(k\) hazard by \(\exp(\alpha_k)\), conditional on the model and other covariates. The unit is the unit used for the longitudinal marker.
This form assumes the specified hazard and longitudinal models are adequate, the cause labels are correctly observed, and the shared random-effects structure and likelihood calculation are appropriate. The parameter’s meaning also depends on the chosen longitudinal feature and its scale. Sparse events do not invalidate the model by themselves; they make the data less informative about some of its parameters.
Information, Instability, and a Useful Approximation
In a likelihood analysis, information about \(\alpha_k\) is related to the curvature of the log likelihood near its maximum. A sharply curved likelihood indicates that substantially different parameter values fit less well; a flat likelihood indicates that a wider range of values may fit similarly. In a multivariate model, the relevant curvature is considered alongside the other parameters: strong correlation between \(\alpha_k\) and baseline, covariate, or random-effects parameters can further weaken its precision.
A simple Poisson rate comparison illustrates why few events can mean high uncertainty. Suppose there are two groups with fixed person-time, and the cause-specific event counts are independent Poisson variables. The log rate ratio is the difference between the log rates. Under this approximation, its estimated variance is the sum of the reciprocals of the two event counts. This is an illustration of sparse-event information, not a substitute for fitting the joint competing-risks likelihood, which also accounts for the modeled longitudinal process and competing events.
Worked Example: A Rare Cause and a Wide Approximate Interval
Consider hypothetical follow-up in two groups defined by a higher versus lower modeled marker value. For one event cause, the higher-marker group contributes 120 person-years and has 6 events; the lower-marker group also contributes 120 person-years and has 2 events. Assume, for this illustration only, independent Poisson counts with constant rates within groups.
The estimated rates are \(6/120=0.0500\) and \(2/120=0.0167\) events per person-year. The estimated rate ratio and log rate ratio are
For independent Poisson counts, the approximate standard error of the log rate ratio is
The approximate 95% Wald interval on the log scale is \(1.099\pm1.96(0.8165)\), giving \([-0.501,\ 2.699]\). Exponentiating its endpoints gives an approximate rate-ratio interval of \([0.606,\ 14.86]\). Although the estimated rate is three times as high in the higher-marker group, these data are compatible with a wide range of rate ratios, including values below one. The estimate is not precise. The calculation is a deliberately simplified approximation: a joint-model interval must come from its fitted likelihood or another method suited to that model.
Diagnosing a Weak Cause-Specific Estimate
Start with cause-specific event counts, not just the total number of participants or total number of events. Examine how many participants experience each cause and how the cause events are distributed across covariate and longitudinal-marker patterns. Also inspect the time at which events occur: a cause with few early events may have little information about later marker values, while a cause with few events overall may leave its association entangled with baseline-hazard or random-effects parameters.
Then assess the fit itself. Useful checks include whether different reasonable starting values lead to materially different estimates; whether the optimizer reports convergence and small gradients; whether the estimated variance-covariance matrix is positive definite; and whether the observed information has very small eigenvalues or a large condition number. A nearly singular information matrix or a very large standard error can indicate weakly identified parameter combinations. In joint models, repeat or refine numerical integration as appropriate and check whether the estimate changes materially; integration error can mimic or compound instability.
Influence checks can reveal whether one or two participants with the rare cause dominate the association estimate. Compare estimates after omitting such participants as a diagnostic, while recognizing that deletion is not a principled replacement for the intended analysis. Likewise, inspect whether the longitudinal trajectories for participants experiencing the rare cause actually span the marker values needed to distinguish a marker association from other model components.
No single event-count threshold guarantees stability. The information depends on the event process, risk sets, predictor distribution, repeated measurements, model specification, and correlation among parameters. A cause with a modest number of events might still yield useful information if the marker signal is strong and well represented; a larger count may remain uninformative if the relevant marker variation is limited.
Three Responses That Solve Different Problems
When an estimate is unstable, three tempting responses are to impose a parameter constraint, simplify the association structure, or report an uncertainty interval. They are not interchangeable. A constraint changes which parameter the model estimates; simplification changes the assumed relationship; an interval describes uncertainty under the fitted model.
| Approach | What changes | What it can help with | What it does not guarantee |
|---|---|---|---|
| Constrain associations across causes | Distinct parameters are replaced by a common parameter or another specified restriction. | Can reduce the number of free parameters and pool information when the restriction is scientifically defensible. | Does not preserve a distinct rare-cause association if the true associations differ. |
| Use a simpler association structure | The form or number of association terms is reduced, while parameters can remain cause-specific. | Can reduce weakly supported complexity, such as several terms for one cause when the data support only a simpler term. | Does not ensure precise estimation or make a misspecified structure correct. |
| Report an uncertainty interval | The estimate is accompanied by a range reflecting uncertainty under a chosen procedure. | Makes imprecision or asymmetry more visible than a point estimate alone. | Does not add information, stabilize a failed fit, or correct model misspecification. |
A shared-parameter constraint is therefore a substantive modeling assumption, not merely a numerical fix. A simpler structure may retain cause-specific associations but omit or combine features that the data cannot support. The choice should be justified by scientific context and the inferential question, not solely by which model produces the smallest standard error.
Worked Example: What Pooling Under a Constraint Changes
Suppose a hypothetical unconstrained fit gives cause-specific association estimates \(\widehat{\alpha}_1=0.20\), \(SE_1=0.15\), and \(\widehat{\alpha}_2=1.10\), \(SE_2=0.65\). Cause 2 is rare. For illustration, suppose the estimates are independent and approximately normal, and consider imposing the restriction \(\alpha_1=\alpha_2=\alpha\). Under these assumptions, the inverse-variance weighted estimate is
Here \(w_1=1/0.15^2=44.44\) and \(w_2=1/0.65^2=2.367\). Thus
The estimated shared hazard ratio per marker unit is \(\exp(0.2455)=1.278\). This is much more precise than the rare-cause estimate, but it estimates a common association, not cause 2’s distinct association. In this illustration, the estimate is pulled toward cause 1 because cause 1 has much greater weight. In an actual joint model, dependence between estimates and the full likelihood mean this elementary weighting calculation need not apply; the constrained model must be fitted directly.
Intervals: Showing Uncertainty, Not Removing It
A Wald interval uses an estimate and its estimated standard error, typically in the form \(\widehat{\alpha}_k\pm1.96SE(\widehat{\alpha}_k)\) for an approximate 95% interval. It relies on a reasonably accurate normal approximation near the maximum. With sparse events, the likelihood may be asymmetric or poorly approximated by a quadratic, so the Wald interval can be misleading. If \(\alpha_k\) is interpreted on the log-hazard scale, exponentiating the endpoints gives an interval for the hazard ratio per marker unit.
A profile-likelihood interval instead retains the parameter of interest at candidate values and maximizes the likelihood over the other parameters at each value. It can reflect asymmetry in the likelihood better than a Wald interval, though it still depends on the specified model and reliable likelihood computation. Neither interval repairs a flat or numerically unreliable likelihood. A bootstrap can be useful for assessing sampling variability, but with very few events of a cause, resamples may contain even fewer events or none; the resampling design and failures must be reported rather than hidden.
Worked Example: Wald and Profile Intervals Can Differ
Suppose a hypothetical joint-model fit for a rare cause yields \(\widehat{\alpha}_k=1.10\) with estimated standard error \(0.65\). The Wald 95% interval is
On the hazard-ratio scale, this is \([\exp(-0.174),\exp(2.374)]=[0.840,\ 10.74]\) per one-unit increase in the marker. Now suppose profiling the full likelihood yields a 95% profile-likelihood interval \([0.05,\ 2.55]\). Its hazard-ratio interval is \([\exp(0.05),\exp(2.55)]=[1.051,\ 12.81]\). The intervals differ because the likelihood is not well summarized by a symmetric quadratic approximation.
The profile interval excludes zero on the log-hazard scale, whereas the Wald interval includes zero. This difference signals that inference is sensitive to the interval method in this hypothetical fit; it is not a reason to select whichever result is preferred. The likelihood profile, numerical accuracy, model assumptions, and sparse-event context should be examined before drawing a conclusion.
Common Pitfall and Practical Interpretation
A common mistake is to treat a smaller standard error after constraining parameters as proof that the rare-cause association is now reliably estimated. The wrong reasoning is: “The standard error fell, so the data now estimate the rare cause more precisely.” The correction is that the constrained fit estimates a different quantity: a shared association under an equality assumption. Its precision may improve because information is pooled, but the assumption can be wrong, and the shared parameter does not answer a question about an unrestricted cause-specific association.
Similarly, a broad interval is not itself a remedy. It is a report of uncertainty under a chosen analysis. If the fit is numerically unstable, first investigate the likelihood calculation and parameterization. If it is stable but imprecise, report that imprecision and consider whether a scientifically supportable simpler model or constraint addresses the intended question. State clearly which association is being estimated and on what marker scale.
Check Your Understanding
Answer each question, then compare with the solutions.
- Why can a study with many total events still have a poorly estimated association for one event cause?
- In the Poisson illustration, suppose the high-marker group has 4 events and the low-marker group has 1 event, with equal person-time. Calculate the log rate ratio and its approximate standard error.
- What does imposing \(\alpha_1=\alpha_2\) change, and why is it not equivalent to reporting a wider interval?
- Name two diagnostics, beyond the standard error, that can reveal instability in a fitted cause-specific association.
- Why might a Wald interval and a profile-likelihood interval differ for a rare cause?
Solutions. (1) Events from other causes do not automatically provide strong information about a distinct association parameter for the rare cause; relevant risk sets, marker variation, and parameter correlations matter. (2) With equal person-time, the rate ratio is \(4/1=4\), so the log rate ratio is \(\log(4)=1.386\). Its approximate standard error is \(\sqrt{1/4+1/1}=\sqrt{1.25}=1.118\). (3) The equality restriction replaces two cause-specific parameters with one common association. A wider interval instead expresses uncertainty about a parameter under the fitted model; it does not impose equality or change the target in that way. (4) Examples include sensitivity to starting values, a nearly singular information matrix, large gradients or convergence warnings, sensitivity to numerical integration, and domination by one or two participants. (5) Sparse-event likelihoods can be asymmetric or poorly approximated by a quadratic, while a Wald interval uses a local normal approximation and a profile interval uses likelihood values across candidate parameter values.