Why a Convergence Warning Matters
A regression routine can stop with a convergence warning, return implausibly large coefficients, or report a very wide interval even when its syntax is valid. These are not merely software inconveniences. They can indicate that the data contain too little information to estimate the requested parameters, that a covariate is numerically awkward, or that event status has been coded incorrectly. A successful optimizer exit does not by itself establish that the model is appropriate or that its estimates are reliable.
This tutorial focuses on diagnosing those problems in competing-risk regression. The same checks apply whether the model is fitted separately for a cause-specific hazard or for a subdistribution hazard. Those are distinct model targets; switching between them is not a numerical repair. The starting point is to state which event is being modeled and confirm that the data and model represent it correctly.
For a cause-specific Cox model, a common specification is
where \(\lambda_k(t\mid X)\) is the instantaneous rate of cause \(k\) at time \(t\) among subjects still event-free, \(X\) is the covariate vector, and \(\lambda_{k0}(t)\) is an unspecified baseline hazard. Its usual partial-likelihood inference assumes independent subjects, correctly constructed risk sets, and proportional cause-specific hazards for the modeled covariates; censoring must be independent of the event process conditional on modeled information. A subdistribution regression has its own proportional subdistribution-hazard assumption and risk-set construction. Neither model can compensate for incorrect event codes or an unsupported number of parameters.
A Diagnostic Sequence
Before changing optimizer controls, check the data and the information available for the model. A practical sequence is:
Identify the event cause and whether the fitted model is cause-specific or subdistribution. Verify that the selected method matches the scientific question rather than choosing whichever fit converges.
Check event counts by cause, censoring counts, time ranges, missing values, and impossible combinations. Verify the software's required status codes and reference levels.
Look for rare causes, empty covariate-by-cause cells, near-constant predictors, extreme values, strong collinearity, and a parameter count large relative to the number of events.
Make one justified change at a time, then compare estimates, standard errors, convergence messages, and fitted predictions. Record whether the change altered the model or only improved its numerical representation.
Sparse Causes and Separation
A cause with few events provides little information for estimating many regression coefficients. Sparse data become especially problematic when a covariate level has no observed events of the modeled cause. In a Cox partial likelihood, separation can occur when the likelihood keeps improving as a coefficient moves toward positive or negative infinity. The optimizer may report a huge estimate, fail to converge, or stop at a large finite value that depends on numerical tolerances.
The principle can be seen from a binary predictor \(X\). Suppose every event of the target cause occurs among \(X=0\), while people with \(X=1\) are still in the corresponding risk sets. A very negative coefficient makes \(\exp(\beta)\), the relative hazard for \(X=1\), approach zero. That increasingly favors the observed ordering, so the partial likelihood may not have a finite maximizing value. The estimate is not evidence that the true effect is infinitely protective; it is evidence that the data do not identify a finite coefficient in that direction.
Worked Example: A Sparse Cause and an Unstable Coefficient
Suppose a hypothetical cohort has 80 subjects: 22 with a binary exposure and 58 without it. During follow-up, the target cause occurs in 6 unexposed subjects and in none of the exposed subjects. Other-cause events and censoring also occur. A cause-specific Cox model is fitted with exposure as its only predictor.
First check the information. There are only 6 target-cause events, and the exposed group has zero. If exposed subjects are in the risk sets at times when unexposed target events occur, their zero target-event count can create separation in the direction of a negative exposure coefficient. The exact risk-set pattern matters, so the event table is a warning that calls for checking the risk sets, not a proof based on the table alone.
Imagine the software returns \(\hat\beta=9.6\) with \(SE=5.1\) for a different coding or parameterization in which the estimated direction is positive. A nominal Wald interval on the log-coefficient scale is
Exponentiating gives a hazard-ratio estimate of \(\exp(9.6)\approx 1.48\times10^4\), with a nominal interval
The calculation is arithmetically correct, but the interval is so wide that it signals very weak information, and a Wald approximation may be unreliable under separation. It should not be presented as a precise estimate. Check the cause coding and risk sets; consider reducing unsupported model complexity or using a suitable penalized method. Any simplification must be scientifically defensible, and causes should not be pooled merely to obtain a smaller standard error.
Possible responses include collecting more data, limiting the number of predictors based on subject-matter priorities, or using a suitable penalized likelihood method. Penalization can stabilize estimation, but it changes the estimation procedure and does not create information absent from the data. Report the method and its implications. Combining event causes is only reasonable if the combined outcome is scientifically meaningful and defined in advance. Changing from cause-specific to subdistribution regression changes the model target; it is not a remedy for sparse data.
Extreme Covariates and Numerical Scaling
A predictor measured on a very large numerical scale, containing extreme values, or having little variation can cause numerical instability. Scaling a continuous predictor changes the numerical value and unit of its coefficient, not the underlying fitted relative hazards, provided the transformation is linear and the model otherwise stays the same. Centering can also make calculations easier to interpret and may improve numerical conditioning when values are large.
For a Cox model with a single predictor \(X\), define \(Z=(X-c)/a\), where \(a>0\) is a chosen scale. Since \(X=c+aZ\),
The constant factor \(\exp(\beta c)\) is absorbed into the unspecified baseline hazard, so the coefficient for \(Z\) is \(a\beta\). The relative hazard for a fixed change in the original \(X\) is unchanged. With an explicit intercept or other model structure, confirm the corresponding parameterization rather than assuming the same invariance.
Worked Example: Rescaling a Large Predictor
In a hypothetical cause-specific Cox analysis, a biomarker is recorded in units that range from about 9,500 to 10,200. Suppose the fitted coefficient per raw unit is \(\hat\beta=0.002\). The raw numbers are large relative to their spread, making the coefficient difficult to read and potentially inconvenient for computation.
Set \(Z=(X-9{,}800)/100\), so one unit of \(Z\) is a 100-unit increase in the original biomarker. Here \(X=9{,}800+100Z\), and the rescaled coefficient is
The estimated relative hazard for a 100-unit increase is \(\exp(0.2)\approx1.221\), or about 22.1% higher, under this fitted model. The corresponding raw-scale calculation is \(\exp(0.002\times100)=\exp(0.2)\), the same result. Rescaling does not make an implausible relationship plausible or remove influential observations; inspect data-entry ranges and fit diagnostics as well.
Incorrect Event Coding: A Data Problem, Not an Optimizer Problem
Competing-risk software differs in its input conventions. Some functions expect a single status variable with 0 for censoring and distinct positive codes for each cause. Others require a separately constructed event indicator for a particular cause, or accept a factor with specified levels. Never infer the correct coding from a function name alone: check its documentation and verify the analysis data that are actually passed to it.
For a cause-specific fit for cause \(k\), the target event is coded as an event, while a competing event occurs at its observed time and removes that person from the cause-\(k\) risk set. It is not ordinary loss to follow-up before the competing event. For a Fine–Gray fit, competing-event subjects are handled by the subdistribution risk-set construction, so recoding them as ordinary independent censoring can produce the wrong analysis. Censoring, a competing event, and the event of interest must remain distinguishable in the source data.
Worked Example: Auditing Cause and Censoring Codes
Consider these hypothetical records, where the intended source coding is 0 for censored, 1 for the target event, and 2 for a competing event.
| Subject | Follow-up time (months) | Source status |
|---|---|---|
| A | 4 | 1 |
| B | 6 | 2 |
| C | 8 | 0 |
| D | 3 | 1 |
For a cause-specific analysis of cause 1, construct an event indicator that is 1 only for source status 1 and 0 otherwise. The records become A: event 1 at 4 months; B: event 0 at 6 months; C: event 0 at 8 months; and D: event 1 at 3 months. Subject B contributes to the cause-1 risk set up to month 6, when the competing event occurs. Subject C is censored at month 8. The cause-1 event count is therefore 2, not 3.
If source status 2 were accidentally treated as the target event, the analysis would instead count B and misidentify the outcome. If status 2 were dropped from the data, the follow-up time and risk sets would also be wrong. For a Fine–Gray model of cause 1, retain the distinction between B's competing event and C's censoring, and supply the status in the form required by the chosen implementation. Check counts and example rows before fitting.
Interpreting Diagnostics and Choosing a Remedy
Useful warning signs include a nonconvergence message, coefficients that change greatly with small optimizer adjustments, enormous standard errors, a nearly singular information matrix, fitted probabilities or hazards at numerical boundaries, and estimates that change sharply when one observation is removed. Examine cause-specific event counts across categorical levels, ranges and distributions of continuous predictors, correlations among predictors, and convergence traces when available. Review risk sets at event times if separation is suspected.
A disciplined response distinguishes a data correction from a model change and a numerical adjustment:
- Correct coding or data errors first. Fix verified transcription, time-unit, reference-level, or status-code errors, then rerun the intended model.
- Address unsupported complexity. Reduce the number of parameters only with a defensible plan, such as removing redundant terms or avoiding an unjustified interaction. Do not select a smaller model solely because it gives a preferred result.
- Improve numerical representation. Center or rescale continuous variables, and check units and extreme values. This should preserve the modeled contrast when transformed correctly.
- Consider penalization when appropriate. State the penalty and recognize that penalized estimates are not identical to ordinary maximum-likelihood estimates. For separated sparse data, ordinary Wald inference can be especially misleading.
- Do not treat more iterations as more information. Adjusting tolerances or iteration limits can help diagnose an optimizer that stopped prematurely, but cannot produce a finite, well-supported estimate when the likelihood has no finite maximum.
The main alternatives have different trade-offs. An ordinary unpenalized fit is straightforward when the model is identifiable and the data are adequately informative. A penalized fit can stabilize estimates but depends on the penalty and may introduce shrinkage. Simplifying the model can improve support but risks omitting important structure. Collecting more events directly improves information but may not be feasible. Changing the event definition or regression estimand changes the question and must be justified scientifically, not selected as a software workaround.
Common Mistakes
- Equating convergence with validity. A routine can converge to a boundary estimate or a result based on incorrect coding. Check data, assumptions, and stability.
- Ignoring a rare cause because the total sample is large. Precision for a cause-specific parameter depends strongly on the number and distribution of events of that cause, not just the cohort size.
- Treating competing events as ordinary censoring in every method. Risk-set handling differs between cause-specific and subdistribution regression; follow the model definition and software requirements.
- Using a huge Wald interval as reassuring evidence. A formula can return an interval even when its approximation is poor. Sparse events and separation require explicit caution.
- Pooling causes or switching estimands to make the fit behave. These choices alter the outcome or target. They require scientific justification and clear definition.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions below.
- A cause-specific model has 5 target events, all in one level of a binary predictor. What should you inspect before interpreting a very large coefficient?
- A biomarker is converted from units to tens of units. If its original coefficient is 0.03 per unit, what is the coefficient per ten-unit increase, and what relative hazard does that represent?
- In a dataset coded 0 for censoring, 1 for cause A, and 2 for cause B, how should status 2 be handled in a cause-specific Cox fit for cause A?
- Why is increasing the maximum number of optimizer iterations not a complete remedy for separation?
Solutions. (1) Check cause counts by predictor level and inspect the target-event risk sets. If subjects in the empty-event level remain at risk at target events, separation may cause the likelihood to favor an unbounded coefficient; also verify coding and model complexity. (2) The coefficient per ten units is \(10(0.03)=0.30\). The relative hazard for a ten-unit increase is \(\exp(0.30)\approx1.350\), under the model. (3) Treat status 2 as a competing event at its observed time: it is not a cause-A event, and the subject leaves the cause-A risk set at that time. (4) More iterations can help an optimizer that stopped too early, but separation can mean that no finite maximizing coefficient exists. Iterations cannot supply the missing information; consider data support, a defensible model change, or suitable penalization.