Separation and Sparse Events in Weighted Pooled Logistic Models

Why Events Can Be Difficult to Model

A pooled logistic survival model represents each at-risk person-period as a binary outcome: an event during the interval or no event during it. In a weighted fit, each row also contributes according to its treatment weight. If an interval contains very few events, or if its event pattern is perfectly aligned with a predictor such as treatment, the fitted coefficient can become highly unstable—or a finite maximum-likelihood estimate may not exist at all.

These are related but distinct problems. Separation is a property of the observed outcomes and model predictors that can make the likelihood increase as a coefficient moves without bound. Sparsity means that event information is limited; it often produces weak curvature and imprecise estimates, but does not by itself imply separation. The distinction matters because the diagnostic evidence and possible responses differ.

Weighted Pooled Logistic Likelihood and Separation

Let \(i\) index individuals and \(t\) index intervals in which an individual is at risk. Write \(Y_{it}=1\) if the event occurs in interval \(t\), and \(Y_{it}=0\) otherwise. Let \(X_{it}\) be the model’s predictor vector, including terms for treatment and the baseline hazard, and let \(\beta\) be its coefficient vector. For rows with positive weight \(w_{it}\), the fitted hazard is \(p_{it}(\beta)=\operatorname{expit}(X_{it}^{\mathsf T}\beta)\), where \(\operatorname{expit}(u)=1/(1+\exp(-u))\).

The condition says that moving in direction \(d\) does not worsen the signed predictor for any row: it raises predictors for events or lowers them for non-events, with improvement for at least one row. The weights need only be positive for the rows under consideration; their sizes affect the amount of likelihood change, not the sign condition. In a pooled survival model, repeated rows from one individual may be dependent. That dependence matters for inference, but it does not change this criterion for the weighted objective being optimized.

Proof. Fix any finite \(\beta\), and consider \(\ell_w(\beta+sd)\) for \(s\geq 0\). For one row, write \(\eta=X_{it}^{\mathsf T}(\beta+sd)\). The derivative of its weighted contribution with respect to \(s\) is \(w_{it}(Y_{it}-p_{it}(\beta+sd))X_{it}^{\mathsf T}d\). If \(Y_{it}=1\), separation gives \(X_{it}^{\mathsf T}d\geq 0\), so this derivative is nonnegative because \(1-p_{it}>0\) at finite \(\eta\). If \(Y_{it}=0\), separation gives \(X_{it}^{\mathsf T}d\leq 0\), so the derivative is again nonnegative because \(-p_{it}<0\). For a row with strict inequality, the derivative is strictly positive at every finite \(s\). Its weight is positive, so summing over rows shows that the total objective increases strictly at every finite \(s\).

Each row’s contribution has a finite limit as \(s\) tends to infinity: on strict rows the fitted probability tends to the observed outcome, and on equality rows the linear predictor is unchanged. There are finitely many rows, so the sum also has a finite limit. Thus the likelihood has a finite supremum approached along the direction, but no finite point can attain it, since moving farther in the direction increases the objective. \(\square\)

Worked Examples: Complete Separation and Sparse Events

How Weight Concentration Changes the Available Information

A low raw event count does not tell the whole story in a weighted fit. The weighted objective’s curvature depends on both the fitted probabilities and the weights. Its Hessian is \(\nabla^2\ell_w(\beta)=-\sum_{i,t}w_{it}p_{it}(1-p_{it})X_{it}X_{it}^{\mathsf T}\). Each row contributes curvature in proportion to \(w_{it}p_{it}(1-p_{it})\). Predicted probabilities near zero or one contribute little curvature; a small number of events can also leave some predictor contrasts weakly informed.

For any vector \(v\), the negative Hessian satisfies \(v^{\mathsf T}[-\nabla^2\ell_w(\beta)]v=\sum_{i,t}w_{it}p_{it}(1-p_{it})(X_{it}^{\mathsf T}v)^2\geq 0\). So the objective is concave. At finite \(\beta\), if the positive-weight design matrix has full column rank, this expression is positive for every nonzero \(v\), and the objective is strictly concave. Strict concavity means that there cannot be two distinct finite maximizers; it does not guarantee that any finite maximizer exists. The separating example has no finite maximizer despite concavity.

Diagnostics, Alternatives, and Limits

A sound assessment separates outcome support from optimization symptoms. For each interval and important treatment or predictor pattern, tabulate the number of at-risk rows and events. Add the sum of weights for event and non-event rows; raw counts alone can hide a weighted pattern close to separation. Then inspect the fitted probabilities, coefficient estimates, and likelihood as the optimizer iterates. A coefficient that grows in magnitude while its fitted hazards approach zero or one and the likelihood changes very little is consistent with separation or near-separation. A nearly flat likelihood in a coefficient direction, or a nearly singular information matrix, can signal weakly identified contrasts even when no exact separating direction exists.

No single diagnostic proves the cause of instability. A large coefficient can also reflect a coding error, an omitted time pattern, or a model with poorly supported predictor combinations. Check the risk-set construction, outcome coding, treatment and interval indicators, and positive-weight rows. Examine weight concentration by interval and outcome, rather than relying only on the overall weight distribution. Numerical convergence messages are not evidence that a well-supported finite estimate exists.

The main alternatives address different problems. A simpler predictor structure may avoid unsupported contrasts, but it changes the model and can impose substantive restrictions. Combining intervals may increase event counts only when the resulting time grouping is scientifically defensible; it changes the time resolution. Penalized or bias-reduced logistic fitting can produce finite estimates under separation, but it changes the estimation procedure and generally the target of the fitted coefficients. These are not automatic repairs for extreme or inappropriate weights. The later material on resolving numerical instability treats computational responses in more detail; here the essential point is first to identify whether the difficulty is separation, sparse support, or weight concentration.

Sparse events also limit what a fitted hazard can credibly say. Even when a numerical estimate exists, a few events may leave treatment contrasts highly uncertain, and fitted probabilities in thinly supported periods can be sensitive to small changes in the data or model. Conversely, an extreme coefficient alone does not establish a causal effect or a failure of causal identification. This tutorial concerns the behavior of the weighted pooled logistic fit, not the causal assumptions that justify its weights or interpretation.

Common Mistakes

• Calling every sparse table separated. Few events can yield a finite estimate, as the second example shows. Separation requires a direction satisfying the signed-predictor inequalities.
• Changing the intercept to construct a separating direction. If untreated rows include both outcomes, their shared predictor cannot be moved in a direction that improves both. In the first example, hold the intercept fixed and decrease the treatment coefficient.
• Treating a converged optimizer as proof of a finite maximum. A program may stop at a large finite coefficient while the likelihood continues to increase toward a supremum. Check coefficient paths and fitted hazards.
• Using raw event counts as the only measure of support. Weights can make the weighted event pattern very different from the raw pattern. Inspect both weighted totals and their distribution across rows.
• Interpreting weighted curvature as an ordinary sample size. The weighted Hessian describes the curvature of the objective; dependence and weight estimation also matter for inference.