When Binary PQL Estimates Are Most Trustworthy

What Makes Binary PQL More Trustworthy?

Penalized quasi-likelihood (PQL) replaces a nonlinear binary-response problem with a sequence of approximate, weighted linear mixed-model fits. That can make estimation practical, but the approximation is not exact. Its performance is generally more defensible when fitted probabilities are not concentrated near zero or one, and when clusters contain enough informative observations to help estimate their random effects. These are favorable conditions, not a certificate of negligible bias.

The key judgment is therefore comparative: does the model present features that keep the approximation local and supply information within clusters, or does it combine extreme probabilities, weak cluster information, and substantial random-effect variation? The answer helps assess trustworthiness, but it cannot determine the actual bias from a fitted model alone.

This model statement does not by itself ensure that PQL is reliable. Its assessment also depends on whether the fixed-effect and random-effect structure is suitable, whether the fitting procedure converges stably, and how much information the binary data provide. The focus here is on conditions that favor the approximation, not on proving that the model is correctly specified.

Why Fitted Probabilities Matter

At an iteration of PQL, let \(\eta_{ij}\) be the current linear predictor and \(p_{ij}=\operatorname{logit}^{-1}(\eta_{ij})\) its fitted conditional probability. A binary PQL fit constructs a working response and a working weight,

These quantities come from a local approximation to the mean and variance relationship. When \(p_{ij}\) is near 0 or 1, \(w_{ij}\) is small and the working response divides by a small number. The corresponding likelihood is highly nonlinear in the predictor, so a local linear approximation has less room to represent the shape accurately. Probabilities in a moderate range avoid the most extreme behavior and generally make PQL more plausible.

The weight illustrates the point directly: it is largest at \(p_{ij}=0.5\), where \(w_{ij}=0.25\), and declines symmetrically toward either endpoint. This is a useful indicator of conditional information about a linear predictor; it is not a standalone measure of PQL bias or a guarantee that the fit is good.

Within-Cluster Information Is More Than a Large Total Sample

Random effects represent cluster-to-cluster variation. Estimating them requires observations within clusters, not merely a large total number of observations pooled across clusters. For a random-intercept model, one rough summary of conditional information about a cluster's linear predictor is

This sum is useful for describing the contribution of a cluster under the fitted model. It is not a complete measure of information about \(\tau^2\), and no universal cutoff makes a cluster “large enough.” Its value depends on fitted probabilities: a cluster can have many observations but little information if their probabilities are mostly extreme. Observed outcomes matter as well. A cluster with only successes or only failures may provide limited evidence about within-cluster variation, especially when that pattern is common across the data.

The number of clusters and the number of observations per cluster answer different questions. More clusters provide more independent realizations of the random effects; more observations per cluster can provide more information about each cluster. Neither dimension substitutes automatically for the other. A large overall sample can still be unhelpful for PQL if most clusters contain few informative binary outcomes.

A Local Approximation Explains Both the Promise and the Limit

The role of random-effect variation can be seen by expanding the logistic mean \(f(\eta)=\operatorname{logit}^{-1}(\eta)\) around a predictor value \(\eta\). Suppose a random perturbation \(\delta\) has mean zero and variance \(v\). For a small, concentrated perturbation, a second-order Taylor expansion gives

The second expression follows by taking expectations: \(E(\delta)=0\) removes the first-order term, while \(E(\delta^2)=v\). For the logistic function, \(f'(\eta)=p(1-p)\) and \(f''(\eta)=p(1-p)(1-2p)\). The curvature term is therefore affected by both the location on the probability scale and the amount of random-effect variation.

If probabilities stay near the middle and the random-effect spread is modest, the local expansion is more plausible. If the spread is large, a substantial range of random effects can move otherwise moderate probabilities toward zero or one. The Taylor expression is a local justification, not an exact formula for PQL bias: when perturbations are large, omitted higher-order terms may matter, and the expansion itself may be inadequate.

Conditions That Favor PQL—and How to Check Them

A more favorable assessment is reasonable when several features occur together. The fitted probabilities should not be dominated by values close to zero or one; clusters should contain enough observations with useful fitted weights and a mix of outcomes; and estimated random-effect variation should not imply that probabilities range widely across clusters. A well-behaved fit should also converge without warnings that suggest numerical instability.

These features should be checked across the model, not inferred from a single average. An average fitted probability of 0.50 could conceal many fitted probabilities near 0.01 and 0.99. Likewise, the overall event count can conceal clusters with almost no events or almost no non-events. Examine the distribution of fitted probabilities, the event and non-event patterns by cluster, and cluster sizes together.

Warning signs include fitted probabilities concentrated near the endpoints, many clusters with only one outcome, very small working weights, large estimated random-effect variation, or unstable estimates when the fitting algorithm is rerun or modestly changed. A failure to converge is not just a cosmetic issue: without a stable solution, estimates cannot be treated as reliable. Convergence, however, is not evidence that the approximation is accurate.

How PQL Compares With the Main Alternative

Likelihood-based integration, discussed in the earlier material on PQL and likelihood-based integration for binary models, is the principal comparison for assessing the approximation. It accounts for the random effects through integration rather than relying on PQL's sequence of local linear fits. Agreement in fixed-effect estimates and variance estimates is reassuring in the limited sense that two approaches give similar answers for this model and data. Disagreement is a reason to investigate, not proof that one method is automatically correct; numerical accuracy, model assumptions, and data information still matter.

PQL may be computationally convenient, particularly as a fitting strategy for complex mixed models. Likelihood-based integration can require more computation, but it avoids the same PQL linearization. Neither method repairs a misspecified random-effects structure or creates information absent from sparse clusters. If both methods are unstable, the issue may lie in the data or model rather than in the choice between them.

Why Favorable Conditions Are Not a Bias Guarantee

Moderate fitted probabilities and ample within-cluster information make the approximation more plausible; they do not establish negligible bias. The size of approximation bias depends on the full data-generating and fitted-model setting, including the random-effect distribution and variance, covariate patterns, cluster-size distribution, and the model's random-effects structure. Finite-sample variability is also distinct from approximation bias: a noisy estimate is not necessarily biased, and a precise-looking estimate can still be affected by approximation.

The practical conclusion should therefore be calibrated. Describe favorable conditions as evidence that supports using PQL cautiously, and unfavorable conditions as reasons for greater concern. Do not turn informal checks into a numerical rule or treat successful convergence as validation. When a substantive conclusion depends on the estimates, compare against likelihood-based integration where feasible and consider whether the data actually support the model being fitted.