Interpreting Conditional PWP Effects

What a Conditional PWP Effect Describes

A PWP model asks about the hazard of a particular recurrence order among subjects who have reached that order. For example, the hazard of a second recurrence is evaluated among subjects who have already experienced a first recurrence and are still under observation and at risk for the second. The comparison is therefore conditional on eligibility for that event order; it is not automatically a comparison of overall recurrence burden in the original cohort.

This conditioning is the central interpretive point. A PWP hazard ratio compares the instantaneous rates of the next specified event for covariate profiles within the relevant event-order risk set, at the same value of the model's time scale and conditional on the modeled covariates. It does not directly compare the probability of ever having that event, the mean number of recurrences, or the time until any recurrence.

Definition: In a PWP model, the event-order-specific hazard is the instantaneous rate of the \(k\)th recurrence among subjects who have experienced \(k-1\) recurrences and remain under observation and at risk for the \(k\)th. A conditional PWP hazard ratio compares that hazard between covariate profiles within this eligible group, conditional on the modeled history and time scale.

Let \(k\) denote event order, \(X_i\) a subject's covariate vector, and \(u\) the chosen analysis time. A proportional-hazards specification for event order \(k\) can be written as

\lambda_{ik}(u \mid X_i, \text{eligible for event } k) = \lambda_{0k}(u)\exp(\beta_k^\mathsf{T}X_i).

Here \(\lambda_{0k}(u)\) is the baseline hazard for event order \(k\), and \(\beta_k\) is the coefficient vector for that order. The clock \(u\) is study time in a total-time formulation or time since the preceding recurrence in a gap-time formulation. Thus, the comparison must be described on the clock actually used in the model. The risk set is also event-order specific: subjects who have not yet experienced the required \(k-1\) recurrences are not eligible for the \(k\)th-event comparison.

From a Coefficient to a Conditional Hazard Ratio

For two covariate profiles \(x_a\) and \(x_b\) in the same event-order risk set, the model's conditional hazard ratio is

\frac{\lambda_k(u \mid x_a,\text{ eligible})} {\lambda_k(u \mid x_b,\text{ eligible})} = \frac{\lambda_{0k}(u)\exp(\beta_k^\mathsf{T}x_a)} {\lambda_{0k}(u)\exp(\beta_k^\mathsf{T}x_b)} = \exp\!\left\{\beta_k^\mathsf{T}(x_a-x_b)\right\}.

The baseline hazard cancels, so under the proportional-hazards model the ratio does not depend on \(u\). For a binary covariate coded 1 versus 0, the hazard ratio is \(\exp(\beta_k)\). For a continuous covariate, \(\exp(\beta_k)\) is the ratio for a one-unit increase, holding the other modeled covariates fixed; a change of \(d\) units has ratio \(\exp(d\beta_k)\).

The proportional-hazards assumption is conditional on event order and covariates: the covariate hazard ratio does not vary over the model's time scale. It does not require the baseline hazards for different event orders to be identical. Nor does a conditional hazard ratio mean that the same people remain in the comparison over time; event occurrence and censoring change who is eligible and at risk.

Worked Example: Treatment and the Second Recurrence

Suppose a hypothetical PWP total-time model for the second recurrence estimates a treatment coefficient of \(\hat{\beta}_2=-0.357\), where treatment is coded 1 and control 0. Assume the model's event-order risk set contains subjects who have experienced a first recurrence and remain under observation and at risk for a second. The model uses study time as its clock and adjusts for age.

The treatment-versus-control conditional hazard ratio is

\widehat{HR}_2=\exp(\hat{\beta}_2)=\exp(-0.357)=0.700.

At the same study time and age, among subjects eligible for a second recurrence, the fitted instantaneous hazard of that recurrence is 0.700 times as high for treated subjects as for control subjects, under the model. Equivalently, the fitted conditional hazard is 30.0% lower, since \(1-0.700=0.300\). This is not a claim that treatment lowers the probability of any recurrence by 30%, or that it causes a 30% reduction in every subject's recurrence risk.

The interpretation relies on the stated risk set and the proportional-hazards model. If the analysis instead used gap time, the comparison would be at the same elapsed time since the first recurrence, not the same study time. The event-order conditioning is still essential in either case.

Uncertainty and Comparisons Across Covariate Values

A coefficient estimate is uncertain, so interpretation should usually include its confidence interval. A coefficient-scale interval is transformed by exponentiating both endpoints; because the exponential function is increasing, the endpoint order is preserved. A confidence interval for a hazard ratio that includes 1 is compatible, at that confidence level, with no conditional hazard difference as well as with the other values it contains. It is not a probability statement about the fixed parameter.

Worked Example: Confidence Interval for a Conditional Effect

In a hypothetical PWP model for the third recurrence, a binary exposure has estimate \(\hat{\beta}_3=0.262\) and standard error \(SE=0.120\). Assume the fitted model is appropriate for the third-event risk set and use a large-sample 95% Wald interval.

On the coefficient scale, the interval is

0.262 \pm 1.96(0.120) =0.262\pm0.2352 =(0.0268,\ 0.4972).

Exponentiating gives the hazard-ratio estimate and interval:

\widehat{HR}_3=\exp(0.262)=1.300,\qquad \left(\exp(0.0268),\exp(0.4972)\right)=(1.027,\ 1.644).

Among subjects eligible for a third recurrence, the estimated conditional hazard is 1.30 times as high for the exposed covariate group as for the unexposed group, at the same modeled time and values of adjusted covariates. The 95% Wald confidence interval is 1.027 to 1.644. Under repeated sampling and the model's assumptions, intervals constructed this way would cover the parameter 95% of the time. This interval excludes 1, indicating a difference at the corresponding two-sided 5% Wald-test level; it does not establish that exposure causes the difference.

For continuous covariates, stating the unit is crucial. If one unit is too small to be clinically meaningful, report a rescaled contrast and calculate its interval on the coefficient scale before exponentiating.

Worked Example: Rescaling a Continuous Covariate

Suppose a hypothetical PWP model for the second recurrence estimates \(\hat{\beta}_2=0.080\) per one-unit increase in a biomarker, with \(SE=0.025\) per unit. Assume the proportional-hazards model is appropriate and the intended contrast is a 10-unit increase, with other covariates held fixed.

The coefficient contrast for 10 units is \(10(0.080)=0.800\), and its standard error is \(10(0.025)=0.250\). Thus,

HR_{10}=\exp(0.800)=2.226, \qquad 0.800\pm1.96(0.250)=(0.310,\ 1.290).

Exponentiating the interval endpoints gives \((\exp(0.310),\exp(1.290))=(1.363,\ 3.634)\). Among subjects eligible for a second recurrence, the fitted hazard for a 10-unit higher biomarker value is 2.23 times the hazard for a value 10 units lower, at the same model time and other covariate values. The confidence interval corresponds to ratios from 1.36 to 3.63. The contrast is conditional and model-based; it is not a comparison of cumulative recurrence probabilities.

Assumptions, Diagnostics, and Limits

A numerical hazard ratio is meaningful only with a defensible model and risk-set construction. The interpretation above assumes that event order and eligibility are recorded correctly, that the chosen clock matches the scientific comparison, and that the proportional-hazards specification is adequate over the times being analyzed. Covariate effects must be correctly represented, including relevant changes over time if covariates vary. Censoring and loss to follow-up must be non-informative for the modeled event process conditional on the variables included, or handled with an appropriate method. There must also be enough information in the event-order risk set to estimate the effect; a coefficient based on very few late-order events can be unstable.

Useful checks include comparing observed event counts and follow-up with the risk-set construction, inspecting whether the estimated covariate effect appears to change over the analysis time, and examining uncertainty and influential observations. A sharply widening interval at later event orders is a warning that precision is limited. A time-varying pattern in residual-based or graphical checks may challenge proportional hazards. These checks can reveal problems, but they cannot prove that the model or censoring assumptions are correct.

Important limitation: Subjects in a later PWP risk set are a selected group: they have already experienced the required number of events and remained observable. Even in a randomized study, conditioning on recurrence history can change the composition of the groups being compared. A conditional PWP hazard ratio is therefore not generally a marginal treatment effect or a causal effect for the original cohort without additional assumptions and a clearly defined causal estimand.

The conditional interpretation differs from the main alternatives in recurrent-event analysis. The Andersen–Gill model uses a broader recurrent-event risk-set formulation and targets a common event-intensity association rather than restricting each comparison to those eligible for a specified event order. A mean cumulative function summarizes the expected accumulated number of events over time, while a marginal event-rate ratio compares average rates in a target population. These quantities answer different questions. A PWP hazard ratio is especially relevant when the event order and eligibility history are scientifically central, but it does not replace a population-level summary when the goal is total event burden.

Common Mistakes

Calling the ratio a recurrence probability. A hazard ratio compares instantaneous event rates among eligible subjects, not probabilities over a follow-up period.
Omitting event order. “Treatment reduces recurrence hazard” is incomplete if the estimate concerns only, for example, the second recurrence among people who have already had a first.
Ignoring the clock. A total-time comparison is made at the same study time; a gap-time comparison is made at the same time since the preceding event.
Interpreting a conditional association as an overall causal effect. The event-order risk set is selected by prior event history and continued observation, so causal interpretation needs more than a fitted PWP model.
Comparing effect estimates across event orders as if they shared a risk set. The eligible populations differ, and estimates may also differ in precision. A difference between estimates alone does not establish a true change in effect.

Key takeaway: State the event order, the eligible population, the modeled clock, the covariate contrast, and the conditional nature of the comparison whenever reporting a PWP hazard ratio.

Check Your Understanding

Answer each question, then compare your reasoning with the solutions below.

A PWP model for the fourth recurrence estimates \(\hat{\beta}_4=-0.20\) for a binary treatment. What is the fitted hazard ratio, and to whom and what comparison does it apply?
A PWP gap-time model estimates a hazard ratio of 1.5 for a covariate. What does “at the same time” mean in this interpretation?
Why is a PWP hazard ratio of 0.8 not a statement that the probability of recurrence is 20% lower?
A coefficient estimate is 0.10 per one-unit increase in a continuous covariate. What is the fitted hazard ratio for a five-unit increase?

Solutions. (1) \(\exp(-0.20)=0.819\). Among subjects who have experienced three recurrences and remain under observation and at risk for a fourth, the fitted treatment-group hazard is 0.819 times the control-group hazard at the same modeled time and adjusted covariate values. (2) It means the same elapsed time since the preceding recurrence, because the model uses a gap-time clock. (3) A hazard is an instantaneous rate among those currently eligible and at risk; a probability over follow-up depends on the hazard over time and on continued eligibility and observation. (4) The five-unit contrast has coefficient \(5(0.10)=0.50\), so its fitted hazard ratio is \(\exp(0.50)=1.649\), conditional on the specified event-order risk set and other modeled covariates.