LWYY Marginal Rate Model

Why Model a Marginal Recurrent-Event Rate?

A recurrent-event analysis may ask how treatment changes the average rate of repeated hospitalizations, infections, or device failures across a population. That question differs from asking how treatment changes the hazard of the next event for a person with a particular event history. The Lin–Wei–Yang–Ying (LWYY) marginal rate model directly targets the former: a covariate’s effect is a proportional change in the population-average event rate over follow-up.

The distinction matters because people can have different numbers and timings of previous events. The LWYY model does not make the covariate effect conditional on that past event history. Its rate is marginal with respect to the recurrent-event history, while still being modeled as a function of specified covariates. Marginal here does not mean that the model ignores covariates or averages over every population’s covariate distribution.

Definition: Let $N_i^*(t)$ be the number of recurrent events experienced by subject $i$ by time $t$ in the absence of loss to observation, and let $Z_i$ be a vector of covariates. The LWYY marginal rate model specifies the mean event increment, given $Z_i$, as $$ E\{dN_i^*(t)\mid Z_i\}=\exp(\beta^\top Z_i)\,d\Lambda_0(t). $$ Here $\Lambda_0(t)$ is an unspecified baseline cumulative rate function, and $\beta$ is a vector of regression coefficients. Equivalently, where a rate exists, the marginal rate is $\lambda_0(t)\exp(\beta^\top Z_i)$. The model does not condition this rate on the subject’s previous recurrent events.

The baseline rate is allowed to vary freely over time; the proportionality assumption is that covariate effects multiply that baseline by a constant factor over follow-up. For a binary covariate $Z$, the rate ratio comparing $Z=1$ with $Z=0$ is $\exp(\beta)$. The coefficient $\beta$ is therefore a log rate ratio. For a continuous covariate, $\exp(\beta)$ is the rate ratio for a one-unit increase, holding the other modeled covariates fixed.

From Rates to Expected Cumulative Counts

Integrating the mean increment gives a useful interpretation of the model. For fixed $Z_i$, and over a common follow-up interval from 0 to $t$,

E\{N_i^*(t)\mid Z_i\} =\int_0^t E\{dN_i^*(u)\mid Z_i\} =\exp(\beta^\top Z_i)\Lambda_0(t),

assuming $N_i^*(0)=0$ and $\Lambda_0(0)=0$. Thus, when two covariate groups are followed over the same interval and satisfy the model, their expected cumulative event counts have the same ratio as their rates. This is not a statement that each person’s realized count is multiplied by that ratio. It compares conditional population means for covariate profiles.

Worked Example: Interpreting a Binary Covariate

Suppose a hypothetical study models recurrent exacerbations using $Z=1$ for a treatment group and $Z=0$ for a comparison group. Its fitted coefficient is $\hat{\beta}=\log(1.50)=0.4055$. At a particular time, the model’s baseline rate is 0.20 events per person-year.

The treatment-to-comparison rate ratio is $\exp(\hat{\beta})=\exp(0.4055)=1.50$. At that time, the modeled marginal rate is 0.20 events per person-year in the comparison group and $0.20(1.50)=0.30$ events per person-year in the treatment group.

If the baseline cumulative rate over one year is 0.40 events per person, the model gives expected cumulative counts of $0.40$ for $Z=0$ and $0.40(1.50)=0.60$ for $Z=1$. The treatment group’s expected count is 1.50 times the comparison group’s expected count over that interval. These are hypothetical model-based averages, not predictions that every treated subject will have exactly 0.60 events, and the association does not by itself establish a causal treatment effect.

The baseline need not be constant. For example, it can represent a higher event rate early after enrollment and a lower rate later. The model requires a covariate’s rate ratio to remain the same across those periods. This makes the proportional-rate assumption distinct from assuming a constant rate over time.

Worked Example: A Changing Baseline Rate With a Constant Rate Ratio

Consider a hypothetical two-period follow-up. The comparison group’s baseline rate is 0.10 events per person-month in months 0–6 and 0.20 events per person-month in months 6–12. A binary covariate has rate ratio 1.40 throughout follow-up.

For the comparison group, the expected event count over the first six months is $6(0.10)=0.60$, and over the second six months it is $6(0.20)=1.20$. The total expected count is $0.60+1.20=1.80$ events per person over 12 months.

For the covariate group, the corresponding rates are $0.10(1.40)=0.14$ and $0.20(1.40)=0.28$ events per person-month. Its expected counts are $6(0.14)=0.84$ and $6(0.28)=1.68$, totaling $2.52$. The ratio is $2.52/1.80=1.40$. The baseline rate doubled between periods, but the covariate rate ratio remained 1.40, as the model requires.

Assumptions and Scope of the Model

The model is a specification of the mean rate, not a complete probability distribution for event histories. Its interpretation relies on a clear definition of the event, the covariates, and the time scale. Key conditions include:

Correct mean-rate form. Conditional on the included covariates, the marginal rate is proportional to the unspecified baseline rate. The coefficients are constant over the modeled follow-up unless the model is expanded to allow otherwise.
Appropriate observation and censoring. Subjects contribute recurrent-event information while under observation. To infer the intended full-follow-up rate from censored records, censoring must be independent of future event increments, at least conditional on the modeled covariates and any required observed history. If dropout is related to an unobserved high event rate, simply treating it as noninformative can distort the target.
Well-defined follow-up and event counting. The analysis must specify when observation begins and ends and what qualifies as a recurrence. Repeated records must represent the intended events rather than duplicate recording.
Suitable covariate specification. The comparison is conditional on the covariates included in the model. Omitted covariates or misspecified functional forms can make the reported rate ratio an inadequate description of the data.

Repeated events from the same subject are generally dependent. The LWYY model’s marginal mean specification does not require those events to be independent. Dependence still matters for uncertainty calculations, so inference must account for clustering by subject. The details of estimating equations and robust variance calculations are separate methodological steps; a model’s mean-rate interpretation does not, by itself, guarantee valid standard errors.

Worked Example: A Continuous Covariate and Its Units

In a hypothetical cohort, let $Z$ be baseline age measured in decades, and suppose the fitted LWYY coefficient for age is $\hat{\beta}=0.18$. Interpret the coefficient for a 10-year difference and for a 5-year difference.

For a 10-year increase, $Z$ increases by one unit, so the modeled rate ratio is $\exp(0.18)=1.197$. Thus, two otherwise comparable covariate profiles differing by 10 years of age have modeled recurrent-event rates in a ratio of about 1.20, or about 19.7% higher for the older profile.

A 5-year increase is half a unit on this scale. Its rate ratio is $\exp(0.18\times0.5)=\exp(0.09)=1.094$, about 9.42% higher. This interpretation assumes the log rate is linear in age over the relevant range. The unit used to code the covariate is essential: if age were recorded in years instead, the coefficient would be one tenth as large for the same fitted relationship.

How the Marginal Target Differs From Other Recurrent-Event Models

The choice among recurrent-event models should follow the scientific estimand. The earlier material on choosing a recurrent-event model by estimand distinguishes population-average questions from questions about event order or a subject’s conditional event history. In broad terms:

Approach	Primary interpretation	How it differs from LWYY
Andersen–Gill (AG)	Recurrent-event intensity conditional on the specified observed history and covariates	LWYY targets a marginal rate not conditioned on prior recurrences.
PWP	Event-specific risk among subjects who have reached the relevant event order	LWYY does not define separate event-order risk sets or conditional event-order effects.
WLW	Marginal event-order-specific outcomes, often represented through ordered event times	LWYY models the recurrent-event mean rate over follow-up rather than a collection of event-order-specific outcomes.

These are not merely different ways to calculate a standard error. They can answer different questions and produce coefficients with different meanings. An AG analysis may be useful when the event-history-conditional intensity is the target; PWP when event order and reaching a prior event define the target; and WLW when marginal event-order-specific comparisons are of interest. LWYY is natural when the target is a covariate-associated ratio of population-average recurrent-event rates over follow-up.

Derivation, Diagnostics, and Common Pitfalls

The cumulative-count result follows directly from the mean-rate model. For fixed covariates, integrate both sides of the increment equation over the interval:

\begin{aligned} E\{N_i^*(t)-N_i^*(0)\mid Z_i\} &= E\left\{\int_0^t dN_i^*(u)\,\middle|\,Z_i\right\}\\ &= \int_0^t \exp(\beta^\top Z_i)\,d\Lambda_0(u)\\ &= \exp(\beta^\top Z_i)\{\Lambda_0(t)-\Lambda_0(0)\}. \end{aligned}

With $N_i^*(0)=0$ and $\Lambda_0(0)=0$, this is $\exp(\beta^\top Z_i)\Lambda_0(t)$. For two covariate profiles $z_a$ and $z_b$, dividing their expected counts gives $\exp\{\beta^\top(z_a-z_b)\}$, provided they are compared over the same interval under the same baseline model. This establishes why the modeled rate ratio also describes a cumulative-mean ratio in that setting.

In practice, inspect whether the model’s proportional-rate summary is plausible. Compare observed recurrent-event accumulation across covariate groups over follow-up, accounting for different amounts of observation time. Marked changes in the between-group rate ratio over time, or clear departures from a common covariate effect across periods, challenge the model. Also inspect whether follow-up differs substantially between groups and whether dropout or a terminal event may selectively remove subjects with different future recurrence rates.

A frequent mistake is to interpret $\exp(\beta)$ as a hazard ratio for the next event, conditional on how many events a person has already had. That is not the LWYY target: its rate is marginal over event history. Another mistake is to assume that “marginal” means an unconditional comparison of all subjects regardless of covariates. The coefficient is a contrast between specified covariate profiles under the model; the population-average event history is not the same as an unadjusted population contrast.

Key takeaway: LWYY specifies a proportional marginal mean rate for recurrent events. Its exponentiated coefficient is a population-average rate ratio over follow-up, not an event-order-specific effect or a hazard ratio conditional on an individual’s prior recurrences.

Check Your Understanding

Answer each question, then compare with the concise solutions below.

A binary covariate has fitted coefficient $\hat{\beta}=\log(0.80)$. What rate ratio does this imply, and how should it be interpreted?
If the comparison group’s expected cumulative count over a specified common interval is 2.5 and the modeled rate ratio is 1.20, what is the other group’s expected cumulative count under the model?
Does an LWYY covariate effect condition on the number of recurrences a subject has already experienced? Explain.
What does a changing baseline rate permit, and what proportionality requirement remains?
Name one observation or follow-up issue that could make the intended marginal-rate interpretation unreliable.

Solutions. (1) $\exp\{\log(0.80)\}=0.80$: the covariate group has a modeled population-average rate 0.80 times that of the comparison group, or 20% lower, over follow-up. (2) $2.5(1.20)=3.0$ expected events over the same interval. (3) No; the model is marginal with respect to prior recurrent-event history. (4) The baseline rate may vary freely over time, but the covariate rate ratio is assumed constant over the modeled follow-up. (5) For example, informative dropout related to future recurrence rates can bias inference if the analysis treats censoring as independent.

What You'll Learn