Recurrent Adverse Events and Exposure

Why Repeated Events Need a Rate Model

A patient may experience several adverse events during follow-up. Treating each event as if it came from a different patient discards the fact that events from one patient may be related. At the same time, patients may contribute different amounts of observation time: someone followed for a year has more opportunity to experience an event than someone followed for a month. A safety analysis should therefore describe both the number of events and the time during which each patient was at risk.

The central quantity is often an event rate: the expected number of recurrent events per unit of time at risk. It is not the same as the probability that a patient experiences at least one event. A rate can exceed one event per person-year, because the same patient can contribute multiple events. This tutorial focuses on modeling recurrent counts and rates in relation to exposure, including methods that account for dependence among events from the same patient.

Time at risk must match the safety question and the event definition. It may exclude periods when an event cannot be observed or when a prespecified risk window is inactive. The risk-window decisions established in the earlier material on defining risk windows for exposure-related safety should be applied consistently. Counting events in one window while using a different window for the time-at-risk denominator can distort rates.

Poisson Rate Models and Their Assumptions

A basic model treats each patient’s event count as Poisson, conditional on that patient’s covariates and observation time. Let \(C_i\) denote exposure, \(T_i\) time at risk, and \(x_i\) other covariates. A log-linear rate model is

Equivalently, the log of expected count is modeled with \(\log(T_i)\) as an offset: \(\log(\mu_i)=\log(T_i)+\beta_0+\beta_1C_i+x_i^\top\gamma\). The offset coefficient is fixed at one; it accounts for the fact that twice as much time at risk gives twice the expected count when the rate is unchanged. The exponentiated linear predictor without the offset is the event rate per unit of time.

Under this model, \(\exp(\beta_1)\) is the conditional rate ratio for a one-unit increase in exposure, holding the other covariates fixed. The unit of exposure must be stated. A coefficient for exposure measured in mg/L, for example, describes a one mg/L increase; changing the exposure scale changes the coefficient but not the fitted rates. A rate ratio compares expected event rates, not event probabilities, and does not establish a causal effect of exposure.

Accounting for Within-Patient Dependence

Repeated events can cluster within patients. Some patients may be more susceptible to adverse events because of measured or unmeasured characteristics, making one event a marker of elevated risk for later events. A basic Poisson model can then underestimate uncertainty if it treats the observed counts as having only Poisson variation. A larger variance than the mean is called overdispersion. It may reflect patient heterogeneity, misspecified exposure effects, omitted covariates, or other features of the data; it is a diagnostic signal, not a diagnosis of its cause.

One response is to retain the Poisson mean model but use a patient-clustered robust (sandwich) variance estimate. This allows the variance estimate to accommodate dependence or extra variation among events from the same patient, provided there are enough independent patients for the approximation to be reliable. It changes the uncertainty estimate, not the fitted mean structure. It does not by itself model why some patients have higher rates, fix a misspecified mean, or remove confounding.

Another approach is a negative binomial count model, which allows the count variance to exceed its mean according to a specified dispersion structure. This may be useful for patient-level counts, but does not automatically describe event timing or the way a patient’s risk changes after an event. A shared frailty model represents latent patient-level rate differences explicitly.

The conditional interpretation of exposure effects matters. In the shared-frailty model, \(\exp(\beta_1)\) compares rates for a one-unit exposure increase among patients with the same frailty value and the same other covariates. The distribution of frailty describes residual variation across patients, not a measured biological mechanism. Its assumptions, including the chosen frailty distribution and its relationship to exposure, should be treated as modeling choices.

Recurrent-Event Intensity Models

When event times and changing covariates are available, a recurrent-event survival formulation can model the event intensity over follow-up rather than only the total count. Let \(N_i(t)\) be the cumulative number of events for patient \(i\) by time \(t\), and let \(Y_i(t)\) indicate whether that patient is under observation and at risk at time \(t\). A common conditional intensity specification is

Here \(\lambda_0(t)\) is a baseline event intensity, \(C_i(t)\) is exposure at time \(t\), and the other covariates may also vary over time. The intensity is the instantaneous expected event count per unit time, conditional on the patient’s observed history and model variables. A recurrent-event model must specify how follow-up and risk status are represented and how event history enters the model, if it does. The displayed form does not automatically make risk depend on the number or timing of prior events.

A shared frailty \(U_i\) links a patient’s event intensities and induces within-patient dependence after the frailty is integrated out. Without a frailty term, a recurrent-event intensity analysis may still use patient-level robust variance estimation to avoid treating a patient’s multiple events as independent contributions to uncertainty. The exposure coefficient describes a conditional intensity ratio under the specified model; interpretation depends on the time scale, risk process, covariate history, and any frailty conditioning.

Choosing and Checking an Approach

The main approaches answer related but not identical needs. Choose the simplest model that reflects the data structure and the estimand, then check whether its assumptions are plausible.

Useful checks include comparing observed and model-predicted counts across exposure groups and follow-up durations; examining the distribution of counts per patient, including the fraction with zero events and the upper tail; and checking whether the variance substantially exceeds the Poisson mean. For intensity models, inspect fit over time and whether event patterns differ after prior events. Assess whether exposure values and time at risk are correctly aligned, whether a few patients dominate the event total, and whether the fitted exposure relationship is plausible over the observed range. These checks can reveal misspecification, but no single diagnostic proves the model is correct.

A recurrent-event analysis is limited by how events are defined and recorded. Events that are not independent clinical episodes, or that are subject to inconsistent reporting, may not support a meaningful count-rate interpretation. If dropout or treatment discontinuation is related to event risk, simply using observed time at risk may not remove selection bias. Exposure itself may also be associated with baseline risk or treatment changes. As emphasized in the earlier material on missing outcomes and informative dropout, the observation process can be informative; the model’s conclusions then require careful assumptions and sensitivity analysis. A rate model does not convert observational exposure comparisons into causal effects.

The practical decision is therefore not simply “Poisson or not.” First define the event and risk window; then determine whether the target is a rate based on total counts or a time-varying intensity; include exposure and relevant covariates with appropriate time alignment; and decide how patient-level dependence will be handled. Report the time unit, exposure contrast, rate ratio or rate estimate, uncertainty, and the modeling approach clearly.