Worked Safety Exposure-Response Analysis

From Safety Records to an Exposure-Response Analysis

A safety exposure-response analysis must connect each event to the time during which it could have occurred and to the exposure relevant at that time. A count of adverse events without its corresponding time at risk can be misleading: two groups with the same event count may have very different rates if one group contributes substantially more observation time. The analysis must also distinguish an adverse event from an event that prevents or changes the opportunity to observe it, such as death or treatment discontinuation.

This worked analysis follows a hypothetical study of an adverse event of interest. It moves from risk-window definition and event-rate construction to rate modeling, competing-event checks, sensitivity analysis, and absolute-risk interpretation. The quantities are illustrative, not results from a real study. The introductory material on safety risk windows and adverse-event models provides the background for defining the analysis population and outcome; here the emphasis is on connecting those decisions coherently in one analysis.

Definition: A risk window is the prespecified period during which an individual is considered at risk for the event being analyzed. Person-time at risk is the total time contributed by individuals while they are in that window and have not yet had the event being counted. An event rate is the number of events divided by the corresponding person-time. Its units are events per unit time, not the proportion of individuals who experience an event.

For a valid rate comparison, event counts and person-time must refer to the same population, outcome definition, and risk-window rules. If exposure changes over time, person-time should be assigned to the exposure level applicable during each interval, rather than labeling a participant by an exposure they achieve later. The analysis should specify how it handles treatment interruption, dose modification, event onset dates, and any post-treatment risk period. Those choices affect both the numerator and denominator.

Constructing Rates and Comparing Exposure Groups

Consider a simple analysis in which each participant contributes time to a high- or low-exposure category. A rate in group \(g\) is \(r_g=Y_g/T_g\), where \(Y_g\) is the event count and \(T_g\) is person-time. The crude incidence rate ratio is \(r_H/r_L\). It compares event rates; it is not a risk ratio and does not directly state the chance that an individual experiences an event over a fixed period.

Worked Example: Building Crude Event Rates

Suppose a hypothetical study has 120 first events during 1,200 person-years in the high-exposure group, and 40 first events during 1,200 person-years in the low-exposure group. Assume the event definition and risk-window rules are identical in both groups, and that each individual contributes time only while event-free and at risk.

The group-specific rates are:

r_H=\frac{120}{1200}=0.1000\ \text{events per person-year},\qquad r_L=\frac{40}{1200}=0.03333\ \text{events per person-year}.

The crude incidence rate ratio is \(0.1000/0.03333=3.000\). In these hypothetical data, the observed event rate is three times as high in the high-exposure group as in the low-exposure group. This descriptive comparison alone does not show that exposure caused the difference; groups may differ in prognostic characteristics or other aspects of care.

For independent Poisson event counts with the person-time treated as fixed, an approximate standard error for the log rate ratio is \(\sqrt{1/Y_H+1/Y_L}\). Here it is:

SE\{\log(r_H/r_L)\} =\sqrt{\frac{1}{120}+\frac{1}{40}} =\sqrt{0.0333333} =0.1826.

Since \(\log(3)=1.0986\), an approximate 95% confidence interval on the log scale is \(1.0986\pm1.96(0.1826)\), or approximately \((0.7408,1.4565)\) using unrounded inputs. Exponentiating the endpoints gives an approximate interval of \((2.097,4.291)\) for the rate ratio. Under the model assumptions, the interval procedure has 95% coverage in repeated samples. The interval is compatible with a higher rate in the high-exposure group, but it does not address confounding or establish causality.

The log-scale interval follows from estimating \(\log(r_H/r_L)\), whose approximate sampling variance for independent Poisson counts is \(1/Y_H+1/Y_L\). Adding and subtracting 1.96 standard errors gives the approximate normal interval; exponentiation returns to the positive rate-ratio scale. The approximation can be unreliable with very small counts, especially zero counts, or when the Poisson variance assumption is poor. In those cases, a likelihood-based interval or an appropriate small-sample method may be preferable.

Adjusted Model and Model Checks

A common adjusted analysis divides follow-up into intervals for each individual and models each interval's event count. Let \(Y_{ij}\) be the count for individual \(i\) in interval \(j\), \(T_{ij}\) its person-time at risk, and \(E_{ij}\) the exposure value or category for that interval. A Poisson rate model has conditional mean proportional to person-time:

Y_{ij}\sim\operatorname{Poisson}(\mu_{ij}),\qquad \log(\mu_{ij})=\log(T_{ij})+\beta_0+\beta_E E_{ij}+\boldsymbol{\gamma}^{\mathsf T}\mathbf{Z}_{ij}.

Here \(\log(T_{ij})\) is an offset, meaning its coefficient is fixed at one. The vector \(\mathbf{Z}_{ij}\) contains specified covariates, and \(\beta_E\) describes the exposure association conditional on those covariates. If \(E\) is a binary high-versus-low indicator, \(\exp(\beta_E)\) is the adjusted rate ratio. If exposure is continuous, \(\exp(\beta_E)\) is the rate ratio for a one-unit increase, so the unit and functional form must be stated. The model assumes a correctly specified mean structure, conditionally independent counts given the modeled covariates, and Poisson variance equal to the conditional mean unless a variance adjustment is used.

Worked Example: Interpreting an Adjusted Rate Model

Suppose an illustrative Poisson model, fitted to interval-level data with person-time as an offset, estimates the high-versus-low exposure coefficient as \(\hat{\beta}_E=0.8755\), with \(SE=0.2500\). Assume the binary exposure coding, covariates, and interval structure were prespecified, and that the model's mean and variance assumptions have been checked sufficiently for this approximate Wald calculation.

The adjusted rate ratio is \(\exp(0.8755)=2.400\). Its approximate 95% confidence interval is obtained by first forming the coefficient interval:

0.8755\pm1.96(0.2500)=(0.3855,1.3655).

Exponentiating gives \((\exp(0.3855),\exp(1.3655))\), approximately \((1.470,3.918)\). Conditional on the covariates included in this hypothetical model, the estimated high-exposure event rate is 2.400 times the low-exposure rate. The interval expresses uncertainty in the estimated association under the model; adjustment does not remove confounding by unmeasured factors or make an observational exposure contrast causal.

The Wald statistic for testing \(H_0:\beta_E=0\) against \(H_a:\beta_E\ne0\) is \(z=0.8755/0.2500=3.502\), with a two-sided normal-approximation p-value of about 0.00046. This quantifies how incompatible the estimate is with a zero coefficient under the null model and the stated assumptions; it is not the probability that the null hypothesis is true.

Model checks should examine whether the fitted mean captures variation across exposure and follow-up, whether residuals or observed-to-predicted counts show systematic patterns, and whether a few participants or intervals dominate the result. If the observed variance exceeds the Poisson mean, conventional Poisson standard errors may be too small. A robust variance estimate or a model that allows overdispersion may be considered, while preserving a justified mean model. Repeated intervals from the same person also require attention to within-person dependence; treating them as independent can understate uncertainty. Sparse data, separation-like behavior, and strong collinearity can make estimates unstable.

Competing Events and Absolute Risk

A competing event prevents the adverse event of interest from occurring later in the same risk process. Death before the adverse event is a clear example; discontinuation may also end observation or alter the treatment context, depending on the estimand and protocol. A cause-specific rate model describes the instantaneous rate of the adverse event among participants who remain alive and free of all relevant events. It does not, by itself, give the probability of experiencing that event by a fixed time when competing events can occur.

For constant cause-specific rates over time, let \(\lambda_1\) be the adverse-event rate and \(\lambda_2\) the competing-event rate. The probability of remaining free of both events through time \(t\) is \(\exp[-(\lambda_1+\lambda_2)t]\). The adverse-event cumulative incidence is the probability that the adverse event occurs first by time \(t\):

F_1(t)=\int_0^t \exp[-(\lambda_1+\lambda_2)u]\lambda_1\,du =\frac{\lambda_1}{\lambda_1+\lambda_2} \left[1-\exp\{-(\lambda_1+\lambda_2)t\}\right].

The integrand is the probability of reaching time \(u\) without either event multiplied by the adverse-event rate at that time. Integrating gives the stated cumulative incidence. This expression relies on constant rates and the competing-event framework described; if rates vary over time, the survival probability and event rate must instead be integrated over their time-varying forms.

Worked Example: Translating Rates Into Absolute Risk

For illustration, suppose the high-exposure group has a constant adverse-event rate of \(0.1200\) per person-year and a competing-event rate of \(0.1800\) per person-year. In the low-exposure group, suppose the corresponding rates are \(0.04000\) and \(0.1000\) per person-year. These are hypothetical modeled rates. Calculate the probability of the adverse event occurring first within one year.

For high exposure, the total rate is \(0.1200+0.1800=0.3000\) per person-year. Thus:

F_{1,H}(1)=\frac{0.1200}{0.3000}\left[1-\exp(-0.3000)\right] =0.4000(0.2592)=0.1037.

The estimated one-year cumulative incidence is 10.37%. For low exposure, the total rate is \(0.04000+0.1000=0.1400\) per person-year:

F_{1,L}(1)=\frac{0.04000}{0.1400}\left[1-\exp(-0.1400)\right] =0.2857(0.1306)=0.03733.

The absolute difference is \(0.1037-0.03733=0.06637\), or about 6.637 percentage points, using the displayed rounded probabilities. These are model-based hypothetical risks over one year, not rate ratios. If the competing rate were ignored and the adverse-event rate alone used to calculate \(1-\exp(-\lambda_1t)\), the results would be 11.31% and 3.921%. Those differ because that calculation treats competing events as if they did not prevent later occurrence of the adverse event. The appropriate absolute-risk method depends on the event process and estimand.

Sensitivity Analysis and Interpretation

A primary result is only as credible as the risk-window and exposure assignments that produced it. Sensitivity analysis should vary consequential, clinically defensible choices rather than search indiscriminately for a preferred estimate. For example, the analyst can examine whether the exposure association changes when a protocol-justified post-treatment window is lengthened or shortened, when a plausible latency lag is applied, or when exposure is assigned using an alternative prespecified summary. A change may reflect genuine timing, misclassification, selection, or sparse-data instability; it should be interpreted, not merely reported.

Worked Example: Checking Dependence on the Risk Window

Suppose the primary analysis uses an on-treatment risk window plus a prespecified 14-day follow-up period and estimates an adjusted rate ratio of 2.40. In a hypothetical sensitivity analysis, the follow-up period is extended to 30 days for a clinically justified reason, with the same model structure, and the estimated rate ratio is 1.80. Assume the alternative definition adds person-time to both exposure groups but also captures events with a plausible post-treatment latency.

The two estimates suggest that the estimated association is sensitive to how post-treatment time is classified. The difference cannot be resolved by choosing the larger estimate: the analysis should show event counts and person-time under both definitions, assess whether added events occur in the added window, and explain which definition best represents the clinical question. The 1.80 estimate is not automatically more valid simply because it is smaller, and the 2.40 estimate is not automatically more credible because it comes from the primary definition.

Useful alternatives answer different questions. A Poisson rate model with person-time offset is convenient for event counts and rate ratios, especially when data are organized in intervals. A time-to-first-event model uses event timing and can accommodate censoring, but its hazard ratio is not an absolute risk and requires careful treatment of competing events. Recurrent-event models are more appropriate when repeated events are the target, but their estimand and dependence assumptions differ from a first-event analysis. Cumulative-incidence estimation or prediction is needed for absolute event probability when competing events matter. No method repairs a poorly defined risk window, inadequate exposure measurement, informative censoring, or residual confounding by itself.

Common Mistakes and Practical Checks

Comparing counts instead of rates. A larger count may result from more person-time. Report both event counts and person-time, then calculate rates.
Calling a rate ratio a risk ratio. A rate ratio compares events per unit time. A fixed-horizon risk is a probability and requires a model of time to event and, where relevant, competing events.
Assigning future exposure to earlier time. This can create exposure misclassification and selection bias. Assign exposure using information available for each risk interval.
Censoring competing events and calling the result actual risk. Censoring a competing event in a Kaplan–Meier calculation estimates a hypothetical net risk under assumptions, not the observed-world cumulative incidence when the competing event occurs.
Interpreting adjustment as causal proof. Covariate adjustment controls only the measured factors represented adequately in the model. Report the result as an adjusted association unless a causal design and assumptions justify more.

Key takeaway: A defensible safety exposure-response result requires aligned event counts and person-time, an explicit exposure and risk-window definition, a model suited to the event structure, checks for dependence and overdispersion, and an absolute-risk calculation that accounts for competing events when they are part of the clinical process.

Check Your Understanding

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

A group has 18 events over 600 person-years. What is its event rate, with units?
In a log-linear Poisson model, why is log person-time included as an offset rather than estimated as an ordinary coefficient?
In one sentence, distinguish an event rate ratio from a one-year risk ratio.
Why can treating death as ordinary censoring overstate the observed-world probability of a nonfatal adverse event?
Name two checks or sensitivity analyses that could help assess whether a safety exposure-response estimate is reliable.

Solutions. (1) \(18/600=0.03000\) events per person-year. (2) The expected count is proportional to time at risk, so log person-time enters with coefficient fixed at one; the modeled covariates describe the rate per unit time. (3) A rate ratio compares events per unit person-time, whereas a one-year risk ratio compares probabilities of experiencing the event within one year. (4) Death prevents the later nonfatal event, so treating it as if it only ended observation ignores a competing event and can yield a probability different from the actual cumulative incidence. (5) Examples include checking observed versus predicted counts and overdispersion, assessing within-person dependence, and repeating the analysis with a clinically justified alternative risk window or exposure lag.