Binary Adverse Events and Exposure

Why Binary Safety Outcomes Need a Defined Risk Window

A binary adverse-event outcome records whether a participant experienced at least one event within a specified period: yes or no. That simplicity can conceal an important issue. Someone followed for 30 days has less opportunity to experience an event than someone followed for 180 days. Therefore, an exposure comparison based on “ever had an event” is interpretable only when the risk windows are comparable or when the model accounts for their durations.

The outcome must also be defined consistently: specify the event, the start of the risk window, the end of the window, and how early discontinuation or missing follow-up is handled. A binary endpoint of “at least one event” is different from the number of events, which is outside the scope here. Exposure should be summarized over a prespecified, relevant period; a summary measured after an event may not represent exposure preceding that event.

Definition: For participant \(i\), let \(Y_i=1\) if the specified adverse event occurs during the defined risk window and \(Y_i=0\) otherwise. An exposure-response model relates \(P(Y_i=1)\) to a selected exposure measure and, where appropriate, other covariates. Its estimate is an exposure association conditional on the model; by itself, it does not establish a causal effect.

Logistic Models for Comparable Risk Windows

When participants have a common, clinically meaningful follow-up duration, a logistic model directly describes the probability of an event by the end of that window. For a continuous exposure \(x_i\) and covariates \(z_i\), one simple model is:

\log\left(\frac{p_i}{1-p_i}\right)=\beta_0+\beta_1x_i+\boldsymbol{\gamma}^{\mathsf T}z_i,\qquad p_i=P(Y_i=1\mid x_i,z_i).

The assumptions include a correctly defined binary outcome, independent participants conditional on the modeled predictors (or an analysis that accounts for dependence), a suitable functional form for exposure and covariates on the log-odds scale, and no complete or near-complete separation that makes coefficient estimates unstable or unbounded. The Bernoulli likelihood also assumes each participant contributes one such binary outcome. If the data arise from unequal follow-up windows, this model does not automatically correct for differing opportunity to experience an event.

For a one-unit increase in exposure, holding the modeled covariates fixed, \(\exp(\beta_1)\) is the conditional odds ratio. It is not generally a risk ratio. Given a fitted linear predictor \(\eta_i\), the event risk is \(p_i=1/(1+\exp(-\eta_i))\). Model-based risks at selected exposure values can be compared as a risk difference, \(p_1-p_0\), or a risk ratio, \(p_1/p_0\). These absolute-risk summaries depend on the covariate values or population over which they are calculated.

Worked Example: Interpreting a Logistic Exposure Effect

Suppose a hypothetical study follows all participants for 30 days. A fitted model uses exposure \(x\), measured in mg/L, and has \(\hat\beta_0=-3.00\) and \(\hat\beta_1=0.180\) per mg/L. For simplicity, there are no additional covariates in this example. Interpret the exposure coefficient, calculate risks at 5 and 10 mg/L, and form a 95% interval for the per-unit odds ratio if the standard error of \(\hat\beta_1\) is 0.0600.

The odds ratio for a 1 mg/L increase is \(\exp(0.180)=1.197\). Thus, under this model, each 1 mg/L increase is associated with 1.197 times the odds of an event over the 30-day window. The estimate is not a 19.7% increase in risk.

At 5 mg/L, the linear predictor is \(-3.00+0.180(5)=-2.10\), so the modeled risk is \(1/[1+\exp(2.10)]=0.1091\), or 10.91%. At 10 mg/L, it is \(-3.00+0.180(10)=-1.20\), giving \(1/[1+\exp(1.20)]=0.2315\), or 23.15%. The modeled risk difference is \(0.2315-0.1091=0.1224\), approximately 12.24 percentage points. These are model-based risks at the specified exposures, not guaranteed rates in a future study.

A Wald interval on the log-odds coefficient is \(0.180\pm1.96(0.0600)=[0.0624,0.2976]\). Exponentiating the endpoints gives an odds-ratio interval of approximately \([1.064,1.347]\). Under the model and interval procedure, the interval is compatible with odds increases of roughly 6.4% to 34.7% per 1 mg/L. The interval concerns the odds ratio, not the absolute risk difference.

Unequal Follow-Up: Complementary Log-Log With Duration

If follow-up durations differ, a binary event indicator can still be analyzed under a specific event-rate model. Suppose the event has a constant instantaneous rate \(\lambda_i\) during participant \(i\)'s observed interval of duration \(t_i\), conditional on exposure and covariates. Then the probability of no event is \(\exp(-\lambda_i t_i)\), and the probability of at least one event is \(p_i=1-\exp(-\lambda_i t_i)\). Taking the complementary log-log transformation yields:

\log[-\log(1-p_i)]=\log(t_i)+\log(\lambda_i).

A linear model for the log rate therefore gives:

\log[-\log(1-p_i)]=\log(t_i)+\beta_0+\beta_1x_i+\boldsymbol{\gamma}^{\mathsf T}z_i.

Here \(\log(t_i)\) is an offset: its coefficient is fixed at 1, rather than estimated. The assumption is that, within the modeled interval and conditional on predictors, the event rate is constant. Follow-up must be measured consistently (for example, in days), and the exposure summary must represent the interval in a defensible way. If the event rate changes over time, or exposure changes substantially over the interval, a single constant-rate summary may be inadequate. The simple offset model also relies on the way observation ends not creating unaccounted selection or informative loss of follow-up.

In this model, \(\exp(\beta_1)\) is a rate ratio for a one-unit exposure increase, conditional on covariates. It is not an odds ratio. For a target duration \(t\), the predicted risk is \(1-\exp\{-t\exp(\beta_0+\beta_1x+\boldsymbol{\gamma}^{\mathsf T}z)\}\). This conversion makes duration-specific absolute risks and risk differences available for interpretation.

Worked Example: Converting Rates to Risks at a Common Duration

Consider a hypothetical constant-rate model with two exposure categories and rates of 0.00200 events per person-day at the lower exposure and 0.00300 events per person-day at the higher exposure. Follow-up durations vary across participants, so the model includes \(\log(t_i)\) as an offset. Calculate the rate ratio and the modeled risks over 20 days.

The rate ratio is \(0.00300/0.00200=1.50\). Over 20 days, the lower-exposure cumulative rate is \(0.00200(20)=0.0400\), giving risk \(1-\exp(-0.0400)=0.03921\), or 3.921%. The higher-exposure cumulative rate is \(0.00300(20)=0.0600\), giving risk \(1-\exp(-0.0600)=0.05824\), or 5.824%.

Using the unrounded risks, the difference is \([1-\exp(-0.0600)]-[1-\exp(-0.0400)]=0.01902\), or 1.902 percentage points. This is the modeled absolute risk increase over 20 days. It is not the rate ratio: the relative rate is 1.50, while the risk ratio over this window is approximately \(0.0582355/0.0392106=1.485\).

Why Odds Ratios Depend on the Risk Window

A logistic coefficient describes odds over the window used to define its binary endpoint. Even if the underlying rate ratio is constant, an odds ratio can change with follow-up duration. The next calculation illustrates why an odds ratio from one risk window should not be treated as a duration-free measure of the exposure association.

Worked Example: The Same Rate Ratio, Different Odds Ratios

Use the hypothetical event rates of 0.00200 and 0.00300 per day from the previous example. The rate ratio is 1.50 at every duration under the constant-rate assumption. At 10 days, the risks are \(1-\exp(-0.0200)=0.01980\) and \(1-\exp(-0.0300)=0.02955\). The odds ratio is:

\frac{0.0295545/(1-0.0295545)}{0.0198013/(1-0.0198013)} =\frac{\exp(0.0300)-1}{\exp(0.0200)-1} \approx 1.508.

At 100 days, the risks are \(1-\exp(-0.200)=0.1813\) and \(1-\exp(-0.300)=0.2592\). The corresponding odds ratio is \([\exp(0.300)-1]/[\exp(0.200)-1]\approx1.580\). Thus the rate ratio remains 1.50, while the odds ratio changes as the risk window lengthens. At low risks, odds and risks are close, so the odds ratio may approximate a risk ratio; that approximation becomes less reliable as risks increase.

Estimation, Diagnostics, and Interpretation

Both models are commonly estimated by maximum likelihood: coefficients are selected to make the observed event and non-event pattern most likely under the assumed model. Inference should reflect the study design and model. For logistic regression, a coefficient interval can be calculated on the log-odds scale and exponentiated for an odds-ratio interval. For a complementary log-log model, exponentiating an exposure coefficient gives a rate-ratio interval. Absolute-risk intervals require transforming the fitted model and accounting for uncertainty in its coefficients; simply transforming a point estimate does not provide an interval.

Before interpreting estimates, check whether the data support the model:

Check event counts and separation. Few events, especially within exposure ranges or covariate patterns, can produce unstable estimates and very wide intervals. If one predictor pattern perfectly separates event from non-event participants, ordinary logistic maximum-likelihood estimates may not be finite.
Check functional form. A straight-line exposure term assumes a linear change in log odds or log rate, not necessarily in risk. Inspect fitted patterns and use a clinically defensible exposure shape if the simple form is inadequate; avoid adding complexity unsupported by the data.
Check follow-up accounting. Compare durations across exposure groups and inspect how early exits and missing event status were handled. An offset is not a general cure for informative loss to follow-up or for rates that change over time.
Check calibration and influential observations. Compare observed and predicted event frequencies across meaningful risk or exposure groupings, and investigate observations with unusual predictor values or strong influence on the fitted coefficients.
Check the exposure definition and timing. Confirm that the chosen summary is measured over the intended risk period and does not incorporate information occurring after the event.

A logistic model is a natural choice for a common fixed risk window and directly estimates event odds. A complementary log-log model with a duration offset is useful when unequal durations are represented by a constant-rate assumption, and its exposure coefficient has a rate-ratio interpretation. Neither approach is automatically preferable in every setting: the outcome definition, follow-up structure, event frequency, and scientific assumptions should determine the choice. If duration varies but the constant-rate assumption is implausible, forcing a single offset model can obscure time-varying risk; a more suitable analysis may be needed.

In either model, an exposure coefficient is not automatically causal. Participants with different exposures may also differ in disease severity, baseline risk, treatment duration, or other factors that influence adverse events. Covariate adjustment can address measured differences only to the extent that the variables, timing, and model are appropriate. Report the modeled population, exposure contrast, risk window, adjustment variables, and uncertainty, and distinguish adjusted association from causal interpretation.

Common Pitfall

Common mistake: “The odds ratio is 1.5, so the event risk is 50% higher, regardless of how long participants were followed.”

Correction: An odds ratio compares odds, not risks, and applies to the endpoint’s specified risk window. With unequal follow-up, a fixed-window logistic comparison may be misleading. A complementary log-log model with a duration offset can account for duration only under its rate assumptions. Translate either model into absolute risks at a stated duration before discussing clinical impact.

Check Your Understanding

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

A logistic exposure coefficient is 0.20 per mg/L. What is the odds ratio for a 1 mg/L increase, and what does it compare?
Why is an offset \(\log(t_i)\) used in the complementary log-log model? State the key rate assumption.
A model predicts 4% risk at one exposure and 7% at another for the same 30-day window. Give the risk difference in percentage points and explain what it describes.
Two groups have the same estimated event rate ratio but different follow-up durations. Must their odds ratios be equal? Explain.
Name one diagnostic concern if almost no events occur in a high-exposure subgroup.

Solutions. (1) \(\exp(0.20)=1.221\). It compares the event odds over the defined risk window for participants differing by 1 mg/L, conditional on modeled covariates. (2) Under a constant event rate during the interval, \(-\log(1-p_i)=\lambda_i t_i\); taking logs gives the fixed-coefficient duration term \(\log(t_i)\). (3) The difference is \(7\%-4\%=3\) percentage points. It is the modeled absolute difference in 30-day event risk for the specified exposure contrast and covariate setting or population. (4) No. The odds ratio can change with the duration even when the rate ratio is constant, as the worked calculation showed. (5) Estimates may be unstable because of sparse data; separation can also make ordinary logistic maximum-likelihood coefficients unbounded if event status is perfectly predicted.