Exposure-Response For A Binary Efficacy Endpoint

Why Model a Binary Efficacy Endpoint?

Efficacy outcomes are often recorded as a response indicator at a prespecified assessment: for example, whether a patient met a clinical response criterion by a specified visit. The outcome has two possible values, so a model for a continuous response is not automatically appropriate. The analysis should estimate response probabilities that remain between zero and one while describing how those probabilities vary with exposure and baseline prognosis.

A logistic regression model provides one such description. It relates the log odds of response to exposure and selected baseline factors. Its coefficients describe changes in odds, while its predictions can be transformed into response probabilities for clinically interpretable exposure values. The predictions are conditional on the specified baseline factors; a separate standardization step can summarize them for a target population.

Definition: Let $Y_i=1$ indicate that patient $i$ meets the prespecified binary efficacy criterion and $Y_i=0$ otherwise. Let $E_i$ be a defined exposure measure, with its unit stated, and let $Z_i$ denote baseline prognostic factors. A logistic exposure–response model specifies $$ P(Y_i=1\mid E_i,Z_i)=p_i,\qquad \operatorname{logit}(p_i)=\log\left(\frac{p_i}{1-p_i}\right) =\beta_0+\beta_E E_i+\boldsymbol{\gamma}^{T}Z_i. $$ Here $\beta_0$, $\beta_E$, and the elements of $\boldsymbol{\gamma}$ are model coefficients. The particular exposure summary, outcome definition, and covariates must be specified for the analysis.

Model Meaning and Assumptions

The endpoint should be defined before fitting: its assessment time, response criterion, and handling of missing or unevaluable outcomes affect what $Y_i$ means. Exposure also needs a clear definition, such as an area-under-the-curve measure over a stated interval. If $E_i$ is measured in units of 10 mg·h/L, then $\beta_E$ is the change in log odds for a 1-unit increase, meaning an increase of 10 mg·h/L. Stating the unit prevents an otherwise correct coefficient from being clinically ambiguous.

The model has several important assumptions. Conditional on the included predictors, each outcome follows a Bernoulli distribution with probability $p_i$; patient outcomes are independent unless the analysis explicitly models dependence; and the log odds are correctly represented by the chosen predictor terms. In the displayed model, the exposure association is linear on the log-odds scale. This does not mean that response probability is linear in exposure. Baseline covariates should be chosen based on clinical knowledge and the analysis question, not solely on a data-driven significance screen.

In an exposure–response analysis, exposure is often not randomly assigned. Disease severity, adherence, dose changes, or other factors can affect both exposure and response. Adjusting for measured baseline prognostic factors can reduce some differences in prognosis, but it does not guarantee that all confounding has been removed. A fitted exposure association alone does not establish that changing exposure would cause the predicted change in response. In particular, variables that occur after treatment begins may be mediators or consequences of response; adding them as ordinary baseline adjustments can change the estimand or introduce bias.

The model is commonly estimated by maximum likelihood. Given predictor values, the Bernoulli likelihood is the probability of the observed outcomes under the model. The estimates are the coefficient values that maximize that likelihood. The model therefore requires enough information in the data to estimate its coefficients: very sparse response patterns or strong separation of responders from nonresponders can make ordinary maximum-likelihood estimates unstable or unbounded.

From a Coefficient to an Odds Ratio

The logistic link makes a coefficient directly interpretable on the log-odds scale. For two otherwise identical patients whose exposures differ by $\Delta E$, the model implies a difference of $\beta_E\Delta E$ in log odds. Exponentiating that difference gives the odds ratio. This follows by writing the odds at each exposure from the model and taking their ratio:

\frac{\text{odds at }E+\Delta E}{\text{odds at }E} =\frac{\exp(\beta_0+\beta_E(E+\Delta E)+\boldsymbol{\gamma}^{T}Z)} {\exp(\beta_0+\beta_E E+\boldsymbol{\gamma}^{T}Z)} =\exp(\beta_E\Delta E).

Thus, $\exp(\beta_E)$ is the odds ratio for a one-unit increase in the stated exposure unit, conditional on the modeled baseline factors. An odds ratio is not a probability ratio or a percentage-point change in response probability. The corresponding probability change depends on the starting probability and can be calculated from the odds or from the model prediction.

Worked Example: Converting an Exposure Coefficient to a Probability

Suppose a hypothetical fitted model estimates $\beta_E=0.55$ per 10 mg·h/L increase in exposure. For a patient at the lower exposure value, suppose the model-predicted response probability is 0.300. Compare that patient with an otherwise identical patient whose exposure is 10 mg·h/L higher.

Odds ratio. The odds ratio for this exposure increase is $ \exp(0.55)=1.7333 $. The lower-exposure odds are $0.300/(1-0.300)=0.4285714$. Therefore, the higher-exposure odds are $0.4285714\times1.7333=0.7428$, using the unrounded exponential in the multiplication.

Higher-exposure probability. Convert the new odds to a probability: $0.7428/(1+0.7428)=0.4262$, using unrounded odds before rounding the result. The predicted probability difference is $0.4262-0.3000=0.1262$, or 12.62 percentage points.

This is a model-based comparison at the specified baseline covariate values. It says that the fitted model predicts a 0.1262 higher response probability for the higher exposure in this example; it does not mean the odds ratio itself is a 73.33 percentage-point increase, nor does it establish a causal effect.

Predicting Across the Observed Exposure Range

A practical use of the fitted model is to calculate probabilities across exposure values that are represented in the data. The following hypothetical model uses exposure in units of 10 mg·h/L and a baseline severity indicator $S$, coded 0 for lower severity and 1 for higher severity:

\operatorname{logit}(p)=-2.00+0.120E-0.700S.

The positive exposure coefficient corresponds to an odds ratio of $\exp(0.120)=1.127$ for each exposure-unit increase, conditional on baseline severity. The negative severity coefficient corresponds to lower response odds for the higher-severity group at the same exposure. These are conditional associations encoded by the hypothetical fitted model.

Worked Example: Comparing Predicted Probabilities by Baseline Severity

Assume the hypothetical model above was fitted to data in which exposure ranged from 10 to 30 units. Calculate predictions at exposures 10, 20, and 30 for each severity group. For any linear predictor $\eta$, convert to probability using $p=\exp(\eta)/(1+\exp(\eta))$.

Lower severity ($S=0$). At $E=10$, $\eta=-2.00+0.120(10)=-0.800$, so $p=\exp(-0.800)/(1+\exp(-0.800))=0.3100$. At $E=20$, $\eta=-2.00+0.120(20)=0.400$, so $p=\exp(0.400)/(1+\exp(0.400))=0.5987$. At $E=30$, $\eta=-2.00+0.120(30)=1.600$, so $p=\exp(1.600)/(1+\exp(1.600))=0.8320$.

Higher severity ($S=1$). At $E=10$, $\eta=-2.00+0.120(10)-0.700=-1.500$, giving $p=0.1824$. At $E=20$, $\eta=-2.00+0.120(20)-0.700=-0.300$, giving $p=0.4256$. At $E=30$, $\eta=-2.00+0.120(30)-0.700=0.900$, giving $p=0.7109$.

The model predicts increasing response probability over this exposure range in both severity groups, with lower probabilities for the higher-severity group at each listed exposure. The predictions are not evidence that the model is correct: they should be examined alongside the observed data and diagnostics.

Standardizing Predictions to a Baseline Population

A prediction for one specified covariate profile answers a conditional question. If the aim is to compare response probabilities for a population with a mixture of baseline prognostic factors, average individual predictions over that population instead. This is often called standardization or marginal prediction. It preserves the baseline covariate distribution while setting exposure to chosen values in the prediction calculation.

Worked Example: Standardized Predictions for Two Exposure Values

Suppose a hypothetical target population has 60% lower-severity patients and 40% higher-severity patients. Use the model and predictions from the preceding example to calculate standardized probabilities at $E=10$ and $E=30$.

At $E=10$, the standardized prediction is $0.60(0.3100)+0.40(0.1824)=0.2590$, using unrounded group predictions before rounding. At $E=30$, it is $0.60(0.8320)+0.40(0.7109)=0.7836$. The difference is $0.7836-0.2590=0.5246$, or 52.46 percentage points.

These are model-based standardized predictions for the stated baseline composition. The contrast summarizes how predicted responses differ when the model's exposure input is set to those two values while retaining the same severity distribution. As with the conditional predictions, a causal interpretation requires additional assumptions, including adequate control of confounding and support for the exposure comparisons in the target population.

Diagnostics, Alternatives, and Limitations

Predicted probabilities are most defensible within the observed exposure range and where patients with comparable baseline profiles are represented. Before interpreting an exposure curve, inspect the distribution of exposure among responders and nonresponders and assess whether there is meaningful overlap across prognostic groups. If high exposure occurs only in one type of patient, the model may be relying on extrapolation when asked to compare patients at the same exposure but different baseline profiles.

The linear exposure term assumes a straight-line relationship between exposure and log odds. Check whether the observed pattern and residual or calibration summaries are compatible with that assumption. A systematic mismatch can indicate a need for a more flexible exposure term, but added complexity also requires enough data to estimate it reliably. Assess calibration by comparing observed and predicted response proportions across suitable groups or using an appropriate calibration display. Large discrepancies, especially concentrated at particular exposure values, suggest that predictions may be unreliable.

Other warning signs include extreme fitted probabilities, sparse cells, influential observations, and coefficient estimates that change markedly under reasonable model checks. Complete separation occurs when a predictor or predictor combination perfectly distinguishes responders from nonresponders; ordinary logistic maximum likelihood can then fail to yield finite estimates. Penalized methods may be considered, but they do not repair a poorly defined endpoint, lack of exposure overlap, or an unsuitable model structure. If patients contribute multiple outcomes or are clustered by site, the independence assumption also needs attention through a model or inference procedure that accounts for that dependence.

A model with exposure grouped into categories can be easier to display and does not impose a linear log-odds trend across exposure, but grouping discards information and results depend on cut points. A linear probability model is simple to interpret in percentage points, but can predict values below zero or above one and typically has nonconstant outcome variance. Logistic regression is therefore a common default for a binary endpoint, not a guarantee of correct form or causal interpretation. In all approaches, uncertainty in fitted coefficients should be reflected when reporting predictions; a point prediction alone does not convey its precision.

Common pitfall: “The exposure odds ratio is 1.73, so response probability rises by 73 percentage points.” This confuses an odds ratio with a probability difference. In the worked example, an odds ratio of 1.7333 moves the probability from 0.3000 to 0.4262, a difference of 0.1262. Convert odds to probabilities at the relevant starting risk, and describe the result as an adjusted model prediction rather than an automatic causal effect.

Key takeaway: Define the binary endpoint and exposure unit, fit and check the logistic model, then report predicted probabilities at clinically relevant exposure values with baseline factors specified or standardized. Keep the predictions within the data's support and distinguish adjusted association from causation.

Check Your Understanding

Work through each question, then compare with the solutions below.

A fitted exposure coefficient is 0.40 per exposure unit. What is the conditional odds ratio for a two-unit increase?
For a hypothetical model $\operatorname{logit}(p)=-1.50+0.080E$, calculate the predicted probability at $E=5$.
Using the severity-specific predictions above, calculate the standardized predicted response at $E=10$ and $E=30$ for a target population that is 50% lower severity and 50% higher severity. What is the difference?
Why does adjustment for baseline severity not, by itself, prove that increasing exposure causes response to improve?

Solutions. (1) The log-odds difference is $0.40(2)=0.80$, so the odds ratio is $\exp(0.80)=2.226$ (approximately). (2) The linear predictor is $-1.50+0.080(5)=-1.10$. Thus $p=\exp(-1.10)/(1+\exp(-1.10))=0.2497$. (3) At $E=10$, the average is $0.50(0.3100)+0.50(0.1824)=0.2462$. At $E=30$, it is $0.50(0.8320)+0.50(0.7109)=0.7715$. The difference is $0.7715-0.2462=0.5253$, or 52.53 percentage points. (4) Exposure may be associated with other causes of response, including factors not measured or not modeled. An adjusted association has a causal interpretation only under additional assumptions, not from baseline adjustment alone.