Explaining Conditional Effects to Clinical Collaborators

What a Conditional Effect Says

A clinical collaborator hearing “the odds ratio was 2.0” needs to know more than the number. Which patients are being compared? What characteristics are held fixed? Does the comparison concern patients in the same clinic, or patients averaged across clinics? In a patient-level logistic generalized linear mixed model (GLMM), these questions matter because the effect is conditional on the model’s random effects as well as on the included patient covariates.

Consider a binary outcome such as whether a patient experiences an adverse event. A logistic random-intercept model can represent both patient-level predictors and differences in baseline outcome propensity across clinics. The fixed-effect coefficient for a patient-level predictor describes a comparison made at a shared value of the clinic random intercept. A clear explanation names that comparison rather than presenting the coefficient as an unqualified clinical effect.

Definition: In a logistic GLMM, a conditional effect is an effect in the outcome model conditional on specified covariates and random effects. For a patient-level predictor in a random-intercept model, its conditional odds ratio compares patients who differ in that predictor while holding the other modeled covariates and the clinic random intercept fixed. It is therefore a comparison within a shared modeled clinic context, not a comparison of different clinics.

The Model and the Patient Comparison

Let \(Y_{ij}\) be a binary outcome for patient \(i\) in clinic \(j\), and let \(x_{ij}\) be a patient-level predictor. A simple model with a random clinic intercept is

Y_{ij}\mid b_j \sim \operatorname{Bernoulli}(p_{ij}),\qquad \operatorname{logit}(p_{ij})=\beta_0+\beta_1 x_{ij}+\boldsymbol{\beta}^{\mathsf T}\mathbf{z}_{ij}+b_j,\qquad b_j\sim N(0,\sigma_b^2).

Here \(\mathbf{z}_{ij}\) denotes other modeled patient covariates, and \(b_j\) is the clinic-specific deviation in baseline log odds. The model assumes that, conditional on the random effect and included predictors, the outcome follows a Bernoulli distribution with the stated probability. It also assumes the clinic random intercepts follow the specified normal distribution and are independent across clinics. In the usual random-intercept formulation, conditional outcomes from different patients are independent given the predictors and their clinic’s random intercept.

For a one-unit increase in \(x_{ij}\), with the other covariates and \(b_j\) fixed, the log odds change by \(\beta_1\). Exponentiating gives the conditional odds ratio. For a two-level predictor coded 0 and 1, it is the conditional odds for the group coded 1 divided by the conditional odds for the group coded 0.

\operatorname{OR}_{\text{conditional}}=\exp(\beta_1).

The cancellation of the clinic intercept is the reason the comparison is made within a shared clinic context. At fixed \(\mathbf{z}_{ij}\) and \(b_j\), the odds at \(x_{ij}=1\) are \(\exp(\beta_1)\) times the odds at \(x_{ij}=0\). This does not mean the two patients are identical in every respect: it means the comparison holds fixed the modeled covariates and clinic-specific intercept, while changing the predictor of interest.

Worked Examples

Worked Example: Explaining a Treatment Odds Ratio

Hypothetical output. In a study of 960 patients from 24 clinics, a logistic random-intercept GLMM models a binary adverse event. Treatment is coded 1 for the new treatment and 0 for usual care. The fitted treatment coefficient is \(0.70\), with \(SE=0.24\). Assume the model form and random-intercept assumptions above are appropriate, the treatment coding is correct, and the reported Wald interval is suitable for this estimate.

The conditional odds ratio is

\exp(0.70)=2.014.

A 95% Wald confidence interval on the log-odds scale is

0.70\pm 1.96(0.24)=0.70\pm 0.4704=(0.2296,\ 1.1704).

Exponentiating both endpoints gives \((\exp(0.2296),\exp(1.1704))=(1.258,\ 3.223)\). The Wald statistic is \(z=0.70/0.24=2.917\), corresponding to a two-sided \(p\)-value of about \(0.0035\).

Clinical wording. “After adjustment for the patient characteristics in the model, the estimated odds of an adverse event for a patient receiving the new treatment were about 2.01 times the odds for a comparable patient receiving usual care in the same modeled clinic context (95% CI 1.26 to 3.22; \(p=0.0035\)).” This describes a conditional odds comparison. It does not say that the probability of an adverse event doubled, and it does not establish that treatment caused the difference.

Worked Example: Interpreting a Continuous Patient Predictor

Hypothetical output. A model for 30-day readmission includes patient age in units of 10 years, other patient covariates, and a clinic random intercept. The age coefficient is \(0.15\), with \(SE=0.06\). Suppose the outcome coding, covariate specification, and random-intercept assumptions are reasonable for this interpretation.

The conditional odds ratio for a 10-year age increase is \(\exp(0.15)=1.162\). A 95% Wald interval on the coefficient scale is \(0.15\pm1.96(0.06)=(0.0324,\ 0.2676)\), which exponentiates to \((1.033,\ 1.307)\). The Wald statistic is \(0.15/0.06=2.50\), giving a two-sided \(p\)-value of approximately \(0.0124\).

Clinical wording. “Holding the modeled patient characteristics and clinic context fixed, a patient who is 10 years older has estimated readmission odds 1.16 times those of a patient 10 years younger (95% CI 1.03 to 1.31; \(p=0.0124\)).” The comparison is per 10 years, not per one year. It describes an adjusted association; it does not imply that age itself can be changed or that the association is causal.

Worked Example: Why the Same Odds Ratio Does Not Mean the Same Probability Change

Hypothetical calculation. Suppose the conditional treatment odds ratio from the first example is 2.014. At one specified clinic context and fixed covariate values, suppose the modeled probability under usual care is 0.10. The corresponding odds are \(0.10/(1-0.10)=0.1111\). Under treatment, the odds are \(2.014(0.1111)=0.2238\), giving a probability of \(0.2238/(1+0.2238)=0.1829\). The probability difference is about 0.0829, or 8.29 percentage points.

At a different clinic context or covariate profile, suppose the usual-care probability is 0.40. Its odds are \(0.40/0.60=0.6667\). Multiplying by the same conditional odds ratio gives treatment odds of \(2.014(0.6667)=1.343\), and a treatment probability of \(1.343/(1+1.343)=0.5732\). Here the probability difference is about 17.3 percentage points.

Both calculations use the same conditional odds ratio, but the probability changes differ because the starting probabilities differ. These are model-based comparisons at the stated clinic contexts and covariate values; they are not population-average treatment effects.

A Short Justification and the Scope of the Claim

The conditional interpretation follows directly from the odds in the model. Write the linear predictor excluding the predictor of interest as \(c=\beta_0+\boldsymbol{\beta}^{\mathsf T}\mathbf{z}_{ij}+b_j\). For \(x_{ij}=0\), the log odds are \(c\), so the odds are \(\exp(c)\). For \(x_{ij}=1\), the log odds are \(c+\beta_1\), so the odds are \(\exp(c+\beta_1)\). Their ratio is

\frac{\exp(c+\beta_1)}{\exp(c)}=\exp(\beta_1).

The clinic-specific intercept cancels, but it has not disappeared from the model: it still changes the baseline odds and hence the probabilities. The cancellation shows that a random-intercept model implies a common conditional odds ratio across clinics for that predictor, provided the model has no predictor-by-clinic random slope or relevant interaction.

This is not automatically the same as an odds ratio calculated after averaging probabilities over clinics. In general, an average of probabilities transformed to odds does not yield the same odds ratio as a comparison conditional on a clinic random effect. Logistic odds ratios are noncollapsible: conditional and marginal odds ratios can differ even without confounding. If the clinical question concerns the average outcome probability in a target population, report a population-average quantity or standardized predicted probabilities appropriate to that question; do not relabel the conditional coefficient as marginal.

How to Check Whether the Wording Is Defensible

A polished sentence cannot repair a model that does not support its comparison. Before explaining a conditional effect, check that the outcome is coded as described, the reference group and units are known, and the fitted model includes the intended clinic structure. Review convergence and whether estimates or standard errors are unstable. Examine whether outcome events and non-events are represented across predictor groups and clinics; sparse data or separation can make estimates unreliable. Consider whether an assumed common patient-level effect across clinics is plausible. If treatment effects are expected to vary by clinic, a random-intercept-only model’s single conditional odds ratio may conceal that variation.

Also check the scale and scope of each statement. An odds ratio above 1 means higher odds, not necessarily a higher probability by the same factor. A confidence interval describes uncertainty in the estimated conditional odds ratio under the model; it is not a probability that the fixed parameter lies inside the particular interval. A small \(p\)-value is evidence against a zero coefficient under the model and test assumptions, not the probability that the null hypothesis is true.

Common pitfall: “The treatment doubled the chance of an adverse event.” This is not justified by a conditional odds ratio of 2.01. The estimate says the modeled odds were about twice as large for the specified patient comparison within a shared clinic context. To describe chance or risk, give probabilities or an appropriate probability contrast, with the covariate and clinic context stated.

A fixed-effect logistic model without clinic random effects may be appropriate when the analysis does not model clinic clustering, but it does not provide the same clinic-conditional interpretation and may fail to account for within-clinic dependence. Conversely, a population-averaged model answers a different question: it targets average associations in the population rather than comparisons conditional on a clinic random effect. Neither is universally preferable. The clinical question, study design, and model assumptions determine which interpretation is relevant.

A useful reporting sentence has four parts: name the outcome and predictor contrast; state the patient-level comparison; say that other modeled covariates and the clinic context are held fixed; then give the conditional odds ratio and uncertainty. Add a probability contrast only when it is calculated for a clearly specified patient profile or population. Avoid causal wording unless the study design and assumptions justify it.

Check Your Understanding

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

A treatment coefficient is \(0.40\) in a patient-level logistic random-intercept model. What is the conditional odds ratio, and what is the clinic context for its interpretation?
Why is “the odds ratio was 1.8, so the probability increased by 80%” incorrect?
For age measured per 10 years, a coefficient is \(-0.20\). State the conditional comparison in words and calculate its odds ratio.
What changes in the interpretation if a clinic-specific random slope for treatment is included?

Solutions. (1) \(\exp(0.40)=1.492\), or about 1.49. At fixed modeled patient covariates and clinic random intercept, the odds for a patient in the treatment group are about 1.49 times the odds for a comparable patient in the reference group. (2) An odds ratio compares odds, not probabilities; the probability change depends on the starting probability and clinic context. (3) For patients differing by 10 years in age, holding the other modeled covariates and clinic context fixed, the older patient’s odds are \(\exp(-0.20)=0.819\) times the younger patient’s odds. This is about 18.1% lower odds, not an 18.1 percentage-point probability decrease. (4) The treatment effect can vary across clinics through the random slope, so one common conditional treatment odds ratio no longer describes every clinic. A reported effect must specify whether it is for a particular clinic effect or summarizes the modeled distribution of clinic-specific effects.