Conditional Odds Ratios in Logistic Random-Intercept Models

Why Condition on the Cluster?

In clustered binary data, observations from the same cluster may share a baseline propensity for the outcome. A logistic random-intercept model represents this cluster-specific baseline with a random intercept. Its covariate coefficient describes how the conditional odds change when the covariate changes while the cluster is held fixed. The key result is that the shared random intercept cancels from this within-cluster odds ratio.

This is an odds comparison, not a probability difference. It is also conditional: it compares observations at the same value of the cluster effect, rather than averaging over a distribution of clusters. The distinction matters because a constant odds ratio does not imply a constant change in probability.

Model: For binary response $Y_{ij}$ from observation $i$ in cluster $j$, let $Y_{ij}\mid b_j,x_{ij}\sim\text{Bernoulli}(p_{ij})$, with $$ \operatorname{logit}(p_{ij})=\log\left(\frac{p_{ij}}{1-p_{ij}}\right)=\beta_0+\beta_1x_{ij}+b_j. $$ Here $x_{ij}$ is a covariate, $\beta_0$ is the fixed intercept, $\beta_1$ is its fixed slope, and $b_j$ is the random intercept for cluster $j$. A standard random-intercept specification assumes $b_j\sim N(0,\sigma_b^2)$, with effects independent across clusters and independent of the covariates. Conditional on the random effects and covariates, responses are assumed independent. The Bernoulli probabilities must lie between 0 and 1, as they do under the logit link.

The normal distribution for $b_j$ is a common modelling assumption used to describe and estimate cluster variation. The algebraic derivation of the conditional odds ratio needs only the shared-intercept structure: the same $b_j$ must apply to both covariate values being compared. The interpretation of $\beta_1$ is conditional on that cluster effect and on other included covariates being held fixed.

Deriving the Conditional Odds Ratio

For a fixed cluster effect $b_j$, the conditional odds at covariate value $x$ are obtained by exponentiating the linear predictor. Write these odds as $O_j(x)$. For a change from $x$ to $x+1$, the odds ratio compares the two odds within that same cluster.

O_j(x)=\frac{p_j(x)}{1-p_j(x)} =\exp(\beta_0+\beta_1x+b_j).

\begin{aligned} \operatorname{OR}_{j}(x+1:x) &=\frac{O_j(x+1)}{O_j(x)}\\ &=\frac{\exp\{\beta_0+\beta_1(x+1)+b_j\}} {\exp(\beta_0+\beta_1x+b_j)}\\ &=\exp(\beta_1). \end{aligned}

Both the fixed intercept and the cluster-specific intercept appear in the numerator and denominator, so both cancel. The covariate contribution changes by exactly $\beta_1$ on the log-odds scale, and exponentiating that difference gives the odds ratio. It follows that in this model the one-unit conditional odds ratio is the same for every cluster, even though clusters can have different baseline odds.

Result: In a logistic random-intercept model, a one-unit increase in $x$, holding the cluster and other covariates fixed, multiplies the conditional odds of $Y_{ij}=1$ by $\exp(\beta_1)$. If $\beta_1>0$, the conditional odds increase; if $\beta_1<0$, they decrease; and if $\beta_1=0$, the conditional odds are unchanged. This result concerns odds, not a fixed increase in probability.

For a change of $d$ units, the same calculation gives

\frac{O_j(x+d)}{O_j(x)}=\exp(d\beta_1).

This form is useful when a one-unit change is not the most meaningful comparison. For example, if $x$ is measured in minutes, a 10-minute change corresponds to $d=10$ when the model uses one minute as its unit. If the model instead defines one unit as 10 minutes, the relevant change is $d=1$. Always match $d$ to the units used to fit the model.

Worked Example: A Continuous Covariate Measured in Ten-Minute Units

Suppose a hypothetical logistic random-intercept model estimates $\hat{\beta}_1=0.12$, where $x$ is the number of 10-minute increments of a measured exposure. Interpret a one-unit increase in $x$.

The model implies a conditional odds ratio of $$ \exp(\hat{\beta}_1)=\exp(0.12)\approx 1.127. $$

Thus, within the same cluster and holding other covariates fixed, a 10-minute increase in the exposure is associated with conditional odds of the outcome about 1.127 times as large, or about 12.7% higher. The percentage calculation is $(1.127-1)\times100\%\approx12.7\%$. This describes a change in odds, not a 12.7-percentage-point increase in probability. Because the example is hypothetical and observational, the wording does not assert a causal effect.

What the Random Intercept Changes

Although $b_j$ cancels from the same-cluster odds ratio, it does not disappear from the model. For a fixed value of $x$, changing $b_j$ shifts the conditional log odds and multiplies the conditional odds by $\exp(b_j)$. Consequently, two clusters can have different conditional odds at the same covariate value. The random intercept captures this cluster-to-cluster baseline variation; it does not alter the common slope $\beta_1$ in a random-intercept-only model.

The odds ratio is conditional on a particular cluster effect. A population-marginal comparison instead averages predicted probabilities over the random-effect distribution before comparing outcomes. Because the logit transformation is nonlinear, that marginal comparison generally does not equal $\exp(\beta_1)$. This is one reason not to describe the conditional odds ratio as an average population odds ratio.

Worked Example: A Binary Covariate and Cancellation of the Cluster Effect

Consider a hypothetical model with $\hat{\beta}_0=-1.20$, $\hat{\beta}_1=-0.45$, and a particular cluster effect $b_j=0.80$. Let $x=0$ denote the reference category and $x=1$ the comparison category. Find the within-cluster conditional odds ratio.

At $x=0$, the conditional log odds are $-1.20+0.80=-0.40$, so the conditional odds are $\exp(-0.40)\approx0.6703$. At $x=1$, the conditional log odds are $-1.20-0.45+0.80=-0.85$, so the conditional odds are $\exp(-0.85)\approx0.4274$.

Using the unrounded expressions, the ratio is $$ \frac{\exp(-0.85)}{\exp(-0.40)} =\exp(-0.45)\approx0.6376. $$

The comparison category therefore has about 0.638 times the conditional odds of the reference category within this cluster, or about 36.2% lower odds. The same ratio applies to every cluster under this model: replacing $b_j=0.80$ with another cluster's random intercept changes both category-specific odds by the same multiplicative factor, which cancels in their ratio.

A binary covariate's one-unit change is the move from 0 to 1. For a categorical covariate represented by indicator variables, each coefficient gives the conditional log-odds contrast between its category and the stated reference category, with other model terms held fixed. The reference category and coding therefore need to be clear before interpreting the exponentiated coefficient.

Worked Example: A Change of More Than One Unit

A hypothetical model estimates a conditional log-odds slope of $\hat{\beta}_1=0.35$ per one-point increase in a score. Interpret a two-point increase in the score.

Here $d=2$, so the conditional odds ratio is $$ \exp(d\hat{\beta}_1)=\exp(2\times0.35)=\exp(0.70)\approx2.014. $$

Within the same cluster, and holding other covariates fixed, a two-point increase in the score corresponds to about 2.014 times the conditional odds of the outcome, or about 101.4% higher odds. It would be incorrect to double the one-point odds ratio: odds ratios for successive changes multiply, so the two-point ratio is the square of $\exp(0.35)$, equivalently $\exp(0.70)$.

Assumptions, Diagnostics, and Limits

The algebraic result is exact for the stated model, but an estimated coefficient is useful only to the extent that the model and data support its estimation. The binary response must be coded consistently, the covariate's scale and reference categories must be documented, and the random-intercept structure must be a defensible representation of clustering. The conditional-independence assumption says that, after accounting for the covariates and cluster random intercept, remaining responses are independent. If important within-cluster dependence remains, the fitted model may not represent the data adequately.

Check the model fit and computation. Convergence warnings, unstable coefficient estimates, or sensitivity to starting values can signal that the likelihood is difficult to estimate. Sparse outcome patterns or separation can produce particularly unstable logistic estimates.
Inspect the cluster structure. A random-intercept model assumes a common covariate slope across clusters. If evidence or subject-matter reasoning suggests the slope itself varies by cluster, the random-intercept-only odds ratio may be too restrictive.
Assess the random-effect specification. The assumed distribution and independence of random effects from covariates are modelling assumptions, not consequences of the odds-ratio derivation. A poor specification can affect estimation and the interpretation of model-based comparisons.
Keep the scale in view. An odds ratio does not report a probability difference. Without a baseline probability, it cannot by itself say how many percentage points the probability changes.

A standard logistic regression without a random intercept may be a simpler alternative when observations can reasonably be treated as independent, but ignoring genuine clustering can give misleading uncertainty estimates and omit cluster-level heterogeneity. A marginal approach such as generalized estimating equations targets population-averaged associations rather than this cluster-conditional coefficient. These approaches answer different questions; neither is automatically a substitute for the other.

The constant conditional odds ratio is also a consequence of the model's linear predictor. It assumes that the covariate enters linearly on the logit scale and has no interaction that changes its coefficient across clusters or covariate values. If those terms are included, a single $\exp(\beta_1)$ may no longer summarize every comparison. In particular, random slopes allow the covariate effect to vary by cluster; the random-intercept cancellation alone does not justify a common odds ratio in that model.

A Common Interpretation Error

A frequent mistake is to say, “The random intercept cancels, so the odds ratio is a population-average effect.” The first part is correct only for a contrast within the same cluster under the random-intercept-only model. The conclusion does not follow: the ratio is conditional on the cluster effect, while a population-marginal comparison averages over clusters and generally differs because logistic probabilities are nonlinear functions of the linear predictor.

A second mistake is to translate “odds are 1.127 times as large” into “probability is 12.7 percentage points higher.” Odds and probabilities are related but are not interchangeable. State the odds ratio as an odds comparison unless probabilities have been calculated at specified covariate and random-effect values.

Key takeaway: In a logistic random-intercept model, compare two covariate values within the same cluster. The common random intercept cancels, leaving the conditional odds ratio $\exp(d\beta_1)$ for a $d$-unit change. Its interpretation depends on the covariate units, model coding, and the conditional—not population-marginal—target.

Check Your Understanding

Answer each question using the conditional interpretation, and then compare with the solutions below.

A hypothetical fitted slope is $0.20$ per unit. What is the conditional odds ratio for a one-unit increase, and how should it be stated?
A binary covariate has estimated coefficient $-0.30$. What is the conditional odds ratio for category 1 versus category 0?
If a covariate slope is $0.10$ per unit, what is the conditional odds ratio for a three-unit increase?
Why does the conditional odds ratio not depend on $b_j$, and why does that not make it a population-marginal odds ratio?

Solutions. (1) $\exp(0.20)\approx1.221$. Within the same cluster and holding other covariates fixed, the conditional odds are about 1.221 times as large, or 22.1% higher. (2) $\exp(-0.30)\approx0.7408$; category 1 has about 0.741 times the conditional odds of category 0, or about 25.9% lower odds. (3) $\exp(3\times0.10)=\exp(0.30)\approx1.350$; the three-unit increase corresponds to about 1.350 times the conditional odds. (4) The same $b_j$ appears in both log odds and cancels in their ratio. The resulting comparison remains conditional on a cluster effect; averaging probabilities over the random-effect distribution is a different, generally non-equivalent operation.