What a Marginal Mean Model Does—and Does Not—Say
A marginal mean model describes the average response at given covariate values in a target population. In repeated-measures and clustered data, the responses from the same person or cluster may be dependent, but that fact does not change what the mean itself represents. Confusion arises when a population-average statement is read as a subject-specific one, or when a model for means is treated as a complete description of the joint outcome distribution.
Let \(Y_{ij}\) denote the response for unit \(j\) in cluster \(i\), and let \(X_{ij}\) denote its covariates. Write the marginal mean as \(\mu_{ij}=E[Y_{ij}\mid X_{ij}]\). A marginal mean model relates this expectation to covariates, often through a link function \(g\), such as \(g(\mu_{ij})=X_{ij}^{T}\beta\). The conditioning here is on the covariates, not on an unobserved subject-specific effect. A conditional model that also conditions on such an effect answers a different question.
Misinterpretation 1: Marginal and Conditional Coefficients Are the Same
A coefficient in a model conditional on a subject-specific effect need not equal the coefficient in the marginal mean model. In general, averaging a nonlinear conditional mean over the distribution of that effect does not preserve the original model’s form or coefficients. Logistic regression provides a direct demonstration: averaging conditional probabilities and then taking their odds is not the same operation as averaging conditional log odds.
Worked Example: A Conditional Logistic Effect and Its Marginal Effect
Suppose a binary outcome has conditional probability \(P(Y=1\mid X=x,U=u)=\operatorname{expit}(\beta x+u)\), where \(\operatorname{expit}(t)=1/(1+e^{-t})\). Let \(\beta=\log 3\), and suppose \(U\) equals \(\log 3\) or \(-\log 3\), each with probability \(1/2\), independently of \(X\). The conditional odds ratio for increasing \(X\) by one unit is \(e^\beta=3\), for either value of \(U\).
At \(x=0\), the two conditional probabilities are \(\operatorname{expit}(\log 3)=3/4\) and \(\operatorname{expit}(-\log 3)=1/4\). Thus the marginal probability is \(P(Y=1\mid X=0)=(3/4+1/4)/2=1/2\), with marginal odds equal to \(1\).
At \(x=1\), the two conditional linear predictors are \(2\log 3=\log 9\) and \(0\). The corresponding probabilities are \(9/10\) and \(1/2\), so \(P(Y=1\mid X=1)=(9/10+1/2)/2=7/10\). The marginal odds are \((7/10)/(3/10)=7/3\). Therefore the marginal odds ratio comparing \(x=1\) with \(x=0\) is \(7/3\), not \(3\).
The example verifies the distinction without changing the conditional effect across values of \(U\): that effect is exactly \(3\). The marginal effect differs because the conditional probabilities are averaged before their odds are formed. It is not generally valid to interpret a conditional coefficient as the coefficient of the marginal mean model.
This is not a claim that the coefficients must always differ. Under particular models, links, covariate distributions, or effect structures, they can coincide. The safe interpretation is to identify what was conditioned on when each coefficient was defined. For a logistic model, odds ratios are also noncollapsible: even in the absence of confounding, a conditional odds ratio and its marginal counterpart can differ.
Misinterpretation 2: A Constant Link-Scale Effect Is a Constant Response-Scale Effect
A linear predictor may change by a constant amount while the corresponding change in the mean depends on the starting mean. For a differentiable inverse link \(g^{-1}\), if \(g(\mu(x))=\alpha+\beta x\), the chain rule gives \(\frac{d\mu(x)}{dx}=\beta (g^{-1})'(\alpha+\beta x)\). Equivalently, when \(g'\) exists and is nonzero, this derivative is \(\beta/g'(\mu(x))\). Unless that derivative is constant over the relevant range, the response-scale slope varies with \(x\).
The result follows by differentiating both sides of \(g(\mu(x))=\alpha+\beta x\): the left-hand derivative is \(g'(\mu(x))\frac{d\mu(x)}{dx}\), and the right-hand derivative is \(\beta\). Solving gives the stated formula wherever \(g'(\mu(x))\ne 0\). For a discrete covariate, the corresponding response-scale contrast is the difference between the two inverse-link predictions; it should be calculated rather than presumed constant. \(\square\)
Worked Example: A Logistic Coefficient Is Not a Constant Risk Difference
Suppose a binary response has \(g(p)=\log\{p/(1-p)\}\), with a one-unit covariate increase changing the log odds by \(\beta=\log 2\). The conditional odds are therefore multiplied by \(2\). Consider two baseline probabilities.
At baseline \(p_0=0.2\), the odds are \(0.2/0.8=1/4\). Doubling them gives \(1/2\), corresponding to \(p_1=(1/2)/(1+1/2)=1/3\). The risk difference is \(1/3-1/5=2/15\), approximately \(0.133\).
At baseline \(p_0=0.5\), the odds are \(1\). Doubling them gives \(2\), corresponding to \(p_1=2/(1+2)=2/3\). The risk difference is \(2/3-1/2=1/6\), approximately \(0.167\).
In both comparisons the odds ratio is \(2\), but the risk differences are not equal. The coefficient is constant on the log-odds scale; it does not describe a constant absolute change in probability.
For a log link, a constant coefficient for a one-unit increase gives a constant ratio of means, since \(e^{\beta}\) multiplies the mean. That still does not imply a constant difference: the difference is \(\mu(e^\beta-1)\), which depends on the starting mean. Always name the scale of an effect—link, ratio, or difference—and calculate the requested response-scale contrast at the covariate values of interest.
Misinterpretation 3: A Mean Model Implies Independence
A mean model specifies expectations, not all features of the joint distribution. In particular, specifying \(E[Y_{i1}\mid X]\) and \(E[Y_{i2}\mid X]\) does not determine \(\operatorname{Cov}(Y_{i1},Y_{i2}\mid X)\). Independence is a much stronger statement about the joint distribution: for two responses, it requires their joint conditional distribution to factor into the product of their conditional marginal distributions. Equal marginal means are compatible with independence and with dependence.
Worked Example: The Same Binary Means, Two Dependence Patterns
Let \(Y_1\) and \(Y_2\) each be binary, with marginal probability \(P(Y_j=1)=1/2\). Under the first joint distribution, each of the four pairs \((0,0),(0,1),(1,0),(1,1)\) has probability \(1/4\). Then \(P(Y_1=1)=P(Y_2=1)=1/2\), and \(E[Y_1Y_2]=1/4\), so \(\operatorname{Cov}(Y_1,Y_2)=1/4-(1/2)(1/2)=0\). In this particular binary case the variables are independent: each joint cell probability is the product \(1/2\cdot1/2=1/4\).
Now instead assign probability \(1/2\) to \((0,0)\), probability \(1/2\) to \((1,1)\), and zero to the other two pairs. The marginal probabilities are still both \(1/2\), so the marginal means have not changed. But \(E[Y_1Y_2]=1/2\), and the covariance is \(1/2-1/4=1/4\). The responses are dependent: for example, \(P(Y_2=1\mid Y_1=1)=1\), not \(1/2\).
Both joint distributions satisfy the same two marginal mean specifications. One is independent and the other is dependent. Therefore, the mean statements alone cannot establish independence or determine the within-pair association.
This distinction matters in clustered data because repeated observations may share unmeasured influences or other sources of dependence. A mean specification can still describe the intended population averages, but claims about uncertainty or joint outcomes require attention to the dependence structure as well. Do not infer independence merely because a model has been written separately for each response’s mean.
Misinterpretation 4: A Fitted Mean Is an Individual Outcome
A fitted mean is an estimate of an expectation, not a prediction that the next observed response must equal that number. For a binary response, a fitted mean of \(0.7\) is a probability; an individual outcome is \(0\) or \(1\). For a count response, a fitted mean may be noninteger even though the realized count is an integer. A mean can summarize the distribution of possible outcomes without being one of them.
Worked Example: Interpreting a Fitted Binary Mean
Suppose a model gives \(E[Y\mid X=x]=0.7\) for a particular covariate value and \(Y\) is binary. The model implies \(P(Y=1\mid X=x)=0.7\) and \(P(Y=0\mid X=x)=0.3\). If the observed outcome for a person with this covariate value is \(Y=0\), that observation is compatible with the model: the model assigns it probability \(0.3\).
The conditional variance is \(0.7(1-0.7)=0.21\), so the conditional standard deviation is \(\sqrt{0.21}\), approximately \(0.458\). The fitted mean \(0.7\) describes the center of the conditional distribution, not an outcome the person is required to have. It is incorrect to say that this person’s outcome is \(0.7\).
A fitted mean can be used as a point prediction when the prediction target is squared-error loss, but it remains a prediction rule rather than an observed response. A different target, such as predicting the most probable binary class, may use a thresholded probability instead. State whether a reported number is a fitted mean, a predicted class, or an observed outcome.
How to Check an Interpretation
When reading a marginal mean model, pause before translating its coefficients into a substantive claim. Identify the response quantity being modeled, the conditioning variables, and the scale of the coefficient. Then separate what the mean model establishes from what would require additional assumptions about the joint distribution or individual outcomes.
Ask whether the quantity is a population-average mean or a mean conditional on subject-specific information.
Determine whether the reported effect is on the link scale, a ratio scale, or an absolute response scale. Use inverse-link predictions to obtain response-scale contrasts.
A list of marginal expectations does not identify within-cluster covariance or establish independence.
Interpret a fitted mean as a distributional average or probability, not as the realized response of every individual.
The most common interpretive error is to move between these levels without saying so: a conditional coefficient becomes a population-average effect, a link-scale change becomes a fixed risk difference, a mean specification becomes an independence assumption, or an expected response becomes an individual’s outcome. Making the conditioning set and scale explicit prevents these shifts. It also clarifies what a model can answer and what requires a separate model or additional information.
Check Your Understanding
For each question, identify the scale or distributional feature needed to give a correct interpretation.
- Why can a conditional logistic coefficient differ from the coefficient describing the marginal odds relationship, even when the conditional effect is the same at every value of a subject-specific effect?
- A logit coefficient is \(0.4\). What does it say about odds for a one-unit covariate increase, and why does it not specify one constant risk difference?
- Two repeated binary responses both have marginal mean \(0.5\). What additional information would be needed to determine whether they are independent?
- A fitted mean for a binary response is \(0.65\), but the observed outcome is \(0\). Is this necessarily a model contradiction? Explain.
- For a log-link model, a one-unit increase in a covariate multiplies the mean by \(1.5\). Does that imply a constant difference in means? Why or why not?