Marginal Mean Specifications for Count Outcomes

Why Specify a Marginal Mean for Counts?

A count outcome records how many events occurred: for example, symptom episodes, hospital visits, or infections during a follow-up period. In clustered data, several such counts may be recorded for the same person, clinic, or community. A marginal mean model describes the expected count for each observation given its covariates, without requiring the observations in a cluster to be independent.

For count outcomes, a log link is useful because it keeps fitted means positive and makes regression coefficients interpretable as multiplicative changes in expected counts. That interpretation concerns the mean. It does not, by itself, specify a Poisson distribution or determine how responses from the same cluster are associated.

Definition: Let cluster \(i\) contain \(n_i\) count responses \(Y_{ij}\), where \(i=1,\ldots,m\) and \(j=1,\ldots,n_i\). Each \(Y_{ij}\) takes values in \(\{0,1,2,\ldots\}\). Let \(X_i\) denote the covariates used to model that cluster. The marginal mean for response \(j\) is \(\mu_{ij}=E[Y_{ij}\mid X_i]\), assumed finite and strictly positive when a log link is used. A log-linked marginal mean specification is \(\log(\mu_{ij})=\eta_{ij}\), where \(\eta_{ij}\) is a specified linear predictor.

The word marginal indicates that the mean is defined by conditioning on the modeled covariates, rather than on an unobserved cluster-specific effect. The linear predictor can include treatment, time, interactions, and other covariates, as appropriate to the scientific question. The count distribution and the joint distribution of responses within a cluster are separate matters.

Log-Linked Means and Exposure

A basic specification is \(\log(\mu_{ij})=x_{ij}^{T}\beta\), where \(x_{ij}\) is the covariate vector for observation \(j\) in cluster \(i\), and \(\beta\) is a vector of regression parameters. Equivalently,

\mu_{ij}=\exp(x_{ij}^{T}\beta).

Because the exponential is positive, this specification assigns a positive expected count for every finite linear predictor. It does not require the observed count to be positive: zero is a valid observed value even when its expected value is positive.

If observations have different positive exposure durations \(t_{ij}\), a common rate model includes \(\log(t_{ij})\) as an offset:

\log(\mu_{ij})=\log(t_{ij})+x_{ij}^{T}\beta, \qquad \mu_{ij}=t_{ij}\exp(x_{ij}^{T}\beta).

The offset has coefficient fixed at one. Under this model, \(\exp(x_{ij}^{T}\beta)\) is the expected count per unit of exposure, and the expected count over exposure \(t_{ij}\) is that rate multiplied by \(t_{ij}\). This specification assumes proportionality of the mean to exposure; it is not an automatic consequence of observing counts. A zero exposure cannot be inserted into the logarithmic offset, so such observations require a design-specific treatment rather than direct use of this formula.

Worked Example: An Exposure Offset and a Treatment Ratio

Suppose \(Y_{ij}\) is the number of symptom episodes during follow-up, \(t_{ij}\) is follow-up time in years, and \(A_{ij}\) is an indicator equal to one for treatment and zero for control. Let \(a_{ij}\) be age in decades, centered at a reference age. Consider \(\log(\mu_{ij})=\log(t_{ij})-0.4+0.3A_{ij}+0.2a_{ij}\). For an observation with two years of exposure and reference age, calculate the expected counts in both treatment groups and interpret the treatment coefficient.

For control, \(A_{ij}=0\) and \(a_{ij}=0\), so \(\mu_{ij}=2\exp(-0.4)=2(0.6703)\approx1.341\) episodes. For treatment, \(A_{ij}=1\), so \(\mu_{ij}=2\exp(-0.4+0.3)=2\exp(-0.1)\approx1.810\) episodes.

The treatment-to-control ratio at the same age and exposure is \(1.810/1.341\approx1.350\), which agrees with \(\exp(0.3)\approx1.350\). Thus the model specifies an expected count about 35% higher under treatment for these covariates. This is an association encoded by the specified model; it is not, by itself, a claim about a causal treatment effect.

Interpreting Coefficients as Mean Ratios

A log-link coefficient is a difference on the log-mean scale. Exponentiating that difference yields a ratio on the expected-count scale. The following result gives the interpretation precisely for a binary predictor when the other covariates are held fixed.

Theorem: Suppose the marginal mean model is \(\log(\mu_{ij})=\beta_0+\beta_A A_{ij}+z_{ij}^{T}\gamma\), with \(A_{ij}\in\{0,1\}\). For any fixed value of \(z_{ij}\), the ratio of the mean at \(A_{ij}=1\) to the mean at \(A_{ij}=0\) is \(\exp(\beta_A)\).

Proof. By the specified model, the mean at \(A_{ij}=1\) is \(\exp(\beta_0+\beta_A+z_{ij}^{T}\gamma)\), while the mean at \(A_{ij}=0\) is \(\exp(\beta_0+z_{ij}^{T}\gamma)\). Dividing cancels the common terms:

\frac{\mu_{ij}(A_{ij}=1,z_{ij})}{\mu_{ij}(A_{ij}=0,z_{ij})} = \frac{\exp(\beta_0+\beta_A+z_{ij}^{T}\gamma)} {\exp(\beta_0+z_{ij}^{T}\gamma)} =\exp(\beta_A).

Thus \(\beta_A\) is a log mean ratio and \(\exp(\beta_A)\) is the corresponding mean ratio, for the stated covariate comparison. If the linear predictor includes an interaction involving \(A_{ij}\), the treatment contrast also depends on the interacting covariate; the simple constant-ratio interpretation then does not apply unchanged. \(\square\)

Worked Example: A Time-Specific Mean Contrast

Suppose a repeated count outcome has mean model \(\log(\mu_{ij})=1.1+0.25A_{ij}-0.12s_{ij}\), where \(A_{ij}=1\) denotes treatment and \(s_{ij}\) is time in visits from baseline. At visit \(s_{ij}=3\), find the control and treatment means and their ratio.

For control, the linear predictor is \(1.1-0.12(3)=0.74\), so \(\mu_{ij}(0,3)=\exp(0.74)\approx2.096\). For treatment, it is \(1.1+0.25-0.12(3)=0.99\), so \(\mu_{ij}(1,3)=\exp(0.99)\approx2.691\).

The ratio is \(2.691/2.096\approx1.284\). Directly, the difference in linear predictors is \(0.25\), giving \(\exp(0.25)\approx1.284\). The treatment mean is therefore about 28.4% higher at visit 3 under this model. Since there is no treatment-by-time interaction in the stated predictor, the same ratio applies at other modeled times; the two means themselves change with time.

Cluster Totals Require Care with Cluster Size

Sometimes the scientific target is the total count in a cluster, \(S_i=\sum_{j=1}^{n_i}Y_{ij}\). To obtain its conditional expectation from the individual marginal means, the conditioning information must also account for the number of terms in the sum. In particular, the familiar sum of conditional means is immediate when \(n_i\) is fixed, or when it is determined by \(X_i\).

Result: If \(n_i\) is fixed, all \(Y_{ij}\) have finite conditional means given \(X_i\), and \(S_i=\sum_{j=1}^{n_i}Y_{ij}\), then \(E[S_i\mid X_i]=\sum_{j=1}^{n_i}E[Y_{ij}\mid X_i]=\sum_{j=1}^{n_i}\mu_{ij}\). The same equality holds conditional on \((X_i,n_i)\) when cluster size is random and the responses have finite conditional means given \((X_i,n_i)\).

Proof. For fixed \(n_i\), the sum defining \(S_i\) has a finite, fixed number of terms. Conditional linearity of expectation therefore gives

E[S_i\mid X_i] = E\left[\sum_{j=1}^{n_i}Y_{ij}\,\middle|\,X_i\right] = \sum_{j=1}^{n_i}E[Y_{ij}\mid X_i] = \sum_{j=1}^{n_i}\mu_{ij}.

If \(n_i\) is random, apply the same finite-sum argument after conditioning on both \(X_i\) and \(n_i\): \(E[S_i\mid X_i,n_i]=\sum_{j=1}^{n_i}E[Y_{ij}\mid X_i,n_i]\). If \(n_i\) is determined by \(X_i\), conditioning on \(X_i\) already fixes the number of terms, and the fixed-size equality applies. Without that condition, the random upper limit cannot simply be carried outside a conditional expectation given \(X_i\); the conditional means given \(X_i\) alone do not in general supply the displayed formula for the random total. \(\square\)

Worked Example: Summing Expected Counts with Fixed Cluster Size

A study records a count at each of three scheduled visits for each participant. For one participant, suppose the fitted marginal means at the visits are \(1.2\), \(1.5\), and \(1.8\). Let \(S_i\) be the sum of the three visit counts. Find the expected total and state what can be concluded about dependence.

The cluster size is fixed at three, so conditional linearity gives \(E[S_i\mid X_i]=1.2+1.5+1.8=4.5\). No independence assumption is needed for this calculation. The three counts may be dependent; their dependence affects the distribution and variance of their sum, but not this sum of means.

The result is an expectation, not a guarantee that any participant will have exactly 4.5 total events. Since \(S_i\) is integer-valued, the noninteger value summarizes an average over the modeled population or repeated realizations.

A Mean Model Is Not a Dependence Model

A marginal mean specification gives the conditional expectation of each response. It does not determine the probability of each possible count, the variance of a count, or the joint probabilities of counts within a cluster. In particular, choosing a log link does not imply that each response is Poisson, and choosing Poisson marginal distributions would still not uniquely determine their joint distribution.

Worked Example: Same Count Means, Different Dependence

Consider a cluster with two count responses, each having a Poisson distribution with mean 2. Compare two possible joint models: in the first, \(Y_{i1}\) and \(Y_{i2}\) are independent; in the second, \(Y_{i1}=Y_{i2}=Z_i\), where \(Z_i\sim\operatorname{Poisson}(2)\).

In both models, each response has marginal mean 2. Thus each has the same log-linked marginal mean, \(\log(\mu_{ij})=\log(2)\). In the independent model, \(\operatorname{Cov}(Y_{i1},Y_{i2})=0\). In the second model, \(\operatorname{Cov}(Y_{i1},Y_{i2})=\operatorname{Cov}(Z_i,Z_i)=\operatorname{Var}(Z_i)=2\), using the Poisson variance.

The mean specification is identical, but the within-cluster dependence is not. Therefore a coefficient or fitted mean from a marginal count model cannot, on its own, be used to infer that responses are independent or to identify their dependence structure.

Common Mistakes and Practical Checks

Reading a coefficient as an additive count change. With a log link, \(\beta_A\) is an additive change in log mean. The corresponding mean ratio is \(\exp(\beta_A)\), not \(\beta_A\) additional events.
Treating the mean model as a Poisson assumption. A log-linked mean is a specification for \(E[Y_{ij}\mid X_i]\). It does not establish a Poisson response distribution or a particular variance.
Calling a mean ratio a probability ratio. For counts, \(\exp(\beta_A)\) compares expected counts. It is not automatically a ratio of probabilities of any event, such as \(P(Y_{ij}>0\mid X_i)\).
Summing means with an uncontrolled random cluster size. The identity for a fixed finite sum applies conditional on the information that fixes the number of terms. If cluster size is random and not determined by the modeled covariates, condition on it as well before using that identity.
Ignoring the offset's role. An exposure offset changes the modeled mean proportionally to exposure. It is not an estimated covariate effect, and the proportionality assumption should be scientifically appropriate.

Key takeaway: A log-linked marginal count model specifies positive conditional expected counts and turns linear predictor contrasts into multiplicative mean contrasts. Keep that mean statement separate from assumptions about count distributions, exposure behavior, and within-cluster dependence.

Check Your Understanding

Answer each question using the distinction between a mean specification and a dependence assumption.

In a log-linked mean model, a binary treatment coefficient is \(0.4\). What is the treatment-to-control ratio of expected counts at fixed values of the other covariates?
A count model includes \(\log(t_{ij})\) as an offset. What proportionality assumption does this make about the expected count as exposure changes?
Does specifying \(\log(\mu_{ij})=x_{ij}^{T}\beta\) imply that \(Y_{ij}\) is Poisson? Explain.
For three fixed visits with marginal means \(0.8\), \(1.1\), and \(1.4\), find the expected total count. Is independence needed?
If cluster size is random and not determined by \(X_i\), what conditioning information should be included before applying the sum-of-conditional-means identity?