Group-Level Covariates and Ecological Limitations

What Grouped Records Can Tell Us

A grouped binomial record reports, for each group, a number of trials and a number of successes. Groups might be hospitals, schools, regions, or treatment strata. Such records can support models relating success probabilities to characteristics recorded for those groups. They do not, by themselves, reveal how successes are distributed among individuals within a group.

That distinction matters whenever an individual characteristic varies within a group. If a study records each region’s average exposure and total number of cases, it can model how the regional case probability relates to the regional average. But the same records do not generally show whether individuals with higher exposure were more likely to be cases. The first is a group-level relationship; the second is an individual-level relationship.

Definition: For group \(j\), let \(n_j\) be the number of trials and \(Y_j\) the number of successes. A group-level covariate \(x_j\) is a predictor recorded once for the group, or one that has the same value for every individual in that group. An individual-level covariate \(z_{ij}\) is recorded for individual \(i\) in group \(j\) and may vary within the group.

A grouped logistic model can use group-level predictors, for example

Y_j \mid x_j \sim \operatorname{Binomial}(n_j,p_j), \qquad \operatorname{logit}(p_j)=\beta_0+\beta_1x_j.

Here \(p_j\) is the modeled success probability for a trial in group \(j\), conditional on the predictors in the model. The coefficient \(\beta_1\) describes how the log odds of that group-level probability change with \(x_j\). If \(x_j\) is a region’s average exposure, \(\beta_1\) describes an association across regions between average exposure and the modeled probability. It is not automatically the effect of an individual’s exposure on that individual’s outcome.

Predictors Available at the Group Level

A predictor is available to a grouped model if its value is known for each group and the model’s interpretation matches that level of measurement. Examples include a hospital’s staffing policy, a school’s average class size, or a region’s proportion of residents in a specified category. A summary such as an average can be a valid group-level predictor even when it was calculated from individual measurements. But replacing individual measurements by their group average changes the question being modeled.

The outcome data also impose a limit. A total \(Y_j\) tells us how many successes occurred in group \(j\), but not which individuals succeeded. If the record contains no cross-tabulation by an individual covariate, it contains no direct information about how successes are divided across that covariate’s values. The group size and total alone do not fill in that missing allocation.

Worked Example: A Group-Level Predictor in a Logistic Model

Suppose a fitted model relates a region’s urban population share \(u_j\) to the probability that a resident experiences a specified event:

\operatorname{logit}(p_j)=-2+3u_j.

For a region with \(u_j=0.20\), the linear predictor is \(-2+3(0.20)=-1.4\). Therefore

p_j=\frac{e^{-1.4}}{1+e^{-1.4}}\approx\frac{0.2466}{1.2466}\approx 0.198.

For a region with \(u_j=0.60\), the linear predictor is \(-2+3(0.60)=-0.2\), giving

p_j=\frac{e^{-0.2}}{1+e^{-0.2}}\approx\frac{0.8187}{1.8187}\approx 0.450.

If each region has \(n_j=100\) residents, the corresponding modeled expected success counts are \(100(0.198)=19.8\) and \(100(0.450)=45.0\). These are predictions from a model using regional urban shares. They do not show that an individual resident’s event probability changes according to that resident’s own urban status: the predictor here is a region-level share.

When a Group Total Has a Binomial Distribution

The binomial representation has a specific implication: within group \(j\), the \(n_j\) trial outcomes are modeled as independent Bernoulli variables with the same success probability \(p_j\), conditional on the model’s predictors. In that setting, their sum is binomial. If individual-level covariates make the probabilities different within a group, the sum generally does not have that binomial distribution.

Result: If \(Y_{ij}\), for \(i=1,\ldots,n_j\), are conditionally independent Bernoulli variables with the same conditional success probability \(p_j\), then \(Y_j=\sum_{i=1}^{n_j}Y_{ij}\) has a \(\operatorname{Binomial}(n_j,p_j)\) distribution.

Proof. For \(0\leq s\leq 1\), the probability generating function of a Bernoulli variable with success probability \(p_j\) is \(1-p_j+p_js\). By conditional independence, the generating function of the sum is the product of the individual generating functions:

E[s^{Y_j}\mid\text{predictors}] =\prod_{i=1}^{n_j} E[s^{Y_{ij}}\mid\text{predictors}] =(1-p_j+p_js)^{n_j}.

Expanding by the binomial theorem gives

(1-p_j+p_js)^{n_j} =\sum_{y=0}^{n_j}{n_j\choose y}(1-p_j)^{n_j-y}p_j^y s^y.

The coefficient of \(s^y\) is the probability that the sum equals \(y\), which is exactly the binomial probability mass function. \(\square\)

This result explains why a common group probability is more than a convenient summary. When individuals have different probabilities, the generating function becomes a product of different factors, not the binomial power above. Knowing only the average of those probabilities does not, in general, restore the binomial distribution.

Worked Example: Different Individual Probabilities, Different Count Distribution

Consider two independent individuals in one group. Suppose their success probabilities are \(0.2\) and \(0.8\), perhaps because they have different values of an individual-level predictor. The expected total is \(0.2+0.8=1\), so the average probability is \(0.5\). But the distribution of the total \(Y\) is

P(Y=0)=(0.8)(0.2)=0.16,\qquad P(Y=2)=(0.2)(0.8)=0.16,

and hence

$$ P(Y=1)=1-0.16-0.16=0.68. $$

A \(\operatorname{Binomial}(2,0.5)\) variable has probabilities \(P(Y=0)=0.25\), \(P(Y=1)=0.50\), and \(P(Y=2)=0.25\). Its mean is also \(1\), but its distribution differs from the total just calculated. Thus, replacing the two individual probabilities by their average preserves the expected count in this example, but does not preserve the count distribution. A grouped binomial likelihood with one common probability would not represent these heterogeneous probabilities exactly.

Why Aggregate Counts Cannot Generally Recover Individual Effects

Even if group sizes and totals are known exactly, they do not generally identify how successes align with a covariate that varies within groups. The following construction shows the problem without relying on any modeling assumptions.

Result: Group size, total successes, and the number of individuals with a binary covariate value do not, in general, determine the association between that covariate and success.

Proof by construction. Consider a group of four individuals. Two have covariate value \(z=1\) and two have \(z=0\); the group has two successes. In one possible individual-level arrangement, both \(z=1\) individuals succeed and both \(z=0\) individuals fail. In another arrangement, both \(z=0\) individuals succeed and both \(z=1\) individuals fail. Both arrangements have the same group size, covariate count, and total successes. Yet in the first arrangement the success proportions are \(2/2=1\) for \(z=1\) and \(0/2=0\) for \(z=0\); in the second they are \(0/2=0\) and \(2/2=1\), respectively. The observed aggregate margins are identical while the individual-level associations point in opposite directions. Therefore those margins cannot determine the individual-level association. \(\square\)

Worked Example: The Same Group Record, Opposite Individual Associations

Suppose the only available record for a group is \(n=4\), \(Y=2\), and two individuals with \(z=1\). The remaining two have \(z=0\). The aggregate success proportion is \(2/4=0.5\).

Arrangement A: The two \(z=1\) individuals are the successes. Then the success proportions are \(2/2=1\) for \(z=1\) and \(0/2=0\) for \(z=0\). The individual-level risk difference, defined here as the \(z=1\) success proportion minus the \(z=0\) success proportion, is \(1-0=1\).

Arrangement B: The two \(z=0\) individuals are the successes. Then the proportions are \(0/2=0\) for \(z=1\) and \(2/2=1\) for \(z=0\). The same risk difference is \(0-1=-1\).

Both arrangements produce the same grouped record, including the same count of individuals in each covariate category. The aggregate record therefore does not tell us which individual-level pattern occurred. If the covariate counts were not recorded either, even less information would be available.

Interpretation and a Common Pitfall

A group-level association is not inherently invalid. If the scientific question concerns groups, or if the predictors are genuinely defined at that level, grouped records can be suitable for the question. The limitation is one of interpretation: a relationship between group summaries and group outcomes does not, without additional information and assumptions, establish the corresponding relationship between individual characteristics and individual outcomes.

A common mistake is to treat a group average as if every individual had that value, then interpret the resulting coefficient as an individual effect. For example, a positive association between a region’s average exposure and its case probability does not by itself show that, within a region, more-exposed individuals have higher case probability. The association could reflect other differences between regions, the distribution of exposure within them, or how outcomes are allocated among individuals. The grouped record alone cannot distinguish these explanations.

A useful check is to ask what the records actually retain. If they include only \(n_j\), \(Y_j\), and group covariates, then they support inference about modeled group probabilities conditional on those covariates. An individual-level effect generally requires individual records, informative within-group cross-tabulations, or additional assumptions that are stated and justified. A fitted group-level coefficient cannot supply those missing data by itself.

Key takeaway: Model predictors at the level at which they are observed and interpret coefficients at that same level. Grouped success totals retain the number of successes, not which individuals succeeded; varying individual predictors therefore cannot generally be recovered from totals alone.

Check Your Understanding

Use the distinction between group-level and individual-level information to answer these questions.

A data set gives each hospital’s patient count, number of readmissions, and nurse-to-patient ratio. What level is the ratio measured at, and what kind of association can a grouped model describe?
Why does a group total of successes not identify which individuals with a binary covariate succeeded?
Two independent individuals have success probabilities \(0.1\) and \(0.9\). Calculate the probabilities of zero, one, and two successes. Would a binomial model with probability \(0.5\) give the same distribution?
What assumption about individual trial probabilities supports a binomial distribution for the group total?
Give an example of an individual-level conclusion that would not follow solely from a positive association between a group average and group success probability.