Modeling School Absence Counts

Why School Absence Counts Need a Clustered Count Model

A student’s absences are counts, and students in the same classroom may share influences on attendance. For example, they may experience the same classroom routines or disruptions. Treating all student records as independent can therefore understate uncertainty. A generalized linear mixed model (GLMM) can represent both the count outcome and this classroom clustering.

The worked case below is entirely hypothetical. Suppose a school district records each student’s number of absences during an observation window, the number of school days the student was enrolled during that window, lunch-eligibility status, commute distance, and classroom. The aim is to describe associations with absences while accounting for classroom-level variation. These are observational associations, not estimates of causal effects.

Prepare the Data and Specify the Fit

Use one row per student and keep the classroom identifier, absence count, exposure, and predictors on that row. The count should be a nonnegative integer. The exposure must be positive: a student with zero observed enrolled days has no usable rate for this model and needs a separately considered handling decision. Scaling distance as \(D_{ij}=\text{distance in km}/5\) means its coefficient compares students whose commute distances differ by 5 km. The reference group for \(L_{ij}\) is students who are not lunch-eligible.

In R, a corresponding fit with lme4 is specified as follows:

The code line continues with + (1 | classroom), family = poisson(link = "log"), data = student_data. In an actual script, write this as a single function call. The term (1 | classroom) supplies a random intercept for each classroom. The model assumes the rate is proportional to enrolled days: doubling exposure doubles the expected count when the other model terms are held fixed. If that assumption is unsuitable, the offset does not fix the problem.

These choices also define what a coefficient means. At the same exposure and commute distance, \(\exp(\beta_L)\) is the conditional ratio of expected absence counts for lunch-eligible versus non-eligible students in the same classroom. For students differing by 5 km in commute distance, \(\exp(\beta_D)\) is the corresponding conditional count ratio, holding lunch status and classroom fixed. Because the offset scales expected counts by exposure, these are also rate ratios per enrolled day.

From Conditional Contrasts to Expected Counts

A coefficient contrast is not itself an expected count. To obtain an expected count, specify the exposure, covariates, and classroom context. Setting \(b_j=0\) gives a prediction for a classroom at the center of the assumed random-effect distribution on the log scale. Averaging predictions over the random-effect distribution instead gives a population-average expected count for the modeled classroom distribution. These are different targets.

Checking the Fit and Recognizing Its Limits

A successful optimizer message is not evidence that every model assumption is adequate. Begin by checking the data coding: absence counts should be integers, the offset should be finite, and the classroom identifier should reflect the actual grouping. Inspect observed counts and fitted values across classrooms, as well as residual or simulation-based checks for systematic discrepancies. Patterns by fitted count or classroom can reveal lack of fit that a single summary statistic misses.

The Poisson assumption implies that, conditional on the covariates and classroom effect, the count variance equals its conditional mean. A classroom random intercept accounts for some shared variation, but it does not guarantee this conditional variance relationship is correct. If residual checks show more variation than expected, or an excess of zeros relative to model predictions, investigate the source and consider whether another count distribution or model structure is justified. Do not add complexity solely because a diagnostic is imperfect; check whether the alternative fits the data and answers the same scientific question.

The random-intercept specification assumes that the classroom effects are normally distributed and that the modeled covariates are appropriately represented. It also assumes observations are conditionally independent after accounting for those effects and predictors. A singular fit, a variance estimate near zero, convergence warnings, or strong residual patterns deserve investigation. With few classrooms, estimation of the random-effect distribution and uncertainty can be fragile. Repeated observations on a student would also require a dependence structure beyond the classroom intercept; that is not the data structure specified here.

The offset is useful when students have unequal observation time, but it imposes proportionality: expected absences scale directly with enrolled days. The model does not account automatically for missing covariates, informative enrollment duration, or unmeasured confounding. The fixed-effect GLM from GLM is a possible alternative when clustering is absent or not part of the target analysis; ignoring genuine classroom dependence can yield misleading standard errors. A random-intercept GLMM instead estimates conditional associations while representing classroom heterogeneity, but its estimates rely on its distributional and conditional-independence assumptions.