What is a multilevel model and when is it appropriate in program evaluation?

• A. A model that handles time-series data with autocorrelation.
• B. A model that assumes independence of all observations.
• C. A model that handles nested data (e.g., students in schools); appropriate when data are hierarchical and outcomes vary at multiple levels.
• D. A model that ignores clustering effects.

Answer: (C)

Multilevel modeling is designed for data that are nested or hierarchically structured, where outcomes can differ across levels and observations within the same group aren’t independent. In program evaluation, you often have data like students within schools, patients within clinics, or repeated measures over time within individuals. In such cases, the performance or outcomes of units in the same group tend to be more similar to each other than to units in other groups, so standard regression that assumes independence of observations can give biased standard errors and misleading conclusions.

A multilevel model explicitly accounts for this clustering by modeling variation at each level. It uses random effects (for example, random intercepts for each school and possibly random slopes) to capture how outcomes differ across groups and how the effects of predictors may vary between groups. This approach also allows predictors to be included at different levels (student-level and school-level), and can handle different types of outcomes with generalized forms.

Because of these features, multilevel modeling is appropriate whenever data are hierarchical and outcomes vary at multiple levels, which is common in program evaluation. It correctly partitions variance into within-group and between-group components and provides valid inferences about predictors at each level.

The other descriptions don’t fit the situation: time-series models address autocorrelation over time rather than nested data; assuming independence ignores clustering; and ignoring clustering effects is precisely what a multilevel approach is designed to avoid.