A topic in the Open Knowledge Graph — a free, open map of 15,290 topics and the order to learn them in.
Extends regression to hierarchical and nested data structures common in social research (students in schools, individuals in organizations, time points in persons). Covers fixed and random effects, intraclass correlations, cross-level interactions, and applications to longitudinal and clustered data.
Identify nested structures in real datasets, compare single-level and multilevel models, interpret variance components and ICC, test cross-level interactions.
Your prerequisite — linear regression — assumes that observations are independent of one another. That assumption is violated whenever data are nested: students within schools, employees within firms, repeated measurements within individuals, citizens within countries. When observations within a group are more similar to each other than to observations in other groups, you have violated the independence assumption. The consequence is not just a technicality: ordinary regression will underestimate standard errors, making effects appear more statistically significant than they are. Multilevel modeling is the correct tool for this structure.
The key concept is the intraclass correlation coefficient (ICC), which measures how much variance in the outcome is attributable to group membership rather than individual differences. An ICC of 0.15 for student test scores means that 15% of the variance in scores is explained simply by which school a student attends — before any predictors are added. This tells you both that schools matter as units and that the independence assumption is meaningfully violated. Ignoring this structure and running ordinary regression treats each student as if they were from a statistically independent draw; multilevel modeling acknowledges that students in the same school share a context.
The core distinction in multilevel models is between fixed effects and random effects. Fixed effects estimate the average relationship across all groups — the typical slope of, say, family income predicting test scores. Random effects allow that relationship to vary across groups: maybe the income-achievement slope is steeper in some schools than others. A random intercept model lets each group have its own baseline level of the outcome. A random slope model additionally lets each group have its own slope for a predictor. Adding random slopes is not always better — it consumes degrees of freedom and can be poorly identified with small group sizes; the decision should be driven by theory about whether the relationship genuinely varies across contexts.
Cross-level interactions are often the most substantively interesting estimates in multilevel models. These ask: does the effect of a Level 1 predictor (individual-level) depend on a Level 2 characteristic (group-level)? For example, does the effect of a tutoring program vary depending on school resources? This is a cross-level interaction between individual treatment assignment and school-level funding. Correctly modeling this requires the multilevel framework — it cannot be estimated in ordinary regression without ad hoc workarounds. The intuition is that the group context moderates individual-level processes, and multilevel models provide the structural apparatus to test this.
Topics in reflective domains aren't scored by quiz answers. Read, reflect, and mark when you've thought it through.