Multilevel Model (Hierarchical Model)

A multilevel model — also called a hierarchical model, or a mixed-effects model — is built for data that comes in nested groups. Real data is rarely a flat list of independent observations. Students sit within schools, patients within hospitals, customers within regions, repeated measurements within the same person, ad impressions within campaigns. Observations inside the same group tend to resemble each other more than observations from different groups, which breaks the assumption of independence that a simple single-level model relies on. A multilevel model takes the grouping seriously, estimating effects at more than one level at once: the overall, population-level pattern and the way individual groups depart from it. It might give every store its own baseline sales level while still estimating an average across all stores, treating the store-level variation as part of the model rather than noise to be ignored.

The signature technique of a multilevel model is partial pooling, and it is what makes the approach powerful. Instead of estimating each group completely on its own, or lumping every group together as if they were identical, a multilevel model does something in between: each group's estimate is pulled toward the overall average by an amount that depends on how much data that group has and how much groups vary. A store with thousands of sales keeps an estimate close to its own data; a store with a handful of sales, whose own numbers are unreliable, borrows strength from the others and sits nearer the overall mean. This shrinkage guards against being fooled by small, noisy groups while still letting well-populated groups speak for themselves — a balance a single-level model cannot strike.

The plainest contrast is with a single-level model that ignores grouping. Such a model either pools everything, pretending all observations are independent and equivalent — which understates uncertainty and can be badly misled by group structure — or it fits every group entirely separately, which throws away the information groups share and overfits small groups. A multilevel model sits between these two extremes through partial pooling, borrowing strength across groups while still letting them differ. So where a single-level model forces a choice between everyone is the same and everyone is unrelated, the multilevel model treats those as the endpoints of a dial and lets the data decide where to sit, which is usually somewhere sensible in the middle.

The trickier contrast is with the older language of fixed and random effects, which multilevel modeling reframes. A fixed-effects approach estimates a separate, independent coefficient for each group and makes no assumption that the groups are related — closer to fitting every group on its own, with no pooling. A random-effects approach treats the group-level deviations as draws from a common distribution, which is exactly the assumption that enables partial pooling. Multilevel models are, in effect, models that include random effects, often alongside fixed ones — hence mixed-effects. The practical difference is that fixed effects give each group its own free estimate with no shrinkage, while the random-effects structure of a multilevel model shares information and shrinks unreliable small-group estimates toward the average. Which to use depends on whether you want the groups treated as isolated or as members of a related family.

Reach for a multilevel model whenever your data is genuinely grouped and you care about both the overall pattern and the groups within it — or when some groups are small and would give unstable estimates on their own. It shines when you have many groups of varying sizes: partial pooling stabilizes the little ones without overriding the big ones. Decide what varies by group: a random intercept lets each group have its own baseline, while random slopes let the effect of a predictor itself differ by group. Check whether the grouping actually matters by examining how much variation lives between groups versus within them. And keep the structure honest to the data's real hierarchy, since inventing levels that do not exist adds complexity without insight. Used this way, a multilevel model respects structure that a flat model pretends away.

The failures come from ignoring structure or misapplying it. Fitting a single-level model to clearly nested data understates uncertainty and can produce a Simpson's-paradox-style reversal when group composition differs — the flat model sees a pooled pattern that the grouped reality contradicts. Fitting every group separately, with no pooling, overfits the small groups and wastes the strength they could borrow. Building elaborate multilevel structures on too few groups gives the model no basis to estimate the between-group variation it depends on. And forgetting to let slopes vary when the effect really does differ by group forces a single relationship onto groups that behave differently. The discipline is to match the model to the data's actual hierarchy, let partial pooling balance the groups, and use the between-group variation the model reveals rather than assuming it away.

A multilevel or hierarchical model represents data nested in groups, sharing information across groups through partial pooling.

Etymology: source.

Search interest for this term over the last five years:

View interest-over-time on Google Trends →

Areas of marketing where multilevel model (hierarchical model) is a core concern: