Multilevel Neighborhood Effects

A compositional effect arises because a neighborhood's residents differ in their individual characteristics, for example being poorer or older, which drives their health regardless of place. A contextual effect is a feature of the place itself, such as deprivation, pollution, or lack of services, that affects residents over and above their personal attributes. A multilevel model separates them by entering individual covariates to absorb composition and area-level variables to capture context; a contextual coefficient is credible only after relevant individual confounders are controlled. Diez Roux emphasizes that failing to model composition inflates apparent place effects.

The variance partition coefficient, equivalent to the intraclass correlation in a linear model, is the proportion of total outcome variation that lies between neighborhoods rather than within them. It also equals the expected correlation in the outcome between two residents of the same neighborhood. Merlo and colleagues treat it as the starting point for any contextual analysis: a sizeable VPC means neighborhoods differ enough that place is worth studying as a level, while a VPC near zero indicates that individual-level variation dominates and contextual modeling will add little.

For binary outcomes fit with logistic regression, the neighborhood random-effect variance is on the log-odds scale and is not directly comparable to the odds ratios reported for individual risk factors. The median odds ratio, introduced by Merlo and colleagues, converts that variance into the median increase in the odds of the outcome a person would face if they moved from a lower- to a higher-risk neighborhood, comparing two randomly selected areas. Because it is expressed as an odds ratio, it lets readers weigh the importance of place against familiar individual risk factors on the same scale.