The impact of centering first-level predictors on individual and contextual effects in multilevel data analysis

Multilevel data analysis is a powerful analytical tool. Properly applying the models and correctly interpreting the findings are two interrelated general issues in using multilevel modeling (MLM). There are two specific issues when using MLM: (a) separating the individual-level effects of a predictor variable from its contextual effects and (b) centering first-level predictor variables. This can have major implications for interpreting the results at higher levels, and its impact on second-level interpretation is not always apparent.

The major purposes of this article are to show how to separate organizational-level effects from individual-level effects and to show how first-level centering decisions affect the interpretation of second-level coefficients.

The hierarchical linear models (HLM) are used to analyze a hypothetical data set with 385 patients nested within 10 hospitals, using uncentered, group-mean-centered, and grand-mean-centered versions of the predictor variable.

Uncentered and grand-mean-centered models are equivalent, but group-mean-centered models are not equivalent to the other two. For the grand-mean-centered and uncentered models, second-level coefficients provide correct estimates of the individual effect and the contextual effect when the contextual predictor variable is included in the second-level model. The group-mean-centered model leads to a second-level coefficient where individual-level effects are confounded with contextual-level effects.

There is no single best answer to the question of whether to use group-mean centering or grand-mean centering. The theory and specific questions to be answered should be the researcher's guide to selecting which centering approach to use. Understanding the implications of first-level centering is essential to interpreting second-level coefficients correctly.