This paper proposes a decomposition of the composition effect, i.e. the part of the observed between-group difference in the distribution of some economic outcome that can be explained by differences in the distribution of covariates. Our decomposition contains three types of components: (i) the "direct contributions" of each covariate due to between-group differences in the respective marginal distributions, (ii) several “two way” and "higher order" interaction effects due to the interplay between two or more covariates' marginal distributions, and (iii) a "dependence effect" accounting for between-group differences in dependence patterns among the covariates. Our methods can be used to decompose differences in arbitrary distributional features, like quantiles or inequality measures, and allows for general nonlinear relationships between the outcome and the covariates. It can easily be implemented in practice using standard econometric techniques. An application to wage data from the US illustrates the empirical relevance of the decomposition’s components.