BackgroundThe impact of cell-composition effects in analysis of DNA methylation data is now widely appreciated. With the availability of a reference data set consisting of DNA methylation measurements on isolated cell types, it is possible to impute cell proportions and adjust for them, but there is increasing interest in methods that adjust for cell composition effects when reference sets are incomplete or unavailable.ResultsIn this article we present a theoretical basis for one such method, showing that the total effect of a phenotype on DNA methylation can be decomposed into orthogonal components, one representing the effect of phenotype on proportions of major cell types, the other representing either subtle effects in composition or global effects at focused loci, and that it is possible to separate these two types of effects in a finite data set. We demonstrate this principle empirically on nine DNA methylation data sets, showing that the first few principal components generally contain a majority of the information on cell-type present in the data, but that later principal components nevertheless contain information about a small number of loci that may represent more focused associations. We also present a new method for determining the number of linear terms to interpret as cell-mixture effects and demonstrate robustness to the choice of this parameter.ConclusionsTaken together, our work demonstrates that reference-free algorithms for cell-mixture adjustment can produce biologically valid results, separating cell-mediated epigenetic effects (i.e. apparent effects arising from differences in cell composition) from those that are not cell mediated, and that in general the interpretation of associations evident from DNA methylation should be carefully considered.Electronic supplementary materialThe online version of this article (doi:10.1186/s12859-015-0527-y) contains supplementary material, which is available to authorized users.