DNA methylation is amongst the best studied of epigenetic mechanisms impacting gene expression. While much attention has been paid to the proper normalization of bioinformatical data produced by DNA methylation assays, linear models remain the current standard for analyzing post-processed methylation data, for the ease they afford for both statistical inference and scientific interpretation. We present a new, general statistical algorithm for the model-free estimation of the differential methylation of DNA CpG sites, complete with straightforward and interpretable statistical inference for such estimates. The new approach leverages variable importance measures, a class of parameters arising in causal inference, in a manner that facilitates their use in obtaining targeted estimates of the importance of each CpG site. The proposed procedure is computationally efficient and self-contained, incorporating techniques to isolate a subset of candidate CpG sites based on cursory evidence of differential methylation and providing a multiple testing correction that appropriately controls the False Discovery Rate in such multi-stage analysis settings. The effectiveness of the new methodology is demonstrated by way of data analysis with real DNA methylation data, and a recently developed R package (
methyvim; available via Bioconductor) that provides support for data analysis with this methodology is introduced.