This can be interpreted as having a proportion (p) of branches at a site evolve neutrally or under negative selection, while the remained (1-p) can also evolve under diversifying selection. To test for evidence of episodic selection, we construct the likelihood ratio test between the above model (alternative) and the nested null, where ω+ is forced to be in [0,1].
Simulations (manuscript in preparation) show that MEME is nearly always preferable to FEL because it matches the performance of FEL when there is no lineage-to-lineage variation in dN/dS, and significantly improves upon it when such variation is present. For instance, if one clade in a large tree is evolving with dN/dS = 5, while the rest of the tree (say 90% of branches) evolve with dN/dS = 0.1, FEL is likely to call such site NEGATIVELY selected, whilst MEME will detect that a small proportion of branches is evolving subject to positive selection.
The method is described in complete detail in this PLoS Genetics paper