In the standard McDonald and Kreitman test, the estimate of adaptive evolution (α) can be easily biased due to the segregation of slightly deleterious non-synonymous substitutions. Specifically, slightly deleterious mutations contribute more to polymorphism than they do to divergence, and thus, lead to an underestimation of α. Because they tend to segregate at lower frequencies than do neutral mutations, they can be partially controlled by removing low frequency polymorphisms from the analysis, known as the Fay, Wycoff, and Wu or FWW correction (Fay et al. 2001). In this case, α is estimated using the standard MKT equation, but considering only those polymorphic sites (for both neutral and selected classes) with a frequency above the established cutoff.
However, Charlesworth and Eyre-Walker (2008) showed that even removing low-frequency variants, the estimate of α is always downwardly biased and only these estimates are reasonable accurate when the rate of adaptive evolution is high and the distribution of fitness effects of slightly deleterious mutations is leptokurtic (because leptokurtic distributions have a smaller proportion of polymorphisms that are slightly deleterious).
Figure 1. DFE distribution of a mutation. Taken from Eyre-Walker (2007). Probability density of the gamma distribution with a mean of 1 and varying shape parameters. The kurtosis of a distribution is the degree to which it is peaked, and it afects the degree to which α is underestimated because more leptokurtic distributions have a smaller proportion of polymorphisms that are slightly deleterious.