Bobaljik & Zocca (2011) argue that ellipsis reveals the existence of (at least) two classes of gender-paired nouns: in the actor/actress class, the grammatically feminine form is specified for conceptual gender, while the unaffixed form is unspecified, exemplifying the classic markedness asymmetry (Jakobson 1932); in the prince/princess class, both forms are specified for conceptual gender. Here we test two theories of this asymmetry: one that encodes markedness in the linguistic representation (e.g. Merchant 2014, Sudo & Spathas 2016, and Saab 2019), and one that traces the asymmetry to differences in the relative frequency of the forms in each pair (Haspelmath 2006). The frequency approach predicts that the size of the asymmetries (as quantified by acceptability judgments) will correlate with the size of the relative frequency ratio for each pair. We test this prediction in two experiments: the first is a curated set of 16 pairs in English, and the second is a test of 58 pairs that nearly exhausts such pairs in English. We use frequencies from COCA (Davies 2008) to test the prediction of the frequency approach. Our results suggest that the relative frequency hypothesis is not an empirically adequate competitor for the explanation of gender asymmetries.