Two broad paradigms exist for inferring dN/dS, the ratio of nonsynonymous to synonymous substitution rates, from coding sequences: i) a one-rate approach, where dN/dS is represented with a single parameter, or ii) a two-rate approach, where dN and dS are estimated separately. The performances of these two approaches have been well-studied in the specific context of proper model specification, i.e. when the inference model matches the simulation model. By contrast, the relative performances of one-rate vs. two-rate parameterizations when applied to data generated according to a different mechanism remains unclear. Here, we compare the relative merits of one-rate and two-rate approaches in the specific context of model misspecification by simulating alignments with mutation–selection models rather than with dN/dS-based models. We find that one-rate frameworks generally infer more accurate dN/dS point estimates, even when dS varies among sites. In other words, modeling dS variation may substantially reduce accuracy of dN/dS point estimates. These results appear to depend on the selective constraint operating at a given site. For sites under strong purifying selection (dN/dS 0.3), one-rate and two-rate models show comparable performances. However, one-rate models significantly outperform two-rate models for sites under moderate-to-weak purifying selection. We attribute this distinction to the fact that, for these more quickly evolving sites, a given substitution is more likely to be nonsynonymous than synonymous. The data will therefore be relatively enriched for nonsynonymous changes, and modeling dS contributes excessive noise to dN/dS estimates. We additionally find that high levels of divergence among sequences, rather than the number of sequences in the alignment, are more critical for obtaining precise point estimates.