We use Bayes factors to compare two alternative characterizations of human working memory load in their ability to predict errors in database query-writing tasks. The first measures working memory load by the number of different features each task contains, while the second attempts instead to measure the complexity of the task by giving more weight to features that require more mental time for their correct execution. We reanalyze data from a previously conducted experiment using two logistic regression models with random subject effects nested within an experimental condition factor. The two models have alternative covariates based on the alternative measures of working memory load. We construct prior distributions based on our subjective knowledge gleaned from related experiments, providing details of the elicitation process. We examine sensitivity of our results to the effects of prior misspecification and case deletion. Asymptotic approximations are used throughout to facilitate computations. Finally, we comment on the strengths and limitations of the approach in light of our experience.