In metamemory research, people are often asked to judge the probability of recalling a studied item in a subsequent test. These judgments of learning (JOL) are assumed to be inferred from probabilistic cues that are to some extent predictive of actual recall performance. Based on Brunswik’s lens model, the matching between the cues used to make a judgment and their respective validities can be quantified by the matching parameter G, which is computed for each person separately by correlating the predicted values from two linear regressions (one for the judgments, one for the actual recall performance) using the cues as predictors. However, this computation of person-wise regressions does not account for the clustered structure of the data. In addition, Undorf & Bröder (2019) found that there is substantial systematic covariation between JOLs and recall which is not accounted for by the matching parameter G. Both issues could potentially lead to inaccurate estimations of G. Therefore, a hierarchical extension is proposed and compared to the conventional approach in a simulation study and in the reanalysis of five JOL datasets. The simulation study revealed that the G parameter of the hierarchical approach more accurately recovered the true matching than the participant-wise analysis, especially when item-effects or omitted cues affected the data. In the application, however, the differences in G were mainly due to hierarchical shrinkage, suggesting that the classical lens model application causes overfitting and, therefore, underestimation of matching. The findings indicate that people’s metamemory ability is higher than previously assumed.