Predicting recall of words and lists

Abstract

For more than a half-century, lists of words have served as the memoranda of choice in studies of human memory. To better understand why some words and lists are easier to recall than others, we estimated multivariate models of word and list recall. In each of the 23 sessions, subjects (N = 98) studied and recalled the same set of 576 words, presented in 24 study-test lists. Fitting a statistical model to these data revealed positive effects of animacy, contextual diversity, valence, arousal, concreteness, and semantic structure on recall of individual words. We next asked whether a similar approach would allow us to account for list-level variability in recall performance. Here we hypothesized that semantically coherent lists would be most memorable. Consistent with this prediction, we found that semantic similarity, weighted by temporal distance, was a strong positive predictor of list-level recall. Additionally, we found significant effects of average contextual diversity, valence, animacy, and concreteness on list-level recall. Our findings extend previous models of item-level recall and show that aggregate measures of item recallability also account for variability in list-level performance. (PsycInfo Database Record (c) 2021 APA, all rights reserved).

Figure 1.. Variability in free recall of words

Average recall probability of each word. Average recall probabilities are sorted from lowest to highest and plotted along with the standard error confidence band around the average values. Blue and gray markers indicate animate and inanimate words respectively.

Figure 2.. Distributions of β values for each predictor variable in the word and list-recall models when fit to each subject separately. A. Word-Recall Model. B. List-Recall Model.

Each circle on the plots denotes the normalized logistic regression coefficient for a single subject, with filled circles indicating coefficients that met an FDR-correct p < 0.05 significance criterion.