Individual Differences in Representational Precision Predict Spatial Working Memory Span

Abstract

Perhaps the signature feature of working memory is that it is limited. In the same subjects, we used two different retrieval tasks to independently measure two different limits of spatial memory. Precision was measured by asking participants to localize a missing target item among a field of other targets and distracters. Capacity was measured with a similar task where participants identified, rather than localized, a set of remembered targets from within a larger set of identical items. Across participants, the precision of localization was positively correlated with the number of successfully retrieved items. These data suggest that an individual's representational capacity may ultimately be constrained by their ability to form precise representations of space.

Keywords::

• individual differences
• memory capacity
• memory precision
• spatial representation
• spatial working memory

Acknowledgments

We thank Jessica Kowalsky for her help with data collection, and Justin Halberda, Leon Gmeindl, and Ed Awh for helpful comments on a previous version of the manuscript.

Notes

^1. If instead we had used precision across all loads, we would expect especially strong correlations between memory and precision, but this would be because at least some variance in estimated precision would be a direct result of guesses (i.e. failures of memory).

^2. One concern with correlations for small ns (e.g., 30) is that significant effects may reflect the undue weight of a small number of outlier observations. To address this potential concern, we repeated this analysis with an alternative method known as robust regression (implemented in MATLAB using the robustfit command), which is less sensitive to the influence of outliers. This technique assigns a weight to each point in the correlation and then iteratively adjusts how much weight that observation contributes to the regression line. Initially, all observations are given equal weights and a regression equation is calculated exactly as normal. Then, observations which are poorly predicted by the regression (i.e. outliers) become downweighted and the regression is recalculated until stable weights are found. Thus, this iteration procedure estimates a set of weights that allows each observation to contribute to the slope of the regression line only to the degree that it agrees with the slope suggested by the other observations—intuitively, this allows each point to have a roughly equal “pull” on the slope of the line. Then a final regression is performed using the estimated weights. Using this procedure, we still observed a significant relationship between localization and memory performance: localization performance at Load 1 significantly predicted accuracy in the Memory trials (β = − 0.007 t(28) = − 2.630, p < 0.05), as did localization performance across all loads (β = − 0.012 t(28) = − 4.429, p < 0.05).

^3. Again, these relationships survive when robust regression was used to mitigate the effects of outliers. Span was significantly predicted by Load 3 localization performance (β = − 0.113; t(28) = − 2.3717, p < 0.05) and also by localization performance across all loads (β = − 0.142, t(28) = − 11.831, p < 0.05).