Abstract
Many ensemble studies utilize uniform distributions, often with fewer than 10 items per set. This poses a challenge to real-world generalizability, given that naturally occurring ensemble distributions can contain many items and can vary considerably in shape. To that end, we evaluated the influence of set size and various statistical moments on ensemble processing. Across three experiments, participants viewed an ensemble of isosceles triangles with varying orientations for 250 ms. In Experiment 1, we examined the effect of varying set size (10, 20, 40, and 50 items) and range (60°, 90°, 120°, 150°, and 180°) on reports of average orientation for uniform distributions. Accuracy increased with larger set sizes and smaller ranges, with no interaction between the two distribution parameters. Next we evaluated performance using more naturalistic and differently shaped distributions. Specifically, we generated normally-distributed ensembles, and then changed their shape by manipulating skewness and kurtosis. In Experiment 2, participants had difficulty explicitly discriminating whether two ensembles had the same or different values of skew or kurtosis. In Experiment 3, we examined the effect of manipulating multiple distribution parameters (set size, range, skewness, and kurtosis) on reports of average orientation. We again found no interaction between range and set size, but interestingly, participants had higher accuracy for leptokurtic compared with platykurtic distributions, and for skewed compared with non-skewed (i.e., normal) distributions, despite the lack of explicit sensitivity to these statistical moments in Experiment 2. Kurtosis interacted with both range and set size, but not with skew. Importantly, performance for skewed, kurtotic, and normal distributions was more accurate than performance for uniform distributions in Experiment 1. These results reveal the differential contributions of various distribution parameters on ensemble encoding, and, importantly, highlight the need to use naturalistic statistics over artificial uniform distributions when studying ensemble processing.