Response time distributions constrain models of visual search

Evan M. Palmer, Jeremy M. Wolfe & Todd S. Horowitz

PURPOSE: Analysis of visual search response time (RT) data usually concentrates on means or medians. Here we analyze the shape of RT distributions in an effort to constrain models of search behavior. METHODS: We analyzed 10 Os’ data from three visual search experiments: feature search (red among green), conjunction search (red vertical among green verticals and red horizontals), and spatial configuration search (2s among 5s). We tested 4 set sizes (3, 6, 12, 18), with 50% target-present (TP) trials and 50% target-absent (TA) trials. Each subject contributed about 500 data points per condition (4000 trials per subject). We normalized the RT distributions by aligning the 25th and 75th percentile points from each distribution and creating RT histograms with equal numbers of bins. This X-normalization technique is less vulnerable to outliers than the standard Z-transform, enabling us to average across subjects and compare the shape of distributions produced by conditions with very different mean RTs. RESULTS: Interestingly, the shape of the X-normalized RT distributions did not vary as a function of set size for any condition except for spatial configuration TA trials. Feature and conjunction tasks produced identical RT distributions. Moreover, TA and TP trials were identical for those tasks. For the spatial configuration search, the TP distribution differed from TA and differed subtly from both the feature and conjunction tasks. CONCLUSIONS: These data falsify models that predict that different set sizes should produce different RT distributions for TP trials. Successful models should be constrained to produce RT distributions of the same shape for feature and conjunction search. Some additional mechanism may be invoked to account for spatial configuration results.
This work was supported by a grant from AFOSR to JMW.