Abstract
I introduce and investigate a user-friendly hierarchical Bayesian method for fitting psychometric functions (PFs) across multiple conditions and subjects simultaneously. The method incorporates the generalized linear model to allow reparameterization of the four parameters of the PF independently across conditions, for example to define main effects and interactions in a factorial design. Posterior distributions across any or all of the four parameters of a PF for individual observers as well as their location and dispersion parameters across observers are derived using Markov Chain Monte Carlo sampling as implemented in JAGS (Plummer, 2003: http://citeseer.ist.psu.edu/plummer03jags.html). Results of simulations indicate that using a hierarchical structure to model PFs across multiple conditions and observers reduces bias in parameter estimates significantly compared to fitting PFs individually. It is further shown that the method converges successfully using priors that are essentially non-informative, even for modestly-sized experiments. This feature makes the method easy to use for those new to Bayesian modeling and perhaps also more acceptable to critics that are concerned by the use of informative prior distributions. The method is further demonstrated by analyzing human data in an experiment that investigated the effect of attention on correspondence matching in an ambiguous long-range motion display. PFs were estimated simultaneously for five observers in each of 24 conditions in a 2×2×2×3 factorial design. Location and slope parameters of the PF were both reparameterized across the 24 conditions in order to code for factor main effects and interactions. The lapse rate parameter was allowed to take on different values across conditions differing in their attentional load but were otherwise constrained. Results show that observers only favored feature-preserving correspondence matches when explicitly asked to consider which token pairs matched featurewise before motion onset, suggesting that the influence of feature matches was realized by a purely conscious and deliberate process.