Abstract
Psychophysical experiments are the standard approach for quantifying functional abilities and properties of the visual system, and for linking observed behaviour to the perception. In most psychological experiments, human observers or animals respond to multiple trials that are presented in a sequence, and it is commonly assumed that these responses are independent of responses on previous trials, as well as of stimuli presented on previous trials. There are, however, multiple reasons to question the ubiquitous assumption of "independent and identically distributed trials". In addition, pronounced inter-trial dependencies have been reported in mice during a visual task (Busse et al, 2011, J Neurosci). These observations raise two central questions: First, how strong are sequential dependencies in psychophysical experiments? Second, what are statistical methods that would allow us to detect these dependencies, and to deal with them appropriately? Here, we present a statistical modelling framework that allows for quantification of sequential dependencies, and for investigating their effect on psychometric functions estimated from data. In particular, we extend a commonly used model for psychometric functions by including additional regressors that model the effect of experimental history on observed responses. We apply our model to both simulated data and multiple real psychophysical data-sets of experienced human observers. We show that our model successfully detects trial by trial dependencies if they are present and allows for a statistical assessment of the significance of these dependencies. We find that, in our data-sets, the majority of human observers displays statistically significant history dependencies. In addition, we show how accounting for history dependencies can lead to changes in the estimated slopes of psychometric functions. As sequential dependencies are presumably stronger in inexperienced observers or behaving animals, we expect that methods like the ones presented here will become important tools for modelling psychophysical data.
Meeting abstract presented at VSS 2012