Purchase this article with an account.
Nicolaas Prins; Optimizing the estimation of differences between psychometric functions. Journal of Vision 2014;14(10):1177. doi: https://doi.org/10.1167/14.10.1177.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Many adaptive testing methods have been proposed to optimize the efficiency with which the parameters (e.g., threshold, slope) of an individual psychometric function (PF) can be estimated. However, researchers are rarely interested in the value of the parameters of an individual PF per se. More often, researchers are interested in detecting an effect that some experimental manipulation has on these parameters. In other words, researchers are generally interested in the difference between parameter values measured under different experimental conditions. Here, I modify the Bayesian adaptive psi-method (Kontsevich & Tyler, 1999, Vision Research, 39, 2729-2737) in order to optimize the estimation of differences between the parameter values of a pair of PFs. The method maintains and updates, on a trial-by-trial basis, a posterior distribution defined across four parameters: the mean of the two thresholds, the difference between the thresholds, the mean of the two slopes, and the difference between the two slopes. On each trial, the method selects both which of the two PFs should be tested and which stimulus intensity should be used. The criterion by which the method makes these selections is the expected gain in information regarding the values of the difference parameters only. Results of computer simulations indicate that the optimization of estimation of difference parameters leads to a somewhat different placement strategy compared to optimization of parameter values of individual PFs. However, this placement strategy does not result in more efficient detection of differences between parameter values. Results also indicate that estimation of difference parameters is very robust to incorrect assumptions regarding the value of the lapse rate. That is, even though the estimates of the mean parameter values are biased when the generating lapse rate differs from the lapse rate assumed by the method, the estimates of the difference parameter values are relatively unbiased.
Meeting abstract presented at VSS 2014
This PDF is available to Subscribers Only