Typically, we want to use psignifit to fit data from a psychophysical experiment. These data can come from different experimental designs and can display different shapes. These points are not directly dependent on the actual fitting process.
If nothing else is requested, psignifit will assume that the data come from a 2 alternatives forced choice experiment. That means, that on each trial two stimuli were presented and the observer knows that one and only one of these stimuli was the target stimulus. Although this design has considerable theoretical advantages [Green_and_Swets_1966] there might be practical reasons to collect data in another way. One modification might be in showing more than two stimuli. In this case, you should create your Inference object with the keyword ‘nafc’ set to the number of stimulus alternatives that were presented. Setting the keyword ‘nafc’ to a value of 2 or larger results in a fixed guessing rate of 1/nafc. These designs would all inherit most of the theoretical advantages from the 2 alternatives forced choice experiment and they would all be called forced choice experiments (with 4 alternatives, we would call it a four alternatives forces choice experiment for example).
Another modification of the standard acquisition procedure could be to present only one stimulus per trial. The observers might then have to indicate whether the target stimulus was presented or not (typically called yes-no task). In some discrimination experiments, the observers have to indicate whether a target stimulus was to the left or the right of a mark or whether it was presented for a longer or shorter time than a reference stimulus. In all these experiments we will record which response an observer chose and we will then fit the number of “stimulus present”, “stimulus left”, “stimulus longer” responses (or whatever is suitable in the present context). We will summarize these designs as “yes-no designs” although the term yes-no is typically restricted to detection like tasks. The crucial difference between yes-no designs and forced choice designs for fitting psychometric functions is that yes-no designs allow for arbitrarily “guessing” rates. For instance, in a detection task, the observer might be very conservative and virtually never report the presence of a target for low stimulus intensities. Or the observer might always respond “stimulus left” if the stimulus is presented sufficiently for to the left of a mark. In all these situations, the lower asymptote of the psychometric function will be a free parameter. As in all these situations only one stimulus is presented, we can make the lower asymptote of the psychometric function a free parameter by setting the keyword ‘nafc’ to 1. Note that in this case you also need to specify priors for four parameters instead of the three parameters in an nAFC experiment.
In a detection experiment, we typically have two types of trials. In one case a target stimulus is presented, in the other case, no target stimulus is presented. In terms of signal detection theory, these two cases are called “signal+noise” and “noise only”. The observer responds to these two cases with either “yes, a signal was present” or “no, only noise was presented”. This results in four different outcomes on each trial: hits (the observer correctly reports the presence of a signal), false alarms (the observer reports a signal although “noise only” was presented), misses (the observer reports no signal although a signal was presented), and correct rejections (the observer correctly reports the absence of a signal). These different experimental outcomes are discussed in large detail in standard books on signal detection theory [Green_and_Swets_1966]. What matters with respect to fitting psychometric functions is that depending on the strategy, both, hit rate as well as the correct response rate might change. So which of these two should be fitted with a psychometric function? There are two general objectives that an optimal observer could follow in a yes-no task.
Signal detection theory also offers a number of criterion free discriminability parameters, like area under the ROC curve and the famous d’ index. However, these indices can not generally be assumed to have binomial variance (or anything similar to that). Therefore, psignifit does not attempt to fit such data.
A variety of different parametric shapes for psychometric functions have been used. In probit analysis for instance, the data are essentially fit by a cumulative gaussian; visual contrast detection data have been reported to be well fit by a weibull distribution function. Fitting visual contrast detection with a weibull function is also theoretically appealing because it corresponds to the Quick pooling model ([Graham_1989] p. 165).
Psignifit supports a relatively large number of psychometric function shapes. These are selected using two keywords: ‘sigmoid’ and ‘core’. To understand the meaning of these two keywords, let us take a look at the model that psignifit tries to fit:
Here, is a parameter vector (in forced choice tasks is fixed). The critical term that determines the shape of the psychometric function is . We decompose in two functions, a scalar function and a higherdimensional function , such that
In many cases (but not all), will be a simple linear transformation, while will inject a nonlinearity. We will call the ‘sigmoid’ and the ‘core’.
The figure illustrates how sigmoid and core are related to each other. A sigmoid does not have any parameters. Thus, fitting a psychometric function with only a sigmoid would always result in the same psychometric function. Two such sigmoids are shown in the left column of the figure: The first is a logistic sigmoid and the second is the cumulative distribution function of the standard exponential distribution. In order to have parameters that describe the shape of the psychometric function, we use a core object. The top row of the figure illustrates two core objects: the first is an abCore that can be requested with the keyword ‘ab’. We can see that the output of this core is simply a linear function of . However, the slope and intercept of this linear function depends on the two parameters and . The second plot in the first row illustrates a polyCore, as requested with the keyword ‘poly’. Note that the poly core is a nonlinear function of . Again, the two parameters and determine the precise form of the nonlinear function. In order to illustrate the fact that each core object represents a large number of different functions in , four different combinations of and have been plotted.
The four plots in the lower right of the figure demonstrate how sigmoids and cores can be combined to allow for a large number of possible psychometric function shapes. For instance, the lower right plot is a combination of the Exponential sigmoid and the poly core. The resulting function is the cumulative distribution function of the weibull distribution. The combination of logistic sigmoid and ab core corresponds to the logistic function that was the default setting in earlier versions of psignifit. The advantage of separating sigmoid and core is that we can now use a different core object, to specify that a function should be fitted on different axes (e.g. logarithmic instead of linear) or in a different parameterization. Also note, that the figure only presents two sigmoids and two cores. This results in two different function families for the psychometric function. Psignifit includes 6 different sigmoids and 5 different cores, resulting in 30 different function families.
The following two sections describe the sigmoids and cores in more detail. Then finally, there is a section about common combinations of sigmoids and cores.
Six different sigmoids can be selected. All of them correspond to cumulative distributions functions.
There are also six different cores to be selected. The first three are simply linear transformations of the stimulus intensities. The remaining three cores are nonlinear transformations. Typically, these will be needed to define a weibull function.
As already mentioned above, combinations of ‘sigmoid’ and ‘core’ determine the shape of the nonlinear function . There are some shapes that are particularly interesting in psychophysical applications. This section explains how to obtain these typical shapes.
In this case, we combine the ‘logistic’ sigmoid with one of the linear cores (ab,mw,linear). Depending on the core used, this results in different parameterizations.
where . This allows to be interpreted as the 75% threshold and as the width of the interval in which rises from to . A typical choice for is 0.1.
This parameterization does not allow a psychophysically meaningful interpretation of the parameters.
The cumulative gaussian is obtained by combining the gauss sigmoid with one of the linear cores (ab,mw,linear). The parameterizations are precisely the same as for the logistic function with one exception: The scaling factor z(alpha) for the mw parameterization is , where is the inverse of the the cumulative gaussian.
Also for the cumulative Gumbel sigmoids, the parameterizations are similar to the logistic function. However, the Gumbel distribution is skewed. This implies that the alpha parameter of the ab parameterization can not be interpreted as a 75% threshold. For the mw parameterization this is solved in a different way. The lgumbel + mw function is parameterized as follows:
There are a number of ways to parameterize the Weibull function.
which is implemented using the combination of an exp-sigmoid and a poly-core.
|[Green_and_Swets_1966]||(1, 2) Green, DM and Swets, JA (1966): Signal Detection Theory and Psychophysics. New York: Wiley.|
|[Graham_1989]||Graham, NVS (1989): Visual Pattern Analyzers. New York: Oxford University.|
|[Kuss_et_al_2005]||(1, 2) Kuss, M and Jäkel, F and Wichmann, FA: Bayesian inference for psychometric functions Journal of Vision, 5, 478-492.|