-
Notifications
You must be signed in to change notification settings - Fork 12
Options Struct
List of all fields of the option struct in the format:
options.[field] = default Value
options.sigmoidName = 'norm'
This sets the type of sigmoid you fit to your data.
The default value 'norm' fits a cumulative Gaussian to your data.
options.sigmoidName = 'norm';
Another standard alternative is the logistic function.
options.sigmoidName = 'logistic';
For data on a log-scale you may want to fit a log-normal distribution or a Weibull which you invoke with:
options.sigmoidName = 'logn';
options.sigmoidName = 'weibull';
We also included the Gumbel and reversed Gumbel functions for asymmetric psychometric functions. The Gumbel has a longer lower tail the reversed Gumbel a longer upper tail.
options.sigmoidName = 'gumbel';
options.sigmoidName = 'rgumbel';
options.sigmoidHandle
Here you may provide a handle to your own sigmoid which takes two parameters as input and hands back a function value. This should be vectorized or even a formula.
However, this is usually obtained from options.sigmoidName.
options.expType = 'YesNo'
This sets which parameters you want to be free and which you fix, for standard experiment types. Look at Experiment Types for details.
'YesNo' (default) sets all parameters free, which is suitable for a standard yes/no paradigm.
options.expType = 'YesNo';
'2AFC' fixes the lower asymptote to .5 and fits the rest, for 2 alternative forced choice experiments.
options.expType = '2AFC';
'nAFC', fixes the lower asymptote to 1/n and fits the rest. For this type of experiment you MUST also provide options.expN the number of alternatives. As an example with 3 alternatives:
options.expType = 'nAFC';
options.expN = 3;
'equalAsymptote' sets the condition that upper and lower asymptote are equal, but may be variable.
options.expType = 'equalAsymptote';
options.estimateType = 'MAP'
which point estimate you want from the analysis
'MAP' The MAP estimator is the maximum a posteriori computed from the posterior.
options.estimateType = 'MAP';
'mean' The posterior mean. For strict Bayesians the best estimate: the expected value of the Posterior. However, in our paper we show that for psychometric function fitting and the small datasets used in experimental psychology and the neurosciences, and for the our priors the MAP is preferable as default!
options.estimateType = 'mean';
options.stepN = [40,40,20,20,20]
options.mbStepN = [25,20,10,10,20]
This sets the number of grid points on each dimension in the final fitting (stepN) and in the moving of borders (mbStepN). The order is [threshold,width,upper asymptote,lower asymptote,variance scaling]
You may change this if you need more accurate estimates on the sparsely sampled parameters. For example to get an more exact estimate on the lapse rate/upper asymptote plug in:
options.stepN=[40,40,50,20,20];
Now the lapse rate is sampled at 50 places giving you a much more exact and smooth curve for comparisons.
options.confP = .95
The confidence level for the computed confidence intervals. This may be set to any number between 0 and 1 excluding.
For example to get 99% confidence intervals try:
options.confP = .99;
You may specify a vector as well. If you do, the conf_intervals in the result will be a 5x2xN array containing the values for the different confidence levels in the 3rd dimension.
options.confP = [.95,.9,.68,.5];
will return 4 confidence intervals for each parameter corresponding to the 95%, 90%, 68% and 50% credible intervals.
options.CImethod ='percentiles'
This sets how the confidence intervals are computed in getConfRegion.m. Possible variants are:
'project' -> project the confidence region on each axis
'stripes' -> find a threshold with (1-alpha) above it
'percentiles' -> find alpha/2 and 1-alpha/2 percentiles
(alpha = 1-confP)
options.threshPC = .5
Which percent correct correspond to the threshold? Given in percent correct on the unscaled sigmoid (reaching from 0 to 1):
For example to define the threshold as 90% correct try:
options.threshPC = .9;
For details have a look at How to Change the Threshold Percent Correct.
options.priors = getStandardPriors()
This field contains a cell array of function handles, which define the priors for each parameter. If you want to set your priors manually, here is the place for it.
For details on how do change these refer to Priors
options.betaPrior = 10 This sets the strength of the Prior in favour of a binomial observer. Larger values correspond to a stronger prior. We choose this value after a rather large number of simulations. Refer to Priors to learn more about this.
options.useGPU = 0;
this option allows you to use your graphics card to do the computations, which can speed them up considerably, if it works. Set it to the ID of the graphics card to use. Usually this is 1 if you do not
have multiple graphic cards in your computer.
options.nblocks = 35;
options.poolMaxGap = inf;
options.poolMaxLength = inf;
options.poolxTol = 0;
These options set how your data is pooled into blocks. Your data is only pooled if your data Matrix has more than nblocks lines. Then we pool together a maximum of poolMaxLength trials, which are separated by a maximum of poolMaxGap trial of other stimulus levels. If you want you may specify a tolerance in stimulus level poolxTol to pool trials, but by default we only pool trials with exactly the same number of trials.
options.instantPlot = 0
A boolean to control whether you immediately get 2 standard plots of your fit. Turn to 1 to see the effect.
options.instantPlot = 1;
options.borders In this field you may provide your own bounds for the parameters. This should be a 5x2 matrix of start and end of the range for the 5 parameters. (threshold,width,upper asymptote,lower asymptote,variance scale)
For example this would set the borders to
options.borders= [ 1,2 ... % threshold between 1 and 2
.1,5 ... % width between .1 and 5
.05,.05... % a fixed lapse rate of .05
.5,.5 ... % a fixed lower asymptote at .05
exp(-20),.2]; % a maximum on the variance scale of .2
NOTE: By this you artificially exclude all values out of this range. Only exclude parameter values, which are impossible!
options.setBordersType = 0
The method to set the outer borders of the grid. You find it's use in _setBorders.m _
options.setBordersType = 0;
This is legacy by now.
options.maxBorderValue = exp(-10)
Parts of the grid which produce marginal values below this are considered 0 and are excluded from the calculation in moveBorders.m ; it should be a very small value and at least smaller than 1/(max(stepN)).
This for example would exclude fewer values and more conservative movement of the borders:
options.maxBorderValue = exp(-20);
options.moveBorders = 1
toggles the movement of borders by moveBorders.m. Usually this is good to concentrate on the right area in the parameter space.
options.moveBorders = 1;
If you set
options.moveBorders = 0;
your posterior will always use the initial setting for the borders. This is useful if you set options.borders by hand and do not want psignifit to move them after this.
options.dynamicGrid = 0 toggles the useage of a dynamic/adaptive grid. There was hope for a more exact estimate by this, but although the curves look smoother the confidence intervals were not more exact. Thus this is deactivated by default.
options.dynamicGrid = 1;
options.dynamicGrid = 0;
options.GridSetEval = 10000
How many Likelihood evaluations are done per dimension to set the adaptive grid. Should be a relatively large number.
options.GridSetEval = 10000;
Only used with dynamic grid,-> by default not at all.
options.UniformWeight = 0.5000
How many times the average is added to each position while setting the adaptive grid. You may increase this number to get a more equally sampled grid or decrease it to get an even stronger focus of the sampling on the
peak.
When you increase this value very much try to set options.dynamicGrid = 0; which produces an equal step size grid right away.
As an example: This setting will produce a more focused grid which leaves the borders very weakly sampled.
options.UniformWeight = 0.01000;
Only used with dynamic grid,-> by default not at all
options.widthalpha = .05
This changes how the width of a psychometric function is defined width= psi^(-1)(1-alpha) - psi^(-1)(alpha)
where psi^(-1) is the inverse of the sigmoid function. widthalpha must be between 0 and .5 excluding
Thus this would enable the usage of the interval from .1 to .9 as the width for example:
options.widthalpha = .1;
options.logspace: = 0 This is triggered when you fit log-normal or Weibull functions, which are fitted in log-space. This is an internal variable which is used to pass this information to all functions. It is of no interest for a user.