QUEST staircase

1
dear larry

i see david has already responded.

in my lab we simply use plain QUEST. its threshold estimates assume a particular beta, which you specify. over the years, beau and i have looked at dependence of the threshold estimates on beta, and find that efficiency does gradually go down as your estimate of beta is off, but there is very little bias in the threshold estimate, so it's fairly innocuous.

it's heavily used and very reliable. it's been years now since anyone has reported anything even faintly surprising about its behavior.

note that the QUEST package, while included in the psychtoolbox, is pure MATLAB, in no way dependent on the psychtoolbox and is available for download independently:

as David says, it has a simple intuitive interface, making it easy to interleave multiple conditions and apply a family of related algorithms, all in the Bayesian mind set.

best

denis

Denis Pelli
Professor of Psychology and Neural Science

On Oct 30, 2006, at 9:14 AM, David Brainard wrote:


On Oct 30, 2006, at 8:42 AM, Laurence T Maloney wrote:



David, Denis,

Do you have staircase code for a simple staircase in Matlab.
I remembered that there was sometihng like that in
the Toolbox but the web site is down and I can't find it in
my release ....

There is a simple interface to Quest, which is what we use for the most part. I don't have a similarly
abstracted version of m-up, n-down type staircases.

Attached is the Quest folder from the current toolbox, there is an example program QuestDemo that shows the usage.

I'm cc'ing this to Jamie Hillis, who did use m-up n-down in his experiments with me. It is possible he has abstracted
his code in some useful way that I don't know about if that's what you want. Jamie? We would fold such code into
the toolbox if we had it.

DB

P.S. We don't use the Quest threshold estimates directly. Rather, we just take the data and then use the psignifit package
to do a maximum likelihood fit with our favorite psychometric function du jour. Also, we often run interleaved
Quest staircases with different target percent correct (e.g. 65, 76, and 85) to space the trials out a little more
around the steep part of the psychometric function, especially in applications where we care about the slope.

The interface in the implementation makes managing interleaved Quest staircases very easy.

If you don't know about psignifit, it is almost surely the way you want to fit your functions:
http://bootstrap-software.org/psignifit/. They seem to have spent a lot of time and care doing
it well.

<Quest.zip>