Re: Psychophysical measurement technique - PEST? (at)

Subject: Re: Psychophysical measurement technique - PEST?
From:    at <spfHOQAA.ATT.COM>
Date:    Mon, 1 Mar 1993 09:24:00 EST

Peter,
  I've generally used a modified up/down procedure, wherein the direction
of adjustment is changed based on the correctness of the responses.  This
is, as I recall, comparable to the gist of PEST.  Anyway, the algorithm is
based on balancing the number of correct and incorrect responses to get the
desired decision level.

  For two-way forced choice designs, the first and second stimulus sets were
about equally likely to contain the signal, the other being noise.  If the
subject made two correct choices consecutively, for a condition at a given
correlation level (the variable I'm usually measuring), the correlation level
was reduced by 0.1 for the next trial of that condition.  If an error was
made, the level was increased by 0.1 for the next trial.  Reversals in the
direction of correlation level modification occurred at various levels, with
the procedure terminating after seven reversals.  The first reversal was
ignored, and the average of the levels at which the remaining six reversals
occurred was taken as an estimate of the threshold of pattern detection.

Since, at the stopping point, the probability of being correct twice
on a pair of trials equals the probability of not being correct twice
on a pair of trials, the probability of being correct on a single trial
is equal to the square root of 0.50, or about 0.71.  To achieve some other
bogey than 0.71, you would adjust the number of trials at each level so that
the appropriate root comes out to your bogey.  Of course, this will increase
the number of trials you need, so if you can live with 0.71 instead of 0.75
you're better off.

The reason for increasing the level immediately after receiving an incorrect
response, even on the first trial of a pair, is that the second trial offers
no information; there is no way to achieve two correct responses.  In this
way, the number of trials presented to the subject is reduced without
effecting the results.

Hope this helps.  I may actually have some C code lying about which does
this, if you're interested.

Steve Frysinger