QUEST doubt about log scale

1
Greetings to everyone.

This is not about the PTB per se, but more about the QUEST procedure for threshold estimation, since a lot of people from this forum use it too, I figured someone might be able to kindly help.

I already went through all the QUEST threads in the forum and I realized this question has already been asked by others, indirectly or directly, but remains Unanswered.

The question is: Can you use QUEST with linear (instead of log) units on the "intensity/contrast" axis and still get meaningful threshold estimates? This means feeding tguess "A" instead of "log10(A)", using the output of "B=QuestMean" directly as stimulus intensity, instead of "10^B".

I have an implementation using linear units for a size aftereffect experiment. In my experiment, intensity codes size difference between two circles in pixel units. Since effect is bidirectional dependent upon preceding stimulus, intensity can vary from negative to positive --i.e. positive means left circle is bigger, negative means right circle is bigger. This results from always adding intensity/2 to the left circle and subtracting intensity/2 right circle. Hence at some point a log10(negative number) must be fed in to QUEST, which is not possible.

The results I got are consistent with the manipulable direction of the effect (l or r) but it seems that the QUEST estimate settles too fast and its more less fixed after that, no matter what. Below I show a typical run, in this run the expected effect is for right to be perceived as bigger, hence intensity should be negative. The response column means 0=left was bigger 1=right was bigger. pthreshold=0.5

Response Intensity
1 -3.42035033843986e-16
0 -11.8372157647458
0 -5.65499019700693
1 -2.71602947163177
0 -4.70142489031052
1 -4.12755712220092
0 -4.68079211146133
0 -4.58322338730821
0 -4.49054347387657
1 -3.98969822979645
0 -4.04276513221456
0 -3.99916687223084
0 -3.99784780941695
0 -3.99687863028556
0 -3.99549447665043
0 -3.99350036654845
0 -3.99063106329750
1 -3.98651037160287
0 -3.99943525326438
1 -3.99918413574450
0 -3.99996587808286
0 -3.99995070005758
1 -3.99992877095315
0 -3.99999700010579
1 -3.99999566598954


One of the other persons that asked the same question had a timing perception experiment, where the threshold could be before or after some stimulus, hence necessarily using negative and positives thresholds.

Thanks for reading, hope this was not too long but I just wanted to make it really clear.

Luis Z.
UC Berkeley
Undergraduate
2
dear luis

> Can you use QUEST with linear (instead of log) units on the "intensity/contrast" axis and still get meaningful threshold estimates?

the documentation in QuestDemo.m says, "You can think of QUEST as a Bayesian toolbox for testing observers and estimating their thresholds. ... It is important to realize that Quest is merely a friendly adviser, cataloging your data in your q structure, and making statistical analyses of it, but never giving you orders."

the point is that Quest doesn't know any psychophysics or visual perception. it merely catalogs your data and makes bayesian inferences from the assumptions and data that you provide to it. when you call QuestCreate you assert that the observer is well-modeled by a Weibull function with a specified steepness and unknown threshold. most users of Quest use an intensity scale corresponding to log contrast. it is well established that the human psychometric function for a wide range of tasks has a fixed steepness as a function of log contrast, independent of threshold. as experimenter, you are free to pick any intensity scale you like and assert that the psychometric function has a fixed steepness as a function of that "intensity". if you're wrong, i.e. the psychometric steepness is actually intensity dependent, contrary to what you told Quest, then Quest's bayesian reasoning will be ill-founded and i'm not sure what could be said about the results. in practice, i'd expect it to still succeed in finding threshold, which might be enough to satisfy your needs.

Andrew "Beau" Watson and I, independently, have each done a fair bit of testing of the effect of assuming the wrong beta (steepness). As far as I know, the results of this testing are unpublished. We find that assuming the wrong beta has hardly any effect on the mean threshold estimate but does increase the s.d. thus, to achieve a needed precision in the fewest trials you should provide an accurate estimate of beta, but it's ok to guess wildly (i.e. be wrong by a factor of 2), running more trials or making repeated threshold estimates to get the s.e. down to what you need. that's not quite the question you asked, but it suggests that if the intensity scale you provide for your experiment does in fact turn out to have a psychometric function whose steepness depends on intensity, Quest will still home in on threshold using the beta you specified, which will probably be different from the actual beta and the observer's particular threshold.

good luck

best

denis

Denis Pelli
Professor of Psychology and Neural Science
New York University
http://psych.nyu.edu/pelli/
blog http://denispelli.com/
@denispelli

On Mar 24, 2011, at 11:05 PM, Luis Zertuche wrote:

> Greetings to everyone.
>
> This is not about the PTB per se, but more about the QUEST procedure for threshold estimation, since a lot of people from this forum use it too, I figured someone might be able to kindly help.
>
> I already went through all the QUEST threads in the forum and I realized this question has already been asked by others, indirectly or directly, but remains Unanswered.
>
> The question is: Can you use QUEST with linear (instead of log) units on the "intensity/contrast" axis and still get meaningful threshold estimates? This means feeding tguess "A" instead of "log10(A)", using the output of "B=QuestMean" directly as stimulus intensity, instead of "10^B".
>
> I have an implementation using linear units for a size aftereffect experiment. In my experiment, intensity codes size difference between two circles in pixel units. Since effect is bidirectional dependent upon preceding stimulus, intensity can vary from negative to positive --i.e. positive means left circle is bigger, negative means right circle is bigger. This results from always adding intensity/2 to the left circle and subtracting intensity/2 right circle. Hence at some point a log10(negative number) must be fed in to QUEST, which is not possible.
>
> The results I got are consistent with the manipulable direction of the effect (l or r) but it seems that the QUEST estimate settles too fast and its more less fixed after that, no matter what. Below I show a typical run, in this run the expected effect is for right to be perceived as bigger, hence intensity should be negative. The response column means 0=left was bigger 1=right was bigger. pthreshold=0.5
>
> Response Intensity
> 1 -3.42035033843986e-16
> 0 -11.8372157647458
> 0 -5.65499019700693
> 1 -2.71602947163177
> 0 -4.70142489031052
> 1 -4.12755712220092
> 0 -4.68079211146133
> 0 -4.58322338730821
> 0 -4.49054347387657
> 1 -3.98969822979645
> 0 -4.04276513221456
> 0 -3.99916687223084
> 0 -3.99784780941695
> 0 -3.99687863028556
> 0 -3.99549447665043
> 0 -3.99350036654845
> 0 -3.99063106329750
> 1 -3.98651037160287
> 0 -3.99943525326438
> 1 -3.99918413574450
> 0 -3.99996587808286
> 0 -3.99995070005758
> 1 -3.99992877095315
> 0 -3.99999700010579
> 1 -3.99999566598954
>
> One of the other persons that asked the same question had a timing perception experiment, where the threshold could be before or after some stimulus, hence necessarily using negative and positives thresholds.
>
> Thanks for reading, hope this was not too long but I just wanted to make it really clear.
>
> Luis Z.
> UC Berkeley
> Undergraduate
>
>