I've been using an interleaved QUEST experiment to measure
discrimination thresholds along a speech sound continuum, and was
hoping someone might be able to answer a couple questions.
1. This is a somewhat naive question, but QUEST typically changes its
threshold estimates according to a log scale. Since my stimulus
continuum is linear, I was wondering if QUEST can also produce linear
estimates?
2. As I have it set up, QUEST seems to be very sensitive to errors,
particularly early errors. For example, subjects may make an error on
their first scored trial (i.e. their first few trials of each
condition are not scored), and the threshold is moved to a large
value. If the subject then performs perfectly, it takes many trials
for the estimate to begin being reducing again. Similarly, after many
correct trials, a single error results in a very large step backward.
I know that the delta parameter is included in order to allow for
this, but raising the value to 10, 15 or even 20% still produces large
effects of errors. I'm hoping someone can explain why errors produce
such large effects, and which parameters effect this (it seems like
the guess SD will be important)?
In my continuum, I have 200 stimuli (0.5% steps). After QUEST returns
its tTest, I first ensure it is above 0, and less than log10(200), and
then round the value 10^tTest to be a value I can produce (i.e. 0.5%
intervals). I am currently using a mean guess of 60% of my continuum
(0.6 * 200 = 120, so mean = log10(120) ~ 2.08), and a SD of 5%
(log10(10) = 1).
Thanks
discrimination thresholds along a speech sound continuum, and was
hoping someone might be able to answer a couple questions.
1. This is a somewhat naive question, but QUEST typically changes its
threshold estimates according to a log scale. Since my stimulus
continuum is linear, I was wondering if QUEST can also produce linear
estimates?
2. As I have it set up, QUEST seems to be very sensitive to errors,
particularly early errors. For example, subjects may make an error on
their first scored trial (i.e. their first few trials of each
condition are not scored), and the threshold is moved to a large
value. If the subject then performs perfectly, it takes many trials
for the estimate to begin being reducing again. Similarly, after many
correct trials, a single error results in a very large step backward.
I know that the delta parameter is included in order to allow for
this, but raising the value to 10, 15 or even 20% still produces large
effects of errors. I'm hoping someone can explain why errors produce
such large effects, and which parameters effect this (it seems like
the guess SD will be important)?
In my continuum, I have 200 stimuli (0.5% steps). After QUEST returns
its tTest, I first ensure it is above 0, and less than log10(200), and
then round the value 10^tTest to be a value I can produce (i.e. 0.5%
intervals). I am currently using a mean guess of 60% of my continuum
(0.6 * 200 = 120, so mean = log10(120) ~ 2.08), and a SD of 5%
(log10(10) = 1).
Thanks