Interleaved Quests

1
Hi,

I'm trying to run 2 QUESTS: one on 63% (d'=0.1) and one on 92% (d'=2) wityh a 2AFC task.
(Moving dots).

However, I can't get the 63QUEST on the 63% level. I tried several settings from total
range:
(q63=QuestCreate(0, 5, 0.63, 3.5, 0.01, 0.5, 0.01, 40) to a more specific setting:
(q63=QuestCreate(-12,6, 0.63, 3.5, 001, 0.02, 44)

I tried changing the grain, range and tGuess, but still QUEST isn't give me the right
performance level. I played with the QuestDemo's but the responses of the Simulated
subect are probably very different of 'real' subjects responses. Even if I try to mimic a
subject as close as possible, it's still not performing good in realtime.
want (almost always around 70%).
I just follow normal steps: create QUESTS, Update them with a response, get a new
intensity level for the next trial (either q63 or q92, pseudorandom list).
I update both QUESTS with every trial. So, they inform each other. However, this was giving
me some problems due to range settings (out of range errors).
When I tried to run the QUESTS interleaved, but seperate, still the q63 did a bad job. When
I increased the range size, the estimate was worse.

The goal is to have different subjects (kids & adults) perform on the same levels and
compare them in an fMRI paradigm. For me, the most important issue is the amount of
correct-trials. And the fixed performance level (accuracy)
For the 63% I have 73 trials, for the 92% 43.

So, the question is: what are the basic rules to set interleaved QUESTS in order to run them
optimal for different subjects (with obviously different thresholds).
- Do I need to update them seperately, or let them benefit from each other (the latter
seems logic).
- What do I do with missed trials (no response or too late)
- I just calculate accuracy on each level by just CorrectTrials/TotalQuestTrials per QUEST,
right?
- Can someone give me some advise in setting the tGuess, SD, range & grain? I understand
that 63% on the PMF maybe is hard to estimate, but with more trials it must be possible I
guess. Also, when I give QUEST the freedom to go, within a large range, it still go's fast to
a certain threshold, but doesn't make it harder even if I'm definitely way above pThreshold.
I checked the code, but I really can't find anything unusual. Especially because the q92
seems to work allright.

Thanks,

Martijn
2
dear martijn

it is very difficult to give advice without a running program. we always suggest that queries about software that doesn't seem to work right be accompanied by a minimal length program that exhibits the problem.

what you describe should work fine. quest does not keep any data internally. all the data are stored in your data structures (q63, etc.). so you're in control. Quest is not aware that you're interleaving as it is aware of only one of your data structures at any particular time.

out of range errors indicate that you are requesting too small a range. select your range generously. there is no good reason to restrict it.

Do I need to update them seperately, or let them benefit from each other (the latter
seems logic).
updating them separately is fine.

- What do I do with missed trials (no response or too late)
ignore them. they don't count.

- I just calculate accuracy on each level by just CorrectTrials/TotalQuestTrials per QUEST,
right?
yes

- Can someone give me some advise in setting the tGuess, SD, range & grain?
typically "intensity" is log10(contrast). In that case tGuess is your best guess, in advance, of threshold, e.g. -1. SD expressed how broad your confidence interval is (in log contrast), e.g. 2. (It is very bad to greatly underestimate this, but costs only a few trials if you overestimate it. So guess high.) Range serves a similar purpose as SD, but is used solely to specify what range of intensities to restrict calculations to. There is no cost (except array length and calculation time, usually negligible) to overestimating this. 6 would be safe. Grain too, doesn't matter much unless you pick too big a value. A value of 0.01 would be good.

best

denis

Denis Pelli
Professor of Psychology and Neural Science
New York University

On Jul 20, 2007, at 2:46 AM, tinus74 wrote:

Hi,

I'm trying to run 2 QUESTS: one on 63% (d'=0.1) and one on 92% (d'=2) wityh a 2AFC task.
(Moving dots).

However, I can't get the 63QUEST on the 63% level. I tried several settings from total
range:
(q63=QuestCreate(0, 5, 0.63, 3.5, 0.01, 0.5, 0.01, 40) to a more specific setting:
(q63=QuestCreate(-12,6, 0.63, 3.5, 001, 0.02, 44)

I tried changing the grain, range and tGuess, but still QUEST isn't give me the right
performance level. I played with the QuestDemo's but the responses of the Simulated
subect are probably very different of 'real' subjects responses. Even if I try to mimic a
subject as close as possible, it's still not performing good in realtime.
want (almost always around 70%).
I just follow normal steps: create QUESTS, Update them with a response, get a new
intensity level for the next trial (either q63 or q92, pseudorandom list).
I update both QUESTS with every trial. So, they inform each other. However, this was giving
me some problems due to range settings (out of range errors).
When I tried to run the QUESTS interleaved, but seperate, still the q63 did a bad job. When
I increased the range size, the estimate was worse.

The goal is to have different subjects (kids & adults) perform on the same levels and
compare them in an fMRI paradigm. For me, the most important issue is the amount of
correct-trials. And the fixed performance level (accuracy)
For the 63% I have 73 trials, for the 92% 43.

So, the question is: what are the basic rules to set interleaved QUESTS in order to run them
optimal for different subjects (with obviously different thresholds).
- Do I need to update them seperately, or let them benefit from each other (the latter
seems logic).
- What do I do with missed trials (no response or too late)
- I just calculate accuracy on each level by just CorrectTrials/TotalQuestTrials per QUEST,
right?
- Can someone give me some advise in setting the tGuess, SD, range & grain? I understand
that 63% on the PMF maybe is hard to estimate, but with more trials it must be possible I
guess. Also, when I give QUEST the freedom to go, within a large range, it still go's fast to
a certain threshold, but doesn't make it harder even if I'm definitely way above pThreshold.
I checked the code, but I really can't find anything unusual. Especially because the q92
seems to work allright.

Thanks,

Martijn