Psychometric Basics
A function that relates stimulus level to the probability of a particular response (e.g., correct).
What is the psychometric function?
A non-adaptive approach where stimulus levels are chosen in advance (not based on responses).
What is a fixed (preselected) stimulus design / method of constant stimuli-style approach?
In a basic up–down staircase, a correct response typically makes the next trial _____.
What is “harder” (lower intensity / smaller difference), depending on the stimulus axis?
What does PEST stand for?
What is Parameter Estimation by Sequential Testing?
These methods choose the next stimulus level using an updated estimate based on ____ data so far.
What is all (the accumulated) data?
Threshold is the stimulus level corresponding to a chosen ____ on the psychometric function.
What is a performance criterion (e.g., 75% correct)?
One major advantage of adaptive methods when you only care about threshold.
What is higher efficiency / fewer trials needed near threshold?
A “reversal” is when the staircase changes direction because the response changes from correct→incorrect or incorrect→correct. True or False?
What is True?
In PEST, P is the target ____ (e.g., 0.75) the procedure tries to achieve.
What is probability (target performance level)?
A major drawback of max-likelihood / QUEST methods is that they rely on assuming a psychometric function ____.
What is its shape/form (and parameter assumptions)?
In a 2-alternative forced-choice (2AFC) task, “midway between chance and perfect” corresponds to what percent correct?
What is 75% correct?
(Chance in 2AFC is 50%; midway between 50 and 100 is 75.)
Name one reason you might still want a classical full-range design (not adaptive).
What is “to estimate slope / fit the full psychometric function / check model fit / detect nonstationarity across the range”?
Why does a double staircase reduce the chance the participant “games” the procedure?
What is “interleaving two tracks makes the next level less predictable, reducing strategy/pattern tracking”?
What is the practical purpose of running blocks of trials at one level (instead of updating every single trial)?
What is “to reduce overreaction to random single-trial noise and stabilize decisions near threshold”?
In QUEST-style logic, adding a “prior” mainly influences what the algorithm does early on: stimulus placement, final estimate, or both?
What is stimulus placement (primarily early), more than the final estimate?
Name the two additional parameters (besides threshold/location) that commonly matter when modeling psychometric data and can bias threshold if ignored.
What are slope and lapse rate (and/or guess rate for forced-choice tasks)?
A participant’s sensitivity is stable, but their response criterion drifts over time in a yes/no task. Which approach (classical vs adaptive) is more likely to hide that problem, and why?
What is adaptive, because it samples narrowly near threshold and can miss drift/criterion shifts that a full-range design might reveal?
You run a 1-up/1-down staircase in a yes/no task. What performance level does it tend to target, and what’s the catch?
What is ~50% (in idealized conditions), but the catch is response bias/criterion issues in yes/no tasks?
In PEST, step size usually shrinks after a reversal. What is the key benefit of shrinking step size as you approach threshold?
What is “finer resolution and reduced oscillation around threshold, improving precision”?
Give one concrete failure mode that would make QUEST give a biased threshold estimate (not just “noise”).
What is learning/fatigue drift, high lapse rate, wrong function form, wrong slope, or nonstationary behavior across trials?
Two studies report thresholds from the same task, but one uses 75% correct and the other uses 70% correct. What additional information would let you convert one threshold to the other (at least approximately)?
What is “the psychometric function shape/slope (and any guess/lapse assumptions)”—so you can map 70% and 75% points on the same curve?
Design an adaptive strategy choice: You have 60 trials, unknown slope, and strong suspicion of fatigue drift. Which method (staircase, PEST, QUEST) do you pick and what one tweak do you add to reduce drift bias?
What is “PEST or a conservative staircase with decreasing step size, plus a drift-control tweak like shorter blocks, breaks, counterbalancing, or inserting catch trials / monitoring lapse”?
Explain why “average the last N reversals” can be biased if your step size does not decrease over time, and propose a fix.
What is “fixed step sizes can overshoot and bias reversal averages; fix by reducing step size over time or fitting a model to the collected trials”?
You set P = 0.75 and N = 2 (tiny blocks). Predict what happens to stability and why this could inflate reversals/step changes.
What is “it becomes jumpy/unstable because each block is too small, so random noise triggers frequent reversals and step changes”?
QUEST assumes a psychometric function form and often a slope. If the true slope is much shallower than assumed, what systematic effect does that tend to have on threshold estimation or stimulus placement?
What is “the algorithm may place stimuli too close to the assumed threshold too early and misestimate threshold (often biasing it), because the data look less informative than expected under the assumed steep slope”?