Hello,
I want to use a paradigm where subjects have to detect near-threshold targets in flickering noise at random moments, but without any classical trial structure, meaning that there is a continuous sequence of stimuli. The continuous structure with no trials has the consequence that I cannot compute sensitivity (d’) in classical signal detection framework: There is no number of stimulus absent trials and thus no probability to respond yes when there is no target (P(yes|absent)). In other words, although I have number of false alarms, I cannot translate this to false alarm rate, and am dependent on hit rate alone.
Does anyone know a potential measure alternative to d’, which could help me in this situation to quantify the participant’s perceptual sensitivity? I would appreciate any hint or idea.
Best regards,
Nicolai
Hi Fernando,
That's an excellent reference for my purpose, thank you very much!
It seems to me that subdivision into time windows is absolutely necessary. This is because the probability of detection PD and false alarm PFA are necessarily conditional on a time constraint.
That is: PD = probability of registering that target is present given (1) that one target is present and (2) given a fixed dwell time T on the scene. PFA = probability of registering that target is present given (1) that no target is present and (2) given a fixed dwell time T on the scene. The time constraint is usually not stated, but in practice it is always an implicit condition of PD and PFA.
T is the "look" or "decision" period, which must be fixed in order to estimate PD and PFA in meaningful ways. Changing T changes PD and PFA because it changes the condition of the probabilities. Given PD1 and PFA1 for look period T1, it should be possible by analysis to estimate what PD2 and PFA2 at look period T2 would be.
d follows from PD and PFA using the usual assumption of Gaussian statistics. So will d also change with look period T? Generally it will, but there should be a regime for your your choice of T over which d will be fairly constant. Perhaps this will be as T becomes long.
The plausible choices for T should be evident from the details of your experiment design and its objectives.
(these kinds of things occur in my own work in target detection by humans using underwater systems, usually sonars)
Ronald
Thanks Ronald! Exaclty this (getting PFA) was my problem, but the cited reference proposes a method that seems to solve this, which results in an estimate of d' that is quite robust to changes in T.
Hello Nicolai - please see also (https://sites.google.com/a/mtu.edu/whynotaprime/). Here, the author proposes an argument for using nonparametric sensitivity (particularly when the Hit or FA rate = 0). The website explains it better, and a formal citation is Zhang & Mueller, 2005 (http://obereed.net/docs/ZhangMueller2005.pdf)
The conventional nonparametric alternative is A', which makes no assumptions about underlying distributions. Kessel's advice above based on human observer target detection from sonar signals seems a pretty close match to your problem.
I am not sure that I understand your paradigm fully, but if there are several different stimuli, an alternative to d' (and SDT altogether) is percent-correct and the body of theory behind that.
- Badges
- Science topic
- Similar topics
- Psychology
- Cognitive Science