I am doing a psycholinguistic experiment, where I want to calculate d’ for each participant’s test scores. In the test, participants listen to 4 familiar words and 8 novel distractor words, and have to decide for each stimulus whether they have heard it before (familiar) or not (novel). So I have N=12 test trials. The number of test trials (at least for the familiar words) cannot really be changed because of the nature of my experimental design. Due to this low number of trials, it happens quite frequently that participants have hits or false alarm scores of either 0 or 100% and I cannot calculate a d’ for them. It seems that a common solution for this issue is to replace proportions of 0 by proportions of 1/(2N) and proportions of 1 by proportions of 1-(1/(2N)), where N is the number of trials, i.e. N = 12 in my case (MacMillan & Creelman 2005, Stanislaw & Todorov 1999).
However, for my low number of trials, it seems to me quite a massive change to my data to e.g. replace a recorded proportion of hits of 1 by 1-(1/(2*12) = 0.96.
I have one condition in my experiment, where I would actually predict an almost perfect discrimination between familiar and novel words, but when I do this adjustment, I am not sure if this perfect discrimination is still reflected in the d’.
So my question is if this procedure of adjusting proportions of 0 and 1 also works for such a low number of trials or if it affects my d’ scores in a way that they become nonsensical? If this procedure is not ideal, what would be a better way to solve this issue?
