A higher d' indicates that they were better able to perform the task, or are more "accurate" with responses (fewer misses or false alarms). The correlation between facial recognition ability and d' would therefore indicate that people with better facial recognition perform better at the task; i.e. are better able to recognise faces they've seen, while correctly rejecting those they haven't.

As for false memories being lower, that's harder to say. Are people with lower d' performing worse because they're getting more false alarms, or because they're missing more? It might be worth looking at the Criterion (c) as well.

Thank you Logan for your answer!

May I ask you, what is the meaning of different scores of d'? For instance, what is the difference between d'=0, d'=1 and d'=2. Which range of d' is good for an experiment?

Hi Mehran Emadi Andani. It depends on the kind of task you are using. If it's a Yes-No discrimination task or a 2AFC. d' values normally go from 0 to 4 (in 2AFC tasks values rarely go beyond 3) and the numbers are just differences between distributions of probabilities for scoring a hit in relation to a false alarm so it doesn't make much sense to estimate d' values if you don't have a condition to compare them with. But any d' value above zero means that the participant can discriminate the signal over noise above chance.

Many thanks Renzo Lanfranco for your reply!

What are the differences between d'=0, d'=1, d'=2, etc.? The answer is not simple. We can begin with the logical and pragmatic range of d', which is from 0 (chance) to 4.65 (based on Hit rate = .99, FA rate = .01). (Anything higher than d' = 4.65 will be getting awfully close to Hit = 1, and FA = 0, either of which give you "undefined" numbers [infinity].)

Importantly, because d' is a distance measure, and therefore has a true 0 value (no sensitivity at all), it is on a ratio scale. So, a d' = 2 is exactly twice as sensitive as d' = 1, and d' = 4 is twice as sensitive as d' = 2, etc. In this way, d' is like weight, and length, which are also ratio scales. Therefore, to take a more interesting example, since d' = 4.65 * 1/10 = 0.465, a d' of 0.465 is 10% of the way from chance (0) to maximum (4.65). That tells us that d' = 0.465 is quite poor. On the other hand, d' = 4.65 * 8/10 = 3.72, so a d' of 3.72 is 80% of the way from chance to maximum, which is really good (and 8x better than d' = 0.465).

There is one other factor to bear in mind, however, which is the Hit and FA rates that would produce a given d'. So, to take the previous example of d' = 3.72, various combinations of Hit and FA rates can produce this value. But they are all getting really close to the maximum pragmatic difference (i.e., H = .99, FA = .01). If you have 0 bias (c = 0), then H = .97, FA = .03 will produce d' = 3.72. This intuitively feels a lot closer to H = .99, FA = .01, which produces d' = 4.65, than only 80% of the way. This is because d' is based on Z scores. So, d' is the number of normal SDs above the mean, which is non-linear. 3.72 SDs above the mean is really high. So, you have to consider the Hit and FA values as well as the d' in judging how good a given level of d' is.

With this in mind, MacMillan and Creelman (2005, p. 9) note that researchers generally try to avoid ceiling or floor effects, and thus, in a 2-alternative forced choice (2-AFC) task they often look for accuracy between 60-90%. 60% accuracy corresponds to roughly d' = 0.5. In terms of H and FA rates, if you are completely unbiased (c = 0), and H = .60, FA = .40, then d' = 0.51. 90% accuracy corresponds to roughly d' = 2.5. In terms of H and FA rates, if you are unbiased, then H = .89, FA = .11 produces d' = 2.5. So, the typical range of d' in studies lacking ceiling or floor effects is d' of 0.5-2.5 (with the latter, again, being 5x more sensitive than the former). I hope this helps!

Get the same question today! Thanks for your replies.

d prime can also be interpreted as the distance of two standard Gaussian distributions (Noise, Signal+Noise) in units of standard deviation. So any d prime higher than 2 is already quite high. As Les said, d prime is typically useful between 0 and 2.5. If you're getting values higher than 2.5 you need a very large number of trials to still measure errors reliably.

If you are more familiar with AUC, you can convert your d prime to AUC. However than you lose the linearity of the scale.

In an expertise study we recently found that expert echocardiographers performed at a d prime of about 2.5, compared to about 1.0 for medical students.