Background:

In signal detection theory d' provides an estimate of memory sensitivity that is supposedly uncontaminated by response bias.  The authors argue that if d' truely provides an uncontaminated estimate of sensitivity then it should provide the same estimates when tested using forced choice tests as when using old/new recognition.  The authors test this assumption by collecting both old/new and forced choice recognition responses.

Experiment 1

Subjects were presented with 280 pictures for study at a rate of 2 seconds per picture.  Another 280 were used as lures on the tests.  Half the 280 items that were studied were tested in a forced choice fashion and half were tested in a old/new recognition fashion.  The first 90 items on the test were used for practice.  Subjects were given feedback about their performance.  The old/new and forced choice test trials were randomly mixed together.

• The main result of this experiment was that d' was higher for old / new recognition than it was for forced choice.  This can be seen in their Figure 1.
• The difference between old/new recognition and forced choice was greatest for subjects with conservative response criterion.  This can be seen in their Figure 2.
• This results can be accomodated by either the unequal variance signal detection model or by the dual process signal detection model.

This was like Expeirment 1 except that subjects provided confidence judgments as well.  These confidence judgments allowed a direct test of the response bias account (i.e. they could look at cases where subjects had high confidence versus lower confidence) and also allowed the authors to construct ROC curves in order to look more closely at the dual process and unequal variance models.  Experiment 2A was just like Experiment 1 except that confidence judgments were collected.  Experiment 2B was made more difficult by creating lures that were related in various ways to the targets.

• The ROC curves were asymmetrical and fit the dual process and unequal variance models better than the signal detection model.
• Estimates of d' calculated at different confidence levels again showed that as subjects became more conservative in their responses the estimated d' was greater for old/new recognition than for forced choice (see Figures 6 and 7).
• Parameter estimates obtained by fitting the dual process model accurately predicted accuracy on the forced choice test.

It was still possible that old/new recognition relies more on recollection than forced choice recognition.  In Experiment 3 remember/know judgments were collected to directly test this claim.  They used the same materials as in Experiment 2A.

• The proportion of "remember" responses on the two types of tests were not significantly different.

However, the results of these three experiments suggest that the same underlying processes (recollection and familiarity) are responsible for people's performance on both tasks.  Measures of d' are misleading, because it is systematically influenced by response bias. The dual process model is a better way of explaining the data than is the unequal variance model.