# Parkin talk on false positive memory

Date: Wed Mar 13 1996 - 12:43:55 GMT

> Date: Tue, 12 Mar 1996 13:28:43 +0000 (GMT)
> From: "E.J.Fletcher" <ejf195@soton.ac.uk>
>
> The cog. sci. talk was very interesting yesterday. The only problem is
> I didn't understand what "pathological liberal" and "pathological
> conservative" meant. can you please explain this when you have time?

Hi Emma,

I think what that meant was the ways in which JB's memory went wrong:
His usual error pattern was overliberal: along with things he really
saw, that he did remember he had seen, were all those false positives:
things he didn't see, but he remembered as if he had seen.

When they were given the "incentive" instructions ("Be careful, because
you'll get a reward each time you say you saw an item you really saw,
but you'll lose the same amount each time you say you saw an item that
you didn't really see") both patients, instead of becoming more
selective and cutting down on false positives, instead simply cut down
on EVERYTHING: Fewer correct positives and fewer false positives, but
in the same proportion, which was roughly equal.

This was the pattern for JB, but what Parkin meant by "pathologically
conservative" was what the other patient did, which was hardly to say
he remembered seeing anything at all under those conditions.

I will invent some numbers to illustrate:

Normal subjects, when asked about 50 items, 25 of which WERE seen
before, 25 not, will remember they saw, say, 20 out of the 25 that they
really saw, and will wrongly remember having seen, say, 1 out of the 25
that they did not see.

The two patients will say they saw 18 out of the 18 they did see, and
about 18 of the 25 they didn't see.

Give them the "incentive" instructions, and JB goes down to, say
10 and 10 (fewer overall, still the same proportion of false positives
to correct positives) and the other patient will hardly say he saw any at
all, say, 2 correct positives and 2 false positives.

Maybe they're now both pathologically conservative under these
conditions, in that they don't become more selective in sorting out
what they did and didn't see, they just cut back on both kinds of
positives (equally).

There is a better way to analyse all this, called "signal detection
analysis." It's based on things like radar signal detection: You have
to decide whether you have or have not seen a very faint "blip" on the
radar screen. You could get it right by saying there was a blip when
there really was one (a correct positive) or saying there was no blip
when there really wasn't one (a correct negative) or you could make 2
kinds of errors: false positives (saying there was a blip when there
wasn't one) or false negatives (saying there was no blip when there was
one).

Signal detection analysis allows you to analyse all 4 cases, and not
just the correct and incorrect positives that Parkin considered. Using
all 4 cases, the analysis allows you to calculate two separate things:
Your sensitivity to the signal (called d') and your cutoff point (i.e.,
whether you prefer to err in the direction of false positives [liberal]
or false negatives [conservative]), called b, for "response bias").

The advantage of this deeper analysis is that a lot is known about how
bias is influenced: It is influenced by how often blips occur (if
they are rare, and non-blips are common, your cutoff point would be
different from what it would be if they were equally frequent, or
non-blips were the rare ones). It is also influenced by the "payoffs":
the rewards for correct positives and correct negatives, and the costs
of false positives and false negatives. (Think of a murder trials, and
punishing innocent people.)

A signal detection analysis would show whether the main difference
between normal subjects and these frontal amnesics is in d' (memory)
or b (bias). It could also show whether the effects of incentives and
instructions and feedback trial-by-trial influence d' or b or both (or
neither).

> (By the way, you didn't read my essay assignment for Matt Dye did you?)

Not yet, but I plan to... Chrs, S

This archive was generated by hypermail 2b30 : Tue Feb 13 2001 - 16:24:16 GMT