Reply to query

[Douglas Creelman <creelman@xxxxxxxxxxxxxxxxx>]

Salvadmr -

This is a not-uncommon issue - it does arise from time to time. Here is 
my take on the "problem"
of negative values of d'.

First, we consider an identification task the same way as we would treat 
a detection task. In
theory, detection was characterized as discriminating "Noise Alone" from 
"Signal plus Noise,"
whereas identification involves two signals perturbed by noise. The 
formal approach is the same.

As always, the ways numbers from data are treated depends on how they 
are to be used. For
instance:
If these d' values are to be averaged with others, I would suggest 
keeping the negative values;
If many observers gave reasonable data and one listener's data are 
anomalous, I might try to find
a reason to discard the data;
If these represent overall data, you could check to see if they are 
significantly different from zero.
This may not be easy; see Chapter 13 of Macmillan & Creelman, 2nd Ed.
  If not significantly different from zero, I might report them as zero
  If significant, I would look for a reason my experimental participant 
labeled, for instance, a
/pa/ as a /ta/ reliably. A negative d' represents as successful 
discrimination as a positive one. So,
if the observer simply used the wrong keys for responding I suppose 
taking the absolute value
would not be out of the question.

This is only the beginning of what could be an extensive discussion. Is 
your continuum really
unidimensional? Do listeners understand their task? Many questions.

Doug Creelman


At 01:22 PM 5/2/2006, silvadmr wrote:

>>Dear list,
>>
>>I have some few small negative values of d' obtained from an
>>identification task in witch subjects label vowel sounds from a
>>continuum with tree phoneme labels. Generally, the graphs and tables
>>in the literature never show negative values of d' ( i think
>>it  occurs here and there).
>>The interpretation of such values from an identification task of
>>that kind must not be the same as the interpretation of negative
>>values obtained in a yes-no experiment...cause in the case of
>>indentification we have "different sounds" rather then a "signal
>>stimulus" and a "signal plus noise stimulus".
>>
>>What is the usually adopted procedure?
>>Should I take the absolute values (given the nature of the
>>identification task)?...should I convert them to 0? should I keep

--
C. Douglas Creelman             416-690-9407 (phone & fax)
9 Fernwood Park Ave.           creelman@xxxxxxxxxxxxxxxxx
Toronto, ON Canada              
M4E 3E8