Analysis of AX discrimination data

1.  AX design, stimuli drawn in pairs from along a continuum

AX: a pair of stimuli; is X the same as A, or different?; response = "same" or "different" . 

roving (vs fixed): the A stimulus varies along the continuum just as the X does, rather than serving as a fixed standard

ISI: InterStimulus Interval, time between the stimuli in a pair.  The longer the ISI, the more likely the response will reflect a categorization of the stimuli.  Half a second is plenty of time for a subject to give a response.

same trials: for some of the trials, the stimuli will be identical.  A difficult design question is what proportion of trials should be same trials.  Many different trials will sound the same to listeners, so if half the trials are same trials, most of the responses will be "same", which is boring and/or alarming for subjects.

step size: the distance between the stimuli in a pair, in terms of steps along the continuum.  1-step: adjacent on the continuum.  Obviously the step-size per se has no importance, as it depends on how the continuum was constructed (e.g. a 2-step pair from a 5 msec VOT continuum is the same as a 1-step pair from a 10 msec VOT continuum); but for a given continuum larger steps are easier to discriminate.

Discrimination experiments last longer than the corresponding identification experiments.  Although there are fewer different pairs than there are singleton stimuli, there are same pairs as well, and each trial is longer (2 stimuli + ISI).
 

2. Tabulating responses

Much as for identification responses, discrimination responses can be tabulated as the proportion of the repetitions of a pair classified as "same" or as "different".  Unlike identification responses, discrimination responses are right or wrong, so they can also be tabulated as the proportion of the repetitions judged correctly.  (For the different trials, these 2 measures are the same, and often the same trials are not tabulated at all.)  Depending on the response bias of a listener, trials heard as "same" might be judged anywhere from 0% different (listener is tends to respond "same") to 50% different (listener responds at chance).

Graphing: The x-axis is the stimulus dimension, and by convention the value of a pair of stimuli is taken to be the mid-point between the values of the 2 stimuli.  E.g. the x-value of a pair (+10, +20) is +15, while the x-value of a pair (+10, +30) is +20.

practice file: discrim_data.xls, sheet 1: pretend data from a VOT continuum, values are proportion "different" out of 10 repetitions.
 

How to analyze?

3.  ANOVA, just as for identification data

By itself a 1-way ANOVA on stimulus pair will tell you if subjects treated the pairs differently (which they virtually always do); more interesting will be post-hoc tests/planned comparisons.  Such tests will tell you which stimulus pairs are perceived differently from which others, and this is a relevant research question for discrimination data.

practice file: discrim_data.xls, sheet 2: ready to open in SPSS  

·  one continuum, 2 groups of subjects (sheet 2)

·  same data, but now from 1 group of subjects, comparing across continua (sheet 3) (if you enter the factors in the order pair(7), continuum(2), you can practice matching factor levels to variable columns by hand)

4.  Fit a curve/function to the responses?  No, never seen.
 

5.  Analysis of the discrimination peak (across continua and/or subject groups)

Discrimination peak: well-discriminated pair(s), usually with higher responses than to other pairs, thus forming a peak in the plotted discrimination responses.

location: which pair along the continuum is discriminated the best?
height: what is the greatest proportion of responses?
width (not generally quantified): how many pairs are discriminated well, e.g. above chance?  This is a measure made up for this exercise, but a sensible one.

practice file: discrim_data.xls, sheet 1 right columns, and last sheet
(above chance here = 8/10 or better, by a binomial test)
 

6.  Relating discrimination responses to identification responses: within- vs. across-category pairs

The discrimination peak is expected to span the identification boundary.  That is, expectations about performance are different for different pairs drawn from within a single phonetic/phonemic category, vs. from 2 different categories.  Since such a classification of the pairs as within vs. across categories can be established a priori, planned comparisons are appropriate.  (Really, the classification should be established for each individual subject from that subject's identification responses.  But it is generally established for the group as a whole; and in one study of normal and dyslexic children, published adult identification data were used to establish the within and across pairs for the child subjects.)   Or, often the responses to all within-category pairs are averaged, and likewise to all across-category pairs, and these averages are compared (e.g. by paired t-test).  Here is an example (p. 256) from Pisoni (1973), Perc. & Psychophys. 13(2): 253-260, in which the x-axis variable is ISI.  The filled datapoints are means of across-category pairs, and the open datapoints are means of within-category pairs, for 4 different continua.

Relating discrimination responses to identification responses: comparing predicted vs. obtained discrimination

7.  Let's look at the relation between some pretend ID and discrimination data, checking out the locations of the categories and, qualitatively, the correspondence between the 2 sets of data.  Recall that categorical perception refers to perception by categories, specifically, discrimination limited by categorization.

practice file: discrim_ID_data.xls, top sheet right columns

In the ideal case, identification will be perfectly categorical (abrupt boundary between 2 adjacent stimuli).  The discrimination peak should cross this 1-step category boundary, and all other pairs should be discriminated at or below chance.  That is, if a pair of stimuli (a,b) have opposite categorizations (0,1 or 1,0) then their discrimination score will be 1, while if a pair of stimuli (a,b) have the same categorization (0,0 or 1,1) then their discrimination score will be 0 (or at chance, depending on response bias).

Haskins researchers (originally, Liberman et al. 1957, for ABX discrimination; see also Pollack & Pisoni 1971, Psych. Sci.) predict the discrimination score for a pair of stimuli, using the identification score of each stimulus.  With this formula, whenever the identification of the 2 stimuli is the same - both 0, or both .5, or any other value - the predicted discrimination is at chance.   Here is a general formula for AX discrimination data, taken from Godfrey et al. (1981). 

Predicted discrimination = (P1a x P2b) + (P1b x P2a), where

·        P1a  = proportion of times stimulus 1 was identified as “a”

·          P2b  = proportion of times stimulus 2 was identified as “b”

·          P1b  = proportion of times stimulus 1 was identified as “b”

·        P2a  = proportion of times stimulus 2 was identified as “a”

Not surprisingly, this is not what real discrimination data usually look like.  (If identification of a stimulus is split between the 2 responses, then discrimination of pairs including that stimulus will be above chance; furthermore, even when the identification is consistent, discrimination can still be above chance, because it can be based on auditory memory for the particular stimuli.)  Here is an example from Best et al. (1981), a rare paper in showing these predicted-discrimination functions:



8.  The relation between obtained and predicted discrimination is then generally tested by ANOVA, with factors "obtained vs. predicted" and "stimulus pairs".  The typical result is that both factors, and their interaction, are significant.  That is, the obtained and predicted functions are different, but only for certain pairs - the within-category pairs, which are discriminated better than predicted.  (The across-category pairs are generally discriminated perfectly, as predicted.)  This result - better obtained than predicted discrimination - goes back to the first paper to present this kind of analysis, Liberman et al. (1957). 




Prepared by Pat Keating, Spring 2004

Back to the UCLA Phonetics Lab statistics page.