Transcript 6.-triad-testsx

Triad tests
Collecting and making similarity data
Background
 The triad test: from personal construct
psychology and the repertory grid developed
by George Kelly in 1955.
 Adopted by cognitive psychologists and
anthropologists to measure similarities
among pairs of objects.
 Informants respond to the same set of cues and data
can be compared across individuals.

Kelly, G. A. 1955. The psychology of personal constructs.New York: Norton.
Triad tests
 In a triad test, you show people three things and tell
them to:
 “Choose the one that doesn’t fit” or
 “Choose the two that seem to go together best,” or
 “Choose the two that are the same.”
 The “things” can be photographs, dried plants, or 3 x 5
cards with names of people on them.
DOLPHIN
MOOSE
WHALE
SHARK
DOLPHIN
MOOSE
 Dolphin is the odd item in triad 1 if people are
thinking about size.
 It’s moose if people are thinking land vs. sea.
 People often ask “What do you mean by things being
‘the same’ or ‘fitting together’?”
 Tell them that you are interested in what they think that
means.
 By doing this for all triples from a list of things or
concepts, you can explore differences in cognition
among individuals, and among cultures and
subcultures.
First use in anthropology
 Romney and D’Andrade (1964) presented
people with triads of American kinship terms
and asked them to choose the term that was
most dissimilar in each triad.
 For the triad “father, son, nephew,” 67%
selected “nephew” as the most different of the
three items; 22% chose “father.”

Romney, A. K., and R. G. D’Andrade, eds. 1964. Cognitive aspects of English kin terms. In
Transcultural studies in cognition. American Anthropologist 66 (3, part 2, entire issue): 146–70
Ask people to explain their choices
 For the triad “grandson, brother, father,” one informant
said that a “grandson is most different because he is
moved down further.”
 There’s a lot of cultural wisdom in that statement.
Bilingualism and St. Lucian disease terms
 Does cognition of disease terms vary with
bilingual proficiency in St. Lucia?
 52 bilingual English-Patois speakers and 10
monolingual Patois speakers.
 Here is the formula for calculating the
number of triples in a set of items:
 With 9 disease terms, there are 84 possible
triads.

Lieberman, D., and W. W. Dressler. 1977. Bilingualism
and cognition of St. Lucian disease terms. Medical
Anthropology 1:81–110.
n(n -1)(n - 2)
6
 For the bilingual informants:
 Two triad tests, a week apart, one in Patois and one in
English.
 They also measured how bilingual their informants were.
 For the monolingual informants”
 The 10 monolingual Patois informants were simply given
the triad test.
There are n(n-1)/2 pairs in a set of items
 For 9 items, there are 9(8)/2=36 pairs of items
 Count the number of times that each of the 36 possible
pairs of terms was chosen as most alike in the 84
triads.
 Each term in a triad test appears n-1 times, and each
pair appears n-2 times, so they divided the total by
seven.
Triads as similarities
 This produced a similarity coefficient, varying between
0.0 and 1.0, in increments of 0.14 (1/7) for each
possible pair of disease terms.
 The larger the coefficient for a pair of terms, the closer in
meaning are the two terms.
 They analyzed these data for English-dominant, Patoisdominant, and monolingual Patois speakers.
 When Patois-dominant and English-dominant
informants took the triad test in English, their cognitive
models of similarities among diseases was similar.
 When Patois-dominant speakers took the Patoislanguage triad test, however, their cognitive model was
similar to that of monolingual Patois informants.
The take-away:
 Patois-dominant bilinguals hold on to two psychological
models about diseases and switch back and forth,
depending on the language they are speaking.
 English-dominant bilinguals employ the same cognitive
model of disease terms, whichever language they
speak.
BIBDs – Balanced incomplete block designs
 Typically, the terms for a triad test are generated by a
free list.
 Free lists of illnesses, ways to prevent pregnancy,
advantages of breast-feeding, places to go on vacation,
and so on easily produce 60 items or more.
 Even an abbreviated list for a pile sort may be 30 items.
 With just 9 terms, there are 84 stimuli in a triad test.
 For 15 items, the number of decisions an informant has
to make jumps to 455.
 At 20 items, it’s a mind-numbing 1,140.
Burton and Nerlove’s solution
 Balanced incomplete block designs take advantage of
the fact that there is a lot of redundancy in a triad test.
 For 4 items, there are 6 pairs. If you ask “Which of
these two foods is more nutritious?” the informant sees
each pair just once.

Burton, M. L., and S. B. Nerlove. 1976. Balanced design for triad tests. Social Science
Research 5:247–67.
 Recall that in a triad test informants see each triad
n(n–1)(n–2)/6 times
 and each pair n–2=2 times.
 With 84 triads for 9 items, each pair of items appears 7
times.
Redundancy and lambda designs
 This redundancy lets us reduce the number of triads in
a triads test.
 If each pair appears once (a lambda 1 design), then
only 12 triads are needed.
 If each pair appears twice (a lambda 2 design), then 24
triads are needed.
 For a lambda 1 design, each pair of items in
the similarity matrix gets a score of either 1 or
0.
 For a lambda 2 design, each pair of items in
the similarity matrix is 0, .5. or 1.
 For 10 items, a lambda 2 design requires 30 triads; for
13 items, 52 triads; for 15 items, 70 triads; for 19 items,
114 triads; and for 25 items, 200 triads.
 Here are the solutions from Burton and Nerlove for 9
and 10 items.
Balanced Incomplete Block Designs for Triad Tests Involving 9 and 10 Items
For 9 items, 24 triads are needed,
as follows:
Items
For 10 items, 30 triads are
needed, as follows:
1, 5, 9
2, 3, 8
4, 6, 7
1, 2, 3
4, 5, 6
7, 8, 9
Items
1, 2, 3
2, 5, 8
3, 7, 4
2, 6, 9
1, 3, 4
5, 7, 8
1, 4, 7
2, 5, 9
3, 6, 8
4, 1, 6
5, 8, 7
6, 4, 9
9, 5, 2
10, 6, 7
1, 3, 5
3, 7, 9
2, 4, 5
1, 6, 9
2, 4, 8
7, 9, 1
8, 10, 2
2, 7, 6
3, 8, 9
1, 6, 8
4, 8, 9
3, 5, 7
1, 5, 8
9, 3, 10
10, 6, 5
4, 2, 10
5, 6, 3
1, 2, 4
6, 1, 8
3, 5, 6
1, 2, 7
6, 8, 9
7, 10, 3
8, 1, 10
2, 6, 8
2, 3, 6
7, 9, 2
2, 4, 8
8, 4, 7
4, 9, 5
5, 7, 1
9, 10, 1
10, 5, 4
3, 4, 9
BIBDs reduce accuracy
 In a L1 design, each pair of items occurs in the context
of just one other item.
 For dog, parakeet, whale, many people pick whale,
even though dog and parakeet are not that similar.
 Thus the similarity between two items is defined by a
single, third item.
Steps in making a triads test
 1. Identify domain.
 2. Do key informant interviewing /free listing.
 3. Select short list from the total list.
 4. Choose a BIBD or use a complete triad test.
 5. Make the triad tests using Anthropac (randomize to
eliminate order effects).
 6. Collect triad data.
 7. Score the data and unrandomize if necessary
 Instructions for data entry in the Anthropac manual
Procedure in Anthropac
 Tell Anthropac the list of items you have.
 Select a design (e.g. L1, L2).
 Tell it the number of informants you want to interview.
 Anthropac prints out a randomized triad test, one for
each informant.
 Randomizing the order in which the triads appear to
informants eliminates order-effects.
 Anthropac unrandomizes the data on import.
Plusses and minuses of triad tests
 Triad tests are easy to create with Anthropac,
easy to administer, and easy to score.
 Use with relatively few items in a domain.
 In literate societies, most people respond to 200
triads in <.5hr, but it’s boring.
 Literate informants easily handle L2 tests with
15 items and 70 triads.
 People prefer—even like—to do pile sorts, but
pile sorts have their own problems.