Download as pdf or txt
Download as pdf or txt
You are on page 1of 14

zyxwv

THE POWER OF SENSORY DISCRIMINATION METHODS

zyxw
USED IN REPLICATED DIFFERENCE AND PREFERENCE

zyxwv
TESTS

JIAN BI
The Institute for Perception
7629 Hull St.. Suite 200
Richmond, VA 23235
and
DANIEL M. ENNIS

zyxw
The Institute for Perception
The University of Illinois and
The Medical College of Virginia

Received for Publication July 6, 1998

ABSTRACT

Research on the power of discrimination methods in difference and


preference tests has both theoretical and practical significance. Power is
important to evaluate the sensitivities of tests and determine sample size. Ennis
and Bi ( I 998, 1999)proposed the beta-binomial distribution to model replicated
difference and preference tests with inter-trial variation and analyzed in general
the power of the tests. In this paper, the power &f discrimination methods for
replicated digerence and preference tests is discussed further. The equationsfor

zyxw
calculating power for methods based on the BB model are given. Eramples with
tables and charts for calculating and comparing the power of the methods are
also given.

INTRODUCTION

Research on the statistical power of discrimination methods in difference

zyxwvut
zyxwv
and preference tests has both theoretical and practical significance. Methods to
calculate and compare the power of discrimination methods are important to
evaluate the sensitivities of testing methods and determine the sample size of the
tests. If a test has too little power, relative to an important difference, the test
may be inconclusive. On the other hand, if a test has too much power, the

zyxwv
sensitivity increases to the point that small differences, which may have no
practical meaning, can be detected.

Journal of Sensory Studies 14 (1999) 289-302. All Righrs Reserved.


OCopyrighi 1999 by Food & Nutrition Press, Inc., Trumbull. Connecticut. 289
290 zyxwvu
zyxwvut J. BI and D.M. ENNIS

Many authors, for example, Hopkins and Gridgeman (1955); Ura (1960);
David and Trivedi (1962); Bradley (1963); Frijters (1979, 1988); O’Mahony and
Odbert (1985); Ennis (1990, 1993) and Bi (1995) have discussed and compared
the power and sensitivity of discrimination methods. However, the analysis done
and results given about the power and sensitivity of discrimination methods by
the authors are all based on a simple random sample. In many situations,
however, the data for difference and preference testing are not from a simple
random sample, but involve replications. Should difference tests be replicated
always to assess the extent of inter-trial variability? If unreplicated tests are
conducted without any knowledge of inter-trial variability, is a binomial test on
the choice proportion reliable? We will address these questions in a future paper
and simply assume here that within-trial replication has been conducted.
Ennis and Bi (1998, 1999) proposed the beta-binomial (BB) distribution to
model replicated difference and preference tests with inter-trial variation and
analyzed, in general, the power of the tests. However, further comparisons and
calculations of powers for various discrimination methods based on the BB
model were not discussed in that paper.
In this paper, the powers of the nondirectional 2-AFC (two-sided paired
comparison), directional 2-AFC (one-sided paired comparison), 3-AFC, Duo-
Trio and Triangular methods in replicated difference and preference tests are
discussed. Equations for calculating and comparing the power of the methods
based on the BB model are given. Numerical examples with corresponding

zyxwvu
tables and charts are also given.

zyxw
zyxw
zyx
zyxwvuts
POWER CALCULATIONS AND COMPARISONS

Let x, be the number of choice responses in the ith trial in the replicated
difference and preference test ( i = 1, 2, ..., k); n is the number of replications,
k is the number of trials. x, conditionally follows a binomial distribution denoted
by B(n, pi)in the ith trial, but follows a beta-binomial distribution denoted by
BB(p, y) in the whole experiment. p and y are parameters of the BB model.
The BB model can also be parameterized by p and 8. The relationship between

y and 8 is y=-. 8 We use y in this paper for convenience, because it is


(1 +0)

p-
k

Cbi
i=l
z
bounded by 0 and 1. p measures the mean of the proportions in the k trials
while y measures the variation among trials. The moment estimate of p is:
k
Cxi
- i-1
k nk
. Because the mean and variance of x,/n = pi are p and v2,
zyxwvu
zyx
zyxwvu
zyxw
POWER OF SENSORY DISCRIMINATION METHODS

zyxw
zyxw
zyxwvu
respectively, and v2=II(1-cI)+yp(1-p)(1--)
n

mean and variance of fi are p and -.V L


k
1
n
(Ennis and Bi 1998, 1999), the

According to the Central Limit

Theorem, the statistic fi asymptotically follows a normal distribution with mean


291

p and variance -.V 2 The moment estimate of y is: 3= nS -- 1


k fi(l-fi)k(n-l) (n-1)'
k
where S=c(ji-fi)2.
i=l
The power of discrimination methods in replicated difference and
I
preference tests is the probability that a null hypothesis of p = p,, = - (for
2

the 2-AFC and the Duo-Trio) or p = po = - 1 (for the 3-AFC and the
3
Triangular) will be rejected when the true parameter value p = p,, p , # p0.

Choice Probability Criterion


One approach to the calculation of power is based on the choice probability
in the difference and preference tests. For the two-sided hedonic and nondirec-
tional2-AFC methods in a replicated difference and preference test, the power
can be determined by (1) for specified a,p , and y values:

where,
u =-+@-'(I--)-,
1 a v2

l 2 2sk

2
v2
which is the ( 1 - a ) percentile of a normal distribution N (0.5,-),
2 k
292 zyxw
zyxwvut J. BI and D.M. ENNIS

zy
zyxwv
which is the variance of (i under the alternative hypothesis, HI: p=pI,

2 1
v*=-+y(---
4n
1 1
4 4n
1,

which is the variance of (i under the null hypothesis, Ho: u = %,

zyxw
+(.) denotes the standard normal distribution function, and
+-I (.) denotes a percentile of a standard normal distribution.

zyxwv
For the one-sided hedonic, directional 2-AFC and Duo-Trio methods in
replicated difference and preference tests, the power of the testing methods can
be determined by (2) for specified a, pI and y values:

2 zyxwvu
where u2=-+@-'(1-a)-,
1 "2

Jk
which is the (l-a)percentile of a normal distribution N(0.5,- v2
).
k
2

For the 3-AFC and Triangular methods in a replicated difference and


preference test, the power of the testing methods can be determined by (3) for
specified a, pl and y values:

1
where u3=-+@-'(1-a)--.v 2 2 2
v 3 2=-+y(---). u3 is the (l-a)percentile of a
3 fi' 9n 9 9n
1 v3'
normal distribution N(-,-). v, is the variance of (i under the null hypothesis
3 k
H, : p = V3.

In the equations above, power was determined based on differences


specified in terms of choice probabilities. It is well known that choice
probabilities do not inherently represent differences consistently for different
POWER OF SENSORY DISCRIMINATION METHODS

methods. For example, although the duo-trio and 2-AFC methods have the same
293 z
guessing probability of 0.5, a choice probability of 0.75 for the 2-AFC
corresponds to a much smaller difference than that corresponding to a 0.75 for
the duo-trio.
In order to compare sensitivities of discrimination methods used in
replicated difference and preference tests, a Thurstonian concept, 6, should be

zyxw
applied. 6 represents the sensory difference in terms of perceptual standard

zyx
deviation units and d' is an estimator of 6. For a specified difference, 6, the

zyxw
different methods will produce varying response proportions, while the same
response proportions produced by the different methods correspond to different
magnitudes of difference.
If the true difference is assumed to be 6 = 6 , , the power of the discrimina-
tion methods used in replicated difference and preference tests can be calculated
using the above equations, but p , will be replaced withf(6,), wheref(6,) is the
response probability corresponding to 6, for different discrimination methods.
f (6,) can be found from Ennis (1993). When y = 0, there is no inter-trial
variation, so the BB model reduces to a traditional binomial model. In this
situation, the equations for the calculation of power given in this paper are

zyxwv
equivalent to the general power equation given by Ennis (1993). Based on the
BB model, the power Eq. (l), (2) and (3) in this paper are extensions and
generalizations of the general power equation in Ennis (1993) for discrimination
methods, based on the binomial model.

zyx
zy
Thurstone's 6 Criterion
An alternative method for calculating power is based on 6 and the variance
of it's estimate, d'. For the two-sided hedonic and nondirectional 2-AFC
methods in replicated difference and preference tests, the power of the testing
methods can be determined by ( l ' ) , for specified a,6, and y values:

zyxwvut
where 6(u,) is the 6 value corresponding to u, and 6(8,) is the standard
deviation of d', evaluated at 6, in replicated difference and preference tests (Bi

I:[
and Ennis 1998). 6*(6,)= - [l+y(n-l)] where the B value can be found from
the B tables (Bi et al. 1997).
294 zyxwvutsrq
zyxw
zyxwvuts J. BI and D.M. ENNlS

For the one-sided hedonic, directional 2-AFC and Duo-Trio methods the
power of the testing methods can be determined by (2') for specified a,6, andzy
y values:

where 6 (uJ is the 6 value corresponding to uz.

zyx
zyxwvu
For the 3-AFC and Triangular methods in a replicated difference and
preference tests, the power of the testing methods can be determined by (3') for
specified a,6, and y values:

power = 1 - f3 = 1 - 0 [ ";;;;']. (3')

where 6 (u,) is the 6 value corresponding to u3.


The power values calculated using the first approach, i.e., Eq. ( l ) , (2) and
(3), differ little from the values calculated using the second approach, i.e., Eq.
( l ' ) , (2') and (3'). Only in the second decimal place, when k is small, were
differences noted.
As in the traditional difference and preference tests, the 3-AFC and 2-AFC
methods have much higher power than the Triangular and Duo-Trio methods
under the same conditions. The smaller the y value, the larger the power. This
means that inter-trial variation decreases the power of tests. The larger the
sample size, (in terms of n and k) the larger the power. However, the
contribution of the number of trials (k) to power is larger than that of the
number of replications (n). Figures 1 and 2 give comparisons of power for
different methods, inter-trial variations (y) and sample sizes (n and k).

EXAMPLES

Example 1
A replicated difference test was conducted to compare the sweetness of a
new formula of beverage with the classical formula. The 3-AFC method was
used in the test. 30 panelists were selected for the test, with 5 replications for
each panelist. The number of correct responses for the panelists were: 2, 1, 5 ,
5,4,1,1,0,0,5,5,3,3,0,2,1,3,5,5,4,1,2,4,5,2,1,2,5,4and1.
POWER OF SENSORY DISCRIMINATION METHODS zyx
z 295

0.0
1

0.2
zyxwvutsrq
0.4
I

0.6
I

0.8
I

1.o

De(ta

0.0
I

0.2
I

0.4
zyxwvu 0.6 0.8 1.o

zyxw
zyxwv
Ddta

FIG. 1. POWER CURVES FOR DISCRIMINATION METHODS USING THE BB MODEL


A (y = 0.1, n = 5 , k = 50)
B ( y = 0.5. n = 5 , k = 50)
296 zyxwvut J . BI and D.M.ENNIS

'9
0

8 0
a

1 I I I 1 I

1 .o

zyxwvuts
0.0 0.2 0.4 0.6 0.8
Ddta

9
--B zyxwvu
zyxwvutsrq
m
0 zyxwvu
8 z-
a

x-
2-

I I I I I I

FIG. 2. POWER CURVES FOR DISCRIMINATION METHODS USING THE BB MODEL


A (y = 0.1, n = 10. k = 200)
B (y = 0.5, n = 10, k = 200)
zyxwvuz
zyx
POWER OF SENSORY DISCRIMINATION METHODS

It is assumed that we have no information about the variation among the

zyxwv
panelists. From the data, we obtain the estimated y value: 3 = 0.381. From
Table 3.4 in Bi and Ennis (1999), we find the minimum number of choice
responses for significance at (11 5 0.05 is 73, the observed number of correct
responses is 82. Hence, we can conclude that the sweetness of the two beverage
types significantly differ.
297

However, the interest here is to evaluate the sensitivity of the test, i.e., the
power of the test. From Table 1 (for y = 0.4) in this paper, we find that the
power of the test is: 0.123 for 6,= 0.1; 0.415 for 6, = 0.3; 0.763 for 6, = 0.5;
0.951 for 6, = 0.7 and 0.996 for 6, = 0.9. This shows that we have 0.763
confidence of finding a difference of 6, = 0.5. If the true difference between the
two beverage types is smaller, for example, 6, = 0.3, the chance of finding the

zyx
difference is only 0.415.

Example 2
In order to determine detectable differences in sweetness of two brands of
cookies, a company is planning to conduct a consumer difference test. The
supervisor of the test sets the allowed Type I error level Q = 0.05, Type I1
error level3!, = 0.2, i.e., the power is no less than 0.8 when the true difference
is 6, 2 0.3. The interest is how many consumers (k) and replications (n)are
needed for the test. Because the Duo-Trio and Triangle methods have much
lower power than the 2-AFC and 3-AFC methods, only the latter two methods
will be considered in the test. Because the test is directional, the nondirectional
2-AFC method will not be considered.
A pilot test with a small sample shows that the variation among consumers
for the test using the 2-AFC and 3-AFC methods are about 0.3 (y = 0.3) and
0.2 (y = 0.2), respectively. From Tables 2-3 and Fig. 3A-3B, we conclude that
many test plans will satisfy the specified conditions. For example, the following
four test plans: (1) k = 90 and n = 6 for the 2-AFC method; (2) k = 103 and
n = 4 for the 2-AFC method; (3) k = 61 and n = 6 for the 3-AFC method; (4)
k = 73 and n = 4 for the 3-AFC method. The supervisor of the test will select
one plan after consideration of the relative cost of different methods, available
consumers and replications.

CONCLUSION

In this paper we provide the analytical techniques needed to determine the


power of commonly used difference testing methods when within-trial replication
has been conducted. Criteria of a product difference may be made in terms of
a choice probability or in terms of Thurstone’s 6.
zyxwvu
J. BI and D.M. ENNIS

zyxwv
zyxw
zyxwvuts
zyxwv
TABLE 1.
POWER VALUES OF THE 3-AFC METHOD BASED ON THE BETA-BINOMIAL MODEL
(y = 0.4, a = 0.05, n = 5 )
6,
-
k
3
.I
,071 .I27
.3 .5
,202 ,294
.7 .9
.4
4 ,074 ,141 ,235 ,351 ,481
5 ,077 .I55 ,266 ,403 ,553
6 .079 .I67 ,296 .452 ,617
7 ,082 .I8 ,324 .498 ,673
8 ,084 ,192 ,352 ,541 ,722
9 ,086 204 ,379 ,581 ,764
10 ,088 ,215 ,404 ,618 ,801
I1 .09 ,227 ,429 ,652 ,832
I2 ,092 ,238 ,454 ,683 ,859
13 ,094 .249 .477 ,712 ,882
14 ,096 .259 ,499 ,739 ,902
15 ,098 .27 ,521 ,763 ,918
16 .I ,281 ,542 ,785 .932
17 ,102 .291 ,562 ,806 ,943
18 ,103 ,301 ,582 ,825 ,953
19 ,105 .3 1 1 .6 ,842 ,961
20 ,107 ,321 ,618 ,857 ,968
21 ,109 ,331 ,636 ,871 ,974
22 .I 1 ,341 .652 ,884 ,978
23 ,112 .35 ,668 ,896 ,982
24 ,113 .36 ,684 ,906 ,985
25 .I15 ,369 ,698 ,916 ,988
26 ,117 ,379 ,712 ,924 .99
27 ,118 ,388 ,726 ,932 ,992
28 .I2 ,397 ,739 ,939 ,993
29 .I21 .406 .751 .946 ,995
30 ,123 ,415 ,763 .95 I ,996
35 .I3 ,457 ,816 ,972 ,998
40 .I38 ,498 ,857 ,984 .999
45 ,145 ,535 .89 .99 I 1
'50 ,152 ,571 ,916 ,995 1
,165 ,635 .95 1 ,999 1

zyxwvu
60
70 ,178 ,691 ,972 1 1
80 ,191 .74 ,984 1 1
90 ,204 ,781 .99 1 I 1
100 ,216 ,817 ,995 I 1
I20 .24 ,873 .999 1 1
I50 ,274 ,928 1 I 1
200 ,329 ,973 1 1 I
Note: n is the number 01 plicat IS; k is the number of trials; 6, is an ass ncd true difference;y is a
spread parameter in the BB model; a is a significance Icvcl.
POWER OF SENSORY DISCRIMINATION METHODS zyx 299

-
- zy
3
zyxwv4 5 6 7
TABLE 2.

- -
POWER VALUES OF DIRECTIONAL 2-AFC METHOD BASED ON THE BETA-BINOMIAL MODEL
(6, = 0.3, y 0.3, a 0.05)
n
8 9 10 12 I5 20

zyxwvuts
k
3 ,103 ,107 ,111 ,113 .I15 ,116 .I17 ,118 .I2 ,122 .i23
4 ,115 .I2 ,124 ,127 ,129 .I31 ,133 .I34 ,136 ,138 .i4

zyxwvutsr
5 .I26 .I32 .I37 .I4 ,143 .145 .I47 ,149 ,151 ,154 ,156
6 . I36 ,144 .I49 ,153 .I57 ,159 ,161 .I63 ,166 .I69 ,172
7 .I47 ,155 .I61 .I66 .I7 ,173 .I75 ,177 .I8 .183 ,187
8 ,157 .I66 ,173 .I78 .I82 ,186 .I88 .I9 ,194 ,198 ,202
9 ,166 ,177 .I85 .I9 .I95 ,198 ,201 ,204 ,208 ,212 ,216
10 ,176 ,188 ,196 ,202 ,207 ,211 ,614 .217 ,221 .226 .23
I1 ,186 ,198 .207 ,214 ,219 ,223 ,227 .23 .234 ,239 ,244
12 195 ,208 ,218 ,225 .23 I ,236 ,239 ,243 ,248 ,253 ,258
13 ,204 ,219 .229 .237 .243 ,248 ,252 ,255 .26 ,266 ,212
14 .2 I3 ,229 .24 .248 .255 .26 ,264 .268 .273 ,279 ,285
I5 222 ,239 .25 ,259 ,266 .212 .276 .28 ,286 ,292 ,299
16 23 1 ,248 ,261 .27 ,277 .283 .288 ,292 ,298 .305 ,312
17 .24 ,258 .271 .281 .289 ,295 .3 .304 .31 ,317 ,325
18 249 ,268 ,281 ,292 .3 .306 ,311 ,316 ,323 .33 ,337
19 257 .277 ,292 .302 .311 ,317 ,323 ,327 .334 ,342 .35
20 266 ,287 .302 .313 ,322 ,328 ,334 ,339 ,346 ,354 ,363
21 275 ,296 .312 .323 ,332 .339 .345 .35 ,358 ,366 ,375
22 283 ,305 ,322 ,334 .343 .35 ,356 ,362 ,369 ,378 ,381
23 291 ,315 .331 .344 .353 .361 ,367 ,373 ,381 .39 ,399
24 .3 ,324 ,341 .354 .364 ,372 ,378 ,384 ,392 ,401 ,411
25 308 ,333 ,351 ,364 ,374 .382 .389 ,394 ,403 ,412 ,422
26 316 ,342 .36 ,374 ,384 .393 ,399 ,405 ,414 ,424 ,434
27 324 ,351 ,369 .383 .394 ,403 .41 ,416 ,425 ,435 .445
28 332 ,359 ,379 ,393 ,404 ,413 .42 ,426 ,436 ,446 ,456
29 .34 ,368 .388 .403 ,414 .423 .43 ,437 ,446 ,456 .467
30 348 ,317 ,397 ,412 ,424 ,433 .44 .447 ,457 ,467 ,478
35 387 ,419 ,441 .458 .47 .48 ,489 ,496 ,506 ,517 ,529
40 424 ,458 ,483 .5 ,514 .525 ,534 .54 1 3 3 ,564 ,577
45 459 ,496 S22 341 3 5 5 ,567 S76 ,584 ,595 ,608 .62
50 493 ,532 ,559 ,579 ,594 ,605 ,615 ,623 ,635 ,647 .66

zyxwvuts
60 555 ,597 ,626 ,647 ,662 ,674 ,684 .692 ,704 ,717 .73
70 612 .656 .685 ,706 .721 .733 .743 .75 ,762 ,774 ,787
80 662 ,706 ,735 .756 ,771 ,782 ,791 .799 .a1 ,821 ,833
90 707 ,751 ,779 .798 A13 ,823 ,832 3 3 9 ,849 ,859 .a7
00 747 ,789 ,816 ,834 ,848 A58 3 6 5 ,871 ,881 .89 ,899
20 813 ,851 374 ,889 .9 ,908 .914 ,919 ,926 ,933 .94
50 883 .913 .93 .94 1 .949 .954 ,958 .96 I .965 .97 ,974
-
00
r ni
949 ,966 ,975 .98 I .984 .986 .988 ,989 ,991 .992 ,994
:number of replications; k is the number of trials; 6,is an assumed true difference; y a
spread prnmeter in the BB model; a is a significance level.
300 zyxwvutsr
zyxwvu
zyxwv
J. 81 and D.M. ENNIS

zy zy TABLE 3.
POWER VALUES OF THE 3-AFC METHOD BASED ON THE BETA-BINOMIAL MODEL

zyxwvuts
- (6, = 0.3, y = 0.2, a = 0.05)
I

- 3 4 5 6 7 8 9 10 12 15 20
k
3 ,132 .I4 ,146 .I5 ,154 ,157 . I 6 ,162 ,166 . I69 . I74
4 ,147 ,157 ,165 .I7 ,175 ,179 ,182 ,185 .I9 ,194 .2
5 ,162 ,174 ,183 .I9 ,195 .2 ,204 ,207 ,212 ,218 ,225
6 ,176 .I9 .2 ,208 ,214 .22 ,224 ,228 ,234 .24 I ,249
7 .I9 ,205 ,217 ,226 ,233 .239 ,244 ,249 ,256 ,263 ,272
8 ,203 .22 ,233 ,243 .25 1 ,258 ,264 ,269 ,276 ,285 ,294
9 ,216 ,234 ,249 .26 ,269 ,276 ,283 ,288 ,297 ,306 ,316
10 ,228 ,249 ,264 ,276 ,286 ,294 ,301 ,307 ,316 ,326 ,338
II .24 I ,263 ,279 .293 ,303 .312 ,319 ,325 ,335 ,346 ,358
12 ,253 ,276 ,294 ,308 3 2 ,329 ,337 ,344 ,354 ,366 ,379
13 ,265 .29 ,309 ,324 ,336 .346 ,354 ,361 ,373 ,385 ,398
14 ,276 ,303 ,323 ,339 ,352 ,362 .371 ,379 .39 ,403 ,418
15 ,288 ,316 ,338 ,354 .368 ,379 ,388 ,396 ,408 ,422 ,436
16 ,299 ,329 .35 I ,369 ,383 .394 ,404 .4 12 ,425 ,439 ,455
17 .31 I ,342 ,365 ,383 3 9 8 .4 1 .42 ,428 ,442 .456 ,472
18 ,322 ,354 ,379 ,398 ,413 ,425 ,435 ,444 ,458 ,473 .49
19 ,333 ,367 .392 ,411 ,427 .44 .45 I .46 ,474 .49 507
20 ,344 ,379 ,405 ,425 .44I ,455 ,466 ,475 .49 ,506 ,523
21 ,354 .39 ,418 ,439 ,455 .469 .48 .49 ,505 ,521 ,539
22 ,365 ,402 .43 ,452 ,469 ,483 ,494 ,504 .52 ,537 ,555
23 ,375 ,414 ,442 ,465 ,482 ,497 ,508 ,518 ,534 ,551 .57
24 ,385 ,425 ,455 ,477 ,495 .5 1 S22 S32 ,548 ,566 ,584
25 ,396 ,436 ,467 .49 ,508 ,523 ,535 ,546 ,562 .58 ,599
26 ,406 ,441 ,478 SO2 ,521 ,536 ,548 ,559 ,576 ,593 ,613
27 ,415 ,458 .49 ,514 ,533 ,548 ,561 ,572 ,589 ,607 ,626
28 ,425 ,469 ,501 ,526 ,545 ,561 ,574 ,584 ,602 .62 ,639
29 ,435 ,479 ,512 ,537 ,557 ,573 ,586 ,597 ,614 ,632 ,652
30 ,444 .49 ,523 ,548 ,568 ,584 ,598 ,609 ,626 ,645 ,664
35 .49 ,539 ,575 ,602 ,622 .639 ,653 ,664 ,682 ,701 .72
40 ,532 ,584 ,622 ,649 .67 1 ,688 ,701 ,713 .73 ,749 ,168
45 ,572 ,626 ,664 ,692 ,714 .73 ,744 ,755 ,772 .79 ,808
50 ,609 ,664 ,702 .73 .752 .768 ,781 ,792 ,808 ,825 ,842
60 674 .73 ,768 .795 .814 ,829 .84 1 ,851 ,865 .88 ,894
70 .73 ,785 .a2 345 A62 ,876 .886 ,894 ,906 ,918 .93
80 778 ,829 ,862 ,884 ,899 .9 1 ,919 ,926 ,936 ,945 ,954
90 ,818 ,865 ,894 .913 .926 ,936 ,943 ,948 ,956 ,963 .97
I00 ,851 ,894 .92 ,936 ,946 .954 .96 ,964 .97 ,976 .98 I
120 ,901 ,936 ,954 .965 ,972 ,977 .98 1 ,983 ,987 .99 ,992
150 ,948 .97 ,981 ,987 .99 ,992 ,994 ,995 ,996 ,997 ,998
200
- ,983 ,992 ,996 ,997 .998 ,999 ,999 .999 1 1 1
P :ni ie number of replications: k is the number of trials; 6, is an assumed h e difference; y a
spreadparameter in the BB model; a is a significance level.
POWER OF SENSORY DISCRIMINATION METHODS zyx
z 301

zyxwvutsrq
zyxwvut I

zyxwv
4 6 8 10

FIG. 3A. EQUAL-POWER CURVES FOR THE DIRECTIONAL 2-AFC METHOD


(y = 0.3, 6, = 0.3, (Y = 0.05)

4 8 10

FIG. 3B. EQUAL-POWER CURVES FOR THE 3-AFC METHOD


(7 = 0.2, 6, = 0.3, (Y = 0.05)
302 zyxwvu
zyxw
zyxwvut J. BI and D.M. ENNIS

ACKNOWLEDGMENT

zyxw
We thank the reviewers for their helpful comments on this paper.

REFERENCES

BI, J. 1995. Nonparametric models for discrimination methods and sensitivity


analysis for triads. J. Sensory Studies 10, 325-334.
BI, J. and ENNIS, D.M. 1998. A Thurstonian variant of the beta-binomial
model for replicated difference tests. J. Sensory Studies 13, 461-466.
BI, J. and ENNIS, D.M. 1999. Beta-binomial tables for replicated difference
and preference tests. J. Sensory Studies 14, 347-369.
BI, J., ENNIS, D.M. and O’MAHONY, M. 1997. How to estimate and use the
variance of d’ from difference tests. J. Sensory Studies 12, 87-104.
BRADLEY, R.D. 1963. Some relationships among sensory difference tests.
Biometrics 29, 385-397.
DAVID, H.A. and TRIVEDI, M.C. 1962. Pair, triangle and duo-trio tests,
(Technical Report 55). Virginia Polytechnic Institute, Department of
Statistics, Blacksburg, VA.
ENNIS, D.M. 1990. Relative power of difference testing methods in sensory
evaluation. Food Technol. 44, 114, 116, 117.
ENNIS, D.M. 1993. The power of sensory discrimination methods. J. Sensory
Studies 8, 353-437.
ENNIS, D.M. and BI, J. 1998. The beta-binomial model: Accounting for inter-
trial variation in replicated difference and preference tests. J. Sensory
Studies 13, 389-412.

zyx
FRIJTERS, J.E.R. 1979. Variations of the triangular method and the relation-
ship of its unidimentional probabilistic model to three-alternative forced
choice signal detection theory models. Brit. J. Math. Stat. Psych. 32,
229-24 1.
FRIJTERS, J.E.R. 1988. Sensory difference testing and the measurement of
sensory discriminability. In Sensory Analysis of Foods, 2nd Ed. (J.R.
Piggott, ed.) pp. 117-214, Elsevier Applied Science Publishers, London.
HOPKINS, J.W. and GRIDGEMAN, N.T. 1955. Comparative sensitivity of
pair and triad flavor intensity difference tests. Biometrics 11, 63-68.
O’MAHONY, M. and ODBERT, N. 1985. A comparison of sensory differenct
testing procedures. Sequential sensitivity analysis and aspects of taste
adaptation. J. Food Sci. 50, 1055-1058.
URA, S. 1960. Pair, triangle and duo-trio test. Reports of statistical application
research, Japanese Union of Scientists and Engineers 7, 107-1 19.

You might also like