Professional Documents
Culture Documents
Newell and MacFarlane
Newell and MacFarlane
Table 1-Critical absolute rank sum differences for “all treatments” com-
ABSTRACT parisons at 5% level of significancea,b
Multiple comparison procedures involving Friedman rank sums for Number of samples
the analysis of ranked data have recently been proposed as alternatives
Panelists 3 4 5 6 7 8 9 10 11 12
to tests developed by Kramer. Expanded tables for these multiple
comparison procedures (involving the “all treatments” and “treat- 3 6 8 11 13 15 18 20 23 25 28
4 7 10 13 15 18 21 24 27 30 33
ments versus control” comparisons) were presented to accomodate 37
5 8 11 14 17 21 24 27 30 34
the range of panelists and samples usually encountered in sensory 6 9 12 15 19 22 26 30 37 42
evaluation experiments involving ranked data. 7 10 13 17 20 24 28 32 9: 40 44
8 10 14 18 22 26 30 34 39 43 47
9 10 15 19 23 27 32 36 41 46 50
INTRODUCTION 10 11 15 20 24 29 34 38 43 48 53
11 11 16 21 26 30 35 40 45 51 56
TESTS developed by Kramer for assessingsamples that have 12 12 17 22 27 32 37 42 48 53 58
been ranked (Bradley and Kramer, 1957; Kramer, 1956,196O) 13 12 18 23 28 33 39 44 50 55
have been used extensively in sensory evaluation for the last 14 13 18 24 29 34 40 46 52 57 :i
15 13 19 24 30 36 42 47 53 59 66
30 years, with corrected and expanded tables available (Kahan 16 14 19 25 31 37 42 49 55 61 67
et al., 1973; Kramer, 1963). Computerized versions of this 17 14 20 26 32 38 44 50 56 63 69
methodology have been developed (McLellan et al., 1984), as 18 15 20 26 32 39 45 51 58 65 71
well as the tests discussed and tables of critical values pre- 19 15 21 27 33 40 46 53 60 66 73
20 15 21 28 34 41 47 54 68 75
sented in the standard sensory evaluation textbooks (Amerine 21 16 22 28 35 42 49 56 El 70 77
et al., 1965; Kramer and Twigg, 1970). 22 16 22 29 36 43 50 71 79
Joanes(1985) has recently suggestedthat these multiple com- 23 16 23 30 37 44 51 ii;: Ez 73 80
parison tests are invalid for their stated purpose becausethe rank 24 17 23 30 37 45 52 59 67 74 82
25 17 24 31 38 46 53 61 68 76 84
sums of different samples are not independent of one another 26 17 24 32 39 46 54 62 70 77 85
and the probability level for one of the tests is incorrect. The 27 18 25 32 40 47 55 63 71 79 87
resulting multiple comparison procedures proposed by Joanes 28 18 25 33 40 48 56 64 72 80 89
(1985) to overcome these weaknessesin Kramer’s tests utilize 18 26 33 41 49 57 65 73 82 90
ii 19 26 34 42 50 58 66 75 83 92
the distribution of the range of rank sums for each rank sum
configuration with these non-parametric multiple comparison i1 1: 9; i"j ff z1 z; ill ;9 ii i:
proceduresdiscussedin Hollander and Wolfe (1973), as well as 61 70 78 87 96
tables of critical values being available (Dunn-Rankin and Wil- ii 2 fS i: ii z3 62 71 79 89 98
35 20 28 37 45 54 63 72 81 90 99
coxon, 1966; Hollander and Wolfe, 1973; McDonald and 36 20 29 37 46 55 63 73 82 91 100
Thompson, 1967). However, these tables only include critical 37 21 29 38 46 55 64 74 92 102
values for up to 15 panelists and as such -are limiting for their 38 21 29 38 47 56 75 :: 103
effective use in sensoryevaluation experimentation. The purpose 39 21 30 39 48 57 :; 76 85 zz 105
76 86 96 106
of this paper was to expand the scope of the available tables for fY ;: i: :: ii 5": :: 77 87 97 107
use in multiple comparisons for ranked data to accomodatethe 42 22 31 40 49 59 69 78 88 98 109
range of panelists (up to 100) and samples (up to 12) usually 43 22 31 41 50 60 69 79 89 99 110
encounteredin sensory evaluations. 44 22 32 41 51 60 70 80 90 101 111
45 23 32 41 51 61 71 81 91 102 112
46 23 32 42 52 62 72 82 92 103 114
47 23 33 42 52 62 72 83 93 104 115
THEORY METHODS 48 23 33 43 53 63 73 84 94 105 116
“All treatments” comparisons 49 24 33 43 53 64 74 85 95 106 117
50 24 34 44 54 64 75 85 96 107 118 I
To test whether any of the samples are superior or inferior to any 55 25 35 46 56 67 78 90 101 112 124
of the other samples, Joanes (1985) proposed the use of a distribution- 60 26 37 48 59 70 82 94 105 117 130
free test based on Friedman’s rank sums (Hollander and Wolfe, 1973) 65 27 38 50 61 73 85 97 110 122 135
involving the distribution of the largest absolute range between rank 70 28 40 52 64 76 88 101 114 127 140
75 29 41 53 66 79 91 105 118 131 145
sums. For a sensory evaluation involving n panelists ranking k sam- 80 30 42 55 68 81 94 108 122 136 150
ples, this test is based on all (k!)” possible rank configurations being 85 31 44 57 70 84 97 111 125 140 154
equally likely, with two samples “a” and “b” considered to be sig- 90 32 45 58 72 86 100 114 129 144 159
nificantly different if: 95 33 46 60 74 88 103 118 133 148 163
100 34 47 61 76 91 105 121 136 151 167
I&- &I 2 4&n) (1) “Exact values adapted from Hollander and Wolfe (1973) are used for up to 15 panel-
ists.
where R,, Rb are the rank sums for samples “a” and “b”, respec- blnterpolation may be used for unspecified table values involving more than 50 panel-
tively; (Y = appropriate experimentwise error rate; and r(cx,k,n) = ists.
table value such that P((R,- Rt,l < r(cu,k,n)) = 1 --(Y.
While tables of critical values are available (Dunn-Rankin and Wil-
coxon, 1966; Hollander and Wolfe, 1973; McDonald and Thompson,
1967), these tables are limited becausethe true probability distribution
of the range of rank totals is generally unknown, resulting from dif- I
ficulties in obtaining a complete enumeration of the (k!)” possible
Authors Newell and MacFarlane are with Hawkesbury Agricul- rank configurations. For example, 10 panelists ranking four samples
tural College, Richmond NSW 2753, Australia
would involve enumerating (4.)t lo = 6.3 x lOI3 possible rank con-
“Treatments versus control” comparisons four samples (up to 5 panelists). Again, these tables are very
To test whether a specified sample (called the “control”) is limiting for use in sensory evaluation experiments. However,
superior to any of the remaining samples, Joanes (1985) pro- by using a large sample approximation developed by Nemenyi
posed the use of a distribution-free test based on Friedman’s (1963) and discussed by Hollander and Wolfe (1973), where
rank sums (Hollander and Wolfe, 1973) involving the distri- the control is significantly better than sample “a” if:
bution of the largest range between rank sums containing the
control. For this one-sided test, the control sample is consid- R,-R control2 (m(cw,k- 1,1/2). (mk(k+ 1)/b): (4)
ered to be significantly better than sample “a” if:
where m(a,k- 1,*/2) = upper (Y percentile point of the max-
R,- Rcontrol2 r* (o,k- Ln) imum of (k- 1) N(O,l) variables with common correlation r/2
as given in Table A.13 (Hollander and Wolfe, 1973), it is
where r * (cw,k- 1,n) = table value such that possible to expand the scope of the available tables to accom-
P Pa - kmtrol < r *(ol,k- 1,n)) = 1 -CL modate the usual range of panelists and samples encountered
in sensory evaluations. Table 5 and 6 present these expanded
Tables of critical values for this test are available (Hollander tables to enable one-sided “treatments versus control” multi-
and Wolfe, 1973) but are very limited becauseof the difficul- ple comparisons to be performed for sensory evaluations in
ties associatedwith the complete enumeration of the rank con- which up to 100 panelists rank up to 11 samples against the
figurations discussedin the previous section, resulting in tables control sample, with critical values available at (approxi-
only being available for three samples (up to 18 panelists) and mately) the 5% and 1% levels of signficance, respectively.
Similarly, to test whether the control sample is inferior to However, by using a large sample approximation developed
any of the remaining samples, R, - Rcontrolshould be replaced by Nemenyi (1963) and discussed by Hollander and Wolfe
by Rcontrol - R, in the above procedure. (1973), where the control is significantly different to sample
Alternatively, to test whether the control is different from “a” if:
any of the remaining samples, a two-sided test is available
(Hollander and Wolfe, 1973) involving the distribution of the
largest absolute range between rank sums containing the con- IRa - Rcontrol I L Irnl (o,k-l,%) . (nk(k+1)/6)$ (6)
trol. In this test, the control sample is considered to be signif-
where Iml(ol,k- 1,1/2)= upper OLpercentile point of maximum
icantly different to sample “a” if: absolute value of (k- l)N(O,l) variables with common cor-
1% - Lntroll 2 r** k-4 - l,n) relation 1/2as given in Table A.14 (Hollander and Wolfe, 1973),
(5) it is possible to expand the scope of the available tables to
where r** (a,k- 1,n) = table value such that accomodate the usual range of panelists and samples encoun-
tered in sensory evaluation. Table 7 and 8 present these ex-
P(IR,-R,,,,,,r 1 < r** (a,k-1,n)) = 1-o. panded tables to enable two-sided “treatments versus control”
multiple comparisons to be performed for sensory evaluations
Tables of critical values (Hollander and Wolfe, 1973) only in which up to 100 panelists rank up to 11 samples against the
cover the same range of panelists and samples available for control sample, with critical values available at (approxi-
the previously discussed one-sided tests and, as such, are too mately) the 5% and 1% levels of significance respectively.
restrictive for effective use in sensory evaluation experiments. As an example of these “treatments versus control” mul-
To test whether the control is different from any of the nol. 27: 61.
Kramer, A. 1956. A quick, rank test for significance of differences in mul-
remaining samples, Tables 7 and 8 give the critical absolute tiple comparisons. Food Technol. 10: 391.
rank sum differences to be 20 at the 5% level and 24 at the Kramer, A. 1960. A rapid method for determining significance of differ-
ences from rank sums. Food Technol. 14: 576.
1% level of significance, thus indicating that the control is Kramer, A. 1963. Revised tables for determining significance of differ-
significantly different from samples A and C at the 1% level ences. Food Technol. 17: 124.
of level of significance. Kramer, A. and Twigg, B. 1970. “Quality Control for the Food Industry.”
AVI Publishing Company, Westport, CT.
McDonald, B. and Thompson., W. 1967. Rank sum multiple comparisons
in one- and two-way classifications. Biometrika 54: 467.
McLellan, M., Way, R., and Lamb, C. 1984. An interactive computerized
CONCLUSION method for rank analysis. Hort. Sci. 19: 634.
Nemenyi, P. 1963. Distribution-free multiple comparisons. Unpublished
THE USE of the expanded tables of critical values presented Ph.D. dissertation, Princeton University.
MS received l/20/87; revised 3130187; accepted 4127187.
in this paper for multiple comparisons using ranked data have
been shown to be more complete and appropriate to the cir-
propriate amino acid source were available could sufficient of a caffeine-resistant mutant of Aspergillus parasiticus. J. Food Sci. 52:
194.
levels of precursor accumulation occur such that a significant Davis, N. D., Diener, U. L., and Eldridge, D. W. 1966. Production of af-
portion was shunted into aflatoxin production. Additional stud- latoxins B1 and G1 by Aspergillus flauus in a semisynthetic medium.
Apt. Microbial. 14: 378.
ies with BCRl that will be reported separately have indicated Dura ovlc, S.,, Durakovlc, Z., Beritic, T., Radic, B., Lalic, L. M., and Delas,
that caffeine does not directly affect aflatoxin synthesis, but F. 1985a. Biosynthesis of aflatoxins by Aspergillusparasiticus on roasted
instead acts by altering some aspect of primary metabolism. coffee beans. Periodicum Biologorum 87: 503.
Durakovic. S.. Durakovic. Z.. Lalic. L. M.. Posnisil. 0.. and Radic. B. 1985b.
However, additional studies will be needed to determine if the Influence of selected cultiiation para&t&s on thk growth df the toxi-
caffeine-dependent, amino acid-dependent nature of aflatoxin genie mold Aspergillus parasiticus on coffee beans and the biosynthesis
of aflatoxins. Microbioloev (Belmade) 22: 1.
synthesis in caffeine-resistant A. parasiticus involves a mech- Lenovich, L. M. 1981. Eff& of caffeine on aflatoxin production in cocoa
anism similar to that hypothesized above. beans. J. Food Sci. 46: 655.
Naik. M.. Modi. V. V.. and Patel. N. C. 1970. Studies on aflatoxin biosvn-
thesis in Asp&gill& flauw. I&an J. Exp. Biol. 8: 345.
Nartowicz, V. B., Buchanan, R. L., and Segall, S. 1979. Aflatoxin produc-
tion in regular and decaffeinated coffee beans. J. Food Sci. 44: 446.
REFERENCES Payne, G. A. and Hagler, W. M. Jr. 1983. Effect of specific amino acids on
growth and aflatoxin production by Aspergihs parasiticus and Asper-
Betancourt, L. E. and Frank, H. K. 1983. Bedingungen des mikrobiellen gillus fZauus. Appl. Environ. Microbial. 46: 805.
verderbs van grunem kaffee. Deut. Lebn.-Rund. 79: 404. Reddy, T. V., Viswanathan, L., and Venkitasubramian, T. A. 1979. Factors
Buchanan, R. L. and Fletcher, A. M. 1978. Methylxanthine inhibition of affecting aflatoxin production by Aspergillus parasrticus in chemically
aflatoxin production. J. Food Sci. 43: 654. defined medium. J. Gen. Microbial. 114: 409.
Buchanan, R. L., Hoover, D. G., and Jones, S. B. 1983. Caffeine inhibition Sahay, M. 1983. Effect of nitrogen sources on growth and aflatoxin pro-
of aflatoxin production: Mode of action. Appl. Environ. Microbial. 46: duction by Aspergillus arasticus (NRRL 3240). In “Proc. Symp. Myco-
1193. toxin in Food and Feed%, K S Bllerami. ed. Allied Press, Bhaealour.
Buchanan, R. L. and Houston, W. M. 1982. Production of blue-fluorescent India. p. 199.
pyrazines by Aspergillus parasiticus. J. Food Sci. 47: 779. MS received 2/10/87; revised 5/2/87; accepted 5/4/87.
Buchana?, R. L. and Lewis, D. F. 1984a. Caffeine inhibition of aflatoxin
synthesis: Probable site of action. Appl. Environ. Microbial. 47: 1216.
Buchanan, R. L. and Lewis, D. F. 1984b. Regulation of aflatoxin biosyn-
thesis: Effect of glucose on the activities of various glycolytic enzymes.
A pl. Environ. Microbial. 48: 306. Presented at the 47th Annual Meeting of the Institute of Food Technologists, Las
But K anan, R. L., Zaika, L. L., Stahl, H. G., and Mertz, S. E. 1987. Isolation Vegas, NV, June 1619, 1987.