Beruflich Dokumente
Kultur Dokumente
ABSTRACT
A sensory threshold can be defined generally as a stimulus intensity that
produces a response in half of the trials. The deflnirion of the population
threshold is discussed. Five main classical statistical procedures for estimating
thresholds are reviewed. They are the probit, the logistic, the Spearman-Karber,
the moving average and the up-and-abwn procedures. Some new developments in
statistical methodr for estimating thresholds are outlined. The newly developed
methoh include the generalized probit and logistic models, the model based on
the Beta-Binomial distribution, the trimmed Spearman-Karber method, the kernel
method and the sigmoidally constrained maximum likelihood estimation method.
The authors propose a new procedure based on the Beta-Binomial distributionfor
estimating population threshold.
INTRODUCTION
Sensory analysis methods can be divided into four categories: sensitivity,
quantitative, qualitative and affective (Pangborn 1984). Sensory threshold is a
measure of human sensitivity to a given stimulus. Determination of sensory
threshold is an essential element in sensory analysis and is important today for a
variety of purposes including the selection of panelists and the study of ingredient
variation limits in products (Meilgaard et al. 1991).
The aims of this paper are to discuss the concept of the threshold and to
review the classical and new statistical methods for estimating thresholds, some
of which are not well known among sensory practitioners. In addition, a
definition of the population threshold and a new model for estimating population
thresholds are given.
Two types of sensitivity tests exist: direct and indirect. In a direct sensitivity
Journal of Sensory Studies 13 (1998) 133-148. All Rights Reserved.
Topyright 1998 by Food & Nutrition Press, Inc.. Trumbull. Connecticut.
133
134
SENSORY THRESHOLDS
135
level is assumed not to have enough effect (detection or difference) on the sensory
system and, therefore, cannot be perceived. The intuitive appeal of this idea is
that there must be some absolute value below which the sensitivity of the sensory
system does not permit detection. Optimally, the threshold is thought of as a
sharp transition point between sensation and no sensation. Inherent in the idea of
a threshold is the assumption that the transition point is independent of conditions.
In practice, the response of the system is affected by many psychological and
physiological inputs, and shifts in the transition point (if it exists) may occur.
This makes measurement of the threshold difficult and the transition point difficult
to be defined. The problem is solved by treating the threshold as a statistical
concept. The threshold is seen as a random variable that varies both between
observers and within a given observer across time. Empirical data demonstrate
that the detectability of stimuli does not follow a step function in which detection
jumps from 0 to 100% at some special value. In fact, the probability of detection
increases gradually when the intensity of a stimulus increases. Thresholds often
are defined as a stimulus intensity that will produce a response in half
the population, i.e., in 50% of the trials. In a bioassay, the threshold is
expressed as the median effective dose (ED,,) or the median lethal dose (LD,,),
using a specified response or death in 50% of the population,
Although the threshold cannot be defined as the stimulus intensity below
which detection never occurs and above which detection always occurs (a
transition point), the concept of the threshold is useful because it affords a
technique for quantifying the sensitivity of an individual or of a specified
population. This information can be used to evaluate individual sensitivity and to
establish quality control limits in food production or in pollution control.
Moreover, it becomes possible to estimate points of interest other than the 50%point, for example, the stimulus intensity that an individual can detect 90% of the
time, or the stimulus intensity that can be detected by 1% of a given population.
136
SENSORY THRESHOLDS
137
the Moving average method are used widely for their theoretical and practical
merits.
Graphic Method.
Estimation of Threshold. Convert the percent correct above chance, P, to
probits Y. Y = zp 5, where zp is P percentile of standard normal distribution.
i = 1,2,. ..k on ordinary rectangular coordinate graph paper,
Plot points ( X I ,
or on normal probability graph paper. Here, X I is stimulus intensity (or log
stimulus intensity). Draw a straight line by computer or by eye to fit the k points.
Read off threshold (or log threshold), the value of stimulus intensity, X,,
corresponding to Y=5 (or P=0.5 on normal probability graph paper).
Estimation of the Confidence Interval of the Threshold. Calculate 6 , the slope
of the fitting line. Mark two convenient points, (Xc, Y,) and (Xd, Y,) on the line,
The equation of the fitted line is as Y=5+b(X-X,). Read
then b=(Y,Y,)/X,-X,).
off r, corresponding to X,from the fitted line (or calculate from the equation of
the fitted line). Read off the weighting coefficients, w, corresponding to Y, (from
Table A-18 (Natrella 1963)). The approximate standard error of X , can be
obtained from the equation: JV(X,)= {b2(Xnw)}-2.
The upper and lower 95%
confidence intervals of threshold can be obtained by the equation:
x),
Exact Method. The graphical solution often is adequate, but the exact
method may be necessary. For example, the points may be too irregular for us
to place any confidence in a line drawn by eye; or, the weights, that should be
attached to each point, may be so different as to make it difficult to adjust for
them visually.
The exact probit method uses a weighted iterative procedure to fit a straight
line. The threshold and its confidence intervals can be estimated on the basis of
138
139
SENSORY THRESHOLDS
The method also requires a wide range of stimulus levels. In other words,
the lowest stimulus X, must be sufficiently low that there are no responses among
the objects tested and the highest stimulus X, must be sufficiently high that all
objects tested respond. A rule sometimes stated is that if the proportion of
responses at the lowest stimulus PpO or the proportion of responses at the highest
stimulus P,+ 1, the next stimulus level in the series, though untested, should be
assumed to have given P,=O or P,+,= 1; the estimation is then completed as for
the longer series (Finney 1978).
For
The log threshold can be estimated by rn=~,=,k{(Pl+,-P,)(X,+X,+,)/2}.
The
equally spaced log stimulus levels, X,+,-Xl = d, rn = X, +O.Sd-c,=,
l * - l )n*=(l-C)n,
~.
C= 1/3
estimation of error is V(m) = ~ ~ l = ~ { P l { l - P , ) / ( ~Where
in the 3-AFC procedure. The 95 percent confidence limits for rn are mu =
m+1.96JV(rn) and m, = rn-1.96JV(rn).
'e.
140
SENSORY THRESHOLDS
141
142
[J
TI
('1'
r(a+x)r(b+n-x)r(a+b)
r(a + b + a)r(a)r(b)
SENSORY THRESHOLDS
143
x,
logL
4, -1
)I,,
-x,, -I
11 C l o g ( p , + r e , > + C
I=]
Ill,
,=I
I$
log(l-p, + r e , ) - C l o g ( l + r B , ) ,
r=O
r=O
r=O
-I
PI log 4WV
where Po and PI are regression parameters; log d , is log dose at the ith dosage
step. The Newton-Raphson procedure can be used to get the maximum likelihood
estimates of parameters , Po, PI and 8 and the covariance matrix of Po, PI, and
8 . Then the log LD,, is estimated from log LD 5o =
. The confidence
intervals for log LD,, with confidence probability 1-a,is given by the set of x
satisfying M(xo) <z2., where
-Po@,
WX,)
=(Po+xoPI)2(v,,+2x,v,,
,,
+X02VJL
and z2, is the upper a point of x2 with 1 degree of freedom, vII is the variance of
vZ2is the variance of P, and vI2 is the covariance of Po and P I .
Po,
144
SENSORY THRESHOLDS
145
The main advantage of the kernel method is its flexibility, which makes it a
valuable exploratory tool in sensitivity tests.
SUMMARY
The sensory threshold is defined as a stimulus intensity that produces a
response in half of the trials. The definition of population threshold based on a
single compound sampling population has been discussed. Five classical
procedures for estimating threshold were reviewed. They are the probit, the
logistic, the Spearman-Karber, the moving average and the up-and-down
procedures. The latest developed statistical models for estimating threshold were
outlined. They are the generalized probit and logit models, the dose-response
model based on the Beta-Binomial distribution, the trimmed Spearman-Karber
method, the kernel method and the sigmoidally constrained maximum likelihood
estimation method. A new procedure based on the BB distribution for estimating
population threshold is proposed.
ACKNOWLEDGMENTS
The authors would like to thank the editor and the referees for constructive
comments and suggestions on an earlier manuscript. Significant initial discussions
were provided by Dr. E. Chambers IV and Dr. M. Meilgaard.
146
REFERENCES
ANDERSON, D.A. 1988. Some models for overdispersed binomial data.
Austral. J. Statist. 30(2), 125-148.
ASTM Committee E-18. 1990. Standard definitions of terms relating to sensory
evaluation of materials and products. In Annual Book of A5734 Standards. E253-90. American Society for Testing and Materials, Philadelphia, PA.
ASTM Committee E-18. 1991. Standard practice for defining and calculating
individual and group sensory thresholds from forced-choice date sets of
intermediate size. In Annual Book of ASTM Standards. E-1432-91.
American Society for Testing and Materials. Philadelphia, PA.
ASTM Committee E-27. 1984. Standard test method for drop weight impact
sensitivity of solid-phase hazardous materials. In Annual Book of ASTM
Standards. E-680-84. American Society for Testing and Materials,
Philadelphia, PA.
BENNETT, B.M. 1952. Estimation of the LD50 by moving averages. J.
Hygiene SO, 157-164.
BENNETT, B.M. 1963. Optimum moving averages for the estimation of median
effective dose in bioassay. J. Hygiene 62, 401-406.
BERKSON, J. 1944. Application of the logistic function to bioassay. J. Am.
Stati. Assoc. 39, 357-365.
BERKSON, J. 1955. Maximum likellhood and minimum x2 estimates of the
logistic function. J. Am. Stati. Assoc. SO, 206-209.
BLISS, C.I. 1934. The method of probits. Sci. 79, 38-39.
BLISS, C.I. 1935a The calculation of the dosage-mortality curve. AM. Appl.
Biol. 22, 134-167.
BLISS, C.I. 1935b. The comparison of dosage-mortality data. AM. Appl. Biol.
22, 307-333.
BLISS, C.I. 1938. The determination of the dosage-mortality curve from small
number. Quart. J. Pharmacol. 11, 192-216.
COPAS, J.B. 1983. Plottingp against x. Applied Statistics 32, 25-31.
CORNSWEET, T.N. 1962. The staircase method in psychophysics. Am. J.
Psychol. 75, 485-491.
COX, D.R. 1983. Some remarks on overdispersion. Biometrika 70(1), 269-274.
DIXON, W.J. and MASSEY, F.J. 1957. Zntroducrion to Staristical Analysis.
McGraw-Hill.
DIXON, W.J. and MOOD, A.M. 1948. A method for obtaining and analyzing
sensitivity data. J. Am. Stati. Assoc. 43, 109-126.
DRAKE, B. 1975. A fortran program, sigmplot, for fitting sigmoid curves to
threshold data and for plotting the result. Chemical Senses and flavor I , 519531.
SENSORY THRESHOLDS
147
ENNIS, D.M. and BI, J. 1998. The beta-binomial model: Accounting for interintra1 variation in replicated difference and preference tests. J. Sensory
Studies (In press).
FINNEY, D.J. 1950. The estimation of the mean of a normal tolerance
distribution. Sankhya 20, 341-360.
FINNEY, D.J. 1971. Probit Analysis 3rd Ed., Cambridge University Press.
FINNEY, D.J. 1978. Statistical Method in Biological Assay, Charles Griffin &
Co., Ltd.
HAMILTON, M.A. 1979. Robust estimates of the ED,,. J. Am. Stati. Assoc.
74, 344-354.
HAMILTON, M.A., RUSSO, R.C. and THURSTON, R.V. 1977. Trimmed
Spearman-Karber method for estimating median lethal concentrations in
toxicity bioassay. Evir. Sci. & Technol. 11, 714-719.
KALISH, L.A. 1990. Efficient design for estimation of median lethal dose and
quantal dose-response curves. Biometrics 46, 737-748.
KAPPENMAN, R.F. 1987. Nonparametric estimation of dose-response curves
with application to ED,, estimation. J. Stati. Computation & Stimulation 28,
1-13.
KARBER, G. 1931. Beitrag zur kollectiven behandlung pharmakologischer
reihenversuche. Arch. Exp. Path. Pharmak 262, 480-487.
KUPPER, L., PORTIER, C., HOGAN, M. andYAMAMOTO, E. 1986. The
impact of litter effects on dose-response modeling in teratology. Biometrics
42, 85-98.
MEILGAARD, M., CIVILLE, G.V. and CARR, B.T. 1991. Sensory Evaluation
Techniques. 2nd Ed, CRC Press.
MILLER, R.G. and HALPERN, J.W. 1980. Robust estimators for quantal
bioassay. Biometrika 67, 103-110.
MULLER, H.G. and SCHMITT, T. 1988. Kernel and probit estimates in
quantal bioassay. J. Am. Stati. Assoc. 83, 750-759.
NATRELLA, M.G. 1963. Experimental Statistics, National Bureau of Standards
Handbook 91.
PANGBORN, R.M. 1984. Sensory techniques of food analysis. In Food
Analysis: Principles and Techniques, Vol. 1, (D.W. Gruenwedel and J.R.
Whitaker, eds.)
POWERS, J.J. and WARE, G.O. 1976. Comparison of sigmplot, probit, and
extreme-value methods for the analysis of threshold data. Chemical Senses
and Flavor 2, 241-253.
PRENTICE, R.L. 1976. Generalization of the probit and logit methods for dose
response curves. Biometrics 32, 761-768.
SCHMOYER, R.L. 1984. Sigmoidally constrained maximum likelihood
estimation in quantal bioassay. J. Am. Stati. Assoc. 79, 448-453.
148
SEGRETI, A.C. and MUNSON, A.E. 1981. Estimation of the median lethal dose
when responses within a litter are correlated. Biometrics 37, 153-156.
SPEARMAN, C. 1908. The method of right and wrong cases (constant stimuli)
without Gausss formulae. Brit. J. Psychol. 2, 227-242.
STANISWALIS, J.G. and COOPER, V. 1986. Kernel estimates of dose
response. Working Papers, Series 2 1, Virginia Commonwealth University.
(See Biometrics 44,(1988) 1103-1119).
STUKEL, T.A. 1988. Generalized logistic models. J. Am. Stati. Assoc. 83,
426-43 1.
THOMPSON, W.R. 1947. Use of moving averages and interpolation to estimate
median effective dose. I: Fundamental forrnuias, estimation of error, and
relation to other methods. Bacteriol. Rev. 11, 116-145.
WEIL, C.S. 1952. Tables for convenient calculation of median-effective dose
(LD50 or ED50) and instructions in their use. Biometrics 8, 249-263.
WILLIAMS, D.A. 1982. Extra-binomial variation in logistic linear models.
Appl. Statisti. 3I(2), 144-148.
WILLIAMS, D.A. 1986. Interval estimation of the median lethal dose.
Biometrics 42, 641-645.