Anal. Chem. 1991. 63.

139-146 139

tetradecanol) are separated and detected, quantitation is and reversed-phase mechanism. Figure 9 shows that the
hindered by poor solubility and the formation of micelles. p-toluenesulfonic acid and the cefazolin are both detected by
Applications. Since a prerequisite for PAD reactivity is UV a t 254 nm (A), and the two diols, which do not have a
adsorption of the analyte, the presence of other surface-ad- chromophore, are easily detected by PAD (B). In addition,
sorbable substances, as well as electroactive compounds, can the cefazolin has a PAD signal, which has a higher limit of
act as interferences. Therefore, general selectivity is achieved detection than UV, but may be utilized for added selectivity.
via chromatographic separation prior to PAD. This conclusion
does not preclude additional selectivity from control of de- LITERATURE CITED
tection parameters. Kisslnger, P. T. I n Laboratory Techniques in Electroanalytical Chemis-
The assay for alcohols was applied to several matrices to try; Klssinger, P. T., Heineman, W. R., Eds.; Marcel Dekker: New
York, 1984; pp 631-32.
illustrate the analytical utility of the procedure. Separation Adams, R. N. Electrochemistry at SolM Electrodes; Marcel Dekker:
on an ion-exclusion column with direct detection is illustrated New York, 1969.
Gllman, S . I n E/ectroana/ytica/ Chemistry; Bard, A. J., Ed.; Marcel
in Figure 8 for various aliphatic alcohols and polyalcohols in Dekker: New York, 1967; Vol. 2, pp 111-92.
toothpaste (A), liquid cold formula (B), brandy (C), and wine Fleet, B.; Little, C. J. J. Chromtogr. Sci. 1874, 72, 747.
cooler (D). The selectivity for alcohols in acidic media at a Van Rooljen, H. W.; Poppe. H. Anal. Chim. Acta 1881, 730,9.
Hughes, S.; Meschi, P. L.; Johnson, D. C. Anal. Chim. Acta 1881.
Pt electrode contributes to decreased time for sample prep- 732,1.
aration and simplified chromatograms. Johnson, D. C.; Lacourse, W. R. Anal. Chem. 1890. 62,589A.
Nachtmann, F.; Budna. K. W. J. Chromtogr. 1977, 736,279.
The versatility of separations on mixed-mode ion-exchange Jupille, T. J. Chromatogr. Sci. 1978, 77, 160.
columns with selective detection is illustrated in Figure 9 by Beden, B.; Cetin, I.; Kahyaoglu. D.;Takky, D.; Lamy, C. J. Catal.
the simultaneous detection of ionic and neutral species in a 1987, 704, 37.
Ocon, P.; Alonso, C.; Celdran, R.; Gonzalez-Velasco, J. J. Electroanal.
pharmaceutical preparation. This experiment utilizes a UV Chem. 1888, 206, 179.
detector and PAD in series after a PCX-500 column. Under Slingsby, R. W.; Rey, M. J. Li9. Chromatogr. 1880, 73(1), 107.
Lacourse, W. R.; Jackson, W.A.; Johnson, D. C. Anal. Chem. 1888,
acidic conditions, the cephalosporin antibacterial consists of 67, 2486.
a cation (i.e., cefazolin) and neutral and anionic compounds
(i.e., lB-hexanediol, 1,4-cyclohexanediol,and p-toluenesulfonic
acid). The neutral and anionic compounds are separated by RECEIVED for review July 20,1990. Accepted October 8,1990.
the reversed-phase character of the column, while the cationic The financial support of Dionex Corporation is acknowledged
compound is separated by a combination of cation exchange with gratitude.

Statistical Treatment for Rejection of Deviant Values: Critical

Values of Dixon’s “ Q ” Parameter and Related Subrange Ratios
at the 95% Confidence Level
David B. Rorabacher
Department of Chemistry, Wayne State University, Detroit, Michigan 48202

CrHlcal values at the 95% Confidence level for the two-tailed is interesting to note that these treatises, as well as essentially
0 test, and related tests based upon subrange ratlos, for the all analytical chemistry textbooks published in the U.S.during
statlstkal rejectlon of outlying data have been Interpolated by the past decade (3),have settled on the use of Dixon’s Q test
applying cublc regresslon analysls to the values orlglnally (and variants thereof) ( 4 ) as the primary method for testing
published by Dlxon. Corrections to errors In Dixon’s orlglnal for the rejection of outlying values.
tables are also Included. The resultant values are judged to Each of the recent statistical treatises written for analytical
be accurate to wlthln f0.002 and corroborate the fact that chemists has attempted to include critical values of Q for the
correspondlng crltlcal values published In recent slatlstlcal 95% confidence level, values that were not included in Dixon’s
treatlses for analytical chemlsts are erroneous. I t Is recom-
publications. However, not only do the 95% confidence values
differ in each treatise but all compilations contain significant
mended that the newly generated 95% crltlcal values be
errors. The most legitimate set of 95% values is that presented
adopted by analytical chemlsts as the general standard for
by Miller and Miller (4 In I10) ( l a ) ,which they attribute
the rejection of outller values. to King (5),but no such values are listed in King’s article, and
the values of Miller and Miller differ by amounts varying from
0.002 to 0.007 from the 95% values presented in the current
Analytical chemists depend upon the generation and in- manuscript. Anderson ( I b ) describes the equations corre-
terpretation of precise experimental data. As a result, they sponding to the two-tailed tests for Dixon’s parameters
are especially cognizant of the value of statistics in data designated as rl0 (for 3 5 n 5 lo), rll (for 8 5 n Ilo), and
treatment, and a number of statistical treatises have recently rZ1(for 11 In I13) and purportedly lists critical values for
been published that are specifically written for the professional the 90%,95%, and 99% confidence levels for these sample
analytical chemist ( I ) . Included in each of these publications sizes, but the values actually listed in his table are Dixon’s
is a brief section dealing with tests for the rejection of grossly values for one-tailed tests. Thus,as applied to two-tailed tests,
deviant values (outliers). Although many statistical tests have Anderson’s confidence levels should be labeled 80%, 9070, and
been proposed to deal with this topic [Barnett and Lewis (2) 98% (vide infra). Caulcutt and Boddy (IC), while describing
discuss 47 different equations designed for this purpose], it only the equation for the Q (i.e., rl0)ratio, accurately list both
0003-2700/91/0363-0139$02.50/0 0 1991 American Chemical Society
140 ANALYTICAL CHEMISTRY, VOL. 63, NO. 2 , JANUARY 15, 1991

95% and 99% confidence level critical values for this ratio such values can be attributed to random variation alone within
for five sample sizes, 3 5 n 5 7 ; the values given in their Table some reasonable level of probability. If the probability of
G (p 248) for 8 5 n 5 12 and for 13 In 5 40 are clearly not obtaining such outliers is determined to be very small, based
Q values but are somewhat similar to (though not identical on the overall population variance, one may reasonably con-
with) the critical values for rI1and r21,respectively. However, clude that this hypothesis is incorrect, i.e., these values cannot
the equations /or these latter parameters are not euen be attributed to simple random variation and are likely the
mentioned in this treatise. result of a determinate error. On this basis, suspected values
Faced with this confusing and conflicting array of critical may be rejected in a statistically legitimate fashion, provided
values for the popular Q test and Dixon’s related tests based that the confidence level chosen is a reasonable one.
on various subrange ratios, it seems likely that many erroneous Confidence levels commonly used for tests involving the
statistical inferences are being made daily in analytical rejection of data are 99%, 95% or 90%, the higher levels
chemistry laboratories throughout the world. To counteract representing the more conservative approach. Operating a t
this possibility and to extend the utility of Dixon’s parameters a high confidence level reduces the likelihood of rejecting a
for two-tailed tests, the current paper provides a reliable set legitimate value containing no determinate error (i.e., it re-
of critical values a t the 95% confidence level for the Q pa- duces the so-called a risk or “Type I error”), but it also in-
rameter and for all of the related subrange ratios proposed creases the likelihood of retaining a value that contains a
by Dixon. I t also corrects several errors found in Dixon‘s determinate error (Le., it increases the so-called /3 risk or “Type
original published tables. At the same time, it is intended I1 error”) ( I O ) . The consensus of most statisticians is that the
to alert analytical chemists to the problems that exist in former error (rejecting a “good” value) is considered more
current treatises and, hopefully, may serve as a useful source serious and, therefore, the use of high confidence levels in the
for future textbook and treatise authors dealing with this topic. rejection of data is greatly preferred. For example, when
operating a t the 99% confidence level ( a = 0.01) with a
CRITERIA FOR REJECTION OF OUTLIERS population containing no outliers, only 1% of the good values
Determinate Errors and Outliers. In any set of ana- will be rejected. In contrast, when operating a t the 90%
lytical measurements, one or more values may incorporate a confidence level ( a = 0.10), 10% of the good values will be
determinate (systematic) error. Such errors may involve either rejected and, for small sample sizes, the effect upon the es-
a “locator” error (displacement of the population mean) or timation of the population mean becomes significant relative
a “scalar” error (increase in the population variance) or a to the accuracy desired by most analytical procedures.
combination of the two ( 4 ) . Values that are affected by a [However, Dixon has argued that the use of rejection tests a t
determinate error belong to a different population of mea- the 9 0 7 ~confidence level is warranted in many cases ( I I ) . ]
surements. Thus, if the presence of a determinate error is Since the use of a 99% confidence level ( a = 0.01) allows
known to the individual making the measurements, the values for the rejection of only the most extreme deviations-
so affected should automatically be rejected, regardless of their particularly when applied to small sample sizes-the /3 risk
magnitude. (For a brief discussion on the detection and of retaining a value that does not belong to the general sample
elimination of determinate errors in analytical measurements, population (Le., a “bad” value) is quite large, and most ex-
see ref 3b.) However, in many instances, determinate errors perimentalists are not satisfied when operating a t this level
are incurred that are not detected by the experimentalist. in testing for the rejection of data. Operating a t the 95%
If an undetected determinate error affects only a fraction confidence level ( a = 0.05) provides a reasonable compromise.
of the values in the total sample and if the error is sizable If values not rejected at the 95% confidence level are viewed
relative to the experimental precision of the remaining values, with suspicion, additional measurements are probably war-
the affected values may be significantly larger or smaller than ranted. It should be noted that no specific test or choice of
the values that are not affected by such errors; ;.e., they will confidence level is ideal, however. As Natrella has noted, “the
be observed as outlier values. Recognizing this, experimen- only sure way to avoid publishing any ‘bad’ results is to throw
talists tend to suspect the presence of determinate errors away all results” (12). The converse, of course, is also
whenever outliers are observed in a set of data, and there is true-the only sure way to avoid discarding any good results
a strong motivation to eliminate such outliers since they is to retain them all.
significantly affect the sample mean and, therefore, the final A number of tests, such as those based on the “Student’s
estimate of the population mean. In the absence of inde- t distribution” (13), require independent knowledge of the
pendent knowledge that determinate errors are responsible population standard deviation (cr) and the population mean
for the appearance of outliers, however, the decision to delete ( k ) or the uncontaminated sample standard deviation (s) and
deviant values must be based upon reasonable statistical the uncontaminated sample mean (8) for their application
evidence. Some statisticians look askance at any attempt to (2). In the absence of such independent information, these
discard deviant values by means of statistical criteria. Deming, tests are of limited use since a decision must be made as to
a highly respected authority in the industrial application of whether the suspected values should be included or excluded
statistics, has concluded that “a point is never to be excluded in the calculation of s and X . Neither decision is satisfactory
on statistical grounds alone” (6). However, the majority of as the choice made automatically biases the outcome of the
those who have addressed themselves to the statistical test. [The often-cited “Chauvenet’s criterion” for the rejection
treatment of scientific measurements would appear to agree of a deviant measurement (2, 7) is also based on the sample
with Parratt’s statement that “rejection on the basis of a hunch standard deviation and, therefore, poses the same problems
or of general fear is not a t all satisfactory, and some sort of when applied to small samples; it has been noted that this
objective criterion is better than none” ( 7 ) . In fact, the sta- criterion is set to reject, on average, half an observation of good
tistical treatment of outlier values has received increasing data per sample, regardless of the sample size (8,14).] For
attention in recent decades and the various statistical tests analytical measurements, where cr and g (and uncontaminated
that have been proposed are described and compared in texts values of s and X ) are generally not known independently,
(2,8) and an extensive review (9) dealing specifically with this tests that do not require the use of these quantities are greatly
topic. preferred.
Tests for Rejection of Grossly Deviant Values. Sta-
tistical approaches used for identifying values that are affected APPLICATION OF DIXON’S RANGE TESTS
by determinate errors are based on testing the hypothesis that The Q Test and Related Subrange Ratios. In a classic
 ANALYTICAL CHEMISTRY, VOL. 63, NO. 2, JANUARY 15, 1991 141

1950 article ( 4 ) ,Dixon investigated the performance of several The rl0 ratio is commonly designated as "Q" and is generally
statistical tests in terms of their ability to reject bad values considered to be the most convenient, legitimate, statistical
in data sets taken from Gaussian populations. The tests test available for the rejection of deviant values from a small
investigated included both those which require independent sample conforming to a Gaussian distribution. (It is equally
knowledge of u or s and those which do not require such well suited to larger data sets if only one outlier is present.)
information. Of the tests included in the latter group, Dixon The fact that small data sets are common in analytical testing
concluded that tests based on ratios of the range and various procedures, in combination with the simplicity of this test,
subranges were to be preferred as a result of their excellent accounts for the fact that the Q test is included in nearly all
performance and ease of calculation. [Dixon also noted that modern statistical treatises and textbooks designed for use
another test which performs well in screening for outliers is in analytical chemistry (1, 3).
a modified F test, in which the ratio of the standard deviations A few authors (17), following Deming's viewpoint (vide
calculated by including and excluding the suspected deviant supra), object to the rejection of data from any small sample
value is compared to critical values of F; however, this latter based on statistical tests, claiming that the amount of in-
test may be "masked" by a second deviant value.] The range formation available is insufficient to establish the distribution
tests, all of which are closely related, include the following pattern; recommended alternatives include dropping the
(where the values are ordered such that x1 < x 2 < ... < x,-~ highest and lowest values (for a sample containing five or more
< x,): values) or reporting the median. However, this would appear
1. For a single outlier x1 to be overly cautious since most series of repetitive analytical
measurements follow a Gaussian distribution (18) provided
that s is small compared to X (thus, the Q test is not applicable
to analytical measurements when operating close to the de-
2. For outlier x1 avoiding x, tection limits). Dixon has tested the relative merits of the
sample mean and median as an estimator of the population
mean under various conditions (11,15)and has concluded that,
for the most part, the sample mean (after the rejection of
outliers) appears to provide a better approximation than does
3. For outlier x1 avoiding x,, x,-, the median.
In his original calculation of the critical values of the various
r criteria (1.9, Dixon was able to obtain exact solutions only
for the case where n = 3 or 4. Critical values for n = 5 , 7 , 10,
4. For outlier x l avoiding x2 15,20,25, and 30 were calculated by using numerical methods.
All other values were obtained by interpolation and were
generally judged to be accurate within fO.OO1.
An irritating feature of the Q test and Dixon's related su-
brange ratio tests, as they currently exist, is the lack of suitable
5. For outlier x1 avoiding x2 and x, critical values of Q (and rll, r21,etc.) a t the 95% confidence
level (a = 0.05), since this confidence level is frequently utilized
for all other statistical tests. The lack of 95% confidence level
values arises from the fact that Dixon generated critical values
a t several standard probability levels ( a = 0.005, 0.01, 0.02,
6. For outlier x1 avoiding x2 and x,, 0.05, etc.) corresponding to 99.5%, 99%, 98%, 95%, etc.,
confidence levels in terms of a one-tailed test (15). As com-
monly applied by analytical chemists and other experimen-
talists, however, these tests are used as two-tailed tests (i.e.,
(The parenthetical equations are designed for testing x,, one is generally interested in testing outlier values at both the
the highest value rather than the lowest value, xl.) In Dixon's upper and lower ends). The probability that an individual
notation, the first digit in the subscript of each ratio, rij, refers good value may lie outside a specified interval in either tail
to the number of possible suspected outliers on the same end (Le., on either the high or low end) is twice as large as the
of the data as the value being tested, while the second digit probability that it lies outside the chosen interval in only one
indicates the number of possible outliers on the opposite end tail. As a result, the a risk doubles for a two-tailed test (18)
of the data from the suspected value. Thus, the ratio rl0 and the confidence level decreases accordingly. Thus, Dixon's
simply compares the difference between a single suspected 99.5%, 99%, 98%, and 95% confidence levels translate into
outlier (xl or x n ) and its nearest-neighboring value to the 99%, 98%, 96%, and 90% levels, respectively, when consid-
overall range of values in the sample-in other words, it de- ering a two-tailed test, a fact not recognized by some authors
termines the fraction of the total range that is attributable ( I b , 3e,m). Since Dixon did not calculate one-tailed critical
to one suspected outlier. The other ratios are similarly for- values at the 97.5% confidence level, there have been no values
mulated except that they use subranges that are specifically available a t the 95% level when using these tests as two-tailed
designed to avoid the influence of additional outliers either tests. Therefore, about half of the current analytical chemistry
on the opposite end of the data (rll and r12),on the same end textbooks (3a-d,h) listing valid critical values of Q list the
of the data (rm),or both (rZ1and rZ2). Clearly, the latter ratios values for the 99%,96%, and 90% confidence levels, despite
require larger sample sizes to perform satisfactorily. Dixon the fact that the 96% confidence level is somewhat unor-
subsequently generated critical values for all of these ratios thodox, while the remainder (3fg,i-l) provide only 90%
(15)for sample sizes of 3 In 5 30 and recommended (based confidence level values. As noted earlier, the treatises that
on a combination of the relative performance of each ratio have attempted to present critical values for Q at the 95%
and its degree of independence from other outlying values) confidence level (I) have invariably listed flawed values.
that, as a general rule, the various ratios be applied as follows [In tests designed to detect the presence of a single outlier,
(16): for 3 In 5 7, use rlo;for 8 In I10, use r l l ; for 11 I King has argued (5) that the effect of running a two-tailed
n I13, use rZl;for n L 14, use rZ2. test is approximately to double the a risk relative to that for
142 ANALYTICAL CHEMISTRY, VOL. 63, NO. 2, JANUARY 15, 1991

Table I. Critical Values of Dixon's r I o( Q ) Parameter As Applied to a Two-Tailed Test at Various Confidence Levels,
Including the 95% Confidence Level"
confidence level
80 7c 90 7c 95% 96 o/o 98% 99 ?&
N b ( a = 0.20) ( a = 0.10) ( a = 0.05) ( a = 0.04) ( a = 0.02) ( a = 0.01)

3 0.886 0.941 0.970 0.976 0.988 0.994

4 0.679 0.765 0.829 0.846 0.889 0.926
5 0.557 0.642 0.7 10 0.729 0.780 0.821
6 0.482 0.560 0.625 0.644 0.698 0.740
?.
, 0.434 0.507 0.568 0.586 0.637 0.680
8 0.399 0.468 0.526 0.543 0.590 0.634
9 0.370 0.437 0.493 0.510 0.555 0.598
10 0.349 0.412 0.466 0.483 0.527 0.568
11 0.332 0.392 0.444 0.460 0.502 0.542
12 0.318 0.376 0.426 0.441 0.482 0.522
13 0.305 0.361 0.410 0.425 0.465 0.503
14 0.294 0.349 0.396 0.411 0.450 0.488
15 0.285 0.338 0.384 0.399 0.438 0.475
16 0.277 0.329 0.374 0.388 0.426 0.463
17 0.269 0.320 0.365 0.379 0.416 0.452
18 0.263 0.313 0.356 0.370 0.407 0.442
19 0.258 0.306 0.349 0.363 0.398 0.433
20 0.252 0.300 0.342 0.356 0.391 0.425
21 0.247 0.295 0.337 0.350 0.384 0.418
22 0.242 0.290 0.331 0.344 0.378 0.411
23 0.238 0.285 0.326 0.338 0.372 0.404
24 0.234 0.281 0.321 0.333 0.367 0.399
25 0.230 0.277 0.3 17 0.329 0.362 0.393
29 0.227 0.273 0.312 0.324 0.357 0.388
2; 0.224 0.269 0.308 0.320 0.353 0.384
28 0.220 0.266 0.305 0.316 0.349 0.380
29 0.218 0.263 0.301 0.312 0.345 0.376
30 0.215 0.260 0.298 0.309 0.341 0.372
In this and the other accompanying tables, the newly generated or corrected values are indicated in boldface. Sample size.

Table 11. Critical Values of Dixon's r l lParameter As Applied to a Two-Tailed Test at Various Confidence Levels, Including
the 95% Confidence Level
confidence level
80 % 90 % 95 % 96 Yo 98% 99%
N O ('N = 0.20) ( a = 0.10) ( a = 0.05) ( n = 0.04) (1 cr = 0.02) ( a = 0.01)

4 0.910 0.955 0.977 0.981 0.991 0.995

5 0.728 0.807 0.863 0.876 0.916 0.937
-
6
I
0.609
0.530
0.689
0.610
0.748
0.673
0.763
0.689
0.805
0.740
0.839
0.782
8 0.479 0.554 0.615 0.631 0.683 0.725
9 0.441 0.512 0.570 0.587 0.635 0.677
10 0.409 0.477 0.534 0.551 0.597 0.639
11 0.385 0.450 0.505 0.521 0.566 0.606
12 0.367 0.428 0.481 0.498 0.541 0.580
13 0.350 0.410 0.461 0.477 0.520 0.558
14 0.336 0.395 0.445 0.460 0.502 0.539
15 0.323 0.381 0.430 0.445 0.486 0.522
16 0.313 0.369 0.4 17 0.432 0.472 0.508
17 0.303 0.359 0.406 0.420 0.460 0.495
18 0.295 0.349 0.396 0.410 0.449 0.484
19 0.288 0.341 0.386 0.400 0.439 0.473
20 0.282 0.334 0.379 0.392 0.430 0.464
21 0.276 0.327 0.371 0.384 0.421 0.455
22 0.270 0.320 0.364 0.377 0.414 0.446
23 0.265 0.314 0.357 0.371 0.407 0.439
24 0.260 0.309 0.352 0.365 0.400 0.432
25 0.255 0.304 0.346 0.359 0.394 0.426
26 0.250 0.299 0.341 0.354 0.389 0.420
27 0.246 0.295 0.337 0.349 0.383 0.414
28 0.243 0.291 0.332 0.344 0.378 0.409
29 0.239 0.287 0.328 0.340 0.374 0.404
30 0.236 0.283 0.324 0.336 0.369 0.399
(I Sample size.

a one-tailed test; but, for very small sample sizes, the effect by independent knowledge of the true distribution of the
m a y be slightly less than double depending upon the dis- population. This argument implies that, at worst, assuming
tribution pattern of the data since the tail containing the a doubling of the a risk upon going from a one-tailed test to
perceived deviant is determined by the sample rather than a two-tailed test will result in critical values that are slightly
 ANALYTICAL CHEMISTRY, VOL. 63, NO. 2, JANUARY 15, 1991 143

Table 111. Critical Values of Dixon’s rI2Parameter As Applied to a Two-TailedTest at Various Confidence Levels, Including
the 95% Confidence Level
confidence level
80 ‘70 90 % 95 % 96 % 98 ‘70 99 %
NQ (a = 0.20) (a = 0.10) (a = 0.05) (a = 0.04) (a = 0.02) (a = 0.01)

5 0.919 0.960 0.980 0.984 0.992 0.996

6 0.745 0.824 0.878 0.891 0.925 0.951
7 0.636 0.712 0.773 0.791 0.836 0.875
8 0.557 0.632 0.692 0.708 0.760 0.797
9. 0.504 0.580 0.639 0.656 0.702 0.739
10 0.464b 0.537 0.594 0.610 0.655 0.694
11 0.431 0.502 0.559 0.575 0.619 0.658
12 0.406 0.473 0.529 0.546 0.590 0.629
13 0.387 0.451 0.505 0.521 0.564c 0.602c
14 0.369 0.432 0.485 0.501 0.542 0.580
15 0.354 0.416 0.467 0.482 0.523 0.560
16 0.341 0.401 0.452 0.467 0.508 0.544
17 0.330 0.388 0.438 0.453 0.493 0.529
18 0.320 0.377 0.426 0.440 0.480 0.516
19 0.311 0.367 0.415 0.429 0.469 0.504
20 0.303 0.358 0.405 0.419 0.458 0.493
21 0.296 0.349 0.396 0.410 0.449 0.483
22 0.290 0.342 0.388 0.402 0.440 0.474
23 0.284 0.336 0.381 0.394 0.432 0.465
24 0.278 0.330 0.374 0.387 0.423 0.457
25 0.273 0.324 0.368 0.381 0.417 0.450
26 0.268 0.319 0.362 0.375 0.411 0.443
27 0.263 0.314 0.357 0.370 0.405 0.437
28 0.259 0.309 0.352 0.365 0.399 0.431
29 0.255 0.305 0.347 0.360 0.394 0.426
30 0.251 0.301 0.343 0.355 0.389 0.420
Sample size. In Dixon’s original table (13),the r12 critical value a t the 80% confidence level for n = 10 is 0.454. However, the cubic
regression curve based on the 40%, 60%, 90%, 96%, 98%, and 99% confidence level critical values for n = 10 as well as a regression curve
fitted to the 80% cricital values versus sample size indicates that this value should be 0.464. Therefore, it is concluded that the 80% critical
value originally published for r12a t n = 10 contains a typographical error. The same r12 critical value for 95% confidence is obtained either
by using the corrected 80% value or by omitting it altogether from the regression analysis. ‘The r12critical values in Dixon’s original table
(13) for n = 13 a t the 98% and 99% confidence levels are 0.554 and 0.612, respectively. However, cubic regression curves fitted to the
critical values at these two confidence levels as a function of sample size yield corrected values of 0.564 and 0.602, respectively, indicating
that the original values contained typographical errors. These latter values were used in resolving the rI2critical value for the 95% confi-
dence level a t n = 13 (although the same 95% value was obtained by omitting these values and including the critical values for the 40% and
60% confidence levels).

too high, and the resulting decisions that are made will be functions could generally be extended to include the data for
overly conservative. As shown by comparison of Dixon’s the 60% and 40% confidence levels but, except as noted
original one-tailed critical values for 95% confidence (15) with below, these values were not included in fitting the cubic
his two-tailed critical values for the 90% confidence level (16b), regression curves since the data in the lower confidence level
it is clear that Dixon assumed a doubling of the a risk.] region did not significantly affect the calculation of the 95%
Interpolation of Critical Values at the 95% Confidence values.
Level. As noted above, Dixon was able to obtain exact so- Based on the foregoing analysis, the critical values of Q a t
lutions for the various critical values only for the cases where the 95% confidence level were then calculated from the cubic
n = 3 or 4. Although the general form of the equation for n regression curve for each sample size and were found to be
2 5 has been presented (15),the specific expressions for the within fO.OO1 of the values obtained graphically. Cubic
central density function vary with each sample size, and these functions were subsequently fitted to the 99%, 98%, 96%,
expressions have not been published. Therefore, in the current 90%, and 80% confidence level data for each of the other ratio
work, appropriate two-tailed critical values of Q at the 95% functions defined by Dixon. To check the veracity of the
confidence level were initially estimated by plotting the generated equations, the critical values for 96% and 90% were
(two-tailed) critical values for the 99%, 98%, 96%, 90%, and also calculated and, in each case (except as noted below), were
80% confidence levels as generated by Dixon. The graphically found to be within *0.001 of Dixon’s tabular values.
interpolated values for 95% Q were then checked by using In using this approach, it was noted that cubic equations
regression analysis to determine the best fitting empirical could not be fitted to the rz0 critical values for n I19. After
polynomial functions to the values that Dixon published; these examining Dixon’s tabular data carefully, it was discovered
functions were then solved for the appropriate critical values that the 90% confidence level critical values showed a dis-
of Q a t the 95% confidence level. continuity in this region. T o circumvent this problem, the
With the exception of cases in which the original Dixon 60% confidence level values were included in fitting the rz0
tables contained apparent errors (vide infra), it was found that data to cubic equations. In the region 4 5 n I18,the critical
a cubic function provided an optimal fit to the Q values in values of rz0at the 95% confidence level obtained by including
this region. Interestingly, the use of a quadratic function and excluding the 60% critical values were within f0.0002,
generally yielded the same values of 95% Q, to three signif- i.e., undetectable to three significant figures (see Figure 1).
icant figures, as those obtained from a cubic function despite For n I19, the 90% values were then omitted. In this manner,
a notably poorer fit; fourth-power functions also produced the cubic expressions were generated that provided excellent fits
same 95% values to within fO.001 but were less sensitive to to the data and permitted both the 95% and new 90% critical
errors in the tabular data. It was also noted that the cubic values to be computed. Interestingly, in comparing the newly
144 ANALYTICAL CHEMISTRY, VOL. 63, NO. 2, JANUARY 15, 1991

Table IV. Critical Values of Dixon's rzoParameter As Applied to a Two-Tailed Test at Various Confidence Levels, Including
the 95% Confidence Level
confidence level
80 % 90% 95 % 96 C7c 98 % 99 %
h'" ( n = 0.20) ( a = 0.10) ( n = 0.05) ( a = 0.04) (a = 0.02) ( a = 0.01)

4 0.935 0.96; 0.983 0.987 0.992 0.996

5 0.782 0.845 0.890 0.901 0.929 0.950
6 0.670 0.736 0.786 0.800 0.836 0.865
I 0.596 0.661 0.716 0.732 0.778 0.814
8 0.545 0.607 0.657 0.670 0.710 0.746
9 0.505 0.565 0.614 0.62; 0.667 0.700
10 0.474 0.531 0.579 0.592 0.632 0.664
11 0.449 0.504 0.551 0.564 0.603 0.627
12 0.429 0.481 0.527 0.540 0.579 0.612
13 0.411 0.461 0.506 0.520 0.557 0.590
14 0.395 0.445 0.489 0.502 0.538 0.571
15 0.382 0.430 0.473 0.486 0.522 0.554
16 0.370 0.418 0.460 0.472 0.508 0.539
17 0.359 0.406 0.447 0.460 0.495 0.526
18 0.350 0.397 0.437 0.449 0.484 0.514
19 0.341 0.3876 0.427b 0.439 0.473 0.503
(0.379)
20 0.333 0.3786 0.41Sb 0.430 0.464 0.494
(0.372L
21 0.326 0.371 0.4106 0.422 0.455 0.485
(0.365)
22 0.320 0.364b 0.402b 0.414 0.447 0.477
(0.358)
23 0.314 0.358b 0.395b 0.407 0.440 0.469
(0.352)
24 0.309 0.352b 0.3906 0.401 0.434 0.462
(0.347)
25 0.304 0.346b 0.383b 0.395 0.428 0.456
(0.343)
26 0.300 0.3426 0.37gb 0.390 0.422 0.450
(0.338)
27 0.296 0.33S6 0.374b 0.385 0.477 0.444
(0.334)
28 0.292 0.333b 0.370b 0.381 0.412 0.439
(0.330)
29 0.288 0.32g6 0.3Mb 0.376 0.407 0.434
(0.326)
30 0.285 0.326O 0.361b 0.372 0.402 0.428
(0.322)
"Sample size. bStarting with n = 19, the r.," critical values for both the 90% and 95% confidence levels were calculated from the cubic
regression curves fitted to the critical values-published by Dixon (13) corresponding to the two-tailed 60%, 80%, 96%, 98%, and 99%
confidence levels (but omitting the published 90% confidence values). For the 90% confidence level, the values originally published by
Dixon are indicated in parentheses underneath the newly generated values. From a comparison of the two sets of values, it is obvious that
the critical values in the original table were shifted up one row in the column corresponding to the two-tailed 90% confidence level (see text).

generated 90% confidence level values of rz0with the original

tabular data, the results reveal that, in Dixon's original table
( I j ) ,the two-tailed 90% confidence level values (corresponding
to Dixon's one-tailed N = 0.05) for 19 In I30 were acci-
dentally displaced upward by one row (see Table IV).
In any set of data in which the critical values a t the 90%
and/or 96% confidence levels, as calculated from the cubic
regression equation, differed by more than h0.001 from
Dixon's tabular values, both the cubic equation and the or-
iginal tabular data were carefully checked for error. In this
manner, five additional typographical errors were uncovered 0.4 0.5 0.6 0.7
in Dixon's original tables. Corrected values were obtained in rz0 Values
two ways: (i) by generating a new cubic equation for the a function of the confidence
Figure 1. Typical plot of critical values as
specific sample size omitting the suspected tabular value (while level. The curve shown is the regression curve for rz0 at n = 10
extending the cubic fit to include the 60%-and, in some cases (including the values for 60%, 80%, 9 0 % , 96%, 98%, and 9 9 % )
the 40%-confidence level value); and (ii) by generating a corresponding to the cubic equation: YO confidence level = -343.7575
cubic or quartic equation to fit the critical values at the specific +
4- 1835.0349r2, - 2548.5722rZo2 1 188.477rZo3.
confidence level as a function of sample size. Identical values
were obtained with both approaches and, in all five cases, the level and similarly generated critical values for the rll, rlz,ra,
corrected values revealed that one digit had been incorrectly and rZ2functions are included in Tables I-VI along with the
typeset in the original paper (15). The corrected values are values published by Dixon corresponding to the 99%, 98%,
indicated in Tables I11 and VI. 96%, 9070, and 80% confidence levels-all confidence levels
The resultant critical values of Q (rl0)at the 95% confidence .shown being applicable t o a two-tailed test. All values have
 ANALYTICAL CHEMISTRY, VOL. 63, NO. 2, JANUARY 15, 1991 145

Table V. Critical Values of Dixon's r I IParameter As Applied to a Two-Tailed Test at Various Confidence Levels, Including
the 95% Confidence Level
confidence level
80 % 90 % 95 % 96 % 98% 99 %
N" (a = 0.20) (a = 0.10) (a = 0.05) (a = 0.04) (a = 0.02) (a = 0.01)

5 0.952 0.976 0.987 0.990 0.995 0.998

6 0.821 0.872 0.913 0.924 0.951 0.970
7 0.725 0.780 0.828 0.842 0.885 0.919
8 0.650 0.710 0.763 0.780 0.829 0.868
9 0.594 0.657 0.710 0.725 0.776 0.816
10 0.551 0.612 0.664 0.678 0.726 0.760
11 0.517 0.576 0.625 0.638 0.679 0.713
12 0.490 0.546 0.592 0.605 0.642 0.675
13 0.467 0.521 0.565 0.578 0.615 0.649
14 0.448 0.501 0.544 0.556 0.593 0.627
15 0.431 0.483 0.525 0.537 0.574 0.607
16 0.416 0.467 0.509 0.521 0.557 0.580
17 0.403 0.453 0.495 0.507 0.542 0.573
18 0.391 0.440 0.482 0.494 0.529 0.559
19 0.380 0.428 0.469 0.482 0.517 0.547
20 0.371 0.419 0.460 0.472 0.506 0.536
21 0.363 0.410 0.450 0.462 0.496 0.526
22 0.356 0.402 0.441 0.453 0.487 0.517
23 0.349 0.395 0.434 0.445 0.479 0.509
24 0.343 0.388 0.427 0.438 0.471 0.501
25 0.337 0.382 0.420 0.431 0.464 0.493
26 0.331 0.376 0.414 0.424 0.457 0.486
27 0.325 0.370 0.407 0.418 0.450 0.479
28 0.320 0.365 0.402 0.412 0.444 0.472
29 0.316 0.360 0.396 0.406 0.438 0.466
30 0.312 0.355 0.391 0.401 0.433 0.460
" Sample size.
Table VI. Critical Values of Dixon's rZ2Parameter As Applied to a Two-Tailed Test at Various Confidence Levels, Including
the 95% Confidence Level
confidence level
80 % 90,370 95% 96% 98% 99 %
N" (a = 0.20) (a = 0.10) (a = 0.05) (a = 0.04) (a = 0.02) (a = 0.01)

6 0.965 0.983 0.990 0.992 0.995 0.998

7 0.850 0.881 0.909 0.919 0.945 0.970
8 0.745 0.803 0.846 0.857 0.890 0.922
9 0.676 0.737 0.787 0.800 0.840 0.873
10 0.620 0.682 0.734 0.749 0.791 0.826
11 0.578 0.637 0.688 0.703 0.745 0.781
12 0.543 0.600 0.648 0.661 0.704 0.740
13 0.515 0.570 0.616 0.628 0.670 0.705
14 0.492 0.546 0.590 0.602 0.641 0.674
15 0.472 0.525 0.568 0.579 0.616 0.647
16 0.454 0.507 0.548 0.559 0.595 0.624
17 0.438 0.490 0.531 0.542 0.577 0.605
18 0.424 0.475 0.516 0.527 0.561 0.589
19 0.412 0.462 0.503 0.514 0.547 0.575
20 0.401 0.450 0.491 0.502 0.535 0.562
21 0.391 0.440 0.480 0.491 0.524 0.551
22 0.382 0.430 0.470 0.481 0.514 0.541
23 0.374 0.421 0.461 0.472 0.505 0.532
24 0.367 0.413 0.452 0.464b 0.497 0.524
25 0.360 0.406 0.445 0.457 0.489 0.516
26 0.354 0.399 0.438 0.450 0.482b 0.508
27 0.348 0.393 0.432 0.443 0.475 0.501
28 0.342 0.387 0.426 0.437 0.469 0.495
29 0.337 0.381 0.419 0.431 0.463 0.489
30 0.332 0.376 0.414 0.425 0.457 0.483
"Sample size. bThe rzz critical values listed in Dixon's original table (13) for n = 24 at the 96% confidence level (two-tailed) and for n =
26 at the 98% confidence level are 0.484 and 0.486, respectively. The values shown in this table (0.464 and 0.482, respectively) were obtained
by a cubic regression curve fitted to the critical values at these two confidence levels as a function of sample size. These same values are also
generated from the cubic regression lines fitted to the critical values for each of these sample sizes omitting the questionable values and
including the values at the 40% and 60% confidence levels. It is concluded that the original tabular critical values for rz2 for these two
sample sizes and confidence levels were the result of typographical errors.

been carefully cross-checked, and the 95% critical values are (15)were themselves generally accurate to within *0.001). The
judged to be accurate within fO.OO1 relative to the accuracy values shown cover the entire range of sample sizes (3 5 n I
of the values at the other confidence levels (which Dixon stated 30) included in Dixon's original article.
146 Anal. Chem. 1991, 63, 146-15 1

It is suggested that the critical values of Q and the related titative Analysis: Theory and Practice; Harper 8 Row: New York,
r criteria which have been generated for the 95% confidence 1987; pp 78-80. (f) Rubinson, K. A. Chemical Analysis; Little, Brown:
Boston, 1987; pp 162-164. (9) Day, R. A., Jr.; Underwood, A. L.
level should be used routinely by practicing analytical chemists Quantitative Analysis, 5th ed.; Prentice-Hall: Englewood Cliffs, NJ,
in testing for the rejection of outliers since this confidence 1986; pp 29-31. (h) Manahan, S.E. Quantitative Chemical Analysis;
Brooks Cole: Monterey. CA, 1986; pp 74-75. (i)Kennedy, J. H. Ana-
level provides a reasonable compromise between ultracon- iyticai Chemistry: Principles, 2nd ed.; Harcourt, Brace, Jovanovich:
servatism and the overzealous rejection of deviant values. New York, 1990; pp 35-39. (i) Harris, D. C. Quantitative Chemical
These values should also be incorporated into future analytical Analysis; Freeman: San Francisco, 1982; pp 51-52. (k) Ramette, R.
W. Chemical Equilibrium and Analysis ; Addison-Wesley: Reading, MA,
chemistry treatises and textbooks dealing with tests for the 1981; pp 53-54. (I) Christian, G. D. AnalyticalChemistry; Wiley: New
rejection of data to provide a uniform set of critical values York, 1980; pp 78-79. (m) Flaschka, H. A,; Barnard. A. J., Jr.; Stur-
rock, P. E. Quantitative Analytical Chemistry; Willard Grant: Boston,
a t this standard confidence level. 1980; pp 19-20.
As a concluding comment, it should be noted that recent Dixon, W. J. Ann. Math. Stat. 1950, 27. 488-506
studies on the variance of the arithmetic mean after rejection King, E. P. J. Am. Stat. Assoc. 1953, 48, 531-533.
Deming, W. E. Statistical Aaustment of Data; Wiley: New York,
of outliers suggest the superiority of two more recently pro- 1943 (republished by Dover: New York. 1964); p 171.
posed criteria (Huber-type skipped mean and Shapiro-Wilk Parratt, L. G. Probability and Experimental Errors in Science ; Wiley:
New York, 1961; pp 176-178.
rules) for rejection decisions (19),particularly for larger sam- Hawkins, D. M. Identification of Outliers; Chapman and Hall: London,
ples containing multiple outliers. Nonetheless, the simplicity 1980.
of Dixon's range ratio tests argues strongly for their continued Beckman, R . J.; Cook, R . D. Technometrics 1983, 25, 119-149.
(a) Burr, I. W. Applied Statistical Methods; Academic: New York.
use in many analytical applications. 1974; pp 194-195. (b) Hamilton, W. C. Statistics in Physical Science;
Ronald: New York. 1964; pp 45-49.
ACKNOWLEDGMENT Dixon, W. J. I n Contributions to Order Statistics; Sarhan, A. E.,
Greenberg, B. G., Eds.; Wiley: New York, 1962; pp 299-342 (see pp
I express my appreciation to my colleagues, George Schenk 314-317).
and David Coleman, for helpful suggestions regarding this Natrelia, M. G. Experimental Statistics ; National Bureau of Standards
Handbook 91; NBS: Washington, DC, 1963; Chapter 17.
manuscript. "Student". [Gossett, W. S.] Biometrika 1908, 6 , 1.
Taylor, J. R. An Introduction to Error Analysis: The Study of Uncer-
LITERATURE CITED tainties in Physical Measurements; University Science: Mill Valley,
(a) Miller, J. C.; Miller, J. N. Statistics for Analytical Chemistry. 2nd ed.; CA, 1982; pp 142-145.
Wiley: New York. 1988; pp 62-64, 218. (b) Anderson, R. L. Practical Dixon, W. J. Ann. Math. Stat. 1951, 22, 68-78.
Statistics for Analytical Chemlsts; Van Nostrand Reinhoid: New York, (a) Dixon, W. J. Biometrics 1953, 9, 74-89. (b) Dean, R . 8.; Dixon.
1987; pp 31-32. (c) Caulcutt, R.;Boddy, R . Statistics for Analytical W. J. Anal. Chem. 1951, 2 3 , 636-638.
Chemists; Chapman and Hall: London, 1983; pp 66-67, 248. Peters, D. G.; Hayes, J. M.; Hieftje, G. M. Chemical Separations and
Measurements : Theory and Practice of Analytical Chemistry; Saun-
Barnett, V.; Lewis, T. Outliers in Statistical Data, 2nd ed.; Wiiey: New ders: Philadelphia, 1974; p 36.
York. 1984.
(a) Skoog, D. A.; West, D. M.; Holler, F. J. Analytical Chemistry: An Youmans, H. L. Statistics for Chemistry; Charles E. Merrill: Columbus,
Introduction, 5th ed.; Saunders: Philadelphia, 1990; p 56. (b) Skoog, OH, 1973; p 65 ff.
D. A.; West, D. M.; Holler, F. J. Fundamentals of Analytical Chemistry, Hampel, F. R. Technomefrics 1985, 27, 95-107.
5th ed.; Saunders: New York, 1988; pp 13-16. (c) Hargis, L. G.
Analytical Chemistry; Prentice-Hall: Englewocd Cliffs, NJ, 1988; p 56.
(d) Fritz, J. S.:Schenk, G. H. Quantitative Analytical Chemistry, 5th RECEIVED
for review June 15, 1990. Accepted October 16,
ed.: Allyn 8 Bacon: Boston, 1987; pp 45-46. (e) Potts, L. W. Quan- 1990.

Frequency-Domain Spectroscopic Study of the Effect of

n-Propanol on the Internal Viscosity of Sodium Dodecyl Sulfate
Micelles
M. J. Wirth,* S.-H. Chou, and D. A. Piasecki
Department of Chemistry and Biochemistry, University of Delaware, Newark, Delaware 19716

The rotational diffusion behavior of tetracene in sodium do- INTRODUCTION

decyi sulfate micelles is studied as a function of the n- The importance of micelles in analytical chemistry has
propanol content of the micellar solution. Fluorescence an- burgeoned both in spectroscopy and in separation science.
isotropy measurements using frequencydomain spectroscopy Micelles are used to enhance fluorescence ( I ) , thermal lensing
show that tetracene reorients faster with increasing concen- (21, and room-temperature phosphorescence ( 3 ) . In liquid
tration of n-propanol. This result Is consistent with micellar chromatography, micelles modify the organic content of the
liquid chromatographic studies and lends insight into the role mobile phase ( 4 ) , allowing more rapid gradient elution ( 5 ) .
of n-propanol as a mobile-phase modifier. The components Micelles can serve as the pseudostationary phase in elec-
of the rotational diffusion tensor are determined from the trokinetic chromatography (6), providing selectivity for sep-
double-exponential anisotropy decay. These confirm the va- aration of polar organic solutes. The nature of solute inter-
lidity of the Debye-Stokes-Einstein model and allow caicu- actions with micelles is thus a vital area of research.
iation of the viscosity of the micelle interior. The viscosity
An important question entails the dynamics of solutes in-
teracting with micelles. One of the drawbacks of micellar
decreases from 8 to 4 CP as the concentration of n-propanol
liquid chromatography has been poor column efficiency due
increases from 0 to 10% (v/v). The components of the ro-
to slow mass transfer between the surfactant-modified sta-
tational diffusion tensor indicate that the solvation environment tionary phase and the mobile phase ( 7 ) . Dorsey et al. dem-
of tetracene in the micelle is structurally disordered. onstrated that the addition of at least 1% of n-propanol gives
a 40% increase in the number of theoretical plates for the
* T o whom correspondence should be addressed. solute benzene in micellar liquid chromatography (8). They
0003-2700/91/0363-0146$02.50/0 1991 American Chemical Society