A Suggestion for Using Powerful and Informative Tests of Normality

Author(s): Ralph B. D'Agostino, Albert Belanger and Ralph B. D'Agostino, Jr.

Reviewed work(s):
Source: The American Statistician, Vol. 44, No. 4 (Nov., 1990), pp. 316-321
Published by: American Statistical Association
Stable URL: http://www.jstor.org/stable/2684359 .
Accessed: 03/06/2012 07:58

Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at .
http://www.jstor.org/page/info/about/policies/terms.jsp

JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of
content in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms
of scholarship. For more information about JSTOR, please contact support@jstor.org.

American Statistical Association is collaborating with JSTOR to digitize, preserve and extend access to The
American Statistician.

http://www.jstor.org
 Commentariesare informativeessays dealing with viewpoints of statis-
tical practice, statistical education, and other topics considered to be
of general interest to the broad readershipof The American Statistician.
Commentaries are similar in spirit to Letters to the Editor, but they
A Suggestion for Using Powerful and Informative
Tests of Normality
RALPH B. D'AGOSTINO, ALBERT BELANGER, AND RALPH B. D'AGOSTINO, JR.*
For testing that an underlyingpopulation is normally dis-
tributedthe skewness and kurtosis statistics, b1 and b2,
andthe D'Agostino-Pearson K2statisticthatcombinesthese
two statistics have been shown to be powerful and infor-
mative tests. Their use, however, has not been as prevalent
as their usefulness. We review these tests and show how
readilyavailableand popularstatisticalsoftwarecan be used
to implement them. Their relationshipto deviations from
linearity in normal probabilityplotting is also presented.
KEY WORDS: Eb, b2; D'Agostino-Pearson K2; Kur-
tosis; Normal probabilityplot; Skewness.
1. INTRODUCTION
Tests of normalityare statisticalinferenceproceduresde-
signed to test that the underlyingdistributionof a random
variable is normally distributed.There is a long history of
these tests, and there are a plethora of them available for
use (D'Agostino 1971; D'Agostinoand Stephens1986, chap.
9). Majorstudies investigatingthe statisticalpower of these
over a wide range of alternative distributionshave been
undertaken,and a reasonablyconsistentpicturehas emerged
as to which of these should be recommendedfor use. See
D'Agostino and Stephens (1986, chap. 9) for a review of
these power studies. The Shapiro-Wilk W test (Shapiroand
Wilk 1965), the third sample moment ( b1) and fourth
sample moment (b2) tests, and the D'Agostino-Pearson K2
test combiningthese (D'Agostino and Pearson1973) emerge
as excellent tests. The WandK2tests sharethe fine property
of being omnibus tests, in that they have good power prop-
erties over a broad range of nonnormaldistributions.The
b1 and b2 tests have excellent properties for detecting
nonnormalityassociated with skewness and nonnormalkur-
*Ralph B. D'Agostino is Chairmanand Professor, MathematicsDe-
partment,Boston University, Boston, MA 02215. Albert Belangeris Stat-
istician, Statistics and Consulting Unit, Boston University, Boston, MA
02215. Ralph B. D'Agostino, Jr., is a graduate student, Statistics De-
partment, Harvard University, Cambridge, MA 02138. This work was
supported in part by National Heart Lung and Blood InstituteGrant 1-
RO -HL-40423-02.
316 The American Statistician, November 1990, Vol. 44, No. 4
involve longer discussions of background, issues, and perspectives. All
commentarieswill be refereed for their merit and compatibilitywith these
criteria.
tosis, respectively. The extensive power studies just men-
tioned have also demonstratedconvincingly that the old
warhorses, the chi-squared test and the Kolmogorov test
(1933), have poor power propertiesand should not be used
when testing for normality.
Unfortunately,the preceding results have not been dis-
seminatedvery well. The chi-squaredand Kolmogorovtests
are still suggested in textbooks for testing for normality.
Majorstatisticalpackages such as SAS and SPSSX perform
the excellent Shapiro-Wilk W test for sample sizes up to
50. For larger samples, however, they supply the poor-
power Kolmogorov test. These statisticalpackages do give
skewness and kurtosis measures. They are not, however,
the b1 and b2 statistics. Ratherthey are functionsof these,
the so-called Fisher g statistics (Fisher 1973). The docu-
mentation on this latter point is very incomplete. In our
experience, many users are unawareof it, and descriptive
evaluationof normalityor nonnormalityis confusedbecause
of it. Hypothesis testing using the powerful Wband b2 is
not presentedor even suggested.
In this article, we discuss the skewness, 4b, and kur-
tosis, b2, statistics and indicate how they are excellent de-
scriptive and inferentialmeasuresfor evaluatingnormality.
Further,we relate the Fisher g skewness and kurtosismea-
sures producedby the SAS and SPSSX software packages
to b1and b2and show how a simple program(SAS macro)
can be used to produce an excellent, informativeanalysis
for investigatingnormality. This analysis contains separate
tests based on Ab7and b2 and the K2 test, which combines
Vb7 and b2 for an omnibus test. Finally, we indicate how
the precedingcan be used in conjunctionwith normalprob-
ability plotting. The latter gives an informativegraphical
component to an analysis for normality.
2. POPULATION-MOMENTS DESCRIPTION OF
NORMALITY AND NONNORMALITY
A population, or its random variable X, is said to be
normally distributedif its density function is given by
1 _ I b-r y m < x < m0
f(x) 1 e 2 -?? < x < m(
V2-iTo- >0
?) 1990 American Statistical Association
Here , and a- are the mean and standarddeviation, respec- tails are lighter than the normal tend to have /2 < 3. In
tively, of it. Of interest here are the third and fourth stan- terms of their peak, it tends to be broaderthan the normal
dardized moments given by (platykurtic).Readers are referredto D'Agostino and Ste-
phens (1986) for furtherdiscussion of these and to Balanda
E(X - ,u)3 E(X -
1 =[E(X a-3 and MacGillivray (1988) for a detailed discussion of kur-
(2)-
-_ tosis. D'Agostino and Stephens (1986) also gave examples
and of well-known nonnormaldistributionsindexedby /i and
P2-
E(X - )4 _E(X
-3
32
132 = 22[E(X
( - 2 (3) 3. SAMPLE MOMENTS AS INDICATORS OF
NONNORMALITY
where E is the expected value operator. These moments
measure skewness and kurtosis, respectively, and for the Karl Pearson (1895) was the first to suggest that the
normal distributionthey are equal to 0 and 3, respectively. sample estimates of f31and /2 could be used to describe
The nonnormalityof a populationcan be describedby values nonnormaldistributionsand used as the bases for tests of
of its central moments differing from the normal values. normality. For a sample of size n, XI, . ., X,1the sample
The normal distribution is symmetric, so 138= 0. A estimates of /31 and /32 are, respectively,
nonnormaldistributionthat is asymmetricalhas a value of m3l 32 (4)
N =
\7,j = 0 (see Fig. 1); 138> 0 correspondsto skewness
to the right and 13f< 0 correspondsto skewness to the and
left. b= m4/m2, (5)
The word kurtosis means "curvature,"and it has tradi-
tionally been measuredby the fourth standardizedmoment where
132. For the normal distribution, its value is 3. Figure 1 mk (Xi - X)kln (6)
displays two nonnormaldistributionsin which 132 3. Un-
imodal distributionswhose tails are heavier or thickerthan and X is the sample mean
the normal have 132 > 3. These distributionsalso tend to X = ZXi,/. (7)
have higher peaks in the center of the distribution,and in
the past these distributionswere often describedin terms of As descriptive statistics, values of Nb and b2 close to 0
the high peaks (leptokurtic).Unimodaldistributionswhose and 3, respectively, indicate normality.To be more precise
the expected values of these are 0 and 3(n - 1)/(n + 1)
undernormality. Values differing from these are indicators
A B C of nonnormality. The signs and magnitudesof these give
informationabout the type of nonnormality[e.g., /b >
ocorrespondsto positive skewness and b2 > 3(n- -)(n
+ 1) relates to heavy tails in the populationdistribution].

4. TESTS OF NORMALITY BASED ON SAMPLE

MOMENTS
The Nb and b2 statistics are the bases for powerful tests
of normality(D'Agostino and Stephens 1986, chap. 9).

4.1 Tests of Skewness ( b1)

C Here the null hypothesis is Ho: normalityversus the al-

ternative;HI: nonnormalitydue to skewness. For alterna-
tives (A/,(3,#&0), a two-sidedtest basedon Nb is performed.
For directional alternatives(V/13 > 0 or /13 < 0), one-
A
sided tests are performed.Tables of criticalvalues are avail-
able (D'Agostino and Stephens 1986, chap. 9). For sample
sizes n > 8, a normal approximationthat is easily com-
puterizedis available. It is obtainedas follows (D'Agostino
1970):
1. Compute Nb from the sample data.
2. Compute

5
Y= b{( 6(n-)( ) n' (8)
Figure 1. Illustrationof DistributionsWith\/ # 0 and 132 # 3;
DistributionsDifferingin Skewness and-Differingin Kurtosisfrom
the NormalDistribution;Top panel: A, Vii > 0; B, X/ = 0; C, 3(n2 ? 27n - 70)(n ? 1)(n ? 3),
V, <0O;Bottompanel: A, 132- 3; B, (32 < 3; C, (32 > 3.

The AmericanStatistician, November 1990, Vol. 44, No. 4 317

 W2= -1 + {2(/2(b) - 1)}1/2, (10) 4.3 Omnibus Test
2= 1/8in VIb, (11) D'Agostino and Pearson (1973) presenteda statistic that
combines b1 and b2 to produce an omnibus test of nor-
and
mality. By omnibus, we mean it is able to detect deviations
a = {2/(W2 - 1)}I/2. (12) from normalitydue to either skewness or kurtosis. The test
statistic is
3. Compute
K2 = Z2(\/'b1) + Z2(b2), (20)
Z( b1) = I1n(Y/a + {(Y/a)2 + 1}1/2). (13)
where Z(V/-b) and Z(b2) are the normalapproximationsto
Z(V/J7) is approximatelynormallydistributedunderthe null
Nb and b2 discussed in Sections 4.1 and 4.2. The K2
hypothesis of populationnormality.
statistic has approximatelya chi-squareddistribution,with
2 df when the populationis normally distributed.
4.2 Tests of Kurtosis (b2)
5. NUMERICAL EXAMPLE
Here the null hypothesis is Ho: normalityversus the al-
ternative;HI: nonnormalitydue to nonnormalkurtosis.Again Table 1 contains a sample of cholesterol values from a
a two-sided test (for 12 # 3) or one-sided tests (for /2 > sample of 62 subjects from the FraminghamHeart Study.
3 or /82 < 3) can be performed.Again elaboratetables are The data are presentedas a stem-and-leafplot. From these
available (D'Agostino and Stephens 1986, chap. 9). More- data we obtain
over, a normal approximationdue to Anscombe and Glynn + = 1.02, Z(Vbi) = 3.14, p = .0017,
(1983) is available. It is valid for n ' 20 and is as follows:
b2 = 4.58, Z(b2) = 2.21, p = .0269,
1. Compute b2 from the sample data.
and
2. Compute the mean and varianceof b2,
K2 = 14.75, p = .0006.
3(n 1)
E(b2) (14) The precedingp values are the levels of significance for the
n?1 correspondingtwo-sided tests. For the Kolmogorov test, p
and = .087. The data are clearly nonnormal.The Vb, and b2
24n(n statistics quantify the natureof the nonnormality.The data
var(b)=
- 2) (n - 3) (5
b2) distributionis skewed to the right and heavy in the tails.
=(n + 1)2(n + 3)(n + 5) (15)
The Kolmogorov test gives no informationabout this non-
3. Compute the standardizedversion of b2, normalityand only indicates marginallynonnormality.
x = (b2 - E(b2))/ varb2) (16) 6. THE FISHER g STATISTICS
4. Compute the third standardizedmoment of b2, Both SAS and SPSSX routinely give skewness and kur-
tosis statistics in their descriptive statistics output. Unfor-
6(n2- 5n + 2) 6(n + 3)(n + 5)
2) - Table 1. Cholesterol Data From the Framingham Heart Study
(n + 7)(n + 9) n(n - 2)(n - 3)
(17) Stem-and-leaf plot Number
39 3 1
5. Compute 38
37
8 36
A =6 +
f31
-(b V8-) 35 3 1
L) f32) 34
33 46 2
32 7 1
? (1I +
i+ 1(b2))l(8
?~,8
9^ (18) 31
30 008 3
29
6. Compute 28 35 2
27 00288 5
// 2 26 347778 6
Z(b2) = ((1 - A) 25 444668 6
24 03678 5
23
F 1 -2/A
[L1J 11d
. 22
0000122244668
0556
13
4
/ V2/(9A (19) 21 7
0125678
20 02 2
Z(b2) is approximatelynormally distributedunderthe null 19 28 2
18 4 1
hypothesis of populationnormality. 17
16 7 1
Both Z( b) andZ(b2) can be used to test one-sided and NOTE:The descriptivestatisticsare sample size, n = 62; mean, X = 250.0; standard
two-sided alternativehypotheses. deviation,S = 41.4; skewness, Vff = 1.02; kurtosis,b2 = 4.58:

318 The American Statistician, Novemnber1990, Vol. 44, No. 4

tunately, neither give VbA and b2. Rather, they give the and Stephens (1986, chap. 2) for a detailed discussion of
Fisher g statistics defined as follows: probabilityplotting.
A normal probabilityplot is simply a plot of the inverse
n
skewness g - ( -- l)(n- 2)3 (21) of the standardnormal cumulative on the horizontal axis
and the ordered observations on the vertical axis. The in-
and verse of the normalcumulative is usually defined in such a
way to enhance the linearity of the plot, and one common
n(n + 1) (X - 4
k
kurtosis g92 procedureis to let the normalprobabilityplot employ Blom's
(n - 1)(n-2)(n- 3)S4 (1958) approximation.In this case, the normal probability
3(n - 1)2 (22) plot is a plot of

(n- 2) (n- 3)
i
where Z = ( n + 1/4) on X(i)' (26)
where X(i) is the ith ordered observationfrom the ordered
n - X)2 (23) sample X(l) ' ' X(1t)and
...
n - I n-1
i -
is the sample variance. Z =F1( 3/8) (27)
These are related to Vbj and b2 via the following:
is the Z value such that
(n 2) (24)
i- 3/8 V 2-
2dx (28)
and n + 1/4 f x 2;;e

(n- 2) (n- 3) 3 (n - 1) for i= 1, ., n.

(n + 1)(n- 1) (n + 1)
The BMDP statistics software package does compute
VW/ and b2. All of the preceding software do not perform
tests of normalitybased on skewness and kurtosis.
7. RECOMMENDATIONS
The tests just describedbasedon \/H andb2areexcellent
and powerful tests. We recommend that for all sample
sizes VN and b2 should be computed and examined as
descriptive statistics. For all sample sizes n - 9, tests of
hypotheses can be based on them. In particular,for n >
50, where the Shapiro-Wilk test is no longer available, we
recommendthese tests and the D'Agostino-Pearson K2 test
as the tests of choice. The justification for this is not only
because of their fine power but also because of the infor-
mation they supply on nonnormality.In conjunction with
the use of standardstatisticalsoftware,such as SAS, SPSSX,
and BMDP, the skewness and kurtosis measuresthey pro-
duce can be used as inputs to simple programs(macros) to
perform these tests. In the appendix, we supply one such
simple macro that can be used with SAS and that will pro-
vide two-tailed tests.
8. NORMAL PROBABILITY PLOT
Another component in a good data analysis for investi-
gating normalityof data and an item again often not well
handledroutinelyin computerpackages is the normalprob-
ability plot. This plot is a graphicalpresentationof the data
that will be approximatelya straight line if the underlying
distributionis normal. Deviations from linearitycorrespond
to various types of nonnormality.Some of these deviations
reflect skewness and/orkurtosis.Othersreflect featuressuch
as the presenceof outliers, mixturesin the data, or truncation
(censoring) in the data. Readersare referredto D'Agostino
Figure2 is a normalprobabilityplot of the dataof Section
5. The expected straightline can be obtainedby connecting
the + 's on the graph. Figure 3 contains a numberof forms
thatthe normalprobabilityplot will producein the presence
of nonnormality.For the present data set, its skewness to
the right is very evident in the plot.
A programfor the normal probabilityplot applicable to
SAS is part of the macro given in the appendix.
9. CONCLUSION
We have discussed the uses of \/N and b2 as descriptive
and inferentialstatisticsfor evaluatingthe normalityof data.
We have made specific recommendationsfor their uses.
Further we have reviewed briefly the normal probability
plot, which can be used in conjunctionwith V/W and b2 for
a graphical analysis. A good complete normality analysis
would consist of the use of the plot plus the statistics. The
use of these is superiorto what is routinelygiven in standard
computer software. Serious investigators should consider
using the materialsof this article in their data analysis.
APPENDIX: A MACRO FOR USE WITH SAS
STATISTICAL SOFTWARE
The following macro takes as input a variablename and
a data set name. It produces as output the results of a uni-
variatedescriptive analysis (PROC UNIVARIATE), skew-
ness andkurtosismeasuresandtest statistics,the D'Agostino-
Pearson omnibus normality test statistic, p levels, and a
normal probabilityplot.
%MACRONORMTEST(VAR,DATA);
PROCUNIVARIATENORMALPLOTDATA==&DATA;
VAR &VAR;OUTPUTOUT= XXSTATN = N
MEAN=XBAR STD=S SKEWNESS=G 1
KURTOSIS-=G2;

The American Statistician, November 1990, Vol. 44, No. 4 319

 375 +

*
350 +

* *

325+ *

H *

0 300- *

L~~~~~~~~~~~~~~~~~~~
E ~~~~~~~~~~~~~~~~~**
275 +
T**
E**
R ***
Q 250 +
**
L ~~~~~~~~~~~~~**
**

225 + **
**
**
* **+

200 + * **

*
*

175 +

150 +

-3 -2 -1 0 1 2 3
NORMALIZED
RANK
Figure2. NormalProbabilityPlot of CholesterolData.

DATA; Z_B 1= DELTA*LOG(Y/ALPHA

SET XXSTAT;
DO _Z_= -1,0,1; _X_=XBAR+_Z_*S;
END;
KEEP_X_ _Z_;
DATA;SET &DATA _LAST_;
PROC RANK TIES = MEAN NORMAL = BLOM; VAR
&VAR;RANKSBLOMRANK;
OPTIONS LS = 80;
PROCPLOTNOLEGEND;
PLOT &VAR*BLOMRANK= '*'
OVERLAYHAXIS= -3 TO 3 BY .5;
LABELBLOMRANK 'NORMALIZEDRANK"
&VAR= 'NORMALPROBABILITY PLOTFOR
&VAR";
DATA;
SET XXSTAT;
SQRTB 1= (N - 2)/SQRT(N*(N - 1))*G1;
Y = SQRTB1*SQRT((N+ 1)*(N+ 3)/(6*(N-2)));
BETA2= 3*(N*N+ 27*N- 70)*(N+ 1)*(N+ 3)/
((N - 2)*(N + 5)*(N + 7)*(N + 9));
W = SQRT(- 1 + SQRT(2*(BETA2-1)));
DELTA= 1/SQRT(LOG(W));
ALPHA = SQRT(2/(W*W - 1));
+ SQRT((Y/
ALPHA)**2+1));
OUTPUT; B2 = 3*(N -1)/(N + 1) + (N -2)*(N -3)/
((N + 1)*(N -1))*G2;
MEANB2=3*(N- 1)/(N+ 1);
VARB2 = 24*N*(N - 2)*(N - 3)/
((N + 1)*(N+ 1)*(N+ 3)*(N+ 5));
X = (B2 - MEANB2)/SQRT(VARB2);
MOMENT = 6*(N*N - 5*N + 2)/
((N + 7)*(N+ 9))*SQRT(6*(N+ 3)*(N+ 5)!
X_*Z= '+ '/ (N*(N - 2)*(N - 3)));
A = 6 + 8/MOMENT*(2/MOMENT
+ SQRT(1+ 4/
(MOMENT**2)));
Z_B2 = (1 - 2/(9*A) -((1 - 2/A)/(1 + X*SQRT(2/
(A-4))))**(1/3))/SQRT(2/(9*A));
PRZB1 = 2*(1-PROBNORM(ABS(Z_B1)));
PRZB2 2*(1- PROBNORM(ABS(Z_B2)));
CHITEST= Z_B 1*Z_B 1+ Z_B2*Z_B2;
PRCI-HI= 1- PROBCHI(CHITEST,2);
FILEPRINT;
PUT "NORMALITYTESTFOR VARIABLE&VAR"
N= /
@20 GI = 8.5 @33 SQRTB1=8.5 @50 "Z=" Z_
B1 8.5 "P=" PRZB1@6.4/

320 The American Statistician, November 1990, Vol. 44, No. 4

 z z
Indication: =/ ?2 3 = o, P?>3
Symmetric Symmetric
Thin Tails Thick Tails

x ~~~~x

Indication: Z> <0 z

Skewed to Right Skewed to Left

x ~~~~~~x x x

z z zz
Indication: Mixture Truncated Truncated Outlier
of Normals at Left at Right at Right

Figure3. Indicationsof DeviationsFromNormalityin a NormalProbabilityPlot.

@20 G2 = 8.5 @33 B2 = 8.5 @50 "Z=" ZB2 8.5" Balanda, K. P., and MacGillivray, H. L. (1988), "Kurtosis:A Critical
P= " PRZB26.4// Review," The American Statistician, 42, 111-1 19.
Blom, G. (1958), Statistical Estimates and Trantsformed Beta Variables,
@33 '(K**2=CmHSQ (2 DF) =" CHITEST 8.5 " P=" New York: John Wiley.
PRCHI6.4; D'Agostino, R. B. (1970), "Transformationto Normalityof the Null Dis-
%MENDNORMTEST; tributionof gl," Biometrika, 57, 679-681.
(1971), "An OmnibusTest of Normalityfor Moderateand Large
/* The SAS optionsMACRO,DQUOTEand *, Sample Size," Biometrika, 58, 341-348.
D'Agostino, R. B., and Pearson, E. S. (1973), "Testing for Departures
/* LINESIZE= 80 must *, From Normality. 1. Fuller Empirical Results for the Distributionof b,
/* be in effect. *, and \/ib," Biometrika, 60, 613-622.
I* *, D'Agostino, R. B., and Stephens, M. A. (1986), Goodness-of-fit Tech-
/* Exampleof a statementto executethe macroabove: *, niques, New York: Marcel Dekker.
/* %NORMTEST(CHOL,DATA1) Fisher, R. A. (1973), StatisticalMethodsfor Research Workers(14th ed.),
*,
New York: Hafner Publishing.
I* *, Kolmogorov, A. (1933), "Sulla DeterminazioneEmpirica di una Legge
[Received April 1989. Revised January 1990.] di Distribuzione," Giornalle dell'InstitutoItaliano degli Attuari, 4, 83-
91.
Pearson, K. (1895), "Contributionsto the MathematicalTheory of Evo-
REFERENCES lution," Philosophical Transactionsof the Royal Society of London, 9 1,
343-358.
Anscombe, F. J., and Glynn, W. J. (1983), "Distributionof the Kurtosis Shapiro, S. S., and Wilk, M. B. (1965), "An Analysis of VarianceTest
Statistic b, for Normal Statistics," Biometrika, 70, 227-234. for Normality (Complete Samples)," Biometrika, 52, 591-61 1.

The AmericanStatistician, November 1990, Vol. 44, No. 4 321