Download as pdf or txt
Download as pdf or txt
You are on page 1of 4

Should the Interquartile Range Divided by the Standard Deviation be Used to Assess

Normality?
Author(s): Richard L. Warr and Roger A. Erich
Source: The American Statistician , NOVEMBER 2013, Vol. 67, No. 4 (NOVEMBER 2013),
pp. 242-244
Published by: Taylor & Francis, Ltd. on behalf of the American Statistical Association
Stable URL: http://www.jstor.com/stable/24591487

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.

Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at
https://about.jstor.org/terms

American Statistical Association and Taylor & Francis, Ltd. are collaborating with JSTOR to
digitize, preserve and extend access to The American Statistician

This content downloaded from


190.237.55.7 on Tue, 11 Aug 2020 21:41:02 UTC
All use subject to https://about.jstor.org/terms
Should the Interquartile Range Divided by the Standard Deviation
be Used to Assess Normality?

Richard L. Warr and Roger A. Erich

have a similar (F_1(0.75) — F-1(0.25))/cr (where F_1 is the


We discourage the use of a diagnostic for normality: the in quantile function). Also we discuss the power of the test where
terquartile range divided by the standard deviation. This statistic the null hypothesis is: the data are normally distributed. When
has been suggested in several introductory statistics books as a compared with the Shapiro-Wilk test for normality (Shapiro
method to assess normality. Through simulation, we explore the and Wilk 1965), we conclude that the IQR/S is not a statistic
rate at which this statistic converges to its asymptotic normal that should be employed either as a formal test or an informal
distribution, and the actual size of tests based on the asymp rule of thumb.

totic distribution at several sample sizes. We show that there are


nonnormal distributions from which this method cannot detect
2. CONVERGENCE AND POWER COMPARISONS
a difference. Additionally, we show the power of this test for
normality is quite poor when compared with the Shapiro-Wilk In this section, we look at the asymptotic distribution o
test. statistic IQR/S given a sample from a normal distribution.
we consider how fast this statistic converges to its asymp
KEY WORDS: Hypothesis testing; Power; Shapiro-Wilk test.
distribution. We point out that there is no standard for calcu
the IQR of a dataset. Hyndman and Fan (1996) identified
methods to estimate quantiles. We use "Type 5" as defined
(R Development Core Team 2011) and referred to as "Defi
1. INTRODUCTION 5" in Hyndman and Fan (1996).
DasGupta and Haff (2006) showed that as n gets large
In an introductory statistics text (McClave, Benson, and Sin
IQR ( 1.566\
cich 2011, p. 219), the authors suggest a method to assess nor —~ N ( 1.35, ). (2)
mality. This method is the sample interquartile range (IQR)
divided by the sample standard deviation, which For
we denote
this as
asymptotic distribution, we examin
S. They stated if the data are approximately normal, then
ples are needed to assume that the asymptoti
reasonable fit. Using Monte Carlo simulation
^*1-34. (1) million samples of size n and estimate the qu
distribution when the sample size is equal to
We were unfamiliar with this assessment of normality
shown in until
Table 1. The table columns contain th
teaching from the above cited text. We have since found that
quantiles with different sample sizes. The "E
this method was also presented by Mendenhall and Sincich
contains the Monte Carlo estimates of the app
(2006), Kokoska (2008), and others. The fundamental problem
These Monte Carlo estimates all have 95% confidence intervals
with this rule of thumb is how close to 1.34 is acceptable to
with half-widths less than 0.0015. The "Asymptotic" column
conclude normality? The only real way to determine what is
shows the 0.025 and 0.975 quantiles for the asymptotic distribu
close is to find the statistic's distribution.
tion suggested in Equation (2). Clearly as n get larger, the limits
In this article, we demonstrate that this statistic should not
of the asymptotic results converge with the actual values. For a
be used as a measure to determine the "normal-ness" of data.
sample size of n = 10, there is a fairly large gap between the
Specifically, we address how fast the statistic converges to its
"Est actual" and "Asymptotic" upper limit. This gap narrows as
asymptotic distribution and other nonnormal distributions
n increases. that

Developing a hypothesis test for the IQR/.S' statistic could be


done in two ways. First, an approximate 0.05 level test can be
Richard L. Warr (E-mail: richard.warr@afit.edu) and Roger A. Erich (E-mail:
roger.erich.l .ctr@us.af.mil), Department of Mathematicsdefined with the rejection
and Statistics, Air region lying to the right of the 0.025
quantile
Force Institute of Technology, Wright-Patterson Air Force Base,andOH
to 45433
the left of the 0.975 quantile of the asymptotic
7765. The views expressed in this article are those of the distribution
authors and indo
Table
not1. The last column in Table 1, "Coverage,"
reflect the official policy or position of the United States Air Force,
shows theDepartment
percentage of samples from IQR/S"s asymptotic nor
of Defense, or the U.S. Government. The authors thank the editor, associate edi
mal distribution that are not rejected. An exact 0.05 level test
tor, and two anonymous referees for their suggestions; these suggestions greatly
could be developed for a particular sample size using Monte
improved the content of this article. The authors also thank Patrick Chapin and
Kevin Pond for their consultation and advice. Carlo simulation as with the "Est actual" columns in Table 1.

242 In the Public Domain DOI: 10.1080/00031305.2013.847385 The American Statistician, November 2013, Vol. 67, No. 4

This content downloaded from


190.237.55.7 on Tue, 11 Aug 2020 21:41:02 UTC
All use subject to https://about.jstor.org/terms
0.8 1.0 1.2 1.4 1.6 1.8 2.0 1.0 1.4
n=40 n=100

Figure 1. The asymptotic normal PDF shown with a histogram of one million Monte Carlo samples for four different sample sizes.

Figure 1 and Table 1 show that using the asymptotic distri


sampling distribution is skewed,
bution for a small sample size would not be appropriate. Later
distribution has fat tails, and the
bution
we show that regardless of the test (exact or approximate), the has (F~'(0.75) — F_1(0.25
IQR/.S' statistic does not produce powerful tests. (1) distribution for the skewed ca
Next we consider other distributions that have (F-1(0.75) —
tion for the case with heavy tails
F-1(0.25))/er «s 1.35, similar to the normal distribution. a Thedistribution with a similar (F-1
Table 2 shows the results of
IQR/.S' statistic is not able to reliably distinguish other distri
butions that have this same ratio. In other words, the prop
and Wei bull cases with three di
erty of (F-1(0.75) — F_1(.25))/ct r* 1.35 is not unique
theto IQR/5 test's power is quit
the normal distribution. For example, consider the Weibull Shapiro-Wilk test in both the ex
(1.67, ß) distribution. This distribution also has (F_1(0.75) —
these three cases, the Shapiro-Wil
F_1(0.25))/or ~ 1.35. Therefore, if we obtained a sample more from powerful than the IQR/5 t
this distribution, on average, this statistic could not distinguish
IQR/5 test performs better when d
it from a normal, even though it is right-skewed. However, if
skewness. However, these results
looking at a Q-Q plot, there may be an apparent way to distin for normality is inappropriate for
guish the difference. Figure 2 shows the Q-Q plot of a sample of
size n — 100 from a Weibull distribution with IQR/.S' = 1.32. 3. DISCUSSION
The plot would lead us to conclude that these data have tails
different from a normal distribution. From the previous section, it is clear that the IQR/S m
To compare the power of this test, we look at three cases
for assessing normality is severely limited in its applica
when the sampling distribution is not normal. We ask how well
presume, in most cases, it is much less powerful than m
does the statistic IQR/S reject the null hypothesis of normality
ular tests for normality. The power of the test is low
when the data are not normal? We use the Shapiro-Wilkthat test any nonnormality might be detected in a histogr
(Shapiro and Wilk 1965) as a comparison, although thereQ-Q are plot before this test is sensitive enough to reject t
many others to choose from. We consider one case where the

Table 2. Proportion of samples where the hypothesis of norm


Table 1. Quantiles of the asymptotic distribution and the estimated
rejected
exact values, along with the percentage of samples that fall within the
asymptotic quantiles
Exponential 1) Cauchy(O.l) Weibull( 1.67,1)

0.025 Quantile 0.975 Quantile n IQRSa IQRSe S-W IQRSa IQRSe S-W IQRSa IQRSe S-W

n Asymptotic
Asymptotic
Est actual
EstAsymptotic
actual Asymptotic
Est actual %Est actual %10
Coveraj 0.076 0.124 0.444 0.481 0.521 0.591 0.026 0.063 0.132
Covera;
20 0.136 0.190 0.836 0.822 0.851 0.866 0.034 0.072 0.288
10 0.574
0.5740.625
0.625 2.126
2.126 1.976
1.976 97.94
97.94 40 0.246 0.273 0.995 0.984 0.986 0.988 0.055 0.076 0.605
20 0.802
0.8020.856
0.856 1.898
1.898 1.807
1.807 97.84
97.84 100 0.469 0.477 1.000 1.000 1.000 1.000 0.069 0.071 0.982
40 0.962
0.9620.980
0.980 1.738
1.738 1.703
1.703 96.51
96.51
100 1.105
1.105 1.108
1.108 1.595
1.595 1.595
1.595 95.62
95.62 NOTE: IQRSa is the test based on the asymptotic distribution of IQR/.V and IQRS
exact test.

The American Statistician, November 2013, Vol. 67, No. 4

This content downloaded from


190.237.55.7 on Tue, 11 Aug 2020 21:41:02 UTC
All use subject to https://about.jstor.org/terms
Normal Q-Q Plot

O
c\i

Q)

'■E LO

1 9
E ^
03
CO
LO
d
Ö

Theoretical Quantiles

Figure 2. The normal Q-Q plot of a sample of size 100 from a Weibull (1.67, 1) distribution.

Hyndman,
hypothesis. Additionally, the statistic cannot distinguish the dif R. J., and Fan, Y. (1996), "Sample Quantiles in Statistica
ference between the normal and other distributions that have The American Statistician, 50, 361-365. [242]

similar (^"'(0.75) — F"'(0.25))/ct ratios. We recommend that S. (2008), Introductory Statistics (Preliminary Edition
Kokoska,
Solving Approach, New York: WH Freeman & Co. [242]
this method should not be advocated, especially in introductory
textbooks. McClave, J., Benson, P., and Sincich, T. (2011), Statistics for B
Economics (11th ed.), Boston, MA: Pearson. [242]
[Received December 2012. Revised July 2013.]
Mendenhall, W„ and Sincich, T. (2006), Statistics for Engineeri
Sciences (5th ed.), Upper Saddle River, NJ: Prentice-Hall, Inc.

REFERENCES R Development Core Team (2011), R: A Language and Environme


tical Computing, Vienna, Austria: R Foundation for Statistical C
Available at http://www.R-project.orgl [242]
DasGupta, A., and Haff, L. (2006), "Asymptotic Values and Expansions for the
Correlation Between Different Measures of Spread," Journal of S.,
Shapiro, Statistical
and Wilk, M. (1965), "An Analysis of Variance Test for
Planning and Inference, 136, 2197—2212. [242] (Complete Samples)," Biometrika, 52, 591-611. [242,243]

244

This content downloaded from


190.237.55.7 on Tue, 11 Aug 2020 21:41:02 UTC
All use subject to https://about.jstor.org/terms

You might also like