See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/309225521

Univariate and multivariate skewness and kurtosis for measuring

nonnormality: Prevalence, influence and estimation

Article  in  Behavior Research Methods · October 2016

DOI: 10.3758/s13428-016-0814-1

CITATIONS READS

259 1,353

3 authors:

Meghan K. Cain Zhiyong Zhang

StataCorp University of Notre Dame
9 PUBLICATIONS   310 CITATIONS    87 PUBLICATIONS   1,896 CITATIONS   

SEE PROFILE SEE PROFILE

Ke-Hai Yuan
University of Notre Dame
176 PUBLICATIONS   7,507 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Exploratory factor analysis View project

SEM diagnosis View project

All content following this page was uploaded by Meghan K. Cain on 15 June 2018.

The user has requested enhancement of the downloaded file.

Behav Res (2017) 49:1716–1735
DOI 10.3758/s13428-016-0814-1

Univariate and multivariate skewness and kurtosis

for measuring nonnormality: Prevalence, influence
and estimation
Meghan K. Cain1 · Zhiyong Zhang1 · Ke-Hai Yuan1

Published online: 17 October 2016

© Psychonomic Society, Inc. 2016

Abstract Nonnormality of univariate data has been Almost all commonly used statistical methods in psychol-
extensively examined previously (Blanca et al., Method- ogy and other social sciences are based on the assumption
ology: European Journal of Research Methods for the that the collected data are normally distributed. For exam-
Behavioral and Social Sciences, 9(2), 78–84, 2013; Miceeri, ple, t- and F-distributions for mean comparison, Fisher Z-
Psychological Bulletin, 105(1), 156, 1989). However, less transformation for inferring correlation, and standard errors
is known of the potential nonnormality of multivariate data and confidence intervals in multivariate statistics are all
although multivariate analysis is commonly used in psy- based on the normality assumption (Tabachnick & Fidell,
chological and educational research. Using univariate and 2012). Researchers rely on these methods to accurately
multivariate skewness and kurtosis as measures of nonnor- portray the effects under investigation, but may not be
mality, this study examined 1,567 univariate distriubtions aware that their data do not meet the normality assump-
and 254 multivariate distributions collected from authors of tion behind these tests or what repercussions they face when
articles published in Psychological Science and the Amer- the assumption is violated. From a methodological perspec-
ican Education Research Journal. We found that 74 % of tive, if quantitative researchers know the type and severity
univariate distributions and 68 % multivariate distributions of nonnormality that researchers are facing, they can exam-
deviated from normal distributions. In a simulation study ine the robustness of normal-based methods as well as
using typical values of skewness and kurtosis that we col- develop new methods that are better suited for the analysis
lected, we found that the resulting type I error rates were of nonnormal data. It is thus critical to understand whether
17 % in a t-test and 30 % in a factor analysis under some practical data satisfy the normality assumption and if not,
conditions. Hence, we argue that it is time to routinely report how severe the nonnormality is, what type of nonnormal-
skewness and kurtosis along with other summary statistics ity it is, what the consequences are, and what can be done
such as means and variances. To facilitate future report of about it.
skewness and kurtosis, we provide a tutorial on how to com- To understand normality or nonnormality, we need to
pute univariate and multivariate skewness and kurtosis by first define a measure of it. Micceri (1989) evaluated
SAS, SPSS, R and a newly developed Web application. deviations from normality based on arbitrary cut-offs of
various measures of nonnormality, including asymmetry,
Keywords Nonnormality · Skewness · Kurtosis · Software tail weight, outliers, and modality. He found that all 440
large-sample achievement and psychometric measures dis-
tributions were nonnormal, 90 % of which had sample sizes
larger than 450. More recently, Blanca et al. (2013) evalu-
ated nonnormality using the skewness and kurtosis1 of 693
 Meghan K. Cain
[email protected]
1 Withoutspecific mention, the skewness and kurtosis refer to the
1 University of Notre Dame, Notre Dame, IN, 46556, USA sample skewness and kurtosis throughout the paper.
Behav Res (2017) 49:1716–1735 1717

small samples, with sample size ranging from 10 to 30. including a guideline on how to report sample statistics in
The study includes many psychological variables, and the empirical research and some possible solutions for nonnor-
authors found that 94.5 % of distributions were outside the mality.
range of [-0.25, 0.25] on either skewness or kurtosis and
therefore violated the normality assumption. However, nei-
ther Micceri nor Blanca et al. discuss the distribution of Univariate and multivariate skewness and kurtosis
skewness or kurtosis, how to test violations of normality,
or how much effect they can have on the typically used Different formulations for skewness and kurtosis exist in
methods such as t-test and factor analysis. the literature. Joanes and Gill (1998) summarize three com-
Scheffe (1959, p.333) has commented that kurtosis and mon formulations for univariate skewness and kurtosis that
skewness are “the most important indicators of the extent they refer to as g1 and g2 , G1 and G2 , and b1 and b2 . The
to which nonnormality affects the usual inferences made in R package moments (Komsta & Novomestky, 2015), SAS
the analysis of variance.” Skewness and kurtosis are also proc means with vardef=n, Mplus, and STATA report g1 and
an intuitive means to understand normality. If skewness is g2 . Excel, SPSS, SAS proc means with vardef=df, and SAS
different from 0, the distribution deviates from symmetry. proc univariate report G1 and G2 . Minitab reports b1 and b2 ,
If kurtosis is different from 0, the distribution deviates from and the R package e1071 (Meyer et al., 2015) can report all
normality in tail mass and shoulder (DeCarlo, 1997b).2 three. There are also several measures of multivariate skew-
In practice, normality measures such as skewness and ness and kurtosis, though Mardia’s measures (Mardia, 1970)
kurtosis are rarely reported. In order to study nonnormal- are by far the most common. These are currently available
ity, we have contacted and obtained responses from 124 in STATA, or as add-on macros multnorm in SAS or mardia
researchers, among whom only three reported skewness in SPSS (DeCarlo, 1997a).
and kurtosis in their papers. The under-report of normal-
ity measures can be due to several reasons. First, many Univariate skewness and kurtosis
researchers are still not aware of the prevalence and influ-
ence of nonnormality. Second, not every researcher is famil- For the univariate case, we adopt Fisher’s skewness (G1 )
iar with skewness and kurtosis or their interpretation. Third, and kurtosis (G2 ). Specifically, the skewness, G1 , is calcu-
extra work is needed to compute skewness and kurtosis lated as
than the commonly used summary statistics such as means √
and standard deviations. Fourth, researchers might worry n(n − 1) m3
G1 = · 3/2 , (1)
about the consequences of reporting large skewness and n−2 m2
kurtosis.
This paper provides a simple and practical response to the and the kurtosis, G2 , as
continuing under-report of nonnormality measures in pub-
lished literature by elucidating the problem of nonnormality    
n−1 m4
and offering feasible recommendations. We begin with an G2 = · (n + 1) −3 +6 , (2)
easy-to-follow introduction to univariate and multivariate (n − 2)(n − 3) m22
skewness and kurtosis, their calculations, and interpreta- 
tions. We then report on a review we conducted assessing where mr = ni=1 (xi − x̄)r /n is the rth central moment
the prevalence and severity of univariate and multivariate with x̄ being the sample mean and n the sample size. The
skewness and kurtosis in recent psychology and education sample skewness G1 can take any value between negative
publications. We also show the influence of skewness and infinity and positive infinity. For a symmetric distribution
kurtosis on commonly used statistical tests in our field using such as a normal distribution, the expectation of skewness is
data of typical skewness, kurtosis, and sample size found 0. A non-zero skewness indicates that a distribution “leans”
in our review. In addition, we offer a tutorial on how to one way or the other and has an asymmetric tail. Distribu-
compute the skewness and kurtosis measures we report here tions with positive skewness have a longer right tail in the
through commonly used software including SAS, SPSS, R, positive direction, and those with negative skewness have a
and a Web application. Finally, we offer practical recom- longer left tail in the negative direction.
mendations for our readers to follow in their own research, Figure 1 portrays three distributions with different values
of skewness. The one in the middle is a normal distribution
and its skewness is 0. The one on the left is a lognormal
2 Kurtosis measures can be centered at either 0 or 3, the former is distribution with a positive skewness = 0.95. A commonly
usually referred to as “excess kurtosis”. This is because the normal used example of a distribution with a long positive tail is
distribution has a kurtosis of 3, and therefore an excess kurtosis of 0. the distribution of income where most households make
1718 Behav Res (2017) 49:1716–1735

expectation of kurtosis of a normal distribution is 0. If a dis-

0.15
Lognormal (skewness=0.95)
Normal (skewness=0)
Skew-normal (skewness= -0.3) tribution is leptokurtic, meaning it has positive kurtosis, the
distribution has a fatter tail than the normal distribution with
the same variance. Generally speaking, if a data set is con-
0.10

taminated or contains extreme values, its kurtosis is positive.

density

If a distribution is platykurtic, meaning it has negative kur-

tosis, the distribution has a relatively flat shoulder and short
tails (e.g., see Fig. 2b). For example, the distribution of age
0.05

of the US population has negative kurtosis because there are

generally the same number of people at each age.5
Because for a normal distribution both skewness and
0.00

kurtosis are equal to 0 in the population, we can conduct

0 5 10 15 20 25 30 hypothesis testing to evaluate whether a given sample devi-
Fig. 1 Illustration of positive and negative skewness ates from a normal population. Specifically, the hypothesis
testing can be conducted in the following way.6 We first cal-
culate the standard errors of skewness (SES) and kurtosis
around $53,000 a year3 and fewer and fewer make more. In (SEK) under the normality assumption (Bliss, 1967, p.144-
psychology, typical response time data often show positive 145),
skewness because much longer response time is less com- 
mon (Palmer et al., 2011). The distribution on the right in 6n(n − 1)
SES = , (3)
Fig. 1 is a skew-normal distribution with a negative skew- (n − 2)(n + 1)(n + 3)
ness = -0.3. For example, high school GPA of students who 
apply for colleges often shows such a distribution because n2 − 1
SEK = 2(SES) . (4)
students with lower GPA are less likely to seek a college (n − 3)(n + 5)
degree. In psychological research, scores on easy cognitive
tasks tend to be negatively skewed because the majority of Note that the standard errors are functions of sample size. In
participants can complete most tasks successfully (Wang particular, standard error decreases as sample size increases,
et al., 2008). and the strictness with which we call a distribution “normal”
Kurtosis is associated with the tail, shoulder and peaked- becomes more and more rigid. This is a natural conse-
ness of a distribution. Generally, kurtosis increases with quence of statistical inference. With these standard errors,
peakedness and decreases with flatness. However, as two statistics,
DeCarlo (1997b) explains, it has as much to do with the ZG1 = G1 /SES (5)
shoulder and tails of a distribution as it does with the
peakedness. This is because peakedness can be masked by and
variance. Figure 2a and b illustrate this relationship clearly.
Figure 2a shows the densities of three normal distribu- ZG2 = G2 /SEK, (6)
tions each with kurtosis of 0 but different variances, and
Fig. 2b shows three distributions with different kurtosis but can be formed for skewness and kurtosis, respectively. Both
the same variance. Normal distributions with low variance of these statistics can be compared against the standard
have high peaks and light tails as in Fig. 2a, while distri- normal distribution, N(0, 1), to obtain a p-value to test
butions with high kurtosis have high peaks and heavy tails a distribution’s departure from normality (Bliss, 1967). If
as in Fig. 2b. Hence, peakedness alone is not indicative of there is a significant departure, the p-value is smaller than
kurtosis, but rather it is the overall shape that is important. .05 and we can infer that the underlying population is non-
Kurtosis must increase as skewness increases, because of normal. If neither test is significant, there is not enough
the relationship: kurtosis ≥ skewness2 − 2 (Shohat, 1929).
Kurtosis has a range of [−2(n − 1)/(n − 3), ∞) in a sam- 5 Based the estimated population by the US Census
ple of size n and a range of [-2,∞] in the population.4 The (https://www.census.gov/popest/data/national/asrh/2014/files/
NC-EST2014-ALLDATA-R-File01.csv), the kurtosis of age is -0.844.
6 Other hypothesis testing methods for skewness and kurtosis are avail-

able (Anscombe and Glynn, 1983; D’Agostino, 1970). The reason for
3 The inflation adjusted medium household income is $53,657 in 2014
adopting the method discussed here is that the standard errors of skew-
based on census. ness and kurtosis are reported in popular statistical software such as
4 Note that if g = m /m2 − 3 is used to estimate kurtosis it also has a SPSS and SAS, and, therefore, it is a feasible method for evaluating
2 4 2
minimum value of -2. skewness and kurtosis through existing software.
Behav Res (2017) 49:1716–1735 1719

(a) (b)
variance=0.5 kurtosis=3
variance=1 kurtosis=0

0.5
variance=2 kurtosis=−1

0.4
0.4

0.3
density
density
0.3

0.2
0.2

0.1
0.1
0.0

0.0
−4 −2 0 2 4 −4 −2 0 2 4

Fig. 2 Illustration of the relationship between kurtosis and variance. In Fig 2a each population has a kurtosis of 0, and variance varies from 0.5
to 2.0. In Fig 2b each population has a variance of 1, and kurtosis varies from -1 to 3.

evidence to reject normality based on skewness or kurtosis Two statistics,

although it may still be nonnormal in other characteristics.
n
z1,p = b1,p (10)
6
Multivariate skewness and kurtosis
and
The univariate skewness and kurtosis have been extended to
multivariate data. Multivariate skewness and kurtosis mea- b2,p − [p(p + 2)(n − 1)] /(n + 1)
z2,p = √ (11)
sure the same shape characteristics as in the univariate case. [8p(p + 2)] /n
However, instead of making the comparison of the distribu-
can be formed for multivariate skewness and kurtosis,
tion of one variable against a univariate normal distribution,
respectively. The statistic z1,p can be compared against
they are comparing the joint distribution of several variables 2
the chi-squared distribution χp(p+1)(p+2)/6 , and the statistic
against a multivariate normal distribution.
z2,p can be compared against the standard normal dis-
In this study, we use Mardia’s measures (Mardia, 1970)
tribution N(0, 1) to test a distribution’s departure from
of multivariate skewness and kurtosis, because they are
normality. If the test statistic z1,p is significant, e.g. the p-
most often included in software packages. Mardia defined
value is smaller than .05, the joint distribution of the set
multivariate skewness and kurtosis, respectively, as
of p variables has significant skewness; if the test statistic
1  3
n n
−1

z2,p is significant, the joint distribution has significant kur-

b1,p = 2
(xi − x̄)  S xj − x̄ , (7) tosis. If at least one of these tests is significant, it is inferred
n
i=1 j =1 that the underlying joint population is nonnormal. As in the
1  2
n univariate case, non-significance does not necessarily imply
−1
b2,p = (xi − x̄)  S (xi − x̄) , (8) normality.
n
i=1

where x is a p × 1 vector of random variables and S is the

biased sample covariance matrix of x defined as Review of skewness and kurtosis in practical data

1 
n
 Although Micceri (1989) and Blanca et al. (2013) have
S= (xi − x̄)(xi − x̄) . (9)
n studied univariate nonnormality, we are not aware of any
i=1
study that has investigated multivariate skewness and kur-
Both measures have a p subscript, so they are specific to a tosis with empirical data or has tested the significance of
set of p variables. The expected Mardia’s skewness is 0 for nonnormality. Therefore, we conducted a study to further
a multivariate normal distribution and higher values indi- evaluate the severity of nonnormality of empirical data,
cate a more severe departure from normality. The expected especially in the multivariate case. Focusing on published
Mardia’s kurtosis is p(p + 2) for a multivarite normal research, we contacted 339 researchers with publications
distribution of p variables. As in the univariate case, val- that appeared in Psychological Science from January 2013
ues under this expectation indicate platykurtism and higher to June 2014 and 164 more researchers with publications
values indicate leptokurtism. that appeared in the American Education Research Journal
1720 Behav Res (2017) 49:1716–1735

from January 2010 to June 2014. The two journals were cho- Because these most extreme values may be outliers, we also
sen due to their prestige in their corresponding fields. We report 1st through 99th percentiles of univariate skewness
asked the researchers to provide the univariate and multi- and kurtosis. Percentiles can be interpreted as the percent
variate skewness and kurtosis of continuous variables used of samples with lower skewness or kurtosis than that value.
in their papers. Binary, categorical, and nominal variables There is clearly a large range from the 1st to the 99th
were excluded, but likert items were included because they percentile, especially for kurtosis. The correlation between
are often treated as normal in data analysis. To help the sample size and skewness is r = −0.005, and with kurtosis
researchers compute the skewness and kurtosis, we provided is r = 0.025. These are comparable to what Blanca et al.
a tutorial for different software as we will present later in (2013) have reported in which correlations between sample
this paper. Our data collection ended in November, 2014, by size and skewness and kurtosis were .03 and -.02, respec-
which point we had obtained 1,567 univariate measures and tively. The results in Table 1 include skewness and kurtosis
254 multivariate measures of skewness and kurtosis from when the sample size is smaller and larger than 106, the
194 studies. Some authors submitted univariate results with- median sample size of all collected data. As shown in this
out multivariate results so some variables for which we have table, negative skewness and kurtosis are much more com-
univariate measures are not included in a multivariate mea- mon than previously reported: 38 % of distributions have
sure. The median sample size for these studies was 106, negative skewness and 47 % have negative kurtosis. This
and the sample size ranged from 10 to 200,000. The median could be due to the number of likert measures provided, but
number of variables included in a multivariate measure was because of the anonymous submission option this cannot
3, and ranged from 1 to 36. Since researchers had the option be confirmed. The mean univariate skewness is 0.51, and
to submit skewness and kurtosis anonymously, it is unclear the sample size-weighted mean is 0.47. The mean univari-
how many authors responded to our request or what their ate kurtosis is 4.29, and the sample size-weighted mean is
study characteristics may be. 8.41. Sample size-weighted means are helpful because we
expect sample measures to better-reflect that of the popula-
Univariate skewness and kurtosis tion as sample size increases. To account for this, measures
from large samples are given higher weight than those from
As shown in Table 1, univariate skewness ranged from smaller samples. Therefore, on average, the skewness and
-10.87 to 25.54 and univariate kurtosis from -2.20 to kurtosis are larger than that of a normal distribution. To
1,093.48, far wider than previously reported or tested. further visualize what these distributions look like, Fig. 3

Table 1 Univariate skewness and kurtosis

n ≤ 106 n > 106 Overall

Percentile Skewness Kurtosis Skewness Kurtosis Skewness Kurtosis

(a) Skewness and kurtosis by sample size

Minimum −4.35 −2.20 −10.87 −1.99 −10.87 −2.20
1st −1.68 −1.79 −2.68 −1.56 −2.08 −1.70
5th −1.10 −1.28 −1.27 −1.28 −1.17 −1.28
25th −0.33 −0.60 −0.33 −0.52 −0.33 −0.57
Median 0.27 0.02 0.15 0.12 0.20 0.07
75th 0.91 1.35 1.00 2.12 0.94 1.62
95th 2.25 5.89 3.56 19.39 2.77 9.48
99th 4.90 30.47 10.81 154.60 6.32 95.75
Maximum 6.32 40.00 25.54 1,093.48 25.54 1,093.48

(b)Percent of significant skewness and kurtosis by sample size

n ≤ 106 n > 106 Overall
Skewness 51 82 66
Kurtosis 33 77 54
Either 56 95 74

There were 805 distributions with n ≤ 106 and 762 with n > 106. Nonnormality is defined by significant statistics ZG1 or ZG2 , p < .05
Behav Res (2017) 49:1716–1735 1721

Fig. 3 Histograms of 20 randomly selected distributions collected for review

shows histograms of 20 randomly selected distributions 1,476. Percentiles of Mardia’s skewness and kurtosis split
from our review. Note that there is no common shape that by median sample size and median number of variables
explains skewness or kurtosis. used in their calculation are presented in Table 2. The
Percentages of univariate distributions with significant correlation between sample size and Mardia’s skewness
skewness or kurtosis by sample size are presented in Table 1. is r = −0.01 and with Mardia’s kurtosis is r = 0.02.
About 66 % of univariate distributions had significant skew- The correlation between the number of variables and Mar-
ness and 54 % had significant kurtosis. Almost 74 % of dis- dia’s skewness is r = 0.58 and with Mardia’s kurtosis is
tributions had either significant skewness or kurtosis and were r = 0.73. After centering Mardia’s kurtosis on p(p + 2),
therefore classified as nonnormal. As expected, it becomes the expected value under normality, the correlation between
easier for tests to become significant with larger sample sizes. kurtosis and the number of variables becomes r = 0.05.
Over 95 % of distributions with sample sizes greater than the The mean multivariate skewness is 32.94, and the sam-
median sample size, 106, were tested as nonnormal. ple size-weighted mean is 28.26. The mean multivariate
Conversely, when the sample size was less than 106 only kurtosis is 78.70, and the sample size-weighted mean is
56 % of distributions were significantly nonnormal. 92.03. Therefore, the average skewness and kurtosis are
greater than that of a multivariate normal distribution. This
Multivariate Skewness and Kurtosis has important ramifications especially for SEM, for which
multiple outcome measures are often used and for which
The 254 collected Mardia’s multivariate skewness ranged multivariate kurtosis can asymptotically affect standard
from 0 to 1,332 and multivariate kurtosis from 1.80 to errors.
1722 Behav Res (2017) 49:1716–1735

Table 2 Mardia’s measures by sample size and number of variables

Percentile By Sample Size By # of Variables Overall

n ≤ 106 n > 106 p≤3 p>3

(a) Mardia’s Skewness

Minimum 0.01 0.00 0.00 0.02 0.00
1st 0.03 0.00 0.00 0.43 0.00
5th 0.23 0.02 0.03 1.08 0.035
25th 1.15 0.35 0.33 5.72 0.76
Median 3.04 3.26 1.14 1.40 3.08
75th 13.91 14.92 2.95 44.43 14.32
95th 124.97 107.54 23.97 211.31 112.82
99th 635.90 496.77 343.60 786.84 610.66
Maximum 1,263.60 796.92 496.77 1,263.60 1,263.60

(b) Mardia’s Kurtosis

b2,p ∗
b2,p b2,p ∗
b2,p b2,p ∗
b2,p b2,p ∗
b2,p b2,p ∗
b2,p
Minimum 2.19 −90.50 1.99 −18.57 2.00 −7.77 15.09 −90.50 1.99 −90.50
1st 2.23 −61.02 2.34 −15.43 2.20 −7.72 18.90 −63.61 2.23 −54.55
5th 3.35 −23.59 2.79 −7.51 2.39 -3.74 22.26 −30.83 2.92 −17.01
25th 8.08 −2.33 8.81 0.26 7.02 −0.82 37.76 −2.38 8.26 −1.35
Median 14.24 −0.70 31.69 5.37 8.71 0.26 60.86 5.55 18.90 0.59
75th 43.00 2.22 90.89 29.32 14.84 2.34 153.3 27.36 56.69 7.47
95th 190.1 28.18 419.4 179.25 52.69 44.54 614.4 119.3 323.1 98.17
99th 942.6 87.45 755.4 732.9 384 369 1,356 719.4 914.9 541
Maximum 1,476 108.1 1,392 1,368 556 541 1,476 1,368 1,476 1,368

∗ is b
There were 136 multivariate distributions with n ≤ 106, 118 with n > 106, 144 with p ≤ 3, and 110 with p > 3. b2,p 2,p centered on p(p +2)

Percentages of multivariate distributions with significant were statistically significant when the sample size was
Mardia’s skewness and kurtosis are presented in Table 3. larger than 106. Similarly, more Mardia’s measures became
About 58 % of multivariate skewness measures and 57 % of significant with more variables.
multivariate kurtosis measures reached significance. Com- To summarize, based on the test of 1,567 univariate
bining these, 68 % of multivariate distributions were signif- and 254 multivariate skewness and kurtosis from real data,
icantly nonnormal. In particular, 94 % of Mardia’s measures we conclude that 74 % of univariate data and 68 % of

Table 3 Percent significant Mardia’s skewness and kurtosis at significance level 0.05

By Sample Size By # of Variables Overall

n ≤ 106 n > 106 p≤3 p>3

Skewness 34 86 53 65 58
Kurtosis 35 82 47 70 57
Either 46 94 60 79 68

There were 136 multivariate distributions with n ≤ 106, 118 with n > 106, 144 with p ≤ 3, and 110 with p > 3. Nonnormality is defined by
significant statistics z1,p or z2,p , p < .05
Behav Res (2017) 49:1716–1735 1723

multivariate data significantly deviated from a univariate or specified skewness and kurtosis from a Pearson distribution
multivariate normal distribution. In examining only those in R (R Core Team, 2016) using the package PearsonDS
univariate measures included in a multivariate measure, (Becker & Klößner, 2016).7 Then, we obtained the empir-
68 % have significant nonnormality. Therefore, nonnormal- ical type I error rate to reject the null hypothesis that
ity is a severe problem in real data, though multivariate the population mean is equal to 0 using the significance
nonnormality does not appear to be a severe problem above level 0.05 in a two-tailed, a lower-tail, and an upper-tail
and beyond that of univariate normality. However, this one-sample t-test.
relationship requires further study to evaluate. Table 4 displays the empirical type I error rate for each
condition. For brevity, type I error rates of just the low-
est sample size are presented for conditions with skewness
Influences of skewness and kurtosis between -1.17 and 0.94 because these conditions did not
present any problems. To better understand the empirical
In order to clearly show the influence of skewness and kur- type I error rate, we bold those that are outside of the range
tosis, we conducted simulations on the one-sample t-test, [0.025, 0.075]. When the skewness and kurtosis are 0, the
simple regression, one-way ANOVA, and confirmatory fac- generated data are from a normal distribution and the empir-
tor analysis (CFA). Simulation studies are useful because ical type I error rate is close to 0.05 even when the sample
when data is generated from a specific model we know what size is as small as 18 for all three tests. When data devi-
results the statistical tests should show, and so we can evalu- ate from normality, the results show that a two-sided test is
ate how nonnormality affects those results. Note that for all more robust than a one-sided test. The two-sided test only
of these models, the interest is in the normality of the depen- has increased type I error rate for a skewness of 6.32, for
dent variable(s). There are no normality assumptions put on which a sample size of 554 is necessary to dissipate the
the independent variables. effect. A lower tail t-test has even higher type 1 error rates at
this skewness, and an upper tail t-test has an increased type
Influence of univariate skewness and kurtosis I error rate with negative skewness and very low rates with
high positive skewness.
Yuan et al. (2005) show that the properties of mean esti- A simple regression and a one-way ANOVA with three
mates are not affected by either skewness or kurtosis asymp- groups were also tested at all of these conditions. The
totically, but that the standard error of sample variance regression was robust to all conditions, even at the lowest
is a function of kurtosis. If normality is assumed (kur- sample size. Type 1 error rate for the ANOVA gets as low as
tosis = 0), the standard error of sample variance will be 0.022 under the most extreme skewness (6.32) at the lowest
underestimated when kurtosis is positive and overestimated sample size (18). However, all other type I error rates were
when kurtosis is negative. In other words, kurtosis will still within the [0.025,0.075] robustness range. Type I error rates
have an effect on variance estimates at very large sam- can increase if each population is from a different distribu-
ple sizes while mean estimates are only affected in small tion, but as long as each distribution has equal variances the
samples. For example, Yanagihara and Yuan (2005) found departures are still not too severe. Influence on power, how-
that the expectation and variance of the t-statistic depends ever, can be immense (See Levine and Dunlap (1982), for
on skewness, but that the effect lessens as sample size example).
increases.
To concretely demonstrate the influence of univariate Influence of multivariate skewness and kurtosis
skewness and kurtosis, we conducted a simulation study on
a one-sample t-test. In the simulation, we set the skewness In order to show the influence of multivariate skewness
to the 1st, 5th, 25th, 50th, 75th, 95th, and 99th percentiles and kurtosis, we conducted simulation studies on CFAs.
of univariate skewness found in our review of practical First, we focus on a one-factor model with four manifest
data. These were tested in sample sizes of the 5th, 25th, variables. For each manifest variable, the factor loading
50th, 75th, and 95th percentiles of sample size found in our is fixed at 0.8 and the uniqueness factor variance is 0.36.
review. Therefore, these conditions should represent typi- The variance of the factor is set to 1. Note that the expected
cal results found in our field. Because kurtosis has little Mardia’s kurtosis is p(p + 2). When kurtosis = 24 data are
influence on the t-test, it was kept at the 99th percentile,
95.75, throughout all conditions. In total, we considered 35 7 Pearson distribution includes a class of distributions. It is used here
conditions for each test. Under each condition, we gener- because it allows us to vary the skewness and kurtosis while the mean
ated 10,000 sets of data with mean 0, variance 1, and the and variance remain the same.
1724 Behav Res (2017) 49:1716–1735

Table 4 Type I error rates of the one-sample t-test

Tail Tested

Sample Size Skewness Two-tailed Lower-tail Upper-tail

18 −2.08 0.057 0.029 0.079

48 −2.08 0.055 0.033 0.072
105 −2.08 0.052 0.037 0.065
555 −2.08 0.050 0.043 0.058
1488 −2.08 0.050 0.046 0.057
18 −1.17 0.048 0.035 0.064
18 −0.33 0.046 0.045 0.053
18 0.20 0.045 0.051 0.046
18 0.94 0.049 0.061 0.038
18 2.77 0.064 0.092 0.023
48 2.77 0.060 0.082 0.027
105 2.77 0.056 0.075 0.031
555 2.77 0.050 0.062 0.039
1488 2.77 0.052 0.059 0.045
18 6.32 0.177 0.216 0.005
48 6.32 0.123 0.157 0.011
105 6.32 0.090 0.120 0.016
555 6.32 0.062 0.081 0.028
1488 6.32 0.055 0.069 0.035

Bolded entries are those outside of the range [0.025,0.075] and are therefore considered different from the nominal 0.05

from a multivariate normal distribution, and the centered review of practical data of four manifest variables (30, 60,
kurtosis is 0. Although in our review of practical data about and 100). Sample sizes of 48, 106, 554, and 1489 were used
half of the data sets had centered Mardia’s kurtosis less than to evaluate these conditions. Because skewness does not
0, 21 is the only multivariate kurtosis less than 24 we were influence SEM asymptotically, it was kept at 0 throughout
able to successfully simulate. Hence, we used these two val- all conditions. In total, 20 conditions were considered. 1,000
ues of Mardia’s kurtosis (21 and 24) along with the 75th, data sets were used to evaluate each condition. The authors
95th, and 99th percentiles of Mardia’s kurtosis found in our are currently unaware of any method to simulate data with

Table 5 Type I error rates of the χ 2 test for factor analysis with 4 manifest variables

Sample Size

Kurtosis Centered Kurtosis 48 106 554 1489

21 −3 0.061 0.058 0.060 0.060

24 0 0.053 0.046 0.048 0.050
30 6 0.055 0.052 0.055 0.056
60 36 0.108 0.121 0.149 0.152
100 76 0.161 0.215 0.287 0.298

Bolded entries are those outside of the range [0.025,0.075] and are therefore considered different from the nominal 0.05
Behav Res (2017) 49:1716–1735 1725

a particular multivariate skewness and kurtosis, so instead The same conditions were tested as in the simulation study
we used the R package lavaan (Rosseel, 2012) to simu- above, with the exception of those with a sample size of 48.
late data from a model with certain univariate skewness and The same univariate kurtoses were used to simulate the data,
kurtosis. Appropriate univariate values were found to simu- though they result in different multivariate kurtosis for eight
late multivariate values of a population by trial and error. variables than they do for four. The resulting empirical type
The influence of skewness and kurtosis is evaluated I error rates of these multivariate kurtoses for both of these
through the empirical type I error rate of rejecting the fac- models are given in Table 6.
tor model using the normal-distribution-based chi-squared Once again, type I error is maintained when the distribu-
goodness-of-fit test. This test is significant when the model tion is multivariate normal (kurtosis = 80), but once kurtosis
does not fit the data. Because the true one-factor model reaches 150 all type I errors are above 0.05. As sample size
was fit to the simulated data, one would expect the empir- increases, the problem worsens. In comparison to the results
ical type I error rate to be close to the nominal level 0.05. shown in Table 5, type I errors are worse with an increase
Deviation from it indicates the influence of skewness and in the number of manifest variables. However, holding the
kurtosis. The empirical type I error rates at different levels number of manifest variables constant, an increase in the
of Mardia’s kurtosis are summarized in Table 5. number of factors lowers type I error rate.
The results show that when the data are from a multivari- In summary, if either univariate or multivariate non-
ate normal distribution (kurtosis = 24), the empirical type I normal data are analyzed using normal-distribution-based
error rates were close to the nominal level 0.05. However, methods, it will lead to incorrect statistical inference. Given
when the data deviate from a multivariate normal distribu- the prevalence of nonnormality as we have shown in the
tion to a Mardia’s kurtosis of 60, the empirical type I error previous section, it is very important to quantify the nonnor-
rates are all greater than 0.05. Unsurprisingly, the problem mality. We suggest using skewness and kurtosis to measure
becomes worse with an increase in sample size. For exam- nonnormality and we will show how to obtain both uni-
ple, when the multivariate kurtosis is 100 and the sample variate and multivariate skewness and kurtosis in the next
size is 1489, the normal-distribution-based chi-squared test section.
rejects the correct one-factor model 29.8 % of the time.
Type 1 error rates were also obtained in a one-factor
model with eight manifest variables and a two-factor model Computing univariate and multivariate skewness
with four manifest variables each to investigate the effects and kurtosis
of an increase in the number of manifest variables or num-
ber of factors. Factor loadings were adjusted to maintain In this section, we illustrate how to compute univariate
uniqueness factor variance at 0.36 and total variance at 1. and multivariate skewness and kurtosis in popular statistical

Table 6 Empirical Type I error rates of the χ 2 test for factor analysis with 8 manifest variables

Sample Size

# of Factors Kurtosis Centered Kurtosis 106 554 1489

1 75 −5 0.0654 0.0695 0.0688

80 0 0.0533 0.0528 0.0502
90 10 0.0546 0.0591 0.0574
150 70 0.191 0.2449 0.2603
250 170 0.4159 0.5847 0.6373
2 75 −5 0.0861 0.0675 0.0609
80 0 0.0729 0.0549 0.0522
90 10 0.0781 0.061 0.0597
150 70 0.1664 0.1695 0.1652
250 170 0.3134 0.3746 0.4126

Bolded entries are those outside of the range [0.025,0.075] and are therefore considered different from the nominal 0.05
1726 Behav Res (2017) 49:1716–1735

software including SAS, SPSS, and R as well as a newly 1999-2000, in 3rd grade in the spring of 2002, in 5th grade
developed Web application. As previously mentioned, dif- in the spring of 2004, and in 8th grade in the spring of 2007.
ferent softwares produce different types of univariate skew- The data used here consist of four consecutive mathemati-
ness and kurtosis. Furthermore, most don’t report tests or cal ability measures of 563 children from kindergarten to 1st
multivariate measures. Using our software and macros for grade. To simplify our discussion, we assume that all files
SAS, SPSS, and R produces consistent and full results to be used are in the folder of “C:\nonnormal”, which needs
across software. Some software requires macros that can to be changed accordingly.
be downloaded from our website at http://w.psychstat.org/
nonnormal. Our Web application can be found at http://w. SAS
psychstat.org/kurtosis. All tools provided perform listwise
deletion before assessing nonnormaltiy. As an example, we To use SAS for computing the univariate and multivariate
use a subset of data from the Early Childhood Longitudinal skewness and kurtosis, first download the mardia.sas macro
Study, Kindergarten Class of 1998-99 (ECLS-K) to show file from our website. Our macro was modified from a SAS
the use of different software. The ECLS-K is a longitudi- macro MULTNORM provided by the SAS company. After
nal study with data collected in kindergarten in the fall and saving the sas macro file, the following code can be used to
spring of 1998-99, in 1st grade in the fall and spring of get the skewness and kurtosis for the ECLS-K data.8

In the SAS input, Line 1 through Line 4 read the ECLS- a standard error 0.21 (Line 8), indicating the kurtosis is
K data in the file “eclsk563.txt” into SAS. Line 5 significantly larger than 0.
includes the SAS macro file downloaded from our web- The second part of the output, from Line 10 to Line
site for use within SAS. The sixth line uses the function 23 includes the information on multivariate skewness and
mardia in the macro to calculate skewness and kurtosis. kurtosis. First, the multivariate skewness is 2.26 (Line 16)
The argument “data=” specifies the SAS database to use with a standardized measure of 212.24 (Line 17). The p-
and “var=” specifies the variables to use in calculating the value for a chi-squared test is approximately 0 (Line 18).
skewness and kurtosis. Therefore, the multivariate skewness is significantly larger
The SAS output from the analysis of the ECLS-K data is than 0. Second, the multivariate kurtosis is 25.47 (Line 21)
given below. The first part of the output, from Line 1 to Line with the standardized measure of 2.51 (Line 22). The p-
8, displays the univariate skewness and kurtosis as well as value for a z-test is approximately 0.01 (Line 23). Therefore,
their corresponding standard error. For example, the skew- the multivariate kurtosis is significantly different from that
ness for the ECLS-K data at time 1 is 0.69 with a standard
error 0.10 (Line 5). Based on a z-test, one would conclude
that the skewness is significantly larger than 0. For another 8 The number on the right is used to identify the code only and is not
example, the kurtosis for the data at time 4 is 1.29 with part of the SAS code.
Behav Res (2017) 49:1716–1735 1727

of a multivariate normal distribution with 4 variables (24). distribution and therefore violate the normality assumption
Consequently, the data do not follow a multivariate normal if used in multivariate analysis.
1728 Behav Res (2017) 49:1716–1735

SPSS The code on the first eight lines in the input is used to read
the ECLS-K data into SPSS. These lines are not necessary
DeCarlo (1997b) has developed an SPSS macro to cal- if your data are already imported into SPSS. Line 10 gets
culate multivariate skewness and kurtosis.9 We slightly the SPSS macro into SPSS for use. The function mardia
modified the macro to make the output of univariate skew- calculates univariate and multivariate skewness and kurtosis
ness and kurtosis consistent to other software. To use the for the variables specified by the vars option on Line 11.
SPSS macro, first download the macro file mardia.sps to Note that the folder to the data file and the SPSS macro
your computer from our website. Then, open a script editor file needs to be modified to reflect the actual location of
(File->New->Syntax) within SPSS and include the follow- them.
ing SPSS script.

The SPSS output from the analysis of the ECLS-K data the output includes univariate skewness and kurtosis and
is given below. Similar to the SAS output, the first part of the second part is for the multivariate skewness and kurto-
sis. SPSS obtained the same skewness and kurtosis as SAS
9 The original macro can be downloaded at http://www.columbia.edu/ because the same definition for skewness and kurtosis was
ld208/Mardia.sps. used.
Behav Res (2017) 49:1716–1735 1729

R ECLS-K data into R and Line 2 provides names for

the variables in the data. The third line loads the R
To use R, first download the R code file mardia.r to function mardia into R. Finally, the last line uses
your computer from our website. Then, in the editor of the function mardia to carry out the analysis on
R, type the following code. The code on Line 1 gets the Line 4.
1730 Behav Res (2017) 49:1716–1735

The output from the R analysis is presented below. skewness and kurtosis as SAS and SPSS.
Clearly, it obtains the same univariate and multivariate
Behav Res (2017) 49:1716–1735 1731

Fig. 4 Interface of the Web application

Web Application for Skewness and Kurtosis

.sas7bdat, and an Excel data file with the extension
To further ease the calculation of univariate and multivariate name .xls or .xlsx. In addition, a CSV file (comma sep-
skewness and kurtosis, we also developed a Web application arated value data file) with the extension name .csv and a
that can work within a Web browser and does not require TXT file (text file) with the extension name .txt can also
knowledge of any specific software. The Web application be used. If a .csv or .txt file is used, the user needs to
utilizes the R function discussed in the previous section to specify whether variable names are included in the file. For
obtain skewness and kurtosis on a Web server and produces Excel data, it requires the first row of the data file to be the
the same results as SAS, SPSS, and R. variable names.
To access the Web application, type the URL http://
psychstat.org/kurtosis in a Web browser and a user will see Select variables to be used
an interface as shown in Fig. 4. To use the Web applica-
tion, the following information needs to be provided on the Skewness and kurtosis can be calculated on either all the
interface. variables or a subset of variables in the data. To use all the
variables, leave this field blank. To select a subset of vari-
Data ables, provide the column numbers separated by comma
“,”. Consecutive variables can be specified using “-”. For
The data file can be chosen by clicking the “Choose File” example, 1, 2-5, 7-9, 11 will select variables 1, 2, 3,
button10 and locating the data set of interest on the local 4, 5, 7, 8, 9, 11.
computer.
Missing data
Type of data
Missing data are allowed in the data although they will be
The Web application allows commonly used data types such removed before the calculation of skewness and kurtosis.
as SPSS, SAS, Excel, and text data. To distinguish the This field should be left blank if the data file has no missing
data used, it recognizes the extension names of the data values. If multiple values are used to denote missing data,
file. For example, a SPSS data file ends with the extension they can be specified all together separated by a comma
name .sav, a SAS data file with the extension name (,). For example, -999, -888, NA will specify all three
values as missing data.
10 Note that different operating systems and/or browsers might show
the button differently. For example, for Internet Explorer, the button After providing the required information, clicking the
reads “Browse...”. “Calculate” button will start the calculation of skewness and
1732 Behav Res (2017) 49:1716–1735

kurtosis. The output of the analysis is provided below. The those of Micceri (1989) and Blanca et al. (2013) strongly
output is identical to the R output except for the variable suggest the prevalence of nonnormality in real data.
names for univariate skewness and kurtosis. This is because Our investigation on the influence of skewness and kurto-
by default the variable names are constructed using “V” and sis involved simulation studies on the one-sample t-test and
an integer in R. factor analysis. Through simulation, we concretely showed
that nonnormality, as measured by skewness and kurto-
sis, exerted great influence on statistical tests that bear the
Discussion and recommendations normality assumption. For example, the use of the t-test
incorrectly rejected a null hypothesis 17 % of the time
The primary goals of this study were to assess the preva- and the chi-squared test incorrectly rejected a correct factor
lence of nonnormality in recent psychology and education model 30 % of the time under some conditions. Therefore,
publications and its influence on statistical inference, as nonnormality can cause severe problems. For example, a
well as to provide a software tutorial on how to compute significant result might be simply an artificial effect caused
univariate and multivariate skewness and kurtosis. First, by nonnormality.
nonnormality clearly exists in real data. Based on the test Given the prevalence of nonnormality and its influence
of skewness and kurtosis of data from 1,567 univariate vari- on statistical inference, it is critical to report statistics such
ables, much more than tested in previous reviews, we found as skewness and kurtosis to understand the violation of nor-
that 74 % of either skewness or kurtosis were significantly mality. Thus, we highly recommend that journal editors
different from that of a normal distribution. Furthermore, and reviewers encourage authors to report skewness and
68 % of 254 multivariate data sets had significant Mardia’s kurtosis in their papers. In Table 7, we list the summary
multivariate skewness or kurtosis. Our results together with statistics that are critical to different statistical methods in
Behav Res (2017) 49:1716–1735 1733

empirical data analysis. For example, mean comparisons

means, variances-covariances, univariate & multivariate skewness & kurtosis

means, variances-covariances, univariate & multivariate skewness & kurtosis
would be influenced by skewness while factor analysis is
more influenced by kurtosis. To facilitate the report of
univariate and multivariate skewness and kurtosis, we have

mean, variance-covariances, univariate & multivariate kurtosis

provided SAS, SPSS, and R code as well as a Web applica-
tion to compute them.

variances-covariances, univariate & multivariate kurtosis

variances-covariances, univariate & multivariate kurtosis
Once nonnormality has been identified as a problem,
means, variances, univariate & multivariate skewness
the main options for handling it in a statistical analysis
include transformation, nonparametric methods, and robust
analysis. Transforming data so that they are closer to nor-
mally distributed is a relatively easy option, because after
means, variances, univariate skewness

transformation the researcher can proceed with whichever

normality-based method they desire. In psychology, log
transformation is a common way to get rid of positive skew-
ness, for example. More generally, the class of Box-Cox
transformations (Box and Cox, 1964) is also very pop-
Recommendations

ular because it’s easy to use and can accomodate many

types of nonnormality. However, it has been suggested that
Box-Cox and other transformations seldom maintain linear-
ity, normality, and homoscedasticity simultaneously (Sakia,
1992, for example), and even if transformation is success-
ful the resulting parameter estimates might be difficult to
interpret.
Corder and Foreman (2014) offer an easy-to-follow
review of nonparametric techniques, including the
Mann–Whitney U-test, Kruskal-Wallis test, and Spearman
rank order correlation, among others. The basic premise
of most of these methods is to perform analysis on ranks
rather than the raw data. This is, of course, a more robust
procedure than assuming normality of raw data, but can be
less powerful in some circumstances and the results can be
less meaningful. However, for data that is already ordinal
Table 7 Critical summary statistics for different methods in empirical research

or ranked these methods can be very useful, and can still be

advantageous in other circumstances, as well.
Robust analysis can work better than transformation and
non-parametric methods in many situations, though histori-
cally it has also been the most difficult to conduct. Robust
analysis can address three points of concern: parameter
estimates, standard errors of those estimates, and test statis-
tics. Within the context of SEM the methods that perform
best in dealing with each of these issues, respectively, are
robust estimation using Huber-type weights (Huber, 1967),
Factor and principal component analysis
Repeated measures ANOVA, MANOVA

sandwich-type standard errors, and a rescaled chi-squared

statistic following robust estimation (Yuan & Bentler,
1998). Alternatively, one can also obtain sandwich-type
Mean comparison (t-test, F-test)

SEM without mean structures

standard errors and rescaled test statistics following normal

Regression and correlation

SEM with mean structures

distribution-based maximum likelihood (Satorra & Bentler,

Growth curve analysis

1988), and such a procedure will still yield consistent results

when data are of heavier tails but do not contain outliers.
Recently, some software packages have begun to include
these procedures, making robust analysis a much easier
Methods

option than it has ever been before. Currently, EQS (Mul-

tivariate Software, Inc.), WebSEM (Zhang & Yuan, 2012),
1734 Behav Res (2017) 49:1716–1735

and the R package rsem (Yuan & Zhang, 2012) are the Methods for Research in the Natural Sciences Statistics in
only softwares to offer truly robust methods that address Biology: Statistical Methods for Research in the Natural Sci-
ences (VOL 1, pp. 140–146). New York McGraw-Hill Book
all three points of concern (parameter estimates, standard
Company.
errors, and test statistics), and WebSEM and rsem offer them Box, G.E., & Cox, D.R. (1964). An analysis of transformations An
for free. Additionally, WebSEM has a user-friendly inter- analysis of transformations. Journal of the Royal Statistical Soci-
face in which researchers can draw the path diagram they ety. Series B (Methodological), 211–252. Retrieved 2016-07-26,
from. http://www.jstor.org/stable/2984418
wish to fit.
Corder, G.W., & Foreman, D.I. (2014). Nonparametric statistics: A
As shown in Fig. 3, there is no common distribution step-by-step approach: Wiley.
of practical data in psychology and education. With such D’Agostino, R.B. (1970). Transformation to Normality of the Null
diversity in data shapes and research goals, it is impossi- Distribution of g 1 Transformation to Normality of the Null
Distribution of g 1. Biometrika, 57(3), 679. doi:10.2307/2334794
ble to create one universal solution. However, we hope that
DeCarlo, L. (1997a). Mardia’s multivariate skew (b1p) and multi-
through this paper we were able to elucidate the problem variate kurtosis (b2p). Retrieved from. http://www.columbia.edu/
through our review of practical data and simulation and ld208/Mardia.sps
offer some feasible recommendations to researchers in our DeCarlo, L. (1997b). On the meaning and use of kurtosis. On the
meaning and use of kurtosis. Psychological methods, 2(3), 292.
field. It is our hope that researchers begin to take nonnor- doi:10.1037/1082-989X.2.3.292
mality seriously and start to report them along with means Huber, P.J. (1967). The behavior of maximum likelihood estimates
and variances that have already been established in data under nonstandard conditions. In Proceedings of the fifth Berkeley
analysis. We believe that reporting skewness and kurtosis symposium on mathematical statistics and probability (VOL. 1, pp.
221–233.
in conjunction with moving toward robust analysis offer Joanes, D., & Gill, C. (1998). Comparing measures of sample
two high-impact changes that can be made in the litera- skewness and kurtosis Comparing measures of sample skew-
ture at this time. These actions will not only increase the ness and kurtosis. Journal of the Royal Statistical Society:
transparency of data analysis, but will also encourage quan- Series D (The Statistician), 47(1), 183–189. 10.1111/1467-9884.
00122
titative methodologists to develop better techniques to deal Komsta, L., & Novomestky, F. (2015). Moments: Moments, cumu-
with nonnormality, improve statistical practices and con- lants, skewness, kurtosis and related tests moments: Moments,
clusions in empirical analysis, and increase awareness and cumulants, skewness, kurtosis and related tests. Retrieved from
knowledge of the nonormality problem for all researchers in https://CRAN.R-project.org/package=moments (R package ver-
sion 0.14.
our field. Levine, D.W., & Dunlap, W.P. (1982). Power of the F test with skewed
data: Should one transform or not? Power of the f test with skewed
data: Should one transform or not?. Psychological Bulletin, 92(1),
272. doi:10.1037/0033-2909.92.1.272
Acknowledgments This research is supported by a grant from the Mardia, K.V. (1970). Measures of Multivariate Skewness and
U.S. Department of Education (R305D140037). However, the contents Kurtosis with Applications Measures of Multivariate Skew-
of the paper do not necessarily represent the policy of the Department ness and Kurtosis with Applications. Biometrika, 57(3), 519.
of Education, and you should not assume endorsement by the Federal doi:10.2307/2334770
Government. Meyer, D., Dimitriadou, E., Hornik, K., Weingessel, A., & Leisch,
Correspondence concerning this article can be addressed to Meghan F. (2015). e1071: Misc Functions of the Department of Statistics,
Cain ([email protected]), Ke-Hai Yuan ([email protected]), or Zhiyong Probability Theory Group (Formerly: E1071), TU Wien. Retrieved
Zhang ([email protected]), Department of Psychology, University of from https://CRAN.R-project.org/package=e1071 (R package ver-
Notre Dame, 118 Haggar Hall, Notre Dame, IN 46556. sion 1.6-7).
Micceri, T. (1989). The unicorn, the normal curve, and other
improbable creatures. The unicorn, the normal curve, and
other improbable creatures. Psychological Bulletin, 105(1), 156.
References doi:10.1037/0033-2909.105.1.156
Palmer, E.M., Horowitz, T.S., Torralba, A., & Wolfe, J.M. (2011).
Anscombe, F.J., & Glynn, W.J. (1983). Distribution of the Kurto- What are the shapes of response time distributions in visual
sis Statistic b2 for Normal Samples Distribution of the Kur- search?. Journal of Experimental Psychology: Human Percep-
tosis Statistic b2 for Normal Samples. Biometrika, 70(1), 227. tion and Performance, 37(1), 58–71. Retrieved 2016-07-26, from
doi:10.2307/2335960 doi:10.1037/a0020747
Becker, M., & Klößner, S. (2016). PearsonDS: Pearson Distribution R Core Team (2016). R: A Language and Environment for Statistical
System PearsonDS: Pearson Distribution System. Retrieved from Computing R: A Language and Environment for Statistical Com-
https://CRAN.R-project.org/package=PearsonDS (R package ver- puting. Vienna, Austria R Foundation for Statistical Computing.
sion 0.98. https://www.R-project.org/
Blanca, M.J., Arnau, J., López-Montiel, D., Bono, R., & Bendayan, R. Rosseel, Y. (2012). lavaan: An R Package for Structural Equation Mod-
(2013). Skewness and Kurtosis in Real Data Samples Skewness eling lavaan: An R Package for Structural Equation Modeling.
and kurtosis in real data samples. Methodology: European Journal Journal of Statistical Software, 48(2), 1–36. http://www.jstatsoft.
of Research Methods for the Behavioral and Social Sciences, 9(2), org/v48/i02/
78–84. doi:10.1027/1614-2241/a000057 Sakia, R.M. (1992). The Box-Cox Transformation Technique: A
Bliss, C. (1967). Statistical Tests of Skewness and Kurtosis Statistical Review. The Statistician, 41(2), 169. Retrieved 2016-07-26, from
Tests of Skewness and Kurtosis. In Statistics in Biology: Statistical doi:10.2307/2348250
Behav Res (2017) 49:1716–1735 1735

Satorra, A., & Bentler, P. (1988). Scaling corrections for statistics of Mathematical and Statistical Psychology, 58(2), 209–237.
in covariance structure analysis (UCLA Statistics Series 2). Los doi:10.1348/000711005X64060
Angeles: University of California at Los Angeles: Department of Yuan, K.H., & Bentler, P.M. (1998). Normal theory based
Psychology. test statistics in structural equation modelling. British Jour-
Scheffe, H. (1959). The analysis of variance. The analysis of variance. nal of Mathematical and Statistical Psychology, 51(2), 289–
New York: Wiley. 309.
Shohat, J. (1929). Inequalities for Moments of Frequency Functions Yuan, K.H., Bentler, P.M., & Zhang, W. (2005). The Effect
and for Various Statistical Constants. Biometrika, 21(1/4), 361. of Skewness and Kurtosis on Mean and Covariance Struc-
doi:10.2307/2332566 ture Analysis: The Univariate Case and Its Multivariate Impli-
Tabachnick, B.G., & Fidell, L.S. (2012). Using Multivariate Statistics cation. Sociological Methods & Research, 34(2), 240–258.
Using Multivariate Statistics (6 ed). Pearson. doi:10.1177/0049124105280200
Wang, L., Zhang, Z., McArdle, J.J., & Salthouse, T.A. (2008). Yuan, K.H., & Zhang, Z. (2012). Robust structural equation modeling
Investigating Ceiling Effects in Longitudinal Data Anal- with missing data and auxiliary variables. Psychometrika, 77(4),
ysis. Multivariate Behavioral Research, 43(3), 476–496. 803–826. doi:10.1007/s11336-012-9282-4
doi:10.1080/00273170802285941 Zhang, Z., & Yuan, K.H. (2012). WebSEM: Conducting structual
Yanagihara, H., & Yuan, K.H. (2005). Four improved statistics for equation modelling online. Notre Dame, IN. Retrieved from.
contrasting means by correcting and kurtosis. British Journal https://websem.psychstat.org

View publication stats