A peer-reviewed electronic journal.

Copyright is retained by the first or sole author, who grants right of first publication to Practical Assessment, Research & Evaluation. Permission
is granted to distribute this article for nonprofit, educational purposes if it is copied in its entirety and the journal is credited. PARE has the
right to authorize third party reproduction of this article in print, electronic and database forms.
Volume 19, Number 17, November 2014 ISSN 1531-7714

A Primer on Multivariate Analysis of Variance (MANOVA) for

Behavioral Scientists

Russell T. Warne, Utah Valley University

Reviews of statistical procedures (e.g., Bangert & Baumberger, 2005; Kieffer, Reese, & Thompson,
2001; Warne, Lazo, Ramos, & Ritter, 2012) show that one of the most common multivariate
statistical methods in psychological research is multivariate analysis of variance (MANOVA).
However, MANOVA and its associated procedures are often not properly understood, as
demonstrated by the fact that few of the MANOVAs published in the scientific literature were
accompanied by the correct post hoc procedure, descriptive discriminant analysis (DDA). The
purpose of this article is to explain the theory behind and meaning of MANOVA and DDA. I also
provide an example of a simple MANOVA with real mental health data from 4,384 adolescents to
show how to interpret MANOVA results.

One of the most common multivariate statistical in only 5 out of 62 articles that used MANOVA
procedures in the social science literature is multivariate (8.1%) in educational psychology journals included the
analysis of variance (MANOVA). In one review of correct post hoc procedure: discriminant descriptive
educational psychology journals, MANOVA was used analysis (DDA). Tonidandel and LeBreton (2013)
in 23.0% quantitative articles, making it the most found that almost every article in the Journal of Applied
common multivariate procedure (Warne, Lazo, Ramos, Psychology in which MANOVA was used did not include
& Ritter, 2012, p. 138)—results that agree with Bangert a follow-up DDA. In my scrutiny of three major
and Baumberger (2005), who found that MANOVA psychology journals—the Journal of Clinical Psychology,
was the most commonly used multivariate statistical Emotion, and the Journal of Counseling Psychology—I found
procedure in a clinical psychology journal. Other a total of 58 articles in which researchers used
reviewers of statistical procedures in psychological MANOVA between 2009 and 2013. However, in none
research (e.g., Kieffer, Reese, & Thompson, 2001; of these articles did researchers use a DDA. Huberty
Zientek, Capraro, & Capraro, 2008) have found and Morris (1989) found similar results when
MANOVA to be the most popular multivariate investigating six leading psychology and education
statistical method in the published literature. research journals. The purpose of this article is to
explain the principles of MANOVA and how to
Despite the popularity of MANOVA, the method
correctly perform and interpret a MANOVA and its
is persistently misunderstood, and researchers are
associate post hoc procedure, DDA. A simple example
particularly unfamiliar with the proper statistical
of a MANOVA and DDA using real data relevant to
procedures after rejecting a multivariate null
hypothesis. Warne et al. (2012), for example, found that
Practical Assessment, Research & Evaluation, Vol 19, No 17 Page 2
Warne, MANOVA

behavioral scientists follows the explanation to make then the null hypothesis is rejected (as would be the
the article more applicable. case for vectors A, C, and D in Figure 2). However, if a
vector does not extend into the rejection region (e.g.,
What is MANOVA? vector B in Figure 2), then the null hypothesis is
MANOVA is a member of the General Linear retained.
Model—a family of statistical procedures that are often
used to quantify the strength between variables
(Zientek & Thompson, 2009). Many of the procedures
in the General Linear Model are hierarchically
organized—i.e., more specific procedures are often
special cases of general procedures (Zientek &
Thompson, 2009). MANOVA, specifically, is an
analysis of variance (ANOVA) that has two or more
dependent variables (Fish, 1988). Because most
behavioral scientists already have a firm grasp of
ANOVA (Aiken, West, & Millsap, 2008), I will use it as
the starting point in this article for explaining Figure 1. Example of the logic of univariate
MANOVA. hypothesis testing.
In an ANOVA the independent variable is a The shaded area is the rejection region. If the
nominal variable that has two or more values, and the test statistic is in the rejection region (shown
dependent variable is intervally or ratio scaled.1 The both on the histogram and on the number line
null hypothesis is that the mean score on the dependent below), then the null hypothesis is rejected.
variable will be statistically equal for every group. As
with any null hypothesis statistical significance testing
procedure, an observed statistic is calculated and then
compared to a sampling distribution. If the observed
statistic is found to be a more extreme value in the
sampling distribution than the critical value (as shown
in Figure 1), then the null hypothesis will be rejected;
otherwise, the null hypothesis is retained. Afterwards,
an effect size—usually η2 in an ANOVA—quantifies
the relationship between the independent and
dependent variables (Thompson, 2006).
MANOVA is merely an ANOVA that has been
mathematically extended to apply to situations where
there are two or more dependent variables (Stevens,
2002). This necessitates several changes to the logic of
ANOVA. The first is displayed in Figure 2. In contrast
to Figure 1, which is a one dimensional number line, Figure 2. Example of the logic of multivariate
Figure 2 shows the rejection region of a MANOVA as analysis of variance hypothesis testing for
a region outside of a circle on a two-dimensional perfectly uncorrelated variables.
Cartesian plane. Therefore, instead of the observed The shaded area is the rejection region. If the
statistic being expressed as a point on a number line, it vector for the test ends in the rejection region,
can be graphically expressed as a vector (Thomas, then the null hypothesis is rejected. If the dotted
1992). If the vector extends into the rejection region, lines originating at the end of a vector meet an
axis in the shaded region, then the ANOVA for
1 Please note that a one-way ANOVA simplifies to a t-test if the
that axis’s dependent variable would also be
independent nominal variable has only two groups (Thompson, rejected.
2006).
Practical Assessment, Research & Evaluation, Vol 19, No 17 Page 3
Warne, MANOVA

same two dependent variables would be only .05. The

severity of Type I error inflation in multiple ANOVAs
Figure 2 also shows how the results of a
depends on how correlated the dependent variables are
MANOVA can vary from the results of two ANOVAs.
with one another, with Type I error inflation being
Each vector in the figure can be decomposed into two
most severe when dependent variables are uncorrelated
pieces (which are represented in the figure by the
(Hummel & Sligo, 1971). Readers will recognize the
dotted lines), one for each dependent variable (Saville
Type I error inflation that occurs with multiple
& Wood, 1986). Notice how vector A extends beyond
ANOVAs because experiment-wise Type I error
the point on both axes where the unshaded region
inflation also occurs when conducting multiple
ends. This means that if an ANOVA were conducted
independent sample t-tests. In fact, one of the main
on either dependent variable (which visually would
rationales for conducting ANOVA is to avoid
collapse the two-dimensional plane into a number line
conducting multiple t-tests, which inflates the Type I
resembling the lower portion of Figure 1), the null
error that occurs when researchers conduct many t-
hypothesis of the ANOVA would be rejected.
tests (Thompson, 2006).2
Likewise, analyzing vector B would produce a non-
statistically significant MANOVA and two ANOVAs Figure 2 represents an ANOVA with two perfectly
that were both also non-statistically significant. On the uncorrelated dependent variables. However, if
other hand, vector C would produce a statistically dependent variables are correlated, then the MANOVA
significant MANOVA and one statistically significant is better represented in Figure 3. Notice how the axes
ANOVA (for the vertical axis). Finally, vector D would are no longer perpendicular with each other. Rather,
produce a statistically significant MANOVA, but in the correlation between the dependent variables is
both ANOVAs the null hypotheses would be retained. represented by the angle of the axes, with a more
oblique angle indicating a higher correlation between
Why Use MANOVA? dependent variables (Saville & Wood, 1986). Although
Given MANOVA’s more complicated nature, the axes in Figures 3 look strange, it is important for
some researchers may question whether it is worth the readers to remember that in statistics all axes and scales
added complexity. The alternative to using MANOVA are arbitrary. Therefore, having nonperpendicular axes
is to conduct an ANOVA for each dependent variable. is acceptable. In exploratory factor analysis, for
However, this approach is not advantageous because example, factor solutions are almost always rotated so
(a) conducting multiple ANOVAs increases the that the pattern of factor loadings is more interpretable
likelihood of committing a Type I error, and (b) (Costello & Osborne, 2005). Indeed, in oblique
multiple ANOVAs cannot determine whether rotation methods, such as promax rotation, the axes are
independent variable(s) are related to combinations of not only rotated, but can be nonperpendicular—just
dependent variables, which is often more useful like the axes in Figure 3 (Thompson, 2004).
information for behavioral scientists who study Although the examples in Figures 2 and 3
correlated dependent variables. represent MANOVAs with two dependent variables,
Avoiding Type I Error Inflation an extension to three dependent variables can be
imagined without difficulty. For three perfectly
When multiple dependent variables are present, uncorrelated dependent variables, the graph would be
conducting a MANOVA is beneficial because it three dimensional, with the unshaded region being a
reduces the likelihood of Type I error (Fish, 1988;
Haase & Ellis, 1987; Huberty & Morris, 1989). The 2
Some researchers who choose to conduct multiple t-tests instead
probability of Type I error at least once in the series of of an ANOVA control Type I error inflation with a Bonferroni
ANOVAs (called experiment-wise error) can be as high correction (Thompson, 2006). The analogous relationship between
as 1 - (1 - .05)k, where k is the number of ANOVAs t-tests and ANOVA and between ANOVA and MANOVA
extend even to this point, and applying a Bonferroni correction
conducted. Therefore, if a researcher chooses the would also control Type I error inflation in a group of ANOVAs.
traditional α value of .05 for two ANOVAs, then the However, Bonferroni methods are overly conservative, especially
experiment-wise Type I error can be as high as .0975— with a large number of tests (Stevens, 2002) or correlated variables
not .05—even though the α for each ANOVA is .05. (Hummel & Sligo, 1971). Moreover, a Bonferroni correction to a
However, the Type I error for a MANOVA on the series of ANOVAs would not provide the benefits of MANOVA
described elsewhere in the article.
Practical Assessment, Research & Evaluation, Vol 19, No 17 Page 4
Warne, MANOVA

perfect sphere. As the variables became more single observed variable—an unrealistic assumption for
correlated, the axes would become more oblique, and many constructs in the behavioral sciences. Therefore,
the sphere would become more distorted and resemble MANOVA is a statistical procedure that is more in
something like a football. If two of the three dependent accordance than ANOVA with behavioral scientists’
variables were perfectly correlated (i.e., r = + 1), then beliefs about the topics they study.
the graph would collapse into a two-dimensional graph
like in Figure 3. A MANOVA with four or more Post Hoc Procedures
dependent variables requires more than three Post hoc procedures are often necessary after the
dimensions and axes—something that most people null hypothesis is rejected in an ANOVA (Thompson,
have difficulty imagining visually, but which a 2006) or a MANOVA (Stevens, 2002). This is because
computer can easily handle mathematically. the null hypotheses for these procedures often do not
provide researchers with all the information that they
desire (Huberty & Morris, 1989; Thomas, 1992). For
example, if the null hypothesis is rejected in an
ANOVA with three or more groups, then the
researcher knows that at least one group mean
statistically differs from at least one other group mean.
However, most researchers will be interested in
learning which mean(s) differ from which other group
mean(s). Many ANOVA post hoc tests have been
invented to provide this information, although Tukey’s
test is by far the most common test reported in the
psychological literature (Warne et al., 2012).
Just like ANOVA, MANOVA has post hoc
procedures to determine why the null hypothesis was
Figure 3. Example of the logic of multivariate rejected. For MANOVA this is usually a DDA, which
analysis of variance hypothesis testing for is a statistical procedure which creates a set of perfectly
moderately correlated variables. uncorrelated linear equations that together model the
The shaded area is the rejection region. differences among groups in the MANOVA (Fish,
1988; Stevens, 2002). The benefit of having
Examining Combinations of Variables uncorrelated equations from a DDA is that each
function will provide unique information about the
ANOVA and MANOVA also differ in that they differences among groups, and the information can be
investigate completely different empirical questions. combined in an additive way for easy interpretation.
Zientek and Thompson (2009, p. 345) explained that, “. Unfortunately, most researchers who use
. . two ANOVAs actually test the differences of the means on MANOVA do not use DDA to interpret their
the observed or measured variables . . . whereas the MANOVA MANOVA results (Huberty & Morris, 1989; Kieffer et
actually tests the differences of the mean of the DDA function al., 2001; Warne et al., 2012). This may be because
scores” of the groups. In other words, MANOVA tests when researchers use SPSS to conduct a MANOVA,
the differences between underlying unobserved latent the computer program automatically conducts an
variables (derived from the variables in the dataset), ANOVA for each dependent variable. However, SPSS
while ANOVA only tests differences among groups on does not automatically conduct a DDA when the null
an observed variable (Haase & Ellis, 1987; Huberty & hypothesis for a MANOVA has been rejected, even
Morris, 1989; Zientek & Thompson, 2009). MANOVA though DDA is a more correct post hoc procedure.
is therefore often more useful to social scientists than
ANOVA because most topics they research are latent Zientek and Thompson (2009) explained that
constructs that are not directly observable, such as because MANOVA analyzes latent variables composed
beliefs and attitudes. With ANOVA it is assumed that of observed variables and ANOVA is only concerned
these constructs are measured without error and with a with observed variables, using ANOVA as a post hoc
Practical Assessment, Research & Evaluation, Vol 19, No 17 Page 5
Warne, MANOVA

procedure following an ANOVA is non-sensical (see Watson, & Clark, 1998), it is theoretically sound to
also Kieffer et al., 2001). This is because ANOVA and conduct a MANOVA to investigate the relationship
MANOVA were developed to answer completely between grade and both outcomes at the same time.
different empirical questions (Huberty & Morris, 1989). The preference of a MANOVA over a univariate test is
Indeed, Tonidandel and LeBreton (2013) explained further strengthened by the moderately strong
that, . . . multivariate theories yield multivariate hypotheses correlation between depression and anxiety (r = .544, p
which necessitate the use of multivariate statistics and < .001) in the dataset. The null hypothesis for this
multivariate interpretations of those statistics. By invoking MANOVA is that the six independent variable groups
univariate ANOVAs as follow-up tests to a significant are equal on both dependent variables. For a
MANOVA, researchers are essentially ignoring the multivariate MANOVA with additional groups or more than two
nature of their theory and data. (p. 475) dependent variables, the null hypothesis would be that
all groups are equal on all dependent variables (Stevens,
Moreover, Fish (1988) and Huberty and Morris
2002).
(1989) showed that the same data can produce different
results when analyzed using a MANOVA or a series of Table 1 shows descriptive statistics for the
ANOVAs—a situation that could potentially be variables in the example data from the Add Health
misleading to researchers who follow MANOVA with study. A MANOVA was conducted to determine the
a series of post hoc ANOVAs. relationship between grade and the combined variables
of depression and anxiety. The results from SPSS 19.0
The purpose of the remainder of this article is to
are shown in Table 2, which I modified slightly to make
provide a real-life example of a MANOVA with a
it easier to read.
single nominal independent variable and two
dependent variables and a post hoc DDA as a simple Table 1. Descriptive Statistics of Variables Used in
blueprint for future researchers who wish to use this the Example MANOVA
procedure. By doing so, I hope to make conducting Depression Anxiety
Grade n
and interpreting MANOVA and DDA as easy as M SD M SD
possible for readers. 7 662 0.88 1.114 0.75 1.048
8 664 1.08 1.188 0.80 1.058
Conducting a MANOVA 9 778 1.17 1.188 0.93 1.080
10 817 1.27 1.232 0.96 1.109
The example data in this study is taken from the 11 790 1.37 1.204 1.12 1.156
National Longitudinal Study of Adolescent Health 12 673 1.34 1.140 1.10 1.105
(Add Health), a longitudinal study of adolescent and Note. Listwise deletion in all analyses, so all results are based
on the 4,384 cases that had data on all three variables.
adult development that started in the 1994-1995 school
year when its participants were in grades 7-12. Data for Readers will notice a few things about Table 2.
this example were taken from a public use file of First, there are two sections: intercept (containing four
variables collected during the initial data collection rows) and grade (containing another four rows). The
period and downloaded from the Inter-University intercept section of the table is necessary to scale the
Consortium for Political and Social Research web site. results and does not provide any substantive
For this example the only independent variable is information for interpretation. The grade section,
grade, and the two dependent variables are participants’ however, displays results from the hypothesis test.
responses to the questions (a) “In the last month, how Second, there are four rows, each of which display four
often did you feel depressed or blue?” and (b) “In the statistical test statistics: (a) Pillai’s Trace, (b) Wilks’
last month, how often did you have trouble relaxing?” Lambda, (c) Hotelling’s Trace, and (d) Roy’s Largest
Both responses were recorded on a 5-point Likert scale Root. All four of these are test statistics for the same
(0 = Never, 1 = Rarely, 2 = Occasionally, 3 = Often, null hypothesis, although their formulas differ (Olson,
and 4 = Every day). For convenience these variables 1976). All four can be converted to an F-statistic, which
will be called “depression” and “anxiety,” respectively. can then be used to calculate a p-value, which are
As previous researchers have found that anxiety and displayed in Table 2.
depressive disorders are often comorbid (e.g., Mineka,
Practical Assessment, Research & Evaluation, Vol 19, No 17 Page 6
Warne, MANOVA

Table 2. Example MANOVA Results (Modified from SPSS Output)

Partial Noncentrality
Effect Test Statistic Value F df Sig. (p)
η2 Parameter
Intercept Pillai’s Trace 0.533 2470.323 2, 4337 < .001 .533 4940.647
Wilks’ Lambda 0.467 2470.323 2, 4337 < .001 .533 4940.647
Hotelling’s Trace 1.139 2470.323 2, 4337 < .001 .533 4940.647
Roy’s Largest Root 1.139 2470.323 2, 4337 < .001 .533 4940.647
Grade Pillai’s Trace 0.022 9.834 10, 8676 < .001 .011 98.340
Wilks’ Lambda 0.978 9.873 10, 8674 < .001 .011 98.726
Hotelling’s Trace 0.023 9.911 10, 8672 < .001 .011 99.112
Roy’s Largest Root 0.021 18.388 5, 4338 < .001 .021 91.939

All four of the MANOVA test statistics for the Table 2 contains additional information of note.
Add Health data are statistically significant (p < .001). First, the effect sizes (i.e., partial η2 values) are all either
This indicates that the null hypothesis is rejected, and .011 and .021—indicating that grade does not have a
grade has a statistically significant relationship with the particularly powerful statistical relationship with the
combined dependent variables of depression and dependent variables. Second, the noncentrality
anxiety. In many cases the results from the four parameter for each test statistic is displayed in Table 2.
statistical tests will be the same (O’Brien & Kaiser, This noncentrality parameter is a statistical parameter
1985). In fact, these statistics will always produce the necessary to estimate a noncentral distribution that
same result when there are only two groups in the models the distribution of test statistics if the null
independent variable (Haase & Ellis, 1987; Huberty & hypothesis is not true (Steiger & Fouladi, 1997). When
Olejnik, 2006) or when there is a single dependent the noncentrality parameter is equal to zero, the null
variable (because this simplifies the test to an ANOVA; hypothesis perfectly fits the data and therefore should
see Olson, 1976). not be rejected. Therefore, it makes sense that all four
test statistics have high noncentrality parameter values
However, it is possible that the four statistical tests
(between 91.939 and 99.112) in Table 2 because all test
will produce contradictory results, with some statistics
lacking statistical significance and others being statistics suggest that the null hypothesis should be
statistically significant (Haase & Ellis, 1987). This is rejected, and each one has a very small p-value (p <
because the four tests differ in their statistical power, .001). Finally, the column labeled “Observed Power”
with Roy’s great root having the most power when gives the a posteriori statistical power of the
MANOVA design for each test statistic. The numbers
dependent variables are highly correlated—possibly
in this column will always vary inversely with the
because they all represent a single construct—and the
corresponding p-value, so they provide no information
other three having the more power for more disparate
beyond that provided by the p-values in the table
outcomes (Huberty & Oljenik, 2006). Additionally,
(Hoenig & Heisey, 2001).
Pillai’s trace is highly robust to many violations of the
assumptions of MANOVA (Olson, 1976). Regardless Post Hoc DDA
of the test statistic(s) that a researcher uses, he or she
should explicitly report which test statistic(s) he or she With the results of the MANOVA in Table 2
has chosen and not merely state that the results are showing that the null hypothesis should be rejected, it
statistically significant (Haase & Ellis, 1987; Olson, is necessary to conduct a post hoc DDA in order to
1976). Readers can see that in this example all four test determine the mathematical function(s) that distinguish
statistics provided evidence that the null hypothesis the grade groups from one another on dependent
should be rejected, even though the values of the variable scores. The number of functions that are
statistics and their degrees of freedom vary because of created will vary, but the minimum will always be the
their varying formulas and theoretical distributions number of dependent variables or the number of
(Haase & Ellis, 1987). groups minus 1—whichever value is smaller (Haase &
Practical Assessment, Research & Evaluation, Vol 19, No 17 Page 7
Warne, MANOVA

Ellis, 1987; Stevens, 2002; Thomas, 1992). Therefore,

Table 3. Example Standardized Discriminant
in the Add Health example there will be two
Coefficients and Structure Coefficients
discriminant functions because there are two
(Modified from SPSS Output)
dependent variables.
In SPSS, a descriptive discriminant analysis can be Standardized Discriminant Coefficientsa
conducted through the “discriminant” tool. The Function 1 Function 2
computer program produces a table labeled “Wilks’ Depression .700 -.958
Lambda,” and displays the results of the significance Anxiety .431 1.105
test for the two discriminant functions. In this example Structure Coefficientsb
the first discriminant function is statistically significant Function 1 Function 2
(p < .001), while the second is not (p = .125), despite Depression .932 -.364
the large sample size (n = 4,384). Therefore, it is not Anxiety .808 .590
necessary to interpret the second discriminant function.
Parallel Discriminant Ratio Coefficients
SPSS displays a number of other tables that aid in Function 1 Function 2
the interpretation of discriminant functions, the two Depression (.700)(.932) = (-.958)(-.364) =
most important of which show the standardized .652 .349
discriminant coefficients (SDCs) and the structure Anxiety (.431)(.808) = (1.105)(.590) =
coefficients for the discriminant functions. These tables .348 .652
are displayed in Table 3. The SDCs can be used to a
SPSS displays this information in a table labeled
calculate a score for each subject for each discriminant
“Standardized Canonical Discriminant Function
function. Given the numbers in Table 3, the function Coefficients.”
score for the first function can be calculated as: b
SPSS displays this information in a table labeled
DDA Score1 = .700Depression + .431Anxiety “Structure Matrix.” Note. Parallel DRCs should always add
to 1.0 (Thomas & Zumbo, 1996), but do not in Function
These SDC values are interpreted the same way as 2 because of rounding.
the standardized regression coefficients in a multiple
regression (Stevens, 2002). Therefore, the .700 in the 1992; Thomas & Zumbo, 1996), which requires the
equation means that for every 1 standard deviation structure coefficients for the variables. The second
increase in examinees’ depression scores their DDA section of Table 3 shows the SPSS output of the
score is predicted to increase by .700 standard structure coefficients, which are the correlations
deviations if all other variables are held constant. More between each observed variable and the DDA scores
substantively, because discriminant functions maximize calculated from each discriminant function. This is
the differences among the grade groups, it can be seen analogous to structure coefficients in multiple
that although both depression and anxiety contribute to regression, which are correlations between independent
group differences, depression shows more between- variables and predicted values for the dependent
group variation. This is supported by the data in Table variable (Thompson, 2006). To calculate a parallel
1, which shows that from Grades 7 to 11 depression DRC, one must multiply each SDC with the
and anxiety scores increase steadily, although the corresponding structure coefficient (Thomas, 1992), as
changes in depression scores are more dramatic. These is demonstrated at the bottom of Table 3. For the first
results also support studies showing that both anxiety function the parallel DRCs are .652 for depression and
disorders and depression often begin in adolescence .348 for anxiety—indicating that depression is clearly
(e.g., Cole, Peeke, Martin, Truglio, & Seroczynski, the more important variable of the two in
1998). distinguishing between independent variable groups in
this example.
However, it is important to recognize that the
SDCs cannot indicate which variable is more Although parallel DRCs are highly useful, their
“important” for distinguishing groups. The easiest way interpretation may be complicated by the presence of a
one can determine variable importance is by calculating suppressor variable, which is a variable that has little
a parallel discriminant ratio coefficient (DRC; Thomas, correlation with a dependent variable but which can
strengthen the relationship that other variables have
Practical Assessment, Research & Evaluation, Vol 19, No 117 Page 8
Warne, MANOVA

with a dependent variable and thereby improve model among average members of their DDA groups, which
fit if included (Smith, Ager, & Williams, 1992; ranged in number from four to eight.
Thompson, 2006; Warne 2011). Because a suppressor
This article on MANOVA and DDA was designed
variable and a useless variable willl both have small to help social scientists better
ter use and understand one
parallel DRCs, it is often necessary to calculate a total of the most common multivariate statistical methods in
DRC to determine if a suppressor variable is present
the published literature. I hope that researchers will use
using the following equation, adapted from Equation
the information in this article to (a) understand the
21 of Thomas and Zumbo (1996):
nature of MANOVA and DDA, (b) serve as a guide to
correctly performing
erforming their own MANOVA and post
hoc DDA analyses, and (c) aid in the interpretation of
their own data and of other studies that use these
methods. Researchers in the behavioral sciences should
use this powerful multivariate analysis method more
In this equation, dfBetween and dfWithin represent the often inn order to take advantage of its benefits. I also
degrees of freedom for the model’s between between- and encourage researchers to investigate other post hoc
within-group
group components, which are the same degrees procedures, such as dominance analysis (Azen &
of freedom as in an ANOVA for the same independent Budescu, 2003, 2006), which has advantages over
variable on a single dependent variable (Olson, 1976). F DDA. However, other methods are more complex and
represents the same F-ratio for that variable, and λ is difficult to perform than a DDA, and many researchers
the discriminant function’s eigenvalue. All information may not find the advantages worth the extra time and
needed to calculate a total DRC for each independent expertise needed to use them. Therefore, I believe that
variable is available to a researcher in the SPSS output DDA is often an expedient choice for researchers who
for a MANOVA and DDA. Calculating total DRCs is wish to understand their multivariate analyses.
not necessary ry for this example because both the Regardless
gardless of the choice of post hoc method,
depression and anxiety variables have high enough researchers should completely avoid using univariate
parallel DRCs that both are clearly useful variables in methods after rejecting a multivariate null hypothesis
the discriminant function, and neither is a plausible (Huberty & Morris, 1989; Kieffer et al., 2001;
suppressor variable. Tonidandel & LeBreton, 2013; Zientek & Thompson,
2009).
Conclusion
References
ferences
Although this MANOVA and itss accompanying
DDAs were simple, analyses with additional Aiken, L. S., West, S. G., & Millsap, R. E. (2008). Doctoral
independent or dependent variables would not be training in statistics, measurement, and methodology in
much more complex. Regardless of the number of psychology: Replication and extension of Aiken, West,
Sechrest, and Reno's (1990) survey of PhD programs in
variables, interpreting the MANOVA is the same as
North America. Americann Psychologist, 63,
63 32-50.
interpreting Tables 2 and 3. More complex models doi:10.1037/0003-066X.63.1.32
066X.63.1.32
couldd produce more than two DDA functions, but
Azen, R., & Budescu, D. V. (2003). The dominance analysis
interpreting additional functions is the same as
approach for comparing predictors in multiple
interpreting the two functions in the Add Health regression. Psychological Methods, 8,
8 129-148.
example. Moreover, with additional functions, SPSS doi:10.1037/1082-989x.8.2.129
989x.8.2.129
often produces interesting figures that can show Azen, R., & Budescu, D. V. (2006). Comparing predictors in
researchers the degree of overlap among groups and multivariate regression models: An extension of
how different their average members are from one dominance analysis. Journal of Educational and Behavioral
another. These figures can help researchers understand Statistics, 31, 157-180.
180. doi:10.3102/10769986031002157
the discriminant functions and how the groups differ Bangert, A. W., & Baumberger, J. P. (2005). Research and
from one another. For example, Shea, Lubinski, and statistical techniques used in the Journal of Counseling &
Benbow (2001, pp. 607-610)
610) presented five graphs that Development: 1990-2001. Journal of Counseling &
efficiently showed in three dimensions the differences Development, 83, 480-487.
487. doi:10.1002/j.1556-
doi:10.1002/j.1556
6678.2005.tb00369.x
Practical Assessment, Research & Evaluation, Vol 19, No 17 Page 9
Warne, MANOVA

Cole, D. A., Leeke, L. G., Martin, J. M., Truglio, R., & Statistician, 40, 205-214.
Seroczynski, A. D. (1998). A longitudinal look at the doi:10.1080/00031305.1986.10475394
relation between depression and anxiety in children and Smith, R. L., Ager, J. W., Jr., & Williams, D. L. (1992).
adolescents. Journal of Consulting and Clinical Psychology, Suppressor variables in multiple regression/correlation.
66, 451-460. doi:10.1037/0022-006X.66.3.451 Educational and Psychological Measurement, 52, 17-29.
Costello, A. B., & Osborne, J. W. (2005). Best practices in doi:10.1177/001316449205200102
exploratory factor analysis: Four recommendations for Steider, J. H., & Fouladi, R. T. (1997). Noncentrality interval
getting the most from your analysis. Practical Assessment, estimation and the evaluation of statistical models. In L. L.
Research & Evaluation, 10(7), 1-9. Available online at Harlow, S. A. Mulaik, & J. H. Steiger (Eds.), What if
http://pareonline.net/getvn.asp?v=10&n=7 there were no significance tests? (pp. 221-257). Mahwah, NJ:
Fish, L. J. (1988). Why multivariate methods are usually vital. Lawrence Erlbaum Associates.
Measurement and Evaluation in Counseling and Development, Stevens, J. P. (2002). Applied multivariate statistics for the social
21, 130-137. sciences. Mahwah, NJ: Lawrence Erlbaum Associates.
Haase, R. F., & Ellis, M. V. (1987). Multivariate analysis of Thomas, D. R. (1992). Interpreting discriminant functions:
variance. Journal of Counseling Psychology, 34, 404-413. A data analytic approach. Multivariate Behavioral Research,
doi:10.1037/0022-0167.34.4.404 27, 335-362. doi:10.1207/s15327906mbr2703_3
Hoenig, J. M., & Heisey, D. M. (2001). The abuse of power. Thomas, D. R., & Zumbo, B. D. (1996). Using a measure of
The American Statistician, 55, 19-24. variable importance to investigate the standardization
doi:10.1198/000313001300339897 of discriminant coefficients. Journal of Educational and
Huberty, C. J., & Morris, J. D. (1989). Multivariate analysis Behavioral Statistics, 21, 110-130.
versus multiple univariate analyses. Psychological Bulletin, doi:10.3102/10769986021002110
105, 302-308. doi:10.1037/0033-2909.105.2.302 Thompson, B. (2004). Exploratory and confirmatory factor
Huberty, C. J., & Olejnik, S. (2006). Applied MANOVA and analysis: Understanding concepts and applications.
discriminant analysis. Hoboken, NJ: John Wiley & Sons. Washington, DC: American Psychological Association.
Hummel, T. J., & Sligo, J. R. (1971). Empirical comparison Thompson, B. (2006). Foundations of behavioral statistics. New
of univariate and multivariate analysis of variance York, NY: The Guilford Press.
procedures. Psychological Bulletin, 76, 49-57. Tonidandel, S., & LeBreton, J. M. (2013). Beyond step-down
doi:10.1037/h0031323 analysis: A new test for decomposing the importance of
Kieffer, K. M., Reese, R. J., & Thompson, B. (2001). dependent variables in MANOVA. Journal of Applied
Statistical Techniques employed in AERJ and JCP Psychology, 98, 469-477. doi:10.1037/a0032001
articles from 1988 to 1997: A methodological review. Warne, R. T. (2011). Beyond multiple regression: Using
Journal of Experimental Education, 69, 280-309. commonality analysis to better understand R2 results.
doi:10.1080/00220970109599489 Gifted Child Quarterly, 55, 313-318.
Mineka, S., Watson, D., & Clark, L. A. (1998). Comorbidity doi:10.1177/0016986211422217
of anxiety and unipolar mood disorders. Annual Review Warne, R. T., Lazo, M., Ramos, T., & Ritter, N. (2012).
of Psychology, 49, 377-412. Statistical methods used in gifted education journals,
doi:10.1146/annurev.psych.49.1.377 2006-2010. Gifted Child Quarterly, 56, 134-149.
O’Brien, R. G., & Kaiser, M. K. (1985). MANOVA method doi:10.1177/0016986212444122
of analyzing repeated measures designs: An extensive Zientek, L. R., Capraro, M. M., & Capraro, R. M. (2008).
primer. Psychological Bulletin, 97, 316-333. Reporting practices in quantitative teacher education
doi:10.1037/0033-2909.97.2.316 research: One look at the evidence cited in the AERA
Olson, C. L. (1976). On choosing a test statistic in panel report. Educational Researcher, 37, 208-216.
multivariate analysis of variance. Psychological Bulletin, 83, doi:0.3102/0013189X08319762
579-586. doi:10.1037/0033-2909.83.4.579 Zientek, L. R., & Thompson, B. (2009). Matrix summaries
Saville, D. J., & Wood, G. R. (1986). A method for teaching improve research reports: Secondary analyses using
statistics using N-dimensional geometry. The American published literature. Educational Researcher, 38, 343-352.
doi:10.3102/0013189X09339056
Practical Assessment, Research & Evaluation, Vol 19, No 17 Page 10
Warne, MANOVA

Acknowledgment:

The author appreciates the feedback provided by Ross Larsen during the writing process.

Citation:
Warne, Russell (2014). A Primer on Multivariate Analysis of Variance (MANOVA) for Behavioral Scientists.
Practical Assessment, Research & Evaluation, 19(17). Available online:
http://pareonline.net/getvn.asp?v=19&n=17

Corresponding Author:
Russell T. Warne
Department of Behavioral Science
Utah Valley University
800 W. University Parkway MC 115
Orem, UT 84058
Email: [email protected]