Version info: Code for this page was tested in Stata 12.
MANOVA is used to model two or more dependent variables that are continuous with one or more categorical predictor variables.
Please note: The purpose of this page is to show how to use various data analysis commands. It does not cover all aspects of the research process which researchers are expected to do. In particular, it does not cover data cleaning and checking, verification of assumptions, model diagnostics or potential follow-up analyses.
Examples of one-way multivariate analysis of variance
Example 1.
A researcher randomly assigns 33 subjects to one of three groups. The first group receives technical dietary information interactively from an on-line website. Group 2 receives the same information from a nurse practitioner, while group 3 receives the information from a video tape made by the same nurse practitioner. The researcher looks at three different ratings of the presentation, difficulty, usefulness and importance, to determine if there is a difference in the modes of presentation. In particular, the researcher is interested in whether the interactive website is superior because that is the most cost-effective way of delivering the information.
Example 2. A clinical psychologist recruits 100 people who suffer from panic disorder into his study. Each subject receives one of four types of treatment for eight weeks. At the end of treatment, each subject participates in a structured interview, during which the clinical psychologist makes three ratings: physiological, emotional and cognitive. The clinical psychologist wants to know which type of treatment most reduces the symptoms of the panic disorder as measured on the physiological, emotional and cognitive scales. (This example was adapted from Grimm and Yarnold, 1995, page 246.)
Description of the data
Let’s pursue Example 1 from above.
We have a data file, manova.dta, with 33 observations on three response variables. The response variables are ratings called useful, difficulty and importance. Level 1 of the group variable is the treatment group, level 2 is control group 1 and level 3 is control group 2.
Let’s look at the data. It is always a good idea to start with descriptive statistics.
use https://stats.idre.ucla.edu/stat/stata/dae/manova, clear
summarize difficulty useful importance
Variable | Obs Mean Std. Dev. Min Max
-------------+--------------------------------------------------------
useful | 33 16.3303 3.292461 11.9 24.3
difficulty | 33 5.715152 2.017598 2.4 10.25
importance | 33 6.475758 3.985131 .2 18.8
tabulate group
group | Freq. Percent Cum.
------------+-----------------------------------
treatment | 11 33.33 33.33
control_1 | 11 33.33 66.67
control_2 | 11 33.33 100.00
------------+-----------------------------------
Total | 33 100.00
tabstat difficulty useful importance, by(group)
Summary statistics: mean
by categories of: group
group | useful diffic~y import~e
----------+------------------------------
treatment | 18.11818 6.190909 8.681818
control_1 | 15.52727 5.581818 5.109091
control_2 | 15.34545 5.372727 5.636364
----------+------------------------------
Total | 16.3303 5.715152 6.475758
correlate useful difficulty importance
(obs=33)
| useful diffic~y import~e
-------------+---------------------------
useful | 1.0000
difficulty | 0.0978 1.0000
importance | -0.3411 0.1978 1.0000
Analysis methods you might consider
Below is a list of some analysis methods you may have encountered. Some of the methods listed are quite reasonable, while others have either fallen out of favor or have limitations.
- MANOVA – This is a good option if there are two or more continuous dependent variables and one categorical predictor variable.
- Discriminant function analysis – This is a reasonable option and is equivalent to a one-way MANOVA.
- The data could be reshaped into long format and analyzed as a multilevel model.
- Separate univariate ANOVAs – You could analyze these data using separate univariate ANOVAs for each response variable. The univariate ANOVA will not produce multivariate results utilizing information from all variables simultaneously. In addition, separate univariate tests are generally less powerful because they do not take into account the inter-correlation of the dependent variables.
One-way MANOVA
We will start by running the manova command.
manova difficulty useful importance = group
Number of obs = 33
W = Wilks' lambda L = Lawley-Hotelling trace
P = Pillai's trace R = Roy's largest root
Source | Statistic df F(df1, df2) = F Prob>F
-----------+--------------------------------------------------
group | W 0.5258 2 6.0 56.0 3.54 0.0049 e
| P 0.4767 6.0 58.0 3.02 0.0122 a
| L 0.8972 6.0 54.0 4.04 0.0021 a
| R 0.8920 3.0 29.0 8.62 0.0003 u
|--------------------------------------------------
Residual | 30
-----------+--------------------------------------------------
Total | 32
--------------------------------------------------------------
e = exact, a = approximate, u = upper bound on F
Stata provides four multivariate tests by default. Each of these tests is statistically significant. For more information on these tests, please see our Stata Annotated Output: MANOVA page.
The overall multivariate test is significant, which means that differences between the levels of the variable group exist. To find where the differences lie, we will follow up with several post-hoc tests. We will begin with the multivariate test of group 1 versus the average of groups 2 and 3. First, we will use the manova, showorder command to determine the order of the elements in the design matrix. Knowing the order of the elements in the design matrix is necessary to run the post-hoc tests. (Note that the order of the elements in the design matrix changed in Stata 11.)
manovatest, showorder
Order of columns in the design matrix
1: (group==1)
2: (group==2)
3: (group==3)
4: _cons
We will begin by comparing the treatment group (group 1) to an average of the control groups (groups 2 and 3). This tests the hypothesis that the mean control groups equals the treatment group. The output above indicates that the fourth element in the matrix is the constant, so in the matrix command below, we will set it to 0. Once we have created a matrix (which we call c1), we can use the manovatest command to test c1.
matrix c1=(2,-1,-1,0)
manovatest, test(c1)
Test constraint
(1) 2*1.group - 2.group - 3.group = 0
W = Wilks' lambda L = Lawley-Hotelling trace
P = Pillai's trace R = Roy's largest root
Source | Statistic df F(df1, df2) = F Prob>F
-----------+--------------------------------------------------
manovatest | W 0.5290 1 3.0 28.0 8.31 0.0004 e
| P 0.4710 3.0 28.0 8.31 0.0004 e
| L 0.8904 3.0 28.0 8.31 0.0004 e
| R 0.8904 3.0 28.0 8.31 0.0004 e
|--------------------------------------------------
Residual | 30
--------------------------------------------------------------
e = exact, a = approximate, u = upper bound on F
These results indicate that group 1 is statistically significantly different from the average of groups 2 and 3.
Now we will compare control group 1 (group 2) to control group 2 (group 3). Again, we need to create a matrix (called c2 in this example) to do this comparison, and then use that matrix in the manovatest command.
matrix c2=(0,1,-1,0)
manovatest, test(c2)
Test constraint
(1) 2.group - 3.group = 0
W = Wilks' lambda L = Lawley-Hotelling trace
P = Pillai's trace R = Roy's largest root
Source | Statistic df F(df1, df2) = F Prob>F
-----------+--------------------------------------------------
manovatest | W 0.9932 1 3.0 28.0 0.06 0.9785 e
| P 0.0068 3.0 28.0 0.06 0.9785 e
| L 0.0068 3.0 28.0 0.06 0.9785 e
| R 0.0068 3.0 28.0 0.06 0.9785 e
|--------------------------------------------------
Residual | 30
--------------------------------------------------------------
e = exact, a = approximate, u = upper bound on F
The results indicate that control group 1 is not statistically significantly different from control group 2.
We can use the margins command to obtain adjusted predicted values for each of the groups. In the first example below, we get the predicted means for the dependent variable difficulty. In the next two examples, we get the predicted means for the dependent variables useful and importance. These values can be helpful in seeing where differences between levels of the predictor variable are and describing the model.
margins group, predict(equation(difficulty))
Adjusted predictions Number of obs = 33
Expression : Linear prediction: difficulty, predict(equation(difficulty))
------------------------------------------------------------------------------
| Delta-method
| Margin Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
group |
1 | 6.190909 .6186184 10.01 0.000 4.978439 7.403379
2 | 5.581818 .6186184 9.02 0.000 4.369349 6.794288
3 | 5.372727 .6186184 8.69 0.000 4.160257 6.585197
------------------------------------------------------------------------------
margins group, predict(equation(useful))
Adjusted predictions Number of obs = 33
Expression : Linear prediction: useful, predict(equation(useful))
------------------------------------------------------------------------------
| Delta-method
| Margin Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
group |
1 | 18.11818 .9438243 19.20 0.000 16.26832 19.96804
2 | 15.52727 .9438243 16.45 0.000 13.67741 17.37713
3 | 15.34545 .9438243 16.26 0.000 13.49559 17.19532
------------------------------------------------------------------------------
margins group, predict(equation(importance))
Adjusted predictions Number of obs = 33
Expression : Linear prediction: importance, predict(equation(importance))
------------------------------------------------------------------------------
| Delta-method
| Margin Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
group |
1 | 8.681818 1.136676 7.64 0.000 6.453973 10.90966
2 | 5.109091 1.136676 4.49 0.000 2.881246 7.336936
3 | 5.636364 1.136676 4.96 0.000 3.408519 7.864208
------------------------------------------------------------------------------
In each of the three outputs above, we see that the predicted means for groups 2 and 3 are very similar; the predicted mean for group 1 is higher than those for groups 2 and 3.
In the examples below, we obtain the differences in the means for each of the dependent variables for each of the control groups (groups 2 and 3) compared to the treatment group (group1). With respect to the dependent variable difficulty, the difference between the means for control group 1 versus the treatment group is approximately -0.61 (5.58 – 6.19). The difference between the means for control group 2 versus the treatment group is approximately -0.82 (5.37 – 6.19).
margins, dydx(group) predict(equation(difficulty))
Conditional marginal effects Number of obs = 33
Expression : Linear prediction: difficulty, predict(equation(difficulty))
dy/dx w.r.t. : 2.group 3.group
------------------------------------------------------------------------------
| Delta-method
| dy/dx Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
group |
2 | -.6090908 .8748585 -0.70 0.486 -2.323782 1.1056
3 | -.8181818 .8748585 -0.94 0.350 -2.532873 .8965094
------------------------------------------------------------------------------
Note: dy/dx for factor levels is the discrete change from the base level.
margins, dydx(group) predict(equation(useful))
Conditional marginal effects Number of obs = 33
Expression : Linear prediction: useful, predict(equation(useful))
dy/dx w.r.t. : 2.group 3.group
------------------------------------------------------------------------------
| Delta-method
| dy/dx Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
group |
2 | -2.590909 1.334769 -1.94 0.052 -5.207008 .0251907
3 | -2.772727 1.334769 -2.08 0.038 -5.388827 -.1566278
------------------------------------------------------------------------------
Note: dy/dx for factor levels is the discrete change from the base level.
margins, dydx(group) predict(equation(importance))
Conditional marginal effects Number of obs = 33
Expression : Linear prediction: importance, predict(equation(importance))
dy/dx w.r.t. : 2.group 3.group
------------------------------------------------------------------------------
| Delta-method
| dy/dx Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
group |
2 | -3.572727 1.607503 -2.22 0.026 -6.723375 -.4220792
3 | -3.045454 1.607503 -1.89 0.058 -6.196103 .1051936
------------------------------------------------------------------------------
Note: dy/dx for factor levels is the discrete change from the base level.
Finally, let’s run separate univariate ANOVAs. We will use a foreach loop to run the ANOVA for each dependent variable.
foreach vname in difficulty useful importance {
anova `vname' group
}
/* useful */
Number of obs = 33 R-squared = 0.1526
Root MSE = 3.13031 Adj R-squared = 0.0961
Source | Partial SS df MS F Prob > F
-----------+----------------------------------------------------
Model | 52.9242378 2 26.4621189 2.70 0.0835
|
group | 52.9242378 2 26.4621189 2.70 0.0835
|
Residual | 293.965442 30 9.79884808
-----------+----------------------------------------------------
Total | 346.88968 32 10.8403025
/* difficulty */
Number of obs = 33 R-squared = 0.0305
Root MSE = 2.05173 Adj R-squared = -0.0341
Source | Partial SS df MS F Prob > F
-----------+----------------------------------------------------
Model | 3.97515121 2 1.9875756 0.47 0.6282
|
group | 3.97515121 2 1.9875756 0.47 0.6282
|
Residual | 126.287277 30 4.20957589
-----------+----------------------------------------------------
Total | 130.262428 32 4.07070087
/* importance */
Number of obs = 33 R-squared = 0.1610
Root MSE = 3.76993 Adj R-squared = 0.1051
Source | Partial SS df MS F Prob > F
-----------+----------------------------------------------------
Model | 81.8296936 2 40.9148468 2.88 0.0718
|
group | 81.8296936 2 40.9148468 2.88 0.0718
|
Residual | 426.370896 30 14.2123632
-----------+----------------------------------------------------
Total | 508.20059 32 15.8812684
While none of the three ANOVAs were statistically significant at the alpha = .05 level, in particular, the F-ratio for difficulty was less than 1.
Things to consider
- One of the assumptions of MANOVA is that the response variables come from group populations that are multivariate normal distributed. This means that each of the dependent variables is normally distributed within group, that any linear combination of the dependent variables is normally distributed, and that all subsets of the variables must be multivariate normal. A partial test of this assumption can be obtained with the mvtest normality command. For example, mvtest normality difficult useful importance. (The mvtest command was introduced in Stata 11.) With respect to Type I error rate, MANOVA tends to be robust to minor violations of the multivariate normality assumption.
- The homogeneity of population covariance matrices (a.k.a. sphericity) is another assumption. This implies that the population variances and covariances of all dependent variables must be equal in all groups formed by the independent variables. A test of this assumption can be obtained with the mvtest covariance command. For example, mvtest covariance difficult useful importance, by(group).
- Small samples can have low power, but if the multivariate normality assumption is met, the MANOVA is generally more powerful than separate univariate tests.
- There are at least five types of follow-up analyses that can be done after a statistically significant MANOVA. These include multiple univariate ANOVAs, stepdown analysis, discriminant analysis, dependent variable contribution, and multivariate contrasts.
See also
- Stata online manual
References
- Grimm, L. G. and Yarnold, P. R. (editors). 1995. Reading and Understanding Multivariate Statistics. Washington, D.C.: American Psychological Association.
- Huberty, C. J. and Olejnik, S. 2006. Applied MANOVA and Discriminant Analysis, Second Edition. Hoboken, New Jersey: John Wiley and Sons, Inc.
- Stevens, J. P. 2002. Applied Multivariate Statistics for the Social Sciences, Fourth Edition. Mahwah, New Jersey: Lawrence Erlbaum Associates, Inc.
- Tatsuoka, M. M. 1971. Multivariate Analysis: Techniques for Educational and Psychological Research. New York: John Wiley and Sons.