Note: This example was done using Mplus version 5.2. The syntax may not work, or may function differently, with other versions of Mplus.
As the name implies, multivariate regression is a technique that estimates a single regression model with more than one outcome variable. When there is more than one predictor variable in a multivariate regression model, the model is a multivariate multiple regression.
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 and potential follow-up analyses.
Examples of multivariate regression analysis
Example 1. A researcher has collected data on three psychological variables, four academic variables (standardized test scores), and the type of educational program the student is in for 600 high school students. She is interested in how the set of psychological variables is related to the academic variables and the type of program the student is in.
Example 2. A doctor has collected data on cholesterol, blood pressure, and weight. She also collected data on the eating habits of the subjects (e.g., how many ounces of red meat, fish, dairy products, and chocolate consumed per week). She wants to investigate the relationship between the three measures of health and eating habits.
Example 3. A researcher is interested in determining what factors influence the health African Violet plants. She collects data on the average leaf diameter, the mass of the root ball, and the average diameter of the blooms, as well as how long the plant has been in its current container. For predictor variables, she measures several elements in the soil, as well as the amount of light and water each plant receives.
Description of the data
Let’s pursue Example 1 from above.
We have a hypothetical dataset with 600 observations on seven variables which can be obtained by clicking on mvreg.dat. The psychological variables are locus of control (locus), self-concept (self), and motivation (motiv). The academic variables are standardized tests scores in reading (read), writing (write), and science (science), as well as a categorical variable (prog) giving the type of program the student is in; general (prog=1), academic (prog=2), or vocational (prog=3). In addition to the three-category variable prog, the dataset contains a dummy variable for each level of prog (prog1, prog2, and prog3), for example, prog1 is equal to 1 when prog=1 (general), and 0 otherwise. You can store the data file anywhere you like, but our examples will assume it has been stored in c:data. (Note that the names of variables should NOT be included at the top of the data file. Instead, the variables are named as part of the variable command.) You may want to do your descriptive statistics in a general use statistics package, such as SAS, Stata or SPSS, because the options for obtaining descriptive statistics are limited in Mplus. Even if you chose to run descriptive statistics in another package, it is a good idea to run a model with type=basic before you do anything else, just to make sure the dataset is being read correctly. The input file below shows such a model.
Data: File is c:datamvreg.dat ; Variable: Names are locus self motiv read write science prog prog1 prog2 prog3; Missing are all (-9999) ; analysis: type = basic;
As we mentioned above, you will want to look at the output from this command carefully to be sure that the dataset was read into Mplus correctly. You will want to make sure that you have the correct number of observations, and that the variables all have means that are close to those from the descriptive statistics generated in a general purpose statistical package. If there are missing values for some or all of the variables, the descriptive statistics generated by Mplus will not match those from a general purpose statistical package exactly, because by default, Mplus versions 5.0 and later use maximum likelihood based procedures for handling missing values.
<output omitted> SUMMARY OF ANALYSIS Number of groups 1 Number of observations 600 <output omitted> SAMPLE STATISTICS Means LOCUS SELF MOTIV READ WRITE ________ ________ ________ ________ ________ 1 0.097 0.005 0.004 51.902 52.385 Means SCIENCE PROG PROG1 PROG2 PROG3 ________ ________ ________ ________ ________ 1 51.763 2.088 0.230 0.452 0.318
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.
- Multivariate multiple regression, the focus of this page.
- Separate OLS Regressions – You could analyze these data using separate OLS regression analyses for each outcome variable. The individual coefficients, as well as their standard errors will be the same as those produced by the multivariate regression. However, the OLS regressions will not produce multivariate results, nor will they allow for testing of coefficients across equations.
- Canonical correlation analysis might be feasible if you don’t want to consider one set of variables as outcome variables and the other set as predictor variables.
Multivariate regression analysis
The input file for our multivariate regression in Mplus is shown below. In the variable command, the usevariables option is used because only some of the variables in our dataset are used in the model. In the model command, each of the outcome variables (i.e., locus, self, and motiv) is predicted by the four predictor variables using the keyword on. In the output command we have requested fully standardized output (in addition to the unstandardized coefficients) using the stdyx option, this will produce standardized estimates of the coefficients, which you may find useful, but it also requests that Mplus produce the R-square statistic for each of the outcome variables.
data: file is C:datamvreg.dat ; variable: names are locus self motiv read write science prog prog1 prog2 prog3; missing are all (-9999) ; usevariables are locus self motiv read write science prog2 prog3; model: locus on read write science prog2 prog3; self on read write science prog2 prog3; motiv on read write science prog2 prog3; output: stdyx;
SUMMARY OF ANALYSIS Number of groups 1 Number of observations 600 Number of dependent variables 3 Number of independent variables 5 Number of continuous latent variables 0 Observed dependent variables Continuous LOCUS SELF MOTIV Observed independent variables READ WRITE SCIENCE PROG2 PROG3 Estimator ML Information matrix OBSERVED Maximum number of iterations 1000 Convergence criterion 0.500D-04 Maximum number of steepest descent iterations 20 Maximum number of iterations for H1 2000 Convergence criterion for H1 0.100D-03
- At the top of the output we see that 600 observations were used.
- From the output we see that the model includes 3 continuous dependent (outcome) variable and 5 independent (predictor) variables.
- The analysis summary is followed by a block of technical information about the model, we won’t discuss most of this information, but we will note that the default estimator is maximum likelihood (i.e., ML).
<output omitted> THE MODEL ESTIMATION TERMINATED NORMALLY TESTS OF MODEL FIT Chi-Square Test of Model Fit Value 0.000 Degrees of Freedom 0 P-Value 0.0000 Chi-Square Test of Model Fit for the Baseline Model Value 311.076 Degrees of Freedom 18 P-Value 0.0000 CFI/TLI CFI 1.000 TLI 1.000 Loglikelihood H0 Value -8807.819 H1 Value -8807.819 Information Criteria Number of Free Parameters 24 Akaike (AIC) 17663.637 Bayesian (BIC) 17769.164 Sample-Size Adjusted BIC 17692.970 (n* = (n + 2) / 24) RMSEA (Root Mean Square Error Of Approximation) Estimate 0.000 90 Percent C.I. 0.000 0.000 Probability RMSEA <= .05 0.000 SRMR (Standardized Root Mean Square Residual) Value 0.000
- Several measures of model fit are included in the output. The log likelihood (-8807.819) can be used in comparisons of nested models, but we won't show an example of that here.
- The Akaike information criterion (AIC) and the Bayesian information criterion (BIC, sometimes also called the Schwarz criterion), can also be used to compare models.
- Because the model is saturated, the chi-square test of model fit (for the current model, not the baseline model), as well as the CFI, TLI, RMSEA, and SRMR, all show perfect fit. This does not necessarily mean that the model fits well in the sense that the model does a good job of predicting the outcomes, just that these measures are not the appropriate measures of fit for a saturated model.
MODEL RESULTS Two-Tailed Estimate S.E. Est./S.E. P-Value LOCUS ON READ 0.013 0.004 3.380 0.001 WRITE 0.012 0.003 3.599 0.000 SCIENCE 0.006 0.004 1.590 0.112 PROG2 0.128 0.064 2.008 0.045 PROG3 0.252 0.068 3.694 0.000 SELF ON READ 0.001 0.004 0.311 0.755 WRITE -0.004 0.004 -1.121 0.262 SCIENCE 0.005 0.004 1.290 0.197 PROG2 0.276 0.072 3.827 0.000 PROG3 0.423 0.077 5.474 0.000 MOTIV ON READ 0.010 0.005 2.085 0.037 WRITE 0.018 0.004 4.143 0.000 SCIENCE -0.009 0.005 -1.981 0.048 PROG2 0.360 0.080 4.514 0.000 PROG3 0.620 0.085 7.252 0.000 SELF WITH LOCUS 0.057 0.017 3.335 0.001 MOTIV WITH LOCUS 0.060 0.019 3.179 0.001 SELF 0.130 0.022 5.935 0.000 Intercepts LOCUS -1.625 0.156 -10.401 0.000 SELF -0.372 0.177 -2.100 0.036 MOTIV -1.311 0.196 -6.689 0.000 Residual Variances LOCUS 0.365 0.021 17.321 0.000 SELF 0.470 0.027 17.320 0.000 MOTIV 0.574 0.033 17.321 0.000
- Under the heading MODEL RESULTS we first see the regression coefficients for the three outcome variables (locus, self, and motiv) regressed on the predictor variables, denoted ON. The output includes the estimate (i.e., the coefficient), its standard error, the est./s.e. (i.e., the z-value), and the associated p-value. The regression coefficients are followed by estimates of the covariances among the residuals of the outcome variables, denoted WITH. Finally, we see the intercepts for the three regression equations and the residual variances of the outcome variables.
- The coefficients are interpreted in the same way coefficients from an OLS regression are interpreted. For example, looking at the top of the MODEL RESULTS, a one unit change in read is associated with a 0.013 unit change in the predicted value of locus_of_control.
Because we used the stdyx option of the output command the output includes standardized coefficients. We did this primarily to obtain the R-square values for the outcome variables, so we have omitted the standardized output to save space.
<output omitted> R-SQUARE Observed Two-Tailed Variable Estimate S.E. Est./S.E. P-Value LOCUS 0.187 0.029 6.508 0.000 SELF 0.054 0.018 3.010 0.003 MOTIV 0.150 0.027 5.580 0.000
- The output above gives the R-square value for each of the outcome variables, as well as the standard error, z-value, and associated p-value for each. Based on the output, the model explains 18% of variance in locus of control (locus).
If you ran a separate OLS regression for each outcome variable, you would get exactly the same coefficients and standard errors as shown above. So why conduct a multivariate regression? One advantage of estimating the series of equations as a single model is that you can conduct tests of the coefficients across the different outcome variables. For example, the input file below uses the model test command to test the null hypothesis that the coefficients for the variable read are equal to 0 in all three equations. Notice that in the model command each of the terms we wish to test (i.e., each instance of read) is followed by a label in parentheses (e.g., "(r1)"). These parameter labels are then used to refer to the associated coefficients in the model test command. There are a few important things to note about parameter labels. First, the labels must always appear at the end of a line (but not necessarily the end of the command). Second, the labels apply to all parameters listed on the line (meaning all of the parameters on the line are constrained to equality). This is why read is the only predictor variable on the line with the label on it. In the model test command, we give the null hypotheses we wish to test together, in this case, that each of the parameters for read (identified as r1, r2, and r3) are simultaneously equal to zero.
data: file is C:datamvreg.dat ; variable: names are locus self motiv read write science prog prog1 prog2 prog3; missing are all (-9999) ; usevariables are locus self motiv read write science prog2 prog3; model: locus on read (r1) write science prog2 prog3; self on read (r2) write science prog2 prog3; motiv on read (r3) write science prog2 prog3; model test: r1 = 0; r2 = 0; r3 = 0; output: stdyx;
The output generated by this syntax will be identical to the output shown above, except that it will include the additional output generated by the model test command, the additional output is shown below (all other output is omitted).
Wald Test of Parameter Constraints Value 14.486 Degrees of Freedom 3 P-Value 0.0023
The Wald test statistic of 14.486 with 3 degrees of freedom has an associated p-value of 0.0023. These results reject the null hypothesis that the coefficients for read across the three equations are simultaneously equal to 0, in other words, the coefficients for read, taken for all three outcomes together, are statistically significant.
We can also test the null hypothesis that the coefficients for prog=2 (prog2) and prog=3 (prog3) are simultaneously equal to 0 in the equation for locus_of_control. When used to test the coefficients for dummy variables that form a single categorical predictor, this type of test is sometimes called an overall test for the effect of the categorical predictor (i.e., prog).
data: file is C:datamvreg.dat ; variable: names are locus self motiv read write science prog prog1 prog2 prog3; missing are all (-9999) ; usevariables are locus self motiv read write science prog2 prog3; model: locus on read write science prog2 (p1) prog3 (p2); self on read write science prog2 prog3; motiv on read write science prog2 prog3; model test: p1 = 0; p2 = 0; output: stdyx;
Wald Test of Parameter Constraints Value 13.788 Degrees of Freedom 2 P-Value 0.0010
The results of the above test indicate that the two coefficients together are significantly different from 0, in other words, the overall effect of prog on locus_of_control is statistically significant.
The next example tests the null hypothesis that the coefficient for the variable write in the equation with locus_of_control as the outcome is equal to the coefficient for write in the equation with self_concept as the outcome. Another way of stating this null hypothesis is that the effect of write on locus_of_control is equal to the effect of write on self_concept.
Data: File is mvreg.dat ; Variable: Names are locus self motiv read write science prog prog1 prog2 prog3; Missing are all (-9999) ; usevariables are locus self motiv read write science prog2 prog3; model: locus on read write (wl) science prog2 prog3; self on read write (ws) science prog2 prog3; motiv on read write science prog2 prog3; model test: wl = ws;
Wald Test of Parameter Constraints Value 12.006 Degrees of Freedom 1 P-Value 0.0005
The results of this test indicate that the coefficients for write with locus_of_control and self_concept as the outcome are significantly different.
Below we test the null hypothesis that the coefficient of science in the equation for locus_of_control is equal to the coefficient for science in the equation for self_concept, and that the coefficient for the variable write in the equation with the outcome variable locus_of_control equals the coefficient for write in the equation with the outcome variable self_concept.
data: file is mvreg.dat ; variable: names are locus self motiv read write science prog prog1 prog2 prog3; missing are all (-9999) ; usevariables are locus self motiv read write science prog2 prog3; model: locus on read write (w1) science (s1) prog2 prog3; self on read write (w2) science (s2) prog2 prog3; motiv on read write science prog2 prog3; model test: w1 = w2; s1 = s2; output: stdyx;
Wald Test of Parameter Constraints Value 12.902 Degrees of Freedom 2 P-Value 0.0016
The results of the above test indicate that taken together the two sets of coefficients are significantly different. Note that the degrees of freedom is now 2, reflecting the fact that we are comparing two sets of coefficients, rather than 1.
Unlike some other packages, Mplus does not automatically provide a test for the overall model. However, we can produce an equivalent test by constraining the regression coefficients to 0 in our model and comparing the fit of that model to the current saturated model. To constrain all of the regression coefficients to 0, we first constrain all of the coefficients by giving them the label n (recall from above that the label applies to all coefficients on the line. Below that, we use the model constraint command to fix n to 0.
Data: File is mvreg.dat ; Variable: Names are locus self motiv read write science prog prog1 prog2 prog3; Missing are all (-9999) ; usevariables are locus self motiv read write science prog1 prog2 ; model: locus on read write science prog1 prog2 (n) ; self on read write science prog1 prog2 (n); motiv on read write science prog1 prog2 (n); model constraint: n = 0;
Because this isn't the model we want to interpret, we have omitted most of the output. Shown below are the chi-square test of model fit (which provides the overall test) and the MODEL RESULTS section so that we can check to see we have estimated the desired model.
TESTS OF MODEL FIT Chi-Square Test of Model Fit Value 214.658 Degrees of Freedom 15 P-Value 0.0000
The chi-square test of model fit compares the fit of the current model to a saturated model. In the models we estimated above (i.e., the unconstrained or saturated models), this value was 0, because the model was saturated (i.e., has 0 degrees of freedom). By adding constraints to the model we have freed up 15 parameters, so now we get a positive value. The chi-square value of 214.658 with 15 degrees of freedom with an associated p-value of less than 0.0001, indicates that the constrained model fits significantly worse than the saturated model. In other words, the saturated model shown above fits significantly better than the model with the regression coefficients constrained to 0.
The MODEL RESULTS are shown below. It can be a good idea to check this section to make sure the model estimated was the desired model. Note that all of the regression coefficients (denoted ON) are constrained to 0, while the residual covariances (denoted WITH) and variances, as well as the intercepts have been estimated.
MODEL RESULTS Two-Tailed Estimate S.E. Est./S.E. P-Value LOCUS ON READ 0.000 0.000 999.000 999.000 WRITE 0.000 0.000 999.000 999.000 SCIENCE 0.000 0.000 999.000 999.000 PROG1 0.000 0.000 999.000 999.000 PROG2 0.000 0.000 999.000 999.000 SELF ON READ 0.000 0.000 999.000 999.000 WRITE 0.000 0.000 999.000 999.000 SCIENCE 0.000 0.000 999.000 999.000 PROG1 0.000 0.000 999.000 999.000 PROG2 0.000 0.000 999.000 999.000 MOTIV ON READ 0.000 0.000 999.000 999.000 WRITE 0.000 0.000 999.000 999.000 SCIENCE 0.000 0.000 999.000 999.000 PROG1 0.000 0.000 999.000 999.000 PROG2 0.000 0.000 999.000 999.000 SELF WITH LOCUS 0.081 0.020 4.133 0.000 MOTIV WITH LOCUS 0.135 0.023 5.832 0.000 SELF 0.167 0.025 6.791 0.000 Intercepts LOCUS 0.097 0.027 3.531 0.000 SELF 0.005 0.029 0.171 0.864 MOTIV 0.004 0.034 0.116 0.907 Residual Variances LOCUS 0.449 0.026 17.321 0.000 SELF 0.497 0.029 17.321 0.000 MOTIV 0.675 0.039 17.321 0.000
Things to consider
- The residuals from multivariate regression models are assumed to be multivariate normal. This is analogous to the assumption of normally distributed errors in univariate linear regression (i.e., OLS regression).
- Multivariate regression analysis is not recommended for small samples.
- The outcome variables should be at least moderately correlated for the multivariate regression analysis to make sense.
Afifi, A., Clark, V. and May, S. 2004. Computer-Aided Multivariate Analysis. 4th ed. Boca Raton, Fl: Chapman & Hall/CRC.