Testing Covariance Patterns
The most basic use of PROC CALIS is testing covariance patterns. Consider a repeated-measures experiment where individuals
are tested for their motor skills at three different time points. No treatments are introduced between these tests. The three
test scores are denoted as , , and , respectively. These test scores are likely correlated because the same set of individuals has been used. More specifically,
the researcher wants to test the following pattern of the population covariance matrix :
Because there are no treatments between the tests, this pattern assumes that the distribution of motor skills stays more
or less the same over time, as represented by the same for the diagonal elements of . The covariances between the test scores for motor skills also stay the same, as represented by the same for all the off-diagonal elements of .
Suppose you summarize your data in a covariance matrix, which is stored in the following SAS data set:
data motor(type=cov);
input _type_ $ _name_ $ x1 x2 x3;
datalines;
COV x1 3.566 1.342 1.114
COV x2 1.342 4.012 1.056
COV x3 1.114 1.056 3.776
N . 36 36 36
;
The diagonal elements are somewhat close to each other but are not the same. The off-diagonal elements are also very close
to each other but are not the same. Could these observed differences be due to chance? Given the sample covariance matrix,
can you test the hypothesized patterned covariance matrix in the population?
Setting up this patterned covariance model in PROC CALIS is straightforward with the MSTRUCT modeling language:
proc calis data=motor;
mstruct var = x1-x3;
matrix _cov_ = phi
theta phi
theta theta phi;
run;
In the VAR= option in the MSTRUCT statement, you specify that x1
–x3
are the variables in the covariance matrix. Next, you specify the elements of the patterned covariance matrix in the MATRIX
statement with the _COV_ keyword. Because the covariance matrix is symmetric, you need to specify only the lower triangular
elements in the MATRIX statement. You use phi
for the parameters of all diagonal elements and theta
for the parameters of all off-diagonal elements. Matrix elements with the same parameter name are implicitly constrained
to be equal. Hence, this is the patterned covariance matrix that you want to test. Some output results from PROC CALIS are
shown in Figure 17.1.
First, PROC CALIS shows that the chi-square test for the model fit is 0.3656 (df = 4, p=0.9852). Because the chi-square test is not significant, it supports the hypothesized patterned covariance model. Next, PROC
CALIS shows the estimates in the covariance matrix under the hypothesized model. The estimates for the diagonal elements are
all 3.7847, and the estimates for off-diagonal elements are all 1.1707. Estimates of standard errors and t values for these covariance and variance parameters are also shown.
The MSTRUCT modeling language in PROC CALIS enables you to test various kinds of covariance and mean patterns, including matrices
with fixed or constrained values. For example, consider a population covariance model in which correlations among the motor
test scores are hypothesized to be zero. In other words, the covariance pattern is:
Essentially, this diagonally-patterned covariance model means that the data are randomly and independently generated for
x1
–x3
under the multivariate normal distribution. Only the variances of the variables are parameters in the model, and the variables
are not correlated at all.
You can use the MSTRUCT modeling language of PROC CALIS to fit this diagonally-patterned covariance matrix to the data for
motor skills, as shown in the following statements:
proc calis data=motor;
mstruct var = x1-x3;
matrix _cov_ = phi1
0. phi2
0. 0. phi3;
run;
Some of the output is shown in Figure 17.2.
PROC CALIS shows that the chi-square test for the model fit is 9.2939 (df=3, p=0.0256). Because the chi-square test is significant, it does not support the patterned covariance model that postulates zero
correlations among the variables. This conclusion is consistent with what is already known—the motor test scores should be
somewhat correlated because they are measurements over time for the same group of individuals.
The output also shows the estimates of variances under the model. Each diagonal element of the covariance matrix has a distinct
estimate because different parameters have been hypothesized under the patterned covariance model.