The VARCOMP Procedure
This example, using data from Hicks (1973), concerns an experiment to determine the sources of variability in cure rates of rubber. The goal of the experiment was
to find out if the different laboratories contributed more to the variance of cure rates than did the different batches of
raw materials. This information would be useful in trying to control the cure rate of the final product because it would provide
insight into the sources of the variability in cure rates. The rubber used was cured at three temperatures, which were taken
to be fixed. Three laboratories were chosen at random, and three different batches of raw material were tested at each combination
of temperature and laboratory. The following statements read the data into the SAS data set Cure.
data Cure; input Lab Temp Batch $ Cure @@; datalines; 1 145 A 18.6 1 145 A 17.0 1 145 A 18.7 1 145 A 18.7 1 145 B 14.5 1 145 B 15.8 1 145 B 16.5 1 145 B 17.6 1 145 C 21.1 1 145 C 20.8 1 145 C 21.8 1 145 C 21.0 1 155 A 9.5 1 155 A 9.4 1 155 A 9.5 1 155 A 10.0 1 155 B 7.8 1 155 B 8.3 1 155 B 8.9 1 155 B 9.1 1 155 C 11.2 1 155 C 10.0 1 155 C 11.5 1 155 C 11.1 1 165 A 5.4 1 165 A 5.3 1 165 A 5.7 1 165 A 5.3 1 165 B 5.2 1 165 B 4.9 1 165 B 4.3 1 165 B 5.2 1 165 C 6.3 1 165 C 6.4 1 165 C 5.8 1 165 C 5.6 2 145 A 20.0 2 145 A 20.1 2 145 A 19.4 2 145 A 20.0 2 145 B 18.4 2 145 B 18.1 2 145 B 16.5 2 145 B 16.7 2 145 C 22.5 2 145 C 22.7 2 145 C 21.5 2 145 C 21.3 2 155 A 11.4 2 155 A 11.5 2 155 A 11.4 2 155 A 11.5 2 155 B 10.8 2 155 B 11.1 2 155 B 9.5 2 155 B 9.7 2 155 C 13.3 2 155 C 14.0 2 155 C 12.0 2 155 C 11.5 2 165 A 6.8 2 165 A 6.9 2 165 A 6.0 2 165 A 5.7 2 165 B 6.0 2 165 B 6.1 2 165 B 5.0 2 165 B 5.2 2 165 C 7.7 2 165 C 8.0 2 165 C 6.6 2 165 C 6.3 3 145 A 19.7 3 145 A 18.3 3 145 A 16.8 3 145 A 17.1 3 145 B 16.3 3 145 B 16.7 3 145 B 14.4 3 145 B 15.2 3 145 C 22.7 3 145 C 21.9 3 145 C 19.3 3 145 C 19.3 3 155 A 9.3 3 155 A 10.2 3 155 A 9.8 3 155 A 9.5 3 155 B 9.1 3 155 B 9.2 3 155 B 8.0 3 155 B 9.0 3 155 C 11.3 3 155 C 11.0 3 155 C 10.9 3 155 C 11.4 3 165 A 6.7 3 165 A 6.0 3 165 A 5.0 3 165 A 4.8 3 165 B 5.7 3 165 B 5.5 3 165 B 4.6 3 165 B 5.4 3 165 C 6.6 3 165 C 6.5 3 165 C 5.9 3 165 C 5.8 ;
The variables Lab, Temp, and Batch contain levels of laboratory, temperature, and batch, respectively. The Cure variable contains the response values.
The following SAS statements perform a restricted maximum likelihood variance component analysis.
title 'Analyzing the Cure Rate of Rubber'; proc varcomp method=reml data=cure; class temp lab batch; model cure=temp|lab batch(lab temp) / fixed=1; run;
The FIXED=1 option indicates that the first factor, Temp, is fixed. The effect specification Temp|Lab is equivalent to putting the three terms Temp, Lab, and Temp*Lab in the model. Batch(Lab Temp) is equivalent to putting Batch(Temp*Lab) in the MODEL
statement. The results of this analysis are displayed in Figure 108.1 through Figure 108.4.
Figure 108.1 provides information about the variables used in the analysis and the number of observations and specifies the dependent variable.
Figure 108.2: Iteration History
| REML Iterations | |||||
|---|---|---|---|---|---|
| Iteration | Objective | Var(Lab) | Var(Temp*Lab) | Var(Batch(Temp*Lab)) | Var(Error) |
| 0 | 13.4500060254 | 0.5094464340 | 0 | 2.4004888633 | 0.5787185225 |
| 1 | 13.0898262160 | 0.3194348317 | 0 | 2.0869636935 | 0.6016005334 |
| 2 | 13.0893125570 | 0.3176048001 | 0 | 2.0738906134 | 0.6026217204 |
| 3 | 13.0893125555 | 0.3176017115 | 0 | 2.0738685461 | 0.6026234568 |
The "REML Iterations" table in Figure 108.2 displays the iteration history, which includes the value of the objective function associated with REML and the values of
the variance components at each iteration.
Figure 108.3 displays the REML estimates of the variance components.
The "Asymptotic Covariance Matrix of Estimates" table in Figure 108.4 displays the asymptotic covariance matrix of the REML estimates.
The results of the analysis show that the variance attributable to Batch(Temp*Lab) (with a variance component of 2.0739) is considerably larger than the variance attributable to Lab (0.3176). Therefore, attempts to reduce the variability of cure rates should concentrate on improving the homogeneity of
the batches of raw material used rather than standardizing the practices or equipment within the laboratories. Also, note
that since the Batch(Temp*Lab) variance is considerably larger than the experimental error (Var(Error)=0.6026), the Batch(Temp*Lab) variability plays an important part in the overall variability of the cure rates.