Example 29.9 Repeated Measures, Two Repeated Measurement Factors

This example, from MacMillan et al. (1981), illustrates a repeated measures analysis in which there are two repeated measurement factors. Two diagnostic procedures (standard and test) are performed on each subject, and the results of both are evaluated at each of two times as being positive or negative. In the following DATA step, std1 and std2 are the two measurements of the standard procedure, and test1 and test2 are the two measurements of the test procedure:

data a;
   input std1 $ test1 $ std2 $ test2 $ wt @@;
   datalines;
neg neg neg neg 509  neg neg neg pos  4  neg neg pos neg  17
neg neg pos pos   3  neg pos neg neg 13  neg pos neg pos   8
neg pos pos pos   8  pos neg neg neg 14  pos neg neg pos   1
pos neg pos neg  17  pos neg pos pos  9  pos pos neg neg   7
pos pos neg pos   4  pos pos pos neg  9  pos pos pos pos 170
;

For the initial model, the response functions are marginal probabilities, and the repeated measurement factors are Time and Treatment. The model is a saturated one, containing effects for Time, Treatment, and Time*Treatment. The following statements produce Output 29.9.1:


proc catmod data=a;
   title2 'Marginal Symmetry, Saturated Model';
   weight wt;
   response marginals;
   model std1*test1*std2*test2=_response_ / freq design noparm;
   repeated Time 2, Treatment 2 / _response_=Time Treatment 
            Time*Treatment;
run;

The analysis of variance table in Output 29.9.1 shows that there is no significant effect of Time, either by itself or in its interaction with Treatment. The second model includes only the Treatment effect. Again, the response functions are marginal probabilities, and the repeated measurement factors are Time and Treatment.

Output 29.9.1 Diagnosis Data: Two Repeated Measurement Factors
Diagnostic Procedure Comparison
Marginal Symmetry, Saturated Model

The CATMOD Procedure

Data Summary
Response std1*test1*std2*test2 Response Levels 15
Weight Variable wt Populations 1
Data Set A Total Frequency 793
Frequency Missing 0 Observations 15

Population Profiles
Sample Sample Size
1 793

Response Profiles
Response std1 test1 std2 test2
1 neg neg neg neg
2 neg neg neg pos
3 neg neg pos neg
4 neg neg pos pos
5 neg pos neg neg
6 neg pos neg pos
7 neg pos pos pos
8 pos neg neg neg
9 pos neg neg pos
10 pos neg pos neg
11 pos neg pos pos
12 pos pos neg neg
13 pos pos neg pos
14 pos pos pos neg
15 pos pos pos pos

Response Frequencies
Sample Response Number
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
1 509 4 17 3 13 8 8 14 1 17 9 7 4 9 170

Response Functions and Design Matrix
Sample Function
Number
Response
Function
Design Matrix
1 2 3 4
1 1 0.70870 1 1 1 1
  2 0.72383 1 1 -1 -1
  3 0.70618 1 -1 1 -1
  4 0.73897 1 -1 -1 1

Analysis of Variance
Source DF Chi-Square Pr > ChiSq
Intercept 1 2385.34 <.0001
Time 1 0.85 0.3570
Treatment 1 8.20 0.0042
Time*Treatment 1 2.40 0.1215
Residual 0 . .

A main effect model with respect to Treatment is fit. The following statements produces Output 29.9.2:

   title2 'Marginal Symmetry, Reduced Model';
   model std1*test1*std2*test2=_response_ / corrb design noprofile;
   repeated Time 2, Treatment 2 / _response_=Treatment;
run;

The analysis of variance table for the reduced model (Output 29.9.2) shows that the model fits (since the residual chi-square is nonsignificant) and that the treatment effect is significant. The negative parameter estimate for Treatment shows that the first level of treatment (std) has a smaller probability of the first response level (neg) than the second level of treatment (test). In other words, the standard diagnostic procedure gives a significantly higher probability of a positive response than the test diagnostic procedure.

Output 29.9.2 Diagnosis Data: Reduced Model
Diagnostic Procedure Comparison
Marginal Symmetry, Reduced Model

The CATMOD Procedure

Data Summary
Response std1*test1*std2*test2 Response Levels 15
Weight Variable wt Populations 1
Data Set A Total Frequency 793
Frequency Missing 0 Observations 15

Response Functions and Design Matrix
Sample Function
Number
Response
Function
Design Matrix
1 2
1 1 0.70870 1 1
  2 0.72383 1 -1
  3 0.70618 1 1
  4 0.73897 1 -1

Analysis of Variance
Source DF Chi-Square Pr > ChiSq
Intercept 1 2386.97 <.0001
Treatment 1 9.55 0.0020
Residual 2 3.51 0.1731

Analysis of Weighted Least Squares Estimates
Effect Parameter Estimate Standard
Error
Chi-
Square
Pr > ChiSq
Intercept 1 0.7196 0.0147 2386.97 <.0001
Treatment 2 -0.0128 0.00416 9.55 0.0020

Correlation Matrix of the Parameter
Estimates
Row Col1 Col2
1 1.00000 0.04194
2 0.04194 1.00000

The next example illustrates a RESPONSE statement that, at each time, computes the sensitivity and specificity of the test diagnostic procedure with respect to the standard procedure. Since these are measures of the relative accuracy of the two diagnostic procedures, the repeated measurement factors in this case are labeled Time and Accuracy. Only 15 of the 16 possible responses are observed, so additional care must be taken in formulating the RESPONSE statement for computation of sensitivity and specificity.

The following statements produce Output 29.9.3 and Output 29.9.4:

   title2 'Sensitivity and Specificity Analysis, ' 
          'Main-Effects Model';
   model std1*test1*std2*test2=_response_ / covb design noprofile;
   repeated Time 2, Accuracy 2 / _response_=Time Accuracy;
   response exp 1 -1  0  0  0  0  0  0,
                0  0  1 -1  0  0  0  0,
                0  0  0  0  1 -1  0  0,
                0  0  0  0  0  0  1 -1

            log 0 0 0 0   0 0  0   0 0 0 0   1 1 1 1,
                0 0 0 0   0 0  0   1 1 1 1   1 1 1 1,
                1 1 1 1   0 0  0   0 0 0 0   0 0 0 0,
                1 1 1 1   1 1  1   0 0 0 0   0 0 0 0,
                0 0 0 1   0 0  1   0 0 0 1   0 0 0 1,
                0 0 1 1   0 0  1   0 0 1 1   0 0 1 1,
                1 0 0 0   1 0  0   1 0 0 0   1 0 0 0,
                1 1 0 0   1 1  0   1 1 0 0   1 1 0 0;
quit;

For the sensitivity and specificity analysis, the four response functions displayed next to the design matrix (Output 29.9.3) represent the following:

  1. sensitivity, time 1

  2. specificity, time 1

  3. sensitivity, time 2

  4. specificity, time 2

The sensitivities and specificities are for the test diagnostic procedure relative to the standard procedure.

Output 29.9.3 Diagnosis Data: Sensitivity and Specificity Analysis
Diagnostic Procedure Comparison
Sensitivity and Specificity Analysis, Main-Effects Model

The CATMOD Procedure

Data Summary
Response std1*test1*std2*test2 Response Levels 15
Weight Variable wt Populations 1
Data Set A Total Frequency 793
Frequency Missing 0 Observations 15

Response Functions and Design Matrix
Sample Function
Number
Response
Function
Design Matrix
1 2 3
1 1 0.82251 1 1 1
  2 0.94840 1 1 -1
  3 0.81545 1 -1 1
  4 0.96964 1 -1 -1

Analysis of Variance
Source DF Chi-Square Pr > ChiSq
Intercept 1 6448.79 <.0001
Time 1 4.10 0.0428
Accuracy 1 38.81 <.0001
Residual 1 1.00 0.3178

The ANOVA table in Output 29.9.3 shows that an additive model fits, that there is a significant effect of time, and that the sensitivity is significantly different from the specificity.

Output 29.9.4 shows that the predicted sensitivities and specificities are lower for time 1 (since parameter 2 is negative). It also shows that the sensitivity is significantly less than the specificity.

Output 29.9.4 Parameter Estimates
Analysis of Weighted Least Squares Estimates
Effect Parameter Estimate Standard
Error
Chi-
Square
Pr > ChiSq
Intercept 1 0.8892 0.0111 6448.79 <.0001
Time 2 -0.00932 0.00460 4.10 0.0428
Accuracy 3 -0.0702 0.0113 38.81 <.0001

Covariance Matrix of the Parameter Estimates
Row Col1 Col2 Col3
1 0.00012260 0.00000229 0.00010137
2 0.00000229 0.00002116 -.00000587
3 0.00010137 -.00000587 0.00012697