The HPMIXED Procedure

Example 48.2 Comparing Results from PROC HPMIXED and PROC MIXED

This example revisits the mixed model problem from the section Getting Started: MIXED Procedure, in Chapter 63: The MIXED Procedure, with the data set shown in the following statements:

data heights;
   input Family Gender$ Height @@;
   datalines;
1 F 67   1 F 66   1 F 64   1 M 71   1 M 72   2 F 63
2 F 63   2 F 67   2 M 69   2 M 68   2 M 70   3 F 63
3 M 64   4 F 67   4 F 66   4 M 67   4 M 67   4 M 69
;

The response variable Height measures the heights (in inches) of 18 individuals. The individuals are classified according to Family and Gender. The following statements fit a mixed model with random effects for Family and the Family*Gender interaction with the MIXED procedure:

proc mixed;
   class Family Gender;
   model Height = Gender / s;
   random Family Family*Gender / s;
run;

The Iteration History and Fit Statistics tables for the optimization in PROC MIXED are shown in Output 48.2.1. The MIXED procedure converges after six iterations and achieves a –2 restricted log likelihood of 71.02246.

Output 48.2.1: Iteration History and Fit Statistics: MIXED Procedure

The Mixed Procedure

Iteration History
Iteration Evaluations -2 Res Log Like Criterion
0 1 74.11074833  
1 2 71.51614003 0.01441208
2 1 71.13845990 0.00412226
3 1 71.03613556 0.00058188
4 1 71.02281757 0.00001689
5 1 71.02245904 0.00000002
6 1 71.02245869 0.00000000

Fit Statistics
-2 Res Log Likelihood 71.0
AIC (Smaller is Better) 77.0
AICC (Smaller is Better) 79.0
BIC (Smaller is Better) 75.2


Output 48.2.2 displays the covariance parameter estimates and the solutions for the fixed and random effects. Because the fixed-effect model contains a classification effect (Gender) and an intercept, the $\bX ’\bX $ matrix is singular. Only two fixed-effect parameters can be estimated in this model. The MIXED procedure, relying on a sweep operation in the order in which effects enter the model, determines that the last column of the $\bX ’\bX $ matrix is a linear function of previous columns. Consequently, the coefficient for the second level of the Gender variable is zero.

Output 48.2.2: Parameter Estimates and Solutions: MIXED Procedure

Covariance Parameter Estimates
Cov Parm Estimate
Family 2.4010
Family*Gender 1.7657
Residual 2.1668

Solution for Fixed Effects
Effect Gender Estimate Standard Error DF t Value Pr > |t|
Intercept   68.2114 1.1477 3 59.43 <.0001
Gender F -3.3621 1.1923 3 -2.82 0.0667
Gender M 0 . . . .

Solution for Random Effects
Effect Gender Family Estimate Std Err Pred DF t Value Pr > |t|
Family   1 1.2680 1.1201 10 1.13 0.2840
Family   2 0.08980 1.1121 10 0.08 0.9372
Family   3 -1.6660 1.1712 10 -1.42 0.1853
Family   4 0.3082 1.1201 10 0.28 0.7888
Family*Gender F 1 -0.3198 1.0810 10 -0.30 0.7734
Family*Gender M 1 1.2523 1.0933 10 1.15 0.2787
Family*Gender F 2 -0.4299 1.0774 10 -0.40 0.6983
Family*Gender M 2 0.4959 1.0774 10 0.46 0.6551
Family*Gender F 3 -0.08229 1.1409 10 -0.07 0.9439
Family*Gender M 3 -1.1429 1.1409 10 -1.00 0.3401
Family*Gender F 4 0.8320 1.0933 10 0.76 0.4642
Family*Gender M 4 -0.6053 1.0810 10 -0.56 0.5878


The Type 3 Tests of Fixed Effects table in Output 48.2.3 is produced by the MIXED procedure by default.

Output 48.2.3: Test of Gender Effect

Type 3 Tests of Fixed Effects
Effect Num DF Den DF F Value Pr > F
Gender 1 3 7.95 0.0667


The same linear mixed model is fit with the HPMIXED procedure with the following statements:

proc hpmixed;
   class Family Gender;
   model Height = Gender / s;
   random Family Family*Gender / s;
   test gender;
run;

Output 48.2.4 displays the Iteration History and Fit Statistics tables. The HPMIXED procedure, with its default quasi-Newton algorithm, achieves the same –2 restricted log likelihood as the MIXED procedure (71.02246; see Output 48.2.1).

Output 48.2.4: Iteration History and Fit Statistics: HPMIXED Procedure

The HPMIXED Procedure

Iteration History
Iteration Evaluations Objective
Function
Change Max
Gradient
0 4 71.023177956 . 0.034074
1 3 71.022519936 0.00065802 0.007839
2 3 71.022477283 0.00004265 0.004674
3 2 71.0224587 0.00001858 0.000168
4 2 71.022458689 0.00000001 3.28E-6

Fit Statistics
-2 Res Log Likelihood 71.02246
AIC (Smaller is Better) 77.02246
AICC (Smaller is Better) 79.02246
BIC (Smaller is Better) 75.18134
CAIC (Smaller is Better) 78.18134
HQIC (Smaller is Better) 72.98226


Output 48.2.5 displays the results that correspond to those in Output 48.2.2 in the MIXED procedure.

Output 48.2.5: Parameter Estimates and Solutions: HPMIXED Procedure

Covariance Parameter Estimates
Cov Parm Estimate
Family 2.4010
Family*Gender 1.7657
Residual 2.1668

Solution for Fixed Effects
Effect Gender Estimate Standard Error DF t Value Pr > |t|
Intercept   0 . . . .
Gender F 64.8493 1.1477 16 56.50 <.0001
Gender M 68.2114 1.1477 16 59.43 <.0001

Solution for Random Effects
Effect Gender Family Estimate Std Err Pred DF t Value Pr > |t|
Family   1 1.2680 1.1201 16 1.13 0.2743
Family   2 0.08980 1.1121 16 0.08 0.9366
Family   3 -1.6660 1.1712 16 -1.42 0.1741
Family   4 0.3082 1.1201 16 0.28 0.7867
Family*Gender F 1 -0.3198 1.0810 16 -0.30 0.7712
Family*Gender M 1 1.2523 1.0933 16 1.15 0.2689
Family*Gender F 2 -0.4299 1.0774 16 -0.40 0.6951
Family*Gender M 2 0.4959 1.0774 16 0.46 0.6515
Family*Gender F 3 -0.08229 1.1409 16 -0.07 0.9434
Family*Gender M 3 -1.1429 1.1409 16 -1.00 0.3314
Family*Gender F 4 0.8320 1.0933 16 0.76 0.4577
Family*Gender M 4 -0.6053 1.0810 16 -0.56 0.5832


A number of points are noteworthy in comparing the results from the procedures. The covariance parameter estimates are the same, yet the solutions for the fixed effects differ. In fact, both solutions are correct. Solving a sparse system of linear equations requires reordering of the mixed model equations to minimize memory consumption in the factorization process. As a consequence, the order in which singularities are detected can differ from the order in which effects enter the model. Mathematically, the two sets of solutions simply correspond to different choices for the generalized inverse in solving a singular linear system. See the sections Generalized Inverse Matrices and Linear Model Theory, in Chapter 3: Introduction to Statistical Modeling with SAS/STAT Software, for more information about the role and importance of generalized inverses in linear model analysis.

Although the two sets of solutions for the fixed effects correspond to different choices of generalized inverses, many important results are invariant to the choice of the g-inverse. For example, the solutions for the random effects in Output 48.2.5 and Output 48.2.2 are identical. Also, the test for the Gender effect yields the same F value in both analyses (compare Output 48.2.6 and Output 48.2.3). However, note that the p-values associated with both F tests and t tests differ between the two procedures. This is due to their different default methods for computing the degrees of freedom. For this model, the HPMIXED procedure use the residual method to determine the denominator degrees of freedom for tests of fixed effects, whereas the MIXED procedure uses the containment method. The containment method is order-dependent, and thus not available in the HPMIXED procedure.

Output 48.2.6: Parameter Estimates and Solutions: HPMIXED Procedure

Type III Tests of Fixed Effects
Effect Num DF Den DF F Value Pr > F
Gender 1 16 7.95 0.0123