The TPSPLINE Procedure

Example 100.2 Spline Model with Higher-Order Penalty

This example continues the analysis of the data set Measure to illustrate how you can use PROC TPSPLINE to fit a spline model with a higher-order penalty term. Spline models with high-order penalty terms move low-order polynomial terms into the polynomial space. Hence, there is no penalty for these terms, and they can vary without constraint.

As shown in the previous analyses, the final model for the data set Measure must include quadratic terms for both $x_1$ and $x_2$. This example fits the following model:

\[  y = \beta _0 + \beta _1 x_1 + \beta _2 x_1^2 + \beta _3 x_2 + \beta _4 x_2^2 + \beta _5 x_1x_2 + f(x_1,x_2)  \]

The model includes quadratic terms for both variables, although it differs from the usual linear model. The nonparametric term $f(x_1,x_2)$ explains the variation of the data that is unaccounted for by a simple quadratic surface.

To modify the order of the derivative in the penalty term, specify the M=  option. The following statements specify the option M=3 in order to include the quadratic terms in the polynomial space:

data Measure;
   set Measure;
   x1sq = x1*x1;
   x2sq = x2*x2;
   x1x2 = x1*x2;
run;

proc tpspline data= Measure;
   model y = (x1 x2) / m=3;
   score data = pred out = predy;
run;

Output 100.2.1 displays the results from these statements.

Output 100.2.1: Output from PROC TPSPLINE with M=3

Raw Data

The TPSPLINE Procedure
Dependent Variable: y

Summary of Input Data Set
Number of Non-Missing Observations 50
Number of Missing Observations 0
Unique Smoothing Design Points 25

Summary of Final Model
Number of Regression Variables 0
Number of Smoothing Variables 2
Order of Derivative in the Penalty 3
Dimension of Polynomial Space 6

Summary Statistics of Final Estimation
log10(n*Lambda) -3.7831
Smoothing Penalty 2092.4495
Residual SS 0.2731
Tr(I-A) 29.1716
Model DF 20.8284
Standard Deviation 0.0968
GCV 0.0160


The model contains six terms in the polynomial space is the number of columns in ($(\{ 1,x_1,x_1^2,x_1x_2,x_2,x_2^2\} )=6$). Compare Output 100.2.1 with Output 100.1.1: the $\log _{10}(n\lambda )$ value and the smoothing penalty differ significantly. In general, these terms are not directly comparable for different models. The final estimate based on this model is close to the estimate based on the model by using the default, M=2.

In the following statements, the REG procedure fits a quadratic surface model to the data set Measure:

proc reg data= Measure;
   model y = x1 x1sq x2 x2sq x1x2;
run;

The results are displayed in Output 100.2.2.

Output 100.2.2: Quadratic Surface Model: The REG Procedure

Raw Data

The REG Procedure
Model: MODEL1
Dependent Variable: y

Analysis of Variance
Source DF Sum of
Squares
Mean
Square
F Value Pr > F
Model 5 443.20502 88.64100 436.33 <.0001
Error 44 8.93874 0.20315    
Corrected Total 49 452.14376      

Root MSE 0.45073 R-Square 0.9802
Dependent Mean 15.08548 Adj R-Sq 0.9780
Coeff Var 2.98781    

Parameter Estimates
Variable DF Parameter
Estimate
Standard
Error
t Value Pr > |t|
Intercept 1 14.90834 0.12519 119.09 <.0001
x1 1 0.01292 0.09015 0.14 0.8867
x1sq 1 -4.85194 0.15237 -31.84 <.0001
x2 1 0.02618 0.09015 0.29 0.7729
x2sq 1 5.20624 0.15237 34.17 <.0001
x1x2 1 -0.04814 0.12748 -0.38 0.7076


The REG procedure produces slightly different results. To fit a similar model with PROC TPSPLINE, you can use a MODEL statement that specifies the degrees of freedom with the DF= option. You can also use a large value for the LOGNLAMBDA0=  option to force a parametric model fit.

Because there is one degree of freedom for each of the terms intercept, x1, x2, x1sq, x2sq, and x1x2, the DF=6 option is used as follows:

proc tpspline data=measure;
   model  y=(x1 x2) /m=3 df=6 lognlambda=(-4 to 1 by 0.5);
   score data = pred
          out = predy;
run;

The fit statistics are displayed in Output 100.2.3.

Output 100.2.3: Output from PROC TPSPLINE Using M=3 and DF=6

Raw Data

The TPSPLINE Procedure
Dependent Variable: y

Summary of Final Model
Number of Regression Variables 0
Number of Smoothing Variables 2
Order of Derivative in the Penalty 3
Dimension of Polynomial Space 6

GCV Function
log10(n*Lambda) GCV  
-4.000000 0.016330 *
-3.500000 0.016889  
-3.000000 0.027496  
-2.500000 0.067672  
-2.000000 0.139642  
-1.500000 0.195727  
-1.000000 0.219512  
-0.500000 0.227306  
0 0.229740  
0.500000 0.230504  
1.000000 0.230745  

Note: * indicates minimum GCV value.


Summary Statistics of Final Estimation
log10(n*Lambda) 2.3830
Smoothing Penalty 0.0000
Residual SS 8.9384
Tr(I-A) 43.9997
Model DF 6.0003
Standard Deviation 0.4507
GCV 0.2309


Output 100.2.4 shows the GCV values for the list of supplied $\log _{10}(n\lambda )$ values in addition to the fitted model with fixed degrees of freedom 6. The fitted model has a larger GCV value than all other examined models.

Output 100.2.4: Criterion Plot


The final estimate is based on 6.000330 degrees of freedom because there are already 6 degrees of freedom in the polynomial space and the search range for $\lambda $ is not large enough (in this case, setting DF=6 is equivalent to setting $\lambda = \infty $).

The standard deviation and RSS (Output 100.2.3) are close to the sum of squares for the error term and the root MSE from the linear regression model (Output 100.2.2), respectively.

For this model, the optimal $\log _{10}(n\lambda )$ is around –3.8, which produces a standard deviation estimate of 0.096765 (see Output 100.2.1) and a GCV value of 0.016051, while the model that specifies DF=6 results in a $\log _{10}(n\lambda )$ larger than 1 and a GCV value larger than 0.23074. The nonparametric model, based on the GCV, should provide better prediction, but the linear regression model can be more easily interpreted.