The HPLOGISTIC Procedure

Binary Logistic Regression

The following DATA step contains 100 observations on a dichotomous response variable (y), a character variable (C), and 10 continuous variables (x1x10):

data getStarted;
   input C$ y x1-x10;
   datalines;
   D  0  10.2  6  1.6  38  15  2.4  20  0.8  8.5  3.9
   F  1  12.2  6  2.6  42  61  1.5  10  0.6  8.5  0.7
   D  1   7.7  1  2.1  38  61    1  90  0.6  7.5  5.2
   J  1  10.9  7  3.5  46  42  0.3   0  0.2    6  3.6
   E  0  17.3  6  3.8  26  47  0.9  10  0.4  1.5  4.7
   A  0  18.7  4  1.8   2  34  1.7  80    1  9.5  2.2
   B  0   7.2  1  0.3  48  61  1.1  10  0.8  3.5    4
   D  0   0.1  3  2.4   0  65  1.6  70  0.8  3.5  0.7
   H  1   2.4  4  0.7  38  22  0.2  20    0    3  4.2
   J  0  15.6  7  1.4   0  98  0.3   0    1    5  5.2
   J  0  11.1  3  2.4  42  55  2.2  60  0.6  4.5  0.7
   F  0     4  6  0.9   4  36  2.1  30  0.8    9  4.6
   A  0   6.2  2  1.8  14  79  1.1  70  0.2    0  5.1
   H  0   3.7  3  0.8  12  66  1.3  40  0.4  0.5  3.3
   A  1   9.2  3  2.3  48  51  2.3  50    0    6  5.4
   G  0    14  3    2  18  12  2.2   0    0    3  3.4
   E  1  19.5  6  3.7  26  81  0.1  30  0.6    5  4.8
   C  0    11  3  2.8  38   9  1.7  50  0.8  6.5  0.9
   I  0  15.3  7  2.2  20  98  2.7 100  0.4    7  0.8
   H  1   7.4  4  0.5  28  65  1.3  60  0.2  9.5  5.4
   F  0  11.4  2  1.4  42  12  2.4  10  0.4    1  4.5
   C  1  19.4  1  0.4  42   4  2.4  10    0  6.5  0.1
   G  0   5.9  4  2.6  12  57  0.8  50  0.4    2  5.8
   G  1  15.8  6  3.7  34   8  1.3  90  0.6  2.5  5.7
   I  0    10  3  1.9  16  80    3  90  0.4  9.5  1.9
   E  0  15.7  1  2.7  32  25  1.7  20  0.2  8.5    6
   G  0    11  5  2.9  48  53  0.1  50    1  3.5  1.2
   J  1  16.8  0  0.9  14  86  1.4  40  0.8    9    5
   D  1    11  4  3.2  48  63  2.8  90  0.6    0  2.2
   J  1   4.8  7  3.6  24   1  2.2  20    1  8.5  0.5
   J  1  10.4  5    2  42  56    1  20    0  3.5  4.2
   G  0  12.7  7  3.6   8  56  2.1  70    1  4.5  1.5
   G  0   6.8  1  3.2  30  27  0.6   0  0.8    2  5.6
   E  0   8.8  0  3.2   2  67  0.7  10  0.4    1    5
   I  1   0.2  0  2.9  10  41  2.3  60  0.2    9  0.3
   J  1   4.6  7  3.9  50  61  2.1  50  0.4    3  4.9
   J  1   2.3  2  3.2  36  98  0.1  40  0.6  4.5  4.3
   I  0  10.8  3  2.7  28  58  0.8  80  0.8    3    6
   B  0   9.3  2  3.3  44  44  0.3  50  0.8  5.5  0.4
   F  0   9.2  6  0.6   4  64  0.1   0  0.6  4.5  3.9
   D  0   7.4  0  2.9  14   0  0.2  30  0.8  7.5  4.5
   G  0  18.3  3  3.1   8  60  0.3  60  0.2    7  1.9
   F  0   5.3  4  0.2  48  63  2.3  80  0.2    8  5.2
   C  0   2.6  5  2.2  24   4  1.3  20    0    2  1.4
   F  0  13.8  4  3.6   4   7  1.1  10  0.4  3.5  1.9
   B  1  12.4  6  1.7  30  44  1.1  60  0.2    6  1.5
   I  0   1.3  1  1.3   8  53  1.1  70  0.6    7  0.8
   F  0  18.2  7  1.7  26  92  2.2  30    1  8.5  4.8
   J  0   5.2  2  2.2  18  12  1.4  90  0.8    4  4.9
   G  1   9.4  2  0.8  22  86  0.4  30  0.4    1  5.9
   J  1  10.4  2  1.7  26  31  2.4  10  0.2    7  1.6
   J  0    13  1  1.8  14  11  2.3  50  0.6  5.5  2.6
   A  0  17.9  4  3.1  46  58  2.6  90  0.6  1.5  3.2
   D  1  19.4  6    3  20  50  2.8 100  0.2    9  1.2
   I  0  19.6  3  3.6  22  19  1.2   0  0.6    5  4.1
   I  1     6  2  1.5  30  30  2.2  20  0.4  8.5  5.3
   G  0  13.8  1  2.7   0  52  2.4  20  0.8    6    2
   B  0  14.3  4  2.9  30  11  0.6  90  0.6  0.5  4.9
   E  0  15.6  0  0.4  38  79  0.4  80  0.4    1  3.3
   D  0    14  2    1  22  61    3  90  0.6    2  0.1
   C  1   9.4  5  0.4  12  53  1.7  40    0    3  1.1
   H  0  13.2  1  1.6  40  15  0.7  40  0.2    9  5.5
   A  0  13.5  5  2.4  18  89  1.6  20  0.4  9.5  4.7
   E  0   2.6  4  2.3  38   6  0.8  20  0.4    5  5.3
   E  0  12.4  3  1.3  26   8  2.8  10  0.8    6  5.8
   D  0   7.6  2  0.9  44  89  1.3  50  0.8    6  0.4
   I  0  12.7  1  2.3  42   6  2.4  10  0.4    1    3
   C  1  10.7  4  3.2  28  23  2.2  90  0.8  5.5  2.8
   H  0  10.1  2  2.3  10  62  0.9  50  0.4  2.5  3.7
   C  1  16.6  1  0.5  12  88  0.1  20  0.6  5.5  1.8
   I  1   0.2  3  2.2   8  71  1.7  80  0.4  0.5  5.5
   C  0  10.8  4  3.5  30  70  2.3  60  0.4  4.5  5.9
   F  0   7.1  4    3  14  63  2.4  70    0    7  3.1
   D  0  16.5  1  3.3  30  80  1.6  40    0  3.5  2.7
   H  0  17.1  7  2.1  30  45  1.5  60  0.6  0.5  2.8
   D  0   4.3  1  1.5  24  44    0  70    0    5  0.5
   H  0    15  2  0.2  14  87  1.8  50    0  4.5  4.7
   G  0  19.7  3  1.9  36  99  1.5  10  0.6    3  1.7
   H  1   2.8  6  0.6  34  21    2  60    1    9  4.7
   G  0  16.6  3  3.3  46   1  1.4  70  0.6  1.5  5.3
   E  0  11.7  5  2.7  48   4  0.9  60  0.8  4.5  1.6
   F  0  15.6  3  0.2   4  79  0.5   0  0.8  1.5  2.9
   C  1   5.3  6  1.4   8  64    2  80  0.4    9  4.2
   B  1   8.1  7  1.7  40  36  1.4  60  0.6    6  3.9
   I  0  14.8  2  3.2   8  37  0.4  10    0  4.5    3
   D  0   7.4  4    3  12   3  0.6  60  0.6    7  0.7
   D  0   4.8  3  2.3  44  41  1.9  60  0.2    3  3.1
   A  0   4.5  0  0.2   4  48  1.7  80  0.8    9  4.2
   D  0   6.9  6  3.3  14  92  0.5  40  0.4  7.5    5
   B  0   4.7  4  0.9  14  99  2.4  80    1  0.5  0.7
   I  1   7.5  4  2.1  20  79  0.4  40  0.4  2.5  0.7
   C  0   6.1  0  1.4  38  18  2.3  60  0.8  4.5  0.7
   C  0  18.3  1    1  26  98  2.7  20    1  8.5  0.5
   F  0  16.4  7  1.2  32  94  2.9  40  0.4  5.5  2.1
   I  0   9.4  2  2.3  32  42  0.2  70  0.4  8.5  0.3
   F  1  17.9  4  1.3  32  42    2  40  0.2    1  5.4
   H  0  14.9  3  1.6  36  74  2.6  60  0.2    1  2.3
   C  0  12.7  0  2.6   0  88  1.1  80  0.8  0.5  2.1
   F  0   5.4  4  1.5   2   1  1.8  70  0.4  5.5  3.6
   J  1  12.1  4  1.8  20  59  1.3  60  0.4    3  3.8
;

The following statements fit a logistic model to these data by using a classification effect for variable C and 10 regressor effects for x1x10:

proc hplogistic data=getStarted;
   class C;
   model y = C x1-x10;
run;

The default output from this analysis is presented in Figure 54.1 through Figure 54.11.

The "Performance Information" table in Figure 54.1 shows that the procedure executes in single-machine mode—that is, the model is fit on the machine where the SAS session executes. This run of the HPLOGISTIC procedure was performed on a multicore machine with the same number of CPUs as there are threads; that is, one computational thread was spawned per CPU.

Figure 54.1: Performance Information

The HPLOGISTIC Procedure

Performance Information
Execution Mode Single-Machine
Number of Threads 4



Figure 54.2 displays the "Model Information" table. The HPLOGISTIC procedure uses a Newton-Raphson algorithm to model a binary distribution for the variable y with a logit link function. The CLASS variable C is parameterized using the GLM parameterization, which is the default.

Figure 54.2: Model Information

Model Information
Data Source WORK.GETSTARTED
Response Variable y
Class Parameterization GLM
Distribution Binary
Link Function Logit
Optimization Technique Newton-Raphson with Ridging



The CLASS variable C has 10 unique formatted levels, and these are displayed in the "Class Level Information" table in Figure 54.3.

Figure 54.3: Class Level Information

Class Level Information
Class Levels Values
C 10 A B C D E F G H I J



Figure 54.4 displays the "Number of Observations" table. All 100 observations in the data set are used in the analysis.

Figure 54.4: Number of Observations

Number of Observations Read 100
Number of Observations Used 100



The "Response Profile" table in Figure 54.5 is produced by default for binary and multinomial response variables. It shows the breakdown of the response variable levels by frequency. By default for binary data, the HPLOGISTIC procedure models the probability of the event with the lower-ordered value in the "Response Profile" table—this is indicated by the note that follows the table. In this example, the values represented by y = '0' are modeled as the "successes" in the Bernoulli experiments.

Figure 54.5: Response Profile

Response Profile
Ordered
Value
y Total
Frequency
1 0 69
2 1 31

You are modeling the probability that y='0'.




You can use the response-variable options in the MODEL statement to affect which value of the response variable is modeled.

Figure 54.6 displays the "Dimensions" table for this model. This table summarizes some important sizes of various model components. For example, it shows that there are 21 columns in the design matrix $\bX $, which correspond to one column for the intercept, 10 columns for the effect associated with the classification variable C, and one column each for the continuous variables x1x10. However, the rank of the crossproducts matrix is only 20. Because the classification variable C uses GLM parameterization and because the model contains an intercept, there is one singularity in the crossproducts matrix of the model. Consequently, only 20 parameters enter the optimization.

Figure 54.6: Dimensions in Binomial Logistic Regression

Dimensions
Columns in X 21
Number of Effects 12
Max Effect Columns 10
Rank of Cross-product Matrix 20
Parameters in Optimization 20



The "Iteration History" table is shown in Figure 54.7. The Newton-Raphson algorithm with ridging converged after four iterations, not counting the initial setup iteration.

Figure 54.7: Iteration History

Iteration History
Iteration Evaluations Objective
Function
Change Max Gradient
0 4 0.4493546916 . 0.410972
1 2 0.4436453992 0.00570929 0.081339
2 2 0.4435038109 0.00014159 0.003302
3 2 0.4435035933 0.00000022 5.623E-6
4 2 0.4435035933 0.00000000 1.59E-11



Figure 54.8 displays the final convergence status of the Newton-Raphson algorithm. The GCONV= relative convergence criterion is satisfied.

Figure 54.8: Convergence Status

Convergence criterion (GCONV=1E-8) satisfied.



The "Fit Statistics" table is shown in Figure 54.9. The –2 log likelihood at the converged estimates is 88.7007. You can use this value to compare the model to nested model alternatives by means of a likelihood-ratio test. To compare models that are not nested, information criteria such as AIC (Akaike’s information criterion), AICC (Akaike’s bias-corrected information criterion), and BIC (Schwarz’ Bayesian information criterion) are used. These criteria penalize the –2 log likelihood for the number of parameters. Because of the large number of parameters relative to the number of observations, the discrepancy between the –2 log likelihood and, say, AIC, is substantial in this case.

Figure 54.9: Fit Statistics

Fit Statistics
-2 Log Likelihood 88.7007
AIC (smaller is better) 128.70
AICC (smaller is better) 139.33
BIC (smaller is better) 180.80



Figure 54.10 shows the global test for the null hypothesis that all model effects jointly do not affect the probability of success of the binary response. The test is significant (p-value = 0.0135). One or more of the model effects thus significantly affects the probability of observing an event.

Figure 54.10: Null Test

Testing Global Null Hypothesis: BETA=0
Test Chi-Square DF Pr > ChiSq
Likelihood Ratio 35.1194 19 0.0135



However, a look at the "Parameter Estimates" table in Figure 54.11 shows that many parameters have fairly large p-values, indicating that one or more of the model effects might not be necessary.

Figure 54.11: Parameter Estimates

Parameter Estimates
Parameter Estimate Standard
Error
DF t Value Pr > |t|
Intercept 1.2101 1.7507 Infty 0.69 0.4894
C A 3.4341 1.6131 Infty 2.13 0.0333
C B 2.1638 1.4271 Infty 1.52 0.1295
C C 0.6552 1.0810 Infty 0.61 0.5445
C D 2.4945 1.1094 Infty 2.25 0.0245
C E 3.2449 1.4321 Infty 2.27 0.0235
C F 3.6054 1.3070 Infty 2.76 0.0058
C G 2.0841 1.1898 Infty 1.75 0.0798
C H 2.9368 1.2939 Infty 2.27 0.0232
C I 1.3785 1.0319 Infty 1.34 0.1816
C J 0 . . . .
x1 0.03218 0.05710 Infty 0.56 0.5730
x2 -0.3677 0.1538 Infty -2.39 0.0168
x3 0.3146 0.3574 Infty 0.88 0.3787
x4 -0.05196 0.02443 Infty -2.13 0.0334
x5 -0.00683 0.01056 Infty -0.65 0.5177
x6 0.2539 0.3785 Infty 0.67 0.5024
x7 -0.00723 0.01073 Infty -0.67 0.5004
x8 2.5370 0.9942 Infty 2.55 0.0107
x9 -0.1675 0.1068 Infty -1.57 0.1168
x10 -0.2222 0.1577 Infty -1.41 0.1590