How can I get out-of-sample predicted values?

Sometimes it is useful to get predicted values for cases that were not used in the regression analysis. There are two ways to do this in SPSS. Let’s use the hsb2 dataset and create some missing values in a variable. Specifically, we will set the first nine values in the variable write to be missing. Then we will use write as our outcome variable in an OLS regression analysis. Of course, the cases with missing values will not be used in the analysis, but we can still get the predicted values for those cases.

get file ='d:https://stats.idre.ucla.edu/wp-content/uploads/2016/02/hsb2.sav'.

sort cases by id.
if id lt 10 write = $sysmis.
list write read math
/cases=from 1 to 12.

   write      read      math 
 
      .       34.00     40.00 
      .       39.00     33.00 
      .       63.00     48.00 
      .       44.00     41.00 
      .       47.00     43.00 
      .       47.00     46.00 
      .       57.00     59.00 
      .       39.00     52.00 
      .       48.00     52.00 
    54.00     47.00     49.00 
    46.00     34.00     45.00 
    44.00     37.00     45.00 
 
Number of cases read:  12    Number of cases listed:  12

Method 1

When running the regression command, we can use the save subcommand to save the predicted values to the current data file. We have supplied the name for the new variable in parentheses after the SPSS keyword pred. After running the regression, we will list the first 12 cases in the data set for the variables write and pred_1.

regression
  /dependent write
  /method = enter read math
  /save pred(pred_1).

<output omitted>

list write pred_1
/cases from 1 to 12.

    write      pred_1
 
      .      42.24554
      .      40.81015
      .      54.03857
      .      45.58411
      .      47.28941
      .      48.53128
      .      56.83733
      .      48.67533
      .      51.30748
    54.00    49.77315
    46.00    44.31532
    44.00    45.19271
 
 Number of cases read:  12    Number of cases listed:  12

Method 2

Another way to get out-of-sample predictions is to save the model information to an .xml file, use the model handle command to name the .xml file, and then use the ApplyModel function of the compute command to create the predicted values. We will list the first 12 cases in the data file for the variables write and yhat.

regression
  /dependent  write
  /method = enter read math
  /outfile=model('d:/data/working/hsb_m1.xml').

<output omitted>

model handle name = m1 file='d:/data/working/hsb_m1.xml'.

compute yhat = ApplyModel(m1,'predict').

list write yhat
/cases from 1 to 12.
    write     yhat
 
      .      42.25
      .      40.81
      .      54.04
      .      45.58
      .      47.29
      .      48.53
      .      56.84
      .      48.68
      .      51.31
    54.00    49.77
    46.00    44.32
    44.00    45.19
 
 Number of cases read:  12    Number of cases listed:  12

Now let’s look at pred_1 and yhat side by side; as you can see, they are the same.

formats pred_1 yhat (f8.5).

list write pred_1 yhat
/cases from 1 to 12.

    write   pred_1     yhat
 
      .   42.24554 42.24554
      .   40.81015 40.81015
      .   54.03857 54.03857
      .   45.58411 45.58411
      .   47.28941 47.28941
      .   48.53128 48.53128
      .   56.83733 56.83733
      .   48.67533 48.67533
      .   51.30748 51.30748
    54.00 49.77315 49.77315
    46.00 44.31532 44.31532
    44.00 45.19271 45.19271
 
 Number of cases read:  12    Number of cases listed:  12