* QUESTION:
>> we have been modeling visitor numbers as linear functions of several
>> independent variables. We now want to validate this model using some
>> extra observations that have become available, without re-fitting the
>> whole model, i.e. we want to apply the GLM we have constructed to new
>> data to make predictions for these (and compare them with the observed
>> values, naturally).
>>
>> Doing this for regression is easy and fairly straightforward: you just
>> set the dependent variable to missing for the cases for which you want
>> to make predictions. This does not work for GLM. Neither does a
>> Select Cases. We are obviously not eager to recode everything for
>> regression, or to explicitly compute predictions from the parameters.
>> Does anyone have suggestions how to avoid this?
* Solution posted to usenet on 2001/02/26.
*This is a work around solution which automatically calculates the predicted values for the missing values of y.
* UNIANOVA: Calculate predicted values.
* Raynald Levesque rlevesque@videotron.ca
DATA LIST LIST /a(f2.0) b(f2.0) y(f2.0) .
BEGIN DATA.
4 7 3
6 4 2
7 8 5
9 15 7
5 4 .
END DATA.
UNIANOVA
y WITH a b
/METHOD = SSTYPE(3)
/INTERCEPT = INCLUDE
/SAVE = PRED RESID
/PRINT = PARAMETER
/CRITERIA = ALPHA(.05)
/DESIGN = a b
/OUTFILE=COVB('C:\\temp\\param.sav').
SAVE OUTFILE='c:\\temp\\mydata.sav'.
GET FILE='C:\\temp\\param.sav'.
SELECT IF (rowtype_="EST").
FORMATS p1 p2 p3 (F14.8).
WRITE OUTFILE="C:\\temp\\calc predicted.sps"
/"COMPUTE mypre=",p1," + a*",p2," + b*",p3,".".
EXECUTE.
GET FILE='c:\\temp\\mydata.sav'.
INCLUDE FILE="C:\\temp\\calc predicted.sps".
EXECUTE.