I am using SPSS to perform binary logistic regression. One of the parameters generated is the prediction probability. Is there a simple mathematical formula that could be used to calculate it manually? e.g. based on the B values generated for each variable in model?
Hello Nasir Al-Allawi. I believe you are talking about the option to save the predicted probability as a new variable in the dataset. If so, here is an example that may help. Apart from a small amount of rounding error, the predicted probabilities I computed "by hand" match those saved by SPSS--see the attached image.
Cheers,
Bruce
* Modify the path to point to the folder where
* the sample datasets are stored on your computer.
GET FILE='C:\SPSSdata\survey_sample.sav'.
DATASET NAME raw.
COMPUTE male = sex EQ 1.
FORMATS male (F1).
CROSSTABS sex BY male.
LOGISTIC REGRESSION VARIABLES male
/METHOD=ENTER age educ
/SAVE=PRED
/PRINT=CI(95)
/CRITERIA=PIN(0.05) POUT(0.10) ITERATE(20) CUT(0.5).
* The coefficients are (approximately):
-.00585306 (age)
.01706050 (highest year of school)
-.22788545 (constant).
* Use the coefficients to compute the fitted log-odds for each record.
COMPUTE LogOdds = -.22788545 -.00585306*age + .01706050*educ.
* Convert Log-odds to odds.
COMPUTE Odds = EXP(LogOdds).
* Convert predicted odds to predicted probability.
COMPUTE PredProb = Odds/(1+Odds).
COMPUTE Diff = PRE_1 - PredProb.
FORMATS PredProb Diff (F11.5).
VARIABLE LABELS
PredProb "Predicted probability BY HAND".
DESCRIPTIVES PRE_1 PredProb Diff.
* Apart from some rounding error, the results match.
Thank you so much Bruce, that was exactly what I was looking for. Much appreciate ur prompt response.
Dear Bruce Weaver, thank you for your earlier answer .. I presume that we could create a scoring system based on the fitted log odds to predict outcome?
People have certainly done that, Nasir Al-Allawi. A Google search on turns up lots of resources. Good luck with your work.
- Badges
- Science method