Dear Fellow Researchers,
We want to access the prediction accuracy for various parameters. We were able to draw ROC curves for each of the parameter separately using SPSS.
Now, we want to see a cumulative prediction accuracy for the multiple parameters. However, some of our parameters are correlated, e.g., age and BMI. How to draw ROC curve using SPSS or other software/method?
Hello Malgorzata,
How are the multiple measures to be combined? For example, if you have a logistic model, then there is a definable linear combination (as regards estimated log-odds of a case being in the target category). If you are using boolean logic to combine (e.g., (score of > 55 on age AND BMI > 30) OR (A1C > 6)), then that has to be identified and applied to the cases.
I would think that if you're trying to apply thresholds derived from the ROCs of different variables/measures on some outcome, then you'd have to: (a) have a data set with scores on all of the measures (and outcome) of interest; (b) try various boolean combinations (as in classification/decision trees) to identify best combinations (and resultant estimates of collective sensitivity/specificity);
OR
run a logistic model with all of the variables/measures and evaluate directly what the "best" linear combination of the measures has to offer with respect to accuracy of classification;
OR
run a discriminant analysis model with all of the variables (and outcome group/condition as the DV), if you're comfortable with the usual parametric assumptions.
Any one of these methods would control for redundancy/overlap/correlation among the measures.
In either case, validation (on additional data) would be important, as there is always the danger of overfitting.
Good luck with your work.
Kindly check https://www.ncbi.nlm.nih.gov/pmc/articles/PMC3755824/
- Badges
- Science topic
- Similar topics
- Mathematical Sciences
- Statistics
- Statistical Software
- SPSS