Suppose, you have measured 4 clinical parameters (A, B, C, D) at the time of admission of 60 patients with the same disease. You observe the outcome as "severe disease" and "non-severe disease". Now you want to calculate the severity predictive values of:
1. Individual parameters: i] A, ii] B, iii] C, iv] D individually
2. Combinations of v] A+B, vi] A+C, vii] A+D, viii] B+C, ix] B+D, x] C+D
3. Combinations of xi] A+B+C, xii] A+B+D, xiii] A+C+D, xiv] B+C+D
4. and xv] A+B+C+D
How can one compare the results of these 15 combinations and tell which combination can give the highest possible specific, sensitive, positive predictive and negative predictive data ?? Please enlighten.
Thank you
You can perform logistic regression analyses, one separate model for every desired combination. The Akaike Information Criterion (AIC) at the end of the model output will tell you how well your model performs. The lower the AIC, the better the model. Hope this will help you.
Best, Bettina
Agree with Bettina - see the following paper for a concrete example:
Knaus, Draper, Wagner, and Zimmerman. APACHE II: A severity of disease classification system. Critical Care Medicine 1985; 13 (10): 818-29.
Good luck with your project.
Thank you very much Bettina and Kenneth. Really appreciate your support.
Stay healthy stay blessed
Just an add-on:
[Bettina wrote:] "The Akaike Information Criterion (AIC) at the end of the model output will tell you how well your model performs."
Yes, but this performance measure is useless for your purpose. You really need to determin your target statistics: Sepcificity, Sensitivity, PPV, NPV. A ROC analysis might be helpful.
Important: you need to split your data in a "training set" and a "test set". The former is used to estimate the cut-off values of your predictors used for calssifying the outcome. However, one should not calculate Sens, Spec, PPV and NPV using these estimates on the same data, because the estimates are tuned to just this data, and it is impossible to judge if or how well this would generalize to other data. Therefore you have the "test set" to get less biased values of target statistics. Ideally (but usually not achieveable) one would really use all the data to estimate the cut-offs and then perform an completely new study to estimate the target statistics.
- Badges
- Science topic
- Similar topics
- Medicine
- Physiology
- Biosignals
- Biomedical Signal Processing
- Statistical Analysis