In this question, we assume we have a health dataset with many triplets of dummy variables. The dataset looks like this:
(existence_of_symptomA (1/0), symptomA_chronic (1/0), symptomA_persistent (1/0), existence_of_symptomB (1/0), symptomB_chronic (1/0), symptomB_persistent (1/0).......)
Each line represents a patient, and, the data are dummy because multiple symptoms may coexist per patient.
The outcome of interest is a dummy variable "hospital death" (1/0).
If you take a look at the data structure, you will notice that semantically the "existence_of_symptom" variables are the main ones, while the "symptom_chronic" and "symptom_persistent" describe characteristics of the "main" dummy variable.
If one wants to study the odds for death solely based on the existence of symptoms (just the existence_of_symptom variables) this would be a multiple binary logistic regression problem. This would create a model with the odds for death, for each symptom.
Here is the question: What would be the best approach to study the predictive contribution of the two extra "symptom_chronic" and "symptom_persistent" dummy variables per symptom? Would you simply add everything together into the list of IVs to run the logistic regression?
Wouldn't this approach be incorrect?
To begin with, everyone without a symptom will always have values of 0 to the chronic and to the persistent variables as well! Also how will the model recognize and "account for" the fact that data should be seen as triplets?
Any insights?
your model and variable definitions make no sense to me. Take a look at these refs: Using R and RStudio for Data Management, Statistical Analysis and Graphics | Nicholas J. Horton, Ken Kleinman | download (b-ok.cc) and R for everyone: advanced analytics and graphics | Lander, Jared P | download (b-ok.cc) Try to define your variables this way. Then use R. These books should tell you what to do with your variables after you get them defined and write a research question. Another valuable source might be : Statistical Analysis with R For Dummies | Joseph Schmuller | download (b-ok.cc) and R For Dummies (uam.mx)
Now write a clear research question and if you still have questions come back here. Best, David Booth
Hello David,
Thank you for your input. In the above research problem, that has been solved a while ago, the dummy variables have to be treated as binary. This is because each one of the disease dimensions (chronicity/persistence) needs to be represented as a separate variable in the dataset. Apparently, the model coefficients are required in order to understand the "contribution" that each of those dimensions have on the outcome under study (dependent variable).
My question was primarily a logical/philosophical one, rather than a statistical:
If a variable contains information which is a dimension of another variable [e.g: "Type of diabetes (1/2)" is a dimension of "Diabetes (y/n)"], these two variables need to be included to the model because we need to know (1) the contribution of the presence of diabetes and ALSO (2) the contribution of different diabetes types, to, (for instance) the development of heart disease.
- Badges
- Science topic