I am having a difficult time modeling a binary dataset that my team has gathered. It is a dataset of different prescriptions per each patient and an outcome of a certain event.
The goal is to determine if certain prescriptions lead to decreased rates of alerts.
However, there are two main problems:
The data is fully binary without much overlap or interaction between columns
There is a lot of noise in the data, and the chance of the alert occurring is due to chance.
I have run the data through different algorithms to determine any trends or factors that may be important, but have not been very successful.
I was thinking of doing a simple bayesian analysis to determine if each setting has an impact on the outcome of the alerts, but would love to be able to involve many features if possible.
Is there anything I am missing or that I could try to determine the influence of different treatments?
Which kind of Bayesian analysis are you planning to do?
Is the outcome of alerts an explicit response variable related to the prescription? If so, then you may try to fit an ordinary logistic regression as an attempt to explain the behavior of such a dichotomous response using the remaining columns as explanatory variables (covariates). In this context, you can evaluate the possible effect of treatments by using several effect estimates, as the odds ratio, the risk difference, the risk ratio, and the hazard ratio.
Good luck!
Hello Christian,
Yes, sparse matrices are difficult to work with (and that's one of the potential drawbacks associated with retrospective studies). As well, noisy data makes model-building a bigger challenge.
You might consider classification trees as an option. The larger your data set, the better the chance for any derived model to be checked against hold-out data (cross-validation). Any "mining" technique will necessarily be opportunistic, so it's important to rule out over-fitting (and cross-validation helps in this regard).
Good luck with your work.
Hi Christian, my suggestion would be to model this with graphs (Neo4J, for example), which are quite useful for such associations. You can try out as much associations as you like to see what produce the best results. These can be structured just as graph queries (rule based) or you can try algorithmic methods as needed to determine the associations. Just a thought, hope that helps. Cheers.
Thanks for responding, all.
David, I have tried an assortment of models from boosted/bagged trees to simple logistic regression, the model choice seems to have little effect on the outcome.
Wesley, I appreciate the input for steps after analysis. That is exactly what I am looking to do because, for this project, the prediction is not the main concern, but the actionable insight to be gained from separating the classes. The only problem I have right now is the data is not showing much separation between classes, it is either too noisy, or I need more information. There are many cases with a patient having differing levels of alerts with the same treatment information (at least what I have now) therefore the model is attempting to choose without any relevant information.
Suresh, I am already boxed into a SQL database and I am unfamiliar with what you are talking about, but I appreciate the feedback.
- Badges
- Science topic
- Similar topics
- Engineering
- Modeling