I am working with a dataset where each row corresponds to a car trip, with less than 20 unique drivers and at least 500 trips for each. Each driver is labeled as either having or not having a certain medical condition. My task is to train a model to predict the presence of this condition in a given participant.
It seems logical to me that, since the outcome variable is driver-specific rather than trip-specific, I should aggregate the dataset by driver, reducing its size to fewer than 20 data points. However, not everyone I have discussed this with agrees.
To my understanding, correlation between data points can be addressed by mixed effects models in situations where the outcome variable varies within groups/clusters. If the 0/1 label weren’t the same across the trips of a specific driver (for example, indicating if a collision occurred during a given trip), then I think mixed-effects models could be used.
I also don’t know if its is sound to use a model like XGBoost and train it on individual trips, even if it is ensured that the trips of the same driver don’t end up in both the training set and the test set. Although if the number of drivers was significantly larger, I can see the model being able to learn some more general, non-driver specific trends, even given this nested structure.
This problem has been on my mind for a while now, and I am not able to properly explain why I think those methods can’t be used. I am starting to suspect I might be wrong. I would really appreciate some clarification.
