The context is fire danger rating: from the operational viewpoint it's important not to miss the days when events (fires) actually occur.
Hi Paulo,
would it be an option to use a dual approach. First a logistic regression on the probability that a fire occurs and then a regular regression on e.g. the intensity of the firs,
Cheers Jetse
Hello Paulo
Your predictor is the probability of having a fire and you decide on the best threshold probability to label as a 'yes'. You can choose any value, which is simply a trade off between false negatives for false positives. I think statistics like Area Under the Curve give you the optimal value to minimise the total number of falses, but you can choose to be more conservative to reduce the false negatives and increase the false positives, thereby reducing your overall accuracy. I dont know of a method to optimise this choice.
Good luck
Owen
To expand on Owen's suggestion, utilizing sensitivity and specificity for individual values of your outcome might be a good approach to take. Where sensitivity and specificity are closest to each other in their value would represent the optimal threshold. This would represent the best balance between the occurence of false positives and false negatives. Alternatively, you may choose a threshold that has higher sensitivity at the expense of decreasing specificity and vice versa.
- Badges
- Science topic
- Similar topics
- Engineering
- Modeling