Estimating sales volume of new drugs to a therapeutic class

Question

I work with pharmacy claims data for an insurance company and wish to estimate how many members on existing products in a therapeutic class will shift to the new product in the same class i.e. insulin once the drug has been covered by the insurance company.

New entrants/products are very common but each one has their own nuances, some are blockbusters while others are flops, some drugs have many indications/uses while others are very specific.

I have access to lots of demographic, historical utilisation, diagnosis code data and formulary information that I could incorporate into a model. I would be keen to hear people’s thoughts and suggestions about how one would tackle such a problem.

Answer 1

The first approach is a supervised machine learning model. If you have historical data, you can label members who transfer to the new drug when a new drug came out as label one, otherwise label zero. Then you have a supervised machine learning setting. Just use your other information as features. In the end, you count how many members are predicted to take the new drug.

The second approach is an unsupervised machine learning model. Using clustering to classifies members into two groups. You then count the group whose members will take new drugs. The problem of this method is that you need to have couple samples's label determined so that you can know which group is the group you need.

The above two methods all require that you have historical data.

I am thinking if there are other methods that can give an expected value and don't require training on the historical data. I will update once I figure out one.