Set 2013 q9 (ii) Aliasing

[r_v.s]

Would you pls explain "if there are no claims for the 17 exposures, and a claims frequency model is built using a log link, we could have large and opposite-signed parameters for Sidious and Unknown number of bedrooms (or other similar example)"? What are the other possible similar examples??

 

[Darren Michaels]

This all really comes down to how the GLM will be fitted.

The factor for Sidious is really only being used to differentiate it between the other 2 brokers (Maul and Vader) in the case there are 4 bedrooms. Otherwise you really don’t need the parameter at all.

This is because otherwise you would have extrinsic aliasing as all properties with unknown bedrooms would be written through Sidious and vice-versa. If the rogue data were not present, you could have dealt with these properties with just the unknown bedrooms parameter only (or equivalently the Sidious one).

In the case of 4 bedrooms, we know in the example given the policies written by Sidious have no claims, so you need to reduce the estimated claims cost relative to Maul and Vader properties. Hence the idea of a (large) negative parameter coming in. However the experience of the Sidious unknown bedrooms properties will be more in line with the others, so you will need a (large) positive parameter to get the overall estimated claims cost back to normal.

Having no claims for the 17 exposures was just one possible example. There are many other alternative examples that could have been suggested that would have gained credit in the exam.