Actuarial Outpost GLMM & Credibility
 Register Blogs Wiki FAQ Calendar Search Today's Posts Mark Forums Read
 FlashChat Actuarial Discussion Preliminary Exams CAS/SOA Exams Cyberchat Around the World Suggestions

Search Actuarial Jobs by State @ DWSimpson.com:
AL AK AR AZ CA CO CT DE FL GA HI ID IL IN IA KS KY LA
ME MD MA MI MN MS MO MT NE NH NJ NM NY NV NC ND
OH OK OR PA RI SC SD TN TX UT VT VA WA WV WI WY

 General Actuarial Non-Specific Actuarial Topics - Before posting a thread, please browse over our other sections to see if there is a better fit, such as Careers - Employment, Actuarial Science Universities Forum or any of our other 100+ forums.

#1
09-22-2019, 09:52 PM
 noonelikesmodules Member SOA Join Date: Jul 2017 Posts: 48
GLMM & Credibility

Hi All,

Apologies if this is the wrong section.

So I am a L&H actuary and I have recently introduced GLMs into our mortality assumption setting process. As part of the implementation process I had to fit these models into the existing limited fluctuation credibility framework we use. Using LFC + factors from a GLM is fine enough for the time being. Backtesting helped my approvers gain enough comfort to let this fly.

Naturally I did some research and ended up coming across the relationship between GLMMs and Buhlmann credibility. There was a section in ASM manual for exam C that covered roughly this I think, that was a long time ago. I also read some CAS papers by Klinker. Those were particularly helpful in my understanding.

My understanding of GLMMs is that the variable modeled as a random effect will have its level's factors shrunk back to the mean. For certain variables in my mortality models this makes a lot of sense but there is a catch with some other variables. Let me sketch out my example below.

Let's say my model is specified as such. Since I am using an expected count offset I am really modeling A/E ratios. Poission family and log link. Also let's assume my aggregate A/E is 100% to simply further.

ActualCount ~ RiskClass + IssueAge + offset(log(ExpectedCount))

So the catch arises from the fact I have an expected basis (call it VBT15). I model IssueAge as a random effect and because my underlying expected mortality basis varies by IssueAge I am content with factors being shrunk to 100%. (ok so here's actually my question) Let's say I used some relative risk tool and came up with an expected A/E by risk class; How would / could / should I adjust my model to account for the the relative risk tool and get my GLMM risk class factors to shrink to the risk class factors recommended my relative risk tool? Is this even statistically sound? Is this too oddly specific to be easily answered? So essentially of shrinking towards the mean, shrink towards predefined values?

Does this make sense?

The ugly work around I see is to model IssueAge as random effect, let GLMM shrink those parameters then model Risk Class as a fixed effect and independently credibility weight my GLMM fixed effect risk class factors to the relative risk tool's predictions. Unsure of how statistically unsound this is but I bet if I backtest this work around it would work well. Aiming to do better though.

Thanks for the help y'all.
#2
09-22-2019, 10:02 PM
 The_Polymath Member CAS SOA Join Date: Jun 2016 Posts: 2,658

I am at a loss as to why you would try to use GLMs to model mortality risk.
__________________
Quote:
 Consistency is the hobgoblin of little minds
#3
09-23-2019, 07:25 AM
 noonelikesmodules Member SOA Join Date: Jul 2017 Posts: 48

Quote:
 Originally Posted by The_Polymath I am at a loss as to why you would try to use GLMs to model mortality risk.

We are setting a mortality assumption for a life product. We have emerging experience. We are using the GLM to model experience. We aren’t directly modeling qx’s, we are modeling A/Es. We are looking to adjust our expected basis for emerging experience.

To be specific a coworker saw higher A/Es on permanent products than term. Turned out to be an unequal weighting of Super Preferred and Standard risks between the two.
#4
09-24-2019, 09:05 AM
 Underpaid Member SOA Join Date: Oct 2016 Posts: 816

I can't understand your post. A lot of times people can leave out background information and I can deduce what their model is doing/how it works, but in this case, I can't. If you can provide a little more background on what it is you are modeling, you might get better answers and someone might chime in with a succinct reason showing that your results are spurious.
#5
09-24-2019, 08:03 PM
 glassjaws Member Non-Actuary Join Date: Mar 2008 Location: 99 Franklin St. Studying for C4 College: Chambana Favorite beer: Whatever you buy me Posts: 46,176

Quote:
 Originally Posted by The_Polymath I am at a loss as to why you would try to use GLMs to model mortality risk.
It's a frequency model. I don't think it's particularly egregious.
__________________
I am the background noise of life
#6
09-24-2019, 08:05 PM
 examsarehard Member CAS Join Date: May 2011 Posts: 598

If you want to use predefined risk relativities as your complement of credibility, you're better off using a Bayesian glm. This way, instead of shrinking towards the grand mean, you shrink towards the relativities given by your risk tool.

As far as I am aware, there aren't any papers that discuss applying Bayesian GLMs in insurance contexts, but there are plenty of references on the model outside of insurance.