Actuarial Outpost
 
Go Back   Actuarial Outpost > Actuarial Discussion Forum > Property - Casualty / General Insurance
FlashChat Actuarial Discussion Preliminary Exams CAS/SOA Exams Cyberchat Around the World Suggestions

DW Simpson International Actuarial Jobs
Canada  Asia  Australia  Bermuda  Latin America  Europe


Reply
 
Thread Tools Search this Thread Display Modes
  #1  
Old 12-09-2019, 03:06 PM
The Obese Dog's Avatar
The Obese Dog The Obese Dog is offline
Member
 
Join Date: Jan 2008
Posts: 537
Default How do P&C actuaries test for bias and discrimination?

Hi all,

I am not a P&C actuary. Nevertheless, I am looking at P&C models and this question came up. I am too unfamiliar with the P&C regulatory environment to have an opinion of how a DOI would think about this issue. The model I am reviewing includes attributes from a credit rating agency, if that makes a difference.

How do I proof my model is not discriminatory towards protected classes? Any good resources exist out there that are helpful? Does the answer change if the model is not a GLM?
Reply With Quote
  #2  
Old 12-09-2019, 03:18 PM
bjc2142's Avatar
bjc2142 bjc2142 is offline
Member
CAS
 
Join Date: Aug 2014
Location: NYC
Posts: 1,215
Default

I thought it was more about the state law. Whether they allow it or disallow it. I may be wrong though, as I have not worked on filing in personal space.
__________________
FCAS
Reply With Quote
  #3  
Old 12-09-2019, 03:26 PM
NormalDan's Avatar
NormalDan NormalDan is offline
Member
CAS
 
Join Date: Dec 2016
Location: NJ
Posts: 10,882
Default

Quote:
Originally Posted by The Obese Dog View Post
Hi all,

I am not a P&C actuary. Nevertheless, I am looking at P&C models and this question came up. I am too unfamiliar with the P&C regulatory environment to have an opinion of how a DOI would think about this issue. The model I am reviewing includes attributes from a credit rating agency, if that makes a difference.

How do I proof my model is not discriminatory towards protected classes? Any good resources exist out there that are helpful? Does the answer change if the model is not a GLM?
IME it's more about what variables you can or can't include, which can vary by state, rather than specifically testing if certain variables correlate closely with a protected class.
Reply With Quote
  #4  
Old 12-09-2019, 04:00 PM
Vorian Atreides's Avatar
Vorian Atreides Vorian Atreides is offline
Wiki/Note Contributor
CAS
 
Join Date: Apr 2005
Location: As far as 3 cups of sugar will take you
Studying for CSPA
College: Hard Knocks
Favorite beer: Most German dark lagers
Posts: 67,613
Default

One way you could do it (may not be all that practical):

Have a database that includes information on the protected status of risks along with the required information to rate the risks.

Apply the model to this database. Then conduct analyses using the protected statuses to see if there is any biases and/or (illegal) discrimination.
__________________
I find your lack of faith disturbing

Why should I worry about dying? Itís not going to happen in my lifetime!


Freedom of speech is not a license to discourtesy

#BLACKMATTERLIVES
Reply With Quote
  #5  
Old 12-09-2019, 08:33 PM
Maphisto's Sidekick's Avatar
Maphisto's Sidekick Maphisto's Sidekick is offline
Member
CAS
 
Join Date: Nov 2001
Location: South Park Genetics Lab
College: Ardnox
Favorite beer: The kind with alcohol
Posts: 3,257
Default

Quote:
Originally Posted by The Obese Dog View Post
How do I proof my model is not discriminatory towards protected classes? Any good resources exist out there that are helpful? Does the answer change if the model is not a GLM?
(Disclaimer: this is written with a US P&C focus; things might be less fraught elsewhere.)

I faced that question almost 20 years ago.

It's the kind of question that you want to run by corporate legal, as discovering that there is inappropriate bias in existing rating/underwriting variables, or collecting/assembling data that could be used to discriminate against protected classes could lead to litigation and reputational risk, even if you actually do nothing wrong.

If you could get over those hurdles, you would want to study the concept of "disparate impact", which would guide how you designed a test.

Maybe things have changed since my exposure to the topic...but lawyers being lawyers, and consumer advocates being consumer advocates, I would be extremely skeptical that the changes have been for the better.
Reply With Quote
  #6  
Old 12-10-2019, 08:49 AM
The Obese Dog's Avatar
The Obese Dog The Obese Dog is offline
Member
 
Join Date: Jan 2008
Posts: 537
Default

Quote:
Originally Posted by Vorian Atreides View Post
One way you could do it (may not be all that practical):

Have a database that includes information on the protected status of risks along with the required information to rate the risks.

Apply the model to this database. Then conduct analyses using the protected statuses to see if there is any biases and/or (illegal) discrimination.
I'm looking for guidance on what a bias even would be considered. What are the things that a regulator would look at and think are problematic? There are all sort of socioeconomic correlations with most all types of individually-generated data such that I'm sure I can find all sorts of statistically significant differences based on the approach you suggested (which is what I was considering as well) but its not clear to me what sort of differences matter.

For instance, if class X has a lower rate than class Y on average because of differences in variables A, B, C is that discriminatory? Or is that cool if A, B, C have no direct connection to X and Y. For instance, if average credit ratings vary by race (and they do) and my model says that a better credit rating is a better risk, my average premiums will also vary by race as well on average. Does that matter? Based on what I am inferring so far, probably not.

Last edited by The Obese Dog; 12-10-2019 at 08:53 AM..
Reply With Quote
  #7  
Old 12-10-2019, 09:36 AM
Vorian Atreides's Avatar
Vorian Atreides Vorian Atreides is offline
Wiki/Note Contributor
CAS
 
Join Date: Apr 2005
Location: As far as 3 cups of sugar will take you
Studying for CSPA
College: Hard Knocks
Favorite beer: Most German dark lagers
Posts: 67,613
Default

I would imagine that the first level of analysis is at the premium level . . . if one class is "obviously*" different from others, then look at if there's an objective reason for that difference (that class has lower limits or smaller AOI than the other classes).

Distributional analysis among key rating variables might also be looked at (e.g., is the distribution of AOI for class A "close" to that for other classes).

If class C has a very large percentage of risks receiving the highest discount for home alert systems (compared to other classes), again, digging in deeper to see if there would be an objective explanation for that.


Absent a compelling, objective explanation, these results could be used as evidence for bias/(illegal) discrimination.

Bottom line, I don't think you're going to find a "simple," straight-forward test to derive a conclusion of bias/discrimination.


*"Obviously" here is to be defined as, "you know it when you see it".
__________________
I find your lack of faith disturbing

Why should I worry about dying? Itís not going to happen in my lifetime!


Freedom of speech is not a license to discourtesy

#BLACKMATTERLIVES
Reply With Quote
  #8  
Old 12-10-2019, 09:41 AM
PAC PAC is offline
Member
CAS AAA
 
Join Date: Jan 2003
Posts: 263
Default

Quote:
Originally Posted by The Obese Dog View Post
I'm looking for guidance on what a bias even would be considered. What are the things that a regulator would look at and think are problematic?
These are questions for your regulator(s) to answer.
Reply With Quote
  #9  
Old 12-10-2019, 09:16 PM
Maphisto's Sidekick's Avatar
Maphisto's Sidekick Maphisto's Sidekick is offline
Member
CAS
 
Join Date: Nov 2001
Location: South Park Genetics Lab
College: Ardnox
Favorite beer: The kind with alcohol
Posts: 3,257
Default

Quote:
Originally Posted by The Obese Dog View Post
I'm looking for guidance on what a bias even would be considered. What are the things that a regulator would look at and think are problematic?
Note: what follows is written from a US viewpoint. Things can be quite different in other countries.

The concept you want to look into is "disparate impact". It's a key concept that has emerged from lawsuits arising from accusations of illegal discrimination in various industries.

The term is not as cleanly defined in practice as an actuary might like, which is why I encourage you to research the concept.

If you found, for example, that drivers of purple cars had loss costs double the average, but it just so happened that drivers of purple cars were almost exclusively members of a protected class, the use of the purple car factor could be considered illegal because of its disparate impact on the group.

Note the word "could". You might be able to argue that the factor is the least-impactful way to account for a perfectly kosher attribute. Or you could perhaps claim ignorance as a defense - you don't collect data on protected group membership, and therefore are unable to illegally discriminate. But a consumer advocate will likely focus on the correlation when arguing against you.

In a sense "disparate impact" is a little like porn - difficult to precisely and objectively define...but folks are pretty sure they know it when they see it.
Reply With Quote
  #10  
Old 12-11-2019, 09:13 AM
The Obese Dog's Avatar
The Obese Dog The Obese Dog is offline
Member
 
Join Date: Jan 2008
Posts: 537
Default

Thanks, this is helpful.

It sounds like the type of guidance I was hoping for does not exist.
Reply With Quote
Reply

Thread Tools Search this Thread
Search this Thread:

Advanced Search
Display Modes

Posting Rules
You may not post new threads
You may not post replies
You may not post attachments
You may not edit your posts

BB code is On
Smilies are On
[IMG] code is On
HTML code is Off


All times are GMT -4. The time now is 05:54 PM.


Powered by vBulletin®
Copyright ©2000 - 2020, Jelsoft Enterprises Ltd.
*PLEASE NOTE: Posts are not checked for accuracy, and do not
represent the views of the Actuarial Outpost or its sponsors.
Page generated in 0.52084 seconds with 11 queries