Actuarial Outpost
 
Go Back   Actuarial Outpost > Actuarial Discussion Forum > Health - Disability
FlashChat Actuarial Discussion Preliminary Exams CAS/SOA Exams Cyberchat Around the World Suggestions


Not looking for a job? Tell us about your ideal job,
and we'll only contact you when it opens up.
https://www.dwsimpson.com/register


Reply
 
Thread Tools Search this Thread Display Modes
  #31  
Old 11-01-2019, 01:19 PM
Pamela Wells Pamela Wells is offline
Member
SOA AAA
 
Join Date: May 2019
Location: Phoenix, Arizona
Posts: 114
Default

Quote:
Originally Posted by WhosOnFirst View Post
The unfortunate thing about this specific model is that the moment I saw that it used cost as an estimate of morbidity, I knew it was a bad model. It is obvious that the developers made no connection between the data they were using and the people it represented. They saw that there was a feature that was significant and they used it with no thought at all about why it was important and what the downstream impacts on actual people would be. Either that, or they didn't care. I don't know if ignorance or apathy is worse but the end result was the same.
I speculate ignorance. If this model was developed by data scientists with no actuarial background, then there could be many factors that haven't been considered. Aside from just the racial treatment disparity, I would expect that it will have a gender treatment disparity too, although likely smaller than the disparity by race. There's also a good chance that there's selection bias skewing it as well - I doubt they adjusted for benefit richness, and thus would have missed the impact of induced demand and point-of-service price elasticity. That elasticity would be unlikely to surface when looking at large claims... but it is likely to be a hidden variable embedded in diagnosis rates, and early identification of conditions.
Reply With Quote
  #32  
Old 11-01-2019, 01:24 PM
Pamela Wells Pamela Wells is offline
Member
SOA AAA
 
Join Date: May 2019
Location: Phoenix, Arizona
Posts: 114
Default

Quote:
Originally Posted by The Obese Dog View Post
However, the model isn't really "biased" in one (of many) senses of the word. It's predictions are probably pretty good. It was built to predict cost and probably does a great job.
It wasn't built to predict cost. It was built to recommend patients for more intense treatment and intervention. It doesn't do a great job at that - it reinforces and exaggerates an existing treatment bias.

In the underlying data, black patients get less treatment than white patients, and this have overall lower costs. The model uses cost to predict which patients are the sickest - and thus, which patients should get the most attention and be prioritized for treatment. Because white patients are likely to have higher costs for the same condition at the same underlying risk... the model prioritizes white patients for medical attention and intervention above black patients.

It's both treatment bias reflected in the underlying data skew and model bias because the model is recommending action in a disparate fashion that exacerbates the treatment bias.
Reply With Quote
  #33  
Old 11-01-2019, 01:30 PM
The_Polymath The_Polymath is offline
Member
CAS SOA
 
Join Date: Jun 2016
Posts: 3,202
Default

Quote:
Originally Posted by Pamela Wells View Post
It wasn't built to predict cost. It was built to recommend patients for more intense treatment and intervention. It doesn't do a great job at that - it reinforces and exaggerates an existing treatment bias.

In the underlying data, black patients get less treatment than white patients, and this have overall lower costs. The model uses cost to predict which patients are the sickest - and thus, which patients should get the most attention and be prioritized for treatment. Because white patients are likely to have higher costs for the same condition at the same underlying risk... the model prioritizes white patients for medical attention and intervention above black patients.

It's both treatment bias reflected in the underlying data skew and model bias because the model is recommending action in a disparate fashion that exacerbates the treatment bias.
The model is not sophisticated enough to give accurate predictive action. Not enough training data.
Reply With Quote
  #34  
Old 11-01-2019, 01:41 PM
Pamela Wells Pamela Wells is offline
Member
SOA AAA
 
Join Date: May 2019
Location: Phoenix, Arizona
Posts: 114
Default

Quote:
Originally Posted by WhosOnFirst View Post
This is where I get frustrated with data scientists. Just because the data justifies doing something, that doesn't mean it is the right thing to do.
I would like to echo this. I have great appreciation and respect for data scientists, and the potential game for ML and AI is enormous. I previously employed two data scientists in a prior job, and they were amazingly valuable to my team. Leveraging the power of data science is something I embrace.

My observation is that many data scientists understand how to do modeling. They often don't understand the underlying data, the external factors that affect that data, the implicit behaviors and biases that skew that data, or the real-world impact of their model results.

A really good and useful model needs all of those things. A mis-step in any of those can at a minimum create bias. Depending on the type of error, it could make the model completely useless... or even result in inappropriate results.
Reply With Quote
  #35  
Old 11-01-2019, 01:45 PM
Pamela Wells Pamela Wells is offline
Member
SOA AAA
 
Join Date: May 2019
Location: Phoenix, Arizona
Posts: 114
Default

Quote:
Originally Posted by The_Polymath View Post
The model is not sophisticated enough to give accurate predictive action. Not enough training data.
In this case, the size of training data is irrelevant. Even if they'd had the entire population of US patients as their source, the model still would have produced biased results.

I don't recall seeing any thing in the articles that addressed the size of the data set on which the model was built. on what are you basing your assumption that it didn't have enough training data?
Reply With Quote
  #36  
Old 11-01-2019, 04:57 PM
The Obese Dog's Avatar
The Obese Dog The Obese Dog is offline
Member
 
Join Date: Jan 2008
Posts: 471
Default

Quote:
Originally Posted by WhosOnFirst View Post
In your example, you reach out to all 60 people the model identified because they are potential candidates. It isn't a proportional thing that creates adverse impact. If you have a type 2 error rate of 1% for class A and a type 2 error rate of 20% for class B, guess what? You have bias. If the difference between class A and class B is a protected class, you may have legal issues depending upon the type of intervention you are doing.
I don't think this is true. They are getting slammed precisely because they reached out to different proportions of people, not the error rate. Hence, reaching out to all 60 would be biased. Also, trying to equalize proportions on accuracy would guarantee differing rates of of reach out and vice-versa (unless the underlying base rates are the same, which they aren't).

Quote:
As far as getting to parity is concerned, you need a different model. Specifically, you need a model that doesn't create adverse impact.
I am specifically questioning whether this is possible. If there are true underlying differences how different segments utilize healthcare (black people are more likely to have missing coded conditions because they get less care), I am saying there DOES NOT exist a dataset that isn't biased and there can't be. Even if you gather race data and force parity on whatever statistic you decree represents bias, this won't go away, because you still won't be as effective across subsets of getting care to the people who need it most (your true goal). In other words, I am saying I believe the proposed "fix" is still biased...maybe much less so than the original model but biased nonetheless (at least to the extent I understand the technical details). Obviously, there are still better and worse answers.

Quote:
However, the cost of that gain comes at the expense of a protected class. That is clearly illegal hence why NY is taking action.
It is not clear that the gain comes at the expense of a protected class, depending in what sense you mean it (I mean in the sense of harming them from what would have happened in absence of a model). It could just be codifying what was already happening without a model. It could be that this model has in fact helped them, just not as much as it could have without the bias and so they proportionally suffered while gaining absolutely. I realize I'm being a bit absurd here by my point is that it's a much easier task to ask models not create bias. It's a much harder task to ask them to undo the bias that already exists throughout the system (although I firmly believe they are the best tool we have for the task used appropriately).

I am not disputing that including allowed costs in a model for the use they intended it for was a bad idea or that they did a "good job". I am trying to point out that creating truly unbiased models is MUCH HARDER than being implied in this thread.
Reply With Quote
  #37  
Old 11-01-2019, 04:59 PM
The Obese Dog's Avatar
The Obese Dog The Obese Dog is offline
Member
 
Join Date: Jan 2008
Posts: 471
Default

Quote:
Originally Posted by The_Polymath View Post
The model is not sophisticated enough to give accurate predictive action. Not enough training data.
This is wrong on multiple levels.
Reply With Quote
  #38  
Old 11-01-2019, 05:00 PM
The Obese Dog's Avatar
The Obese Dog The Obese Dog is offline
Member
 
Join Date: Jan 2008
Posts: 471
Default

Quote:
Originally Posted by Pamela Wells View Post
It wasn't built to predict cost. It was built to recommend patients for more intense treatment and intervention. It doesn't do a great job at that - it reinforces and exaggerates an existing treatment bias.

In the underlying data, black patients get less treatment than white patients, and this have overall lower costs. The model uses cost to predict which patients are the sickest - and thus, which patients should get the most attention and be prioritized for treatment. Because white patients are likely to have higher costs for the same condition at the same underlying risk... the model prioritizes white patients for medical attention and intervention above black patients.

It's both treatment bias reflected in the underlying data skew and model bias because the model is recommending action in a disparate fashion that exacerbates the treatment bias.
I don't disagree with any of that and neither does my post (as it existed in my head). I should have just used a different word choice.

By "built to predict cost", my understanding was that literally the target outcome was allowed cost. Maybe I am misunderstanding and prior year cost was a feature but the target variable was something else.

Last edited by The Obese Dog; 11-01-2019 at 06:02 PM..
Reply With Quote
  #39  
Old 11-01-2019, 06:37 PM
Pamela Wells Pamela Wells is offline
Member
SOA AAA
 
Join Date: May 2019
Location: Phoenix, Arizona
Posts: 114
Default

Quote:
Originally Posted by The Obese Dog View Post
I don't disagree with any of that and neither does my post (as it existed in my head). I should have just used a different word choice.

By "built to predict cost", my understanding was that literally the target outcome was allowed cost. Maybe I am misunderstanding and prior year cost was a feature but the target variable was something else.
From the descriptions given, the target variable was "need". The model was designed to identify those patients with the greatest need for medical attention. They used cost as a variable, because they assumed that cost was a good proxy to need. They failed to account for a known treatment bias, wherein black patients receive less service than white patients.

As a result, their model produces outcomes that treat black patients as if they have lower need and prioritizes white patients.

Their core assumption was false - cost is not a good proxy for need.
Cost has a known bias that reflects racially disparate treatment approaches.
The mode exacerbates that bias and has outcomes that suggest that black patients have lower need.
The model suggests more medical intervention and higher prioritization of white patients over black patients with the same conditions and severity.

There is bias in the underlying data, and because of a bad assumption, the model produces biased results as well.
Reply With Quote
  #40  
Old 11-01-2019, 07:31 PM
The Obese Dog's Avatar
The Obese Dog The Obese Dog is offline
Member
 
Join Date: Jan 2008
Posts: 471
Default

Quote:
Originally Posted by Pamela Wells View Post
From the descriptions given, the target variable was "need".
When I say the target variable I mean the literal thing they put into the number and said "predict this", not what someone had in their head as what they were trying to actually achieve. I doubt that was "need". At least I can't think of any way to represent that as a numerical vector. Whether or not it was allowed cost, it was almost undoubtedly a proxy. We never observe and don't know "need" - either before or after the fact. We only have better and worse proxies.

I'll grant that removing the connection to allowed cost may make for a less biased model - I don't think I've contested that point. What I am protesting is that doing so means will no longer have bias. It is easy to say models shouldn't have bias, but in reality it is very hard to keep them from creeping in, possibly intractable. If you collect race data and modify the model to achieve parity on race, you might achieve parity across a visible, measurable dimension but still not true parity. That doesn't mean we shouldn't try but it is a point worth appreciating, and we shouldn't feel that our models are beyond such influences. In fact, I haven't seen an actuarial model yet that incorporates race. We just completely ignore it (like UHC did) and hope that's okay. That makes sense where we legally have to but surely there are models out there where that's not true.

Last edited by The Obese Dog; 11-01-2019 at 07:39 PM..
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 04:06 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 0.16412 seconds with 11 queries