

FlashChat  Actuarial Discussion  Preliminary Exams  CAS/SOA Exams  Cyberchat  Around the World  Suggestions 
DW Simpson 
Actuarial Salary Surveys 
Actuarial Meeting Schedule 
Contact DW Simpson 
General Actuarial NonSpecific 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. 

Thread Tools  Search this Thread  Display Modes 
#1




GLM for mortality
The applied statistics module uses a poisson distribution with a loglink to model mortality.
Why would that be more appropriate than using logistic regression? Is there a good online (free) resource for guidance on GLM parameters by use case? By parameters  I am referring to distribution and link function. Thank you
__________________

#2




If by "model mortality" you mean "model expected time (from now) until death" . . . then logistic regression isn't the correct set up as it generally models binary results (e.g., will you die in the next x years where x is a fixed value).
I've found the following material (from CAS MASI exam) to be helpful: http://wwwbcf.usc.edu/~gareth/ISL/I...20Printing.pdf
__________________
I find your lack of faith disturbing Why should I worry about dying? It’s not going to happen in my lifetime! Freedom of speech is not a license to discourtesy #BLACKMATTERLIVES 
#3




Yea  I'm not looking at a classification model  I am looking to model a qx  which is between 0 and 1. Seems like a perfect candidate for logistic  but maybe the interpretation is difficult to model projections?
Anyway  thanks for the material  I will have a look...
__________________

#4




You can do Binomial distribution where X ~ Bi(N, p) and use the logistic regression to get "p". It's just with a high N and low p, it converges to a Poisson anyway. From Wikipedia:
Quote:
Last edited by AMedActuary; 06152018 at 02:57 PM.. 
#5




Also, if you want to run this as a logistic model, you would have to change your data from the number of people in one column and number of deaths in another to all 1's and 0's. So if you have 10000 people and 100 deaths and want the deaths to be zeros, then you need 100 0's and 9900 1's in your dataset instead with the predictors that you'll be using for each member.

#7




I've supervised interns doing lapse models with a GLM . . . logistic regression.
But the question was looking at what were some key indicators of new business not renewing with the company; so GLM was the route of choice. However, if you're just looking at a "final answer," you could use some additional "data science" type analysis as outlined in the reference I linked above.
__________________
I find your lack of faith disturbing Why should I worry about dying? It’s not going to happen in my lifetime! Freedom of speech is not a license to discourtesy #BLACKMATTERLIVES 
#8




I believe the primary advantage of using a Poisson model with a log link for mortality is that then the modeled mortality naturally decomposes as a product of factors, which is consistent with how mortality assumptions have traditionally been expressed (e.g. preferred mortality is x% of standard mortality).
It's theoretically unreasonable, of course  you're using a model that says that it's possible for a cell that contains N lives to experience more than N deaths. But practically speaking the mortality rates being considered are rather small, and in that case there isn't really that much difference between a Poisson distribution with mean Nq and a binomial distribution N events with probability q.
__________________
The Poisson distribution wasn't named after a fish  it was named after a man ... who was named after a fish. 
#9




Quote:
Is there a reason why you're limiting yourself to GLMs? I think there are many reasons why this occurs, but just wanted to see what your constraints are (regulatory, software, etc). I personally don't use GLMs very often for modeling mortality, churn, retention, etc. I'm guessing you mention quasiPoisson because there are overdispersion problems. In those cases, I think a decent fix is to switch over to a negative binomial distribution where you can set up differing means/variances. Another typical approach outside of GLMs is going to be a survival model, such as proportional hazards, which feels similar to a Poisson regression and is a little more computationally heavy, but should give you better results. State of the art stuff for "timetoevent" is something like this: https://ragulpr.github.io/2016/12/22...hurnmodeling/ But this requires more data at different times than what you're probably working with. Riley 
Thread Tools  Search this Thread 
Display Modes  

