Actuarial Outpost
 
Go Back   Actuarial Outpost > Exams - Please Limit Discussion to Exam-Related Topics > SoA/CAS Preliminary Exams > Exam PA: Predictive Analytics
FlashChat Actuarial Discussion Preliminary Exams CAS/SOA Exams Cyberchat Around the World Suggestions


Upload your resume securely at https://www.dwsimpson.com
to be contacted when new jobs meet your skills and objectives.


Reply
 
Thread Tools Search this Thread Display Modes
  #401  
Old Yesterday, 07:29 PM
JRemy511 JRemy511 is offline
Member
SOA
 
Join Date: Oct 2011
College: University of Rochester
Posts: 58
Default

Quote:
Originally Posted by windows7forever View Post
In the June exam, log transformation was observed in Task 1 after viewed the target variable's right skewness. However it did not get used again until Task 4 to identify potential interaction feature to add to GLM model. After that, we use GLM model with log link function to ensure positive predictions only, so we abandon the use of log transformation on the target variable anymore.

I do not see the log transformation on the target variable is useful in that exam. Especially the modeling tasks use log link on the target mean that has nothing to do with log transformation on the target variable. If you do log transformation on the target variable, you need to make sure the target variable has all values greater than 1. Otherwise you can get negative predictions possibly.

If we do GLM, we do not need to log transformation on the target variable at all no matter it's skewed or not. We do log transformation on the target variable when we handle standard linear model like OLS.
I think the usage of the log transformation in the beginning was to create a more intuitive visual appearance of the graphs looking for interactions. With the initial value, all of the means were scrunched down at the bottom with many high outliers. Once the transformation happened the interactions between target and predictors was easier (albeit none of them really stand out) to detect.

The GLM used the Gamma family and log link to essentially do the same thing as log transforming but with the added restrictions of always producing a positive value. If you log transformed the variable you could you a regular gaussian GLM with the identity link but that model could produce negative values (not allowed by the target variable).
__________________
P, FM, MFE, MLC, C
Reply With Quote
  #402  
Old Yesterday, 07:38 PM
Gemii Gemii is offline
SOA
 
Join Date: Nov 2019
Posts: 7
Default

What is the difference between offset and weight? How do we interpret the coefficients resulting from a GLM with offset or weight?
Reply With Quote
  #403  
Old Yesterday, 07:51 PM
JRemy511 JRemy511 is offline
Member
SOA
 
Join Date: Oct 2011
College: University of Rochester
Posts: 58
Default

Quote:
Originally Posted by Gemii View Post
What is the difference between offset and weight? How do we interpret the coefficients resulting from a GLM with offset or weight?
From my understanding (anyone feel free to correct this), and offset deals with the data going into the model, and a weight deals with the final interpretation of the model.

For an example

Company A had 500 workers and 100 injuries
Company B had 5 workers and 5 injuries

If you were predicting the amount of injuries a company has per year, an offset could be applied so that the model takes into account the number of workers the company has when predicting the rate of injuries. In this example company A has 1 in 5 employees being injured, Company B has 1 in 1 employees being injured. The offset makes sure the model does not treat these as equal.

The weight is about applying a singular predictor to a group of individuals with varying degrees of experience.

For example a model predicts that a person has a 10% chance of going to the hospital within a given year.

Person A was enrolled for 6 months
Person B was enrolled for 12 months

The weight given to these two individuals is .5 for Person A and 1 for person B. So looking at Person A there's only a 5% chance they went to the hospital, while Person B remains at 10% chance they went.

Again, someone please correct me if I'm wrong.
__________________
P, FM, MFE, MLC, C
Reply With Quote
  #404  
Old Yesterday, 07:59 PM
Gemii Gemii is offline
SOA
 
Join Date: Nov 2019
Posts: 7
Default

Quote:
Originally Posted by JRemy511 View Post
From my understanding (anyone feel free to correct this), and offset deals with the data going into the model, and a weight deals with the final interpretation of the model.

For an example

Company A had 500 workers and 100 injuries
Company B had 5 workers and 5 injuries

If you were predicting the amount of injuries a company has per year, an offset could be applied so that the model takes into account the number of workers the company has when predicting the rate of injuries. In this example company A has 1 in 5 employees being injured, Company B has 1 in 1 employees being injured. The offset makes sure the model does not treat these as equal.

The weight is about applying a singular predictor to a group of individuals with varying degrees of experience.

For example a model predicts that a person has a 10% chance of going to the hospital within a given year.

Person A was enrolled for 6 months
Person B was enrolled for 12 months

The weight given to these two individuals is .5 for Person A and 1 for person B. So looking at Person A there's only a 5% chance they went to the hospital, while Person B remains at 10% chance they went.

Again, someone please correct me if I'm wrong.
Gotcha, thanks! Do you know when we would add offset as an additional predictor (module 6) and when we would just specify it as a parameter (December 2018 Exam)?
Reply With Quote
  #405  
Old Yesterday, 08:21 PM
JRemy511 JRemy511 is offline
Member
SOA
 
Join Date: Oct 2011
College: University of Rochester
Posts: 58
Default

Quote:
Originally Posted by Gemii View Post
Gotcha, thanks! Do you know when we would add offset as an additional predictor (module 6) and when we would just specify it as a parameter (December 2018 Exam)?
I'm not exactly certain, but looking at the code in the two examples you provided I'm not certain that either produces a different result. The GLM in the December exam defines an offset but I'm not versed enough in R coding to know if it has the same result as the GLM in module 6.
__________________
P, FM, MFE, MLC, C
Reply With Quote
  #406  
Old Yesterday, 10:14 PM
DukeSilver DukeSilver is offline
SOA
 
Join Date: Dec 2018
Posts: 12
Default When to binarize?

This might be a dumb question, but I'm wondering why they binarize for fitting the glm in the Hospital Readmissions but not June 2019's PA. Is it because we have too many factor variables in the June 2019 exam so it'd be way too much? Thanks to anyone who knows and can explain
Reply With Quote
  #407  
Old Yesterday, 10:24 PM
Awktuary916 Awktuary916 is offline
CAS SOA
 
Join Date: Apr 2017
Posts: 2
Default

Does anyone know if we can do split screen during the exam? Like Word on the left and RStudio on the right?
Reply With Quote
  #408  
Old Yesterday, 10:38 PM
JRemy511 JRemy511 is offline
Member
SOA
 
Join Date: Oct 2011
College: University of Rochester
Posts: 58
Default

Quote:
Originally Posted by DukeSilver View Post
This might be a dumb question, but I'm wondering why they binarize for fitting the glm in the Hospital Readmissions but not June 2019's PA. Is it because we have too many factor variables in the June 2019 exam so it'd be way too much? Thanks to anyone who knows and can explain
I think there are reasons to choose to do it in both cases or not. It could be the number of variables that would result is rather large, but if you've condensed the groups many of the variables are binary already and wouldn't need to be binarized. Another is that the variables that have many factors are not signifiant and the person doing to model did not consider it important to split out into individual factors.

Another reason could be the June exam was going to involve lasso/ridge regularization and that automatically does the binarization and would indicate any variables that might require that sort of split. This suggests the test creators thinking about later tasks while doing model work which isn't the best real world thinking.
__________________
P, FM, MFE, MLC, C
Reply With Quote
  #409  
Old Yesterday, 11:03 PM
Niccalis Niccalis is offline
SOA
 
Join Date: Dec 2012
Posts: 23
Default

We don't need the confirmation email from Prometric to get in do we? Just our ID?
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 02:15 AM.


Powered by vBulletin®
Copyright ©2000 - 2019, 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 0.28420 seconds with 10 queries