Agenda Day Two

Friday, October 5
6:45 a.m. – 7:30 a.m.

Session Coordinator(s)

Facilitator(s)

7:30 a.m. – 9:00 a.m.

Presenter(s): Ian Duncan, FSA, CSPA, FCA, FCIA, FIA, MAAA ; Nhan Huynh, MS

Survival models tend to be studied for exam purposes but are not used much by health care actuaries. Yet they are powerful techniques with many applications. In this case study we will apply survival model techniques to a problem. 

To legally self-insure in California, employers must comply with the reserving policy of the CA Office of self-insured plans. The reserving methodology for future medical benefits is: average of last three years of payments x life expectancy, where life expectancy is determined according to the 2011 Unites States Life Tables. Our hypothesis is that the United States Life Tables over-state disabled work comp claimant survival. We will:

  • Test this hypothesis using a dataset;
  • Develop appropriate models that better fit the empirical data; and
  • Develop semi-parametric models that allow us to predict the survival of individual claimants based on covariates.

Workers Comp. survival data; R code. 

Session Coordinator(s)

Facilitator(s)

9:00 a.m. – 9:30 a.m.

Session Coordinator(s)

Facilitator(s)

9:30 a.m. – 11:00 a.m.

Presenter(s): Sreenivas (Si) Konda, Ph.D

Time Series: Brief introduction, Trends, Seasonality, Forecasting, Model building. Example: Google flu time series data to forecast flu outbreaks.

R code

Relevant articles (including problems with Google’s first model)

Flu database

Per capita cost data

Session Coordinator(s)

Facilitator(s)

11:00 a.m. – 12:00 p.m.

Session Coordinator(s)

Facilitator(s)

12:00 p.m. – 2:30 p.m.

Presenter(s): Sreenivas (Si) Konda, Ph.D

  • Use of big data in health care - exercise levels are measured and recorded by a device (Fitbit wristbands). The device records the number and intensity of workouts, classified as either light or standard.  We will develop models to predict improvements in clinical measures depending on demographic factors, initial health status and number/intensity of workouts.
  • Machine learning application to high cost prediction.
  •  
  • Machine learning
    • Variable selection methods: Random Forest, Bagging, Boosting, LASSO, Elasticnet (as time permits);
    • Review of multinomial logistic regression, Decision Tree, Dimension reduction: Clustering , PCA

BMI Exercise data; Medical cost data.  R code

Session Coordinator(s)

Facilitator(s)

2:30 p.m. – 2:30 p.m.

Session Coordinator(s)

Facilitator(s)