Analyzing Workers Compensation Reform Impacts on Loss Development Patterns in NCCI Rate-Making

1. Analyzing Workers Compensation Reform Impacts on Loss Development Patterns in NCCI Rate-Making Frank Schmid Director and Senior Economist

2. Outline • The Objective • The Statistical Framework • The Example of an Unidentified State • Conclusion Page 2 of 39

3. The Objective • Regulatory reforms may affect the pattern with which losses develop, starting with the initial (first) payment • Loss development patterns matter in ratemaking and reserving • Changing run-off patterns bear on the tail factor • The impact of reforms on the tail factor is difficult to gauge (because both numerator and denominator are affected) Page 3 of 39

4. The Statistical Framework • Loss development can be modeled as a time series problem • Once loss development is cast into a time series framework, the statistical technique of state-space modeling can be applied • State-space models are flexible (by allowing fortime-variation of parameters) and accommodating(to regulatory details) Page 4 of 39

5. The Statistical Framework • There are three dimensions of time in a loss triangle • Exposure time (exposure growth across accident or policy years) • Calendar time (calendar year effect) • Development time (run-off, that is, decline in incremental payments, net of the calendar year effect) Page 5 of 39

6. The Statistical Framework • The model is written in terms of (logarithmic) growth rates of incremental payments—these growth rates are allowed to be time-varying Page 6 of 39

7. The Statistical Framework • The model is Bayesian • A (posterior) parameter estimate is the result of a prior that is taken to the data • All prior distributions are conjugate, that is, they are from the same family as the posterior distribution • Expert priors are used for the calendar year effect—to be discussed below Page 7 of 39

8. The Statistical Framework • The model is estimated using the Metropolis-Hastings algorithm • The technique is also known as MCMC (Markov-chain Monte-Carlo simulation) • We use WinBUGS 1.4.2 and OpenBUGS 2.2.0 (the latter within the R package BRUGS) Page 8 of 39

9. The Statistical Framework • The model fits to the logarithm of incremental payments • Negative incremental payments are coded as missing values • In Bayesian models, missing values are treated as parameters that need to be estimated Page 9 of 39

10. The Statistical Framework • There is a stochastic add-up constraint in the model • This constraint ensures that for every development year, the sum of estimated incremental payments lines up with the observed cumulative payments • This technique, which is known as the cusum (cumulative sum) chart technique, is critical for interpolation when there are negative incremental payments Page 10 of 39

11. The Statistical Framework • The rate of exposure growth (eta) • eta normalizes to the level of exposure the incremental payments within a given accident/policy year • eta is modeled as a random walk • eta has no bearing on the tail factor Page 11 of 39

12. The Statistical Framework • The calendar year effect (kappa) • An expert prior is used for the calendar year effect • Rate of CPI Medical Care inflation (“M-CPI”) for medical claims • Average weekly wage (QCEW), CPI, or fixed rate for escalating indemnity claims, depending the regulatory stipulation • Zero for non-escalating indemnity claims Page 12 of 39

13. The Statistical Framework • The calendar year effect (kappa), cont’d. • The fraction of the incremental payment that goes to escalating indemnity claims is allowed to vary across development years • The model can handle up to two non-zero inflation rates (as demonstrated below) • The calendar year effect varies along the diagonal (as opposed to being constant on a given diagonal) Page 13 of 39

14. The Statistical Framework • The calendar year effect (kappa), cont’d. • The inflation rate pertinent to workers compensation (WC) claims is known up to a constant • WC Infl. Rate = kappa + constant + error term • For instance, if the WC-pertinent rate of medical inflation differs systematically to M-CPI inflation, then this difference (the “constant”) feeds into the run-off rate (delta) Page 14 of 39

15. The Statistical Framework • The calendar year effect (kappa), cont’d. • Because any systematic difference between theWC-pertinent rate of inflation and the official rate of inflation feeds into the run-off rate (delta), it is this official rate of inflation (e.g., the M-CPI) that is relevant when projecting payments into the future • It is known that rates of inflation are close to random walks, which implies that the best forecast for any future rate of inflation is the current rate Page 15 of 39

16. The Statistical Framework • The run-off rate (delta) • We assume a stationary rate of run-off for the unobserved development years • The projected rate of run-off merges with the rate of mortality (www.ssa.gov) in development year 60, unless the run-off is faster • No dynamic mortality model is used • According to a special report in the New England Journal of Medicine 352(11), pp.1138-1145, there is little ground for assuming continued gains in life expectancy Page 16 of 39

17. The Example of an Unidentified State • Regulatory reforms • 1982 • 1986 (minor; effect is modeled but not broken out) • 1990 • 1992 Page 17 of 39

18. The Example of an Unidentified State • The object is to model the effect of the 1990/92 reform cluster on the loss development pattern • Pre-reform: Policy years 1983 through 1989 • Post-reform: Policy years 1993-2004 Page 18 of 39

19. The Example of an Unidentified State • Major reform items • Introduction of escalation of indemnity benefits at the rate of the CPI (regardless of the date of injury) for PTD claims, effective May 1991 • Indemnity benefits for Fatal claims had been escalating at a fixed rate of 4 percent since June 1986 • The model accounts for the escalation of Fatal claims, but the effect of this reform is not broken out in the following analysis (as mentioned) Page 19 of 39

20. The Example of an Unidentified State • Major reform items, cont’d. • Limitation of TTD claims to 52 weeks • Tightening of standards for continued eligibility of indemnity benefits • For injuries past age 55, there is an immediate retirement offset; otherwise, there is a retirement offset starting five years prior to the official retirement age Page 20 of 39

21. The Example of an Unidentified State • Expected effects • Faster run-off due to faster closing of claims • Spill-over into medical, as these claims may now close faster as well • Escalation contributes to an increase of the tail, all else being equal • Retirement offset will (in part) be picked up by the run-off rate; in addition, we reduce the benefits for development years 36 and later by 50 percent Page 21 of 39

22. The Example of an Unidentified State • Pre-reform and post-reform “triangles” Page 22 of 39

23. The Example of an Unidentified State • The pre-reform triangle consists of pre-1990 policy years only • At the same time, this triangle includes elements from diagonals through calendar year 2004 • To the degree that the 1990/92 reform cluster affected existing claims (for instance by accelerating their closure), the model may underestimate the reform impact Page 23 of 39

24. The Example of an Unidentified State • The post-reform triangle consists of post-1992 diagonals only • Yet, only in the first column are all observations from the post-reform period • As development time progresses, the post-reform triangle phases in observations from the previous regulatory setting Page 24 of 39

25. The Example of an Unidentified State • Indemnity: delta (“9”: pre-reform; “8”: post-reform) Page 25 of 39

26. The Example of an Unidentified State • Medical: delta (“9”: pre-reform; “8”: post-reform) Page 26 of 39

27. The Example of an Unidentified State • Indemnity: Change in Ultimate Loss (per $1 of Initial Payment) Page 27 of 39

28. The Example of an Unidentified State • Medical: Change in Ultimate Loss (per $1 of Initial Payment) Page 28 of 39

29. The Example of an Unidentified State • Indemnity: Tail Factors by Regulatory Regime Page 29 of 39

30. The Example of an Unidentified State • Medical: Tail Factors by Regulatory Regime Page 30 of 39

31. The Example of an Unidentified State • Indemnity: Calendar Year Effect in First Column Page 31 of 39

32. The Example of an Unidentified State • Medical: Calendar Year Effect in First Column Page 32 of 39

33. The Example of an Unidentified State • Indemnity: Calendar Year Effect on Final Diagonal Page 33 of 39

34. The Example of an Unidentified State • Medical: Calendar Year Effect on Final Diagonal Page 34 of 39

35. The Example of an Unidentified State • Indemnity: Diagnostic, Pre-Reform Triangle Page 35 of 39

36. The Example of an Unidentified State • Indemnity: Diagnostic, Post-Reform Triangle Page 36 of 39

37. The Example of an Unidentified State • Medical: Diagnostic, Pre-Reform Triangle Page 37 of 39

38. The Example of an Unidentified State • Medical: Diagnostic, Post-Reform Triangle Page 38 of 39

39. Conclusion • NCCI has devised a loss development model that is capable of incorporating detailed statutory provisions • The model allows the estimation of tail factors according to the applicable regulatory setting • The model is capable of quantifying the impact of regulatory reforms on the ultimate loss and, hence, the tail factor Page 39 of 39