SOA - Factor Affecting Retirement Mortality (FARM)

Abstract | Bibliography | Table of Summaries

This paper was commissioned by the Society of Actuaries Social Security committee.

Mortality assumptions are a basic input to valuing life insurance, annuities, pension plans and even Social Security. Generally, the mortality rates used differ by only age and gender. When are further refinements appropriate? Which of the factors that can influence mortality experience are appropriate to include? How can these factors be fairly and systematically reflected in the mortality assumptions used in the valuation of benefits paid to retirees? How might these factors be expected to influence future improvements in mortality? This paper is the starting point for answering these questions and others related to risk classification of retiree mortality.

One of the more interesting applications of the work presented herein relates to the proposals to reform many Social Security systems into mandatory individual accounts. One of the problems with such individual accounts is the ability to annuitize the account values at retirement at "fair market" rates. At the moment, there is very little to no underwriting in the individual annuity market. If a person volunteers to buy an individual annuity, it is assumed that that person is healthy–in fact, very healthy. Except for structured settlements (income for people severely impaired mostly because of auto accidents) there exists almost a single mortality assumption for individual annuities in Canada and the United States.

If more countries move to mandatory individual accounts to provide retirement income, or if annuitization of retirement savings is made mandatory (these are separate possibilities), then there will be a wider spectrum of mortality among the prospective annuitants than exists today. This, in turn, will create the need for more risk classification in the annuitant market. This is already happening in the UK where pensioners who have used tax–advantaged systems to save for retirement must buy some form of life annuity by age 75. In response, there is now a growing "impaired annuity" market in the UK. Using UK mortality and an interest rate of 5 percent per annum, the following table shows the increase in annuity income justified by various increases in the mortality rate, q x.

	Increase in Annuity
Increase in qx	Age 65	Age 75	Age 85
+25%	+ 7%	+11%	+15%
+50	+13	+21	+30
+100	+24	+40	+58

"Impaired annuities" met a market that had become disappointed with ever– decreasing annuity income per dollar of premium because of decreasing interest rates combined with decreasing mortality rates (i.e. increasing life expectancy).

"Impaired annuities" have found demand in two separate market niches. It has enhanced incomes for those who retire with profiles of high mortality (e.g. those in poor health). It has also worked to provide annuities to fund nursing home care and other long term care costs for ill or frail elderly people who are near the end of their lives.

In a competitive annuity market, impaired annuities represent an opportunity for some insurers to carve out a profitable, sizeable niche at the expense of their competitors. New players in the market can do this without threatening the turnover of their existing book of business.

This is especially timely as the massive baby–boom generation is just now starting to think about the issue of retirement income security and the potential annuitization of retirement savings.

In the UK, risk classification in annuity pricing has created new rates for classes that include: smokers, those with medical impairments, (diabetics, high blood pressure, high cholesterol, stroke or heart attack victims), the overweight, and, more recently, manual workers living in geographic areas displaying higher than average mortality.

It is the hope of the authors of this paper that broader risk classification will come to North American annuity markets. If that were to occur, more retirees could buy life annuities at a fair market value. This is just not possible today and is a hindrance to the expansion of this important market. The review of the factors included in this paper will assist in the future creation of a wider variety of risk classes.

In addition to refinements in annuity pricing, there is also the issue of proper evaluation of liabilities held by defined benefit pension plans, whether privately or publicly sponsored. For example, is it appropriate for all plans to use the same mortality assumption? What variation might we expect from plan to plan, reflecting differences in location, industry, jobs covered, etc.? Poor estimates of mortality, and therefore liabilties, could risk plan solvency or, conversely, unduly constrict plan sponsor and contributing participant assets.

The authors have reviewed 45 papers addressing 12 different risk factors that influence mortality after retirement. The factors reviewed include: age, gender, race and ethnicity, education, income, occupation, marital status, religion, health behaviors, smoking, alcohol, and obesity. For each factor, the pertinent results from the 45 papers reviewed are gathered and presented. The 45 papers have also been summarized. The summaries of the research papers and the summaries of the different risk factors are available to interested readers in the Table of Summaries.

As is virtually always the case, this research is presented with one very large caveat. Our summary review of factors affecting retirement mortality focuses on 12 risk factors as previously listed. This list could be expanded to include dozens of other factors – both proximate to and remote from the time of death. For example, Hurd et al. (1999) evaluated the impact of 13 health indicators that are strong predictors of mortality.This raises two questions: (1) What should one do about the potential lack of independence between the risk factors that are included in an actuarial model? (2) What should one do about important risk factors that are not included in an actuarial model? From an actuarial perspective, these questions relate to the construction of an initial risk classification system and any subsequent refinement of that system.

(1) Our review indicates that all 12 risk factors are important enough to be included in an actuarial model to the extent that such data are available for risk classification. Practically, one might expect to have information on age and gender, and possibly one or two others, but not on all 12 risk factors. This means that one needs to be careful to ensure that the numerical values of parameters used for select subsets of the 12 risk factors are consistent between the estimation model (in the published paper) and the actuarial model. For example, if the estimation model jointly included age, gender, marital status, education, income, and occupation while the actuarial model included age and occupation, the parameter values from the estimation model would likely not be directly applicable to the actuarial model. An important exception would occur if the risk effects of age and occupation were independent of the risk effects of gender, marital status, education and income.

(2) As a general rule, the larger the number of risk factors used in an actuarial model, the less significant are the effects of any omitted variables, and the greater the likelihood that the assumption of independence of the omitted variables would be a reasonable approximation to reality. If an important risk factor is missing from an actuarial model, then the most direct response is to re–estimate the published estimation model with that risk factor omitted (assuming that the risk factor was originally included in the published model). This would yield parameter estimates for the actuarial model that were consistent with the retained risk factors, even though it was recognized that the parameter estimates were somewhat biased. This would be satisfactory if the distribution of the omitted factor in the estimation sample was the same (or approximately the same) as in the insured population. An important exception would occur if the omitted risk factor were subject to significant antiselection (which would be more likely for life insurance than retirement annuity models).

Because of the complex issues involved in estimating numerical values for parameters of actuarial models, including choice and measurement of risk factors, assumptions about independence/dependence and antiselection, and other related issues, we focused our review on the qualitative impact of the 12 main risk factors, on the consistency of the evidence supporting conclusions about that impact, and on specific interactions between the risk factors that have been identified and documented to date.

Research