Making sense of . . . Traditional survival models: aggregate, select, or select & ultimate Excerpt #2b from Jim Daniel’s LTAM seminars James W. Daniel Jim Daniel’s Actuarial Seminars www.actuarialseminars.com August 7, 2020 c Copyright 2020 by James W. Daniel; reproduction in whole or in part without the express permission of the author is forbidden. Foreword This document briefly describes aggregate, select, or select & ultimate survival models. Essentially it’s part of one of the lessons I gave in my face-to-face LTAM exam-prep seminars. NOTE: The numbered items—Examples, Equations, Definitions, Sections, et cetera—can be reached using a modern pdf reader or browser by clicking on the number in the reference. For example, if some text mentions Equation 1.4, if you click on the 1.4 you should be taken directly to that equation; some readers and browsers will return you to where you were if you click on the Back arrow. 2 Chapter 1 Traditional survival models: aggregate, select, or select & ultimate 1.1 Aggregate models versus select models As you’ll see in detail below, an aggregate model for future-lifetime random variables allows you to deduce the probability distribution of Tx for any age x from knowledge of the probability distribution of T0 , the future-lifetime random variable of a newborn. On the other hand, with a select model the probability distribution of the future-lifetime random variable at one age cannot in general be deduced from the probability distribution of the future-lifetime random variable at a different age. A select & ultimate model is a special type of select model that is a sort of blend of a select and an aggregate model. In an aggregate model, a life aged x is usually denoted by (x) and its future-lifetime random variable by Tx ; in a select model or select & ultimate model, a life aged x is usually denoted by [x] and its future-lifetime random variable by T[x] . 1.1.1 Aggregate models Suppose that the probability distribution of the future-lifetime random variable T0 of a newborn life (0) in an aggregate model is known. In particular, suppose that its survival function S0 , for which S0 (t) = Pr[T0 > t], is known. I want to deduce the probability distribution of Tx for any x > 0. The idea in an aggregate model is that when you consider an (x), a life aged x, absolutely the only thing you know is that (x) was once a typical ordinary newborn (0) who has survived to age x—you have no additional information, such as that (x) has just been diagnosed with a serious disease or that (x) just passed a rigorous physical with flying colors. At first glance it might then seem as though saying that (x) survives at least t years is equivalent to saying that (0) survives x + t years—so it might seem as though t px , the probability that (x) survives at least t years, should equal S0 (x + t), the probability that a newborn survives at least x + t years. But this thought ignores the fact that you know that the newborn has already fought off Mother Nature and managed to survive the first x years. In fact, the number of years (x) lives is the number of years the newborn lives minus the age x, conditioned on knowing that the newborn has survived x years: Tx = T0 − x T0 > x. 3 From this, you can find t px : t px = Pr[Tx > t] = Pr[T0 − x > t T0 > x] = Pr[T0 > x + t T0 > x] Pr[T0 > x + t and T0 > x] Pr[T0 > x + t] = Pr[T0 > x] Pr[T0 > x] S0 (x + t) = . S0 (x) = This shows that the probability distribution of Tx is determined by that of T0 in an aggregate model. It’s also easy to find the force of mortality µx (t). Intuitively, this measures how hard Mother Nature is trying to kill off the x-year-old after t years, which seems as though that should be the same as how hard she is trying to kill off a newborn after x + t years—which is just µ0 (x + t), also denoted by µ(x + t) or µx+t . This is in fact correct: d S0 (x+t) d − dt − dt t px S0 (x) = S (x+t) µx (t) = 0 p t x S0 (x) S 0 (x + t) =− 0 = µ0 (x + t) = µ(x + t) = µx+t . S0 (x + t) Since in general t px = e− Rt 0 µx (z) dz , it follows that for an aggregate model t px = e− Rt 0 µ(x+z) dz . In summary: KEY ⇒ Fact 1.1 (Aggregate models) Suppose that µ0 (t) = µ(t) = µt is the force of mortality for a newborn after t years, and that S0 is the survival function for a newborn. Then, in an aggregate model, Rt S0 (x + t) = e− 0 µ(x+z) dz t px = S0 (x) and µx (t) = µ0 (x + t) = µ(x + t) = µx+t . KEY ⇒ Exams! ⇒ Example 1.2 A commonly tested aggregate survival model is the Exponential model or forever constant force model. In this model, the force of mortality for a newborn is some positive constant µ, which by Key Fact 1.1 makes t px = e−µt for all ages x. This means that Tx is an Exponential random variable with mean µ1 , for all ages x. This is an unrealistic model for human lives, but an easy model for computations on exams. End of Example 1.2 1.1.2 Select models In a select model, a life aged x is usually denoted by [x] to distinguish it from a life aged x in an aggregate model, where the notation is (x). The intuitive idea is that you have some special information about [x], such as that [x] was just diagnosed with some chronic illness that makes the survival probabilities for [x] quite different from those of a normal x-year-old in an aggregate model. Thus the formulas in Key Fact 1.1 do not hold. Note that the force of mortality µ[x] (t) is sometimes denoted by µ[x]+t . 4 Example 1.3 Suppose that for each age x ≥ 0 the force of mortality for a select life aged x is constant and is given by µ[x] (t) = 0.01(x + 1). Then µ[9] (t) = 0.1 and µ[0] (9 + t) = 0.01 are quite different from one another, unlike for an aggregate model in which µ9 (t) = µ0 (9 + t). End of Example 1.3 1.2 Select & ultimate models In the previous Subsection on select models, Subsection 1.1.2, I mentioned that with select models you have some special information about the select life [x] that makes the survival probabilities for [x] quite different from those for a normal x-year-old in an aggregate model. In some cases—such as when what makes [x] special is that [x] just passed a demanding physical exam including a thorough medical history—the impact of the special information may only last a few years; just because you were in great shape at 25 and had lower mortality probabilities than the average 25-year-old for the next few years doesn’t mean that that will still be true in 10 or 15 years. In such a case, a select & ultimate model might be appropriate, with special one-year survival probabilities (the select probabilities) for a few years, after which the one-year survival probabilities (the ultimate probabilities) are the same as for a typical person of that age. This means, in particular, that the force of mortality will be special for [x] for the first few years during the so-called select period. The length of that select period—the number of years with special survival and death probabilities—is usually denoted by d. One way to define a select & ultimate model—an approach not commonly seen on Exam LTAM— is to provide one formula for the force of mortality during the d-year select period and another after that period. That is, µ[x] (t) would be given by a special formula for 0 ≤ t ≤ d but then would equal µ(x + t) for t > d, where µ(z) is the force of mortality for a typical newborn. Then the future lifetime of the life [x] + d would have the same distribution as that of an ordinary (x + d)-year-old. Example 1.4 Suppose that the force of mortality for a normal life is forever constant at µ = 0.006, so that µ(z) = 0.006 at all ages z. Suppose that a select & ultimate model has a two-year select period, and that µ[x] (t) = 0.006(3 − t) for t ≤ 2 while µ[x] (t) = µ(x + t) = 0.006 for t > 2. Then t p[x] = e− Rt 0 µ[x] (z) dz = e− Rt 0 0.006(3−z) dz 2 = e−0.018t+0.003t for 0 ≤ t ≤ 2, while t p[x] = 2 p[x] t−2 p[x]+2 = 2 p[x] t−2 px+2 = e−0.024 e−0.006(t−2) for t > 2. End of Example 1.4 On Exam LTAM, it’s more common for a problem to describe a select & ultimate model by using either a mortality table or a life table. 1.2.1 Select & ultimate mortality tables One of the common ways questions on Exam LTAM describe select & ultimate models is to provide a mortality table—a table of one-year death probabilities for a select life [x] at each age x. With a d-year select period, the death probabilities q[x]+j are special for the d years with 0 ≤ j ≤ d − 1 and then revert to ordinary probabilities after d years, so that q[x]+j = qx+j for j ≥ d. 5 Example 1.5 Here’s an example of a part of a select & ultimate mortality table with a two-year select period. x 90 91 92 q[x] 0.20 0.30 0.40 q[x]+1 0.25 0.35 0.45 qx+2 0.21 0.32 0.45 x+2 92 93 94 Notice that to find the sequence of mortality or survival probabilities that apply to a select life [x], you move across the row corresponding to the age x and then down the column of ultimate (or ordinary) probabilities. For instance, 4 p[90] = (1 − 0.2) × (1 − 0.25) × (1 − 0.21) × (1 − 0.32) = 0.32232. For more complicated probabilities such as 1|3 q[91] , it’s usually simpler to work with the life table corresponding to a mortality table, as indicated in the following Subsection 1.2.2. End of Example 1.5 1.2.2 Select & ultimate life tables It’s often easier to compute probabilities intuitively using the life table that corresponds to the mortality table provided in an Exam LTAM question; if you’re unfamiliar with computing probabilities intuitively using life tables, I suggest you take a look at my LTAM Seminar Excerpt Traditional survival models: intuitive calculations using life tables before continuing here. I’ll illustrate constructing a corresponding select & ultimate life table by using the select & ultimate mortality table above in Example 1.5. As usual in a life table, I can choose the value of one ` arbitrarily; I’ll choose `[90] = 10000. Using the survival probabilities starting with [90] deduced from the mortality table in Example 1.5 produces part of the corresponding life table: x 90 91 92 93 `[x] 10000 — — — `[x]+1 8000 — — — `x+2 6000 4740 3223.2 1772.76 x+2 92 93 94 95 Now I can’t choose an arbitrary value for `[91] , because after two years it has to lead to `[91]+2 = `91+2 = `93 that has already been computed to be `93 = 4740. To correspond to the mortality table in Example 1.5, I first need to choose `[91]+1 so that p[91]+1 = 1 − q[91]+1 = 1 − 0.35 = `[91]+2 4740 = , `[91]+1 `[91]+1 which gives `[91]+1 = 7292.31 approximately. Then I need to choose `[91] so that p[91] = 1 − q[91] = 1 − 0.30 = `[91]+1 7292.31 = , `[91] `[91] which gives `[91] = 10417.58 approximately. Inserting these values into the life table gives x 90 91 92 93 `[x] 10000 10417.58 — — `[x]+1 8000 7292.31 — — `x+2 6000 4740 3223.2 1772.76 x+2 92 93 94 95 6 Similarly, I can’t choose an arbitrary value for `[92] , because after two years it has to lead to `[92]+2 = `92+2 = `94 that has already been computed to be `94 = 3223.2. Instead, I have to make my way backwards from right to left just as I did for the line starting at age 91. That same process first produces `[92]+1 = 5860.36 approximately and then `[92] = 9767.27 approximately. The complete select & ultimate life table corresponding to the select & ultimate mortality table in Example 1.5 is thus x 90 91 92 93 `[x] 10000 10417.58 9767.27 — `[x]+1 8000 7292.31 5860.36 — `x+2 6000 4740 3223.2 1772.76 x+2 92 93 . 94 95 Complicated probabilities are easy to compute intuitively in the usual manner with life tables. Example 1.6 Suppose that an Exam LTAM problem gave me the select & ultimate mortality table given in Example 1.5 above and asked me to compute 1|3 q[91] . My approach would be to first create the part of the select & ultimate life table applying to [91] as found above and then compute intuitively. Although 1|3 q[91] is of course the probability that [91] survives one year and then dies in the following three years, I think of it intuitively as the fraction of select 91-year-olds that die between ages 92 and 95. The number of select 91-year-olds that die between ages 92 and 95 is (expected to be) `[91]+1 − `[91]+4 = `[91]+1 − `95 = 7292.31 − 1772.76 = 5519.55. The number of select 91-year-olds is `[91] = 10417.58, so the fraction of select 91-year-olds that die between ages 92 5519.55 and 95 is 10417.58 = 0.52983. That is, 1|3 q[91] = 0.52983. End of Example 1.6 7