You are on page 1of 21

rs.

If the
lived and

^ration as
boration.
M u ltip le D e c re m e n t P ro c e s s e s
pnuously
neasured

measure
what we

4.1 M ultiple Decrem ent Tables for a Real Cohort
4.2 M ultiple Decrem ent Life Tables for Periods
4.3 Som e Basic M athem atics of M ultiple Decrem ent Processes
4.4 Associated Single Decrem ent Tables from Period Data
4.5 Cause-specific Decom position of Differences in Life Expectancies
4.6 Associated Single Decrem ent Tables from Current Status Data
he time
4.7 Stationary Populations w ith M ultiple Sources of Decrem ent
if some
1 as the
parried In the previous chapter, we defined a single decrement process as one in which individuals
have only one mode of exit from a defined state. A multiple decrement process is one in which
individuals have more than one mode of exit. As we noted, multiple decrement processes are
far more common in demography than are single decrement processes. These situations arise,
for example, in fertility analysis when individuals are viewed as being exposed to the risk
of pregnancy and to discontinuance of use of contraception; in migration, when individuals
are exposed to the risk of moving to different places; in nuptiality, when married persons are
exposed to the risks of divorce and widowhood; and in many other circumstances. And as we
noted earlier, persons in a particular state in a real cohort are always exposed to the risk of death,
in addition to whatever other risks they may have of leaving that state. Multiple decrement
processes are sometimes referred to as situations of “competing risks.”

4.1 M ultiple Decrem ent Tables for a Real Cohort

Conceptually, the construction of a multiple decrement table for a real cohort is no more chal­
lenging than the construction of a single decrement table. It is only necessary to add columns
equivalent to other columns in a single decrement life table but which pertain exclusively to
particular causes of decrement (i.e., to particular modes of exit from the table). The functions
in these columns have a straightforward interpretation:

n d' x = number of decrements from cause i in the age interval x to x + n

lt q'x = probability of leaving the table from cause i between ages x and x + n for
someone who reached age x

= ,,4/ix

72 MULTIPLE DECREMENT PROCESSES

„m' x — rate of decrement from cause i in the age interval x to x + n

= n^x/nL x
l x = number of persons reaching age x who will eventually succumb to cause i
OO

The l x values in these formulas pertain to the number of cohort members who reach age
x, i.e., who have survived all causes of decrement before age x. Likewise, the „L X column
pertains to all person-years lived between x and x + n by the cohort members who have
survived all causes of decrement. All of the columns customarily found in a single decre­
ment table are also found in a multiple decrement table, where they refer to “all causes of
decrement combined.” Although we are using age x as the basic dimension of the multiple
decrement table for illustrative purposes, it is representing the more general dimension of
duration since entry into the state.
Note that we have not defined columns for n L' x , T x , n a' x , or e' x . The reason is that e' x (for
which the other columns are needed as input) does not admit to a straightforward interpretation.
In one sense, it could be “the life expectancy at age x for persons who will succumb to cause
/.” But those who will later succumb to i cannot be identified at age x. They will only be
identifiable at later stages as the competing risks work themselves out. It is pointless to compute
an expectation for unidentifiable people. A similar problem of interpretation arises in regard
to /'., the number of persons aged x who will eventually exit from cause i. But this is a useful
column because it can be used to calculate l x /l x , the proportion of persons aged x who will
eventually leave the table from cause i. One of its major uses, for example, is to calculate the
probability that a marriage will end in a divorce. The base of this proportion is l x, all persons
surviving to age x, and this number is indeed identifiable by age x. It is when l' x is made the
base of a probability or expectation that conceptual difficulties become intractable.
Added up over all causes /, the decrements must sum to the total number leaving the defined
state:

By our formulas for n m' x and n c/[., these must also sum to the equivalent function in the life
table for all causes combined:

and

. 1800. and then adding the columns that pertain to decrements from the individual causes. Almost always.1.1) simply says that all of the survivors to age x in the cohort must leave the defined state from one or another recognized cause of decrement above that age. MULTIPLE DECREMENT PROCESSES 73 iuse i i reach age h.e.v column who have igle decre- causes of e multiple tension of Age in years —*- tat e‘ x (for Figure 4.1 presents life-lines for a cohort of 10 individuals each of whom was born on January 1. This latter task is simply a matter of recording what has happened.1800 b to cause II only be ) compute Also. pretation. So the basic problem is one of converting these observed rates into the probabilities of exiting the table from various causes. Figure 4.2 M ultiple Decrem ent Life Tables for Periods Constructing a multiple decrement table for real cohorts involves preparing a basic life table for all causes of decrement combined. The life lines are converted into certain columns of a multiple decrement life table in table 4.. these intensities will be represented by sets of decrement rates from various causes. The state of interest is being single. Exits from the state are to marriage and death. as indicated on the figure. date of birth.1 Life-lines in the single state for a hypothetical cohort of 10 births. January 1. since in regard s a useful who will l'x = 'En4> pulate the ci=x I persons oo oo made the K = EE^ = E'^'=Z* (4-1) i i II. i. n M‘ x . the life 4. never-married.-V ci—x e defined The relation in (4. But very often the analyst wants to draw forth the implications of the intensities of a multiple decrement process that are recorded during some specific period of time.

1 Age Number Number Number Number Probability Probability Probability Number Number Person- remaining dying marrying leaving of dying of of leaving reaching reaching years in between between single state between marrying single state age x who age x who lived in single ages x and ages x and between ages x and between between will will single state state at x + n while x+n agesx and x + n while agesx and ages x and eventually eventually between age x in single x+n in single x+n x+n die while in marry ages x and state state single state x+n M M lD X lx nd° nudx ndx =„4? nix n'lx nix — n1? Lx lx n^x +~nduxM «M +'n^ix 0 10 1 0 1 1/10 0 1/10 4 6 9.Table 4.00 50 1 1 0 1 1/1 0 1/1 1 0 9. tz> -+h? .07 l 9 1 0 1 1/9 0 1/9 3 6 32.00 10 8 1 3 4 1/8 3/8 4/8 2 6 70.60 60 0 0 0 g 9 9 •obabilii —. ££ _ ™ CD 2 S3. gS o P5 CD° CD-• ^ P 5 3 ►n Q-o xCP 3. —■ o H ^ In orde *o >-• a 'c cd o p^ Z CD C/5CDH t.22 5 8 0 0 0 0 0 0 2 6 40. cd ^ P £ »« B* »^ "Ooo " 3 ”*~3s —•9 o 9.1: Life table in the single state for a hypothetical cohort of 10 births shown in figure 4.00 40 1 0 0 0 0 0 0 1 0 10.88 20 4 0 3 3 0 3/4 3/4 1 3 23. aS st zr o s P M < ^r.9 3 ~ SO31o—.36 30 1 0 0 0 0 0 0 1 0 10.

” But nothing requires that the rate of cancer death per person-year of exposure be increased when other causes decline. Holding constant n m' x .3). that “people must die of something. more of the potential victims of cancer will be carried off by other causes in the age interval. and hence the probability of cancer death.1) for „l/v. What is clear is that the number of person- years of exposure to the risk of cancer deaths will increase when other causes decline. So we can use the relation between l x and „ L x derived in chapter 3 to develop a conversion formula. l x . So „m x = n m‘ x + „m~ l .„a x ) n m x This formula for nq'x is very similar to formula (3. after all. MULTIPLE DECREMENT PROCESSES 75 In order to make this conversion. and if the probability of exiting from one cause declines.nax)(nmx + nmx ') Now the competing nature of the multiple decrements comes into full view. the probability of exiting from some other cause(s) must increase. the proportion of 65-year-olds who die from cancer between ages 65 and 70 will be lower the higher are death rates from other causes. refer back to the relation between rates of decrement and probabilities of decrement for a cohort: and A n lx Note that. will be increased in the interval. In other words.2) are not specific to cause i but are the values pertaining to all causes of decrement combined. all persons starting life must die of something. when „m x ‘ is higher. Inserting this relation in formula (4. The denominators are identical.n a x )„d x )/n for l x in the expression for n q' x gives: nix = (4. It is common to write the rate of decrement from causes other than / in the age interval x to x + n as „m x'.2) has n m l x in the numerator where (3. Because of this dependence. n q x is commonly referred to as a “dependent probability.2) gives: nix = (4. given a certain death rate from cancer in the age interval 65-9. the only difference between the formulas is that (4. for example. Given the same number starting out an age interval (l x). Substituting ( n i x + (n . The values of „a x and n m x in the denominator of (4. the number of deaths from cancer.1) has „m x .3) 1 + (n . just as in the basic life table of chapter 3.2) 1 + In. the lower will be n q' x .” Does this dependence mean that when the death rate from other causes declines the death rate from cancer must rise? Not at all. so that the number of cancer deaths will increase. the numerators of n m x and n q' x are the same whereas the denominators of the former is „ L x and of the latter. the higher is n m~‘ in (4. . The necessary relationship that governs the multiple decrement process relates not to the rates but to the probabilities'. This common confusion results from a mistaken view that the rates of decrement are necessarily dependent on one another. The reason for this dependence is that.

= nm'x.A n c l\ — nQx ' = nqx n dx Once we have computed the life table for all causes combined.76 MULTIPLE DECREMENT PROCESSES Formula (4. as described in chapter 3. The basic ingredient in this table is nmx = the rate of decrement from all causes combined in the age interval a: to x + n.2) for n q‘ x will provide a straightforward way of converting „m' x to ^ thereby completing the multiple decrement life table. c) Compute the number of decrements from cause i in the age interval x to x + n as: n dx n qix ' lx d) Compute the number of persons aged x* who will eventually leave the table from cause i as: l'x* — X] a d'x . This must be converted to „q x . t qx n dx •ll>1x So: . where n M x is the observed decrement rate from cause i between ages x and x + n in the population. then the relationship becomes: — n . by dividing n q' x by n q x . a -V _ M‘ n.a ' D' n q x — a q. themselves. we get: . So the steps for constructing a period multiple decrement life table are the following: a) Compute a life table for all causes of decrement combined. or the recorded decrements.//v . we can simply take the n q s series from that table and apportion it to various causes of decrement according to their relative rates of decrement. it is unnecessarily tedious to implement. — nq x Mx n n n ux where „ D' x is the observed number of decrements from cause i between ages x and x + n in the population and n D x is the observed number of decrements from all causes combined in the interval.v).x . However. Note that. b) Compute the probability of exit from cause i in the age interval x to x + n as: n c Lx '■ nq X Note that if we have accepted nM[.A . The usual procedure consists of assuming for each cause that n M\ = n m‘ x (which also implies n M x = „/n. since the probabilities are in the same ratio as the rates.

958 97.959 0.160 20 4.00236 0.027 3.532 97.909 13.217 0.823 er3.554 36. A=’'c.948 467 98.431 25.00314 304 20.000).489 15.00425 0.00963 915 19.264 All 1.07669 0. 1996.573 217 98. MULTIPLE DECREMENT PROCESSES 77 D n 4 and tedious to Box 4.618 0. The state of interest is being alive and decrements from that state are attributed to neoplasms (cancer) and all other causes of death.17427 0. 45 19.205 I x +n ibined Proportion of female newborns that will eventually die from neoplasms under the US age-cause-specific death rates of 1991: 21.00019 19 21.249 0.653 0.571 69.756 1..00092 0.211 17.00025 24 21.050 0.00003 3 21. 000 = 21.'K 1. 4 = £"4 (l=X Example: US.634 70 114.008 0.00015 15 21.201 5 1.03371 0.091 10.187 10 1. „f(v = As 'l* 3.637 0.00015 14 21.578 32.205/100.01330 0. from Box 4.714 246.739 41.264 4.969 17.172 15 3.076 35 12.439 19.000 0.882 0. females. l x = "master" US . using the lx and nqx columns for the “master” life table in the US in 1991.00584 0.05155 0.02095 0.27363 0.263 89. the nqx life table for 1991.205 ft in this 3.00171 167 20. National Center for Health Statistics.01547 1.043 60 59.x' „DX 2.924 98.2% Proportion of females who survive to age 75 that will die from neoplasms: 10.00168 0. The chance that a newborn would die from neoplasms is 0. 618 = 14.018 0.04439 3.695 78.141 jte usual 25 6.1 presents several columns for a multiple decrement life table for US females in 1991.673 95.00818 0.7% Data source'.00012 12 21.146 85 320.1 Multiple Decrement Life Table .02191 1.212 (21.870 0.03696 2.145 75 143. 236/69.10212 4.00090 0.491 856 98.543 5.078 93.070 0.00000 0. Cause / = death from neoplasms.289 0.486 0.428 1.711 0.02920 2.085 21.00783 0.087 33.236 80 164. females.519 50 25.693 36.00314 0. 1991.958 * 55 37.083 0.991 ages x 40 15.264 8.220 57. 205/100.434 96.00087 85 21.881 7.379 0.11552 0.986 30. for all causes combined r relative Agex ii Dx „D‘X lx n Qx X „4 l‘x recorded AH deaths Deaths from neoplasms ng: 0 15.00582 561 20.047.00262 0.169 1 275 99.955 318 98.05012 2.00041 41 21.117 implies 30 9.384 11.758 63 100. .534 85.604 65 88.634 268 99.

then the probability that no event will occur in three tosses is the product of the probabilities . If we define an event as tossing a head...(x) = lim „m x We can correspondingly define the force of decrement from cause i as: |x'(x) = lim n m' x The force of decrement from cause i at age x is simply the rate at which persons are leaving the defined state from cause i in the tiny interval of age from x to x + clx..' (y) over all i.....” The force of decrement functions from various causes are additive and their sum equals the force of decrement from all causes combined as long as we define the set of decrements to be mutually exclusive and exhaustive (i.. The analogy to coin-tossing is obvious..h p-* (x) = p(x) where |x(x) is the “force of decrement from all causes combined. decrements) per person-year of exposure. = e-/.e. e-/. e -f x x + " vHy)dy . so that: +" [M-lOO+tr20-)+-+lr*(v)]4y n P x = e-/..78 MULTIPLE DECREMENT PROCESSES 4.4) is the probability of remaining in the defined state during the interval x to x + n if only decrement i were operating. we haven’t left any out and we haven’t counted any decrements twice). with k causes of decrement..e. It is an annualized rate because its units are number of events (i..+B d y or.3 Som e Basic M athem atics of M ultiple Decrem ent Processes Recall that in chapter 3 we defined the force of mortality as: p. We asserted in chapter 3 that one of the most important relations in demography is: In a multiple decrement process we can express |x(y) as the sum of p. Since.4) says that the probability of remaining in the defined state between x and x + n when many causes are operating (i.. Equation (4.+" n'Mdy ..... the probability of surviving all decrements) is the product of each of the probabilities of remaining in that state if individual decrements were acting alone.e... + nmx + • • • + nmx = n "br­ as we take the limit of both sides as n approaches zero we must have |x](x) + |a2(x) H---------. nPx = where (4..

does not mean that the underlying processes that they represent are independent. MULTIPLE DECREMENT PROCESSES 79 than no head will appear on each of the three tosses: 0. since Its were both numerator and denominator include these terms. For example.admit no overlap or combinations or synergistic relations .5 • 0.4) will continue to hold. That is. Then: is: x+n „ 4 _ fx+ "l(< a +(fl)da j e -J>W d yp}{ a )da n <7. the process of assignment of a cause to each particular decrement created a set of wholly separate and “independent” entities. a head. i.4) The formula for n q‘ x shows again why the probability of succumbing to cause i in a discrete age if only interval depends upon the value of other decrements.5. equation (4.V\x k (a)iovx <a < x+n.4). In a cohort. the assumption of independence has also slipped into our derivation of (4.Im-1 u/j-*(>’)!dy (jT (fl) da f x + n e-/“lLr10.2(n)-|-----------.1 {x ) + |x2(x) + • • • + u.e. i over a small age interval da are l{a)p! (a) da. Clearly. on p.1). This multiplicative property ■tains only when the outcomes of the three tosses are statistically independent: when one outcome does not depend on the others. separated. n x and The formula for n m‘ x shows that its value is also “dependent” on the force of decrement S) is the from causes other than i.)+(x2(>’)+-+M-t0’)Kv da JX (4. However.1(fl) + p. That these statistical entities are independent . the direction of this dependence is not predictable.5 • 0. functions from single decrement i can be derived by noting that deaths from cause 'ice). it is very likely that an increase in the incidence of influenza in a population will raise death rates from certain cardiovascular diseases as well as from influenza. It entered at the point where we defined members of the set of decrements to be mutually exclusive and exhaustive. Other functions in the life table from all decrements combined can be similarly expressed in terms of the various force of decrement functions by substituting p.v l(x) l(x) X . But whatever 8 are leaving this synergistic relation among disease processes. For example: OO l e-firta)da dx = J oo e ~fo [»i. so that: x+n = J e-fil*'W + *2 W+"-+*k W d ypJ(a)da X i n4 J+ n e-f> ( y )d y i x i (a)da X ~ nL x ~ f*+ n e-i: m-0 ’)dy da fx+" g.*(x) for p(x) in the corresponding continuous notation formula (see appendix 3. the age composition of person-years lived in the age interval x to x+n will depend ibilities only on the 1(a) function in that interval.l(o)+l<-2(«)+""l-M-*(a)14n^JC decrement 0 0 nient from dusive and In addition. the higher is any other force of decrement. . the data will always come to the analyst in i annualized a set of cause-of-death assignments in which influenza and cardiovascular diseases are tidily xposure. the lower will be n q' x .

Keyfitz. The formulas of appendix 3. since cause i has been arbitrarily deleted from the set of multiple decrements. For example. we don’t observe directly associated single decrement processes: processes in which one decrement alone is operating. p'(x). nM'x. Almost always. depends not only on those force of decrement functions but also on the preexisting age composition of the population. ” In an associated single decrement table for a period.). the activity of a particular decrement i will be observed as it works itself out within a multiple decrement process. 4. and Schoen. Functions referring to one of the decrements in a multiple decre­ ment life table have a “mixed” structure since survivors to any age must be computed with the p(x) functions whereas deaths from cause i must be computed using p' (x). will be: (4. it need not be the case that n M‘ x in a population equal nm‘x in a cohort with the same force of decrement functions from i and —i.4) expresses the formula for *p' x .in a multiple decrement process.80 MULTIPLE DECREMENT PROCESSES For a population at a moment in time. associated single decrement functions are obtained by computing both survivors and deaths with p' (x) alone. N(a).1 can readily be adapted for these functions. It is easy to specify what the functions in an associated single decrement table should be in terms of p'(x). It is not necessary to repeat all these formulae but it might be useful to review the three types of functions encountered so far. But. it is common to assume that the life table n m' x in a population equal n M\ in the population (for an exception. the probability of surviving from age x to x + n in an associated single decrement life table where cause i is the only decrement. We denote all functions of an associated single decrement table with a as a superscript on the left. Very often the analyst is interested in knowing what a life table would look like if only that cause of decrement were operating to diminish a cohort. The construction of an associated single decrement table therefore involves a thought experiment in which we ask “what would happen if . If a table is constructed based on p _ '(x). . the age composition. it is sometimes called a “cause-deleted” table. see Preston. Sim­ ilarly..5) where c(a) is the proportion of the population aged a to a + da within the age group x to x + n. 1972). Equation (4. this feature involves an additional layer of hypothesizing beyond that involved in constructing a hypothetical cohort. as in the case of the basic life table.4 Associated Single Decrem ent Tables from Period Data Associated with each decrement i in a multiple decrement process is a force of decrement function.. The life table that results from asking this question is called an associated single decrement life table (ASDT).along with i . Functions involving all decrements combined are obtained by computing both survivors and deaths with p(x) = p 1 (x) + • • • + p^(x). Normally. by replacing p(x) by p'(x). Such a table would still be considered an associated single decrement table since cause — i can be considered one of two decrements . As in the case of the basic life table. They are the same functions defined for the basic life table in chapter 3. The decrement we are interested in may equally well be —i (all decrements other than . So the mortality rate from decrement i between age x and x + n observed in the population.

n a x) n m' x le decre- ted with The remaining problem of life table construction is then to adopt a set of *a x values.n x the three and lined are r). as in X population 1972). How can we then turn around and say that we want to reproduce that same which we n M[ function in the associated single decrement table to estimate *m' x ? But if we don’t start triod. this with „M 'x. Since at all idered one ages |i(n) > pi (a). In this case we would proceed surviving as in chapter 3 for the basic life table. We can observe.table that whereas e (ASDT). If able. since In general. the age-distribution of person-years lived in the interval is older when a single decrement is operating than when multiple decrements are operating. If we use conventional jted single procedures. the rate of decrement from jx' (x) if i were the only decrement (*mjt) differs from Ible would what it would be if i were working in the presence of other decrements (n m 'x). MULTIPLE DECREMENT PROCESSES 81 y oil those nl compare: population. The „m x -> „q x conversion would be: the only * M l 1 11mY ir Y1— x nin. We n chapter have generally ignored this latter disparity in producing a basic life table.5) X =* l‘(x) J e-f^'W^iiJWda X group x to A+/Z population „ 4 = l ( x ) f e-/>0’)‘V(fl)r/a ■ But. We observe n m‘ x in a cohort but we need *m' x for the ASDT: Mould look f x + "e-f° v'Wdy^Wda f^+" e-f'My'Idy^ (a) da . But it is important to note that the value of „ M' x we observe when multiple decrements are operating to generate the observation would not typically be the same value we would observe if only decrement decrement i were operating. we will already have accepted the n M‘ x function to estimate n m' x in the multiple hstruction decrement table. Graduation . Sim- vors and n4x = 1 + (n .ml.. lem which could be solved by procedures similar to those outlined in chapter 3. f^'e-f'^'^dyda f x +n e~ EvWdy da than ().. The disparity is produced by a process exactly rscript on analogous to the age distributional disturbances that can produce a disparity between n m x and 1e should n M x even when the force of mortality functions are the same in a cohort as in a population. in a cohort (or n M x in a population). This weighting s itself out differential raises *m‘ x relative to n m x if pi (x) is rising in the interval. n d x =l(x ) J e-/r ^ dy ii(a) da X A-f-/? (4. so we might do the replacing same in moving from a multiple decrement table to an ASDT. a prob­ ixample. none of them completely satisfactory: (a) Ignore the disparity between *m‘ x and n m\. where do we begin? (ructing a There are several solutions to this dilemma. Although we can readily define associated single decrement functions in terms of pi (x ). X+ll population. we cannot directly observe pi (x).

By substituting y for the p' (x) function in the expressions for *m x and n m‘ x we find that *m' x and n m‘ x are both also equal to y. above 65). if the force of mortality function is constant in the age interval at x .r J x . . when the causes deleted in the ASDT are relatively minor. means that p' (a)and p ' (a) have exactly the same shape between ages x and x + n. by assumption. But the assumption is rarely very satisfactory unless the age interval is small. It is equal to the n p x in the parent life table raised to the power Rj. although their levels will generally differ. So in this case. it is the preferred approach. mortality is high from many causes and is rising rapidly with age. *p>x — e-!x+" R-'(a)da _ e~ f*+" R‘ y(a)da then nP'jc = e~Ri'fx+" v¥)da _ |^g-//+" = \nPx] R ' So in this case. When the assumption is tenable. If the force of mortality function is constant with age in the interval. the n p x function in the ASDT bears a simple relation to the „p x function in the “parent” life table for all causes combined. In cause- of-death life tables at high ages (say. However. (b) A second solution is to make the assumption that the force of decrement function for i is constant at some value y in the interval x to x + n. The changes in age structure induced through the thought experiment performed in producing an ASDT can be relatively large if the causes deleted have high death rates. Since.th e n : * n Px o . dx ■e *// Lx lx+n * n n Ml IVI x and the rest of the table can be readily completed. n m‘ x = pn' x — n M‘ x . of course. This procedure is logically consistent and simple to apply. Also. (c) A third approach is proposed by Chiang (1968).82 MULTIPLE DECREMENT PROCESSES procedures become relatively more attractive since “borrowing” *a' x functions from other populations is difficult (practically none are published) and hazardous. Procedure (a) risks committing relatively large errors in some circumstances. where R' is the constant of proportionality for decrement i in the interval. then the rate of decrement is undisturbed by any age distributional changes induced by deleting a cause. Assume that the force of decrement func­ tion from cause i is proportional to the force of decrement function from all causes combined in the age interval x t o x + n : p' (a) = R 1 • p(a) for x < a < x + n. So fairly substantial changes can be induced in the death rates from remaining causes. this approach should work quite well because the age structural changes induced by the deletion within 5-year-wide intervals could not be very large. This assumption.

the values of *a‘ x were set equal for all i to „a x in the parent life table. it does not tell us what value of * n m' x should be used in the life table or. combined: s and is leriment n D‘ x = f x +n N(a)R' \x(a)da ve high it D. An approach that is adaptable to intervals of irregular length is to approximate the value of *a x by interpolation between two extreme situations. what amounts in the to the same thing.7) . In fact. in the interval will simply equal the ratio of the observed decrements from cause i to the decrements for all causes i cause. But a moment’s thought suggests that. the two extreme situations correspond to R 1 =0 and R' = 1 respectively. which used rval at this approach.x f x + " N(a)[L(a) da paining So making this substitution for R 1 gives: should itr-wide >n for i Jnction This ingenious device solves the „m x —> n q x conversion problem in producing the ASDT.6) No iteration is required in this case because the *cl' x function is directly calculable without n. but clearly the number of years lived in the interval by those alive at the beginning of the interval must be n. Nor is there any obvious solution to this iduced problem. a life table that reflects the operation of a decrement with higher values of |i' («) in the interval will have a younger age distribution of person-years lived and hence a lower value of „ci x . is then undetermined. 1968).2). In the official US cause-of-death life tables of 1959-61 (NCHS. similar to an approach suggested for all causes (section 3. then the simple graduation formula for data tabulated in 5-year age or duration intervals is: (unc­ oined (4. The most satisfactory approach to estimating *a[ is probably to graduate the *q' x function in successive intervals and infer the value of *a x from the general conformation of this schedule. we can interpolate between these extremes: . MULTIPLE DECREMENT PROCESSES 83 mi other Furthermore. ■ So in However. If we assume that the distribution of deaths from cause i follows a quadratic function over the age interval x — 5 to x + 10.2. the n a x value in the parent life table must fall short of the *a x values for all i. with the same shape of the force of mortality functions as for all causes combined. *a x . by the Chiang assumption. Since R' represents the proportion of deaths in the age interval due to cause i. The other extreme situation is when all deaths in the age interval are from cause i. The first one is when there are no deaths from cause i in that age interval. what value of *a x should be used. by the assumption of proportionality. in ble (4. The average number of years lived in the interval by those dying from cause i. In the intermediate cases. the value of R. Then. the average number of years lived in the interval by those alive at the beginning of the interval must be the same in the ASDT and in the parent life table. of knowing F n.

54 years relative to the life table with all causes present. . we presented a method for estimating the contribution of age-specific mortality differences to differences between two life expectancies. approach (b) is best since it is completely consistent and fully developed.84 MULTIPLE DECREMENT PROCESSES In both the ASDT and the parent life table the following relationship must hold: II RX . a gain of 3. The table is constructed by Chiang’s method (approach c). 1. Under this assumption. in absence of neoplasms. _ —. we will investigate how to obtain a better estimate of Joa 85 • The table shows that. If the interval n is short.5 Cause-specific Decom position of Differences in Life Expectancies In section 3. In chapter 7. can be estimated with the following equation: „<(2) -»<(!) nm x(2 ) — ntWx(l) nR 1 (2 ) • n m x ( 2 ) ~ „#x(l) .. life expectancy at birth would have been 82. Which of the general approaches to constructing an ASDT should be used in any particular application obviously depends upon the nature of the data and tenability of assumptions. as should have been anticipated on the basis of the discussion in chapter 3. For the last age group. then approach (c) is best because the assumption of proportionality must be very good (the decrements that remain are a very high proportion of total decrements to which they are assumed proportional).46 years. In other situations. the contribution of differences in all-cause mortality in a specific age group can be distributed proportionately to the difference in cause-specific mortality in the corresponding age group (Arriaga.= 10 to 75.— n ■ n P x + n a x • n < lx l(x) = n — (n — nax) ' nQx So equation (4. Box 4.10. we use here the assumption that *m x = n m 'x. Elandt- Johnson and Johnson (1980) have compared several of these approaches and concluded that results are not very sensitive to the procedure used.9) n m x(2) n (1 ) . This method can be easily extended to estimate the contribution of differences in cause-specific death rates by assuming that the distribution of deaths by cause is constant within each age group in each population.8) This equation can also be derived through a Taylor expansion. It is based upon the data for US females in 1991 that were presented in box 4.6) for a. in which case ^is simply e% 5 /R ~'.8) for x = 0.n m x ( \ ) (4. n A'Y . If one is constructing cause-deleted tables in which relatively small causes of decrement are deleted. The specific contribution of differences in mortality rates from cause i between ages x and x + n. 5. convenience may dictate the approach selected.1. 1989).2 presents an associated single decrement table for causes of deaths other than neo­ plasms.7) can be written as: (4. and using equation (4. with the *a~' values developed by graduation (equation 4. 4. and 80.

e the high Agex R~‘ * -/ */— i *a~ i lx u Px nClx 4 II l x l x 11 lix V lions.42 0.08 0.547 59.61668 97.605 82.95178 90.54 0.99182 2.646 22.231 2.618 0.325 2.657 64.80 0.637 44.700 2.628 15.000 0.34 :i 80.46 neo.91960 98.05 0.99220 100.86205 98.54 0.67 0.99763 98.582 48.72637 2.2 Associated Single Decrement Life Table for Causes of Death other than Neoplasms (via approach c) n Dx — n D'x RII= UD.217 0.98162 95.068 2.74525 69.631 19.23 The 10 0.46 years.99250 96.605 78.83599 99.462 2. Box 4. 0 0.46 35 0. 1996.99727 98.16 : the 75 0.86701 79.875 73.91991 86.68006 78.969 7.99661 98.29 d by 15 0.97003 93.198 2. 1991.v = 0.54014 95.46 0.008 0.91322 99.97905 2.776 2.99783 98.99764 2.88448 2.94845 2. Life expectancy at birth in the absence of neoplasms: 82.152 82.616 50.8) for .90562 20 98.10 0.289 0.637 10.242 2.70633 97.969 2.486 0.070 0.74 0.711 0.42 Probability of surviving to age 85 in the absence of neoplasms: 0.653 0. (4.96629 2.00000 52.570 9.655 31.000 0.653 68. .585 53.399 2.54105 93.09 0.36 50 0.539 6.26 0.48 0.618 12. 80 andt- from equation (4.99922 98.148 2.672 39.152 78.8) l x .00 60 0.15 45 0.696 21.5.249 0.63 a 85 ' 30 0.48 B5 is 25 0.10 5 0.637 0.96 40 0.56219 96.686 17.992 2.550 55.92 0.22 ided 70 0.866 2.676 13.050 0.695 25.99575 2.03 0.99908 2.82573 2.99495 97.703 30.89788 41.42 this 80 0.61931 85.843 69.95 0.81683 57.6) for x = 10 to 75 :1 that in the 85 c85 ice it >85 R• !es in Example: US.600 2.99846 99.77017 68.018 0. Data source: National Center for Health Statistics.57492 89. 1 0.685 40.647 27.870 0.99587 97. 0.959 0.92 years.99910 2.92331 2.283 7.548 63.79593 98.28 ling rom Probability of surviving to age 85 for all causes combined: 0.64 55 0.99686 2.275 78.98670 2.220 1.nPx> n a x* e x W = {nPxf * r 4-11 — lX " ' I tin icular *a-‘ : calculated from equation (4.275 74.99923 99.882 0.10 0.00000 6. 1.99217 0. females.07 n be 85 0.51 ality 65 0.083 0.53 Life expectancy at birth for all cause combined: 78.72 0.99738 2.681 36.379 0.95 0.467 2.78 2.98863 96.695 34.677 45.99832 1. cause i = neoplasms .99600 100.99416 2.577 58.915 2.86813 98.756 0.

In our example.6 Associated Single Decrem ent Tables from Current Status Data We noted in chapter 3 that a restricted associated single decrement process can be directly observed when cohorts are asked about their event histories. it is possible to proceed directly to the construction of an ASDT for the process under investigation. which can be obtained by question on current status in a survey. and „ A x = contribution of all-cause mortality differences in age group * to * + n to differences in life expectancies. the proportion never-married at age *.3 presents an application of the age and cause decomposition method to analyzing the difference in male life expectancies at birth between China and India in 1990. It shows that about 68 percent of the 8. However. Their force of decrement function may differ from that of the original members of the cohort. it will be possible to estimate ^ p‘ x .11) It can be easily shown that „ A . It should be reiterated that tables constructed from such data will pertain only to the surviving members of the cohort who did not emigrate. John Hajnal (1953) was apparently the first to recognize that certain ASDT’s could also be constructed from current status data.and cause-specific contribution to differences in life expectancies can thus be pre­ sented in a two-by-two table where the elementary contributions add up to the total difference in life expectancies. From the 5-year question.e. For example. we need not view them as subject to a multiple decrement process.M(*). a question might be asked on “what was your marital status five years ago. as expressed in equation (3. Such applications require retrospective questions.” rather than the more complex set of questions needed to construct a complete marital history. The only restriction on the kind of functions that can be displayed will result from limitations in the data. the probability that a person who was aged * five years earlier (i. In particular. Since the members of the cohort who respond in the survey have clearly had a force of mortality function of zero. without any resort to retrospective questions. he proposed that an associated single decrement table for first marriages be constructed from data on the current marital status of the population. Therefore. But no information will be available on |. = £ „ A ' . 4. To see this more formally. the l x column of the cohort’s ASDT (with a radix of one) will exactly equal the proportion single at age *.86 MULTIPLE DECREMENT PROCESSES where n R'x 0) = proportion of deaths from cause/ in age group* to* +n ( n D' x / n D x ) in population j (or at time j). and eg(2)-eg(l) = X ) « A * = ££»A!c -V *i The age. . the total difference in efj is reduced by about one year due to lower rates of noncommunicable diseases in India in 1990.°s(*) + p.2 additional years of life expectancy at birth in China is attributable to lower rates of communicable diseases below age 5 in this country. is closely related to the l x column of the birth cohort’s ASDT. Under the assumption that there is no differential mortality or migration by first-marriage status. who is aged * + 5 at the time of survey) will have remained single to the time of the survey. Box 4. let us assume the population to be closed to migration. Then the force of decrement from all causes combined for the single population at age * will be: V? (*) = p.

will be found by j the cumulating the forces of decrement: lion ctly S(x) = 5(0) ■ xp0 = 5(0) • e~fo M-’W the = 5(0) • e~f° [pDl(«)+M.068 0.1 0. From Cause 3 = Injuries. Harvard iginal University.0388 0.2 15 0.315 0.0 0.5 — 58..504 0.m x ( 2).337 0.0028 0.223 0.109 0. .6 5.5 -0.304 0.v + n at time 1 and time 2 differences (or in population 1 and 2) r'(1). 0 0. MULTIPLE DECREMENT PROCESSES 87 population Box 4.0102 0.051 0.0009 0. School of Public Health. 1990* i in 1990. perinatal and nutritional conditions.1003 0.2 xiction ♦In this example: la.248 0. and Risk Factors in 1990 and Projected to 2020. males.592 0.v nmx nRl n Rx nRl unicable .677 0.429 0.309 0. By relations n be developed above.025 0. Injuries.728 0.4 -0. D.073 0.(!) ference in ■' .070 0.174 0.vP) = ad-cause mortality rate between ages x and .3 Age and Cause Decomposition of Difference in Life Expectancies at Birth //i v(f)> n«. 10 was 70 3 single Total Difference = (China) —ejj(India) = 66.573 0.0084 0.0139 0.1 - iroceed Sum 8.394 0.v=0i=l W'ated Data source: Murray.1 -0.039 -0.552 0.877 0.174 0.v .1 -0.3 0. the number of single persons in a cohort at age x.0025 0.0267 0.380 0."(u)]r/fl where 5(0) is the number who were single in the cohort at birth.0 30 0.1 -0.084 0.046 0.6 -0..0 0.0015 0.0 -1. n R'x{2) = proportion of deaths from cause i between ages x and x + n at time 1 and time 2 (or in population 1 and 2) — contribution of all-cause mortality differences in age group x to x + n to differences in e" (from equation 3. maternal.879 0.6 0. rather Cause 2 = Noncommunicable diseases. 5(a).2 9.7 0.n m x (\) ii method Example: India and China.X(2 ) •„m. and A.882 0.8 0.0 'e need 70 0.8 0.188 0.4 (1 ) •„/».9 0.382 0.0043 0.257 0.2 years= /C .722 0.0021 0.5 0.v(2 ).6 0.A ? . R3 nR'x A2 n iV.3 = 8. A ? Jowever. 1996.0342 0.Ai .149 5.1 0.247 0.1 ! cohort 60 0. The Global Burden of Disease: A Comprehensive Assessment of tohort Mortality and Disability front Diseases. For Cause 1 = Communicable diseases.7 -0.3 0.101 0. Lopez.1 directly 45 0.3 0. C.488 0..796 0.A« .11) is be pre­ . Boston.104 0.030 0.095 0. j. y the data. Ls be where \x Ds {x) is the force of mortality at age x for the single population and |x'M(x) is the rt to force of first marriage (defined of course only for the never-married population).. in China Agex India China II A.0 5 0.326 0. .0929 0.

Nevertheless. N(x). Now form the ratio of the single to the total population at age x: S(x) S(0)-<r/o N(x) N(0) ■ e~fo R-0' ( a '> da _ e ~ fo [M.11) N(x) where *p^ is the probability of remaining in the single state between ages 0 and x in the cohort’s ASDT. Define the proportion single (i. the difference between the force of mortality functions for the single and total populations. never married) at ages x to x+5 as: . This expression includes a term in [p.Ds(a)-n-Dr(«))]^« (4. S(x)/N(x) is an unbiased estimate of *p^ for the cohort members who survived. If the assumption of no differential mortality by marital status is wrong. This is the logic of Hajnal’s procedure. Then each cohort would trace out the same history as every other cohort and features of that history could be inferred by comparisons across cohorts.. based on the force of first marriage alone.then: S(x) = ~fo \x M (a)da _ * M e p (4.there is no differential mortality by marital status . If we assume that this difference is zero at all ages .10) since 5(0) = N( 0). So we see a very simple connection between an observed proportion single for a cohort at age x. Then S(x)/N(x) will be biased as an estimate of * by the factor: exp The proportion single at x will underestimate the probability of remaining single in the cohort ASDT because single persons have higher mortality. can be similarly derived as: where p.e. S(x)/N(x).M(a)+((Ji.Dr(a) is the force of mortality function for the total population at age a.88 MULTIPLE DECREMENT PROCESSES The total population aged a: in the cohort. A single-round survey asking about current marital status thus yields one piece of information about the ASDT for each cohort interviewed.Di(a) — \x DI (a) ]. Hajnal’s method is typically applied in 5-year age intervals. Suppose that the death rate for the single population exceeds that of the total population at some or all ages. then S(x)/N(x) is biased as an estimator of the cohort’s * p q . But suppose we can assume that the force of nuptiality has been constant over time. a smaller fraction of them will have survived to report their marital status in the survey. and the ASDT for that cohort.

Since we have already shown that a stationary population with . entering school. v . Although we have illustrated the concept of deriving ASDTs from current-status data with examples drawn from first marriage.10) and (4. the procedure can be applied to a wide range of processes. Hajnal estimates the proportion who will not marry by age 50 as: n (50) = .(5 EI45 + 5 fl5 o) He then calculates the mean age at marriage over the base of those who do in fact marry by age 50: 5'E. Among these are the process of becoming a mother. the number of years the cohort will spend in each age interval: F 45 PY S (0. SMAM is called “the singulate mean age at marriage. a constant set of age-specific mortality rates. Its computation from current-status data in a single census or survey assumes that first marriage rates have been constant over time and that differences in mortality or migration rates by marital status are negligible. MULTIPLE DECREMENT PROCESSES 89 The person-years lived in the single state below age 50 could then be estimated bv addins no successivevalues of 5nv and multiplying by 5. Nevertheless. 4. of ever using contraception. •v=0.n ( 5 0 ) SMAM = 1 .” because it refers only to marriage occurring to the single population. we introduced one of the two most basic models in demography. and other processes. But they can be informative if interpreted correctly. as shown in equations (4. Let us now extend the assumption about mortality rates to include constancy of age-specific risks from each of the multiple sources of decrement.7 Stationary Populations w ith M ultiple Sources of Decrem ent In section 7 of chapter 3. Current-status data are not as informative about the timing of events as vital statistics or retrospective survey data. because it requires only a census tabulation of marital status by age. If rates have been changing.5 0 . of moving from place of birth.5 This value could serve as an estimate of the mean age at marriage for the hypothetical cohort except that not all members of the cohort will have married by age 50. Box 4. entering the work force. 50) = 5 • 5 nA. of becoming sterilized.11). The proportion of a cohort who have remained in a state is clearly a product of its cumulative rates of leaving the state.rio. it is the only measure of marriage age available in many historic populations.4 illustrates the computation of SMAM for Turkish males in 1990. and sometimes they are the only data available. Recall that a stationary population would result from the maintenance of three conditions: a constant flow of births per unit of time. The value of SMAM is the mean age at first marriage for a cohort of women (or men) who marry by age 50. and zero net migration rates at all ages. the value of SMAM is a hodge-podge of rates in the recent and distant past. the stationary population model.5 5 n .n(50) Those not marrying are excluded from the denominator and their person-years lived in the single state below age 50 are excluded from the numerator.

the number of persons in an annual cohort of births who will eventually succumb to cause i.853.258 30 2.541.023 45 Esn.026 50 980.000 10 3.784 43.50-.0245 Data source: Turkey.086 35 1. Thus. = 5.560. . it must also be the case that each year it would have „ L x ■ n m\ = n d' x deaths from cause i at age x to x + n.435. [Turkiye istatistik yilligi..052.043 40 1.000 5 3.418.113 28.052.077 0.581.096.899 180.203 0.541. males.ii9 A=0 0(50) = m+m _ .765 629.119. 1995. the sum of annual deaths from cause i over all ages must be equal to /q. 1990 Agex sNx 5 Sx 0 3.409 1.255 3. 1 0 annual births would have „L.4 Associated Single Decrement Tables from Current-status Data: Calculation tion of Singulate Mean Age at Marriage 5 Nx = Total population aged x to x + n 5 X = Number of single persons aged x to x + n S 5nv = = proportion single at ages x to .V persons at each age (where each function pertains to the same life table).409 3.061 3.90 MULTIPLE DECREMENT PROCESSES Box 4. 1995.0245 5-5. Turkey.627 0.718 25 2. Statistical Yearbook of Turkey.0245 SMAM = 25.115 22.153 1.784.255 1.0 years 1 .222 0.000 15 3.957 20 2. Basbakanlik Devlet Istatistik Enstitusu. 1996. Basbakanlik Devlet Istatistik Enstitusu.412 0.527 0.v + n 5n x 51145 +5 n 50 n (50) = 2 5-Ef=0S n x-50-0(50) SMAM = i .] Ankara.031 45 1.030.767 0.n(50) Example: Turkey.111.165.121 77.134 0.900 1.560.900 3.

5.D. 6 . If the number having the disease at a moment in time is H and the annual number of new diagnoses of the disease is / .will eventual^ die of cause / is equal to the proportion of annual deaths at ages x+ that are attributable to cause i ■ These relations have wide applicability in populations that can be assumed to be stationary: I The probability that a marriage will end in divorce could be estimated by the ratio of annual divorces to annual marriages.s by the number of entrants to graduate school. could be found by dividing the annual number of Ph. The probability that someone entering graduate school will eventually receive a Ph. MULTIPLE DECREMENT PROCESSES 91 In other words. 4. 3. This probability is typically called a “case-fatality ratio” in epidemiology and its value is often estimated by resorting to stationary population assumptions. The probability that a newborn will be diagnosed with cancer at some point in his or her life will be equal to the number of annual cancer diagnoses divided by the annual number of births. The “prevalence” of a disease can be defined as the proportion of the population having the disease at a moment in time. This is equivalent to life expectancy at birth being equal to population size divided by the annual number of births.D. then in a stationary population H 1H H To ' to 'T77' prevalence incidence expected duration of the disease itatistik same from ^t be umb . the probability that someone aged a. The “incidence” of a disease can be defined as the ratio of new cases of the disease diagnosed in a particular period divided by the person-years lived in the population during that period. Likewise. in a stationary population the probability that a newborn will evei die from cause i is equal to the annual number of deaths from cause i divided by t h e a n n JB— number of births (or of deaths from all causes. 2. since the annual number of births miL'e™-'] the annual number of deaths). The expected number of years spent with morbidity from a newly-diagnosed disease would equal the number of persons suffering from the disease at a moment in time divided by the annual number of new diagnoses of that disease. The probability that someone with newly diagnosed cancer will eventually die from cancer would equal the ratio of cancer deaths to cancer diagnoses.