0Journal of Epidemiology and Community Health 1997;51:80-86

How good is Sullivan's method for monitoringchanges in population health expectancies

Colin D Mathers, Jean-Marie Robine

AbstractStudy objective - To compare health ex-

pectancies calculated by Sullivan's methodand the multistate life table method inorder to identify the magnitude ofthe biasin Sullivan's method and assess how ser-

iously this limits its use for monitoringpopulation health expectancies.Design - A simulation model was usedto compare health expectancies calculatedusing Sullivan's method and the multistatelife table method under various scenariosfor the evolution of disability over time inpopulations. The simulation model was

based on abridged cohort life tables usingdata on French mortality from 1825-90 anddisability prevalence data from the 1982French health survey.

Main results - The Sullivan method couldnot detect a sudden change in disabilitytransition rates, but the simulations sug-

gested that it provides a good estimateof the true multistate value if there are

smooth and relatively regular changes indisability prevalence over the longer term.When disability incidence rates are in-creasing or decreasing smoothly over time,the absolute bias in the Sullivan estimate ofdisability free life expectancy is relativelyconstant with age. The relative bias thusincreases at older ages as disability freelife expectancy decreases.Conclusions - The difference between theestimates produced by the two methodswas small for realistic scenarios for theevolution of population health and Sul-livan's method is thus generally acceptablefor monitoring relatively smooth long termtrends in health expectancies for popu-lations.

(_T Epidemiol Community Health 1997;51:80-86)

Australian Institute ofHealth and Welfare,GPO Box 570,Canberra ACT 2601,AustraliaColin D Mathers

Laboratoired'Epidemiologie etd'Economie de laSante, Universite deMontpellier, EquipeINSERM, 34094Montpellier Cedex 5,FranceJ-M Robine

Correspondence to:Dr C Mathers.

Accepted for publicationJuly 1996

The concept of a health indicator which com-

bined information on mortality and morbiditywas first proposed by Sanders' and the firstexample of such an indicator was publishedin a report of the US Department of HealthEducation and Welfare,2 which contained pre-liminary estimates of "disability free life ex-

pectancy" (DFLE) calculated using a methoddevised by Sullivan.3 This involved using theobserved prevalence of disability at each age inthe current population (at a given point oftime) to divide the years of life lived by a periodlife table cohort at different ages into years withand without disability. During the first halfof the 1980s, Sullivan's method was used to

estimate DFLE for the USA,4 Canada,5 and

France.6 Estimates of active life expectancy forthe USA had also been made using doubledecrement life table methods.7

Health expectancy indices which combinemortality and morbidity into a single compositeindicator are a very attractive tool for moni-toring long term trends in the evolution ofpopulation health and for addressing the ques-tion of compression or expansion of morbidityin populations. During the second half of the1980s, there was a dramatic increase in thenumber of health expectancy calculationscarried out, almost all using the Sullivanmethod,89 which uses information on the pre-valence of health states in the population. Morerecently, multistate life table methods havebeen developed which require information ontransitions between health states. Various au-thors have claimed that Sullivan's method pro-duces biased or incorrect estimates and cannotbe used to monitor health expectancies ofpopu-lations over time.' 1-2 This paper compareshealth expectancies calculated by Sullivan'smethod and the multistate life table methodusing a simulation model under various scen-arios for the evolution of disability over time.

The problem with Sullivan's methodSullivan's method uses the observed age spe-cific prevalences ofhealth states in a populationat a given point in time ("sectional" pre-valences) to calculate the years of life lived inthe various health states at each age by a periodlife table cohort.'3

Multistate life table methods were proposedby Rogers, Rogers and Belanger'° 4 to take intoaccount reversible transitions between goodhealth and one or more disability (or otherhealth) states. Their data for people aged 70years or more from the US longitudinal surveyon ageing (LSOA) show that transition ratesfrom dependence to independence can be sur-prisingly high, even for the older old. In ad-dition, the multistate life table method allowsone to calculate health expectancies for popu-lation subgroups in a specific health state at agiven age, eg those not disabled at age 65,whereas the Sullivan method gives only theaverage health expectancy for the entire popu-lation at a given age.

Problems relating to the validity of the Sul-livan method were first raised in 1989 byBebbington" and Brouard and Robine.'Bebbington compared the Sullivan methodwith the double decrement life table methodusing data where the disability incidence ratewas rising over time. He demonstrated thatthe Sullivan method gives a lower estimate of

80

Assessment of Sullivan method in France

disability than the double decrement methodbecause, in effect, the disability prevalence rateused in the Sullivan method reflects the pastexperience of each cohort and not the currentincidence rates. Brouard and Robine'5 similarlyargued that the prevalence of disability is astock dependent on past history, whereas theincidence of disability is a flow which can beused to compute a "pure period proportion" ofdisabled people, not dependent on past flows,which in turn could be used to compute a pureperiod indicator ofDFLE. They noted that thequestion of whether to measure stocks or flowsis a common dilemma in demography and inother disciplines such as economics. Brouard'6had previously demonstrated the large bias re-sulting from the use of stocks rather than flowsto calculate trends in working life expectancyfor French women and has reminded thosedebating the merits of the Sullivan and mul-tistate methods of the very similar discussionsrelating to the use of prevalence versus lon-gitudinal data in "working life tables".17-19

Rogers, Rogers and Belanger'o argued thatthe Sullivan method produces biased estimatesof active life expectancy and that this bias,which is pessimistic, leads to the conclusionthat active life expectancy is declining in apopulation whose transition probabilities be-tween the independent and dependent statesare held constant. They argued that the Sullivanindex is biased in the direction of increaseddependency whenever the independent popu-lation is very much larger than the cor-responding dependent population. It wassubsequently demonstrated that the dis-crepancy between estimates made using the twomethods results not from bias in the prevalenceindex, but from the use of prevalence estimatesfrom a non-stationary population, that is, apopulation in which the prevalence of de-pendency has not reached the equilibrium valueassociated with current transition rates."3

It is now well understood that Sullivan'smethod, unlike the standard life table methodfor calculating period life expectancy, does notproduce a pure cross sectional indicator derivedfrom the current health conditions of the popu-lation.20 This is because the prevalence rates arepartly dependent on earlier health conditions ofeach age cohort, that is, incidence, recovery,and state specific mortality rates applying atearlier times (or ages). To construct a purelycross sectional indicator, one would have touse the equilibrium prevalences observed in afictitious cohort which had always been exposedto the observed cross sectional transition ratesbetween health states. In an equilibrium orstationary population, where all transition ratesare constant over time, Sullivan's method givesthe same health expectancies as the multistatemethods. 3 The problems with Sullivan'smethod arise not because it uses prevalenceand mortality data averaged over all healthstates, but because the data it uses are de-pendent on past conditions in the population.This problem has been illustrated in a scenariosimulating large changes in cardiovascular casefatality rates over a short time period,'2 al-though the authors incorrectly generalise from

this extreme case to conclude that Sullivan'smethod is always inaccurate.2'

Theoretically, the multistate life tablemethod is to be preferred for calculating healthexpectancies, but its use requires longitudinaldata which are expensive and time consumingto collect and are rarely available. Hence wehave developed a simulation model usingFrench data, which allows us to compare theSullivan estimate with the pure period estimatefrom the multistate life table for a non-sta-tionary population which has experienced real-istic changes in transition rates over time. Ouraim is to identify the magnitude of the bias inSullivan's method and assess how seriously itlimits the use of the method for monitoringtrends in population health.

The multistate life table methodFor these simulations, we work with abridgedlife tables using five year age groups throughoutthe lifespan, and terminating in a final openended age interval of 85 years and over. It hasbeen shown that abridged life tables provideestimates of health expectancies that are ofacceptable accuracy compared to estimatesbased on full life tables.'322 The starting pointfor the Sullivan method is the standardabridged life table,23 and particularly the lifetable functions lx (the number of persons sur-viving to exact age x) and Lx (the total numberofyears lived in the age interval (x, x+ 5)). Theobserved prevalence of disability, dx, in the ageinterval (x, x +5) is used to divide the yearslived in the age interval into years spent withdisability, dXLx, and years without disability(1 -dx)Lx.The average expectation of life free of dis-

ability at age x, or DFLE at age x (DFLEX), isthe total years of life lived free of disabilityfrom age x onwards divided by the number ofpersons alive at age x:

w

DFLEX = (1- di)Lillxi.=x

where the summation on index i is across thefive year age intervals of the abridged life table(x, x+ 5,, . . ., w), where w = 85 denotes the finalopen ended age interval.The multistate life table generalises the single

state life table (which analyses the transitionfrom a single "alive" state to the "absorbing"death state) to include reversible transitionsbetween two or more non-absorbing "alive"states. In order to calculate DFLE, we considera two state life table with a non-disabled state(denoted by subscript 1) and a disabled state(denoted by subscript 2).The number of survivors in state k (k = 1, 2)

and age x is denoted by lxk (see figure 1). Thetransition probability qx of the single state lifetable becomes a transition matrix giving theprobabilities of transition between states j andk in the age interval (x, x+ 5). We ignore thecomplexities of multiple transitions and con-sider only the probabilities of transitions be-tween an initial state at age x and a final state

81

Mathers, Robine

Figure 1 Transitions in the multistate life table.

at age x+5. The transition probability i4, theprobability of a person not disabled at exactage x being disabled at exact age x+ 5, is closelyrelated to the incidence rate of disability forthe age interval (x, x + 5). The transition prob-ability rx, the probability of a person disabledat exact age x being free of disability at exactage x + 5, is closely related to the recovery ratefrom disability for the age interval (x, x+ 5).The transition rates qxl and qx2 are the prob-

abilities of dying within the interval (x, x + 5)for a non-disabled and disabled person re-spectively.

Ifwe assume that persons who make a trans-ition between states (1, 2 or 3) spend on averageone half of the interval (x, x +5) in the ori-ginating state (this could be generalised todifferent fractions for each pair of states ifsuch information were available), then it isstraightforward to calculate the survivors x+ 5,1and 4+5,2 in each state at age x + 5 and the yearsof life Lx1 and LX2 lived in states 1 and 2 duringthe interval (x, x + 5).The "disability free" life expectancy for the

entire cohort at age x is

w

DFLEX = E il(lxl + 1x2)i=x

Experimentation with different values for theproportion of the interval (x, x + 5) lived in the

Incidence------ Recovery

40 60 80 100

originating state has shown that the simulationresults are not highly sensitive to the value ofthis parameter, and that the assumption that itis 0.5 for all transitions does not significantlyaffect the validity of the comparisons betweenthe Sullivan and multistate methods for thescenarios examined below.

The multistate simulation modelESTIMATION OF TRANSITION RATES FOR FRENCHMALES IN 1982Robine et al6 calculated the DFLE for Frenchmales using data on the prevalence of disability(and institutionalisation) by five year age groupin 1982. We have chosen a set of transitionrates to and from disability which give a goodfit of the calculated prevalence of disability tothe observed prevalence at each age using thefollowing assumptions:

1. 1.05% of the life table cohort of 100 000are born disabled (so 10I = 98 950 and 102 =1050).

2. The relative risk of death for disabledcompared to non-disabled persons is as-sumed to be R=3.66 at each age. Thisallows the computation of the state spe-cific mortality risks q1 and q2 from /H, 4,2,and q,, the proportion of the cohort ineach state and the total life table mortalityrisk, as follows:

q, = 4x*qJ(41. + R*4x2)

q2= R*ql

Under this assumption, the state specificmortality risks will vary not only in accordancewith trends in the overall mortality risk, butalso with the relative prevalence of disability(4.2/14) -

The value 3.66 was taken from LSOA datafor persons aged 70 years or more.24 The sim-ulations are not highly sensitive to the valueassumed over a range of R= 1.0 to 10.0, or tothe assumption that the relative risk is the sameat all ages.Figure 2 illustrates the transition rates chosenfor the stationary population scenario (in-cidence and recovery transition rates constantover time). The "recovery" rates were initiallyset at r,=0.2 for ages 65 and over, based onestimated two year recovery rates from theLSOA,25 and then incidence rates chosen toresult in life table prevalence rates L.2/L, closeto the observed prevalence rates. It was foundnecessary to vary the recovery rates also toobtain reasonable fits. Note that there is not aunique solution for the incidence and recoveryrates, although they are constrained severelyby the observed prevalences. The rates chosenare purely for the purpose of carrying out thesimulations described below.

THE SIMULATION MODELIn order to compare the Sullivan and multistatemethods, we set up a model containing 34fictitious cohorts of 100 000 males born at five

1 2 3Age Not disabled Disabled Dead

Ix IXF/ i ,;1xr.

x + 5 Ix+5,1 Ix+5,2

0.1a)

4

c0

0.01

0.0010

Age (y)

Figure 2 Transition rates chosen for the stationary population scenario to give theobserved prevalence of disability in the French male population in 1982.

82

Assessment of Sullivan' method in France

4- Sullivan DFLE------- Multistate DFLE

Total LE at age 0(French males)

1960 1980Calendar year

Figure 3 Comparison of Sullivan and multistate life table methods for calculatingdisability free life expectancy (DFLE) under scenario 1: constant mortality rates,disability incidence rates increase by 50% from 1945.

75 r

70 H

65 H

For each scenario for the evolution of dis-ability transition rates over time (describedbelow), a set of assumed transition rates werespecified for each period from 1825 to 1890.These transition rates were chosen to give agood fit of the calculated period prevalenceof disability by age in 1982 to the observedprevalence in 1982 as described above. Theseperiod transition rates together with the periodlife tables were used to construct multistate lifetables for calendar years from 1920 to 1990 atfive year intervals. The multistate life tableswere used to calculate the total years lived ineach state and hence the DFLE. This is a pureperiod expectancy based on the assumed periodtransition rates. The life tables for each periodand birth cohort were all constructed usingthe three assumptions outlined in the previoussection.

2000

CALCULATING THE DFLE USING THE SULLIVANMETHODFrom the age specific transition rates assumedfor the various periods in the model, we canobtain the age specific transition rates ex-perienced by each birth cohort as it ages withtime. Together with the cohort life tables, wecan thus construct the multistate life table foreach birth cohort and calculate the prevalenceof disability at each age in the cohort. Theseprevalence rates at given ages are influencedby all the transition rates experienced by thecohort at earlier ages in earlier years.These were then used to estimate the period

prevalence of disability by age for each calendaryear. Note that these prevalences come from adifferent birth cohort for each age group andhence are different functions of transition ratesfor prior periods through which the cohortshave lived. Finally, these simulated "observed"period prevalences of disability by age are usedtogether with the period life table to calculateDFLE using the Sullivan method.

40n1900

II

1920 1940 1960

Calendar year

1980

Figure 4 Comparison of the Sullivan and multistate life table methods for calculatidisability free life expectancy (DFLE) under scenario 2: age specific disability incideand recovery rates are constant over time.

year intervals from 1825 to 1990. The abriperiod life tables for French males for ea

these calendar years was obtained fromlished sources.'$'8 For some of these ywhere life tables were not available foispecific year, a life table was constructedlog-linear interpolation of the mortality iniq_ from the life tables for the nearest avaiyear on either side. The five year interv:lowed us to ignore the effects of the two v

wars during which life expectancy decresubstantially.29 These period life tablesthen used to construct cohort life tableeach of the 34 birth cohorts, representiniactual survival experience of French malecohorts.

2000 CONSTRUCTION OF SCENARIOS

The following section examines a number ofillustrative scenarios in which age specific trans-

ing ition rates from non-disabled to disablednce change over time but, for simplicity, recovery

rates are assumed to be constant over timefor each age. Note that simple but reasonablyrealistic scenarios have been chosen which il-

idged lustrate relatively regular long term trends inch of population health and also some short term andpub- drastic changes in population health. Separaterears, simulations were carried out to test the sensi-r the tivity of the results reported here to the as-

using sumption that recovery rates are constant over

dices time. It was found that, while time trends inilable recovery rates contributed to differences be-al al- tween Sullivan and multistate estimates ofvorld DFLE, the Sullivan method is less sensitive to-ased variations in recovery rates than disability rateswere for the types of scenarios examined here. Thiss for is because the recovery rates are generallyg the higher in magnitude than incidence rates andbirth so their time variations are more rapidly re-

flected in the resulting prevalence rates.

75 r

70 H

65 f

60K

C.)cCIDC.)a1)0.xa)

I 55K

50K

1900 1920l

1940

C.)

41)

CL

xa)

I

60 H

55K

50 K

45 H

83

An L

Mathers, Robine

75

co 60CD,x

CD° 55_/,

a)

50 -4----Sullivan DFLE------- Multistate DFLE

Total LE at age45 (French males)

401900 1920 1940 1960 1980

Calendar year

Figure 5 Comparison of the Sullivan and multistate life table methods for calculatidisability free life expectancy (DFLE) under scenario 3: disability incidence rate decby twice the amount of the mortality decline from 1945 onwards.

75 n

70 H

65K

60

55

50 *- Sullivan DFLE--- Multistate DFLE

Total LE at age ((French males)45

1900 1920 1940 1960

Calendar year

1980

Figure 6 Comparison of the Sullivan and multistate life table methods for calculatidisability free life expectancy (DFLE) under scenario 4: disability incidence rate inc

by twice the amount of the mortality decline from 1945 onwards.

prevalence for French males in 1982. This isbecause the computed prevalences of disabilityfor 1982 in the simulation model are dependenton the changes in transition rates assumedunder the various scenarios.

After examining a first scenario which graph-ically illustrates the limitations of the Sullivanmethod, we examine an equilibrium scenarioin which disability transition rates are constantover time while mortality rates decline as ob-served historically in the French male popu-lation. We then examine two extreme scenariosin which disability incidence rates increase anddecrease at twice the absolute magnitude ofthe change in mortality rate at each age. These

D scenarios illustrate strong versions of the twomajor hypotheses for the evolution of health:compression of morbidity and expansion ofmorbidity. It is very likely that the real evolution

2000 of morbidity in the populations of developedcountries falls somewhere between these ex-tremes.

ingreases

ResultsSCENARIO 1

Mortality rates are assumed constant (at 1982values) and in 1945 the incidence rate of dis-ability is assumed to increase by 50% at eachage.Up to 1945, Sullivan's method and the mul-

tistate method give the same value for DFLEat birth (see figure 3). In 1945, the multistateestimate faithfully follows the drop in incidencerates, but the Sullivan estimate takes 30 yearsto reach the same value. This graphically il-lustrates how Sullivan's method has averagedthe change in incidence rates in 1945 over 30years through the use ofperiod prevalence rateswhich take decades to reach their equilibriumvalues for the new transition rates.

)

SCENARIO 2Age specific incidence and recovery rates are

iJ assumed constant with time. Mortality rates2000 decline as observed historically in the French

male population. As shown in figure 4, Sul-ing livan's method gives almost the same results as

reases the pure period estimate from the multistatelife tables.

It is important to note that, while simplescenarios have been chosen, and constant val-ues assumed for some transition rates, the mul-tistate method itself has not been simplifiedapart from the assumption that half of an ageinterval is lived in the originating health state.Separate simulations were also carried out toensure that this assumption did not affect thevalidity of the comparisons between the Sul-livan and multistate methods.Where the "incidence" rate is assumed to be

increasing or decreasing as a multiple f of thechange in mortality rate, it was necessary toalter slightly the age specific incidence andrecovery rates chosen for 1982 to maintain a

good fit of the calculated period prevalenceof disability for 1982 to the observed period

SCENARIO 3

Starting in 1945, at each age from 5 onwards,the disability incidence rate is assumed to de-crease by twice the amount which the death rateof non-disabled persons decreases. Mortalityrates decline as observed historically in theFrench male population. In this scenario, itis assumed that the "prevention of disease"reduces equally both the incidence of deathfrom a healthy state and the incidence of non-fatal disability.

Sullivan's method and the multistate methodgive results that are reasonably close (see figure5). The value given by Sullivan's method islower than the multistate estimate, but gives a

reasonably good indication of the trend over

time.

C-)C

C.)a)xI)

-C4UI

84

85Assessment of Sullivan s method in France

3

2.5 Age U to /D4a) Age 80

Age 85+

c 2 -

a)

cr

t 1.5

(V

0)

1900 1920 1940 1960 1980

Calendar year

Figure 7 Postulated changes in disability rates over time for scenario 5: the disabiliincidence rate decreases except during the two world wars and the 1960s.

75 r

70

65

60

55

50

45

40 *- Sullivan DFLE------- Multistate DFLE

Total LE at age(French males)

25 1 l1900 1920 1940 1960

Calendar year

1980

Figure 8 Comparison of the Sullivan and multistate table methods for calculatingdisability free life expectancy (DFLE) under scenario 5: the disability incidence ratedecreases except during the two world wars and the 1960s.

SCENARIO 4

At each age from 5 onwards, the disalincidence rate is assumed to increase bythe amount which the death rate of nonabled persons decreases. Mortality rates deas observed historically in the Frenchpopulation. In this scenario, it is assumedall people "saved" from death are left discand that an equal extra number of non-dis,persons become disabled. This expansion ofbidity is assumed to start in 1945 (any e

and disability incidence rates would go nega

SCENARIO 5

Sullivan's method and the multistate methocresults that are still reasonably close (see f

6), but in this case, the value given by Sullivan'smethod is higher than the multistate estimate.For scenarios 3 and 4 (disability incidence

rates decreasing and increasing by twice themortality trend respectively), the SullivanDFLE in 1990 are approximately 0.9 yearslower and higher respectively than the mul-tistate DFLE for ages 0 to around 65. As theabsolute bias in the Sullivan estimate of DFLEis relatively constant with age, the relative biasincreases from around 1.5% at birth to around7% at older ages where DFLE is much lower.

DiscussionData requirements for multistate methods areconsiderable and there are very few countrieswhere national data is available or likely to beavailable for some time. Sullivan's method is

2000 being used more and more widely. The Sullivanmethod is not capable of detecting a suddenchange in disability transition rates, but our

ity simulations suggest it provides a good estimateof the true period value if there are smooth andrelatively regular changes over the longer term,as postulated by the principal scenarios forevolution of population health. We illustrateboth these conclusions using a more complexbut not unrealistic scenario for French males(see figure 7). In scenario 5, disability incidencerates jump substantially during the two worldwars, but otherwise decline at a constant annualpercentage except during the 1960s when theepidemic ofcardiovascular disease was peaking.In order to illustrate clearly the differencesbetween the methods, we ignore any effects ofthe two world wars on mortality. The resultingestimates of DFLE obtained by the Sullivanmethod and the multistate method are shownin figure 8. We conclude that Sullivan's methodseems to be a quite acceptable method for

0 monitoring long term trends in health ex-

pectancies at the population level.Barendregt et al'2 also carried out a sim-

ulation comparing the Sullivan method with the2000 multistate method. Their hypothetical example

was based on a sudden and very large changein survival rates, stable incidence rates andvarying prevalence rates and shows that theSullivan method is quite incapable for moni-toring the resulting changes in health ex-pectancy. This confirms the conclusion reachedhere using scenario 5, that Sullivan's methodis not appropriate for detecting sudden changes

bility in population health. However, they generalisetwice their conclusion to state that Sullivan's methodt-dis- is not appropriate for the analysis of changes incline health expectancy over time in any cir-male cumstances. We disagree with the conclusion:that the simulations presented here have shown that

abled Sullivan's method provides acceptable estimatesabled of the true period value of health expectancy ifmor- there are smooth and relatively regular changesarlier in transition rates over a reasonably long termtive). period, and in particular, provides good estimates

of trends in health expectancy.We fully agree that the multistate life table

method is to be technically preferred for cal-l give culating period health expectancies, since it isigure based on period transition rates only and will

C

0I)

x01).C

0)CD

35

30

Mathers, Robine

detect sudden or gradual changes in disabilitytransition rates. However, its use requires lon-gitudinal data which are expensive and timeconsuming to collect and are rarely available.The LSOA has opened new prospects for cal-culation of truly cross sectional health ex-

pectancies, as evidenced by recent papers,242530but it should be noted that in order to avoidbias, institutionalised persons must be in-terviewed beginning with the first round of thelongitudinal survey - which was not the case

with the LSOA.It may be possible to develop alternative

methods to correct for the conservative bias in-herent in Sullivan's method. One possibility isto include a retrospective question on health stateone or two years ago in cross sectional healthsurveys to obtain information about flows intoand out of health states. There are two problemswith this, long term recall of health states maybe biassed and the question can only be askedof survivors. Pollard, Golini, and Milella3' havedeveloped a method for calculating transitionrates from such cross sectional survey data con-

taining information on the time of onset of dis-ease. It may also be possible to obtain informationon differential survival through record linkage or

use of a national death index.Methods of analysis and calculation have

proliferated at a far greater speed than theavailability of good quality population healthdata, and we feel that the simulations reportedhere should lend confidence to the continueduse of Sullivan's method to calculate populationhealth expectancy time series and assist in iden-tifying the circumstances in which Sullivan'smethod provides an acceptable alternative to

the technically exact multistate method for

monitoring trends in population health.

CONCLUSIONS

The Sullivan method is not capable of detectinga sudden change in disability transition rates,but it provides a very good estimate of themultistate value if there are smooth and rel-

atively regular changes over the longer term

and we conclude that Sullivan's method seems

to be a quite acceptable method for monitoringrelatively smooth long term trends in health

expectancies at the population level in non-

stationary populations.Data requirements for multistate methods

are considerable and there are very few coun-

tries where national data is available or likelyto be available for some time. Sullivan's methodis being used more and more widely. Sullivan'smethod provides a useful indicator which can

be used with confidence for monitoring trends,as long as its limitations are understood.

This work was assisted by the National Health and Medical

Research Council (Australia) which provided a Travelling Fel-

lowship to Colin Mathers, Naturalia et Biologica (France), and

a grant from CNAMTS (France).

1 Sanders BS. Measuring community health levels. Am 7Public Health 1964;54:1063-70.

2 US Department of Health, Education, and Welfare. Toward

a social report. Washington, DC: US Department of Health,

Education, and Welfare, 1969.

3 Sullivan DF. A single index of mortality and morbidity.HSMHA Health Reports 1971;86:347-54.

4 Colvez A. Blanchet M. Potential gains in life expectancyfree of disability: a tool for health planning. Intj Epidemiol1983;12:224-9.

5 Wilkins R, Adams OB. Health expectancy in Canada, late1970s: demographic, regional and social dimensions. AmJ Public Health 1983;73:1073-80.

6 Robine JM, Colvez A, Bucquet D, Hatton F, Morel B,Lelaidier S. L'esperance de vie sans incapacite en Franceen 1982. Population 1986;41:1025-42.

7 Katz S, Branch LG, Branson MH, Papsidero JA, Beck JC,Greer DS. Active life expectancy. N Engl J Med 1983;309: 1218-24.

8 Network on Health Expectancy. Statistical world yearbook"health expectancy". Montpellier: REVES, 1993.

9 Mathers C, Robine JM. Health expectancy indicators: areview of the work of REVES to date. In: Robine JM,Mathers CD, Bone MR, Romieu I, eds. Calculation ofhealth expectancies, harmonization, consensus achieved andfuture perspectives. Colloques INSERM Vol 226, John Lib-bey Eurotext and Les Editions INSERM, France, 1993.

10 Rogers RG, Rogers A, Belanger A. Longer life but worsehealth? Measurement and dynamics. Gerontologist 1990;30:640-9.

11 Bebbington AC. Expectation of life without disability meas-ured from the OPCS disability surveys. In: Robine JM,Blanchet M, Dowd EJ, eds. Health expectancy. London:HMSO, 1992;23-34.

12 Barendregt JJ, Bonneux L, Van der Maas PJ. Health ex-pectancy: an indicator for change? J Epidemiol CommunityHealth 1994;48:482-7.

13 Mathers C. Health expectancies in Australia, 1981 and 1988.Canberra, ACT: Australian Institute of Health 199 1; Ap-pendix B.

14 Rogers RG, Rogers A, Belanger A. Active life among theelderly in the United States: multistate life-table estimatesand population projections. Milbank Memorial FoundationQuarterlylHealth and Society 1989;67:370-411.

15 Brouard N, Robine JM. A method for calculation of healthexpectancy applied to longitudinal surveys of the elderlyin France. In: Robine JM, Blanchet M, Dowd E, eds.Health expectancy. London: HMSO, 1992;87-97.

16 Brouard, N. Esperance de vie active, reprise d'activite fem-inine: un modele. Revue Economique 1980;31:1260-87.

17 Schoen R, Woodrow K. Labor force status life tables forthe United States, 1972. Demography 1980;17:297-322.

18 Bonneuil N, Brouard N, Robine JM. Active life expectancy:a method of calculation. REVES Paper No 92. Proceedingsof the 5th International Meeting of the Network on HealthExpectancy (REVES). Ottawa, 1992.

19 Brouard N. Evaluation of existing analytic approaches. RE-VES Paper No 45. Proceedings of the 3rd InternationalMeeting of the Network on Health Expectancy (REVES).Durham, 1990.

20 Crimmins EM, Saito Y, Hayward MD. Sullivan and multi-state methods of estimating active life expectancy: two

methods, two questions, two answers? In: Robine JM,Mathers CD, Bone MR, Romieu I, eds. Calculation ofhealth expectancies, harmonization, consensus achieved andfuture perspectives. Colloques INSERM Vol. 226, JohnLibbey Eurotext and Les Editions INSERM, France,1993.

21 Van de Water HPA, Boshuizen HC, Perenboom RJM, Ma-thers CD, Robine J-M. Health expectancy: an indicatorfor change? (Letter). J Epidemiol Community Health 1995;49:330.

22 Boshuizen HC, van de Water HPA, Perenboom RM. Inter-national comparison of health expectancy: preliminaryresults. In: Mathers CD, McCallum J, Robine J-M, eds.Advances in health expectancies: Proceedings of the 7th Meetingof the International Network on Health Expectancy (REVES),Canberra, February 1994. Canberra: Australian Instituteof Health and Welfare, 1994.

23 Chiang CL. Life table and mortality analysis. Geneva: WorldHealth Organisation, 1978.

24 Crimmins E, Hayward MD, Saito Y. The relationship be-tween changing mortality rates, changing morbidity rates

and the health status of the population. REVES Paper No91. Proceedings ofthe 5th International Meeting of the Networkon Health Expectancy (REVES). Ottawa, 1992.

25 Rogers RG, Belanger A, Rogers A. Recent developmentsin multistate modelling of active life expectancy. Paperpresented to 5th International Meeting of the Network on

Health Expectancy (REVES). Ottawa, 1992.26 Institute National de la Statistique et des Etudes Econ-

omiques. Annuaire Statistique de la France: Retrospectif.Paris: PUF, 1961.

27 Institute National de la Statistique et des Etudes Econ-omiques. Annuaire Retrospectif de la France; 1948-1988.Paris: INSEE, 1990.

28 Institute National de la Statistique et des Etudes Econ-omiques. Couet C, Court Y, eds. La situation de-mographique en 1990. INSEE Resultats No 193-194,Demographie-Societe No 16-17, Paris, 1992.

29Vallin J. Tables de mortalite du moment et par generation 1899-1981: mise a jour des tables annexes du cahier no 63. Vol 93.Paris: INED, 1984.

30 Belanger A, Rogers A, Rogers RG. Recent developments inmulti-state modeling of active life expectancy. REVESPaper No 90. Proceedings of the 5th International Meetingof the Network on Health Expectancy (REVES). Ottawa,1992.

31 Pollard JH, Golini A, Milella G. On the use ofhealth surveysfor estimating transition rates for morbidity processes.Genus 1991;XLVII(3-4):63-77.

86