The Epidemiological Transition: Policy and Planning Implications for Developing Countries (1993)

Kenneth G.Manton and Eric Stallard

Forecasting chronic disease and health problems in adulthood requires different models and data than those for infant mortality and infectious disease because chronic diseases are often associated with long-term exposure to risk factors and age changes in host physiology. Many infectious diseases have both acute and chronic health consequences, and often are cofactors for chronic disease incidence and progression. Lifestyle and dietary habits are also important risk factors for chronic and degenerative diseases. Thus, in projecting the prevalence of these diseases, as well as resulting disabilities and deaths, it is necessary to consider the level of exposure to the risk factors and the consequences of other diseases and conditions of poor health.

Two examples of infectious diseases that contribute to chronic illnesses are malaria and helicobacter pylori. Chronic malarial infection alters the immune system so that the Epstein-Barr virus (EBV) causes Burkitt’s lymphoma in East Africa, rather than infectious mononucleosis as in developed countries (Lam et al., 1991; Prevot et al., 1992). Helicobacter pylori (H.pylori), a waterborne bacteria, causes an infection that may be asymptomatic for decades (Barthel et al., 1988). It is a major causative agent for duodenal ulcers (Blaser, 1988) and gastric cancer (Forman, 1991; Talley et al., 1991), one of the most prevalent cancers in developing countries (Forman et al., 1990; Parsonnet et al., 1991). Prevention or treatment of H. pylori could

reduce gastric cancer (and possibly other chronic gastric disorder) rates by up to 60 percent in developing countries (Forman, 1991).

Communicable viral infections also contribute to chronic diseases. Although atherosclerosis and coronary heart disease (CHD) are etiologically linked to blood lipids, other factors explain more of the geographic, temporal, and cultural distribution of CHD. An etiology of cardiovascular disease (CVD) based on food-borne viral infection appears to be associated with the global distribution of risk factors and their change over time. There is evidence of viral activity in atherosclerosis’ association with immunological factors (Muscari et al., 1990; Tertov et al., 1990), which might explain why Chinese men with heart attacks have average cholesterol levels (i.e., 194 milligrams per deciliter (mg/dl)) lower than the “desired” level of 200 mg/dl (Schwartzkopff et al., 1990). Thus, viral and bacterial diseases have chronic effects on cancer (e.g., EBV and retroviruses), neurological disease (e.g., poliomyelitis and late effects of encephalitis), CVD, and autoimmune disorders (e.g., rheumatoid arthritis and rheumatic heart fever).

Predisposition to a number of chronic diseases may be the result of nutritional insults during gestation and infancy. Maternal “deprivation,” for example, increases the risk of type I diabetes in children (Crow et al., 1991). Other studies show that the ratios of birth to placental weight, and to body weight at 1 year, have strong associations with several chronic diseases (including diabetes) at later ages (Barker and Martyn, 1992). A low ratio (and low weight at 1 year) implies fetal growth dysfunction, favoring the development of the brain and heart at the expense of other organs (e.g., lungs or liver). Thus, malnutrition may cause fetal growth dysfunction, which contributes to the risk of CVD, pulmonary, and hepatic diseases through adulthood (Barker et al., 1991a,b, 1992a,b).

Many chronic diseases (cancer, CVD, stroke, diabetes mellitus) are associated with nutritional or lifestyle factors: fat and protein consumption, micronutrient deficiencies (Choi et al., 1990; Tonglet et al., 1992), alcohol use, smoking, and levels of physical activity (Dodu, 1988; Trowell and Burkitt, 1991). In developing countries, especially rural areas, fat intake and serum cholesterol tend to be low, thus reducing CHD risk. Among males in rural Nigeria, the mean cholesterol level was almost 40 mg/dl lower than in men living in urban areas, and similar differences were noted in Ghana and Côte d’Ivoire (Kesteloot et al., 1985; Knuiman et al., 1982). Mean cholesterol is low in China (e.g., Chen et al., 1991) where CHD is 7 percent of total mortality. Among the Tarahumara Indians of Mexico, cholesterol rose from 121 to 159 mg/dl over five weeks after changing from a “traditional” (2,700 kilocalories (kcal), low cholesterol and fat, high complex carbohydrate and fiber) to an “affluent” (4,100 kcal, high fat, cholesterol, sugar, and energy) diet (McMurray et al., 1991). Though cholesterol increased in the Tarahumara, the ratio of high- to low-density cholesterol did not change because of high levels of physical activity.

On the other hand, low levels of exposure to risk factors can also have adverse effects on health. Low serum cholesterol levels (especially low-density lipoprotein) is associated with gallbladder disease (Mohr et al., 1991), and with liver (Chen et al., 1991), colon (Cowan et al., 1990; Lee et al., 1991; Nomura et al., 1991), and lung (Isles et al., 1989) cancer. These associations may be mediated by low vitamin C intake (Choi et al., 1990; Jacques et al., 1987), low serum albumin (Kimura et al., 1979), viral infection (Mozar et al., 1990), rapid weight fluctuation (Hamm et al., 1989), or free-iron overload (Sullivan, 1989). The association of low cholesterol with hemorrhagic stroke and cancer is found in select population studies. Thus, diets consumed in developing countries need to be carefully examined for favorable and unfavorable elements, especially during pregnancy, infancy, and childhood, and old age (McGill, 1988; Tonglet et al., 1992).

Physical activity is a problem for the elderly in developing countries because after “retirement,” activity declines rapidly (e.g., Dowd and Manton, 1992; Wilson et al., 1991). In Zimbabwe, mortality rises after retirement because of the rapid onset of impairments of activities of daily living (ADLs) due to activity reduction and malnutrition, especially vitamin B deficiency (Wilson et al., 1991; Evans, 1990). Zimmet et al. (1991) found that reduced physical activity increased risk factors for CHD and diabetes in Mauritius. Maintenance of physical activity with limited impairment is difficult in developing countries because of a lack of the physical and housing aids available in developed countries. Thus, promotion of physical activity (with nutritional supplementation) is important at postreproductive ages, both for risk factor reduction and for health maintenance.

Nutrition, physical activity, and metabolic factors are important in chronic disease forecasts. Exposures such as viral and bacterial infection, alcohol, and smoking, also need to be modeled either directly or through effects on measurable parameters (e.g., effect of smoking on pulmonary function). Once disease (and mortality) is predictable from measured risk factors, it can be related to interventions and used in policy development. Disease burden can be translated into effects on human capital (Manton et al., 1991d) by forecasting the physical and cognitive impairments generated by chronic disease. This paper discusses a three-part model to forecast chronic disease, disability, and mortality: the first part describes changes in risk factors; the second predicts disability, morbidity, or mortality as functions of risk factors; and the third assigns costs for health events that reduce productivity or incur medical costs.

To conduct health forecasts and simulations, data are needed. Below we review types of data available in developed and developing countries.

Developed countries have national health surveys and longitudinal studies of select populations. Some U.S. health surveys are longitudinal (e.g., the Longitudinal Studies of Aging in 1986, 1988, and 1990, based on the 1984 Supplement on Aging to the National Health Interview Survey (NHIS); National Center for Health Statistics (NCHS), 1987). Some surveys measure risk factors (e.g., National Health Examination Survey (NHES), National Health and Nutrition Examination Survey (NHANES) and the NHANES-I ten year follow-up). Some community studies longitudinally follow risk factors and health outcomes (e.g., studies in Framingham, Mass; Evans County, Ga.; Albany, N.Y.; and Charleston, S.C.). Epidemiological and health survey data are common in Britain and Scandinavian countries (e.g., the Swedish Götebörg Study; the Finnish North Karelia project and East-West studies). Scandinavian countries have population-based disease registries (e.g., the Swedish Tumor Registry; Manton et al., 1986). The National Cancer Institute’s SEER program covers a large portion of the United States.

A number of surveys now describe the functional status of the elderly either by supplementing the sample with older persons (e.g., the Supplement on Aging—Longitudinal Survey on Aging (SOA-LSOA) or use of specialized survey designs (e.g., the 1982 to 1994 National Long Term Care Surveys (NLTCS)). Epidemiological studies of elderly populations were started (e.g., National Institute on Aging’s Establishment of a Population for Epidemiologic Studies of the Elderly (EPESE) program) and instrumentation was added to existing studies (e.g., functional assessment in Framingham and the 30-year follow-ups of the Finnish, Dutch, and Italian components of the Seven Countries Study). In addition, vital statistics data on mortality and administrative data on health service use are generally of good quality. Because of the availability of data, health can be compared across culturally and socioeconomically diverse populations in developed countries (e.g., persons of Japanese ancestry in Hawaii: Reed et al., 1988; blacks and whites in Charleston, S.C.; Lackland et al., 1992).

Developing countries often have health survey data due to World Health Organization (WHO) and United Nations programs. Most surveys rely on self-reported conditions and symptoms, and lack risk factor measurement. The 1976–1977 Indonesian disability survey used local physicians as interviewers but did not have formal clinical exams or measure risk factors (Dowd and Manton, 1992). There are now health surveys of the elderly in 17 countries in three WHO regions (e.g., Andrews et al., 1986; Manton et al., 1987). WHO’s noncommunicable disease program sponsored Monitor-

ing of Trends and Determinants in Cardiovascular Disease (MONICA) risk factor studies in 45 sites and cross-sectional surveys of CHD risk factors in Cuba, Ghana, Mauritius, Sri Lanka, Tanzania, and Thailand (Dowd and Manton, 1990). WHO and the National Institutes of Health sponsored stroke surveys in Taiwan, China, India, the United States, Nigeria, Colombia, Ecuador, Mexico, Venezuela, and Peru (Bharucha et al., 1988). However, with few longitudinal studies in developing countries, cross-sectional data on risk factor distributions may have to be combined with parameter estimates made from longitudinal studies in developed countries. Such a combination assumes that relations estimated from risk factor time-series data in developed countries represent biologically invariant characteristics of chronic disease processes. If this assumption “holds” (relative to the precision of other data employed) the model can legitimately integrate data and parameters from multiple sources.

A major methodological problem is to “mix” parameters estimated from ancillary longitudinal data sets to describe risk conditions in the reference population. Such a procedure is logically similar to “indirect standardization” (i.e., distributional differences are controlled by reweighting cell-specific rates) except that, within each cell population, we have a model describing the evolution of risk factors. Using models within each cell population exploits far more information than modeling a cell population by a single probability assuming the population is homogeneous. If the process is well described, inferences about health changes may be unconfounded from differences in the distribution of risk factors; that is, aggregate-individual interactions (conditional on individual parameters; Hoem, 1985, 1989) produce biases that are “negligible” relative to the precision of other estimates.

To forecast health changes for n years, we use a cohort component projection model,.

where P_t is a vector of age-specific population counts. The G matrix contains probabilities of surviving from one age category to the next for A age categories. One can produce projections for males and females or for ethnic groups by having Gs and Ps for each group. Survival probabilities are predicted from information on specific diseases and risk factors. The five-year probability of death is the sum of the probabilities of death for each of K causes. If we can eliminate a cause of death by intervention (e.g., influenza deaths may be eliminated among young and old persons by vaccination, nutritional supplements (e.g., vitamin A), and antiviral drugs) then we

can eliminate (or reduce by a fixed amount) the probability of dying from influenza. How information is used in predicting each element of G differentiates forecasting models. We used a flexible model (Manton et al., 1992a) with components for risk factor change, i.e., the relation of risk factors to mortality. Below, the components are integrated in a comprehensive model to forecast health status.

Risk factor regressions describe changes over time. Changes are due to lifestyle modification (cessation of smoking, less physical labor at work, reduction of stress); nutrition (change in protein-energy balance, fat consumption, micronutrient intake); environment (both general and job-related exposure); and health care access. The regressions describe changes over a fixed time period for each of J risk factors (and variables that either can be modified or are naturally undergoing changes we want to anticipate) as a function of age (many risk factors such as disability or cholesterol change with age), the prior value of risk factors (i.e., the person’s health at the beginning of the period), factors affecting a person’s ability to change (e.g., education, income), and exogenous factors (e.g., health programs).

The regression predicts risk factor values as a linear function of current age, risk state, and factors that can be changed to improve health:

In the equation, u₀ is a constant (it may reflect genetic determinants of risk factor level), u_l represents age effects, x_it are risk factor values at time t; d is the parameter that allows diffusion effects, e_it, to change with age. Many risk factors (R) interact and are correlated over time so all risk factors must be included in each equation. For example, smoking may raise blood pressure and reduce lung capacity, but reduce weight. Weight gain may increase blood glucose (a measure of diabetes and a CVD risk factor; Modan et al., 1991, 1992), serum cholesterol, and blood pressure. Thus, it is unrealistic to model changes in one risk factor at a time. For example, the Multiple Risk Factor Intervention Trial (MRFIT) Research Group (1990) found no improvement in mortality in 6.8 years of intervention; by 10.5 years, the medications had been changed and total mortality had dropped significantly. The original antihypertensive drug worsened cardiac arrythmias and adversely altered glucose metabolism, raising certain mortality risks.

We include interactions with age (i.e., x_it · age_it) because risk factor effects may change with age. For example, reducing blood pressure may have adverse effects in an elderly person with preexisting heart disease. The z_it represents other factors such as education, income, employment, and public health efforts at disease prevention. Thus, Equation (2) both de-

scribes risk factor change with time and variables through which one might intervene. Only if intervention variables are modeled can their effects be assessed. This methodology differs from strict “cause elimination” in which all (or a fixed proportion of) deaths from a cause are assumed to be averted by an intervention. Cause elimination indicates the potential effect of intervention. Risk factor regressions forecast the size of the health effect and, possibly, help us understand why some effects fail to materialize.

The last term in Equation (2) represents the interaction of error with age (e.g., there may be less stability of blood pressure in elderly persons (McLean et al., 1992)) and adjusts regressions for age-related heteroscedasticity (systematically unequal variances).

To estimate each regression requires longitudinal data. Though such data are rare in developing countries, often age- and sex-specific distributions of risk factors are available from surveys (e.g., the six countries in Dowd and Manton, 1990; the 45 MONICA sites; or the 13 countries in Knuiman et al., 1982). Initial conditions (i.e., age-specific and sex-specific risk factor means and variances) can be combined with regression coefficients describing risk factor changes estimated from longitudinal studies in developed countries (e.g., Framingham, Kaunas, and Finnish East-West studies; Dowd and Manton, (1992)). In addition, socioeconomic factors may improve a regression’s applicability to the population of interest.

Studies of risk factors in special populations can be used to assess assumptions of physiological invariance of the cross-temporal regressions (e.g., Lackland et al., 1992). Thus, coefficients describing the relation of two physiological variables, if socioeconomic status and demographic factors are controlled, may be less variable across countries than age- and sex-specific risk factor means and variances. Thus, if no longitudinal data exist, regressions estimated from studies in developed countries may be combined with data on age, sex, and socioeconomic population distributions, and age-specific and sex-specific risk factor distributions in the country. Later, we show how life tables are adjusted for country-specific mortality differences. Model assumptions can be verified in cross-national studies (Reed et al., 1991; Choi et al., 1990).

Multivariate hazard functions describe mortality as functions of age and risk factors. The model below is based on the Gompertz function used to describe adult mortality:

where α determines the mortality level (the scale parameter) and θ determines the shape of its age dependence (i.e., θ×100 is the percentage that

mortality increases per year). A θ of 0.0805 means mortality increases 8.05 percent per year. Equation (3) does not include risk factors. We generalized Equation (3) by substituting a scalar function of risk factor values (predicted from the regressions) for α. Thus, instead of a fixed-scale parameter, we use a scalar function of J time-varying risk factors, or

(4)

where a quadratic function of risk factors is substituted for α. In Equation (4) there are terms describing mortality rates independent of age and risk factors (i.e., μ₀, expressed as a rate of deaths per 100,000 persons in a year); linear risk factor effects (i.e., b, a vector containing a linear coefficient for each risk factor scaled as a change in the annual mortality rate); and nonlinear effects (i.e., the matrix B, whose diagonals represent the effect on the mortality rate of risk factor values squared and whose off-diagonals represent the effect on mortality of the pairwise product of risk factors). The T superscript indicates that the vector or matrix should be transposed. In a Gompertz function, α is time invariant. In Equation (4) at each time, a new set of risk factor values (with or without interventions) estimated from the regressions is substituted. Each term is multiplied by the exponential function, representing the percentage increase in mortality per year of age. Thus, the constant can be evaluated for persons aged 45, 60, or 75 by substituting a value for age_it in the exponential and multiplying it by μ₀. This procedure is done for every coefficient in the hazard so that all risk factor effects are age dependent. Thus, we can examine the effect on the annual mortality rate of a risk factor change predicted by the regressions (e.g., a 5 percent reduction in cholesterol over two years due to dietary changes) at specific ages.

The addition of the J time-varying risk factors affects the values of other coefficients (i.e., if risk factors explain some of the age dependence of mortality, θ is correspondingly reduced). With a lot of risk factor information, mortality might be modeled with θ=0 (i.e., no age effect). It is more likely that some age-dependent risk factors are unmeasured, and country-specific unmeasured factors affect the age dependence of mortality. When we do not have measures of the risk factors causing the age dependence of mortality, their effect must be represented by μ₀ (the constant) and θ (i.e., μ₀ and θ represent the average constant and age-dependent effects, respectively, of unobserved risk factors on mortality).

The quadratic form of the hazard allows risk factor functions to be estimated for each cause of death and corresponding coefficients added to obtain a total mortality function:

(5)

Thus, if K=3 (e.g., cancer, heart disease, and other), we estimate three equations like Equation (5) where μ_0k, b_k, B_k differ by cause. They are summed to obtain total mortality. The same risk factor values are used in each equation.

In developing countries, age-specific and sex-specific means and variances of risk factors (e.g., smoking, body weight, blood pressure, cholesterol, blood glucose), estimates of the age dependence of total mortality, and the proportion of all deaths due to (exhaustive and exclusive) causes of death can be combined with regressions and hazard coefficients estimated with data from longitudinal studies in developed countries. The output is a cohort life table that shows, for example, how changing a risk factor affects life expectancy, the risk of death from each cause, and ages at which mortality is affected by direct and indirect risk factor effects. The life table also describes age changes in risk factor means and variances due to the deaths of persons with adverse risk factor values, as well as risk factor means and variances for survivors. Interventions are simulated by changing the initial risk factor distribution (a short-lived effect) or coefficients in the regression or mortality functions.

Forecasts are useful for public health planning in developing countries because adverse lifestyle habits (smoking) and dietary practices (fat consumption) may not have been adopted and may be prevented by central action (e.g., import taxes on cigarettes or alcohol). Clinical and epidemiological studies suggest that early (and low-cost) action on nutrition and physical activity can reduce disease and mortality at later ages (Blair et al., 1989; Lindsted et al., 1991; Paffenbarger et al., 1986; Zimmet et al., 1991).

Many developing countries lack adequate resources to deal with current health problems (e.g., infant and maternal health)—let alone future problems. To prioritize the allocation of scarce resources we must compare diverse, long-range outcomes. Two metrics for comparison are life expectancy (active life expectancy, if adjusted for disability), and changes in risk factors. Both represent health-adjusted “human capital.” Active life expectancy is the productive potential of the population over its remaining lifetime. Risk factors represent the earliest factors in the causal chain leading to disease and disability.

To make decisions about interventions, the costs of different scenarios must be evaluated—possibly assigning different values to each year of life gained or lost. For example, one way to assign value is to see what an average person earns at all ages past t. If a person dies at t then lost wages, representing the value of the worker’s production to the economy (i.e., what society was willing to pay for the labor), represent the indirect cost of

death. Direct costs are the cost of trying to avert death and of the disposition of the individual and his property at death.

There are problems in translating human capital into monetary terms in economies with surplus labor. Human capital may be measured directly by a person’s functional capacity. Such calculations are important in evaluating interventions in groups not active in the labor force (i.e., the young and the elderly). Thus, independent of labor force conditions, human capital is assumed to have intrinsic productive value. The problem is to estimate this intrinsic, but latent, value. The measure considered here is “active” life expectancy—life expectancy decomposed into disability categories.

The regression and hazard coefficients are estimated from longitudinal data in developed countries and combined with country-specific mortality and risk factor data from developing countries to make forecasts. Illustrative life tables and age-specific risk factor values projected from parameters estimated from longitudinal data are given in Table 1.

The survival probability (l_t), life expectancy (e_t), and means (v_t) for eight risk factors are presented for survivors to age t. Estimates of age-specific risk factor means are presented for both independent and dependent elimination of CVD. In independent elimination, the resulting change in overall mortality does not change risk factor means and variances. Dependent elimination means that we calculate cause-specific life tables where risk-factor change and mortality functions interact, i.e., risk factor values for persons who would have died from the eliminated cause of death no longer died and are not subtracted from the risk factor means and variances so that mortality for the non-eliminated causes is increased.

Risk factor means differ over age between the two computations because eliminating CVD reduces mortality selection at later ages. For males at age 90, dependent elimination increases life expectancy 3.1 years (6.0–2.9). Under independence it increases 4.3 years (7.2–2.9). The decrease is 1.2 years less under dependence because persons with adverse risk factor values for CVD (e.g., smokers) now no longer die of CVD but die at later ages from the remaining diseases—if we assume that the eliminated cause had risk factors in common with one or more of the retained causes. Ignoring the effect of CVD mortality on risk factor means and variances overestimates CVD’s effect on life expectancy by 26 percent.

In Tables 2A and 2B (Dowd and Manton, 1990), descriptive and risk factor statistics are presented for six developing countries, along with composite statistics for developed countries. Male life expectancy varies from

49 years in Ghana to 72 years in Cuba. Proportionate mortality for CVD is lowest (21 percent) in Sri Lanka and highest (50 percent) in the developed country composite. Cancer risks are high in Cuba (22 percent; with the most smoking) and Tanzania (18 percent; with the lowest cholesterol). Cholesterol varies from 135 mg/dl in Tanzania to 215 mg/dl in Mauritius. The U.S. National Cholesterol Education Program identified cholesterol levels below 200 mg/dl as desirable, with values between 200 and 239 mg/dl as borderline but not requiring treatment unless other risk factors (e.g., hypertension, obesity, smoking) are present. This borderline area occurs because much of the risk of CVD comes from the effect of jointly elevated risk factors. Per capita smoking varies from 1.4 cigarettes per day in Sri Lanka to 11 per day in Cuba. Body mass index (BMI) varies little (206–220 hectograms per

square meter), but these levels are low (e.g., the average for developed countries is 260). Blood pressure is highest in Ghana (134/82). CVD in developing countries is often due to untreated hypertension

In Table 3 the economic benefits of interventions are illustrated along with changes in total and cause-specific mortality achieved by restricting risk factors to an “ideal” profile. These calculations were carried out for males age 30 years and older in six countries with cross-sectional (age-specific and sex-specific) risk factor and proportional mortality data. The population was projected to extinction by using Equation (1), both with (“ideal”) and without (baseline) interventions in each of the cohort equa-

tions for risk factor changes and mortality used to generate elements of G. The six developing countries have life expectancies from near those of the United States (i.e., Cuba) to those of extremely poor countries. Risk factor effects vary in different country conditions. The greatest gain in median age at death for the ideal risk factor profile is in Tanzania (6.4 years). In Ghana the gain in median age at death is less (4.5 years) but the proportionate gain in GNP is highest, 41.5 percent. This results occurs because the distribution of cause-specific mortality was close to that predicted by the ideal profile (a cancer excess did not exist); the population was younger than in Tanzania so that young persons were affected. Thus, the most important risk factors are associated with differences in the age-specific and sex-specific means and variance of risk factors across countries (e.g., in the least developed countries, cholesterol and BMI are low and blood pressure is elevated; in developed countries the level of detection and treatment of hypertension is high). Economic impact depends on the age profile of mortality (cancer generally affects younger persons than CVD) and the population age distribution (i.e., demographic parameters relevant to labor force activity interact with health parameters in determining economic effects). The ideal risk factor profile improved total and cause-specific mortality for all countries. GNP (converted to U.S. dollars) also increased in all countries.

In evaluating interventions, the gain in life expectancy may not be the most relevant measure. A measure directly representing a person’s physical and mental capacity may be more appropriate. One measure weights health by subjective factors (e.g., quality adjusted life years, QALYs; Wright, 1990). QALYs are criticized because weights are determined subjectively (LaPuma and Lawlor, 1990) and calibrating instruments are not validated (Carr-Hill and Morris, 1991).

A second approach decomposes life expectancy into the time expected to be lived in states defined by level and type of functional ability. To define these states we calculate scores on multiple dimensions of physical and cognitive functioning identified from self-reported items. The scores are then used in the forecasts. Table 4 presents male and female active life expectancy (ALE) estimates based on functional dimensions identified from (a) 27 measures of physical and mental performance for 25,541 respondents to the 1982 and 1984 NLTCS, and (b) 31 health and functioning measures on 16,600 Medicare-eligible, noninstitutionalized U.S. respondents in four sites. By presenting results from two studies, we can compare functional dimensions derived from different survey instruments and different populations (the Medicare sample excluded institutional residents and was not a national sample).

The dimensions of ability/disability are identified by a multivariate procedure (i.e., grade of membership analysis; Woodbury and Clive, 1974; Woodbury et al., 1978) from measures reported with error. If the 27 and 31 measures define a small number of “similar” dimensions, scores may be compared between populations. “Comparison” of dimensions in two populations can be done either by matching the association of each self-reported item with each dimension in the two analyses or by conducting an analysis on data pooled from the two data sets. In the first case, validation depends on the correlation of the dimensions with independent measures (e.g., mortality, acute and long-term care (LTC) service use, age and sex; Manton et al., 1991c). In the second, the analysis is performed by using the missing-information principle (Orchard and Woodbury, 1971) to infer scores for the developing country. The procedure used is robust to differences in sampling; other multivariate procedures such as principle components are not. If dimensions that represent functioning in the combined populations are defined, the disability scores can be compared (Manton et al., 1991a) and used to project change in function.

The dimensions derived from the two analyses were similar. They are less subject to measurement error than the original items; because they are statistically weighted averages, averaging over multiple items by using optimized weights improves reliability. These dimensions have excellent predictive validity for health service use, mortality, and sociodemographic measures in the United States (e.g., Manton and Stallard, 1990, 1991; Manton et al., 1992b). Life tables calculated from the two populations (labeled Medicare and NLTCS) appear in Table 4 (Manton et al., 1991a).

Both the total life expectancy (e.g., male e_t in the NLTCS is 15.4 years at age 65) and the decomposition of e_t into the proportion expected to be lived with given levels and types of impairment are presented. Of the 15.4 years in NLTCS, about 93 percent, or 14.3 years, is expected to be lived free of disability. These people could continue to work or perform social functions (e.g., provide child care within the family or care for other disabled persons). As in the case of risk factors, we can “intervene” in the progression of disability with aging to forecast what life expectancy and costs would be if intervention goals are met.

For the purposes of illustration, we present ALE as estimated for U.S. cohorts representing all ages 65 and above for a given date, for 1990 and 2020.

We conducted two simulations: persons remaining “active” to death, and persons remaining frail to death (both from age 65). In the first case, male life expectancy was 23.6 years and female life expectancy was 33.4 years. For frail males, life expectancy was 5.1 years and for frail females, life expectancy was 7.4 years.

The strategy used for risk factors can be used to estimate ALE for

developing countries when longitudinal data on disability are missing. From cross-sectional surveys of functioning in developing countries (which exist in many cases), we calculate scores specific to age and sex that may be similar in substantive content to scores calculated from longitudinal surveys of disability in developed countries. The age and sex distribution of scores in the developing country can be combined with parameters describing age-specific rates of change in disability, and the dependence of mortality on the scores, estimated from a longitudinal study in a developed country.

A concern is that functional measures may be more culturally dependent than risk factors such as blood pressure (although there are measurement effects for risk factors such as “white coat” hypertension; anxiety over a medical visit also acutely raises hematocrit and cholesterol). The 27 measures in the NLTCS involve well-defined physical tasks (e.g., hold a 10-pound package, climb one flight of stairs, rise from a seated position) and ADLs (Katz and Akpom, 1976; Katz et al., 1983). ADLs are based on a sociobiological model of the acquisition during childhood (and loss with age in reverse order) of biologically fundamental self-maintenance skills (e.g., toileting, bathing, eating, dressing, and walking). Thus, these functions have to be performed in developing countries (they are basic to human existence), but the social and physical environment may make some harder or easier to do. Instrumental activities of daily living (IADLs) also represent tasks necessary for self-maintenance (e.g., laundry, cooking). With appropriate adjustments, ADLs and IADLs could be meaningfully assessed in rural and developing country settings.

We have discussed strategies for combining multiple data sets to forecast morbidity, disability, and mortality, and presented a three-part model based on mortality and cross-sectional risk factor or disability data from a developing country, and longitudinal risk factor and survival parameter estimates from a developed country. This model assumes invariance of health and functional process parameters (but not distributions) across socioeconomic and cultural conditions. Such assumptions are reasonable for physiological risk factors. For functional measures, the validity of the assumption is improved if socioeconomic covariates are available.

A health forecasting model is useful in planning health interventions for developing countries. A model is necessary because successful strategies should consider all elements of a health care system simultaneously—public health efforts at primary prevention; access to primary health care facilities; and modern health care facilities to treat serious diseases. Often strategies in developing countries have emphasized one sector of health care over another at different times. This single-sector, sequential strategy

has seldom proved successful (Cox, 1992). To identify the optimal balance and timing of interventions in different health service sectors requires evaluating multiple demographic, health, and economic parameters. This complexity requires a model to reflect interrelations of factors to determine the most efficient service mix for specific conditions.

Without a dynamic model it is difficult to guage accurately the quantitative effects of intervention. Clearly, reducing smoking has broad and large health benefits. However, there are small (relative to benefits) but real adverse consequences of smoking reduction that must be considered (e.g., persons who quit smoking may gain weight, leading to increases in blood pressure, cholesterol level, and blood sugar). Furthermore, without a model, one will have difficulty in assessing the “true” level of risk for an individual from a single risk factor measurement (Keli et al., 1992). People may be misclassified in risk categories and estimates may be distorted so that choosing between such qualitatively different health strategies as “high-risk” (i.e., screening and then treatment) versus “population-based” (i.e., reducing risk factor levels in the population) approaches is difficult. The evaluation of these two strategies is affected by the form of the risk function and the population distribution of risk factors (Strachan and Rose, 1991). Thus, biologically accurate models are necessary to make good qualitative policy choices.

Finally, models provide tools to monitor population health change and intervention effects in a systematic way. One may initially have to select a strategy using “uncertain” long-range projections. The use of models in surveillance can aid in the management of programs and in the “fine tuning” of interventions by tracking real-time effects. For example, the elderly in the United States and other countries are very health conscious (e.g., the greatest changes in risk factors have been for persons aged 65 to 74). However, U.S. public health interventions have been directed toward avoiding risk factors rather than optimizing nutrition in target groups. Thus, although cholesterol levels declined in the United States between 1977 and 1987, there was little recognition of the special nutritional needs of the elderly—who are often malnourished (Popkin et al., 1992). Thus, health interventions have to be crafted for specific country conditions and for the special health problems of target groups.

Andrews, G.R., A.J.Esterman, A.J.Braunack-Mayer, and C.Runge 1986 Aging in the Western Pacific: A Four-Country Study. Manila: Western Region World Health Organization.

Barker, D.J.P., and C.N.Martyn 1992 The maternal and fetal origins of cardiovascular disease. Journal of Epidemiology and Community Health 46:8–11.

Barker, D.J.P., A.R.Bull, C.Osmond, and S J.Simmonds 1991a Fetal and placental size and risk of hypertension in adult life. British Medical Journal 301:259–262.

Barker, D.J.P, K.M.Godfrey, C.Fall, C.Osmond, P.D.Winter, and S.O.Shaheen 1991b Relation of birth weight and childhood respiratory infection to adult lung function and death from chronic obstructive airways disease. British Medical Journal 303:672–675 .

Barker, D.J.P., K.M.Godfrey, C.Osmond, and A.Bull 1992a The relation of fetal length, ponderal index and head circumference to bloodpressure and the risk of hypertension in adult life. Pediatric and Perinatal Epidemiology 6:35–44.

Barker, D.J.P., T.W.Meade, C.H.D.Fall, A.Lee, C.Osmond, K.Phipps, and Y.Stirling 1992b Relation of fetal and infant growth to plasma fibrinogen and factor VII concentrations in adult life. British Medical Journal 304:148–152.

Barthel, J.S., T.U.Westblom, A.D.Havey, F.Gonzalez, and E.D.Everett 1988 Gastritis and Campylobacter pylori in healthy, asymptomatic volunteers. Archives of Internal Medicine 148:1149–1151.

Bharucha, N.E., E.P.Bharucha, A.E.Bharucha, A.V.Bhise, and B.S.Schoenberg 1988 Prevalence of stroke in the Parsi community of Bombay. Stroke 19(1):60–62.

Blair, S.N., H.W.Kohl, R.S.Paffenbarger, D.G.Clark, K.H.Cooper, and L.W.Gibbons 1989 Physical fitness and all-cause mortality: A prospective study of healthy men and women. Journal of the American Medical Association 262:2395–2401.

Blaser, M.J. 1988 Type B gastritis, aging, and Campylobacter pylori. Archives of Internal Medicine 148:1021–1022.

Carr-Hill, R.A., and J.Morris 1991 Current practice in obtaining the “Q” in QALYs: A cautionary note. British Medical Journal 303:699–701.

Chen, Z., R.Peto, R.Collins, S.MacMahon, J.Lu, and W.Li 1991 Serum cholesterol concentration and coronary heart disease in population with low cholesterol concentrations. British Medical Journal 303:276–282.

Chiang, C.L. 1968 Introduction to Stochastic Processes in Biostatistics. New York: John Wiley & Sons.

Choi, E.S., R.B.McGandy, G.E.Dallal, R.M.Russell, R.A.Jacob, E.J.Schaefer, and J.A. Sadowski 1990 The prevalence of cardiovascular risk factors among elderly Chinese Americans. Archives of Internal Medicine 150:413–420.

Cowan, L.C., D.O’Connell, M.H.Criqui, et al. 1990 Cancer mortality and lipid and lipoprotein levels. American Journal of Epidemiology 131(3):468.

Cox, P.S.V. 1992 Prevention vs cure in developing countries: The pendulum syndrome. Lancet 339:664–665.

Crow, Y.J., K.G.M.Alberti, and J.M.Parkin 1991 Insulin dependent diabetes in childhood and maternal deprivation in northern England, 1977–86. British Medical Journal 303:158–160.

Dodu, S.R.A. 1988 Emergence of CVD in developing countries. Cardiology 75:56–64.

Dowd, J.E., and K.G.Manton 1990 Forecasting chronic disease risks in developing countries. International Journal of Epidemiology 19(4):1018–1036.

1992 Projections of disability consequences in Indonesia. Journal of Cross-Cultural Gerontology 7(3):237–258.

Evans, J. 1990 The economic status of older men and women in the Javanese household and the influence of this upon their nutritional level. Journal of Cross-Cultural Gerontology 5:217–242.

Forman, D. 1991 Helicobacter pylori infection: A novel risk factor in the etiology of gastric cancer. Journal of the National Cancer Institute 83(23):1702–1703.

Forman, D., F.Sitas, N.G.Newell, A.R.Stacey, J.Boreham, R.Peto, T.C.Campbell, J.Li, and J.Chen 1990 Geographic association of Helicobacter pylori antibody prevalence and gastric cancer mortality in rural China. International Journal of Cancer 46:608–611.

Hamm, P., R.B.Shekelle, and J.Stamler 1989 Large fluctuations in body weight during young adulthood and twenty-five-year risk of coronary death in men. American Journal of Epidemiology 129:312–318.

Hoem, J.M. 1985 Weighting, misclassification and other issues in the analysis of survey samples of life histories. Pp. 249–293 in J.J.Heckman and B.Singer, eds., Longitudinal Analysis of Labor Market Data. Cambridge, Mass.: Cambridge University Press.

1989 The issue of weights in panel surveys of individual behavior. Pp. 539–559 in D. Kasprzyk, G.Duncan, G.Kalton, and M.P.Singh, eds., Panel Surveys. New York: John Wiley & Sons.

Isles, C.G., D.J.Hole, C.R.Gillis, V.M.Hawthorne, and A.F.Lever 1989 Plasma cholesterol, coronary heart disease, and cancer in the Renfrew and Paisley survey. British Medical Journal 298:920–928.

Jacques, P.F., S.C.Hartz, R.B.McGandy, R.A.Jacob, and R.M.Russell 1987 Ascorbic acid, HDL, and total plasma cholesterol in the elderly. Journal of the American College of Nutrition 6:169–174.

Katz, S., and C.A.Akpom 1976 A measure of primary sociobiological functions. International Journal of Health Services 6:493–508.

Katz, S., L.G.Branch, M.H.Branson, J.A.Papsidero, J.C.Beck, and D.S.Greer 1983 Active life expectancy. New England Journal of Medicine 309:1218–1223.

Keli, S., B.Bloemberg, and D.Kromhout 1992 Predictive value of repeated systolic blood pressure measurements for stroke risk. Stroke 23(3):347–351.

Kesteloot, H., D.X.Huang, X.S.Yang, J.Claes, M.Rosseneu, J.Goboers, and J.V.Joossens 1985 Serum lipids in the People’s Republic of China: Comparison of western and eastern populations. Arteriosclerosis 5:427–433.

Kimura, N., H.Toshima, Y.Nakayama, K.Takayama, H.Tashiro, and M.Takagi 1979 Fifteen-year follow-up population survey on stroke: A multivariate analysis of the risk of stroke in farmers of Tanushimaru and fisherman of Ushibuka. Pp. 505–510 in Y.Yamori, W.Lovenberg, and E.D.Freis, eds., Prophylatic Approach to Hypertensive Diseases. New York: Raven Press.

Knuiman, J.T., C.E.West, and J.Burema 1982 Serum total and high density lipoprotein cholesterol concentrations and body mass

index in adult men from 15 countries. American Journal of Epidemiology 116(4):631–642.

Lackland, D.T., J.J.Orchard, J.E.Keil, D.E.Saunders, F.C.Wheeler, L.L.Adams-Campbell, R.H.McDonald, and R.G.Knapp 1992 Are race differences in the prevalence of hypertension explained by body mass and fat distribution? A survey in a biracial population. International Journal of Epidemiology 21:236–245.

Lam, K.M.C., N.Syed, H.Whittle, and D.H.Crawford 1991 Circulating Epstein-Barr virus-carrying B cells in acute malaria. Lancet 337:876–878.

LaPuma, J., and E.F.Lawlor 1990 Quality-adjusted life-years. Journal of the American Medical Association 263:2917–2921.

Lee, I-M., R.S.Paffenbarger, and C-C Hsieh 1991 Physical activity and risk of developing colorectal cancer among college alumni. Journal of the National Cancer Institute 83:1324–1329.

Lindsted, K.D., S.Tonstad, and J.W.Kuzma 1991 Self-report of physical activity and patterns of mortality in Seventh-Day Adventist men. Journal of Clinical Epidemiology 44:355–364.

Manton, K.G., H.Malker, and B.Malker 1986 Comparison of temporal changes in U.S. and Swedish lung cancer, 1950–51 to 1981–82. Journal of the National Cancer Institute 77(3):665–675.

Manton, K.G., G.C.Myers, and G.Andrews 1987 Morbidity and disability patterns in four developing nations: Their implications for social and economic integration of the elderly. Journal of Cross-Cultural Gerontology 2:115–129.

Manton, K.G., and E.Stallard 1990 Changes in health functioning and mortality. Pp. 140–162 in S.Stahl, ed., The Legacy of Longevity: Health and Health Care in Later Life. Newbury Park, Calif.: Sage Publications.

1991 Cross-sectional estimates of active life expectancy for the U.S. elderly and oldest-old populations. Journal of Gerontology 48:S170–182.

Manton, K.G., J.E.Dowd, and M.A.Woodbury 1991a Methods to identify geographic and social clustering of disability and disease burden. Pp. 61–84 in A.G.Hill, and J.Caldwell, eds., Measurement of Health Transition Concepts. Canberra: Australian National University Press.

Manton, K.G., E.Stallard, and S.Wing 1991b Analyses of black and white differentials in the age trajectory of mortality in two closed cohort studies. Statistics in Medicine 10:1043–1059.

Manton, K.G., E.Stallard, and M.A.Woodbury 1991c A multivariate event history model based upon fuzzy states: Estimation from longitudinal surveys with informative nonresponse. Journal of Official Statistics 7:261–293.

Manton, K.G., M.A.Woodbury, and E.Stallard 1991d Statistical and measurement issues in assessing the welfare status of aged individuals and populations. Journal of Econometrics 50:151–181.

Manton, K.G., E.Stallard, and B.Singer 1992a Sources of uncertainty in projecting the size and health status of the U.S. elderly population. International Journal of Forecasting 8(3):433–458.

Manton, K.G., M.A.Woodbury, E.Stallard, and L.S.Corder 1992b The use of Grade of Membership techniques to estimate regression relationships. In P.Marsden, ed., Sociological Methodology. London: Blackwell.

Manton, K.G., J.E.Dowd, and E.Stallard 1993 The effects of risk factors on male and female cardiovascular risks in middle and late age. In K.G.Manton, B.Singer, and R.Suzman, eds., Forecasting the Health of Elderly Populations. New York: Springer-Verlag.

McGill, H.C. 1988 Cerebral artery atherosclerosis and diet. Stroke 19(7):801.

McLean, K.A., P.A.O’Neill, I.Davies, and J.Morris 1992 Influence of age on plasma osmolality: A community study. Age and Ageing 21:56–60.

McMurray, M.P., M.T.Cerqueira, S.L.Connor, and W.E.Connor 1991 Changes in lipid and lipoprotein levels and body weight in Tarahumara Indians after consumption of an affluent diet. New England Journal of Medicine 325(24):1704–1708.

Modan, M., S.Almog, Z.Fuchs, A.Chetrit, A.Lusky, and H.Halkin 1991 Obesity, glucose intolerance, hyperinsulinemia, and response to antihypertensive drugs. Hypertension 17(4):565–573.

Modan, M., J.Or, A.Krasik, Y.Drory, Z.Fucha, A.Lusky, A.Chetrit, and H.Halkin 1992 Cardiovascular disease in men. Circulation 85(3):1220.

Mohr, G.C., D.Kritz-Silverstein, and E.Barrett-Connor 1991 Plasma lipids and gallbladder disease. American Journal of Epidemiology 134(1):78–85.

Mozar, H.N., D.G.Bal, and S.A.Farag 1990 The natural history of atherosclerosis: An ecologic perspective. Atherosclerosis 82:157–164.

Multiple Risk Factor Intervention Trial (MRFIT) Research Group 1990 Mortality rates after 10.5 years for participants in the multiple risk factor intervention trial. Journal of the American Medical Association 263:1795–1801.

Muscari, A., C.Bozzoli, G.M.Puddu, C.Rovinetti, G.P.Fiorentini, R.A.Roversi, and P. Puddu 1990 Correlations between serum lipids and complement components in adults without demonstrated atherosclerotic disease. Atherosclerosis 81:111–118.

National Center for Health Statistics 1987 The Supplement on Aging to the 1984 National Health Interview Survey. In J. Fitti and M.G.Kovar, eds., Vital and Health Statistics. Series 1, No. 21. DHHS Pub. No. (PHS) 87–1323, Public Health Service. Washington, D.C.: U.S. Government Printing Office.

Nomura, A.M.Y., G.N.Stemmermann, and P-H.Chyou 1991 Prospective study of serum cholesterol levels and large-bowel cancer. Journal of the National Cancer Institute 83:1403–1407.

Orchard, R., and M.A.Woodbury 1971 A missing information principle: Theory and application. Pp. 697–715 in L.M. LeCam, J.Neyman, and E.L.Scott, eds., Proceedings Sixth Berkeley Symposium on Mathematical Statistics and Probability. Berkeley: University of California Press.

Paffenbarger, R.S., R.T.Hyde, A.L.Wing, and C.C.Hsieh 1986 Physical activity, all-cause mortality, and longevity of college alumni. New England Journal of Medicine 314:605–613.

Parsonnet, J., G.D.Friedman, D.P.Vandersteen, Y.Chang, J.H.Vogelman, N.Orentreich, and R.K.Sibley 1991 Helicobacter pylori infection and the risk of gastric carcinoma. New England Journal of Medicine 325(16):1127–1131.

Popkin, B.M., P.S.Haines, and R.E.Patterson 1992 Dietary changes in older Americans, 1977–1987. American Journal of Clinical Nutrition 55:823–830.

Prevot, S., S.Hamilton-Dutoit, J.Audoouin, P.Walter, G.Pallesen, and J.Diebold 1992 Analysis of African Burkitt’s and high-grade B cell non-Burkitt’s lymphoma for Epstein-Barr virus genomes using in situ hybridization. British Journal of Haematology 80:27–32.

Reed, D.M., J.A.Resch, T.Hayashi, C.MacLean, and K.Yano 1988 A prospective study of cerebral artery atherosclerosis. Stroke 19:820–825.

Reed, T., J.Quiroga, J.V.Selby, D.Carmelli, J.C.Christian, R.R.Fabsitz, and C.E.Grim 1991 Concordance of ischemic heart disease in the NHLBI twin study after 14–18 years of follow-up. Journal of Clinical Epidemiology 44:797–805.

Schwartzkopff, W., J.Schleicher, I.Pottins, S-B. Yu, C-Z Han, and D-Y.Du 1990 Lipids, lipoproteins, apolipoproteins and other risk factors in Chinese men and women with and without myocardial infarction. Atherosclerosis 82:253–259.

Strachan, D., and G.Rose 1991 Strategies of prevention revisited: Effects of imprecise measurement of risk factors on the evaluation of “high-risk” and “population-based” approaches to prevention of cardiovascular disease. Journal of Clinical Epidemiology 44(11):1187–1196.

Sullivan, J.L. 1989 The iron paradigm of ischemic heart disease. American Heart Journal 117:1177–1188.

Talley, N.J., A.R.Zinsmeister, A.Weaver, E.P.DiMagno, H.A.Carpenter, G.I.Perez-Perez, and M.J.Blaser 1991 Gastric adenocarcinoma and helicobacter pylori infection. Journal of the National Cancer Institute 83:1734–1739.

Tertov, V.V., A.N.Orekhov, K.S.Sayadyan, S.G.Serebrennikov, A.G.Kacharava, A.A.Lyakishev, and V.N.Smirnov 1990 Correlation between cholesterol content in circulating immune complexes and atherogenic properties of CHD patients’ serum manifested in cell culture. Atherosclerosis 81:183–189.

Tonglet, R., P.Bourdoux, T.Minga, and A-M.Ermans 1992 Efficacy of low oral doses of iodized oil in the control of iodine deficiency in Zaire. New England Journal of Medicine 326:236–241.

Trowell, H.C., and D.P.Burkitt 1991 Western Diseases: Their Emergence and Prevention. Cambridge, Mass.: Harvard University Press.

Wilson, A.O., D.J.Adamchak, and A.C.Nyanguru 1991 A study of well-being of elderly people in three communities in Zimbabwe. Age and Ageing 20:275–279.

Woodbury, M.A., and J.Clive 1974 Clinical pure types as a fuzzy partition. Journal of Cybernetics 4:111–121.

Woodbury, M.A., J.Clive, and A.Garson 1978 Mathematical typology: A grade of membership technique for obtaining disease definition. Computers and Biomedical Research 11:277–298.

World Health Organization 1987 World Health Statistics Annual 1987. Geneva: World Health Organization.

Wright, L. 1990 The role of aging in establishing social priorities: An economic perspective. Pp. 647–658 in R.L.Kane, J.G.Evans, and D.MacFadyen, eds., Improving the Health of Older People: A World View. Oxford: World Health Organization and Oxford University Press.

Zimmet, P.Z., V.R.Collins, G.K.Dowse, G.M.M.Alberti, Z.Tuomilehto, H.Gareeboo, and P. Chitson 1991 The relation of physical activity to cardiovascular disease risk factors in Mauritians. American Journal of Epidemiology 134:862–875.