Setting

Matlab is a rural area about 55 km from the Dhaka, the capital of Bangladesh. ICDDR,B has long maintained a research station that sponsors the HDSS and extensive research on the health of a population of over 200,000 people. It also provides health services that supplement the government system of health care. The overwhelming majority of older individuals live with adult children (mostly sons), and alternative sources of support—financial and otherwise—outside the family are scarce. Bangladesh per capita income in 2000 was US$370/year. The predominant occupation for rural men is agriculture, with labor force participation rates remaining very high even for older men. Women are largely restricted by convention to activities in the home, with relatively little opportunity to venture outside the homestead, although these restrictions have decreased in recent years. Given the high level of poverty and the scarcity of health providers (4,671 persons/physician, 3,312 persons/hospital bed), contact with the formal health care system is thought to be relatively infrequent. National life expectancy in 2000 was 58 for both men and women; at age 50 it was approximately 30 years, again, with no significant gender differences. Approximately 10 percent of the population was over age 50 (Bangladesh Bureau of Statistics, 2002).

Sampling

The HDSS provided the sampling frame for the MHSS (Rahman et al., 1999). The Matlab surveillance area consists of 8,640 baris or residential compounds, of which roughly one-third (31.1 percent) or 2,687 baris were randomly selected. The bari is the basic unit of social organization in rural Bangladesh and in Matlab in particular (Aziz, 1979; Rahman, 1986). Baris usually consist of a cluster of households linked in many instances in a kin network (however, about 16 percent of baris consist of only a single household and, even in multihousehold baris, kin networks may exist only for subclusters of households). Sampling baris rather than households provides a better representation of family networks, a major focus of the MHSS survey. In single-household baris, that household was selected for detailed interviews. In all other baris, two households were selected. All individuals in these households who were age 50 and over were interviewed. For those below age 50, certain criteria were followed to reduce the interviewing load vis-à-vis large households. Probability weights reflecting this multistage sampling scheme were assigned to each individual and are included in the MHSS public-use data set (Rahman et al., 1999; Rahman, Menken, Foster, and Gertler, 2001).

There were 11,150 individuals age 15 and over in the bari sample. People were eliminated from this analysis (see Table 10-1) if the individual probability weight could not be calculated (54) or the MHSS identification number was erroneous and could not be matched in the HDSS (165). Thus we begin with 10,931 respondents who could be followed in the HDSS. Individuals were dropped if information was missing on self-reported activities of daily living, the major disease index, or observed activities of daily living, or if BMI was either missing or out of range. Most nonresponses stem from the fact that there were special teams for collection of observed health indicators and for anthropometric measurement. They visited only subsets of all MHSS households. Very few respondents refused to participate. Thus for the purposes of this study, we focus on 9,536 respondents age 15 and older (4,399 men, and 5,137 women) for whom we have complete information. Rahman and Barsky (2003) found little difference between individuals who had complete information on health indicators and all respondents.

TABLE 10-1

Sample Exclusions.

Mortality Analytic Plan

Longitudinal models of mortality subsequent to MHSS fieldwork were estimated to investigate the predictive power of the health measures. Each respondent contributed an observation for each person-year from 1996 to 2000 unless censored by death or out-migration. The logistic hazards model of mortality tested has the following form:

where µ_ik is the risk that person i who survived to the start of follow-up year k dies in that year, Female_i is 1 if person i is female and 0 if male, Age_ikis person i’s age at the start of follow-up year k, SR_ij is the jth of J self-reported health indicators, and OBS_im is the mth of M observed health indicators for person i. The health indicators were recorded once—in the MHSS. Therefore, Age is the only time-varying variable.

Estimation of models incorporates a Huber-White correction for intracluster correlation in the distribution of the probability of mortality across observation years for a given respondent. Results presented in the tables display the relative odds associated with a variable, or the relative change in

resulting from a one-unit change in a variable.

Table 10-2 presents the number of MHSS respondents by sex and broad age group and the number of deaths recorded for each sex/age group in the HDSS over the subsequent 5 years. Deaths of 10.7 percent of those over age 50 in the MHSS were recorded in the HDSS within the follow-up period, compared with only 0.7 percent of those under age 50 and 3.7 percent overall. These figures, however, do not adjust for out-migration (censoring). Among those over 50 in the MHSS, only 3.2 percent out-migrated during the follow-up period, compared with 15.1 percent of those under age 50 and 11.5 percent of all respondents (not shown). Out-migrants were censored at the end of the last full year of their observation. Those who died contributed a person-year for each year in which they were alive at the start. Those for whom neither death nor out-migration was recorded in the HDSS in the follow-up period each contributed 5 person-years of observation. The number of person-years of observation in which age was under 50, over 50, and all ages at the start of the year is given in Table 10-2 along with the yearly death rates (deaths/person-years). This death rate was 0.8 percent for all ages. While there was a relatively high rate in the older group (2.0 percent), death was a quite rare event in the younger subsample (0.1 percent). For this reason, all mortality analyses were replicated using the Rare Events Logistic Regression package for STATA (King, 2004). In all cases, rare event models resulted in higher coefficient estimates and lower standard errors, so we choose to report the more conservative logistic regression results in our later tables. Table 10-2 also shows lower mortality rates for women than for men; however, as is shown later, with more precise controls for age, sex differences are not significant.

TABLE 10-2

Respondents, Deaths, Person-Years, and Death Rates.

Extent to Which the Various Health Indicators Are Interrelated

Bivariate and multiple correlations (when each indicator is estimated from the other four) are shown in Table 10-3. Clearly, while these indicators are related, none is very closely determined by the others. These results are confirmed in a principal component analysis (not shown). In that analysis, the first component, which accounts for close to 50 percent of the variation, shows approximately equal weightings from all five indicators. Yet it is not until four factors are included that the variation accounted for reaches 89 percent. Thus, while the indicators are interrelated, each has a sizeable component unrelated to the others.

TABLE 10-3

Interrelationships Between Health Measures: 1996 Matlab Health and Socioeconomic Survey.

Mortality in the Five Years Subsequent to the MHSS in Relation to the Health Indicators

Health status, as measured by each of the five indicators treated dichotomously, is substantially related to subsequent mortality (Table 10-4). Of those who self-reported poor health, 10.2 percent died versus 1.9 percent of those who self-reported good or fair health. ADL scores, whether self-reported or observed, showed greater differentials in outcome by health status, while major disease and BMI indicators differentiated risk of dying less well. Figure 10-3 works backward from survival outcome. It provides evidence that those who died, compared with those who survived the five-year follow-up period, were much more likely to have been in poor health at the time of the MHSS. Using SRH, 53 percent of those who died had said they were in poor health, compared with only 16 percent of those who survived. For reported ADL limitations, 72 percent of those who died had reported some limitation, compared with 19 percent of those who lived. For observed ADL limitations, less common overall, 31 percent of those who died had been observed to be mobility-impaired, compared with only 5 percent of those who survived. The greatest difference clearly was in the reported ADL measure.¹

TABLE 10-4

Subsequent Mortality by Health Status.

FIGURE 10-3

Health status indicators by subsequent survival. SRH = self-reported health; ADL = activities of daily living; BMI = body mass index.

Figure 10-4 demonstrates the predictive power of the ADL scores— which characterize people along a continuum rather than a simple dichotomy. The likelihood of dying over the 5-year follow-up period shows a generally monotonic increase with ADL score, whether self-reported or observed. Furthermore, the two measures appear to capture slightly different mortality trajectories. Self-reported ADL shows a greater increase in mortality as it moves from representing people who can do all tasks easily (score = 0) to those who score difficulty with only one or two. It may be that individuals recognize the level or change in level more acutely than can be picked up on the kinds of measures that went into the observed ADL score. Observed ADL scores may capture extreme levels of poor health better: respondents who scored five or more experienced a subsequent mortality risk that approached 30 percent.

FIGURE 10-4

Proportion dying by self-reported and observed ADL scores.

Multivariate Models

We turn to multivariate models of mortality to address several questions. First, which indicator adds greatest explanatory power when age and sex are taken into account? Second, does the predictive power of these health measures differ for older and younger adults? Initial models predicting the odds of dying are shown in Table 10-5 for the total sample of person-years, person-years for which the individual was under 50 at the start of the year, and person-years for which the individual was 50 or over at the start. The first line reports on the model in which age is the only predictor. The pseudo-R² is .18 for total person-years. Interestingly, age is not a significant predictor of mortality in adults under 50. Adding sex does not significantly improve the predictive power of the model, as shown in line 2. Therefore, once age-in-years is in the model, no sex difference is detected. The remaining lines show the improvement in pseudo-R² when a single health indicator is added. For total person-years, SRH adds .01 to the pseudo-R² and those with poor SRH have odds of dying 2.28 times the odds of those of the same age and sex who do not report poor SRH. In all three models, poor SRH is significantly associated with increased subsequent mortality. A much larger coefficient is estimated for the young subsample of person-years, in which there is a low baseline level of mortality. The yearly probability of dying for this subsample is 0.12 percent (Table 10-2). Poor SRH in the young group (OR = 5.6), when age and sex are held constant at their means, raises the probability of death from 7 to 41 in 10,000. This is a large effect, especially considering that it results from a single question. For the over-50 subsample, the effect of poor SRH is smaller but highly significant (OR = 2.1) and the probability of death, again estimated at the means of age and sex, rises from 1 to 2 percent. While each indicator is significant in the total and age 50 and over models, only the self-reports—whether of general health or ADLs or major disease—add predictive power for younger people.

TABLE 10-5

Logistic Regression Models of Mortality Adding Age, Sex, and One Health Measure.

Stepwise regressions were carried out (Table 10-6) in which age and sex were entered first and then health indicators entered in the order of the magnitude of their additional predictive power. For total person-years and the over-50 subsample, observed ADLs entered first, followed by self-reported ADLs, BMI, and SRH. The index of major disease did not enter at all. For the under-50 subsample, only self-reported ADLs served to predict mortality, whereas for those over 50, all indicators except major disease increased the fit to the data.

TABLE 10-6

Stepwise Logistic Regression Models of Mortality Adding Health Measures to Age and Sex.

Finally, the complete models, in which age, sex, and the four indicators that were significant in any of the Table 10-6 models are included, are shown in Table 10-7. For the total person-years and over-50 models, mortality increases significantly with age, with each additional year associated with a 6-7 percent increase in the odds of dying. Controlling for age and health, the odds of dying for a woman in the same samples are less than half (.44-.46) those of a man. The sizable advantage in mortality controlling for health that women have is substantial enough that, despite having substantially worse health on all observed and self-reported health indicators, their mortality (adjusted only for age and shown in Table 10-4) was not significantly different from men’s.

TABLE 10-7

Logistic Regression Model of Mortality with Age, Sex, and Health Measures as Predictors.

This disjunction between sex differences in health and mortality introduces the possibility that, as in other studies, women may overreport poor health and greater disability. If this were the case, the inclusion of interactions between sex and each of the significant health measures included in Table 10-7 would show significant negative interactions between health measures and the female main effect, indicating that the health measure was a worse predictor of health for women. Yet when sex−health measure interactions were entered, either individually or as a group, none approached statistical significance (models not shown, see Kuhn, Rahman, and Menken, 2004).

The absence of any significant interactions between the respondent’s sex and the predictive power of SRH measures is an extremely robust finding. The effects do not even approach significance, and these results do not change when nonsignificant interaction terms for sex and observed health measures are dropped (models not shown). Women’s adjusted risk of mortality, controlling for other factors, is much lower than men’s. Women’s low baseline mortality risks are multiplied by risks due to poor health, which are more common among women; the result is overall equality in mortality of men and women. The process is not driven by a less substantial relationship between women’s poor health and women’s survival, but by the fact that the same poor health multiplier effect is applied to a lower baseline mortality effect. This result holds in rare events logistic regression models and probit regression models, both of which also apply the very realistic assumption that mortality risks are multiplicative in nature, not additive. These models were also run using a linear probability model in order to estimate the size of interaction effects if mortality risks were indeed additive. Interactions between sex and SRH were found to be highly nonsignificant in these models as well.

Discussion

This analysis demonstrates both the multidimensional nature of the relatively simple and low-cost health measures in the MHSS and their substantial value and validity in a low-income setting such as Bangladesh. We believe this is the only study to examine subsequent mortality in relation to these multiple measures of health. Each measure is a significant predictor of subsequent mortality, particularly for elders. With the exception of major disease, each appears to capture, at least in part, a different dimension of health from the others. These findings support the notion that individuals can effectively assess their own health status even in settings of poor education and low levels of interaction with modern health systems.

Furthermore, the measures seem equally appropriate for men and women. The higher age-specific prevalence of poor health, whether reported or observed, among women does not appear to result from their overreporting poor health compared with men. The proportional increase of mortality for those whose health was poor versus not poor, whichever the health measure used, did not differ for women and men. At the same level of health, women had lower mortality risks than men; however, these lower risks were combined with much higher rates of poor health. These differences countered one another; as a result, we found no sex differences in overall age-adjusted mortality.

Objective health measures are frequently assumed to be superior to self-reports because they minimize concerns over reporting bias and international comparability. We find, however, that they are not necessarily better predictors of mortality. The self-reported and observed measures were similar in predicting mortality at older ages, while self-reports were the only significant predictors of mortality at younger ages.

The marginal time and financial costs of collecting each measure vary substantially. Cost estimates based on the MHSS are given in Table 10-8, along with the improvement in mortality prediction found when each health measure was added to the model based on age and sex (repeated from Table 10-5). For comparison, we consider the high expected costs of drawing and analyzing a blood sample. For the second round of the MHSS, we have proposed a battery of seven biomarker analyses.² The estimated costs are included in Table 10-8. Self-reported measures take less survey time than observational tests or drawing a blood sample and incur no additional capital costs; in addition, they can be collected by less skilled interviewers. The observed indicators cost at least 2.5 times more to collect than the self-reports. The current cost of collecting biomarkers is extremely high, but is likely to decrease in the future for two reasons. First, as collection is standardized, less expensive technology may be developed. Second, with continuing research, a small set of essential indicators is likely to be agreed upon.

TABLE 10-8

Cost of Health Measures and Increase in Pseudo-R².

These results raise questions regarding a minimal set of survey-based health measures that ensures precision and accordance with international standards. The least expensive measure, a single SRH question requiring $0.01 per respondent, adds significantly in predicting subsequent mortality for respondents of all ages. In fact, it predicts mortality better than a number of more complex health indicators.

The self-reported ADL index, the next least expensive indicator to collect, adds even more to the predictive power of our models. This indicator seems to work at least as well, and possibly better, than observed ADLs and anthropometric measures that are far more demanding of resources. However, as for SRH, it is important to ask whether self-reported ADLs produce results that are truly comparable across different populations and subgroups within populations. The results here indicate they are promising enough to be included in studies in other countries in order to assess their usefulness.

The detailed questions on specific conditions that went into the major disease index were far more complicated to implement than the ADL questions and the index added no explanatory power to the analysis of mortality. It may be that the questions on specific conditions are better predictors than the overall index; we plan to address this issue in further analyses of the MHSS.

The observed ADL index and BMI add significantly in predicting mortality for elders. They are not significant for those under 50. It may be that the activities that go into the MHSS ADL index are not sufficiently difficult to detect disabilities that contribute to higher short-term mortality risk in younger adults. Neither the observed ADL index nor BMI picks up the increase in poor health with age for those under 50, especially for females, that is self-reported. In contrast, SRH and the self-reported ADL index increase with age and are significant mortality predictors, perhaps because younger adults recognize change in their own abilities or differences in comparison with peers even when their disabilities do not put them in a range considered unusual by the observer. We suspect that, unless the set of tasks is changed to detect more subtle disability, observed ADL measures for younger adults should be included in survey data collection only as a baseline for longitudinal follow-up rather than as valid health indicators.

The four measures discussed here—SRH, self-reported and observed ADLs, and BMI—are strong predictors of survival over the next 5 years in the rural Bangladesh context, although the observed measures work best for elders. They are therefore good markers of general health and the effects of the health system and can be used to examine basic gender and socioeconomic differentials in that society. The high prevalence of poor health based on the self-reports, even at low ages, in Bangladesh makes these measures perhaps even more valuable than they would be in more advanced societies. More importantly, compared with complex measures such as biomarkers or disability adjusted life years, these measures can be obtained more frequently and in much larger samples for the same cost. If resources are especially constrained, it may be possible to restrict the set of measures collected further and still be able to capture levels, differentials, and trends in overall health over time.

Given the urgent need to assess health and track its trends over time in sub-Saharan Africa, we strongly recommend including a range of health indicators in data collection—in a way that permits research on indicators in this region. Studies are needed that collect, in selected settings, simple and low-cost health indicators like those examined in this paper; expanded versions of these indicators (e.g., SRH with more categories); other objective measures (e.g., lung capacity, grip strength) that are more costly in terms of survey-team time, equipment, and training; and selected biomarkers that are far more expensive to collect and more difficult to analyze. It would also be appropriate to address issues of comparability of measures across population through techniques such as calibrating vignettes. Simultaneous collection—in a few selected studies—of all of these indicators in comparable manner will permit the kind of “head-to-head” analysis performed here for one country and for the measures included in the MHSS. This research would be most informative if carried out where it is possible to follow respondents for some period of time to observe survival outcomes.

The expected outcomes of this research include

• Identification of measures that provide valid information on prevalence of specific conditions or diseases and those which provide information on overall health.
• Estimates of costs and difficulties of collecting each indicator.
• Cost-benefit analyses, which would help investigators decide on the package of indicators to be included in future data collection.

One obvious locus for this type of research is the INDEPTH network. Its Demographic Surveillance System sites offer crucial advantages. Since the sites carry out continuous longitudinal follow-up of large populations, they provide data on mortality trends prior to collection of the health measures. While few sites have been operating as long as the Matlab HDSS (which is a member of INDEPTH), a number, such as Agincourt in South Africa and Navrongo in Ghana, have been operating for at least 10 years. A set of health indicators measured in a one-time survey would contribute to the assessment of measures described in this paper. INDEPTH sites routinely collect the follow-up data needed to study relationships of health measures to subsequent mortality. In addition, if these measures were included periodically in subsequent DSS rounds, their sensitivity to trends and emerging differentials in health could be assessed.

The issue of comparability between populations could be addressed in two ways—for which DSS sites are again especially appropriate. Following the work on vignettes (see Salomon et al., 2004), calibrating vignettes could be included in at least one DSS round that includes health indicators. These vignettes could be extended to include ADLs as well as other aspects of health for which self-reports are collected. In addition, mortality prediction across sites could be analyzed to determine if, as for men and women in Bangladesh, the increase in the odds of mortality for those in poor health as measured by a specific indicator is the same across countries.

Development of low-cost and informative health status modules could be guided by this research. They could be regularly included in DSS data collection rounds, providing valuable comparable information across sites on health levels and trends over time. These modules could also be included in standalone surveys covering larger geographic areas. Nationally representative surveys could then provide the basis for comparative studies of countries, with data from INDEPTH sites providing greater detail and longitudinal validation.

In addition to tracking changes in health over time in defined populations, DSS sites have demonstrated they are ideal settings for conducting and evaluating health and social interventions intended to improve health either directly (e.g., family planning programs, oral rehydration programs) or indirectly (education of women, micro-credit programs). They offer well-established sampling frames with regular follow-up so that program and control groups can be established and compared over time. Increasingly, programs include direct adult health interventions, such as Agincourt’s South African Stroke Prevention Initiative. Regular observation of simple health measures can provide a baseline for assessing differences between experimental and control groups, and follow-up information for assessing the long-term impact of interventions on comparable health status measures. At the same time, repeated observation of health status allows researchers in DSS sites to measure the health impact of programs not aimed directly at health, including education, family planning, and women’s status interventions. Finally, such measures facilitate analysis of the health effects of unexpected economic or political crises and other changing social or health patterns.

As societies in sub-Saharan Africa grow older and continue to absorb the consequences of the HIV/AIDS epidemic, it will grow increasingly important to assess their health and health needs on a regular basis and to evaluate the impact of intervention programs. For these reasons, further research on the promising set of simple indicators evaluated here for one developing country setting and on the more detailed and expensive measures discussed here is an essential part of a research agenda on aging in Africa.