Address for correspondence: Steven Prus, Department of Sociology and Anthropology, Carleton University, 1125 Colonel By Drive, Ottawa, Ontario, Canada K1S 5B6 e-mail: email@example.com
This paper tests two competing hypotheses on the relationship between age, SES, and health inequality at the cohort/population level. The accumulation hypothesis predicts that the level of SES-based health inequality, and consequently the overall level of health inequality, within a cohort progressively increases as it ages. The divergence-convergence hypothesis predicts that these inequalities increase only up to early-old age then decrease. Data from a Canadian national health survey are used in this study, and are adjusted for SES-biases in mortality. Bootstrap methods are employed to assess the statistical precision and significance of the results. The Gini coefficient is used to estimate change in the overall level of health inequality with age, and the Concentration coefficient estimates the contribution of SES-based health inequalities to this change. Health is measured using the Health Utilities Index, and income and education provide the measure of SES. First, the findings show that the Gini coefficient progressively increases from 0.048 (95% CI: 0.045, 0.051) at ages 15–29 to 0.147 (95% CI: 0.131, 0.163) at ages 80+. Second, the data reveal that health inequalities between SES groups (Concentration coefficients for income and education) tend to follow a similar pattern of divergence. Together these findings provide support for the accumulation hypothesis. A notable implication of the study's findings is that the level of health inequality increases when compensating for age-specific socio-economic differences in mortality. These selective effects of mortality should be considered in future research on health inequalities and the lifecourse.
Material factors comprise the direct effects of SES on health, while lifestyle and psychosocial factors are the indirect effects (Segall and Chappell 2000). Those with higher education, for example, tend to have higher occupational status and earnings and, thus, adequate financial resources to support the purchase of good housing, nutrition, and private healthcare, all of which are directly tied to better health. SES influences health indirectly, as position in the socio-economic structure affects psychosocial experiences (e.g. exposure to negative life events and chronic stressors, self-mastery and -coherence, coping skills, and social support) and health-related lifestyle preferences and behaviours (e.g. cigarette smoking, excessive alcohol and refined-food consumption, leisure-time exercise, access to and use of preventative healthcare services, and acquisition/interpretation of health-education information), which in turn affect health.
Age, SES, and health: divergence versus divergence-convergence
It is further argued that the strength of the relationship between SES and health changes over the lifecourse, as the health of lower and higher SES persons generally declines at different rates. There are two main, competing hypotheses about the relationship between age, SES, and health.
The cumulative effects of healthier living, coupled with other advantages in economic, social, and psychosocial resources, over the lifecourse help postpone or compress morbidity and disability into a shorter period of the last years of life for persons with higher SES. Individuals with lower SES by contrast tend to experience increasingly poorer health over the lifecourse, which reflects negative cumulative effects of less healthy lifestyles and economic, social, psychosocial disadvantages on their health with age. As health advantages and disadvantages associated with these resources (or the lack of them) cumulate with age, health differences between socio-economic groups grow and the SES-health relationship strengthens.
The divergence-convergence hypothesis, on the other hand, maintains that the health gap between SES groups diverges only up to middle age and early-old age, then converges. This pattern reflects socio-economic differences in the extent of exposure to health-related psychosocial and behavioral risk factors and their impact on health at various stages of life (House et al. 1990, 1994). House and Robbins (1983) explain that the size of SES differentials in the exposure to psychosocial/behavioural risk factors associated with morbidity and disability are greatest in middle and early-old age. The impact of many of these factors on health is also greatest at these ages, as people become more biologically vulnerable to disease and illness as they grow older, and as the lack of social support, mastery, and competence become more challenging with age. Since exposure to risky health behaviours, lack of social support, high stress, low mastery/competence, and other psychosocial risk factors among lower relative to higher SES groups (and their impact on health) are greatest in middle and early-old age, the socioeconomic-based divergence in health should also be largest at this point in the lifecourse.
Health is however less stratified along socio-economic lines (i.e. SES-based gaps in health converge) among old adults. This reflects the fact that SES differences in exposure to psychosocial/behavioural risk factors fade away among old adults, though their impact on health is still strong. SES differences in exposure to some health risks are much smaller in old age compared to other ages because of extensive public welfare policies (principally Medicare and Social Security) aimed at reducing healthcare and economic, and thus health, inequalities in old age, as well as changes in lifestyle (e.g. persons with low SES are more likely to have retired and/or quit smoking and drinking alcohol) (House et al. 1994).
Age, SES, and health: from individual to population level health dynamics
Health dynamics occur on two distinct, yet related, levels: an individual level and a population level. Most research on health dynamics, such as that described above, focuses on the individual level of analysis – it examines the extent to which individuals with early-life health advantages generally maintain their health status relative to those with early-life health disadvantages over the adult lifecourse. Analysis at the population level considers the collective aspects of these individual level processes.
According to the accumulation hypothesis, a direct implication of cumulative health advantage and disadvantage for individuals with higher and lower SES respectively (i.e. individual level health dynamics) is that SES-based and thus overall levels of health inequality within a cohort (i.e. population level health dynamics) increase as it ages (Hart et al. 1998). Differences in average health status between SES groups and thus inequality in the total distribution of health outcomes therefore widen with age as a result of the more rapid decline in the health of lower compared to higher SES individuals over the lifecourse. By contrast, it is deduced from the divergence-convergence hypothesis that SES-based, and consequently overall, levels of health inequality increase then decrease as a cohort grows old as the health of lower and higher SES individuals diverges then converges over the later parts of their lifecourse. The objective of the current study is to apply and test the accumulation and divergence-convergence hypotheses at the population/cohort level of analysis.
The first and primary research question asks: how does the overall level of health inequality within a cohort change as it ages – does the overall level of inequality in the distribution of health outcomes increase or increase then decrease with age as implied by the accumulation and divergence-convergence hypotheses respectively? The second research question asks: to what extent do SES health inequalities (i.e. differences in average health status between SES groups) account for overall health inequalities? As suggested by the accumulation hypothesis, an increase in the overall level of health inequality with age is tied to an increase in the SES-based level of health inequality. The divergence-convergence hypothesis predicts that SES-based (and thus overall) levels of health inequality increase then decrease as a cohort grows old.
This study is based on cross-sectional data from the public-use version of the 1994/1995 National Population Health Survey (NPHS), which covers a representative sample of private Canadian households (excluding those on Indian Reserves and Canadian Forces Bases and in some remote areas of Quebec and Ontario). The NPHS collects information on health and illness, use of health services, determinants of health, and demographic and economic characteristics of individuals, and is based on a multistage stratified cluster probability sampling design developed by Statistics Canada. Sample weights, which were adjusted to sum to sample size, are used in all data analyses here to account for unequal probabilities of selection as a result of the multistage sampling design employed in the NPHS.
The household response rate for the 1994/1995 NPHS was 88.7 per cent. In each sampled household, some limited information was collected from all household members (n = 58,439) and one person, aged 12 years and over, was randomly selected for a more in-depth interview. These in-depth interviews, which comprise the data used in this paper, were obtained from 17,626 individuals, for a response rate of about 96.1 per cent. At the Canada level, these yield a combined response rate of about 85 per cent for the 1994/1995 NPHS (Statistics Canada 2005).
This study focuses on the adult lifecourse from age 15 and over. Age is a categorical variable divided into five-year intervals in the NPHS, and collapsed here into six age groups: 15–29 (sample size: 4,014), 30–39 (sample size: 3,592), 40–49 (sample size: 2,756), 50–64 (sample size: 2,873), 65–79 (sample size: 2,361), and 80+ (sample size: 591). The age variable in the NPHS is top-coded at 80 years of age to guard against disclosure.
Approximately 800 cases (or four per cent of the total sample) contained missing data on the variables (largely the SES – income and education – measures) used here, but they were randomly scattered through the data. Analysis shows no statistically significant differences between the average health of missing income cases and valid cases (p= 0.945), as well as between missing education cases and valid cases (p= 0.245). Missing data are therefore excluded from the analyses. The final sample size is 16,187 persons.
Health is measured using the widely-used Health Utilities Index Mark 3 (HUI), which is the most comprehensive and global measure of health status in the NPHS. HUI is an index of an individual's overall functional health based on eight self-reported attributes: vision, hearing, speech, mobility, dexterity, cognition, emotion, and pain/discomfort. Respondents are asked up to several questions per attribute (see Appendix A for the entire HUI module for the 1994/1995 NPHS questionnaire) about their usual abilities or day-to-day health.
These attributes are weighted and organised into a single numerical value using a multi-attribute utility theory, based on preference measures for health states derived from an Ontario, Canada community sample survey (i.e. respondents in this survey were asked to rank various health conditions in order of the severity of their impact on one's health). Values, which reflect health utilities, range from about 0 (i.e. utility of being dead or completely unfunctional) to 1 (i.e. utility of being healthy or perfect functional health) in increments of 0.001 (Feeny et al. 2002, Furlong et al. 2001). For example, a respondent who is near-sighted, yet fully healthy on the other seven attributes, receives a score of 0.973 or 97.3 per cent of full health.
More generally speaking, an HUI score of 0.80 or greater indicates very good health while scores below 0.80 indicate moderate or severe functional health problems (Roberge et al. 1995). Relatedly, differences of greater than 0.03 between HUI scores are deemed to be unconditionally (clinically) important and meaningful, and differences between 0.01–0.03 may be important in various situations (Drummond 2001, Feeny et al. 2002, Grootendorst et al. 2000, Schultz and Kopec 2003). Based on the intra-class correlation coefficient, HUI scores have a test-retest reliability of 0.77 (Feeny et al. 2002).
HUI is also highly correlated with other commonly-used indicators of global health such as self-rated health. Humphries and van Doorslaer (2000) show that the level of SES-based inequality in self-rated health is not significantly different from that in HUI. Lima and Kopec (2005) point out that the HUI is also statistically associated with drug use and hospitalisation, and is reactive to changes in health status due to serious illness or disability.
Overall, the HUI provides a rather objective measure of functional limitations and disabilities, and is often considered a measure of individual as well as population health (Roberge et al. 1995). The HUI also provides a comprehensive, global measure of health. This is important for this study because the hypotheses used here assume that lower SES persons are more likely to experience a general susceptibility to disease and illness or multiple health problems as opposed to condition-specific health problems. Finally, and importantly, the HUI is one of the few available health measures that are appropriate for use in most measures of inequality since it is based on a continuous scale.
This study uses a dual-indicator of SES: income and education (occupation is not appropriate for this study since the NPHS excludes previous occupational information for those no longer in the work force such as retirees). A dual-indicator approach is employed since income and education (while highly correlated) have unique characteristics (e.g. unlike income, education does not change significantly over the middle and later life course) and they reflect different aspects of the social class structure (e.g. income is more likely to represent purchasing power, while education better reflects acquisition and interpretation of health information) (e.g. Davey Smith et al. 1998, Oakes and Rossi 2003, Winkleby et al. 1992). Using both education and income provides a broader measure of SES.
It is argued that conventional measures of SES such as education and current income may be less suitable for old ages than younger ages (Kaplan et al. 1987, Matthews et al. 2005), and that alternative measures like long-term income (average income over many years) or net worth/total financial assets (which are not available in the NPHS) may better capture the cumulative effects of lifetime SES on health status in old age (Benzeval and Judge 2001, Robert and House 1996). Conventional SES measures, however, are the most commonly used variables in studies of SES and health. Using education and income thus facilities a comparison of this study to other studies that use similar measures of SES and health (e.g. Humphries and van Doorslaer 2000, Wagstaff and Watanabe 2003). Furthermore, as demonstrated by Duncan et al. (2002), conventional SES indicators tend to be strongly correlated with health.
The highest level of education obtained is a categorical variable in the NPHS and is rank-ordered here as follows: (1) doctorate, masters, or medical degree; (2) bachelors degree; (3) some university; (4) community college diploma; (5) some community college; (6) trade school diploma; (7) some trade school (or other schooling beyond high school); (8) high school graduate; (9) some high school; (10) elementary or some elementary school; and (11) no schooling. The analysis is based on a standardised version of this variable – education is collapsed into age-specific quartiles to reduce the impact of cohort effects (e.g. young adults are better educated than old adults). Each education quartile represents 25 per cent of the cases for a given age. For those aged 15–29, for example, respondents are rank-ordered by education and then divided into four equal groups, where the first quartile (symbolised as Q1) is made-up of 15–29-year-old respondents with the lowest 25 per cent of education, the second quartile (Q2) comprises 15–29-year-olds with the next lowest 25 per cent of educational attainment, and so on. This procedure is replicated for each age group; hence, every respondent in the sample is assigned to one of four education quartiles based on his/her educational ranking within a particular age group. Table 1 shows the age-specific education quartile thresholds.
Table 1. Income and education quartile thresholds by age*
Income quartile thresholds (in Canadian dollars) are shown in the first row for each age. Education quartile thresholds are shown in the second row, where: Elementary = Elementary School; Some HS = Some High School; Some TS = Some Trade School; TS = Trade School Diploma; Some CC = Some Community College; CC = Community College Diploma; Some Univ = Some University.
Q1 = persons with the lowest 25 per cent of scores . . . and Q4 = persons with the highest 25 per cent of scores.
15–29 (n= 4,014)
30–39 (n= 3,592)
40–49 (n= 2,756)
50–64 (n= 2,873)
65–79 (n= 2,361)
80+ (n= 591)
Income is based on total annual household income (in Canadian dollars) before taxes and deductions. It is also a categorical variable that is divided into numerous income intervals (e.g. no income; $1–$4,999; $5,000–$9,999; $10,000–$14,999 . . . ). The standardisation process discussed above was repeated for income (i.e. income was collapsed into age-specific quartiles to reduce the impact of cohort effects). Table 1 shows the income quartile thresholds for each age group.
Patterns of age, SES and health may be influenced by a SES-bias in mortality (Beckett 2000, House et al. 1994, Lynch 2003, Wolfson et al. 1993). This bias may alter distributions of education, income, and other measures of SES within age groups and consequently how the effect of SES on health is conditioned by age. SES-based health inequalities (as well as overall health inequalities), especially in old age, therefore may be underestimated because a disproportionate amount of those with lower SES have died, leaving a relatively smaller but healthier population of lower status seniors.
To help deal with this challenge the data are weighted to compensate for the effects of SES differences in mortality. In particular, Mustard et al. (1997) provide estimates of age-specific socio-economic differences in mortality based on a representative sample of deaths that occurred over a two-year period (from June 1986 to May 1988) in the Canadian province of Manitoba. Using the same approach discussed above to standardise household income and highest level of education, they find a statistically significant relationship between mortality and income quartile rank among individuals aged 30–49 and 50–64 and a significant association between mortality and education quartile rank only for those aged 65 and over. The odds of mortality with a one-level decrease in income quartile (e.g. from the fourth quartile to the third quartile) are 1.34 and 1.36 within the 30–49 and 50–64 age groups respectively. The odds of death with a one-level decrease in education quartile among persons aged 65+ are 1.13. Using these odds, the data here are weighted to compensate for these age-specific socio-economic differences in mortality. In addition, the data are also weighted to account for the sampling design discussed above.
Two statistical techniques are used to analyse the research questions: (1) Gini coefficient and (2) Concentration coefficient. The Gini coefficient is used to measure the overall level of health inequality within a cohort (i.e. to answer the first question). The Concentration coefficient is used to estimate the extent of health inequalities between SES groups, which provides an answer to the second question.
The Gini coefficient (symbolised here as G) is a summary device that provides a single number measure of relative (as opposed to absolute) inequality. G ranges from zero to one. If everyone had the same health, G would be zero; conversely, if just one individual was healthy and all others unhealthy, the coefficient would be one. Hence, the higher the G, the more health inequality that exists.
The Gini coefficient expresses the degree of inequality in the distribution of HUI as a single number. To elaborate on the Gini-coefficient findings, health quartile distributions are also calculated. In a health quartile distribution persons are ranked according to their HUI score and divided into quartiles, where the first quartile (Q1) comprises persons with the lowest 25 per cent of HUI scores . . . and the fourth quartile (Q4) represents those with the highest 25 per cent of HUI scores. Each quartile's health share is then calculated by summing HUI scores of all persons in that quartile and dividing this figure by the sum of all HUI scores. By comparing health quartiles across age groups it is possible to see how shares of health have changed between health groups.
The Concentration coefficient (CB), which is a modified version of the Gini coefficient, is used here to estimate differences in health status between SES groups (i.e. SES-based health inequality) (Wagstaff and van Doorslaer 2004). Two different CB values are calculated: one for education and one for income. CB for education measures the contribution of health inequalities between each education quartile to the overall level of health inequality (i.e. G), and CB for income measures this contribution for income quartiles. Appendix B provides a discussion of the mathematical calculation of G and CB.
It is common for the Gini (or Concentration) coefficient to be reported without information on its sampling variance (standard error), even though most studies are based on sample data. When this information is provided it is often based on the bootstrap method (Efron and Tibshirani 1993). The bootstrap technique is used here to estimate standard errors and obtain confidence intervals for the Gini and Concentration coefficients. Null hypotheses of no difference between two given coefficients are also tested. If the confidence intervals of two coefficients do not overlap with each other at a given level of significance (e.g. 95 per cent), the difference between these coefficients is statistically significant at that level (Moran 2005).
The bootstrap method involves repeated random sampling with replacement (i.e. each case has a chance of being selected more than once in each bootstrap sample) from the data at hand. This produces a series of random samples from which the statistic of interest is computed for each of these bootstrap samples. This process of repeated sampling produces estimates of the standard error and thus confidence interval of the statistic (Walters and Campbell 2004). In the current study, this process is repeated 1,000 times and the normal-approximation method is used to produce the 95 per cent confidence interval using Stata (StataCorp 2005). Jolliffe and Krushelnytskyy (1999), Mills and Zandvakili (1997), Moran (2005), and Sosa Escudero and Gasparini (2000) provide further information and discussion on statistical inference through bootstrap techniques for measures of inequality.
Overall health inequality
The first research question asks: does the overall level of health inequality within a cohort increase as it ages? Table 2, which is graphically presented in Figure 1, provides the data to answer this question – it shows the Gini coefficient (G) across age groups. These data are calculated before and after adjustments for the SES-bias in mortality.
Table 2. Gini coefficients (G) of HUI inequality, standard errors (SE) and 95 per cent confidence intervals (95% CI) by age*
Data weighted for sampling design and the SES-bias in mortality (Gini coefficients weighted for sampling design only are in brackets)
As expected, inequality rates are higher after these adjustments are made. By ‘re-introducing’ deceased persons with lower SES, and generally poorer health, back into the sample, the overall level of health inequality (G) increases by more than 10 per cent (from 0.061 to 0.068) at ages 40–49 for example. Since true levels of health inequalities are probably underestimated by unadjusted data, all results discussed below are weighted for the SES-bias in mortality.
Table 2 also shows that total inequality in the distribution of health increases with age, supporting the accumulation hypothesis. Inequality changes steadily, but moderately, up to ages 40–49. There is then a trend toward even greater dispersion in health outcomes, especially during old age. The overall rate of health inequality increases more than three-fold from ages 15–29 (G = 0.048) to 80+ (G = 0.147).
Table 2 provides the bootstrapped standard error and 95 per cent confidence interval for each Gini coefficient. The standard errors are small relative to their Gini coefficients, indicating that these coefficients are estimated with a high degree of precision. The null hypothesis that two health distributions (i.e. Gini coefficients) are equal can also be assessed with this information. Since none of the confidence intervals for the Gini coefficients overlap with each other, the difference (i.e. increase) in the Gini coefficient from one age group to the next is statistically significant.
Table 3 provides insight into the above findings. First, it reveals that total health inequality (G) primarily reflects the differences in health shares between the bottom and top quartiles. The bottom quartile (i.e. Q1 or those with the lowest 25 per cent of HUI scores) is the only group that has a smaller proportion of total health at each age. For example, their share of all health is just 21.7 per cent at ages 15–29, yet they constitute 25 per cent of this population. This finding is logical since the bottom quartile also represents those with the poorest health.
Table 3. Share of total health (HUI) in percentages by health quartile*
Data weighted for sampling design and the SES-bias in mortality.
Q1 = persons with the lowest 25 per cent of HUI scores . . . and Q4 = persons with the highest 25 per cent of HUI scores.
Second, it shows that the increase in total health inequality observed in Table 2 mainly reflects the decrease in the bottom quartile's (Q1) share of health over the lifecourse. By ages 80+, their share of health drops to just 13.9 per cent. The relative situation of the second quartile (Q2) stays almost the same with age. The decrease in health shares for the bottom quartile benefits the top two quartiles (Q3 and Q4), which possess 30.4 per cent and 33.6 per cent of all health at ages 80+ compared to 26 per cent and 27.2 per cent respectively at ages 15–29.
SES-based health inequality
The overall level of health inequality (G) increases with age as suggested by the accumulation model. It is also suggested that this increase is a consequence of widening health inequalities between SES groups over the lifecourse (i.e. the second research question). To test this assumption Table 4 (displayed graphically in Figure 2) shows the pattern of both between-income group/quartile health inequality (i.e. CB for income) and between-education group/quartile health inequality (CB for education) across age groups.
Table 4. Between–income and between-education group/quartile Concentration coefficients (CB) of HUI inequality, standard errors (SE) and 95 per cent confidence intervals (95% CI) by age*
Data weighted for sampling design and the SES-bias in mortality.
The data in Table 4 again tend to support the accumulation hypothesis. They show that increased heterogeneity in health between both income and education quartiles parallels the increase in overall health inequality (G) with age as reported in Table 2. The level of between-income quartile inequality in absolute terms tends to increase over the adult lifecourse (CB ranges from 0.002 at ages 15–29 to 0.0111 at ages 80+), but does moderate somewhat during early old age (ages 65–79). Indeed, the difference between all pairs of (CB for income) values is statistically significant except for the difference in coefficients between ages 50–64 and 65–79. In relative terms, the contribution of income-related health inequality to total health inequality (i.e. CB as a per cent of G) also tends to increase with age. There is a similar pattern for health differences by education.
This paper uses the accumulation and divergence-convergence hypotheses to explain population level health dynamics. At the individual level, the accumulation hypothesis maintains that health disadvantages attached to early-life risky lifestyle, and lack of material and psychosocial resources of individuals with lower SES, cumulate with age. While morbidity and disability are increasingly experienced by lower SES persons from middle age and onward, higher SES individuals – who tend to have less exposure to these circumstances – experience a ‘compression of morbidity’ into a short period at the end of life. The implication of cumulate health advantages and disadvantages for population-level health dynamics is that differences in health between SES groups, and thus in the overall distribution of health outcomes, widen with age.
The divergence-convergence hypothesis also maintains that health disadvantages of lower SES persons associated with less healthy lifestyles and economic, social and psychosocial deficiencies cumulate with age, but that their impact on health are greatest in middle and early-old age. SES differences in exposure to some health risk factors are much smaller in old age because of extensive old-age welfare policies aimed at reducing economic, healthcare, and health inequalities, as well as changes in lifestyle, material, and social circumstances of persons with lower SES. Health is therefore less stratified along socio-economic lines in old age.
The findings presented here provide support for the accumulation hypothesis that an increase in the overall level of health inequality with age is tied to an increase in the SES-based level of health inequality. Both SES (CB) and total (G) health inequality tend to increase hand in hand with age.
The accumulation hypothesis does not assume that social policies and programmes are completely ineffective at countering health inequalities in later life. Without such intervention the divergence observed here would certainly be much wider. Cumulative advantage/disadvantage processes, however, outweigh the redistributive function of public health and income programmes. This would suggest that to achieve postponement of morbidity and disability for all persons, efforts need to focus more forcefully on, or target, lower SES groups, especially the poorest of the poor and at earlier stages of the lifecourse.
Public policies can help to reduce socioeconomic-based health inequalities, notably in middle and early-old age, by reducing the exposure to and impact of health-related risk factors among lower SES groups. Health promotion policies have potential for compressing morbidity and disability among entire cohorts, and not just those with economic advantage within cohorts. Many conventional health promotion policies however overlook the socio-economic factors that produce the problem at hand. Tobacco-control policies for example are typically aimed at the general population; yet socioeconomic-centered tobacco policies that reflect the economic, cultural, and social/physical environmental factors that underlie the above average smoking rates of lower SES persons must also be considered. Health promotion policies, programmes, and services aimed at modifying rates of obesity, exercise, stress, alcohol consumption, and so on should also be responsive to the needs of lower SES persons.
Strengthening public policies aimed at reducing economic inequalities in the total population would also probably change how the relation of age to health varies across socio-economic groups. The widening health gap between SES groups observed here suggests that policies of income redistribution should be targeted at the lowest SES groups. The potential for healthy ageing in old age hinges on economic resources since old-age income security programmes help individuals maintain pre-retirement standings of living and prevent poverty, both of which are associated with health (Wolfson et al. 1993). Yet seven per cent of all Canadian seniors live below the poverty line, with substantially higher poverty rates for females and those living alone (NACA 2005). Increasing absolute income or providing a guaranteed income above the poverty line to the poorest seniors would be likely to help to smooth inequalities in the overall distribution of health as cohorts enter old age.
Research limitations and implications
This paper provides a general theoretical, methodological, and empirical framework for further research on population health dynamics. This research would benefit from long-term longitudinal health data, which are not generally available in Canada.
First, longitudinal health data could establish a causal link between changes in SES-based and overall health inequalities with age. The data used here show that the increase in overall health inequality with age is accompanied by a similar change in health inequalities between SES groups. Longitudinal data could help to verify this link.
Second, it is often difficult in cross-sectional studies, such as the present one, to isolate the effects of cohort and age. While the SES standardisation approach employed here helps to reduce the influence of cohort effects, longitudinal data could help to further disentangle confounding age and cohort effects.
Third, the data show that health inequalities continue to increase during later life, but are top-coded at 80 years of age. Longitudinal data that are not top-coded would make it possible to determine the exact patterns in the health gap between status groups and hence the overall level of health inequality within a cohort during very old age.
Any future research on age, SES and health needs to consider the SES-bias in mortality. The findings here show that when compensating for the selective effects of mortality, inequalities in health increase. Future research on health inequalities and the lifecourse must consider age-specific socio-economic differences in mortality. True levels of health inequalities between SES groups and their contribution to inequalities in the overall distribution of health would otherwise probably be underestimated.
The author wishes to thank Wallace Clement and the anonymous reviewers for their exceptionally helpful comments and suggestions on an earlier draft of this article. The work underlying the paper was carried out as part of the SEDAP (Social and Economic Dimensions of an Aging Population) Research Program supported by the Social Sciences and Humanities Research Council of Canada.
Appendix A: Health Utilities Index (HUI) Module, 1994/1995 NPHS Questionnaire
The next set of questions asks about day to day health. The questions are not about illnesses like colds that affect people for short periods of time. They are concerned with a person's usual abilities. You may feel that some of these questions do not apply to you, but it is important that we ask the same questions of everyone.
Q1 Are you usually able to see well enough to read ordinary newsprint without glasses or contact lenses?
Q2 Are you usually able to see well enough to read ordinary newsprint with glasses or contact lenses?
Q3 Are you able to see at all?
Q4 Are you able to see well enough to recognize a friend on the other side of the street without glasses or contact lenses?
Q5 Are you usually able to see well enough to recognize a friend on the other side of the street with glasses or contact lenses?
Q6 Are you usually able to hear what is said in a group conversation with at least three other people without a hearing aid?
Q7 Are you usually able to hear what is said in a group conversation with at least three other people with a hearing aid?
Q7a Are you able to hear at all?
Q8 Are you usually able to hear what is said in a conversation with one other person in a quiet room without a hearing aid?
Q9 Are you usually able to hear what is said in a conversation with one other person in a quiet room with a hearing aid?
Q10 Are you usually able to be understood completely when speaking with strangers in your own language?
Q11 Are you able to be understood partially when speaking with strangers?
Q12 Are you able to be understood completely when speaking with those who know you well?
Q13 Are you able to be understood partially when speaking with those who know you well?
Q14 Are you usually able to walk around the neighbourhood without difficulty and without mechanical support such as braces, a cane or crutches?
Q15 Are you able to walk at all?
Q16 Do you require mechanical support such as braces, a cane or crutches to be able to walk around the neighbourhood?
Q17 Do you require the help of another person to be able to walk?
Q18 Do you require a wheelchair to get around?
Q19 How often do you use a wheelchair?
Q20 Do you need the help of another person to get around in the wheelchair?
Hands and fingers
Q21 Are you usually able to grasp and handle small objects such as a pencil and scissors?
Q22 Do you require the help of another person because of limitations in the use of hands or fingers?
Q23 Do you require the help of another person with:
___ Some tasks?
___ Most tasks?
___ Almost all tasks?
___ All tasks?
Q24 Do you require special equipment, for example, devices to assist in dressing because of limitations in the use of hands or fingers?
Q25 Would you describe yourself as being usually:
___ Happy and interested in life?
___ Somewhat happy?
___ Somewhat unhappy?
___ Unhappy with little interest in life?
___ So unhappy that life is not worthwhile?
Memory and thinking
Q26 How would you describe your usual ability to remember things?
___ Able to remember most things?
___ Somewhat forgetful?
___ Very forgetful?
___ Unable to remember anything at all?
Q27 How would you describe your usual ability to think and solve day to day problems?
___ Able to think clearly and solve problems?
___ Having a little difficulty?
___ Having some difficulty?
___ Having a great deal of difficulty?
___ Unable to think or solve problems?
Pain and discomfort
Q28 Are you usually free of pain or discomfort?
Q29 How would you describe the usual intensity of your pain or discomfort?
Q30 How many activities does your pain or discomfort prevent?
Inequalities in the distribution of health outcomes can be measured both at the univariate or marginal level (i.e. the basic level of health inequality within a population) and at the bivariate or conditional level (i.e. the extent of health inequalities between groups, such as SES groups) (Wolfson and Rowe 2001). The Gini coefficient (G) measures inequality at the univariate/marginal level, and a modified version of the Gini coefficient, often called the Concentration coefficient (C), measures inequality at the bivariate/conditional level.
The mathematical expression for the weighted G (i.e. weighted in this paper to take into consideration the sampling design and the SES-bias in mortality as discussed above), as provided by Crystal and Waehrer (1996), is:
In this formula let i= 1, . . . , k index individual observations in the data, where the data are ranked by health (i.e. HUI score) and k is the number of observations. The health (HUI score) and (sample and mortality) weight of the ith observation are denoted by ni and wi respectively.
The Gini coefficient is a ‘pure’ or ‘overall’ (i.e. univariate) measure of health inequality because individuals are ranked by health. If individuals are ranked by SES (starting with the most disadvantaged person) as opposed to health, the corresponding Gini coefficient [called the Concentration coefficient (C)] provides a (bivariate) measure of the level of SES health inequality (Wagstaff and van Doorslaer 2004).
If individuals happen to be ranked the same in terms of health and SES, G will also equal C. Any difference in rankings will result in G exceeding C, which is always the case in reality since some persons disadvantaged in SES are advantaged in health and vice-versa. Wagstaff and van Doorslaer (2004) denote the difference between G and C with the symbol R. Overall health inequality (G) is therefore comprised of two parts: C (reflecting the similarity in rankings in the SES and health distributions) and R (reflecting the difference in rankings in the SES and health distributions). Hence, G = C + R.
Wagstaff and van Doorslaer (2004) point out that the above statement assumes that SES is measured on a continuous scale. They offer a similar decomposition of G for categorical SES data (as is the case in the present study). However, a new term is added: G = CB+ CW+ R, where C is further decomposed into a between-SES group health inequality term (CB) [which measures the contribution of health inequalities between each SES group to the overall level of health inequality; i.e. G] and a within-SES group health inequality term (CW) [which measures the contribution of health inequalities within each SES group to overall health inequality]. Specifically, the CB term shows the level of health inequality if everyone in each SES group receives the mean health score for that group, while CW reflects the level of health inequality within each SES group.