Summary
Our paper shows that the magnitude of both relative and absolute socio-economic inequalities in health-related outcomes is empirically related to the overall level of these outcomes. Relative inequalities, using the Rate Ratio as measure, tend to be larger at lower overall levels (e.g. of mortality). Absolute inequalities, using the Rate Difference as measure, tend to be low at both very low and very high overall levels. Our paper demonstrates that the magnitude of the RR and the RD is bound by mathematical ceilings. These ceilings partly explain the empirical patterns described above. Low RRs at very high overall levels, for instance, are a necessity, not an accomplishment. They reflect the fact that rates in all wealth layers need to be very high in order to uphold a very high overall level. Yet, even where mathematically-defined ceilings do not play a role, the magnitude of absolute and relative inequalities is correlated with the overall level.
Rising RRs with declining overall levels are, however, not a necessity. There are countries with low mortality rates and low RRs. In Uzbekistan and Kazakhstan, for example, the RR in under-5 mortality is low (RR = 1.15 and 1.01 respectively), despite the comparatively low overall under-5 mortality levels in these countries (55/1,000 and 48/1,000 respectively). Similarly, the RD showed variation around the trend-line at most overall levels. Moreover, the exact empirical patterns varied between the specific health-related outcomes, showing that the relationship between relative and absolute inequalities on the one hand, and overall levels on the other, is not as rigid as sometimes suggested.
Evaluation of methodology
Our results are based on DHS data, which uses standardized core questionnaires that generally allow for comparisons across countries. Although there is some uncertainty around the precise estimates for individual countries, it seems unlikely that this explains the systematic patterns observed. As DHS comprises a broad set of countries (representing various regions, and political, economic and cultural contexts), we expect that the patterns described are not dependent on the selection of countries for which DHS data are available. Also, we examined a broad set of outcomes. We expect an approximately similar range of patterns for other outcomes that are associated with socio-economic status.
A wealth index, based on household ownership of assets, was the socio-economic characteristic used in this study. When using maternal education, we found similar patterns (results available upon request).
Our empirical findings are based on a cross-sectional cross-national analysis, and are therefore directly relevant for international comparative studies. Patterns across countries in one period of time can, however, not necessarily be interpreted as also reflecting changes over time within countries. There are, however, indications that the observed tendencies of the RR and RD are also seen over time. In Western Europe, declines in total mortality among adults between the 1980s and 1990s were accompanied by increasing relative inequalities in mortality between socio-economic groups [
2,
14,
15]. In developing countries, there is evidence that the decline in childhood mortality between the 1970s and the 1990s was accompanied by declining absolute socio-economic mortality inequalities, and stable or widening relative inequalities [
16,
17].
Our findings are important, not only for international comparisons of low and middle-income countries, but for all studies in which (health-related) inequalities are compared between populations. When comparing mortality inequalities between European countries [
18], for example, or when monitoring time-trends in inequality [
7], differences in overall mortality levels need to be taken into account. Also when comparing health inequalities between age groups it is important to take into account the fact that overall mortality rises with age. Indeed, relative inequalities tend to decline with age, while absolute differences increase dramatically [
19‐
21].
We used the most simple measures of relative and absolute inequality (the Rate Ratio and the Rate Difference) to illustrate the general tendencies and mathematical ceilings. Our findings can most likely be generalised to more sophisticated measures of relative and absolute inequality, such as the Relative Index of Inequality [
4], the Slope Index of Inequality (SII) [
4], and the Generalized Concentration Index [
5]. The mathematical ceilings to the Concentration Index have been described elsewhere [
22]. In a previous study we have reported a similar relationship between the SII and the overall level to the one reported here for the RD [
23].
Implications for monitoring health inequalities
Our study shows that not only the RR [
7], and not only the RD [
8], but both are associated with the overall level of the outcome. Preference for either measure can therefore not be based on (supposed lack of) these general tendencies.
At the same time these tendencies are not necessities. Scanlan argues that increasing RRs are nearly inevitable as mortality rates decline [
7]. Positive examples, however, demonstrate that keeping relative inequalities low when mortality levels decline, is attainable. This is important, both for policy makers and researchers, especially those who assume that rising inequalities with declining mortality levels are inevitable. Also the RD varies around the trend-line at most overall levels. This implies that both the RR and the RD are not entirely determined by overall levels and that both can be meaningful measures for monitoring inequality.
Conversely, small RRs at high overall levels are almost inevitable, as are low RDs at very low and very high overall rates. Ultimately, very low mortality levels are only attainable when absolute mortality inequalities are low. This should be taken into account when monitoring inequalities. The RR and the RD are therefore only useful for monitoring when the relationship of these measures with the overall level of the outcome is taken into account. Also when setting targets for reducing health inequalities, e.g. a 25% reduction in health inequalities in Europe [
3], it is important to take into account the context in terms of overall rates, and to carefully consider the measure used for monitoring progress.
Whereas there are no standard recipes, we will give some suggestions on how the overall level of the outcome can be taken into account when monitoring inequalities.
When populations with similar overall levels of the outcome are compared, the RR and RD are both meaningful measures for monitoring. Malawi and Peru, for example, exhibit a similar overall level of professional delivery attendance (ca. 55%). Yet, Malawi is doing substantially better in terms of equity in the provision of such care (RD = 23) than Peru (RD = 60). When using the RR, one should, however, be aware that its magnitude can be highly sensitive to whether the outcome is defined positively or negatively, as we demonstrated for skilled delivery attendance. For certain outcomes (e.g. mortality), a negative definition is conventionally used, whereas for others (e.g. immunisation) a positive definition is more common. We warn against uncritical use of common but arbitrary definitions of health-related outcomes in either positive or negative terms. Each definition describes another aspect of the empirical reality, and it can be meaningful to describe inequalities according to both.
When populations with different overall levels are compared, one can assess whether the population with smallest inequalities theoretically could, given its corresponding mathematically-defined ceiling, have reached the higher inequality observed in the population with which it is compared. If the magnitude of inequality of one of the populations seems to be restricted by the mathematically-defined ceilings, such direct comparisons may not be very meaningful. For example, whereas the RD in professional delivery care in Bangladesh (RD = 17) is lower than in India (RD = 48), a direct comparison between the two on basis of the RD may not be very meaningful as absolute inequalities in Bangladesh are necessarily low given its overall level of delivery care (8% vs. 34% in India).
A solution to both of the above issues would be to use Odds Ratio-based measures of inequality. These measures are not bound by mathematically-defined ceilings, and they are insensitive to whether the outcome is defined positively or negatively. While these are obvious advantages of the Odds Ratio, it has the disadvantage that it is hard to interpret by non-researchers [
27], who may tend to misinterpret this measure as a RR [
28]. Moreover, while the insensitivity of the OR to positive or negative health outcomes makes it immune to arbitrary decisions on outcome measures, it does not stimulate the researcher to be explicit in choosing for either a positive or a negative outcome indicator. An explicit choice is valuable in cases where positive and negative indicators, such the immunisation rate versus the non-immunisation rate have different policy implications.
International patterns, as presented in this paper, can also be used for monitoring. A country's performance in terms of health inequality can be assessed with reference to other countries with similar overall levels of the outcome. The trend-line, representing the average performance of countries at a given overall level, can be used as benchmark. For example, Malawi, with an overall level of professional delivery attendance of 55% and an RD of 23, is doing well compared to the trend-line presented in Figure
2c. Alternatively, the best possible attainment at a given overall level, or a predefined target may be used as reference. Finally, expectations based on the diffusion of innovations theory, can be used as framework for evaluating observed inequalities.
It can be useful to assess group-specific rates in addition to summary measures of inequality, for example when monitoring differential diffusion of innovations through a population. Again, it is important to take the overall level of the outcome into account. If not, group specific rates may become an indicator of the overall performance of a country, rather than being an indicator of its distribution. Group-specific rates can be benchmarked similarly as described above, using international comparisons.
Summarizing, both absolute and relative inequality measures can be meaningful for monitoring socio-economic health inequalities, provided that differences or changes in the overall level of the outcome are carefully taken into account. Our paper gives advice on how to take this overall level into account when monitoring these inequalities and presents data that can be used for benchmarking of low- and middle-income countries in the field of maternal and child health.