^{*}

The authors have declared that no competing interests exist.

Analyzed the data: AJB CGV. Wrote the first draft of the manuscript: AJB CGV. Contributed to the writing of the manuscript: AJB CGV.

In a

To monitor progress towards the Millennium Development Goals, it is essential to monitor the coverage of health interventions in subgroups of the population, because national averages can hide important inequalities. In this review, we provide a practical guide to measuring and interpreting inequalities based on surveys carried out in low- and middle-income countries, with a focus on the health of mothers and children. Relevant stratification variables include urban/rural residence, geographic region, and educational level, but breakdowns by wealth status are increasingly popular. For the latter, a classification based on an asset index is the most appropriate for national surveys. The measurement of intervention coverage can be made by single indicators, but the use of combined measures has important advantages, and we advocate two summary measures (the composite coverage index and the co-coverage indicator) for the study of time trends and for cross-country comparisons. We highlight the need for inequality measures that take the whole socioeconomic distribution into account, such as the relative concentration index and the slope index of inequality, although simpler measures such as the ratio and difference between the richest and poorest groups may also be presented for non-technical audiences. Finally, we present a framework for the analysis of time trends in inequalities, arguing that it is essential to study both absolute and relative indicators, and we provide guidance to the joint interpretation of these results.

Equity in health has been part of the public health agenda for quite some time in the US, Europe, and Latin America

In spite of recent developments, descriptive cross-sectional studies of health inequalities are still the most common and useful type of study for the design and implementation of public health policies aimed at improving equity. Such studies require the measurement, presentation, comparison, and interpretation of inequalities in health. In this article, which is part of the

There are multiple dimensions to health equity according to gender, wealth, education, place of residence, ethnicity, and sexual orientation, among other factors. In this article we focus on “socioeconomic position," a term that is now preferred over “socioeconomic status" in the equity literature

Socioeconomic position can be ascertained using different types of indicators, each reflecting slightly—or sometimes markedly—different underlying constructs. From the standpoint of statistical analyses, an indicator should be easy to measure in a valid way during surveys, should not change rapidly over time, should allow breakdown into several categories (preferably of similar size), and should be comparable over time and place. No single measure of socioeconomic position fulfills all these criteria in a satisfactory way. Howe et al. have recently reviewed the advantages and limitations of socioeconomic position indicators in low- and middle-income countries

Easy to measure but can also have a direct effect on health

Often results in unbalanced groups. In poor countries, a large proportion of women may have no education, whereas in wealthier countries most will have completed secondary school.

Size of the categories will vary over time, as more women are educated, which affects the comparison of time trends.

May be difficult to use in country comparisons because of different schooling structure, level names, and content.

Requires several questions to be asked about different sources of income.

Misreporting is frequent, and monthly variability may be important in low-income societies where casual labor and agricultural production are common.

More positively, income is a continuous variable that can be broken down in groups of uniform size, which allows comparisons over time.

Reflects what people spend rather than what they earn.

Difficult to measure, requiring respondents to keep diaries and to answer long questionnaires, and requiring multiple visits by interviewers.

Affected by misreporting, seasonality, in-kind exchanges, and domestic production of goods

If properly measured, consumption expenditure is a useful indicator, but its practical limitations have so far restricted its use in health research in low- and middle-income countries.

Commonly used in high-income country studies, this measure of socioeconomic position is problematic in low- and middle-income countries, where changes in occupation and multiple jobs are common and unemployment or informal jobs predominate.

Long questionnaires and complex post-processing are required to capture all the subtleties of occupation in low- and middle-income countries, where large proportions of the population may fall into a single category. “Farmers," for example, may include anyone from a landless laborer to a plantation owner.

Several classifications used in countries are not ordinal, making it impossible to rank groups.

In light of the problems with the above socioeconomic position indicators, an alternative was proposed by Filmer and Pritchett in 1998

Like all other socioeconomic position indicators, asset indices have limitations

These limitations do not, however, preclude the widespread and valid use of asset indices for documenting the wide gaps between rich and poor that are present in most low- and middle-income countries, as is evident by the consistent associations between asset indices and more complex measures of socioeconomic position

There are two basic approaches for measuring inequalities in intervention coverage. The first is to carry out separate analyses for each relevant coverage indicator, such as contraceptive use, presence of skilled birth attendant, measles vaccine coverage, oral rehydration therapy, etc. The Countdown to 2015 initiative

To avoid the problems of studying one coverage indicator at a time, two related measures have been proposed that combine the coverage of several interventions (

Based on how many preventative interventions each mother/child pair received, out of a set of 8–9 essential interventions

Calculation of co-coverage requires reanalysis of original survey data, which is time-consuming, but because co-coverage is measured at the individual level, standard errors and confidence intervals can be calculated.

Co-coverage is often reported as the percentage of children covered by at least three or six interventions but can also be presented through stacked bar graphs that show the percentage of children in the population covered by a given number of interventions, usually stratified by wealth quintiles (

Based on the weighted average of coverage of a set of eight preventative and curative interventions; the CCI gives equal weight to four stages in the continuum of care: family planning, maternal and newborn care, immunization, and case management of sick children

The weighted average for a group (e.g., a country or a wealth quintile) is calculated as

Because the CCI is a group indicator, jackknife or similar resampling methods are required to estimate its standard error

See

There is no consensus on the ideal measure for expressing the magnitude of inequalities. In 1991, Wagstaff et al.

Despite their simplicity, these measures, which take into account only the top (Q5) and bottom (Q1) quintiles of the population under study, have important limitations. First, these measures are sensitive to changes in the number of individuals in each stratification category. For example, the rich/poor ratio for coverage with skilled birth attendants based on the 2008 Nigeria DHS survey

More importantly, the intermediate population groups (e.g., Q2 to Q4) will not be captured in these simple measures of inequality

The CIX is related to the Gini coefficient

Typically, however, health interventions are more concentrated towards the richer groups, and the CIX assumes a positive value, as the curve is below the diagonal.

Conc. index, concentration index.

The main downside of the CIX is the lack of direct interpretability of its values. Clearly, a value of 20 means more inequality than a value of eight, but these numbers lack a clear meaning, unlike Q5/Q1 ratios, which are easily interpretable.

Alternative formulations for CIX can be used to reflect absolute inequalities

There are two potential problems with a linear regression approach like this when used with an indicator, such as intervention coverage, that has a minimum of 0% and a maximum of 100%. The first is that it assumes a linear relationship between outcome and predictor, which is not always the case, particularly when a “top inequality" or “bottom inequality" pattern is present (see

Inspecting the distance between groups in an inequality graph (such as the five-dot plots in

Three types of patterns of inequality have been described as “linear," “bottom" and “top" inequality patterns by Victora et al.

Under usual conditions, low-coverage countries tend to show a top inequality pattern, with the richest quintile way ahead of the rest. As coverage increases they move to the linear pattern, where the distance between groups is similar. When higher levels of coverage are attained, a bottom inequality pattern usually appears, with the poorest lagging behind

Where there is a linear pattern of inequality (Gambia,

Where there is a bottom inequality pattern (Bolivia,

Where there is a top inequality pattern (Bangladesh,

See

A measure that is closely related to the SII is the relative index of inequality, or RII. The curve-fitting procedure used to calculate the RII is the same as for the SII, but instead of calculating the difference between the fitted values for one and zero, the RII is the ratio between the two. In the Nigeria measles vaccination example, the RII equals 10.6 (80% divided by 7.6%) when the linear regression approach is used. The estimate reduces to 6.4 if logistic regression is used. Given the potential problems associated with linear regression, we strongly advise that logistic regression should always be used in the calculation of SII or RII for coverage indicators.

There is near consensus in the recent literature that no single measure of inequality reveals the full picture, and that authors should report both absolute and relative measures

The debate on absolute versus relative measures of inequality alluded to above is particularly controversial when it relates to the issue of whether inequalities are increasing or declining over time

In

(A) Situation 1—increasing rates of a health indicator, typical of a preventive intervention, such as immunization, or a desirable behavior such as exclusive breastfeeding. (B) Situation 2—declining rates of a health indicator, typical of an ill-health indicator, such as undernutrition or mortality, or a risk factor, such as high parity. “d” indicates the difference in coverage between the top and bottom quintiles; “r” indicates the ratio of the coverage in the top and bottom quintiles.

Let us assume that in situation 1 the richest quintile starts at 40% coverage, and the poorest at 20%. The baseline difference equals 20 percentage points, and the ratio equals two. Let us also assume that coverage among the richest increases to 80% at end point. We can then explore two alternatives for coverage among the poorest: coverage “A," where absolute inequality remains unchanged (the difference is the same as at baseline), and coverage “B," where relative inequality remains constant over time (same ratio). The worst scenario in terms of inequality is an end point for the bottom quintile where coverage is below B, because in this region of the chart, inequality will have increased both in terms of the difference and the ratio. The ideal scenario is an end point where coverage is above A. Here, both the difference and the ratio will have decreased. Finally, there are intermediate situations, where coverage is in between A and B. Here, results are apparently inconsistent. Compared to baseline, the difference between the extreme quintiles (absolute inequality) will have increased and the ratio (relative inequality) will have decreased.

In situation 2 where the outcome is declining, we have similar results for the worst- and best-case scenarios (coverage above A and below B, respectively) as in situation 1, with both the difference and ratio increasing or decreasing. The intermediate scenario, on the other hand, is different: the difference will have decreased but the ratio will have increased.

In other words, the apparent conflict between changes in absolute and relative inequalities reflects scenarios where inequalities have been reduced, but not so much that both absolute and relative measures have decreased. In

This example corresponds to situation 1 in Figure 4. CAR, Central African Republic.

Sample variability has been often overlooked in inequality analysis, irrespective of the measure used, which is problematic when looking at trends. The convenient regression approach to the estimation of the CIX presented by O'Donnell et al.

In this article, we have provided practical guidance on assessing inequalities in coverage of health and nutrition interventions, with emphasis on survey data from low- and middle-income countries. From our own experience, we make several recommendations about how best to assess inequalities in health intervention coverage. First, we conclude that there is no single best measure of inequality, and recommend that at least one absolute and one relative measure should be presented when describing inequalities at a given point in time, as well as when reporting trends over time. Second, when comparing time points or countries, we emphasize how important it is to calculate measures that take the whole population into account, and advocate the use of the CIX and the SII. In addition, we strongly advise the use of logit-based SII for the measurement of absolute inequalities. Because the presentation of these indices is particularly appropriate for academic audiences, we also recommend calculation of the differences and ratios among extreme quintiles, because these are easy to convey to general audiences. Third, when assessing change in inequalities, we argue that it is essential not only to evaluate both absolute and relative changes, but also to report how they evolve jointly. Finally, in situations where conflicting results are provided by absolute and relative measures, we stress that it is essential that researchers spell out the different interpretations of these measures to public health experts, because these interpretations are affected by value judgments and are likely to affect the approaches taken to reduce inequalities in the coverage of health interventions.

Recent international calls for increased accountability in measuring progress towards the Millennium Development Goals demand analyses of health indicators stratified by socioeconomic position and other equityrelated variables.

Socioeconomic position can be ascertained using many different indicators, but the use of a wealth classification based on assets is the best option for national surveys, being feasible and reliable.

Intervention coverage can be assessed by individual indicators, but because combined measures are important for the study of time trends and for cross-country comparisons, we advocate the use of the composite coverage index and the co-coverage indicator.

Inequality measures that take the whole socioeconomic distribution into account are essential, and at least one absolute (the slope index of inequality) and one relative (the relative concentration index) measure should always be presented.

When analyzing inequality trends, absolute and relative inequality must be studied jointly because there is a clear interacting pattern of reduction or increase in inequality that sometimes produces apparently contradictory changes in absolute and relative inequality.

We wish to thank the members of Child Health Epidemiology Reference Group (CHERG) for their insightful comments on the draft version of this manuscript, and especially Jennifer Bryce for putting together this series, encouraging us in writing this piece, and making everything happen. We also thank the Pelotas International Center for Equity in Health (Federal University of Pelotas, Brazil), which produced all the equity data breakdowns used in our analyses and examples.

composite coverage index

concentration index

Demographic and Health Survey/s

Multiple Indicator Cluster Survey/s

maternal, newborn, and child health

relative index of inequality

slope index of inequality