Monitoring subnational regional inequalities in health: measurement approaches and challenges

Subnational regional health inequality is defined as the variability in a given health
indicator between populations living in geographically-defined regions (provinces,
states, etc.). The rationale for measuring subnational regional-based inequality derives
from the assumption that populations of a region share similar conditions that directly
or indirectly affect health. These may include health system inputs and processes,
the availability of other services (e.g. education), local infrastructure, climate,
environmental contaminants, proximity to facilities, or the acceptability of services
(e.g. local culture). Furthermore, regions are the administrative units linked to
resource allocation. Thus, monitoring health inequalities between regions can generate
important evidence and support for targeting of health programs and policies, especially
when disparities are substantial 1]. We note that this is a distinct concept from measuring total inequality within a
population, which is a univariate measure of the distribution of health within a population.

As a starting point, disaggregated data from regions should be presented for visual
inspection, but may be cumbersome to interpret when several health indicators are
presented over a number of years for multiple regions 2]. Moreover, the interpretation of time trends becomes further complicated when the
relative size of regions varies over time. Building on disaggregated data, measuring
and describing regional inequality can be done in a number of ways using summary measures.

Summary measures of inequality condense disaggregated data into concise outputs, and
can thus be used to show trends and make comparisons. The selection of appropriate
regional summary measures entails a few considerations 2]–4]. First, measures may demonstrate absolute inequality (i.e., the absolute magnitude
of difference, retaining the unit of measure of the health indicator) or relative
inequality (i.e., proportional differences that do not retain the unit of measure).
Several summary measures of inequality have both absolute and relative versions. Second,
measures that facilitate pairwise comparisons between two regions can be distinguished
from measures that simultaneously take all regions into account. Third, measures of
inequality may be based on weighted or unweighted data, according to whether or not
the population size in each region is taken into account. Finally, the choice of the
reference point should be justified based on the intended purposes of the analysis.
Reference points are commonly defined as the level of health in the best performing
region, health in a region with special significance (such as the capital region),
the overall mean health of all regions (i.e. national average), or a predetermined
standard level of health. The choice of such a point has important implications when
interpreting inequality measures.

Based on these considerations, each type of summary measure has implicit advantages
and disadvantages, and some are more intuitive than others. A review of the published
literature identified four main categories of summary measures applicable to regional
inequality (Panel 1).

The overall objective of the paper is to describe, compare and contrast current methods
of measuring subnational regional inequality. We use empirical data from four countries
to calculate inequality estimates for reproductive, maternal, newborn, and child health
services generated from 11 summary measures. We identify criteria for determining
the most robust, simplest, and consistent set of inequality measures, and discuss
the implications for reporting subnational regional inequality.

Panel 1. Typology of summary measures of regional inequalities

Pairwise measures

The most basic measures of subnational regional inequality include pairwise measures
such as differences and ratios. For example, the mean level (or a proportion or rate)
of a health indicator in region A may be compared to the mean in region B, or mean
in region A may be compared to the overall national mean. Because pairwise measures
are straightforward and comprehensible, they are ideal when only two areas are being
compared. However, these cannot be used to generate a single summary estimate for
multiple areas. In this case, pairwise comparisons are still possible but a reference
group must be defined, as was done to compare infant mortality rates in the five regions
of Brazil. The rate ratio between the region with the highest mortality (the Northeast)
and the region with the lowest (the South) decreased from 2.6 in 1990 to 2.2 in 2007;
the absolute difference between these two regions decreased from 47.1 deaths per 1000
live births in 1990 to 15.3 deaths per 1000 live births in 2007 5].

Measures of disproportionality

Measures of disproportionality look at the ‘share of health’ in a population that
is experienced by a given share of a subpopulation. (The share of health may encompass
health outcomes, health services or other health indicators.) For instance, the index
of dissimilarity shows the proportion or number of people who would have to move to
a different region to achieve a uniform distribution of health across a country 6]. It may be expressed in absolute (actual number of individuals) or relative (proportion
of the population) scales. The relative version of the index was calculated for four
maternal health service coverage indicators across 94 counties/cities in China. The
smallest inequalities were in hospital delivery rate (index of dissimilarity?=?6 %,
meaning that 6 % of the population would need to be redistributed to achieve a uniform
distribution of coverage across regions). The other indicators had indices of dissimilarity
of 11 % (for examinations rate), 18 % (for more than four postnatal examinations),
and 21 % (for more than four prenatal examinations). These analyses motivated the
integration of subsequent maternal health programmes and policies with a regional
focus 7].

The Theil index is also based on the concept of disproportionality, measuring relative
inequality. It takes into account the proportion of the population in each region
and the prevalence ratios of the health indicators in each region to the national
mean value. The Theil index has a minimum calculated value of 0 (no regional inequality);
as relative inequality increases, the value becomes larger, with no upper bound. If
a populous region has a much higher level of health than the national average the
Theil index will be inflated, indicating greater inequality. Theil index values may
be scaled (for example, uniformly multiplied by 1000) to facilitate interpretation.
For example, the Theil index was used to measure inequalities in the availability
of health workers among 22 provinces in China. It showed greater inequality in per-head
availability of nurses (Theil index value?=?0.067) than doctors (0.043); this was
also observed when inequality was analyzed at county level, with a higher Theil index
value for nurse (0.408) than doctor availability (0.235) 8].

Measures of impact

When applied to subnational regional inequalities, population attributable risk shows
the total health improvement expected at national level if all regions had the same
level of health as the reference group (often defined as the best performing region).
The measure takes into account the population size. A relative version, population
attributable risk percentage, shows the proportional improvement possible if all regions
attained the same level of health as the best performing region. Absolute and relative
versions of population attributable risk have been used to show that the number of
smokers in Montreal could be reduced by 176,869 people (population attributable risk)
or 55 % (population attributable risk percentage) if all city neighborhoods matched
the one with the lowest smoking prevalence 9].

Measures of variance

Measures based on the principle of variance aim to show how widely spread are the
levels of a health indicator in multiple geographical areas. Variance is the sum of
the squared differences between the level of health in each region and the overall
level, divided by the number of regions. It provides an absolute estimate of inequality,
which may be unweighted or weighted. The weighted (or between-group) variance approach
was applied to road traffic injury mortality across 22 cities (or counties) in Taiwan.
Differences between mortality in each city and the overall mean were squared and multiplied
by the city population size; the resulting value, divided by the national population
produced the between group variance, which decreased from 179 in 1997 to 49 in 2008
10].

The standard deviation, or square root of unweighted variance, was used to track regional
fertility inequalities in rural Iran. The standard deviation of the percentage of
births attended by unskilled personnel fell from 15.3 to 10.9 percentage points between
1996–2000 and 2001–2005, indicating decreased inequality 11].

Coefficient of variation is a relative version of standard deviation, expressed as
a percentage of the overall mean 12]. Being a relative measure, it allows comparison of the magnitude of inequalities
for different health indicators–even those that have different units of measurement–which
is not possible with the variance or standard deviation approaches. Additionally,
coefficient of variation takes into account the overall mean, allowing comparisons
over time when the overall mean may have changed. A study in 17 countries from the
Middle East and North Africa from 1980 to 1994 showed that while the mean under-five
mortality rate in the region decreased from 144.5 to 62.4 deaths per 1000 live births,
the coefficient of variation increased from 28.8 % in 1980 to a maximum of 52.3 %
in 1992 13].

Measures of mean differences from mean show how each region differs from a reference
point. The measure expressed as absolute or relative inequality, and may be weighted
or unweighted. Depending on the purposes of the comparison, reference points may include
the mean level of the whole population (a measure referred to as ‘mean difference
from overall mean’), the level of health in the best-performing region (a measure
referred to as ‘mean difference from best’), or a predetermined target level of health
2]. One specific formulation is known as the index of disparity, calculated as the average
of the absolute differences between the levels in each region and the overall mean,
divided by the overall mean and expressed as a percentage 6], 14]. The index of disparity was used to summarize regional inequalities in under-five
death rates in Iran over 1993–2009, and spanned from 24.4 % in 1995 to 17.6 % in 2007
15].