Data sources
This study used two successive nationally representative household surveys: the 2003 and 2008 GDHS [
6],[
7]. The two datasets have comparable information on household characteristics and skilled attendants at birth at the time of the survey. The survey was designed to provide information to monitor the population and health situation in Ghana. The survey used a two-stage sample design to produce separate estimates for key indicators for each of the ten regions in Ghana. The first stage involved selecting clusters (called enumeration areas) from an updated master sampling frame constructed from the recent Ghana Population and Housing Census [
12]. A complete household listing operation was conducted in all the selected clusters to provide a sampling frame for the second stage selection of households. The second stage of selection involved the systematic sampling of the households listed in each cluster. Each household selected was eligible for interview. In these surveys a household was defined as a person or a group of persons, related or unrelated, who live together in the same house or compound, share the same housekeeping arrangements, and eat together as a unit. Further details of the sample design and questionnaire are described elsewhere [
6],[
7].
The 2003 GDHS database included information on 6,251 households and 3639 live births in the five years preceding the survey, whereas the 2008 GDHS database had information on 11, 778 households and 2909 live births in the five years preceding the survey. The two surveys offered the opportunity for analysing coverage trends in the proportion of births attended by a skilled professional.
Statistical analysis
Examining coverage trends is essential for assessing country progress. Information on trends requires at least two separate and comparable measurements at two points in time. A measure of progress - coverage gap - defined as how much coverage would need to increase from 2003 level to reach universal coverage was estimated to examine coverage trends. The change from 2003 to 2008 was then expressed as a percentage of this gap.
To explain the observed change in percentage of skilled birth attendants, the decomposition approach was used. Several regression decomposition approaches exist in the literature. The conventional Blinder-Oaxaca [
14],[
15] decomposition is based on two linear regression models that are fitted separately for the groups A and B:
For these models, Blinder [
14] and Oaxaca [
15] proposed the decomposition equations:
and
where Y
A
− Y
B
is the mean outcome difference, and X
A
and X
B
are mean vectors of the estimated coefficient vectors b
A
and b
B
for the two groups. In both equations, the first term on the right-hand side displays the difference in the outcome variable between the two groups due to differences in observable characteristics, whereas the second term shows the differential that is due to differences in coefficient estimates.
The approach used in this paper divides the change in percentage of skilled birth attendants into change in population structure and change in health behaviour and/or public health over the two time periods (or groups) 2003 to 2008 [
16],[
17]. The population structure was defined as the ratio of number of births in each category or level of the exposure of interest to the sample size expressed as a percentage. The decomposition analyses were performed using national level data disaggregated by birth order, maternal education, and household wealth index. This method assumes that the historical change in the proportion of births attended by skilled professional depends on: 1) change in distribution of maternal education, birth order, maternal age and household wealth index over time (i.e. composition effect); 2) actual change in the proportion of births attended by skilled professional due to change in health behavior or improvement in public health (i.e. basic effect - the regression intercept when x = 0 (
α)); and 3) variation of the proportion of births attended by skilled professional by exposure variables (
β), and the residual effect of other variables not considered as the error term (μ) [
16]. This can be specified mathematically as follows:
where Δ denotes change,
S = percentage of skilled birth attendants,
= arithmetic mean of percentage of skilled birth attendants for the j
th
category of the exposure variable,
w
j
= the population structure for the j
th
category of the exposure variable.
= arithmetic mean of the population structure for the j
th
category of the exposure variable,
Δw
j
= change in population structure expressed as a fraction for the j
th
category of the exposure variable, and,
x = the level of an exposure variable, e.g. maternal age was categorized as 1 (<20 years), 2 (20–34 years), and 3 (35–49 years). So x = 1, 2, 3.
The parameters in the mathematical model were estimated using an Excel spreadsheet program developed for this analysis and it is attached as an Additional file
1 to this manuscript [
16].
In a separate analysis, the effect of changes in population structure in respect of the mother’s characteristics (i.e. exposure of interest) on percentage of skilled birth attendants over the period was explored using the random-effects generalized least squares (GLS) regression. Independent panel datasets were constructed for each of the characteristics (birth order, mothers education, and household wealth index). The panel was defined as the category of the exposure of interest and the order of observations within panel was considered ordered by the number of surveys (two in this case). For example, the birth order and maternal education datasets each contains eight observations, while the household wealth index dataset contains 10 observations. Each dataset contains the following variables: the panel identifier, time identifier, proportion of births in each category of the characteristics (i.e. population structure), and percentage of skilled birth attendants. The analyses were performed with Stata 12 for Mac (StataCorp, College Station, USA) [
18].
The random-effects GLS regression model can be specified as follows:
In this model,
v
i
+
ϵ
it
is the residual that we have little interest in; the interest is to estimate
β (the effect of change in the structure of the exposure of interest).
w
it
is the population structure of the exposure of interest in the
i th panel at survey time,
t. v
i
is the unit-specific residual; it differs between units, but for any particular unit, its value is constant.
ϵ
it
is the “usual” residual with the usual properties (mean 0, uncorrelated with itself, uncorrelated with
w, uncorrelated with
v , and homoskedastic) [
18].
In all analyses, key survey characteristics such as sampling weight, stratification and clustering were accounted for.