Study setting
The study was conducted in nine predominantly rural districts in Midwestern Uganda where the iCCM programme had been implemented by Malaria Consortium in collaboration with the Ministry of Health since 2010. The districts had an estimated population of 2.2 million inhabitants and approximately 23.3 % were children below the age of five years. By 2012 approximately 6800 community health workers (CHWs) were trained to deliver iCCM in the area. The CHWs complemented the formal healthcare sector in an area where health markets for children range from informal care at home and in the private sector to formal care from often distant and poorly functional health facilities. More than half of children being taken for care outside the home in Uganda are known to go to a drug shop or private clinic [
15].
In Uganda, CHWs working from home, locally known as village health team (VHT) members are an important and integral part of the healthcare system and represent the lowest health centre level offering promotive, preventive, and curative health services. In a typical village, there is a team of five to six CHWs offering a package of promotive and preventive health services, with two of them trained on iCCM for a duration of six days. Curative services offered under iCCM include provision of colour-coded artemether/lumefantrine combination to children aged 4–59 months confirmed to have uncomplicated malaria through a rapid diagnostic test, colour-coded amoxicillin to children aged 2–59 months with pneumonia classification based on cough and fast breathing, and zinc and ORS to children aged 2–59 months with uncomplicated diarrhoea. Children showing danger signs are referred to the nearest health facility, from which the CHWs receive supervision. In 2009 the recommended first-line antibiotic treatment for pneumonia was trimethoprim-sulfamethoxazole, while that for diarrhoea was ORS. Following a policy change in 2010, treatment guidelines changed to amoxicillin, and ORS-zinc combination for pneumonia and diarrhoea, respectively.
Study design and participants
The outcome of interest in this study was receiving appropriate treatment for either pneumonia or diarrhoea and has been defined as a) a child receiving an appropriate antibiotic for pneumonia if he or she is reported to have had cough and fast breathing (fast breathing pneumonia) and b) a child receiving either ORS alone or a combination of ORS and zinc if he or she is reported to have had watery stools or abnormally frequent bowel motions. The exposure variable of interest in this study has been defined as obtaining treatment from a community health worker (CHW) delivering iCCM services.
In order to obtain data on the outcome and exposure variable of interest, the study draws from before (2009) and after (2012) population based household surveys. The surveys sought to a) evaluate effect of iCCM on the proportion of children with access to appropriate care for fever, diarrhoea and fast breathing pneumonia through assessment of caregivers health-seeking behaviour around the illnesses, and b) provide an indirect estimate of infant and under-five mortality before and after implementation of iCCM compared to control areas (data not presented in this paper but elsewhere [
16]). With respect to health seeking behaviour, the surveys employed a facilitator-administered a questionnaire that asked about the symptoms experienced by eligible children, having sought for health care from outside the home, places from which health care was sought and any treatments obtained. The treatments obtained, when necessary were identified by the facilitators through examination of drug leftovers, prescription notes and with the use of ‘Drug cards’, which were a collection of pictures of common medicines created to assist caretakers in the identification of treatments provided for their children.
Study participants included primary caregivers of under-fives to whom the adapted questionnaire was administered. The sample size was therefore calculated based on a population-based, representative household interview survey for children under five, with a primary sample of 100 clusters for which household and birth history data were collected, and a smaller, nested sample of 40 clusters for the health-seeking behaviour component. Participant sampling was done with two-stage cluster sampling using the census database of the respective districts as a sampling frame, without restriction or exclusion. The primary sampling units were enumeration areas (village or cluster of household), which were selected with probability proportionate to size. First, the total sample needed for mortality estimation was drawn jointly with the Uganda Bureau of Statistics (UBOS) followed by a sub-set of clusters selected with equal probability, which were used for the health-seeking behaviour assessment. In the second stage of sampling, all households in the selected villages were listed and the number of households needed was selected by systematic sampling. For villages with more than 200 households, an equal size section approach was used whereby the village was divided into 2-4 sections with an approximately equal number of households, and one of the sections was selected using simple random sampling.
Statistical analysis
The main objectives of the analysis were to measure the impact of iCCM on uptake of appropriate treatments for pneumonia and diarrhoea, and to measure the magnitude and distribution of socioeconomic inequality in use of iCCM. Uptake was defined as the proportion of children eligible for treatment who were reported to have received treatment. Children were eligible for pneumonia treatment if they were reported to have had cough and fast breathing while for diarrhoea it was children reported to have had watery stools or abnormally frequent bowel motions. Appropriate treatment for pneumonia comprised of child receiving any antibiotic that was recommended as a first line treatment according to Uganda’s treatment guidelines or iCCM protocol [
19]. Similarly the national treatment guidelines and iCCM protocol recommend ORS and Zinc for diarrhoea and therefore appropriate treatment for diarrhoea comprised of ORS or a combination of ORS and zinc. For the uptake analysis, only the first source of health care from which the child’s caregiver sought for treatment outside the home was considered. Differences in the proportions of children treated appropriately at different time points and between the different groups were determined using Chi-square (
χ
2) tests. The study probed for inequalities in both the disease prevalence and reception of appropriate. During the inequality analysis, a socioeconomic status index was generated through principle components analysis of household assets and income generating activities that are recommended by the Uganda Bureau of Statistics [
20,
21]. The variables used in the analysis included building materials for the household, water source, type of toilet, type of fuel used for cooking and occupation of the head of the household. The socioeconomic status index was then divided into wealth quintiles. The individual indicators that weighted heaviest in the analysis were house construction materials. The socioeconomic status of the population was ranked from poorest to least poor, and age-sex indirectly standardized concentration indices (CI) were used to measure socioeconomic related inequality in uptake of antibiotics for pneumonia and ORS and zinc for diarrhoea. The CI is a measure of socioeconomic related inequality, which takes on values between -1 and +1. In the absence of inequalities, the concentration index takes on the value of zero. Positive values of the CI imply that a situation disproportionately affects the richer groups and negative values imply that a situation affects the poorer groups more and the larger the value, the larger the degree of inequality. Since receiving appropriate treatment and using iCCM are both binary variables, the corrected version of the CI, known as Erreygers concentration index (CCI) recommended for bound outcomes, was computed together with its standard error [
22,
23]. The index is denoted as:
$$ \mathrm{C}\mathrm{C}\mathrm{I}\ \left(\mathrm{y}\right) = 8\mathrm{c}\mathrm{o}\mathrm{v}\left(\mathrm{y}\mathrm{i},\mathrm{R}\mathrm{i}\right) $$
(1)
where Yi = binary outcome and Ri = Fractional rank is SES distribution. If inequality can be explained by linear regression of K (justifiable standardising factors) and J (non-justifiable inequality), this can be denoted as eq. (
2)
$$ {\mathrm{y}}_{\mathrm{i}}={\upbeta}_0+{\displaystyle \sum_{\mathrm{k}=1}^{\mathrm{k}}}{\upbeta}_{\mathrm{k}}{\mathrm{x}}_{\mathrm{i}\mathrm{k}}+{\displaystyle \sum_{\mathrm{j}=1}^{\mathrm{J}}}{\upbeta}_{\mathrm{j}}{\mathrm{z}}_{\mathrm{i}\mathrm{j}}{\upvarepsilon}_{\mathrm{i}} $$
(2)
then equation (i) can be rewritten as equation (iii) to show socioeconomic related inequality in uptake as expressed as a weighted sum of inequalities in its determinants and a residual term.
$$ \mathrm{C}\mathrm{C}\mathrm{I}\left(\mathrm{y}\right)=4\left[{\displaystyle {\sum}_{\mathrm{k}=1}^{\mathrm{k}}}{\upbeta}_{\mathrm{k}\ }{\overline{\mathrm{x}}}_{\mathrm{k}}\mathrm{C}\mathrm{I}\left({\mathrm{x}}_{\mathrm{k}}\right) + {\displaystyle {\sum}_{\mathrm{j}=1}^{\mathrm{J}}}{\upbeta}_{\mathrm{j}}{\overline{\mathrm{z}}}_{\mathrm{j}}\mathrm{C}\mathrm{I}\left({\mathrm{z}}_{\mathrm{j}}\right)+{\mathrm{GC}}_{\upvarepsilon}\right] $$
(3)
With the means of k
x and Z
j denoted as
\( {\bar{\mathrm{x}}}_{\mathrm{k}} \) and
\( {\overline{\mathrm{z}}}_{\mathrm{j}} \) respectively, and their concentration indices denoted as CI (X
k) and CI (Z
j) respectively, where
\( \mathrm{G}{\mathrm{C}}_{\upvarepsilon} \) is a residual term, the contribution of each factor to the inequalities can thus be established. Unfair inequality can thus be measured by subtracting fair inequality from the corrected concentration index to give an index of horizontal inequity (I) [
24].
$$ \mathrm{I}=\mathrm{C}\mathrm{C}\mathrm{I}\left(\mathrm{y}\right)\hbox{-} 4{\displaystyle {\sum}_{\mathrm{k}=1}^{\mathrm{k}}{\upbeta}_{\mathrm{k}\ }}{\overline{\mathrm{x}}}_{\mathrm{k}}\mathrm{C}\mathrm{I}\left({\mathrm{x}}_{\mathrm{k}}\right) $$
Due to the binary nature of the outcome variables, unjustifiable inequality was estimated through a linear model. Justifiable causes introduced in the model included: age and sex while non-justifiable causes included socioeconomic status, source of care, urban-location, education level and occupation of caregivers. Horizontal inequity was only calculated for the variable use of iCCM and not the variables relating to uptake of appropriate treatment. This is because all children fitting the case definition would require treatment as per the treatment guidelines. Therefore presence of any inequalities in treatment uptake would be equated to inequity.
In order to assess the effect of iCCM on appropriate treatment for pneumonia and diarrhoea, logistic regression models followed by PSM models were applied to the 2012 survey data for health seeking behaviour. The logistic regression models were used to estimate the marginal impact of exposure to iCCM on appropriate treatment for pneumonia and diarrhea while controlling for observed covariates. Despite their ability to account for possible confounders, logistic regression models are prone to the risk of selection bias. Since children exposed to iCCM may differ from those exposed elsewhere in several ways that can also influence uptake of appropriate treatment. PSM models were therefore constructed to eliminate the risk of selection bias. The propensity score generated by PSM methods is a balancing score that is believed to imitate a randomised controlled trial by generating a control and treatment group known as the ‘treated’ and ‘control’ respectively. This is because systematic differences in the characteristics of individuals who choose to participate or not to participate in a programme are removed [
25,
26]. Rosenbaum (1983) showed that if one conditions on the probability that a person participates in a programme (which is iCCM in this case), based on a set observable characteristics (X) that influence programme participation (P), known as the “propensity score (Pr(X) = Probability (P =1| X))”, this person’s participation decision (P) is also independent of the potential outcomes E(Y1 and Y0) where Y1 is the expected outcome in the presence of the intervention and Y0 is the expected outcome in the absence of the intervention [
25,
26]. Propensity scores are predicted probabilities, which have a continuous range known as “support” between 0 and 1 and thus the matching is done based on some intervals of this “support”, such as (0,0.1), (0.1,0.2) for individuals whose propensity score lies within the “common support”. The average treatment effects on the treated (ATT) is the mean of individual differences between the outcome of individual participants and their matched pairs and are denoted as E[Y1 – Y0| X, P = 1].
In this paper, the impact evaluation section refers to children who sought for care from a CHW for diarrhoea or pneumonia (used iCCM) as the ‘treated group’ and those who sought care from elsewhere as the ‘control group’. Therefore the term average treatment effects on the treated (ATT) has been used to refer to the effect of iCCM on appropriate treatment for pneumonia or diarrhoea.
The covariates (observable variables) included in the final PSM models included child’s age, sex of the respondent, education level of primary caretaker, socioeconomic status, concurrent infection with fever or diarrhoea, living in an peri-urban area, having no mode of transport at home, knowledge that CHWs have medicines, knowledge of danger signs for pneumonia, number of previous visits to CHWs in the past 3 months, history of hospitalisation in the last three months, and duration of illness. Logit regression was used to predict the propensity score and ATT for participation in iCCM were computed using different PSM techniques including a) kernel matching with bootstrap standard errors, STATA command attk b) stratification on the propensity score matching, STATA command atts c) nearest neighbour matching, STATA command attnd and d) radius matching of propensity scores within a calliper of 0.01, STATA command attr. Further more, sensitivity analysis using Rosenbaum bounds (mhbounds command in STATA) was conducted on the consistently significant ATT results. The purpose of the sensitivity analysis was to examine the robustness of the ATT results to unobserved confounding variables. In all of the analysis, sample weights were applied and cluster robust standard errors are reported. All computations were done in STATA 12 (College Station, TX).