We used Markov modeling in the software program TreeAge to model health and economic impacts of scaling up coverage of CCM from baseline to a target coverage of 90 percent in each of the eleven major regions. Each regional model was populated with 2016 region-specific data on cost per treatment, incidence of childhood pneumonia, background mortality, and baseline coverage of the intervention. Untreated case fatality rate and treatment effect were assumed to be similar in all the regions. Region-specific inputs are displayed in Table 1 and fixed model inputs are displayed in Table 2. We assumed that all treatments were provided outpatient under universal public coverage.
Markov modeling
Figure 1 provides an overview of our Markov model. The model has two arms labelled baseline and target coverage. The branches within the arms represent the probabilities that individuals within the modeled 2016 birth cohort would progress through the health events labeled at each branch. The CCM intervention is present during only the first five cycles. At the end of each cycle, one proportion of the cohort moves back to the “Alive and well” status, whereas the other proportion moves to the “Death” status. Those who returned to the “Death” status are removed from the model, whereas each return to the “Alive and well” status is counted as one life year completed (illustrated by the survival curves in Figure 2). We extrapolated results by running the model for a total of 120 cycles until everyone in the model cohort was dead. Since the model starts at age zero and runs for 120 cycles, output adds up to life expectancy. Model inputs were specific for each region, but the structure of the models did not change.
Model inputs on incidence of pneumonia
Data on prevalence of pneumonia, all-cause, under-five mortality rates and baseline coverage in each region were collected from the Demographic and Health Survey 2016 (DHS 2016). The DHS 2016 estimated the prevalence of pneumonia by asking mothers of children under five years of age whether their child had experienced clinical symptoms of pneumonia during the two weeks prior to the interview. Clinical symptoms are defined as cough accompanied by short, rapid breathing that was chest-related, and/or difficult breathing that was chest-related [2]. In this survey, baseline treatment coverage was estimated as the percentage of children with symptoms of pneumonia who received clinical examinations and oral antibiotics from a health professional.
We used the following formula to estimate the annual incidence of pneumonia among children below the age of five:
Incidence = prevalence/duration of disease [10]
The estimated incidence of pneumonia varied from 0.453 cases per year in the Amhara region to 0.040 cases per year in the Harari region (Table 2). Our calculations were comparable to estimates of incidence of childhood pneumonia in comparable settings [11].
Model inputs on background mortality
Age-specific mortality rates among adults affect the incremental life years gained by reducing U5MRs, and since U5MR is an acknowledged predictor of general population health, the large regional inequalities in U5MRs are likely to be reflected in mortality among older age groups. However, only life tables representing national averages were available for age-specific mortality rates among adults and children older than five years of age. Therefore, we modeled adult morality rates based on the assumption that the U5MR is associated with mortality among older age groups of the same population. In practice, we selected Ethiopian abridged life tables from the 2015 UN world population prospect [12] and matched each region to a national life table in a time period with a similar U5MR as the one observed for the region in 2016 [13]. The life table from this time period was used as a proxy for adult mortality rates in that region.
There are no census data from 2016. Without taking into account the effect of still births, we used regional fertility rates and data on the total number of women in each region to yield a rough estimate of the number of births in 2016 [2, 14]. We used these estimates as size variables in the calculations of weighted averages of the effects observed in each region, the budget impact and the geographical Gini coefficients.
Model inputs on the effectiveness and unit costs of CCM
Data on effectiveness of the treatment were collected from a previously published systematic review of studies assessing CCM of pneumonia in developing countries [15]. The review concluded that CCM on average reduces the case fatality rate of pneumonia in children less than five years of age by 70% (Table 2). We applied this as our input for treatment effects in all the regions.
Data on treatment costs were collected through literature review. We did not encounter any data on regional cost per treatment. However, previous studies indicate that there are significant differences in costs of providing community health services in different geographical contexts [9, 16]. The inputs for costs per treatment in each region were therefore adjusted for rural and urban residency.
One study from Kenya showed that it was 7.2 times more expensive to provide community health services in rural areas compared to urban areas [9]. We assumed that rural Ethiopia would observe a similar increase in cost per treatment compared to urban Ethiopia. We applied a cost per treatment provided for patients with urban residency of 45 USD [8, 17], and the cost per treatment provided for patients with rural residency was modelled to be 7.2 times more expensive (Table 1). However, most regions of Ethiopia have both rural and urban population [18]. The following formula was used to estimate average costs per treatment in each region:
Average cost per treatment = (45 ∗ x) + (45 ∗ 7.2 ∗ y),
where x is the proportion with urban residency while y is the proportion with rural residency.
The studies we relied on to estimate the costs per treatment costed the intervention from a providers’ perspective. Cost items classified as personnel costs, capital costs or supply costs were included. These were further divided into patient care costs and overhead costs [17]. In our Markov modeling, we did not include initial training of health personnel or capital costs. Costs were discounted at a 3 percent rate.
Estimation of the intervention effects on health and health inequalities
We modeled effects of scaling up coverage of CCM as life expectancy gains for children less than five years of age. The children who recovered from pneumonia were assumed to continue to live with the same health risks as the overall population. We did not apply disability weights to the effect measure as we were primary interested in mortality reduction afforded by the intervention. Hence, the incremental effects of the intervention represent gains in life expectancy at birth. We half-cycle corrected and discounted the effects at a 3% rate. The incremental cost-effectiveness ratios were calculated by dividing incremental costs by incremental life years gained.
The Gini coefficient is a measure of inequality (here applied to health), represented by a number between 0 and 1 where 0 represents absolute equality in the distribution of a chosen variable, and 1 represents a situation in which all of the chosen variable belong to one individual. Gini coefficients can be used to describe inequalities in life expectancy between individuals or population groups [19]. We used the DASP extension of the STATA software to calculate the Gini coefficients quantifying inequalities in life expectancy between the regions and between individuals within each region. In our calculations of geographical inequalities, we applied model results on regional life expectancies at birth as the health variable, and estimated numbers for children born in 2016 as the size variables. Data from survival curves provided by the Markov models were applied as the size and health variables for calculation of interindividual health inequalities.
Scale-up scenarios
In the regional scenario analysis, we explored three possible objectives: health maximization, reducing geographic inequality, and universal scale-up. The first two scenarios have lower costs and could be seen as possible pathways to universal scale-up. We estimated incremental costs, reduction in national U5MR, incremental effects, and interindividual GINI impacts of 1) maximizing health by scaling up to 90% coverage in the six regions where the intervention is the most cost-effective, 2) reducing geographic inequality by scaling up to 90% coverage in the three regions with the highest under five mortality rates, and 3) universally scaling up to 90% coverage in all regions. For all scenarios, we estimated expected health impacts across Ethiopia by adding the weighted averages of the effects observed in each region.