Mapping tuberculosis treatment outcomes in Ethiopia

Background: Tuberculosis (TB) is the leading cause of death from an infectious disease in Ethiopia, killing more than 30 thousand people every year. This study aimed to determine whether the rates of poor TB treatment outcome varied geographically across Ethiopia at district and zone levels and whether such variability was associated with socioeconomic, behavioural, health care access, or climatic conditions. Methods: A geospatial analysis was conducted using national TB data reported to the health management information system (HMIS), for the period 2015-2017. The prevalence of poor TB treatment outcomes was calculated by dividing the sum of treatment failure, death and loss to follow-up by the total number of TB patients. Binomial logistic regression models were computed and a spatial analysis was performed using a Bayesian framework. Estimates of parameters were generated using Markov chain Monte Carlo (MCMC) simulation. Geographic clustering was assessed using the Getis-Ord Gistatistic, and global and local Moran’s I statistics. Results: A total of 223,244 TB patients were reported from 722 districts in Ethiopia during the study period. Of these, 63,556 (28.5%) were cured, 139,633 (62.4%) completed treatment, 6716 (3.0%) died, 1459 (0.7%) had treatment failure, and 12,200 (5.5%) were lost to follow-up. The overall prevalence of a poor TB treatment outcome was 9.0% (range, 1-58%). Hot-spots and clustering of poor TB treatment outcomes were detected in districts near the international borders in Afar, Gambelia, and Somali regions and cold spots were detected in Oromia and Amhara regions. Spatial clustering of poor TB treatment outcomes was positively associated with the proportion of the population with low wealth index (OR: 1.01; 95%CI: 1.0, 1.01), the proportion of the population with poor knowledge about TB (OR: 1.02; 95%CI: 1.01, 1.03), and higher annual mean temperature per degree Celsius (OR: 1.15; 95% CI: 1.08, 1.21). Conclusions: This study showed significant spatial variation in poor TB treatment outcomes in Ethiopia that was related to underlying socioeconomic status, knowledge about TB, and climatic conditions. Clinical and public health interventions should be targeted in hot spot areas to reduce poor TB treatment outcomes and to achieve the national End-TB Strategy targets. Ethiopia is located in East Africa and shares boundaries with Eritrea to the North, Djibouti and Somalia to the east, Sudan and South Sudan to the West, and Kenya to the South. It is the second-most populous country in Africa. The last official census report, which was undertaken in 2007, estimated a total population of 74 million [12]. Subsequent projections raised the population estimate to 102 million people in 2017 [13]. Approximately 85% of the population live in rural areas [12]. With a total area of about 1.1 million square kilometres, Ethiopia is the 10th and 27th largest country in Africa and in the world, respectively. The country has a variety of geographical features; its altitude ranges from 125 m below sea level at the Danakil Depression to 4620 m above sea level at mount Ras Dashen, it contains the source of the Blue Nile and is bisected by the Great Rift Valley. Ethiopia is administratively divided into nine regional states and two city administrative councils (Addis Ababa and Dire Dawa). Each regional state is further divided into zones, districts (“woreda”), and neighbourhoods (“kebeles”). The districts are the decentralized administrative level and kebeles, within districts, are the lowest administrative unit in Ethiopia. TB is the leading infectious cause of death in Ethiopia [14], killing more than 30 thousand people annually [15, 16]. Ethiopia is one of 30 high TB and MDR-TB burden countries [17]. The national population based TB prevalence survey conducted in 2010–2011 revealed that the prevalence of smear-positive TB among adults and all age group was found to be 108 and 63 per 100,000 head of population, respectively [18]. According to WHO’s 2014 Global TB report, Ethiopia has achieved all the three targets of the Millennium Development Goals (MDG) for TB prevention and control [19]. Mortality and prevalence due to TB has declined by more than 50% and the incidence rate is falling significantly [19]. However, in the last few years, the TB treatment success rate has decreased in Ethiopia (from 91% in 2012 to 84% in 2016) [2, 20, 21]. District-level TB data (reported from June 2015 to June 2017) were obtained from the Health Management Information System (HMIS) managed by the national TB and leprosy control program (NTLCP). The data include the number of TB patients enrolled in directly observed therapy, short-course (DOTS) centres in each district with their treatment outcomes (including the number of patients with the TB treatment outcomes of cured, treatment completed, died, treatment failed, and lost to follow-up). TB treatment outcomes were defined according to NTLCP guideline definitions [22], which have been adopted from WHO TB treatment guidelines [23]. In this study, a poor treatment outcome was defined as the sum of failure, death and lost to follow up, and treatment success was defined as the sum of cured and treatment completed. The study included pulmonary and extra-pulmonary drug-susceptible TB. Table 1 shows the definition of TB treatment outcomes. definitions of treatment outcomes for tuberculosis patients Data sources for the independent variables are provided in Table 2. Data on the knowledge and attitudes of the population regarding TB were obtained from the EDHS 2011 [24]. Data on health care access and data on behavioural factors such as chat chewing and alcohol drinking were obtained from the 2016 EDHS [25]. Socio-economic data such as population density, dependency ratio, average number of people in the household, unemployed rate, and literacy rate were obtained from the Ethiopia Atlas of Population Density, and the wealth index was obtained from the 2016 EDHS [25]. Climatic and environmental data such as Enhanced Vegetation Index, rainfall, aridity, and mean temperature were obtained from the EDHS Spatial Analysis repository [26]. Summary of independent variables, sources of data and definition of variables Data on the knowledge and attitudes of the population regarding TB were obtained from the EDHS 2011 survey (as the EDHS 2016 survey did not include these data) [24, 25]. These data were collected by semi-structured questions by means of an interviewer-administered questionnaire. The TB knowledge of a person was assessed by three questions: 1) “have you ever heard of an illness called tuberculosis or TB (yes/no)”; 2) “how can a person get tuberculosis or TB?” and 3) “what symptoms will a person with tuberculosis or TB have?” A person was categorised as having “good” knowledge about TB if the person had ever heard about TB, correctly mentioned the route of transmission (i.e. TB is transmitted through the air when coughing or sneezing or through drinking of unboiled milk), and if the person mentioned at least one TB symptom (i.e. persistent cough for more than 2 weeks, weight loss, poor appetite, night sweats, chest pain or fever). Those who missed one or more of these three items were categorized as having ‘poor’ knowledge about TB. The attitude of a person towards TB was measured by two questions: 1) can tuberculosis or TB be cured (yes, no, or don’t know) and 2) if a member of your family got tuberculosis or TB, would you want it to remain a secret? A “good” attitude was defined by a person believing that TB can be cured and not wanting a family member’s TB to be kept a secret. A “poor” attitude was defined by a person believing that TB cannot be cured or wanting a family member’s TB diagnosis to be kept a secret. The health care access data were obtained from the 2016 EDHS [25]. As maternal and newborn health are priorities for the Government of Ethiopia [27], in the 2016 EDHS, the health care access data were collected only from women. Women (aged 15–49 years) were asked whether payment for advice or treatment, or the distance to a health facility, presented major problems in seeking medical advice or treatment for themselves when they were sick. We considered that healthcare access issues existed if either of these challenges were identified by the participants. A total of 15,683 women responded to these questions, of whom 9479 (60%) reported having at least one of the specified problems in accessing health care. Data on behavioural factors such as chat chewing and alcohol drinking were obtained from the 2016 EDHS [25]. Chat chewing was assessed in the 2016 EHDS by asking the questions: 1) “have you ever chewed chat (yes/no)”; and 2) “during the last 30 days how many days did you chew chat?” Similarly, alcohol drinking was assessed by questions such as: 1) “have you ever taken a drink that contains alcohol (Tella/Tegi/Areke/Beer/Wine, etc…) (yes/ no)”; 2) “during the last 30 days, how many days did you have a drink that contains alcohol?” and 3) “during the last 13 months, how often did you take a drink that contains alcohol (almost every day, at least once a week, less than once a week, none in the last 13 months)?”. Since the smoking of cigarettes was rare among the 2016 EDHS participants (less than 1% of women and 4% of men smoke any type of tobacco), we did not include cigarette smoking in our study as an independent variable. Socio-economic data such population density, dependency ratio (defined as number of children (aged under 15 years) and elderly (aged 65+) divided by the working-age population (aged 15–64 years)), average number of people in the household, unemployed rate, and literacy rate were obtained from the Ethiopia Atlas of Population Density, and the wealth index was obtained from the 2016 EDHS [25]. The wealth index was calculated in the 2016 EDHS at the household level. Households were given scores based on the number and kinds of assets they own, ranging from a television to a bicycle or car, in addition to housing characteristics such as source of drinking water, toilet facilities, and flooring materials. These scores were derived using principal component analysis. National wealth quintiles were compiled by assigning the household score to each usual (de jure) household member, ranking each person in the household population by her or his score, and then dividing the distribution into five equal categories, each comprising 20% of the population. We considered those in the poorest and poorer wealth index quintiles as having a low wealth index and those in the middle, richer, and richest wealth quintiles as having a high wealth index. We then calculated the percentage of patients with a low wealth index in each zone as the total number of people with a low wealth index divided by the total number of people who had information on assets in the zone. Climatic and environmental data such as Enhanced Vegetation Index, rainfall, aridity, and mean temperature were obtained from the EDHS Spatial Analysis repository [26]. The EDHS program georeferenced these climatic and environmental data with the existing demographic and health survey data to use for spatial analysis. The prevalence of poor TB treatment outcomes was calculated by dividing the sum of death, treatment failure and lost to follow up by the total number of TB patients enrolled in the DOTS program. First, a univariate analysis was performed by taking the prevalence of poor TB treatment outcomes as the dependant variable and the geographical covariates as independent variables. Since the prevalence of poor TB treatment outcomes at zonal level was skewed to the right, the log transformed prevalence of poor TB treatment outcome was used in the univariate analysis. Variables with a p-value less than 0.2 in the univariate analysis were selected for the final model and checked for the presence of multi-collinearity using variance inflation factors (VIF), excluding variables with a VIF > 7. When we checked for the presence of multi-collinearity, a high degree of collinearity was observed among variables from within the same group of independent variables. Thus, we selected one variable from each group for the final model to avoid the observed multi-collinearity. The variables for the final model were selected first by eliminating implausible variables and then by using a p-value. The variable that was found to be more statistically significant was selected for the final spatial models. Low wealth index, chat chewing, poor knowledge towards TB, and annual mean temperature were selected for the final spatial models. Spatial clustering of poor TB treatment outcomes was explored at a global scale using Moran’s I statistic and at a local scale using Local indicators of spatial association (LISA), estimated using the Anselin Local Moran’s I statistic, and the Getis-Ord statistic. The global Moran’s I statistic was used to assess the presence, strength and direction of spatial autocorrelation over the whole study area and to test the assumption of spatial independence in the implementation of the spatial pattern analysis. The LISA and the Getis-Ord statistics were used to detect local clustering of poor TB treatment outcomes and to identify the locations of hot-spots. These analyses were conducted using tools provided in ArcGIS. Four different Bayesian models were constructed: unstructured model (Model I), spatially structured model without covariates (Model II), spatially structured model with covariates (Model III), and spatially structured and unstructured model with covariates (Model IV). Model IV includes all the components in the preceding models. These models were constructed using WinBUGS version 1.4.3 software (Medical Research Council Biostatistics Unit, Cambridge, United Kingdom). The details of the models are presented in the Additional file 1: Table S1. We assumed that the observed prevalence of poor TB treatment outcome (r) at zone (i) had a binomial distribution with a total number of TB patients enrolled for treatment at the zone (ni) and the predicted prevalence of poor treatment outcomes at the zone (pi): ri ~ Binomial (ni, pi). Model IV for pi was specified as follows: Where pi is the probability of a poor treatment outcome in zone i; α is the intercept; ∑Nβn∗Xn,i is the matrix of independent zone-specific variables (X, i.e. proportion of the population with a low wealth index, chat chewing and poor knowledge towards TB, and mean annual temperature) measured at each zone i, multiplied by their coefficients (β); Ui are unstructured random effects; Vi are the spatially structured random effects, modelled using conditional autoregressive (CAR) approach. The CAR component was defined using an adjacency matrix to determine the spatial relationships between zones. The adjacency matrix for each zone was generated using ArcGIS based on the queen definition, whereby two areas are considered neighbours if they share a common boundary or vertex. A weight of 1 was given if two zones were neighbours and a weight of 0 was given if two zones were not neighbours. The posterior parameters were estimated using a Bayesian Markov Chain Monte Carlo (MCMC) simulation. Non-informative priors were specified for the intercept α (a non-informative, improper prior with bounds – ∞ and + ∞) and the coefficients (normal prior with mean = 0 and precision 1× 10− 6). The priors for the precision of the unstructured and spatially structured random effects were assigned non-informative gamma distributions with shape and scale parameters set at 0.001. The deviance information criterion (DIC) statistic was calculated to select the best-fitting model (models with a lower DIC statistic are considered to show a better compromise between model fit and parsimony). The model was run for 1,000,000 iterations and convergence was successfully achieved after 900,000 iterations. Convergence of the models was determined by visual inspection of posterior kernel density and history plots.