Treatment-seeking rates in malaria endemic countries

Many DHS, MIS and MICS ask questions to determine the prevalence and treatment of fever in children less than 5 years of age. Here it was assumed that the treatment-seeking rates observed in children would be similar in older age groups, as supported by evidence from India, Indonesia and Ethiopia [16, 25]. For a handful of DHS surveys (n = 4) and earlier MICS rounds (MICS 3 and 4, n = 45), data on treatment-seeking for fever were not available, so data on treatment-seeking for acute respiratory infection (ARI) were used, which have previously been shown to correlate strongly with treatment-seeking for fever [25]. For surveys that contained treatment-seeking data, the survey codes for questions regarding where treatment was sought were reviewed and categorized into public/government facilities (likely to have been captured by reporting systems) or ‘any’ medical treatment, which included private or NGO facilities, but excluded non-medical categories of care such as homeopathic doctors or ‘healers’. To generate a comprehensive database of treatment-seeking information, no time restriction was placed on the year of the survey. However, only MICS conducted in the third survey round or later (from 2005) were included due to inconsistency in interview questions on where treatment was sought within earlier surveys.

From each survey containing data on fever (or ARI) treatment-seeking, the total number of children reported to have fever was summarized by cluster and region. The numbers seeking treatment at either category of facility were also totalled to obtain the proportion of those who seek public/government facility-based treatment or any treatment. These data extractions from the DHS and MICS websites [26, 28] and summaries were all automated using Feature Manipulation Engine (FME) version 2015 by Safe Software [29]. MIS data were extracted from the DHS platform and will be discussed as part of the DHS survey from here onward. The total number of children, fever cases and cases that sought treatment were then summarized nationally in order to generate predictions for those countries where survey data were unavailable. National survey data require the incorporation of sampling weights, which are provided in the survey results and help to adjust for differences in the probability of selection to produce a more accurate representation of population-based metrics. Sampling weights were extracted and applied at the individual level following DHS guidelines [30], such that the totals and proportions reported here were consistent with the same metrics (percentage of fever or illness) provided in the official survey reports.

Potential covariates were identified for inclusion in a predictive model using a review of the literature for ‘treatment-seeking’ and ‘care-seeking’ for both ‘malaria’ and ‘fever’ in PubMed on 18 July 2015 [31]. The following have been reported as determinants of care-seeking rates: household wealth [15, 17, 19, 20, 22, 32‐38], care-giver education [15, 19, 32, 34, 39] and household location (rural or urban) or access to health facilities [15, 20‐22, 32, 33, 40‐45]. The World Bank provides freely available national-level indicator data [46], and several indicators that were in keeping with the themes identified by the literature were downloaded: access to electricity (as a proxy for wealth and access to health facilities), gross domestic product (GDP; current US$), GDP per capita, GDP growth (annual %), gross national income (GNI) per capita (current US$), total health expenditure (% of GDP), public health expenditure (% of total), primary education completion rate (% of relevant age group), and rural population (% of total population). In addition to these metrics, the number of health workers was thought to influence treatment-seeking rates. However, there were not sufficient data on the number of community health workers per 1000 people available from the World Bank, so the number of nurses and midwives per 1000 population was used instead. Finally, explicit care-seeking variables were also included: the percentage of pregnant women receiving prenatal care and the percentage of children aged 12–23 months who were immunized against diphtheria, pertussis and tetanus (DPT).

Matching covariate data to national-level survey data was also achieved using FME. When available, covariate data were used from the same year as the survey, and otherwise were matched to the closest year. Covariate data were also matched to those countries without treatment data available. For these countries, covariate data was matched to 2013, which was the most recent year of indicator data available, or the closest year to that in order to generate the most up to date estimates of treatment-seeking rates.

In addition to the social and economic covariate data, countries were grouped geographically based on their WHO regional offices: Region of the Americas (PAHO), Eastern Mediterranean Region (EMRO), European Region (EURO), Southeast Asia Region (SEARO), and Western Pacific Region (WPRO). Countries in the African region were separated into the sub-African regions reported in the World Malaria Report: West Africa (AFRO-W), Central Africa (AFRO-C), East Africa and high-transmission areas in Southern Africa (AFRO-E), and low-transmission Southern African countries (AFRO-S) [1]. These regions then formed strata within the model, as explained below. A map of georeferenced treatment-seeking data availability from national surveys in MECs is shown in Fig. 1.

Two generalized additive mixed models (GAMMs) were developed to predict the proportion of patients that sought treatment from (1) a facility covered by the government reporting system, and, (2) any type of medical care provider [47, 48]. Where multiple surveys were used from the same country, this was accounted for by adding a country-level random effects term. Both year and WHO region were included as fixed effect terms, along with the suite of country-level covariates. The list of the covariates described above was limited after testing for collinearity among all the potential variables. The variables included in GAMMs were: year of survey, WHO region, GDP growth, health expenditure, prenatal care rates, primary education, DPT immunization rates, nurses and midwives per population and proportion rural population. All statistical analyses were performed in the R statistical computing environment [49]. Full details of the model development are provided in Additional file 1.

Existing treatment-seeking data were mapped at the regional level (boundaries shown in Fig. 1) for the 76 countries with available survey data. Although some of these data were available at a point level, only regional maps were generated for consistency with the other adjustment parameters. For the few countries with DHS cluster data and no regional data available, cluster level data were summarized and mapped to the first- (Colombia) or second-level (Zambia) administrative units [50]. Predicted and observed values were then combined into a single, geographically complete, map of treatment-seeking for all MECs showing the most recent available data in each country.

Treatment-seeking data were collected from DHS (n = 195, 13 of which were MIS) and MICS (n = 59) for 76 countries. The number and year of surveys available from MECs are shown in Additional file 1: Figure S1. Data from DHS and MICS round 5 collected data on treatment-seeking for fever, whereas MICS rounds 3 and 4 assess care-seeking for cough (respiratory infection). Additional file 1: Figure S2 confirms that rates observed from the earlier MICS rounds are comparable to the treatment-seeking for subsequent surveys. Indicator data were available for all but nine surveys (there was no health expenditure data available for Zimbabwe and Somalia); these nine surveys could not therefore be included in the analysis. DHS surveys conducted prior to 1990 were excluded as they reported that no individuals sought treatment, and were therefore considered to be non-representative. Two MICS surveys, Yemen (2006) and Burkina Faso (2006), also reported zero treatment-seeking and were excluded. The GAMMs were thus fitted to 228 records from 72 countries. This left 22 MECs for which predictions of treatment-seeking outcomes were required. However, Brazil and Paraguay were also added to the prediction list since Brazil only had national level data available (DHS 1996) and Paraguay had large areas of missing data (DHS 1990). This increased the number of prediction countries to 24.

Two models were chosen as the best candidate models for treatment-seeking from public facilities and five for seeking any treatment. The fitted coefficients, AIC, ΔAIC and ω_i values for each of the best models are shown in Table 1. Model average coefficient values for each geographic region are shown in Additional file 1: Table S1. The parameter with greatest model influence was the percentage of pregnant women receiving prenatal care, followed by DPT immunization rates for pubic treatment-seeking and primary education completion for any treatment. Geographic regions were also statistically significant, as was public health expenditure. Time was also included as an indicator variable because all available surveys with treatment-seeking information were included (1990 and 2005 onwards for DHS and MICS, respectively. Plots of observed treatment-seeking rates over time (see Additional file 1: Figure S3) confirmed that rates of treatment-seeking from any type of facility and particularly from government-run facilities increased over time. Model validation showed good prediction performance (Additional file 1: Figures S4–S6). The root mean square error (RMSE) for the percentage seeking treatment from government facilities was 11.9 and 12.3 % for any treatment facility type.

All observed and predicted values, as well as covariate data, are shown in Additional file 2. There were five countries for which insufficient covariate data precluded predictions. These countries were therefore assigned estimates from a similar and ideally neighbouring country in the same region (rounded to two decimal places to show that they are less precise). Myanmar (Burma) was matched to Timor-Leste, which had the most similar health indicator variables (DPT and pregnant women care). In the absence of indicator data or neighbouring countries in the same WHO region the Democratic People’s Republic of Korea (DPRK) was given the same estimate as Myanmar (Burma). Eritrea was given the same value of neighbouring Ethiopia, French Guiana was given Suriname’s values, and the Republic of Korea was given China’s estimate, which was closest to Korea for the indicators that were available.

The mean predicted values for the proportion seeking treatment (government-based or other) in the 24 countries without survey data available are shown alongside the post-2010 observed treatment-seeking values in Fig. 2. The 95 % CI of the observed measures were obtained from the standard error of national weighted mean. The predicted measures are shown with black points, but are also distinguishable by wider uncertainty resulting from the model predictions being made from sampling within the confidence intervals of the observed measures. In spite of their greater uncertainty, the predicted point estimates follow the pattern of the observed estimates in each region. These plots reveal some areas where the proportion of treatment-seeking from government-based facilities is markedly lower, notably in SEARO. Greater heterogeneity in estimates of both types of treatment-seeking was observed among the countries in the African regions.

Despite variability between geographic regions, treatment-seeking rates have generally increased over time. The non-linear relationship between treatment-seeking and time is shown in Additional file 1: Figure S3. The lowest rates of access to government treatment (<10 %) were observed prior to the year 2000, with the exception of findings from DHS surveys from Chad (2004), Bangladesh (2004 and 2007) and Pakistan (2007) and a MICS from Somalia in 2006 which were all also <10 %. The lowest reported rate of seeking any treatment was 24 % in Togo in 2006. All observations of treatment-seeking rates less than 40 % (from any health facility) were from the African continent (including Somalia, from EMRO). The highest measures of access to public facilities (>65 %) were observed in several regions: AFRO-C (Nigeria and Sierra Leone), AFRO-E (Zambia and Mozambique), PAHO (Costa Rica, Dominica Republic and Peru) and SEARO (Bangladesh (in 2011, but not in the earlier years referenced above), Bhutan, Indonesia and Timor-Leste). Bhutan had the highest reported measure of 74 % of individuals seeking care at government-based facilities (MICS 2010). High rates of access to any treatment type likewise spanned various regions, but values were far greater than public-facility data only. Regions with observations of access rates to any treatment greater than 80 % include: AFRO-C (Nigeria), AFRO-E (Tanzania and Uganda), EMRO (Pakistan), PAHO (Argentina, Belize, Costa Rica, Dominican Republic and Panama), SEARO (Indonesia (since 1991) and Thailand) and WPRO (Cambodia and Viet Nam).

Regional comparisons of treatment-seeking rates are shown in Table 2. Weighted estimates were obtained by multiplying the most recent observed or predicted value for each of the 98 MECs by the proportion of the regional population in each country. Population data from 2013 [51] were summarized nationally in ArcGIS [52]. While SEARO has lowest overall government-based treatment-seeking behaviours, this region also has the highest access to any treatment. Patterns within the other regions were not as marked, but access to treatment including facilities outside the formal government sector was overall higher.

The 19 predicted and five assigned national treatment-seeking estimates were combined with those countries that already had data available. Figure 3 shows maps for both the proportion seeking government treatment (Panel a) and those seeking any treatment (Panel b). Note that for both regional comparisons and mapping, priority was given to DHS surveys if both a MICS and DHS survey were reported within the same four years. DHS data are georeferenced which allows for sub-national mapping and are thus considered more representative and facilitate more extensive potential downstream analyses. The maps illustrated the patterns exhibited in the country-level (Additional file 2) and regional estimates (Table 2). Government-based treatment was shown to not be well accessed in large parts of SEARO, EMRO, AFRO and small patches of PAHO. The map of treatment rates from any sector highlighted areas where treatment of was low overall.