Ethical considerations
Data used in these analyses were available through the DHS Programme [20]. Ethical review and approval for procedures and questionnaires for standard DHS surveys is provided by the ICF Institutional Review Board (IRB). Country-specific DHS survey protocols are reviewed by the ICF IRB and typically by an IRB in the host country. Verbal consent is obtained from the participant and a signature is provided by the interviewer to acknowledge that this event has taken place. Displaced geographical coordinates were obtained following approval from the DHS Programme. Data were securely stored separately from individual and household data.
Data sources
The DHS programme conducts standardized, nationally-representative surveys in over 90 countries worldwide, collecting data pertaining to the broad themes of fertility, family planning, maternal and child health, human immunodeficiency virus (HIV), malaria, and nutrition [21]. The methods of the 2013 Namibia DHS are detailed elsewhere [18]. In summary, the survey used a two-stage stratified cluster design, which involved dividing each administrative region into enumeration areas (EAs) and then classifying these EAs as either urban or rural. EAs were then selected from the urban and rural strata and around 20 households per EA were selected for the survey [18]. The DHS involved three surveys: the Household survey, the Woman’s survey and the Man’s survey [18]. The household wealth index was calculated using principal component analysis involving economic indicators such as household assets [22, 23].
Available data on vector control indicators, collected as part of the DHS Household survey, included data pertaining to ITNs and IRS. A household member was asked to show all the mosquito nets to the interviewer and identify which household members slept under each net the night before the survey. IRS coverage was determined by asking a household member if the dwelling had been sprayed against mosquitoes in the last 12 months. DHS definitions of IRS and ITN were as follows:
-
Indoor residual spraying Spraying of the interior walls of the dwelling with an insecticide against mosquitoes.
-
Insecticide-treated net A factory-treated net that does not require any further treatment (LLIN), or a pre-treated net obtained in the past 12 months, or a net that has been soaked with insecticide within the past 12 months.
Households were classified as not having an ITN if the household did not have any mosquito net or only had untreated nets. Households with at least one ITN per two people who slept in the household the night before the survey were classified as having a sufficient number of ITNs.
EA coordinates were obtained from the DHS Programme. EA coordinates represent a group of up to 20 households and are randomly displaced. Rural EAs are randomly displaced by up to 5 km and urban EAs are displaced by up to 2 km [24].
The indicator Plasmodium falciparum parasite rate (PfPR) is a commonly used indicator of malaria transmission intensity. PfPR2–10 is the proportion of the population aged 2–10 years carrying asexual blood parasites [25]. Modelled malaria parasite prevalence data for the year 2013 were obtained from the Malaria Atlas Project (MAP) portal, made available under the Creative Commons Attribution 3.0 Unported License [26, 27]. MAP PfPR2–10 estimates were derived from data collected across 27,573 population clusters from 1995 to 2014, which were adjusted for age, season and the diagnostic test used [28]. This model was used to predict PfPR2–10 for malaria-endemic countries across Africa, including Namibia, from the year 2000 to 2015, at a resolution of 5 × 5 km [28].
Malaria zones were assigned in line with MSP district strata outlined in the MSP documentation [12]. As part of Namibia’s 2010–2016 MSP, the objective for integrated vector control was to achieve at least 95% coverage with a combination of vector control interventions in all malaria endemic areas and identified transmission foci by 2013 [12]. The country was divided into three Zones, with Zone 1 representing the highest transmission areas (moderate transmission risk), Zone 2 representing low transmission risk and Zone 3 for “risk free” areas [12]. Vector control targets were set for each zone. In Zone 1 the aim was to achieve 95% coverage of a combination of IRS and ITNs in addition to winter larviciding [12]. In Zone 2 IRS, ITNs and larviciding were to be targeted to selected foci [12].
For spatial representations of data, shapefiles for Namibia were downloaded from DIVA-GIS [29], originally sourced from the Database of Global Administrative Areas (GADM) [30].
Data analysis and statistical methods
Quantum GIS (QGIS) 2.14.1 was used for all maps and spatial analyses. All statistical analyses were carried out using STATA 14.0 software package (StataCorp: College Station, TX, USA). All households captured in the survey period (May to September 2013) were included in the subsequent analyses, giving a total of 9846 households and a population of 41,314 individuals.
Three models of transmission intensity were constructed. The first classified households according to weighted regional PfPR2–10 values obtained from MAP for the year 2013. Regions were classified into three categories based on their PfPR2–10 values. The < 1% category constitutes very low transmission risk or malaria-free areas, the 1 to < 5 % category represents low transmission risk and the ≥ 5% category signifies moderate risk of transmission. Regions with PfPR2–10 estimates of zero (malaria-free) were classified into the < 1% category.
The second model used raster data for PfPR2–10 obtained from MAP for the year 2013. PfPR2–10 values for each EA were assigned using the “Point Sampling Tool” in QGIS 2.14.1 [31]. Raster values were converted to percentages and were similarly classified into three PfPR2–10 categories: < 1; 1 to < 5 and ≥ 5 %. Where no raster values were available for EAs because they were located in areas where no transmission was predicted to occur, the EAs were assigned the value of zero. To account for random displacement in DHS data, Euclidean buffers were drawn around EA points of 2 km for urban EAs and 5 km for rural EAs. The MAP PfPR2–10 raster surface was overlaid with buffered EA locations and the mean PfPR2–10 value was extracted. A high correlation between extracted mean PfPR2–10 values and extracted point PfPR2–10 values was observed. EAs were re-categorized into PfPR2–10 categories (< 1, 1 to < 5, > 5%) according to the mean PfPR2–10 values.
In additional sensitivity analyses, EAs outside of the boundary of the PfPR2–10 raster were assigned the value of the nearest raster cell up to 5 km away, relative to the maximum EA displacement distance. This was repeated to assign EAs up to 10 and 20 km outside of the raster boundary the value of the nearest cell. EAs were re-categorized into PfPR2–10 categories (< 1, 1 to < 5, > 5%) and explored the coverage of IRS, having an ITN and having either intervention for the three models respectively (assigning raster cell values to EAs up to 5, 10 and 20 km away).
The third model classified households according to MSP zones. Zones were assigned using QGIS 2.14.1. Administrative districts were assigned zones 1, 2 or 3, as defined by the MSP, and EAs were mapped. To assign zones to EAs, polygon attributes were assigned to the EA points using the QGIS 2.14.1 “Join Attributes by Location” tool.
Categorical data are presented as a frequency and percentage. p-values were calculated using a Chi squared test and p < 0.05 was considered statistically significant. Primary analyses were unweighted but additional weighted analyses were carried out to make the data representative of the whole population. Weighted analyses used the DHS weight variable as per DHS Programme guidance [32]. First a univariable Poisson model (STATA ‘poisson’ function) was used to test for the association between IRS and regional PfPR2–10. In the second model, EA and region were added as mixed effects (STATA ‘mepoisson’ function). In the third model, wealth and residence type covariates were additionally adjusted for. These analyses were carried out for the other outcomes of interest: whether a household owned at least one ITN, and whether a household had at least one intervention (ITN and/or IRS). Risk ratios are presented with 95% confidence intervals and the p-value.
Log-likelihood ratio tests were carried out to compare regional PfPR2–10, EA PfPR2–10 and MSP zones. The first model tested the association between regional PfPR2–10 and IRS, adjusted for covariates (wealth and residence type) and accounted for regional and EA clustering. The second model additionally adjusted for EA PfPR2–10. The third model adjusted for MSP zones in addition to model 1. Log-likelihood ratio tests were carried out with models 2 and 3, respectively nested in model 1.
Log-likelihood ratio tests were repeated for the additional models of EA PfPR2–10. The mean EA PfPR2–10 model was compared to the regional PfPR2–10 model for each intervention using log-likelihood ratio tests. First, the association between regional PfPR2–10 and IRS was tested, adjusting for regional and EA clustering, wealth and residence type. The second model additionally adjusted for the mean EA PfPR2–10 and a log-likelihood ratio test was conducted with the second model nested in the first. This was repeated for the association with having an ITN and either intervention.
Further, EA PfPR2–10 models, where EAs were assigned raster cell values at up to 5, 10 and 20 km away, were compared to the regional PfPR2–10 model, respectively, for each intervention (IRS, ITN and either intervention). First, the association between regional PfPR2–10 and having IRS was tested, adjusting for regional and EA clustering, wealth and residence type. The second model additionally adjusted for EA PfPR2–10 and a log-likelihood ratio test was carried out with the second model nested in the first. This was repeated for each model of EA PfPR2–10 and for each intervention.