Study population
The malaria prevention programme was implemented in all villages in a pre-specified 100 × 40 kilometre corridor along the Angolan–Namibian border in the Trans-Kunene region. The Trans-Kunene region comprises of two Angolan provinces—Cunene and Namibe—and three Namibian regions—Kunene, Ohangwena, and Omusati. Target districts for the malaria prevention programme was jointly decided upon by the Ministry of Health in Angola and Namibia. Following this, the intervention corridor was defined by the study team under the assumption that most cross-border infections would occur in this spatial area as movement across the border happens primarily on foot or bicycle. The intervention corridor encompassed Cunene province in Angola, and Ohangwena and Omusati regions in Namibia. Additional file 1 highlights the corridor in the Trans-Kunene region in which the programme was implemented.
The two Angolan provinces in the TKMI represent areas of lower and unstable malaria endemicity in comparison to the rest of the country [11]. In 2015–2016, malaria prevalence among children under-five was reported to be 1% or less in Cunene and Namibe provinces, which was in stark contrast with the neighbouring province of Cuando-Cubango where prevalence was reported to be approximately 38% [12]. Access to preventive measures such as insecticide-treated nets, indoor residual spraying, and malaria treatment was particularly low in the two Angolan provinces in the TKMI.
The three TKMI regions in Namibia have sustained levels of malaria receptivity and account for a majority of malaria cases reported in the country [13, 14]. In contrast, the south of the country is malaria free while the incidence of malaria in the central regions is low [13]. Overall, malaria incidence and mortality has been decreasing in Namibia, which is largely attributable to increased distribution of long-lasting insecticide-treated bed nets (LLITNs), and improved access to malaria treatment [14].
Community-based malaria prevention intervention
The malaria prevention programme was conducted by community-based volunteers and involved the distribution of LLITNs and behavior change programming. The volunteers did not receive any monetary incentives; instead, they were given an end-of-year food basket, TKMI t-shirts, and light refreshments on the day of net distribution.
For bed nets, community volunteers first compiled a listing of all inhabited structures in programme villages independent of construction type (households). Following this, they delivered and assisted with the installation of rectangular, World Health Organization Pesticides Evaluation Scheme (WHOPES) approved LLITNs in all listed households. Each household received one LLITN per sleeping space. Informal sleeping spaces outside households were not considered during distribution as all families sleep inside their homes at night because of proximity to the land they work in. The distributed LLITNs have an expected lifespan of 3–4 years and have been demonstrated to retain effectiveness up to 20 washes [15]. Systematic reviews have also shown that insecticide treated nets can reduce incidence of Plasmodium falciparum malaria episodes by 50% [16].
Behaviour change programming involved community-based volunteers making monthly visits to households to provide guidance on the use and maintenance of LLITNs. In addition, they also provided important malaria prevention information such as strategies for eliminating standing water.
The malaria prevention programme followed a two-phased block rollout schedule between 2014 and 2016. During Phase I, which occurred between 2014 and 2015, all 35 Namibian villages in the intervention corridor as well as 50% of purposely selected areas in Angola received the treatment. As illustrated in Additional file 2, the Angolan corridor was divided into four approximately equally sized zones. For the Phase I rollout, two of these zones were selected to receive the treatment while the remaining zones were used as control areas. There were no buffer zones between these areas as the primary objective of this study was to measure treatment effect spillovers. Instead, rollout was blocked, so that we could directly measure spillovers in Phase I villages based on whether they bordered areas targeted in Phase II.
The rollout of bed nets was centrally controlled and coordinated across both countries. Thus, during Phase I, the delivery and installation of nets occurred at the same time on both sides of the border. Villages that did not receive the treatment in Phase I received it in Phase II, which was conducted between 2015 and 2016. In both phases, most bed nets were distributed and installed before the peak rainfall months of February and March, which roughly coincides with the period of highest malaria burden in the TKMI area.
Sampling
A random sample of 64 villages in the 100 × 40 km corridor was selected for evaluation prior to programme rollout. The sampling of villages differed by country: in Namibia, a complete list of all 35 village names in the intervention corridor was provided by the local Ministry of Health and a random sample of 26 villages was selected from this sampling frame to get a sample size of at least 250 households. In Angola, the study team conducted an independent listing of all 38 villages in the intervention corridor at the beginning of the study. Due to the larger study team on the Angolan side, all 38 villages were surveyed.
All study villages in the intervention corridor were situated in rural areas, with most villages comprising of 50–100 households. In each of the 64 study villages, a complete household listing was made prior to each survey round. A random sample of 10% of households was then surveyed in each village in each round.
Data collection
All 64 study villages were surveyed three times over the evaluation period. Additional file 3 illustrates the rollout of the evaluation surveys and the TKMI interventions. The first survey was conducted between May 2014 and July 2014 (baseline) prior to intervention launch. The second survey was conducted after Phase I in September 2015 (midline). The final survey was conducted after the completion of Phase II in July 2016 (endline). Additional file 4 contextualizes the rollout of the surveys in relation to the distribution of bed nets. Most of the questions in the surveys were adapted from the Demographic and Health Surveys, with some additions made by the study team. No formal validation was done for the additional questions.
Each survey consisted of a general household and knowledge module and an under-five child module focused on fever incidence and treatment. Due to budgetary and logistical constraints, malaria status of children under-five was not measured or verified by rapid diagnostic tests (RDTs). A primary respondent, typically a female household member, provided responses to questions in both modules. Of the 2184 surveys attempted over the three rounds, 97% were completed. The most common reasons for non-completion were refusals and temporary unavailability of families.
Outcome measures
The primary outcome of interest was caregiver-reported child fever. As part of all surveys, respondents were asked to first list all children in their household under the age of five, and then indicate fever episodes and treatment-seeking in the 2 weeks preceding the survey.
The secondary outcomes analysed were malaria knowledge, LLITN utilization among children under-five, and household LLITN ownership. Malaria knowledge was evaluated by asking each respondent 20 true/false questions on the nature, consequences, and treatments for malaria. For the empirical analysis, the percentage of correct responses was computed and converted into a z-score to facilitate interpretation of the estimated coefficients. To determine child LLITN utilization, respondent reports of whether each listed child slept under a LLITN on the night before the survey were used. Household LLITN ownership was defined as the number of LLITNs owned by a household and was determined based on respondent reports as well. The complete English version of the survey questionnaire is included in “Appendix 1”.
Statistical methods
Summary statistics of key household, respondent, and under-five children characteristics as well as primary and secondary outcomes of interest at baseline were computed. To assess the effectiveness of the malaria prevention programme, data across the three waves were pooled and multivariable linear and logistic regression models that controlled for survey-round fixed effects were estimated. Additionally, the overall implementation period was classified into an immediate period, i.e., the time period within 1 month of intervention implementation, and a 1-year follow-up period. Programme impact was evaluated over both timeframes as well. To determine if programme impact on child morbidity varied by age or gender, sub-group regression analyses were conducted.
Two distinct sources of variation were used to assess the extent to which programme impact depends on coordinated cross-border programming (spillover effects). For Namibian villages, variation in programme implementation in Angola during Phase I was exploited to conduct a standard difference-in-differences analysis, interacting Phase I impact in Namibia with Phase I programme activities in Angola. To allow for both linear and multiplicative interactions, linear probability and logistic regression models were estimated. In both models, the main hypothesis tested was that Namibian villages adjacent to Phase I Angolan villages would experience greater programme impact than villages adjacent to Phase II Angolan villages at midline. Figure 1 illustrates this comparison.
For Angolan villages, spatial variation in Phase I treatment status was not available since all Namibian villages received treatment in Phase I. However, since malaria programmes were largely absent north of the TKMI corridor in Angola, programme villages close to Namibia (with extensive malaria control) can be compared with Angolan villages closer to the rest of Angola (with limited programme efforts). To evaluate if distance to the border modified programme impact in Angola, a three-step process was followed: first, kilometre distance between the border and village centroids was computed; second, data were restricted to Phase I Angolan villages; and third, linear and logistic regression models interacted with distance as well as stratified regression models were estimated. For the stratified regressions, the 20 km width of the Angolan half of the corridor was divided into four roughly equally sized strata, and households within 5, 10, 15, or 20 km of the border were considered separately in each model.
To account for correlated outcomes within households and study villages, standard errors in all regression models were estimated using Huber’s cluster-robust variance estimator [17]. This procedure allows correct inference independent of model specification. Missing responses were excluded from the calculation of descriptive statistics and coefficient estimates in the regression models.
Software
All statistical analyses were conducted in Stata/MP 15.0 [18]. ArcMap 10.3.1 was used to create maps and calculate kilometre distances between villages [19].
Ethical considerations
This study was approved by the Institutional Review Board at the Harvard T.H. Chan School of Public Health, the Ministry of Health and Social Services, Republic of Namibia, and the Comité de Ética do Ministerio da Saude, Ministério da Saúde, República de Angola (Ethics Committee of the Ministry of Health, Ministry of Health, Republic of Angola).