Data sources/measurement
All two-stage cluster surveys were considered for the analysis that were (1) performed in sub-Saharan Africa in 2010 or later were, (2) measured indoor residual spray coverage, and (3) publicly available in October 2016. Specifically, Demographic and Health Surveys (DHS), Malaria Indicator Surveys (MIS), AIDS Indicator Surveys (AIS) and Multiple Indicator Cluster Surveys (MICS) were examined for suitability. These surveys are typically powered to determine child mortality and fertility trends at the regional or provincial level. The samples are selected in two stages. First standard enumeration areas typically defined by the country’s Central Office of Statistics and based upon the previous census are selected probability proportionate to size. The size of enumeration areas varies, but would typically not be larger than a few square kilometres. Second, households are randomly selected within each selected enumeration area. For the remainder of this manuscript the enumeration areas will be referred to as community.
Potential bias
Questions exist about the ability of nationally-representative surveys to measure IRS coverage with accuracy. In particular, the wording of the survey has been called into question as respondents may confuse “sprayed the walls in the house with insecticide” with a self-application of commercially-bought sprays such as Doom or Raid (Kilian, pers. comm). This type of misclassification error is present with all analyses of survey data, and is likely to be randomly distributed throughout the survey. To account for the potential misclassification bias of including community in the analysis that was not targeted for IRS intervention, the analysis of EA-level IRS coverage was limited to EAs where at least 5 households reported receiving IRS. Furthermore, the EA-level analysis of community-level coverage of IRS was limited to houses reporting they were sprayed by a government or non-governmental organization.
Quantitative variables
The surveys ask two separate questions about IRS coverage, first whether the house was sprayed in the previous 12 months and second who conducted the spraying (government, NGO, or private). To determine the drivers of household-level coverage of IRS households were categorized as covered by IRS if they reported being sprayed in the previous 12 months without regard to who conducted the sprayed. To determine the drivers of community-level coverage of IRS households were categorized as covered by IRS if they reported being sprayed in the previous 12 months and reported that the spraying was done by either the government or an NGO.
In addition to the spray variables numerous potential factors associated with IRS coverage were considered a priori, namely: wealth quintile, urban or rural, the number of children in the household under 5 years of age, and the number of women in the household of reproductive age (15–49 years old). These factors were aggregated to the community as mean wealth quintile, urban or rural, mean number of children under 5 years of age per household, and mean number of women of reproductive age per household.
Lot quality assurance sampling—community-level IRS coverage
Although the surveys are not powered to estimate IRS coverage within the primary sampling unit, lot quality assurance sampling (LQAS) can give a probability of a community surveyed achieving a threshold coverage such as the 85% threshold recommended by the WHO [13]. LQAS is commonly used to estimate vaccination coverage [15], a similar intervention to IRS in that population-level coverage is perhaps more important for preventing transmission of disease than individual or household-level coverage [16]. For IRS, the malaria vector is killed while resting on the wall of a household after the mosquito has taken a blood meal thereby making household-level coverage somewhat beneficial to neighbours and community-level coverage more important for controlling malaria transmission.
Three separate thresholds were set for IRS success at the community level: at least 50% coverage, at least 75% coverage and at least 85% coverage as recommended by the WHO [13]. Equation 1 presents the probability that a given community with a number of sampled houses (n) would achieve a threshold (p) given the number of sampled houses reporting their house being sprayed the previous month (a). Probabilities of misclassification <10% were deemed acceptable, and EAs were categorized as having no sprayed houses, <50% coverage, 50–75% coverage, 75–85% coverage, and >85% coverage.
$$P(a) = \frac{n!}{a!(n - a)!}p^{a} q^{n - a}$$
(1)
where p = the proportion of houses sprayed in the community in previous 12 months, q = (1−p) or the proportion of houses not sprayed in the community in the previous 12 months, n = the number of houses samples in the EA, a = the number of houses in the sample sprayed in the previous 12 months, and n−a = the number of houses in the sample not sprayed in the previous 12 months.
The ability to predict >85% coverage is partially dependent upon the number of households sampled within an EA. Predicting >85% coverage with 90% confidence requires a minimum of 15 houses sampled within an EA, and only at 23 houses sampled within a community can a single house not report IRS coverage and the community still be classified at >75% coverage (for a spreadsheet on LQAS decisions see Additional file 1). Therefore, the analysis was limited to EAs where information on IRS was known for at least 15 houses.
Statistical methods—household-level analysis
To assess household-level drivers of IRS coverage first an equity analysis at the community-level was conducted wherein household-IRS coverage was estimated as a function of both wealth quintile and urban/rural for each survey identified. The pooled data were then use to regress the probability of a house receiving IRS as a function of household wealth quintile, household ownership of at least one ITN, urban or rural, number of children under the age of five (categorized as 0, 1, or 2+), and the number of women of reproductive age (categorized as 0, 1, or 2+). A logistic regression was used with the community as a random intercept and included dataset as a covariate.
Statistical methods—community-level analysis
To assess community-level IRS coverage across sub-Saharan Africa EAs were first described as having no IRS coverage (no houses reporting IRS performed by the government or an NGO), <50% IRS coverage (at least 5 houses reporting IRS performed by the government or an NGO but unable to exceed the 50% threshold as set by LQAS), 50–75% IRS coverage (a sufficient number of houses reporting IRS performed by the government or an NGO to meet the 50% threshold as set by LQAS), 75–85% coverage (a sufficient number of houses reporting IRS performed by the government or an NGO to meet the 75% threshold as set by LQAS), or >85% coverage (a sufficient number of houses reporting IRS performed by the government or an NGO to meet the 85% threshold as set by LQAS). EAs were then categorized as having <50% coverage or >50% coverage (excluding EAs reporting no IRS) and the probability of a community having >50% coverage was regressed on community-level factors of wealth, urban or rural, number of children under 5 years of age, and number of women of reproductive age, and ITN coverage. A logistic regression was used with robust standard errors adjusted for clustering at the level of the survey dataset and included household-level IRS coverage for the survey as a covariate categorized as <10%, 10–20%, 20–30%, and >30%. The number of households sampled in the community was also included as a covariate to determine the influence that sample size had on a community being classified as >50% coverage.