The primary malaria-transmitting vectors in Malawi include Anopheles arabiensis, An. funestus, and An. gambiae s.s.10 Vectorial capacity is high because of ample rainfall (725 - > 2,000 mm/),11 high year-round temperatures, and high humidity, especially in low-lying areas along the lakeshore, Shire River Valley, and central plains.12 Transmission is perennial, with seasonal increases in disease incidence after rains during November–April. Administratively, Malawi is divided into 28 districts among three regions (Northern: 6 districts, Central: 9 districts, and Southern: 13 districts).

Malawi has a strong history of malaria control. Malawi began piloting a subsidized ITN program in 1998,13 and by 2003, subsidized ITNs were available through a nationwide social marketing campaign, the first of its kind in sub-Saharan Africa.14 As in several neighboring countries, the bulk of scale-up occurred after 2005 coincidental with increases in development partner assistance. The 2005–2010 Malaria Strategic Plan set goals of 85% coverage among high risk groups (children less than five years of age and pregnant women) through rapid scale-up of ITNs, indoor residual spraying (IRS), and prompt access to artemisinin-combination therapies.14 From 2007 to early 2010, approximately four million LLINs were distributed through antenatal care clinics and the Expanded Program on Immunization. Household ownership of at least one net (treated or untreated) was estimated as 13% in 2000,15 and ownership of at least one ITN was estimated as 27% in 2004,3 38% in 2006,16 and 57% in 2010.4 Since 2007, more than 21 million courses of artemether-lumefantrine have been distributed, and treatment has been available free to fever patients nationwide. Indoor residual spraying has been piloted since 2007 in Nkhotakota District, and in late 2010 was expanded to six other high-prevalence districts.12

Cross-sectional community Plasmodium falciparum parasite rate (PfPR) data for 2000–2011 for Malawi were assembled by year from a combination of published and unpublished sources. These sources included data from national micro-nutrient surveys conducted in 2001 and 2009 by the Ministry of Health (MoH) in collaboration with the Centers for Disease Control and Prevention, Irish Aid, the United Nations Children's Fund,17,18 the National Malaria Indicator Survey conducted in 2010,6 anemia and parasitemia surveys conducted by the College of Medicine Malaria Alert Center in eight districts annually during 2005–2009,19,20 MoH reports, peer-reviewed journals, conference abstracts, and unpublished data obtained through direct contacts with researchers and program staff.

Most surveys used two-stage cluster sampling, where clusters represented by villages or census standard enumeration areas were randomly selected at the first-stage, and households randomly selected at the second stage within clusters. For all surveys, households were classified as belonging to the standard enumeration area within which they were located. For each survey cluster, household data were summarized on the number of persons examined, number positive for P. falciparum malaria, age range, month and year of survey, and method for malaria testing, and linked spatially to the geographic location of the cluster centroid. All malaria testing was conducted by using either rapid diagnostic tests or microscopy; where rapid diagnostic tests and quality-assured microscopy results were available for the same persons, results of microscopy were chosen. Survey cluster locations were geo-coded by using combinations of global positioning systems, electronic gazetteers (Google Earth, Encarta, and Alexandria), and other sources of longitude and latitude. The assembled PfPR data were standardized to the classical age-range of 2 to > 10 years (PfPR_2–10) by using an algorithm based on modified catalytic conversion models.21

Data at 1 × 1 km spatial resolution on urbanization,22 temperature suitability index,23 elevation,24 annual mean precipitation,25,26 and enhanced vegetation index27 were assembled for the period of the study. The values of these underlying ecologic and climatic covariates were extracted to each survey location by using the ArcGIS 10 Spatial Analyst Tool (ESRI, Redlands, CA).28 These covariates were then included in a total-sets analysis, which is an automatic model selection process based on a generalized linear regression model and implemented by using the bestglm package in R.29,30 This approach selects the best combination of covariates based on the value of the Bayesian Information Criteria statistic,31 where the lowest Bayesian Information Criteria indicates the best model fit.

The continuous surfaces of the age-standardized data (PfPR_2–10) were generated by using a space-time MBG framework,32 whereby Bayesian inference was implemented using the Markov Chain Monte Carlo algorithm within the open-source statistics package PyMC.33 Details of model code32 and statistical procedures34 are provided in Supplemental Information 1. In brief, the value of PfPR_2–10 was modeled as a transformation of a spatiotemporally structured field superimposed with unstructured (random) variation on a regular 1 × 1 km grid during 2000–2011. The number of P. falciparum-positive responses from the total sample at each survey location was modeled as a conditionally independent binomial variate given the unobserved underlying age-standardized PfPR_2–10 value21 and a linear function of the climatic and environmental predictors. The unstructured component was represented by a Gaussian distribution with zero mean. The spatiotemporal component was represented by a stationary Gaussian process35 with covariance defined by a spatially anisotropic version of the space-time covariance function proposed by Stein.36 To partly model seasonality, the covariance function was modified to enable the time-marginal model to include a periodic component of wavelength 12 months in the temporal covariance structure. Each survey was referenced temporally by using the mid-point (in decimal years) between the recorded start and end months. For each grid location, samples of the annual mean of the full posterior distribution of PfPR_2–10 for each year were generated. These PfPR_2–10 samples were then used to generate continuous maps of the annual mean for 2000, 2005, and 2010. The continuous maps were also binned into the following PfPR_2–10 classes: 10% to < 20%; 20% to < 30%, 30% to < 40%, and 40–50%.

As a first step for assessing uncertainty around predictions of PfPR_2–10 by using the Bayesian geostatistical model, the continuous mean maps were accompanied by estimates of the posterior standard deviation. In addition, a spatially representative validation set of PfPR_2–10 survey data was also selected by using a spatially de-clustered sampling algorithm.32 The annual predictions were then repeated in full using the remaining data to predict mean PfPR_2–10 at the validation locations. The ability of the model to predict point-values of PfPR at unsampled locations was quantified by using two simple summary statistics: the mean prediction error (MPE) and the mean absolute prediction error (MAPE). The MPE provides a measure of the model bias, and the MAPE is a measure of the average accuracy of individual predictions.

Populations at risk were estimated from the 2010 map by PfPR_2–10 class (defined here as lower meso-endemic [10–40%] and higher meso-endemic [40–50%]) and district for the total population and for children less than five years of age. Totals were calculated from Afripop rasters22 for total populations and populations less than five years of age by using ArcGIS 10 Spatial Analyst Tools. Population-adjusted prevalences were also calculated by district by multiplying the continuous PfPR_2–10 rasters by the population rasters and computing estimated infected populations out of total populations.

The community survey data assemblage included 1,057 P. falciparum survey clusters, from which 33,041 persons were examined and 9,239 were positive during 2000–2011 (Table 1 and Figure 1). Most (88%, n = 933) survey data were for 2005–2011 (Table 1). Most blood examinations for P. falciparum infection (76%, n = 805) were performed by microscopy. In most (88%, n = 930) clusters, sample sizes for malaria testing were less than 50 persons; none were less than five persons. Almost all (98%, n = 1,032) survey clusters included testing among children < 14 years of age, of which 82% (n = 844) included testing only among children less than five years of age.

Results of the total-set analysis showed that the model with urbanization and the temperature suitability index was the best fit in predicting PfPR_2–10. These variables were subsequently included in the malaria prediction model (Supplemental Information 2). The spatial distribution of malaria risk remained largely consistent throughout the recent scale-up period, with the highest predicted prevalence (40–50%) along the shore of Lake Malawi, along the Shire River Valley, and portions of the central plains (Figures 2 and 3). Compared with 2000 and 2005, there was some evidence of increased prevalence along the Zambian border in Mchinji and Kasungu Districts and the western portions of Rumphi and Mzimba Districts in 2010. Similarly, predicted prevalence increased slightly around Mulange District and nearby lowlands in 2010. Across the entire period, prevalence was lowest in urban areas (notably urban areas within Lilongwe, Blantyre, and Mzuzu Districts, where prevalence was 10–20%) and along the northern and central highlands. There was little evidence of a decrease in prevalence during 2005–2010.

The MPE and MAPE associated with the full space-time geostatistical model were 0.04% and 2.7%, respectively, indicating low bias and a slight over-prediction of risk. The standard deviations of the annual mean PfPR_2–10 predictions were similar across the predictions years and ranged from 20% to approximately 30% (Figure 4). Uncertainty appeared to be lowest for predictions to urban areas.

The entire population of Malawi is under at least moderate transmission risk (historically meso-endemic or 10–50% PfPR_2–10). Of a total national population of 13.6 million and a population of children less than five years of age of 2.7 million in 2010, 6.5 million (48%) persons and 1.3 million (49%) children less than five years of age were estimated as residing in the higher transmission intensity areas (40–50% PfPR_2–10) (Table 2), yet no areas had a predicted prevalence > 50%. The national population-adjusted prevalence (PAPfPR_2–10) was 37.4%. Districts with the highest proportion of the population under higher transmission intensity included Karonga (82%) in the Northern region, Nkhotakota (84%) and Mchinji (85%) in the Central region, and Machinga (95%), Chikwawa (96%), Mulanje (94%), and Phalombe (93%) in the Southern region (Figure 5); in close to half (46%) of the districts, greater than 75% of the population resided in areas with > 40% predicted prevalence. Similarly, the highest PAPfPR_2–10 was found in Chikwawa (44%), Machinga (42%), Nsanje (42%), and Karonga (42%). Districts with the lowest proportions of the population under higher transmission intensity included Mzimba (4%), Lilongwe (4%), and Chitipa (12%). The lowest PAPfPR_2–10 was found in Blantyre (26%), Lilongwe (32%), and Mzimba (35%). Comparisons of district PAPfPR_2–10 across the three prediction years (2000, 2005, and 2010) showed only slight (≤ 1%) differences by year. Therefore, only data for 2010 are presented (PAPfPR_2–10 for 2000 and 2005 are shown in Supplemental Information 3).