Background
Variations in the intensity of malaria transmission in countries requires tailoring of interventions appropriate to the corresponding level of transmission. The World Health Organization Global technical strategy for malaria 2016–2030 [
1] requires National Malaria Control Programmes (NMCPs) to stratify their sub-national malaria burden based on the analysis of past and contemporary malaria data, risk factors and the environment. Cartographies of malaria risk obtained through novel and robust approaches are, therefore, required to assess the impact of control and identify areas where targeted malaria control strategies require adaptation to maximize future impact [
2].
Malaria risk mapping in Kenya is not new. Maps of malaria risk were developed as early as the 1950s based on the length of the presumed malaria season [
3]. In the 1970s, topography, climate, and approximations of spleen rates in children were used to classify Kenya into different endemic zones [
4]. Twenty years later climate and empirical
Plasmodium falciparum survey data were used to provide an updated cartography [
5,
6]. The first attempt to apply the principles of model based geostatistics (MBG) to malaria prevalence survey data from Kenya between 1975 and 2009, at 2095 unique locations was undertaken to provide a risk map for the year 2009 [
7]. This map was used to define Kenya’s unmet needs for vector control [
8], future strategic planning [
9] and funding [
10] from 2010. This proved to be a milestone example of how applications of MBG can influence health policy planning and value for money allocation of resources to areas most in need.
However, harnessing the full value of information on malaria prevalence in time and space to provide an understanding of the fine temporal and spatial resolution changes in malaria risk at national or sub-national scales and provision of probability metrics for important programmatic policy relevant thresholds has not been attempted. Such approaches are often limited by a paucity of input data over time; Kenya however, is a country with a rich history of malaria surveys and provides a unique opportunity to explore patterns of malaria endemicity since 1990. Spatio-temporal methods were applied to understand the changing landscape of malaria transmission in Kenya since 1990 and used the statistical certainty in these models to provide insights into the future investments in control during an era of maximizing value for money.
For the first time in Kenya, a MBG framework was used to provide statistical certainty to identify areas that represent policy relevant thresholds, allowing the government to make informed choices on a more efficient future control strategy.
Discussion
The work presented here is an extension of the 2009 map [
7], incorporating more data, using a different model structure and predicting over 26 years (Fig.
2). The analysis considers a temporal presentation of how malaria transmission has changed over 26 years against the changing landscape of disease management, vector control and climate anomalies, allowing reflection on the impact of these associated covariates of
PfPR
2–10 (Fig.
3). Finally, the precision in the contemporary, 2013–2015, model outputs was considered as a vital component of future decision-making (Fig.
5).
Kenya has made substantial progress in reducing infection prevalence (Figs.
2 and
3), the precise contribution of intervention versus climate are hard to disentangle. In addition, it remains difficult to distinguish whether a decrease or increase in prevalence was directly or indirectly related to an intervention being deployed or removed. Clearly, reductions were observed before the implementation of optimized treatment and vector control in 2006. The timing of this initial decline has been demonstrated at a smaller spatial scale along the Kenyan coast [
40] and at a continental scale [
15]. It remains uncertain as to what contributed to this initial decline in
PfPR
2–10 post 2003, however reductions were accelerated and sustained after 2006, which shows continued reductions in national infection rates (Fig.
3), and continued shrinking of the high-intensity areas (Fig.
2). This occurred during a period when sustained efforts to ensure continued replacement of LLINs as part of mass campaigns and routine delivery to pregnant women and infants were high, and treatment regimens for uncomplicated malaria switched to ACT (Fig.
3). The slight rise in 2014 cannot be entirely explained by the stopping of IRS in 20 counties in 2013, since the rise had already started in 2011. This was also observed on the Kenyan coast [
40] where IRS has not been implemented and nationally returned to levels similar to those during IRS campaigns in 2015.
The heterogeneous nature of
P. falciparum transmission in Kenya continues to be reflected in present-day (2013–2015) descriptions of risk nationwide. A large swathe of the country is occupied by areas predicted to have a
PfPR
2–10 less than 1% with a probability of at least 80%, covering approximately 68% (297,497 km
2) of the populated areas and 61% (27.8 million people) of Kenya’s 2015 population. At a higher probability (≥ 90%) at least half (51%) of Kenya’s populated areas, occupied by 53% of Kenya’s population has a prevalence of less than 1%. In such populations where the infection prevalence over the period 2013–2015 is < 1%, should be an indication for possible migration to a pre-elimination phase by the NMCP [
35]. In these areas the coverage of good quality laboratory and clinical services, reporting and, surveillance should be reinforced. Strengthening of surveillance systems will allow quick detection of infections and prompt treatment with effective anti-malarials to prevent onward transmission within this band of low transmission [
36].
The unexpected
PfPR
2–10 observed in Nairobi (1.1%), might be due to a combination of locally acquired and imported malaria [
41]. A population-based infectious disease surveillance over a 5-year period (2007–2011) in Nairobi (Kibera slums) reported that about two-thirds of patients with malaria had traveled to high malarious areas of Western Kenya [
42]. It seems reasonable to assume that Nairobi continues to be exceptionally low prevalence, and where transmission occurs likely limited to the peripheral areas, for example, at a probability of 90%, 68% of county was likely to have a prevalence < 1% while at a probability of 80% the entire county was likely to have < 1%
PfPR
2–10 2013–2015 (Fig.
5).
There continues to be areas of Kenya, which over the last 26 years appear to be intractable to current coverage levels, and approaches to vector control. Areas that on average continue to support
PfPR
2–10 levels of transmission ≥ 30% are located around Lake Victoria, inland toward the highlands and along the southern coast of the Indian Ocean (Fig.
2). While smaller in their geographic extent (8515 km
2), compared to low transmission, these areas encompass 3.9 million people, 8.5% of Kenya’s 2015 population. The counties affected by this elevated level of
PfPR
2–10 transmission are Kilifi, Kwale, Migori, Homa Bay, Kisumu, Siaya, Kakamega, Vihiga, and Busia (Figs.
2 and
4), however, none of the counties are entirely covered by the 80% exceedance probability that it completely belongs to this endemicity class (Fig.
5). It would, therefore, seem reasonable to expand vector control since the current coverages are still low and below NMCP targets, and introduce other possible innovative approaches to parasite control in these nine counties and might include the use intermittent preventive treatment of infants [
43] and/or the use of RTS, S vaccine [
44].
Spatio-temporal geostatistical models of sparse malaria input data have used multiple, dynamic [
45] or long-term averaged covariates [
46] in the prediction of malaria risk. However, caution is urged in the use of multiple covariates in malaria risk mapping. The inclusion of covariates (climate, land use, social economic status and intervention) to assist predictions at locations without data presume: clearly defined and uniform biological relationship with prevalence; the veracity of the averaged or temporally varying covariate data is often not tested; and including covariates related to intervention coverage precludes any further analysis of the impact of intervention on infection prevalence. The present Kenya analysis avoids the use of covariates because, unlike many other countries, there is a large volume of empirical input data, and the empirical prevalence data are a product of all the possible covariate influences of climate and intervention coverage, allowing a plausibility analysis of the role of climate and intervention, thus avoiding circularity. Caution should be extended beyond Kenya, countries without empirical data on prevalence should not be modelled on the basis of presumed covariate associations with malaria or prediction made in data rich countries to years beyond the last available empirical data.
The novelty of non-exceedance probabilities will allow the NMCP in Kenya, and other malaria endemic countries, to implement control measures that are congruent to malaria risk. This may involve re-orientation of resources allowing optimal utilization of funds in a time of competing health agendas and limited resources. The global momentum is to stratify national malaria control because a blanket cover of intervention is no longer appropriate in increasingly heterogenous settings [
1]. The work presented here highlights the statistical value of NEPs and EPs as a tool for future policy formation.
Authors’ contributions
PMM undertook the development of the models, analysis, and writing of the first drafts of the manuscript. EG provided leadership in the model development and scripting. AMN provided support for data assembly and policy implications. EW and RK provided the policy implications context. EAO supported the evolution of the project. RWS conceived the experiments and provided overall scientific direction. All authors reviewed the final manuscript pre-submission. All authors read and approved the final manuscript.
Acknowledgements
The many scientists, archivists, institutions and national control programmes, who have helped assemble malaria data from across Kenya over the last 21 years. The following individuals were instrumental in providing assistance in identifying, and sharing unpublished survey data or have provided assistance in geo-coding of the assembled survey data: Timothy Abuya, Kubaje Adazu, Willis Akhwale, Pauline Andang’o, Ken Awuondo, Fred Baliraine, Nabie Bayoh, Philip Bejon, Simon Brooker, Maria Pia Chaparro, Jon Cox, Meghna Desai, Mark Divall, Ulrike Fillinger, Lia Smith Florey, Priscilla Gikandi, Andrew Githeko, Carol Gitonga, Joana Greenfield, Helen Guyatt, Katherine Halliday, Mary Hamel, Laura Hammitt, Allen Hightower, Tobias Homan, Susan Imbahale, Rachel Jenkins, Chandy John, Elizabeth Juma, Caroline Kabaria, Lydia Kaduka, Jimmy Kahara, Akira Kaneko, Simon Kariuki, Christine Kerubo, Charles King, Chris King, Rebecca Kiptui, Damaris Kinyoki, Astrid Knoblauch, Yeri Kombe, Feiko ter Kuile, Kayla Laserson, Tjalling Leenstra, Eugiena Lo, Brett Lowe, Betsy Makena, Hortance Manda, Charles Mbogo, Margaret McKinnon, Noboru Minakawa, Sue Montgomery, Eric Muchiri, Richard Mukabana, John Muriuki, Winne Musivo, Charles Mwandawiro, Joseph Mwangangi, Lydiah Mwangi, Tabitha Mwangi, Miriam Mwjame, Charlotte Neumann, Emmily Ngetich, Patricia Njuguna, Abdisalan Mohamed Noor, Oscar Nyangari, George Nyangweso, Christopher Nyundo, Christopher Odero, Edna Ogada, Bernards Ogutu, Bernard Okech, George Okello, Stephen Oloo, Maurice Ombok, Raymond Omollo, Simon Omollo, Monica Omondi, Judy Omumbo, Milka Owuor, Viola Otieno, Beth Rapuoda, Evan Secor, Dennis Shanks, Larry Slutsker, Bob Snow, David Soti, Jennifer Stevenson, Willem Takken, Feiko Ter Kuile, Jacobien Veenemans, Juliana Wambua, Vincent Were, Tom Williams, Shona Wilson, Guiyun Yan, Guofa Zhou and Dejan Zurovac. The authors thank Pamela Thuranira, David Kyalo and Joseph Maina for their help with the assembly of the data, shapefiles, and preparation of graphics. Peter Diggle, Michael Chipeta, Dejan Zurovac and Benn Sartorius for valuable methodological discussions and review of earlier versions of the manuscript.