Changing Malaria Prevalence on the Kenyan Coast since 1974: Climate, Drugs and Vector Control

Abstract

Background: Progress toward reducing the malaria burden in Africa has been measured, or modeled, using datasets with relatively short time-windows. These restricted temporal analyses may miss the wider context of longer-term cycles of malaria risk and hence may lead to incorrect inferences regarding the impact of intervention.

Methods: 1147 age-corrected Plasmodium falciparum parasite prevalence (PfPR2-10) surveys among rural communities along the Kenyan coast were assembled from 1974 to 2014. A Bayesian conditional autoregressive generalized linear mixed model was used to interpolate to 279 small areas for each of the 41 years since 1974. Best-fit polynomial splined curves of changing PfPR2-10 were compared to a sequence of plausible explanatory variables related to rainfall, drug resistance and insecticide-treated bed net (ITN) use.

Results: P. falciparum parasite prevalence initially rose from 1974 to 1987, dipped in 1991-92 but remained high until 1998. From 1998 onwards prevalence began to decline until 2011, then began to rise through to 2014. This major decline occurred before ITNs were widely distributed and variation in rainfall coincided with some, but not all, short-term transmission cycles. Emerging resistance to chloroquine and introduction of sulfadoxine/pyrimethamine provided plausible explanations for the rise and fall of malaria transmission along the Kenyan coast.

Conclusions: Progress towards elimination might not be as predictable as we would like, where natural and extrinsic cycles of transmission confound evaluations of the effect of interventions. Deciding where a country lies on an elimination pathway requires careful empiric observation of the long-term epidemiology of malaria transmission.

Fig 1. The Kenyan coast comprising of three counties (Kilifi, Mombasa and Kwale).

Showing population density per 100 m² (yellow 0 through dark blue 203 people per 100 m²) developed from high spatial resolution 1999 census data [13]; urban centres (Grey) defined by the national census bureau [14] where digitized boundaries undertaken using Google Earth, and used to exclude parasite prevalence data; the location of meteorological stations (Black triangles); major river systems (Blue).

Fig 2. Location of age-corrected parasite prevalence (PfPR_2-10) with the highest recorded estimate of prevalence on top (Left hand panel) and lowest PfPR_2-10 estimate on top (Right hand panel) to distinguish prevalence at similar locations with time.

Data displayed against 279 fifth level census administrative units used to make monthly median malaria predictions (see Methods).

Fig 3. Median small area annual estimates of PfPR_2-10 for the 279 sub-locations fitted using a Generalized Additive Regression model between January 1974 and December 2014 (black line).

The 25% and 75% inter-quartile range and the 2.5% and 97.5% credible intervals are shown as solid and dashed red lines respectively.

Fig 4. Plausibility framework.

Extracted monthly GAM fitted median PfPR_2-10 (red line) shown in relation to annual and long rains (March-June) percentage anomalies in precipitation (dark and light grey bars respectively); cumulative "effective" mid-year ITN distribution data (dark green likely efficacious, light green <50% efficacious, see methods); estimated day 7 anti-malarial drug failures to clear parasitaemia based on information provided in text (blue triangles); and malaria policy milestones for standard treatment guidelines and mass ITN distribution dates (FMD), including reported ITN use among all age groups (S3 Data).