We conducted a genome-wide association study of host resistance to severe Plasmodium falciparum malaria in over 17,000 individuals from 11 malariaendemic countries, undertaking a wide ranging analysis which identifies five replicable associations with genome-wide levels of evidence. Our findings include a newly implicated variant on chromosome 6 associated with risk of cerebral malaria, and the discovery of an erythroid-specific transcription start site underlying the association in ATP2B4. Previously reported HLA associations cannot be replicated in this dataset. We estimate substantial heritability of severe malaria (h 2 ~ 23%), of which around 10% is explained by the currently identified associations. Our dataset will provide a major building block for future research on the genetic determinants of disease in these diverse human populations.a single set of estimated haplotypes for 17,960 individuals at the set of over 1.5M SNPs genome-wide that passed our quality control process (Methods). This dataset includes 6,888 individuals from Mali, Burkina Faso, Ghana, Nigeria, Cameroon, Tanzania, Vietnam and Papua New Guinea that have not previously been included in meta-analysis, as well as previously reported data from The Gambia, Malawi and Kenya 9,10 , and thus reflects the haplotype diversity of a substantial portion of the malaria-endemic world.A reference panel enriched for African DNA improves imputation accuracy across the genome The ethnically diverse nature of our study provides challenges for genomic inference, including for our ability to impute genotypes at potentially relevant untyped loci 11 . To address this, we sequenced the genomes of 773 individuals from ten ethnic groups in east and west Africa (specifically from the Gambia, Burkina Faso, Cameroon and Tanzania), including 207 family trios (Figure 1). We combined genotypes at single nucleotide polymorphisms (SNPs) in these data with Phase 3 of the 1000 Genomes Project to form an imputation reference panel which covers most common genetic variation 12 and in which two-fifths of the donor families are of African ancestry (1,203 of 3,046 individuals). In principle, the additional representation of African DNA in this panel should lead to improvements in imputation accuracy for African study populations, and we found that this was indeed the case, with use of our panel leading to a large increase in accuracy relative to panels used in our previous GWAS 9-11 and a more modest improvement relative to using the 1000 Genomes Phase 3 panel alone (Figure S1). Imputation of Vietnamese individuals, and those from Papua New Guinea (PNG) which is substantially diverged from any reference panel population (Figure 1d), were less affected by the inclusion of these additional haplotypes.We specifically examined imputation of malaria-protective alleles in the HBB gene, which have previously been found difficult to impute 9,11 . The SNP encoding the sickle cell mutation (rs334, chr11:5248232) was imputed with r>0.9 in all African populations, as compared with genotype...