Raw genotypes vs haplotype blocks for genome wide association studies by random forests

We consider two different representations of the input data for genome-wide association studies using random forests, namely raw genotypes described by a few thousand to a few hundred thousand discrete variables each one describing a single nucleotide polymorphism, and haplotype block contents, represented by the combinations of about 10 to 100 adjacent and correlated genotypes. We adapt random forests to exploit haplotype blocks, and compare this with the use of raw genotypes, in terms of predictive power and localization of causal mutations, by using simulated datasets with one or two interacting effects.