In Quantitative Trait Locus detection, selective genotyping is a way to reduce costs due to genotyping : only individuals with extreme phenotypes are genotyped. We focus here on statistical inference for selective genotyping. We study, in a very large framework, the performances of different tests suitable for selective genotyping. We proof that we have to genotype symetrically, that is to say the same percentage of large and small phenotypes whatever the proportions of the two genotypes in the population. Besides, we proof that the non extreme phenotypes (ie. the phenotypes for which genotypes are missing) don't bring any information for statistical inference. Same results are obtained in the case of a selective genotyping with two phenotypes correlated.