Genetics datasets contain measurements for millions of single nucleotide polymorphisms (SNPs). Some (usually small) percentage of these values are of low confidence, and are labeled as missing values. It is common to impute these missing values using statistical relationships within the high-confidence samples as well as relationships mined from public datasets. This makes sense, to get an estimate of what that small percentage of missing data likely is in reality.
Sometimes, analyses are performed that impute up to a much larger number of SNPs - for example, a genotyping platform might measure 2.5 million SNPs, but imputation is performed to get a larger sample of 6 million SNPs. My question is, what value do the extra SNPs that were never even measured have from an analysis point of view? When performing feature selection or predictive analysis, it would seem that the imputed SNPs are really just encoding statistical relationships that are present in the dataset to begin with. What does it mean to find a highly associated or predictive SNP in the imputed set?