Imputation is a commonly used technique that exploits linkage disequilibrium to infer missing genotypes in genetic datasets, using a well characterized reference population. While there is agreement that the reference population has to match the ethnicity of the query dataset, it is common practice to use the same reference to impute genotypes for a wide variety of phenotypes. We hypothesized that using a reference composed of samples with a different phenotype than the query dataset would introduce imputation bias.To test this hypothesis we used GWAS datasets from amyotrophic lateral sclerosis, Parkinson disease, and Crohn disease. First, we masked and then performed imputation of 100 disease-associated markers and 100 non-associated markers from each study. Two references for imputation were used in parallel: one consisting of healthy controls and another consisting of patients with the same disease. We assessed the discordance (imprecision) and bias (inaccuracy) of imputation by comparing predicted genotypes to those assayed by SNP-chip. We also assessed the bias on the observed effect size when the predicted genotypes were used in a GWAS study.When healthy controls were used as reference for imputation, a significant bias was observed, particularly in the disease-associated markers. Using cases as reference significantly attenuated this bias. For nearly all markers, the direction of the bias favored the non-risk allele. In GWAS studies of the three diseases (with healthy reference controls from the 1000 genomes as reference), the mean OR for disease-associated markers obtained by imputation was lower than that obtained using original assayed genotypes.We found that the bias is inherent to imputation as using different methods did not alter the results. In conclusion, imputation is a powerful method to predict genotypes and estimate genetic risk for GWAS. However, a careful choice of reference population is needed to minimize biases inherent to this approach