A variety of modern software packages are available for genotype imputation relying on advanced concepts such as pre-phasing of the target dataset or utilization of admixed reference panels. In this study, we performed a comprehensive evaluation of the accuracy of modern imputation methods on the basis of the publicly available POPRES samples. Good quality genotypes were masked and re-imputed by different imputation frameworks: namely MaCH, IMPUTE2, MaCH-Minimac, SHAPEIT-IMPUTE2 and MaCH-Admix. Results were compared to evaluate the relative merit of pre-phasing and the usage of admixed references. We showed that the pre-phasing framework SHAPEIT-IMPUTE2 can overestimate the certainty of genotype distributions resulting in the lowest percentage of correctly imputed genotypes in our case. MaCH-Minimac performed better than SHAPEIT-IMPUTE2. Pre-phasing always reduced imputation accuracy. IMPUTE2 and MaCH-Admix, both relying on admixed-reference panels, showed comparable results. MaCH showed superior results if well-matched references were available (Nei's GST ≤ 0.010). For small to medium datasets, frameworks using genetically closest reference panel are recommended if the genetic distance between target and reference data set is small. Our results are valid for small to medium data sets. As shown on a larger data set of population based German samples, the disadvantage of pre-phasing decreases for larger sample sizes.