Genotype imputation is an important tool in human genetics studies, which uses reference sets with known genotypes and prior knowledge on linkage disequilibrium and recombination rates to infer un-typed alleles for human genetic variations at a low cost. The reference sets used by current imputation approaches are based on HapMap data, and/or based on recently available next-generation sequencing (NGS) data such as data generated by the 1000 Genomes Project. However, with different coverage and call rates for different NGS data sets, how to integrate NGS data sets of different accuracy as well as previously available reference data as references in imputation is not an easy task and has not been systematically investigated. In this study, we performed a comprehensive assessment of three strategies on using NGS data and previously available reference data in genotype imputation for both simulated data and empirical data, in order to obtain guidelines for optimal reference set construction. Briefly, we considered three strategies: strategy 1 uses one NGS data as a reference; strategy 2 imputes samples by using multiple individual data sets of different accuracy as independent references and then combines the imputed samples with samples based on the high accuracy reference selected when overlapping occurs; and strategy 3 combines multiple available data sets as a single reference after imputing each other. We used three software (MACH, IMPUTE2 and BEAGLE) for assessing the performances of these three strategies. Our results show that strategy 2 and strategy 3 have higher imputation accuracy than strategy 1. Particularly, strategy 2 is the best strategy across all the conditions that we have investigated, producing the best accuracy of imputation for rare variant. Our study is helpful in guiding application of imputation methods in next generation association analyses.