Statistical imputation applied to genome-wide array data is the most cost-effective approach to complete the catalog of genetic variation in a study population. However, imputed genotypes in underrepresented populations incur greater inaccuracies due to ascertainment bias and a lack of representation among reference individuals,, further contributing to the obstacles to study these populations. Here we examined the consequences due to the lack of representation by genotyping a functionally important, Polynesian-specific variant, rs373863828, in the CREBRF gene, in a large number of self-reported Native Hawai'ians (N=3,693) from the Multiethnic Cohort. We found the derived allele of rs373863828 was significantly associated with several adiposity traits with large effects (e.g. 0.214 s.d., or approximately 1.28 kg/m 2 , per allele, in BMI as the most significant; P = 7.5x10 -5 ). Due to the current absence of Polynesian representation in publicly accessible reference sequences, rs373863828 or any of its proxies could not be tested through imputation using these existing resources. Moreover, the association signals at this Polynesianspecific variant could not be captured by alternative approaches, such as admixture mapping. In contrast, highly accurate imputation can be achieved even if a small number (<200) of Polynesian reference individuals were available. By constructing an internal set of Polynesian reference individuals, we were able to increase sample size for analysis up to 3,936 individuals, and improved the statistical evidence of association (e.g. p = 1.5x10 -7 , 3x10 -6 , and 1.4x10 -4 for BMI, hip circumference, and T2D, respectively). Taken together, our results suggest the alarming possibility that lack of representation in reference panels would inhibit discovery of functionally important, population-specific loci such as CREBRF. Yet, they could be easily detected and prioritized with improved representation of diverse populations in sequencing studies.poor clustering after visual inspection were removed. Problematic samples with a call rate <0.95 or gender/sex mismatches were removed.We applied additional uniform quality controls as follows: All variant names were updated to dbSNP v144; duplicated loci and indels were removed; triallelic variants or variants with nonmatching alleles to 1000 Genomes Project phase 3 (1KGP) 27 were discarded; loci with unique positions not found in 1KGP were removed from the dataset; alleles were standardized to the positive strand by comparing to 1KGP. Finally, a genotype missingness filter of 5% was applied.Additionally, we genotyped rs373863828 in CREBRF in a total of 4,331 MEC Native Hawai'ians using a Taqman Assay; genotypes for 4,214 of these individuals were called successfully, 3,693 of which also have genome-wide array data and thus formed the dataset of this study. This variant was also genotyped in an additional 407 European Americans, 313 African Americans, 432 Japanese Americans, and 386 Latinos (both American and non-American born Hispanics) from the MEC;...