The hypothesis is that if the global haplotype association disappears in the omnibus test when conditioned on SNP “A” but remains significant under the control of other SNPs, then SNP “A” accounts for the observed association. The age, height, weight, and gender were included as covariates in all of the association analyses. Statistical

tests were performed for both LS and FN BMD. The false discovery rate (FDR) method, which is an effective way to address the problems of multiple comparisons, was used in Cell Cycle inhibitor this study to correct for multiple testing. The imputation of genotypes for untyped SNPs from HapMap in the POSTN gene and its flanking regions, approximately 5 kb upstream and downstream, was conducted by a hidden Markov model programmed in MACH v1.0 [22]. We used the phase II HapMap Asian data (CHB and JPT) as the reference panel. In brief, this buy PHA-848125 method combines genotypic data of studied samples with the reference genotype data and then infers genotypes of untyped

SNPs based on probability. The most frequently sampled genotype will be the final imputed one. We used the most likely genotype for the association analysis. The estimated squared correlation (r 2) between imputed and true genotypes was used to assess the imputation quality in MACH. SNPs with r 2 < 0.3 were defined as low imputation quality and were excluded. The most significant untyped SNP was Bortezomib validated by direct genotyping in the HKSC extreme cohort and was replicated in the HKOS prospective cohort. The weighted z-transform test was used in the meta-analysis of SNP with BMD variation in this study. The interactive effect between POSTN and SOST genes was evaluated using our GWAS data with about 500K SNPs in 800 female subjects with extreme BMD that has been described in detail previously [18]. These 800 GWAS extreme subjects belong to the HKSC extreme cohort, which was used as the discovery cohort in this study (n = 1,572). Several

polymorphisms in these two genes showed nominally significant association with BMD in our GWAS (P < 0.05), although they failed to reach the genome-wide significant level (Table Dynein S3, ESM 1). The most significant SNP of POSTN from this candidate gene study and four SNPs (rs9899889, rs865429, rs1234612, and rs2301682) in the SOST and ∼20 kb flanking regions from the GWAS data were used for the interaction analysis. The interactions were assessed by the MDR program [23]. MDR is a nonparametric data mining approach, which pools multi-locus genotypes with high dimensions into one dimension model. It evaluated the predictor using cross-validation method and permutation testing. The combinatorial examination by these two approaches would minimize false positive rates. Cross-validation consistency and testing accuracy were calculated for each combination of tested SNPs. The final best model was the one with maximal cross-validation consistency and minimal prediction error.