Comparison of statistical models for analyzing genotype, inferred haplotype, and molecular haplotype data.
This report compares statistical models based on molecular and inferred haplotypes of the human paraoxonase-1 gene (PON1). In a study of 402 women comprising three race/ethnicities, 137 women had ambiguous inferred haplotypes. The inferred haplotypes (the one with highest posterior probability) for 20 of these women differed from molecular haplotypes, while based on the posterior distribution from the imputation method, 30 discrepancies were expected. We examined the proportion of the variance in PON1 enzymatic activity (phenotype) explained by genotype, and by inferred and molecular haplotype information. For Caucasians, there was an improvement in adjusted R(2) from 16% for the genotype count model, to 29% for imputed haplotypes, and a further improvement to 33% for molecular haplotypes. For Hispanics and African-Americans, there was no indication that haplotypes helped in explaining PON1 activity, and the imputed model gave essentially the same R(2) as the molecular model. For African-Americans, none of the models had adjusted R(2) that exceeded 4%, while for Hispanics they were all about 21-22%. We propose a new parsimonious model which uses all the genotype information and selected haplotype information. For PON1, this model achieves essentially the same adjusted R(2) as the all-haplotype model, with a potential cost savings and without giving the extreme predictions for uncommon haplotype combinations that the all-haplotype models provides.