Summary. Population-based linkage disequilibrium mapping permits finer scale mapping than linkage analysis. However, the population-based association mapping is subject to false positive results due to the population structure and the kinship between the samples. Although there is interest in simultaneously testing the association between a candidate gene and the multiple phenotypes of interest, the currently available association mapping methods are limited to univariate traits only. Here we present a new method for population-based multitrait candidate gene association mapping as a Bayesian semiparametric approach, where the error distribution is flexibly modelled via a multivariate mixture of Polya trees centred on the family of multivariate normal distributions. The method that we develop accounts for the population structure and the complex relatedness between the samples. We compare the new proposal in type I error rate and power with the existing multivariate version of the parametric model of Yu and co-workers and Li's univariate semiparametric model by using the previously published two type Arabidopsis thaliana flowering data sets of association mapping, as well as simulated data.