Direct analyses of haplotype effects can be used to identify those specific combinations of alleles that are associated with a specific phenotype. We introduce a method for direct haplotype analysis that solves two problems that arise when haplotypes are analysed in populations of unrelated subjects. Instead of assigning a single, most likely, haplotype pair to multiple heterozygous subjects, all haplotype pairs compatible with their genotype were determined and the posterior probabilities of these pairs were calculated using Bayes' theorem and estimated haplotype frequencies. For the individual patients, all possible haplotype pairs were included in the statistical analysis using the posterior probabilities as weights, which were re-estimated in an iterative process together with the haplotype effects. The second problem of unstable haplotype effect estimates, due to the numerous haplotypes and the low frequency at which some occur, was solved by assuming that haplotypes sharing the same alleles show a similar effect and that the extent of this similarity relates to the number of alleles shared. These assumptions were incorporated in a weighted log-likelihood model by introducing a penalty, where differences in effects of similar haplotypes were penalised. Using CETP gene haplotypes, consisting of five closely linked polymorphisms, and baseline CETP and HDL-C concentrations from the REGRESS population, we demonstrated that the model resulted in more stable effects than estimates based on unambiguous patients only.