The analysis of loss of heterozygosity (LOH) is perhaps the most widely used technique in cancer genetics. In primary tumors, however, the analysis of LOH is fraught with technical problems that have limited its reproducibility and interpretation. In particular, tumors are mixtures of neoplastic and nonneoplastic cells, and the DNA from the nonneoplastic cells can mask LOH. We here describe a new experimental approach, involving two components, to overcome these problems. First, a form of digital PCR was employed to directly count, one by one, the number of each of the two alleles in tumor samples. Second, Bayesian-type likelihood methods were used to measure the strength of the evidence for the allele distribution being different from normal. This approach imparts a rigorous statistical basis to LOH analyses, and should be able to provide more reliable information than heretofore possible in LOH studies of diverse tumor types.