Abstract
Many methods have been proposed to reconstruct the pedigree of a sample of individuals from their multilocus marker genotypes. These methods, like those in other fields of statistical inferences, may suffer from both type I (falsely related) and type II (falsely unrelated) errors. In sibship reconstruction, type I errors come from the spurious fusion of two or more small sibships into a single sibship, and type II errors originate from the spurious splitting of a large sibship into two or more small sibships. In this study I investigate the tendencies of both types of errors made by the likelihood methods in sibship reconstruction, using both analytical and simulation approaches. I propose an improvement on the likelihood methods to reduce sibship splitting, and thus type II errors by downscaling the number of inferred siblings sharing the same genotype at a locus. Simulations are then conducted to compare the accuracy of the original and improved likelihood methods in sibship reconstruction of a large sample of individuals in full-sib families of the same small size, the same large size and highly variable sizes, using a variable number of loci with a variable number of alleles per locus. The methods were also applied to the analysis of a salmon data set. I show that my scaling scheme prevents effectively the splitting of large sibships, and reduces type II errors greatly with little increase in type I errors. As a result, it improves the overall accuracy of sibship assignments, except when sibships are expected to be uniformly small or marker information is unrealistically scarce.