Until recently, the estimation of the heritability of a trait has required
knowledge of the pedigree within a population. In natural populations such
knowledge is often unknown. Two techniques have been developed which use ma
rker information to estimate heritabilities without reference to the exact
nature of the relationships: a regression-based estimator that regresses ph
enotypic similarity for a pair of individuals against an estimate of their
relationship and a likelihood-based estimator that maximizes the probabilit
y of the genotypic and phenotypic data given a known population structure.
Computer simulation was used to compare the behaviour of these estimators.
Bias in estimates of heritability decreased with increasing marker informat
ion, decreasing simulated heritability, increasing relatedness and increasi
ng sample size. The techniques displayed reasonable tolerance to the percen
tage of missing data. The regression-based technique shows least average bi
as, but largest variance over simulations. Likelihood-based techniques show
larger average bias, but smaller variances over estimates. A modified form
of the likelihood technique, requiring fewer initial assumptions about pop
ulation parameters, is presented. The modified form shows less bias in its
estimates of heritability than the likelihood technique originally proposed
.