Estimating effects of latent and measured genotypes in multilevel models.

Multilevel modelling is a data analysis technique for analysing linear models in samples with a hierarchical or clustered structure. Clustered data are often present in genetic research where family members may either be required or serve a methodological purpose to study hereditary factors. These samples imply a natural hierarchy because genetically related individuals are grouped within families. We first demonstrate the use of multilevel modelling to study latent genetic and environmental components of variance in extended families where subjects may be related as twins, full siblings, half siblings, or cousins. Next, measured genotypes are included to estimate locus effects. Because the model accounts for the clustering of observations by estimating a random intercept at the family level, it tests for genotype effects on the phenotype within families so that possible population stratification effects cannot cause false positive results. Several extensions are discussed such as testing for genotype-environment interactions, analysing different types of response scales, or tailoring the model to other sample structures. To illustrate the approach we used birth weight data of 5562 children from 3643 fathers from 3186 mothers in 2873 extended families to which simulated genotypes of a hypothetical locus were added.