All models were fitted by maximizing the standard multivariate normal likelihood [66], which depends on a linear model for expected values (incorporating a trait-specific mean as well as age, age2, and sex) and a model for variances and covariances. Likelihoods were maximized using a computationally efficient scoring method [64] implemented in the program POLY (freely available with source code from http://www.sph.umich.edu/csg/chen/poly). The program uses Generalized Estimating Equations estimates (GEE) [67] to select starting values for iterative likelihood maximization, implements various diagnostic techniques, and provides standard errors for all parameter estimates. The current version of the program can handle any non-inbred pedigree. For a subset of the results reported here, estimates were compared with those obtained from SOLAR [68] and QTDT [69], with identical results. We used a likelihood ratio test to compare nested models. To compare non-nested models, we used BIC [70].
