I am also trying to run a random regression analysis with ASREML. The data I have include body weight observations on heifers. Alle heifers (n=7192) were measured once. I included a extensive pedigree (n=23,434). I fitted orthogonal polynomials (TD0,TD1, etc) to the data.
When I compared the likelihoods some questions arised.
Model 1: fitting only random intercept (TD0)
log lik = - 30347.51
Model 2: fitting random intercept and linear regression
(TD0,TD1). In this situation I assumed that the random variables associated with the polynomials were incorrelated.
log lik = - 30344.05
Model 3: same as model2, but now I assumed a correlation between the two variables.(see code below)
log lik = - 30601.27
The lik. gets better from model 1 to model 2, which I expected. However, going to model 3 the likelihood gets worse. This lik. is even worse than in situation1.
What would be a logical explanation for this ?
I was first thinking of overparameterisation as I only had 1
record per animal. However, most animals are genetically related, so that it should have been possible to estimate
I would appreciate comments or suggestions on this problem.
BODY WEIGHT ANALYSIS MODEL 3
SEA 7 !I
LS 21 !I
PS 6 !I
HF 4 !I
PARI 2 !I
UBN 560 !I
BW ~ mu SEA LS PS HF AGE PARI !r UBN 1.0,
animal.TD0 1.80 animal.TD1 0.08
0 0 1
2 0 US 1.80 0.01 0.08