Table 10.

Model selection for gene expression using a backwards step-wise procedure.

	GHR			IGF-1
Model	Log likelihood	Test	P	Log likelihood	Test	P
Gene ~Env + Sex + EBV·Env + EBV·Sex +Env·Sex +EBV·Env·Sex (1)	-82.617			-198.918
~Env + Sex + EBV·Env + EBV·Sex +Env·Sex (2)	-83.111	1 vs. 2	0.320	-199.142	1 vs. 2	0.502
~Env + Sex + EBV·Env + EBV·Sex (3)	-83.460	2 vs. 3	0.403	-199.170	2 vs. 3	0.815
~Env + Sex + EBV·Env (4)	-84.513	3 vs. 4	0.147	-200.233	3 vs. 4	0.145
~Env + Sex (5)	-87.055	4 vs. 5	0.079	-203.634	4 vs. 5	0.033

Progressively simpler models are subsequently defined by removing terms from previous models beginning with interaction effects (e.g. 1 vs. 2). Nested models were tested evaluated using likelihood ratio tests to select the most parsimonious model. The reported p-values correspond to results of the likelihood ratio tests wherein a value greater than the nominal level of significance indicates that removal of the model term does not increase model deviance significantly. The final, parsimonious model that was selected is indicated in bold characters (Env = environment, EBV = EBV group). In the case of GHR, removal of the EBVxEnv term (model 5) resulted in a marginally non-significant likelihood ratio test. However, the interaction term was retained in the final model (model 4 AIC = 185.02), given that its exclusion resulted in a greater penalised likelihood score (model 5 AIC = 186.11).