I have growth data for 5 time points. In my unconditional model the quadratic trend best fits the data (likelihood ratio test). However, the random effect for the quadratic term is not significant but the fixed effect is significant. I also have trouble when adding covariates - the g matrix is no longer positive-definite. Does it make sense to remove the quadratic random effect but leave the fixed effect in the model? Or just estimate a linear trend instead?