estimating family blups

[Toni K Clarke]

Hi, 

I'm trying to calculate family blups for several different traits in 1440 full-sibs comprising 48 families (33 dams and 19 sires). I've used the following animal model with spatial covariates. 

  res = breedR::remlf90(
    fixed = trait ~1,
   genetic = list(model = "add_animal", pedigree = ped, id = "self"),
       spatial = list(model = 'AR',                      
                  coord = x[, c('plot_x','plot_y')]),
   method="ai",data = x)

If I create a family ID (dam:sire) and fit this as a random effect (random=~family) then the model tends not to converge or there is almost no variance associated with family (i presume all accounted for by the genetic effect). Is it OK to just average the individual blups for each family and perform some bootstrapping to get CI's rather than formally fitting family in the model?

Also, I've been using the multi-trait model to estimate heritabilities and they are predictably higher. I'd like to include spatial covariates in this model but they take too long to run and often don't converge. I saw in one of your previous papers you had run single trait models to estimate the spatial effects and then adjusted your phenotypes before running a multi-trait model. Was this just a case of subtracting the spatial blups from the phenotype of interest before estimating heritability?

Many Thanks

Toni

[Facundo Muñoz]

Dear Toni,

sorry for the delay, I'm on leave this week.

Let me address your questions in different messages.

For the first one, are you including the family effect (using the random component) in addition to the genetic effect? In that case, that is very likely the cause of the non-convergence. Indeed, the genetic and family effects are both trying to explain the same thing. The model is thus not identifiable.

If, instead, you are replacing the genetic effect by the family effect, then there could be a number of causes for the convergence problem. Perhaps many families with only one member or even no member at all? I would need to check the data to be sure.

About averaging the individual blups as a mean to obtain family effects and the bootstrap the estimates... yeah, in principle that would be ok. But resampling is delicate as you want to make sure to keep the number of observations in each family constant over replicates, and you need several observations per family to avoid resampling the same individual too much and biasing the estimation of the within-family variance down. If you don't need individual blups, I would rather estimate family effects directly.

Hope it helps.

ƒacu.-

--
Report issues at https://github.com/famuvie/breedR/issues
---
You received this message because you are subscribed to the Google Groups "breedR" group.
To unsubscribe from this group and stop receiving emails from it, send an email to breedr+un...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/breedr/e8ec591f-1ec5-4a86-b77b-cbfba654cb81n%40googlegroups.com.

[Facundo Muñoz]

Dear Toni,

Concerning your second question about multi-trait models with spatial covariates, what breedR will try to do is to fit a spatial model to each trait, either independently from each other, or with a fully parametrised covariance matrix. These are both relatively complex (many parameters) and failure to converge suggests that the data is not sufficiently informative for learning about all those parameters.

As you say, adjusting the phenotypes by removing the spatial effect prior to fitting the genetic model is not ideal, but is a feasible workaround. Regarding the impact on the heritability, some authors consider that the spatial effect should be included in the denominator, whereas others prefer to exclude it. I guess it depends on your objectives. In any case, it is important to disclose with respect to which sources of variation you are reporting the genetic variability, in order to interpret the results correctly.

Best

ƒacu.-

On 31/10/2022 09:59, 'Toni K Clarke' via breedR wrote:

[Toni K Clarke]

Thanks for replying - this makes sense. I will just fit family as a random effect for estimating the family blups and ignore the genetic effect as this seems simpler than bootstrapping.