Hi,
I have around nine traits, which I would like to use in the animal model (estimate heritabilities and genetic, phenotypic correlations). Haven't worked with these models before, I found out that increasing number of traits in the model results in convergence problems.
Is there a way to improve the convergence (I use vce), or what would be the best design for my study, should I split the traits into several models? Perhaps if someone could give me some information on the main pitfalls when working with animal models that would be great.
Thank you.
Dear Julija,
Unfortunately, there is no unique solution for your problem. Indeed, a multi-trait analysis can be problematic as the number of traits analyzed increase.
To begin with, you would have to have lots of data to make sure you can estimate all variance components in a 9x9 variance-covariance matrix, which are 9 + (9*(9-1)/2) = 45 parameters.
Another problem is about using "wrong" starting values for these parameters. It is good to use reasonable starting values (e.g. based on single and two-trait analyses, or based on other datasets, etc..) when estimating genetic parameters.
Another issue could be the use of one trait as a explanatory variable of another trait. For example, when calculating genetic correlations between average daily gain (ADG) and Residual Feed Intake (RFI), we usually have problems because ADG is the trait in one model (ADG) and one of the explanatory for another model (RFI).
So, one suggestion is for you to start slow and estimate genetic parameters for each trait. Then you start doing two traits at a time, and so on... get a feeling for the data first!
Nick
Dear Julija,
Unfortunately, there is no unique solution for your problem. Indeed, a multi-trait analysis can be problematic as the number of traits analyzed increase.
To begin with, you would have to have lots of data to make sure you can estimate all variance components in a 9x9 variance-covariance matrix, which are 9 + (9*(9-1)/2) = 45 parameters.
Another problem is about using "wrong" starting values for these parameters. It is good to use reasonable starting values (e.g. based on single and two-trait analyses, or based on other datasets, etc..) when estimating genetic parameters.
Another issue could be the use of one trait as a explanatory variable of another trait. For example, when calculating genetic correlations between average daily gain (ADG) and Residual Feed Intake (RFI), we usually have problems because ADG is the trait in one model (ADG) and one of the explanatory for another model (RFI).
So, one suggestion is for you to start slow and estimate genetic parameters for each trait. Then you start doing two traits at a time, and so on... get a feeling for the data first!
Nick
Julija,
I would like to echo Nicks recommendation: Start with estimating single trait parameters, use these to estimated bi-variate covariances between traits and (try to) confirm them in tri-variate analyses.
In my experience you will have probably exhausted the effectiveness of VCE. This is not a specific problem to VCE, by the way. Other REML based software (ASREML comes to mind) more often the not have issues when you go beyond tri-variate analyses. They either take forever to iterate, or won't converge, or both.
If you do want to analyse all 9 traits simultaneously, you should look into gibbs sampling methods. We've done analyses using gibbs samplers estimating parameter-matrices for up to 25 traits at a time. And they seem to give satisfactory answers (most of the time).
- Badges
- Science topic
- Similar topics
- Bioinformatics
- Bioinformatics and Computational Biology