Re: Binomial data
[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: Binomial data

> I am running some analysis with binary data in trees, using 0/1 data
> directly (not proportion or percent) and probit as link function 
> (model: x !bin !probit !wt=0 ~ mu !r fam)
> Problem 1:  I am confuse with the interpretation of variance components. 
> Source        Model  terms     Gamma     Component    Compnt/StndErr  
> fam                   80       80    0.165261      0.165261          4.39   
> Variance       1920   1919     1.00000       1.00000            0.00   
> I understand the variance components are expressed on underlying scale
> (threshold model). In the example above the family variance component is
> 0.165261. But what about the error variance component? Is it just 1-
> 0.165261=0.834739? If this is true, then, ASREML just sets the phenotypic
> variance to 1 without estimation of the other portion of the variance. Is
> that right?  

The model is probit(x) = mu + random(fam) + error.

The probit transformation is based on the standard Normal distribution
which has a variance of 1.  Therefore the error variance
on the underlying scale is 1.  The family variance is .165
So the phenotypic variance is 1.165

heritability  is  C times .165 / 1.165  where C is 4 for halfsib
families but often taken as 2.5 in trees.

> Problem 2: When dealing with binomial data, does Asreml use the methods
> described in Gilmour et al 1985 (The analysis of binomial data by a
> generalized linear mixed model)?   Or other methods are used?
This is not the method of Gilmour et al 1985
Rather, it is the method of Schall Biometrika 78 719-727 (1991)

> Thank you very much all of you. 
>   Have a happy new year.

I also wish you all a very happy time over christmas and new year.
I'll be having a fortnight off.
Thank you all for friendship and encouragement and patient understanding.

>     Uilson Lopes