[Developers] Need for simpler way to describe the kind of simple mixed models most people are used to.

[H. Skaug]

Dave,

I may not see the full picture, but I do not think it is necessary
in general to use contrasts with random effects. Clearly, the model
easily becomes overparameterized when you add hierarchies of
random effects, but that is taken care
of by the random effects distribution which acts as a penalty.

However, I must admit that I have been wondering about whether
there are situations where you want the random effects
to add up to zero.

Hans


>I think we should develop the machinery needed to construct linear
>predictors from data in a csv file. I wonder what all is involved.
>One thing we need is a "contrast" object.  This is a vector v
>parameterized in such a way that sum(v)=0.  For the separable function
>to be efficient one want to pass only the  necessary v_i's to the
>separable function.  So for example we can parameterize v as
>
>    v_i =w_i -mean(w)
>
>
>the w_i are the independent variables but the derivative stuff can be
>done in terms of the v_i and the derivatives for the w_i calculated from
>the v_i just when they are needed (ie some kind of interface needs to
>accomplish this.)  Also a penalty like
>
>        100*square(sum(w))
>
>is added so that the problem is not degenerate.
>
>One problem is how to parse the csv file. Hopefully should is existing
>software for this we can use.
>
>I think if we can make this all pretty simple we can get a lot more
>interest from R users.
>--
>David A. Fournier