[R-lang] Re: Investigating random slope variance

[Scott Jackson]

Hmm, this is an interesting discussion!  I think something like a weird,
asymmetrical bimodal distribution would be more informative as a test case.
 A nearly-uniform distribution could also be estimated as similar to a
normal with a very large SD.  The technical definition of a BLUP is that
it's the conditional mean/mode (which are the same in a linear mem), which
I understand to be conditional on the data.  The way I understand that is
that it's a bit like reverse-engineering the subject/item effect given the
data, random effect estimate (which is a Gaussian), and the other
parameters (fixed effects, etc.).  But my math isn't good enough for this
definition to clarify my intuitions much, so I think simulating a few cases
is a good idea.  But from everything I've seen, the amount of data plays a
major role, definitely.


On Fri, Apr 4, 2014 at 12:38 PM, Daniel Ezra Johnson <
danielezrajohnson@gmail.com> wrote:

> but maybe if there's a lot less data, it does do it?
>
>
> On Fri, Apr 4, 2014 at 5:28 PM, Daniel Ezra Johnson <
> danielezrajohnson@gmail.com> wrote:
>
>> as far as i can tell, it doesn't shrink them towards a normal
>> distribution. for example, I tried a simulation where the underlying
>> "subjects" were uniformly distributed in their effect - their effects were
>> 1:100 - and the BLUPs came out in almost exactly that same uniform
>> distribution.
>>
>> dan
>>
>>
>>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://mailman.ucsd.edu/pipermail/ling-r-lang-l/attachments/20140404/5d8a6d94/attachment.html