[R-sig-ME] how to compare between models with replacements of a given predictor

[Juan Kamienkowski]

Dear all,

I have the following problem, and I would appreciate very much if someone
on this list could give me some insight or reference to orient my search.

For example, I have this simple model for lmer():

Y ~ A + B + (1|D)

But B is very difficult and expensive to measure, so I want to replace it
with another predictors that I can compute automatically (B1, B2, ...).
Then, I ran these models,

Y ~ A + B1 + (1|D)
Y ~ A + B2 + (1|D)
...

But when I compared them with the original model, all these models were
worse. Thus, I tried these models,

Y ~ A + B + B1 + (1|D)
Y ~ A + B + B2 + (1|D)
...

And in some cases the absolute value of the estimate of B (and the t-value)
decreased more than others... But I'm not convinced this could be really
conclusive that some replacements are better than others.


My intuition is that this should be a common problem in many fields, but I
couldn't find the answer in the bibliography (probably I didn't try the
good keywords). Thus, I will be very grateful if someone could guide me a
little bit.

Many thanks in advance,

Best,

Juan



-- 
Juan E Kamienkowski
@ Departamento de Fisica, FCEN-UBA http://df.uba.ar/
@ Laboratorio de Neurociencia Integrativa http://neuro.org.ar/
@ Laboratorio de Inteligencia Artificial Aplicada http://liaa.dc.uba.ar/

	[[alternative HTML version deleted]]