Re: st: lincom vs nlcom

[Maarten buis <maartenbuis@yahoo.co.uk>]

--- On Thu, 3/3/11, Maarten buis wrote:
> One way to resolve this problem (for as far as it is
> possible) is really pin down the hypothesis that you want
> to test. "Testing an interaction effect" can mean too
> many things in this type of models. What is your dependent
> variable? A latent score or a probabilty of choosing a
> category, or the odds of moving one category up? Do you
> want to see how the effect of height changes with age or
> how the effect of age changes with height? Do you define
> the effect in absolute terms (difference) or in relative 
> terms (ratios). 

One of the most useful changes of habits that I went through(*)
is from "finding and reporting the effects" to "explore the
consequences of my model". Especially non-linear models like
logit or ordered probit tend to have many effects, depending
on how you define your effect of interest. The process of 
looking at these, trying to reconcile seemingly conflicting
results, and pinning down exactly what I want to know have 
often been the most fruitfull part of an analysis. That often
makes the difference between an effect as a "semi-magical 
quantity" to a logical argument that moves from a question, 
to observing stuff, to summarizing the observed stuff, and 
finaly an answer to the question (with limitations).

-- Maarten

(*) That sentense makes me sound old and wise. I am neither,
I just could not find a better phrase for it.

--------------------------
Maarten L. Buis
Institut fuer Soziologie
Universitaet Tuebingen
Wilhelmstrasse 36
72074 Tuebingen
Germany

http://www.maartenbuis.nl
--------------------------




      

*
*   For searches and help try:
*   http://www.stata.com/help.cgi?search
*   http://www.stata.com/support/statalist/faq
*   http://www.ats.ucla.edu/stat/stata/