[R] Logistic Regression

[Frank Harrell]

The "10% change" idea was never a good one and has not been backed up by
simulations.  It is quite arbitrary and results in optimistic standard
errors of remaining variables.  In fact a paper presented at the Joint
Statistical Meetings about 3 years ago (I'm sorry I've forgotten the names
of the authors) showed that conflicting results are obtained according to
whether you apply the 10% to the coefficients or to the odds ratios, and
there is no theory to guide the choice.  Why risk residual confounding? 
Form a good model apriori and adjust for all potential confounders; don't
base the choice on P-values.  Use propensity scores if overfitting is an
issue.
Frank


farahnazlakhani wrote:
> 
> I am working on my thesis in which i have couple of independent variables
> that are categorical in nature and the depndent variable is dichotomus. 
> Initially I run univariate analysis and added the variables with
> significant p-values (p<0.25) in my full model. 
> I have three confusions. Firstly, I am looking for confounding variables
> by using formula "(crude beta-cofficient - adjusted beta-cofficient)/
> crude beta-cofficient x 100" as per rule if the percentage of any variable
> is >10% than I have considered that as confounder. I wanted to know that
> from initial model i have deducted one variable with insignificant p-value
> to form adjusted model. Now how will i know if the variable that i
> deducted from initial model was confounder or not? 
> Secondly, I wanted to know if the percentage comes in negative like
> (-17.84%) than will it be considered as confounder or not? I also wanted
> to know that confounders should be removed from model? or should be kept
> in model?
> Lastly, I wanted to know that I am running likelihood ratio test to
> identify if the value is falling in critical region or not. So if the
> value doesnot fall in critical region than what does it show? what should
> I do in this case? In my final reduced model all p-values are significant
> but still the value identified via likelihood ratio test is not falling in
> critical region. So what does that show?
> 


-----
Frank Harrell
Department of Biostatistics, Vanderbilt University
--
View this message in context: http://r.789695.n4.nabble.com/Logistic-Regression-tp3578962p3579392.html
Sent from the R help mailing list archive at Nabble.com.