Statalist The Stata Listserver

[Date Prev][Date Next][Thread Prev][Thread Next][Date index][Thread index]

RE: st:Transformation for skewed variables with negative values?

From   "Nick Cox" <[email protected]>
To   <[email protected]>
Subject   RE: st:Transformation for skewed variables with negative values?
Date   Tue, 15 Aug 2006 22:16:18 +0100

I agree with Joseph's general stance on this question. 
Also, consider the alternative of a non-identity link 
and a -glm-, which often offers the advantages of
a transformation without the disadvantages. 

However, more generally, I will add a plug for 
my package -transint- from SSC, which is just 
a help file with various comments on transformations. 
You can install it using -ssc-. 

[email protected] 

Joseph Coveney
> Woong Chung wrote:
> I need following help. I have panel dataset for estimating a 
> simple linear
> equation.
> The problem is that my all variables have sknewness and big 
> variation(large
> std).
> In particualr, the dependent variable and one of independant 
> variables have
> a negative sknewness, while all other independant variables 
> are shown by
> positive sknewness. My first intension is using a log 
> transformation of all
> variables  but seems not to be a good idea since all 
> variables have negative
> values (around 20%)
> Besides, all variables except one of independant variables 
> are ratio, thus
> that idea would make worse.
> I would be so glad if anyone has suggestions to solve this problem
> --------------------------------------------------------------
> ------------------
> It's not clear that you actually have a problem.
> It shouldn't be a problem that your independent variables are 
> skewed or have 
> a wide distribution.  There isn't any assumption their 
> distribution, and it 
> is considered better to for them to cover more ground.  
> They're only assumed 
> not to comprise a linear combination within machine 
> precision.  (There are 
> other assumptions about them, in particular, about their 
> relation to the 
> random effects, but that's another matter.)
> Fit the model as-is.  Examine the residuals and empirical 
> Bayes predictions.
> If these do not have a reasonably normal-appearing distribution, then
> transform the dependent variable in accordance with shaping-up their
> distributions, and not the dependent variable's distribution per se.
> Also, from your description, it seems that your dependent 
> variable is a 
> ratio.  Consider sticking its denominator in the model as a 
> predictor and 
> using its numerator as the dependent variable.

*   For searches and help try:

© Copyright 1996–2024 StataCorp LLC   |   Terms of use   |   Privacy   |   Contact us   |   What's new   |   Site index