Statalist


[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: st: Interpretation of regressionmodel of ln-transformed variable


From   "roland andersson" <rolandersson@gmail.com>
To   statalist@hsphsun2.harvard.edu
Subject   Re: st: Interpretation of regressionmodel of ln-transformed variable
Date   Thu, 6 Nov 2008 08:00:28 +0100

There are a total of 303 of 37736 patients with LOS=0. I looked for
the different means in the two groups. Obviously the difference is
largest for the arithmetic means and almost absent for the harmonic.
Can you give an advice on what mean to use when reporting on LOS?
						
->	lapscopiintention = 0

	Variable     Type	Obs	Mean	[95% Conf.	Interval]
					
	vtid  Arithmetic	28808	2.985178	2.924702	3.045654
	Geometric	28586	2.322214	2.304817	2.339742
	Harmonic	28586	1.931199	1.918483	1.944085
					

						
->	lapscopiintention = 1

	Variable     Type	Obs	Mean	[95% Conf.	Interval]
					
	vtid  Arithmetic	8928	2.792451	2.745438	2.839463
	Geometric	8847	2.294519	2.26541	2.324002
	Harmonic	8847	1.933647	1.911211	1.956615
					



2008/11/5 Maarten buis <maartenbuis@yahoo.co.uk>:
> --- "Lachenbruch, Peter" <Peter.Lachenbruch@oregonstate.edu> wrote:
>> You can expect differences since your model transforms the response
>> variable, while the glm transforms the mean function. The model you
>> cite below fits log(mu)=XB, while your other model fit
>> E(log(y))=XB.  For non-linear functions these won't be the same.
>
> To expand a bit on that, there are two reasons why the models give
> different answers:
>
> 1) In case of the log transformed y, -regress- with the -eform()-
> option will give you a model for the geometric mean, while -glm- with
> the -link(log)- option will give you a model of the arithmetic mean.
> The two are different but the results should in most cases be pretty
> close.
>
> 2) -regress- with log transformed y will ignore all observations with
> an y equal to 0. The reason is that ln(0) is not defined so will give
> you a missing value. -glm- models the average y, and an average of 0 is
> perfectly legal, so -glm- can handle a LOS of 0 without problem. This
> could lead to larger differences between the two models. If you have
> observations whose value on the dependent variable is 0, than -glm- is
> the preferred method.
>
> Hope this helps,
> Maarten
>
> -----------------------------------------
> Maarten L. Buis
> Department of Social Research Methodology
> Vrije Universiteit Amsterdam
> Boelelaan 1081
> 1081 HV Amsterdam
> The Netherlands
>
> visiting address:
> Buitenveldertselaan 3 (Metropolitan), room N515
>
> +31 20 5986715
>
> http://home.fsw.vu.nl/m.buis/
> -----------------------------------------
>
>
>
> *
> *   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/
>
*
*   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/



© Copyright 1996–2014 StataCorp LP   |   Terms of use   |   Privacy   |   Contact us   |   What's new   |   Site index