[R-sig-ME] request

[David Winsemius]

On May 28, 2013, at 3:49 PM, Steven McKinney wrote:

> 
> 
> What appears to be a preprint copy is available at
> 
>   http://www.ats.ucla.edu/stat/paperexamples/atkins/Modeling_Infrequent_Counts_JFP2007_-_WEB.pdf
> 

This article contains what I consider to be unfortunate advice:

"One consequence of this is that using a simple transformation allows us to interpret regression coefficients in the Poisson model as the percentage change in the expected counts:

100*[eβ*δ −1] (5) where β is the regression coefficient from the Poisson regression and δ is the units of change in

the predictor (e.g., for one unit of change in the predictor, δ = 1)."

This advice (which is repeating a misinterpretation foisted by the SAS Stats Manual) fails to recognize that the scale of interpretation is not symmetric upon "inversion" or "complementation" of the predictors. So if the predictor is a 1/0 variable, then one coding of a variable with a coefficient of log(2)  might be "interpreted" as a 100% change(say for "married"==1),  whereas the reverse coding ( alternately coded "not married" ==1)  would be "interpreted" as a 50% change. Ironically this advice occurs immediately after the distinction between linear scales and multiplicative scales. It is the multiplicative scale that invalidates that "percentage change" interpretation simply because the range from a "null"-value of 0 to the low end of possible values os "100%" whereas the range to the high end is unlimited upward.

(The correct interpretation is for the exponentiated coefficient to be seen as a relative risk or a multiplicative factor that creates a predicted count or rate relative to the Intercept or baseline estimate. There is really no need to subtract one if one realizes that a factor of 1 will not change any estimate if the result is being multiplied by a baseline value.)

If the advice were couched in more limited manner such that it were resticted to small coefficients and sufficient caveats about its approximate nature were offers, I would be less offended. It bothers me to see this further source of misinterpretation cited as an authority.

> 
> 
> Steven McKinney
> 
> Statistician
> Molecular Oncology and Breast Cancer Program
> British Columbia Cancer Research Centre
> 
> 
> ________________________________________
> From: r-sig-mixed-models-bounces at r-project.org [r-sig-mixed-models-bounces at r-project.org] On Behalf Of Seth Bigelow [seth at swbigelow.net]
> Sent: May 28, 2013 7:19 AM
> To: 'Champika Shyamalie Kariyawasam'; r-sig-mixed-models at r-project.org
> Subject: Re: [R-sig-ME] request
> 
> Shyam, I have found Dave Atkin's tutorial to be very helpful in a situation
> that may be similar to the one you describe. I was at a loss to understand
> the estimated coefficients in poisson regression before reading it. It seems
> to be behind a paywall ($11.95), though it was not the last time I checked.
> --Seth
> 
> A tutorial on count regression and zero-altered count models for
> longitudinal substance use data.
> By Atkins, David C.; Baldwin, Scott A.; Zheng, Cheng; Gallop, Robert J.;
> Neighbors, Clayton
> Psychology of Addictive Behaviors, Vol 27(1), Mar 2013, 166-177.
> 
> -----Original Message-----
> From: r-sig-mixed-models-bounces at r-project.org
> [mailto:r-sig-mixed-models-bounces at r-project.org] On Behalf Of Champika
> Shyamalie Kariyawasam
> Sent: Monday, May 27, 2013 11:50 PM
> To: r-sig-mixed-models at r-project.org
> Subject: [R-sig-ME] request
> 
> Hi all
> 
> I am using generalized linear mixed model fit by the laplace approximation
> (family poisson) to analize my data. I have seed number (dependent variable)
> as a function of study site in two lations in two countries. i ran the model
> in R . But i need some assistance to interpret my data. Any source,
> reference or result of previous work welcome.
> 
> thanks in advance
> 
> shyam


David Winsemius
Alameda, CA, USA