
From  Suzy <scott_788@wowway.com> 
To  statalist@hsphsun2.harvard.edu 
Subject  Re: st: RE: transformation of a continuous variable for a logisticregression model 
Date  Wed, 19 Apr 2006 23:17:02 0400 
It's good that you consider this makes biological sense. My main concern was that you were focusing on the statistical results alone. In curve fitting
there is often a tendency to overfit and ignore
substantive or scientific considerations.
I have only two detailed comments to add:
1. Terminology. You call "quadratic" what [R] fracpoly (and presumably Patrick Royston
and coauthors) would call "degree 2" and what the paper cited here appears to call "secondorder". This may sound like a parade of synonyms, but my strong guess
is that it is not. With fractional polynomials the degree is the number of powers, and _not_ the highest power. In your case, your term "quadratic" appears quite wrong therefore, especially for polynomials in which none of the individual powers
is 2. I was reacting to your term and not looking
carefully at the documentation which explains this
terminology.
2. I have not tried to understand what you are doing with boxtid (which is a userwritten command). But in very general terms my understanding is that
although quite different polynomials may give similar overall fits the individual terms in those polynomials may be not at all comparable. The basic underlying issue is very likely that both these kinds
of polynomials are not orthogonal.
Note that attachments should not be sent to Statalist. This is explicit in the FAQ.
Nick n.j.cox@durham.ac.uk
Suzy
Nick  Not to beat a dead horse, but I just thought I'd share this with you  from:<>*
Vincenzo Bagnardi, Antonella Zambon, Piero Quatto and Giovanni Corrao. Flexible MetaRegression Functions for Modeling Aggregate DoseResponse Data, with an Application to Alcohol and Mortality. Am J Epidemiol 2004; 159:10771086.
"Although it is rather simple, the family of secondorder fractional polynomial models offers considerably flexibility. In particular, by choosing p1 and p2 from a predefined set P = {–2, –1, –0.5, 0, 0.5, 1, 2, 3}, a very rich set of possible functions, including some socalled Ushaped and Jshaped relations, may be accommodated. The powers are expressed according to the BoxTidwell transformation (12 <http://aje.oxfordjournals.org/cgi/content/full/159/11/1077#KW
H142C12>), in which denotes if pi != 0 and log x if pi = 0. When p1 = p2 = p, the model becomes log(RR½x) = ß1xp + ß2(xp log x)."
I thought that a second order polynomial = "degree of 2" (M=2) = quadratic as shown in my output from fracpoly below (M=2). I had also emailed the fracplot to show the quadratic curve, but for some reason, it was deleted via transport. In any case, the age variable transformations (age_1 and age_2) from the fracgen command were calculated using the the formulas above  ß1age3 + ß2(age3 log age).
Thus, I still respectfully do not understand why the fracpoly and boxtid results are not consistent with this variable. As far as a theoretical justification of the functional form of age and the response variable  it does make sense for these data.
Nick Cox wrote:
Sorry, but this to me is just a restatement of your previous posting, and addresses none of the points I raised.functional
That aside,
I don't understand how a quadratic function can have powers 3 3. Cubics in my experience are never appropriate for global fits unless there are clear dimensional grounds for using them, which seems unlikely here.
Nick n.j.cox@durham.ac.uk
Suzy
Thanks for your response Nick. In a nutshell, age is not linear in the logit. I'm using the fracpoly command to identify the best
(quadratic withform for age in the full model. The result returned from Fracpoly was a quadratic function with powers 3 3 (which also looks good with fracplot). However, when I further assessed the model using the Boxtid command, the results with the new age transformation  the results were not favorable (the Ho was rejected). When I transformed another continuous variable in the same full logistic model
the Boxtidpowers 1 2 by Fracpoly), the Boxtid results were favorable, all graphs looked very good, and the diagnostics were good (linktest, etc...). I'm trying to understand why my results aren't consistent (Fracpoly and Boxtid) with the age variable, but is with all other continuous variables?
Nick Cox wrote:
I am not clear what you think Statalist members know[95% Conf.
that can help you here. For example, the field in which you are working, what the response variable dmcat means, and what other predictors there may be are all
hidden from view, so the chance of giving opinions drawing on substantive expertise is zero. Otherwise
put, you appear to be assuming that the choices
here can all be made on purely statistical criteria, an attitude which always worries me greatly.
What I have observed, as a kind of anthropologist of
statistical science, is that age plays very different
roles in different fields. Economists often seem to find that a quadratic in age does very nicely, whereas biostatisticians often seem to need more complicated representations, which seems
perfectly plausible given the complexities of
childhood, adolescence, etc.
Either way, fracpoly like other programs has
no inbuilt sensor (or censor) selecting theoretically or scientifically sensible functional forms. So, I suggest that you plot the curve implied against
age and think about it as something that needs justification
or interpretation independently from the data.
Nick n.j.cox@durham.ac.uk
Suzy
I am trying to transform one final continuous independent variable (age) in a logistic regression model. I've tried what I know that's available via Stata. For example, I used the fracpoly command and the best transformation was a second order polynomial with powers 3 3.
Fractional polynomial model comparisons:

age df Deviance Gain P(term) Powers

Not in model 0 2098.129  
Linear 1 1834.224 0.000 0.000 1
m = 1 2 1805.957 28.267 0.000 1
m = 2 4 1791.327 42.897 0.001 3 3
m = 3 6 1790.526 43.699 0.670 2 3 3
m = 4 8 1788.431 45.793 0.351 2 2 3 3

I then used fracgen to generate the new age variables  age_1 and age_2.
fracgen age 3 3
> gen double age_1 = X^3 > gen double age_2 = X^3*ln(X) (where: X = (age+1)/10)
The coefficients for age_1 and age_2 from the full logistic regression model:


Y var  Odds Ratio Std. Err. z P>z
1.06986Interval]
+

age_1  1.087994 .0093302 9.83 0.000
.95709551.106436
age_2  .9644247 .0037538 9.31 0.000
[95% Conf..9718101
However the boxtid command rejected the null for both age_1 and age_2....
age_1  .0100805 .0007172 14.055 Nonlin. dev. 24.646 (P = 0.000)
p1  .0535714 .2122906 0.252


age_2  .0021756 .0004885 4.453 Nonlin. dev. 7.894 (P = 0.005)
p1  3.864227 2.133377 1.811
In all other respects, the preliminary diagnostics look good...
Linktest:


dmcat  Coef. Std. Err. z P>z
.6639337Interval]
+

_hat  .8900851 .1153855 7.71 0.000
.09217931.116236
_hatsq  .0319886 .0307101 1.04 0.298
.2546606.0282022
_cons  .0450195 .1069617 0.42 0.674
.1646215


lroc
Logistic model for dmcat
number of observations = 3354
area under ROC curve = 0.8647
etc...etc...etc...
My question is should I be concerned with the results of
something else Icommand? Is there something I've done incorrectly or
can do/should do?
* For searches and help try:
* http://www.stata.com/support/faqs/res/findit.html
* http://www.stata.com/support/statalist/faq
* http://www.ats.ucla.edu/stat/stata/
* * For searches and help try: * http://www.stata.com/support/faqs/res/findit.html * http://www.stata.com/support/statalist/faq * http://www.ats.ucla.edu/stat/stata/
© Copyright 1996–2016 StataCorp LP  Terms of use  Privacy  Contact us  What's new  Site index 