s-news
[Top] [All Lists]

Re: Interpretation of 'interquartile effects' in summary.lrm

To: "Stephen D.Weigand" <weigand.stephen@charter.net>
Subject: Re: Interpretation of 'interquartile effects' in summary.lrm
From: Frank E Harrell Jr <f.harrell@vanderbilt.edu>
Date: Tue, 28 Nov 2006 07:17:46 -0600
Cc: "Inman, Brant A. M.D." <Inman.Brant@mayo.edu>, s-news@lists.biostat.wustl.edu
In-reply-to: <c739eff2035350065b7e1c0130bb56e9@charter.net>
References: <6021CA6EF4C8374084D4F5A141F1CBBB6648DE@msgebe23.mfad.mfroot.org> <456BBFD6.7010203@vanderbilt.edu> <c739eff2035350065b7e1c0130bb56e9@charter.net>
User-agent: Thunderbird 1.5.0.5 (X11/20060728)
Stephen D.Weigand wrote:

On Nov 27, 2006, at 10:49 PM, Frank E Harrell Jr wrote:

Inman, Brant A. M.D. wrote:
S-experts: I am using the Design and Hmisc libraries to fit a logistic regression
model in a manner similar to that given below (note that this model may
not make sense, I just use it to demonstrate my question):


[...]

ddist <- datadist(pbc)
options(datadist='ddist')
fit <- lrm(edema ~ age + alb + ascites + hepmeg, data=pbc)
--------------------------
Since my goal is to extract the odds ratios from this model, I have
tried two different methods: the obvious method of exponentiation of the
regression coefficients and using the summary.Design function.
--------------------------
exp(fit$coef)
summary(fit)
--------------------------
The problem that I have is with the "summary(fit)" output, namely that
its odds ratios are different from those of "exp(fit$coef)".  From the
documentation that is provided by the function's author, I surmise that
the reason the odds ratios for the continuous variables are different is
that the function calculates the odds ratios using "inter-quartile
effects".  I have not previously encountered this method of computing
odds ratios and would appreciate any advice from the experts regarding
the reasons for using the inter-quartile method for calculating the odds
ratios rather than the usual method.  Which method should be reported in
a paper and is there evidence/publications to support this choice?

Please see my book Regression Modeling Strategies which details the reasons for this and for not doing exp(coef). A one-unit change is not meaningful for many of the variables we see in biomedical research.

Frank

Brant Inman
Mayo Clinic


--
Frank E Harrell Jr   Professor and Chair           School of Medicine
                     Department of Biostatistics   Vanderbilt University

Brant,

My additional comments:

For continuous variables that are found to be approximately linear in the logit
(perhaps only due to limited power to detect non-linearity),  I report an
odds ratio based on a difference that is easy to understand or clinically
meaningful: e.g., a 10-year difference in age, or a 0.5 cm difference in
lesion size.

If it's hard to come up with a meaningful difference, if I've transformed a
variable, or if it enters into the model non-linearly, then I follow the above reference and report an odds ratio based on Q3 vs. Q1 since a k-unit change in the variable does not have the same effect across the distribution of values.

I've had some success explaining the use of Q3 vs. Q1 as "comparing a person with a typical above-average value to one with a typical below-average value".

The one drawback to the Q3 vs. Q1 approach is that a careless reviewer may think you've cut the variable into quartiles, fit the model, and reported ORs based on this. So a few sentences justifying Q3 vs. Q1 ORs may be worth including in your
report.

Hope this helps,

Stephen Weigand
Rochester, Minnesota, USA

Nicely said Stephen. I would just add that most relationships are nonlinear and I don't base modeling on the power to detect nonlinearity but on a tradeoff between underfitting and overfitting. And when the model is more than mildly nonlinear, the Q1 Q3 approach isn't as good as showing the continuous partial effect of the predictor holding other predictors constant.

Frank
--
Frank E Harrell Jr   Professor and Chair           School of Medicine
                     Department of Biostatistics   Vanderbilt University

<Prev in Thread] Current Thread [Next in Thread>