Notice: On April 23, 2014, Statalist moved from an email list to a forum, based at statalist.org.

# Re: st: epiconf: counfounding

 From Svend Juul To "statalist@hsphsun2.harvard.edu" Subject Re: st: epiconf: counfounding Date Wed, 9 Feb 2011 10:34:28 +0100

```SR Millis wrote:

I've been working with your Stata program, epiconf. Using theaccompanying
dataset and fitting the following logistic regression model:

. epiconf dead ab_uria, con(age weight) cat(hichol hypert smoke sex) backward

I obtained the following results.  How do you interpret the model when the
changes in the absolute values of the change in the odds ratios are not
monotonically increasing (in the case of backward selection) or, in the
case of forward selection, monotonically decreasing?  That is, in this model,
if we take 10% as the cut-point of importance, we need only adjust for age,
smoking, and weight?  If so, how do you obtain the adjusted rate ratio and
its confidence interval?

Assessment of Confounding Effects Using Change-in-Estimate Method
Exposure:   "ab_uria"
N =          743

Backward approach
Potential confounders were removed one at a time sequentially

Adj Var              95% CI     Change in OR      p

Crude         2.35   1.24, 4.44          .         .
-i.sex        2.26   1.21, 4.24       -3.6    0.47664
-i.hichol     2.08   1.12, 3.86       -8.0    0.15144
-weight       1.88   1.04, 3.40      -10.0    0.21915
-i.hypert     2.00   1.13, 3.55        6.6    0.42812
-i.smoke      2.24   1.28, 3.92       11.9    0.06456
-age*         3.90   2.34, 6.52       74.4    0.00000
*Crude estimate

===================================================================

First comment: Don't. Data-driven model building leads to wrong conclusions;
see, for example, this FAQ:
http://www.stata.com/support/faqs/stat/stepwise.html
Using a change-in-estimate criterion is no better than using a significance
criterion.

Second comment:
Obviously, the "Crude" estimate in the first line is not the crude estimate,
but the starting estimate, and with backward selection it is the model
including all variables.

If you (for legitimate reasons) want to adjust for age, smoking, and weight:
. logistic dead ab_uria age i.smoke weight
or, if your Stata is prior to version 11:
. xi: logistic dead ab_uria age i.smoke weight

Svend
________________________________________________________

Svend Juul
School of Public Health, Department of Epidemiology
Bartholins Allé 2
DK-8000 Aarhus C,  Denmark
sj@soci.au.dk<mailto:sj@soci.au.dk>
_________________________________________________________

*
*   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/
```