[Date Prev][Date Next][Thread Prev][Thread Next][Date index][Thread index]

Re: st: confidence intervals for ratio of predictions-- bootstrap vs. parametric methods?

From   "Joseph Coveney" <>
To   "Statalist" <>
Subject   Re: st: confidence intervals for ratio of predictions-- bootstrap vs. parametric methods?
Date   Thu, 11 Oct 2007 09:45:47 -0700

Daniel Waxman wrote (excerpted):

A couple of months ago, Maarten Buis was kind enough to answer my questions:
How to correctly use -predictnl- to calculate confidence intervals for a
ratio of two adjusted predictions  (i.e. relative risk) after logistic


The problem is that these confidence intervals appear unreasonably wide, and
the lower bound can be negative, which is nonsensical.

So I did a bootstrap of the following program, and the bias-corrected
bootstraps give much happier results:
(note that I am bootstrapping the relative risk at a specific value of the
covariates (zero=1, zlog=`1') )


The sample sizes range from 1500 to 6000, with an event rate of ~ 1.5% -
3.5% depending on the population.  Most subjects are at low risk (i.e. the
distribution of the predictors is highly skewed).
Example of the different results: parametric: 3.1 (0.7 ,5.4); bc-bootstrap:
3.7 (2.2,7.4)

My questions:
1.  Can anybody explain why the results are so different, and whether the
bias-corrected bootstraps can reasonably be thought to be much closer to the
2.  If reporting parametric CIs, what to do when the results get negative?
3.  In one of my subpopulations, the bootstrap process returned a few red
'x's instead of dots, meaning, I think that in some samples, one of the
covariate patterns didn't exist or the regression couldn't be performed.  No
bc-CI was calculated.  Any thoughts on the real meaning of this?  (are the
CIs truly infinite?)


What do you get with, say, -glm . . . , family(binomial) link(log)-,
and -adjust- or -lincom , rrr-?

The topic of parameter estimates falling outside of the parameter space (and
by extension, their confidence intervals) when modeling risk ratio and risk
difference versus logit came up on the list in the past month or so.  It
seems like you're asking for trouble when modeling risk ratios in the
neighborhood of 1.5% event rates--the negative parametric lower confidence
limits and red Xs during bootstrapping seem to bear this out.  Is there a
reason why you cannot use the canonical link or, perhaps, the complementary
log-log link?

Joseph Coveney

*   For searches and help try:

© Copyright 1996–2017 StataCorp LLC   |   Terms of use   |   Privacy   |   Contact us   |   What's new   |   Site index