 |
[Date Prev][Date Next][Thread Prev][Thread Next][Date index][Thread index]
Re: st: Binomial Regression
On 8 Aug 2007, at 14:33, Jay Kaufman wrote:
Journal of Clinical Epidemiology
Volume 60, Issue 9, September 2007, Pages 874-882
Relative risks and confidence intervals were easily computed
indirectly from multivariable logistic regression
Interestingly, they sound the same notes of caution around Poisson
regression with robust variance estimation as around binomial
regression -
(Extracted from the discussion)
Unlike other recent articles on this subject [references omitted], we
are less optimistic about the use of log-linear models, either
Poisson or log binomial, to estimate relative risks directly. Log-
binomial regression, as others and we [31] have found, will not
converge routinely. Failure of the log-binomial model to converge
should not, as has been suggested, point to Poisson regression as an
alternative. Rather, it should signal a fundamental shortcoming of
any log-link model—the failure to fit the data and to bound estimates
of risk by the interval [0,1]. The chances of exceeding these bounds
are especially high when outcomes are common, the situation that
prompted this and other articles. Log-link models by definition
assume a constant relative risk across different patterns of
covariates. But this assumption can fail when the reference risk (of
outcome in the unexposed) is high and relative risk is not small. In
addition, unlike logit models for which recoding the 0/1 outcomes to
1/0 merely inverts the resulting ORs, log-linear models with recoded
outcomes will not generate inverted relative risks.
P Before printing, think about the environment
=================================
Ronan Conroy
[email protected]
Royal College of Surgeons in Ireland
120 St Stephen's Green, Dublin 2, Ireland
+353 (0)1 402 2431
+353 (0)87 799 97 95
*
* 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/