home *** CD-ROM | disk | FTP | other *** search
- Comments: Gated by NETNEWS@AUVM.AMERICAN.EDU
- Path: sparky!uunet!zaphod.mps.ohio-state.edu!usc!howland.reston.ans.net!paladin.american.edu!auvm!UI.URBAN.ORG!PETRONIS
- X-Vmsmail-To: SMTP%"stat-l%mcgill1.bitnet@psuvm.psu.edu"
- Message-ID: <921221170949.2060092b@UI.URBAN.ORG>
- Newsgroups: bit.listserv.stat-l
- Date: Mon, 21 Dec 1992 17:09:49 -0500
- Sender: STATISTICAL CONSULTING <STAT-L@MCGILL1.BITNET>
- From: PETRONIS@UI.URBAN.ORG
- Subject: Poisson regr versus Normal transform
- Lines: 37
-
- Subscribers to STAT-L,
-
- A researcher asked me to review a regression analysis. The author is
- studying the relationship between a count dependent variable (number of
- cigarettes smoked daily) and several covariates.
-
- The author chose multiple regression (ols) to model the count outcome. This
- choice is clearly problematic as this outcome is a count, which has a poisson
- distribution, and multiple regression assumes the outcome has a normal
- distribution.
-
- My initial reaction was to suggest poisson regression. Given the
- researcher is unfamiliar with poisson regression, I thought perhaps a
- transformation of the outcome for normality might be more easily understood.
- Not having the data in hand, I cannot assess the degree of non-normality.
-
- Given these two choices, poisson regression versus transformation for
- normality, what is the better choice? By using multiple regression instead
- of poisson regression for a count outcome, what is the error in the results?
- Simply that the beta's are biased? If the outcome variable was transformed
- such that it was normally distributed and the regression rerun, would this
- provide as good a solution to the poisson error problem? Would the results
- be comparable to those from a poisson regression?
-
- Any insights would be appreciated. I am pretty much looking to fill in a
- blank --
- "You employed multiple ols regression, which assumes the outcome has
- a normal distribution, to model a count outcome which has a poisson
- distribution. Given this misspecification of the error term, the results you
- presented are ... FILL IN THE BLANK. This problem can be surmounted by ...
- FILL IN THE BLANK".
-
- Thanks for reading this far. Please respond to me directly. If there is a
- variety of opinion, I will summarize for the list.
-
-
- PETRONIS@UI.URBAN.ORG
-