Date: Wed, 24 Sep 1997 16:32:11 -0500
Reply-To: "Nichols, David" <nichols@SPSS.COM>
Sender: "SPSSX(r) Discussion" <SPSSX-L@UGA.CC.UGA.EDU>
From: "Nichols, David" <nichols@SPSS.COM>
Subject: Re: NLR, Poisson, and help!!
Fitting a linear model to transformed data and then back-
transforming the resulting coefficients doesn't give you
the same results as fitting the nonlinear model directly.
David Nichols
Senior Support Statistician
SPSS Inc.
nichols@spss.com
>----------
>From: Dale Glaser[SMTP:dale.glaser@SHARP.COM]
>Sent: Wednesday, September 17, 1997 1:19 PM
>To: SPSSX-L@UGA.CC.UGA.EDU
>Subject: NLR, Poisson, and help!!
>
>two part question and any help would be most appreciated:
>
>1)
>
>Analyzing a variable (consisting of counts) called number of
>readmissions:
>
>breakdown is as follows:
>
>value frequency
>0 55
>1 25
>2 10
>3 4
>4 1
>5 1
>
>.......per Neter, et al , Applied Linear Statistical Models, this variable
>seems to follow a poisson distribution, in that the "outcomes are
>counts....with a large number of occurrences being a rare event" (p.
>609)......in the context of comparing an intervention vs control group on
>number of readmissions, and the attendant violation of assumptions (e.g.,
>normality) parametric statistics (e.g., oneway ANOVA) seems
>inappropriate, thus, one alternative is the nonparametric analogue (e.g.,
>Mann-Whitney) or possibly dichotmize the variable (0-no readmit; 1
>-readmit) and conduct logistic regression.....however, my brief perusal of
>nonlinear regression seems to indicate that this option may be the most
>appropriate analytical avenue?........so, for this type of data is NLR
>appropriate? overkill? inappropriate?...any suggestions greatly
>appreciated.....
>
>
>2) deciding to practice NLR, using some data from Neter et al (pg. 536),
>they use days hospitalized to predict prognostic index.........a log
>transform of the criterion variable was conducted and then they used
>OLS to obtain the following solution: y'= 4.0371 + -.03797; in their
>example they use the following start values:
>
>G0 = exp(4.0371) = 56.6646
>G1 = -.03797
>
>they use the following prediction formula:
>
>f (X, G)= (56.6646)exp[-.03797(2)]
>
>......so, I used the following syntax:
>
>* NonLinear Regression.
>MODEL PROGRAM G0=56.665 G1=-.0380 .
>COMPUTE PRED_ = G0*exp(G1*2).
>NLR logprog
> /OUTFILE='C:\WINDOWS\TEMP\SPSSFNLR.TMP'
> /PRED PRED_
> /SAVE PRED RESID DERIVATIVES
> /CRITERIA SSCONVERGENCE 1E-8 PCON 1E-8 .
>
>
>...and got nothing close to their results on page 545!!!.....five model
>iterations resulted in the following: Y' = 4048.95 + -35.73(X)
>........(their
>final result is: Y' = (58.6065)exp(-.03959X)
>
>............so encountering a domain that I am admittedly a complete
>neophyte, what do most of you do when encountering an analyses
>where there is an abundance of 0 values?..........if you use NLR, do you
>use coefficients from the regression equation as your starting values?
>
>thank you!!
>
>Dale Glaser, Ph.D.
>Clinical Outcomes Research/Dept. of Psychology
>Sharp HealthCare/SDSU
>San Diego, CA
>
| 
