TASK: problem with "glm" with binomial errors
STATUS: Open
FROM: p.dalgaard@kubism.ku.dk
in glm(,binomial) it's possible that loss of significant
digits make expected values 0 or 1 even though there's no
divergence of the fit. (Happened to me with menarche data,
infants and grown-ups included)
[ Need the example data. Glm needs a complete overhaul. ]
OK, here's a way to fake some data (simulated as if menarche occurs
uniformly between 10 an 15 years of age) There's no real problem with
fitting a logistic regression to this, but the logistic curve gets too
close to 0 and 1 for the very young and the very old.
BTW: rbinom gives NA for p=0 and p=1, which causes some problems with
the simulations.
BTW2: Doesn't "occurred" take a double r?
fake.age<-runif(500,0,50) fake.menarche<-rbinom(500,1,.99999 *(fake.age > 15) + .99999 * (fake.age > 10 & fake.age <= 15)/5 * (fake.age - 10)+.000005) glm(fake.menarche~fake.age,binomial)
Error: fitted probabilities of 0 or 1 occured
fake2.age<-fake.age[fake.age<25&fake.age>5] fake2.menarche<-fake.menarche[fake.age<25&fake.age>5] glm(fake2.menarche~fake2.age,binomial)
Call: glm(fake2.menarche ~ fake2.age, binomial) Coefficients: (Intercept) fake2.age -15.1061 1.2011 Degrees of Freedom: 189 Total; 187 Residual Null Deviance: 253.9 Residual Deviance: 48.73
O__ ---- Peter Dalgaard Blegdamsvej 3 c/ /'_ --- Dept. of Biostatistics 2200 Cph. N (*) \(*) -- University of Copenhagen Denmark Ph: (+45) 35327918 ~~~~~~~~~~ - (p.dalgaard@biostat.ku.dk) FAX: (+45) 35327907 =-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=- r-devel mailing list -- Read http://www.ci.tuwien.ac.at/~hornik/R/R-FAQ.html Send "info", "help", or "[un]subscribe" (in the "body", not the subject !) To: r-devel-request@stat.math.ethz.ch =-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-