Dear Ben, Le samedi 18 avril 2009 à 23:37 +0000, Ben Bolker a écrit : > Emmanuel Charpentier <charpent <at> bacbuc.dyndns.org> writes: > > > > > I forgot to add that yes, I've done my homework, and that it seems to me > > that answers pointing to zero-inflated Poisson (and negative binomial) > > are irrelevant ; I do not have a mixture of distributions but only part > > of one distribution, or, if you'll have it, a "zero-deflated Poisson". > > > > An answer by Brian Ripley > > (http://finzi.psych.upenn.edu/R/Rhelp02/archive/11029.html) to a similar > > question leaves me a bit dismayed : if it is easy to compute the > > probability function of this zero-deflated RV (off the top of my head, > > Pr(X=x)=e^-lambda.lambda^x/(x!.(1-e^-lambda))), and if I think that I'm > > (still) able to use optim to ML-estimate lambda, using this to > > (correctly) model my problem set and test it amounts to re-writing some > > (large) part of glm. Furthermore, I'd be a bit embarrassed to test it > > cleanly (i. e. justifiably) : out of the top of my head, only the > > likelihood ration test seems readily applicable to my problem. Testing > > contrasts in my covariates ... hum ! > > > > So if someone has written something to that effect, I'd be awfully glad > > to use it. A not-so-cursory look at the existing packages did not ring > > any bells to my (admittedly untrained) ears... > > > > Of course, I could also bootstrap the damn thing and study the > > distribution of my contrasts. I'd still been hard pressed to formally > > test hypotheses on factors... > > > > I would call this a truncated Poisson distribution, related > to hurdle models. You could probably use the hurdle function > in the pscl package to do this, by ignoring the fitting to > the zero part of the model. On the other hand, it might break > if there are no zeros at all (adding some zeros would be a > pretty awful hack/workaround).
Indeed ... > If you defined a dtpoisson() for the distribution of the > truncated Poisson model, you could probably also use bbmle > with the formula interface and the "parameters" argument. This I was not aware of... Thank you for the tip ! By the way, I never delved into stats4, relying (erroneously) on its one-liner description in CRAN, which is (more or less) an implementation of stats with S4 classes. Therefore mle escaped me also... May I suggest a better one-liner ? (This goes also for bbmle...) > The likelihood ratio test seems absolutely appropriate for > this case. Why not? Few data, therefore a bit far from the asymptotic condition... Anyway, a better look at my data after discovering a bag'o bugs in the original files led me to reconsider my model : I wanted to assess, among other things, the effect of a boolean (really a two-class variable). After the *right* graph, I now tend to think that my distribution is indeed zero-deflated Poisson in one group and ... zero-deflated negative binomial in the other (still might be zero-deflated Poisson with very small mean, hard to say graphically...). Which gives me some insights on the mechanisms at work (yippeee !!) but will require some nonparametric beast for assessing central value differences (yes, this still has a meaning...), such as bootstrap. ______________________________________________ R-help@r-project.org mailing list https://stat.ethz.ch/mailman/listinfo/r-help PLEASE do read the posting guide http://www.R-project.org/posting-guide.html and provide commented, minimal, self-contained, reproducible code.