Quantum Forest

I was conducting some preliminary analysis of regeneration data: counts of seedlings in a series of plots. The distribution was certainly non-normal, with a strong ‘Poisson look’, so I went for a generalised linear model with a Poisson family and a log link. Lately, I’ve been running analysis in parallel using Splus and Genstat. I was quite surprised with the results: the coefficients of the model were different but the predicted values were identical. After meditating about the wonders of statistical software (read #@$) for a while, I did remember about the default Helmert contrasts in Splus. In Helmert contrasts, the j^th linear combination is the difference between the j+1^st and the average of the first j levels. However, most statistical software (including Genstat and SAS) reports the results using Treatment contrasts.

Note to self: use treatment contrasts as default, so results from Splus are comparable with the ones obtained with other software. Change the defaults using the following code:

options(contrasts = c(factor="contr.treatment",
+ ordered="contr.poly"))