Re: which model (GLMs)is the best?

Epistat:

With all respect to you (after all, this is your data, not mine) recall
that I analyzed this data to my satisfaction and found that it will
support only a modest model in about three of your cited variables. The
bottom line is that a lengthy model forced on these data will have low
t-ratios for the model coefficients because of the intercorrelations
among the variables. This means, among other things, those model
coefficients will be very poorly estimated. Where a model coefficient
has been estimated to have a positive value it may actually be
negative. It's that bad. Ultimately, the completed model will be
worthless for predicting snail counts. The problem lies in the
intercorrelations among the "independent variables". That cannot be
overcome by using more exotic software.

I still have most of my analysis of these data if you want me to post
it here. It shows, for instance, the correlation matrix. Without that
matrix you cannot see how badly those "independent" variables are
correlated among themselves. Also, several "all combinations" analyses
that show the futility of trying to include more variables than your
data can support.

I'm wondering whether you understand the "numbers" that are being
listed out from your GLM software?

I certainly hope you are not intending to publish any of your cited
models. IMHO they are meaningless. OMU





epistat wrote:
Dear friends,
I have analyze my data with the models of generalized linear models
using S-PLUS, and found three models were relatively good, but i can't
decide which is the best,how should i dicide ?

Model1:
glm(formula = snail ~ grass + gheight + humidity + altitude + soiltem
+ airtem + grass:altitude, family = Gamma(link = inverse), data =
model, na.action = na.exclude, control = list(epsilon = 1e-04, maxit =
50, trace = T))

(Dispersion parameter for Gamma family taken to be 0.2644025)

Null deviance: 63.635 on 161 degrees of freedom

Residual deviance: 42.324 on 151 degrees of freedom

AIC: 1528.1

Model2:

glm(formula = snail ~ grass + gheight + humidity + altitude + soiltem +
airtem + grass:altitude, family = quasi(link = inverse, variance =
"mu^2") , data = model, na.action = na.exclude, control = list(epsilon
= 1e-04, maxit = 50, trace = F))

(Dispersion parameter for quasi family taken to be 0.2644025)
Deviance Residuals:
Null deviance: 63.635 on 161 degrees of freedom
Residual deviance: 42.324 on 151 degrees of freedom
AIC: NA


Model3:

glm(formula = snail ~ grass + gheight + humidity + altitude + soiltem +
airtem + grass:altitude, family = quasi(link = log, variance = "mu^3"),
data = model, na.action = na.exclude, control = list(epsilon = 1e-04,
maxit = 50, trace = F))

(Dispersion parameter for quasi family taken to be 0.005042872)

Deviance Residuals:

Null deviance: 1.4113 on 161 degrees of freedom

Residual deviance: 1.0080 on 151 degrees of freedom

AIC: NA

How should i evaluate my models ? Thanks very much!

.