R help - Mar 2013 - Poisson and negbin gamm in mgcv - overdispersion and theta

Dear R users,

I am trying to use "gamm" from package "mgcv" to model
results from a mesocosm experiment.  My model is of type

M1 <-  gamm(Resp ~ s(Day, k=8) + s(Day, by=C, k=8) + Flow + offset(LogVol),
data=MyResp,
                        correlation = corAR1(form= ~ Day|Mesocosm),
            family=poisson(link=log))

where the response variable is counts, offset by the log of sample volume.

Unfortunately, the residuals from the model show heteroscedasticity. While
trying to follow up on this, I have run into following problems:

1) How to estimate the overdispersion parameter from a (Poisson) gamm?
I have not been able to extract residual degrees of freedom from M1.

2) How to manually estimate theta for a negative binomial gamm?
I would like to see if applying a negative binomial distribution with log link
(model below) would solve the problem. However, negbin in gamm requires a known
theta...

M2 <-  gamm(Resp ~ s(Day, k=8) + s(Day, by=C, k=8) + Flow + offset(LogVol),
data=MyResp,
                        correlation = corAR1(form= ~ Day|Mesocosm),
            family= negbin(THETA, link="log"))

3) And finally, can I somehow compare the models M1 and M2? Trying anova(M1,M2)
gives the message: "Error in eval(expr, envir, enclos) : object
'fixed' not found" (and I am anyway not sure if this is a valid
approach between Poisson and negbin gamms).

I am most grateful for any help!

Aino

Aino Hosia
Postdoc
Havforskningsinstituttet/Institute of Marine Research
PO Box 1870 Nordnes, N-5817 Bergen, Norway
(Nordnesgaten 50)
Tel: +47 55 23 53 49
E-mail: aino.hosia@imr.no<mailto:aino.hosia@imr.no>
www.imr.no<http://www.imr.no/>






	[[alternative HTML version deleted]]