Dear R users,
I have developed a model
I have compared several options of obtaining p-values for
poisson lmer model including Marlov chain monty carlo methods, single
term deletions and summary.>
> However, I encountered several problems that can be classified as
> (1) the p values from the summary command are total different from
those derived from Marlov chain monty carlo methods
> How can I proceed, left with these uncertainties in the estimations of
> the p-values?
>
> Below is the corresponding R code with some output so that you can see:
>
> ##
> fit<-lmer(End~Treatment+offset(log(Area)+(1|Site/Treatment),
family=poisson)
>
Summary
Intercept
> The p-values from mcmc are:
>
> ##
> markov1=mcmcsamp(m2,5000)
>
> HPDinterval(markov1)
> lower upper
> (Intercept) -1.394287660 0.6023229
> logpatch 0.031154910 0.1906861
> loghab 0.002961281 0.2165285
> landscape_diversity 0.245623183 1.6442544
> log(site.(In)) -41.156007604 -1.6993996
> attr(,"Probability")
> [1] 0.95
>
> ##
>
> mcmcpvalue(as.matrix(markov1[,1])) #i.e. the p value for the intercept
> [1] 0.3668
> > mcmcpvalue(as.matrix(markov1[,2])) #i.e. the p-value for logpatch
> [1] 0.004
> > mcmcpvalue(as.matrix(markov1[,3])) #i.e. the p-value for loghab
> [1] 0.0598
> > mcmcpvalue(as.matrix(markov1[,4])) #i.e. the p-value for landscape.div
> [1] 0.0074
>
> If one runs the mcmcsamp function for, say, 50,000 runs, the p-values
> are slightly different (not shown here).
>
> ##here are the p-values summarized in tabular form:
<snip>
[MCMC]> logpatch 0.004
> loghab 0.0598
> landscape_diversity 0.0074
<snip>
[single-term deletions]> logpatch 0.007106
> loghab 0.1704
> landscape_diversity 0.01276
<snip>> To summarize, at least for quasipoisson models, the p-values obtained
> from mcmcpvalue() are quite different from those obtained using
> single-term deletions followed by a chisquare test.
>
> Especially in the case of "loghab", the difference is so huge
that one
> could tend to interpret one of the p-values as "marginally
significant".
