[R-sig-ME] lme4, glmmPQL and calculating QAICc

[Ben Bolker]

Antonio.Gasparrini at lshtm.ac.uk wrote:
> Dear R users,
>  
> following the recent discussion on calculating QAICc in lme4, I report the weird results I got comparing glmer and glmmPQL.
>  
> I ran these models: 
>  
> pql.model <- glmmPQL(outcome ~ offset(log(pop)) + time + 
>  harmonic(month,3,12), random=list(region=pdSymm(~time)), family=poisson, data)
>  
> glmer.model <- glmer(outcome ~ offset(log(pop)) + time + 
>  harmonic(mm,3,12) + (time|region), family=poisson, data)
> The first model with glmmPQL estimates a sigma (within-group error) anyway, both with poisson or quasipoisson family.
> Its value is 1.40
> The value of sigma^2 is equal to the overdispersion parameter of simpler glm-gam models (~1.96), which makes sense.
>  
> The second model with glmer doesn't estimate a sigma (correctly), but when the family is set to quasipoisson the estimated sigma [lme4:::sigma(glmer.model)] is 15.8, which is simply unbelievable. The standard errors are therefore incredibly huge.
>  
> I couldn't find a reason for that.
> Any comment/suggestion is more than welcome.
> Thanks

  This is interesting.  Can you provide a reproducible example?  Is
random=list(region=pdSymm(~time)) really equivalent to (time|region)?