[R] What computer power for GAMM models

[Gavin Simpson]

1) The model is far too complex. I have trouble imagining a system
where the trend potentially uses 49 degrees of freedom and yet the
unexplained variance within years follows an AR(7). It looks like you
have time points within years over many years; is that why you are
trying to use such a complex trend, to capture the repeated seasonal
patterns? If so, break that down and fit a spline using `bs = "cc"`
(or one of the other cyclic splines in mgcv) and a separate spline for
the trend. If you want to allow the seasonal pattern to vary with
time, then you can do that with multivariate splines, eg, this way you
can compare a model 1

m1 <- gamm(yi ~ ti(DayOfYear, bs = "cc") + ti(Time, bs = "cr"), ....)

with model 2

m2 <- gamm(yi ~ ti(DayOfYear, bs = "cc") + ti(Time, bs = "cr") +
ti(DayOfYear, Time, bs = c("cc","cr")), ....)

Of course, you'll need to fix the the ARAM structure at know
coefficient (currently you are estimating it) or you aren't really
just comparing the "fixed" part of the models.

2) You aren't heeding Simon's warnings about fiddling with lmeControl.
You don't want EM iterations, so you need at least to have
`lmeControl(niterEM = 0)`, alongside other options

A general word of advice with using gamm(), start simply and make sure
you have lots of RAM. These models are very complex so if you start
with simple models you will be less likely to run into convergence
issues and you get results more quickly. Get lots of RAM because
models with `correlation` arguments involve inverting potentially big
covariance matrices and that can quite quickly chew through RAM.

I've had pretty good success fitting models with non-linear trends etc
to environmental time series using gamm(), but the only time things
went a bit awry was when too complex a model was required (where I had
data covering hundreds of thousands of years from an ocean core and
the trend really was very wiggly, and then the correlation structure
and the trend fought for control as there was an identifiability
issue).

HTH

Gavin

On 25 October 2013 07:19, Sylvie Martin <smartin_sepia at orange.fr> wrote:
>
>
> Hello,
>
>  I use GAMM function (MGCV package) in R software to study relationship between pollen and pollinosis. My models include autoregressive terms. Here is an exemple:
>
> CupAR7_2<-gamm(nb_rca_2~s(Trend,bs="ps",k=49) +Semaine
>             +s(TX03,bs="ps",k=8) +s(UN03,bs="ps",k=4)  ......+pollen01
> ,correlation=corARMA(p=7,q=0,form=~Trend|Annees),
> niter=40,control=lmeControl(maxIter=100000,msMaxIter=100000)
>         ,family=quasipoisson, data=donnees, na=na.omit)
>
> Each model need 1/2 hour to 1 hour to get a result and I frequently get the following error messages:
>
> Erreur dans lme.formula(fixed = fixed, random = random, data = data, correlation = correlation,  :
>   nlminb problem, convergence error code = 1
>   message = singular convergence (7)
>
> Erreur dans na.fail.default(list(Xr.4 = c(0.00374377951791214, 0.00438373736920182,  :
>   missing values in object
>
>
>  Does my computer lack power,(if so,what is required?) or is R  limited in speed of execution or is my syntax not ok. Here are the features of my computer:
>
> processor: 3.10 GHz  i5-2400 CPU
> RAM: 4 Go (3.87 Go)
> OS: 64 bits
> Windows 7 professionnel (Pack 1)
>
> Another question: how can we add a smoothing parameter in the model GAMM to help convergence ("sp =" is not accepted)?
>
> Thanks for your help
> Best regards
>
> Sylvie Martin
> SEPIA-Santé
>
> Bureau d'études en épidémiologie,
> biostatistiques, environnement
>
> 31 rue de Pontivy
>
> 56 150 BAUD
>
> FRANCE
> tél : 02 97 28 80 38
> fax: 02 97 28 81 10
>         [[alternative HTML version deleted]]
>
>
> ______________________________________________
> R-help at r-project.org mailing list
> https://stat.ethz.ch/mailman/listinfo/r-help
> PLEASE do read the posting guide http://www.R-project.org/posting-guide.html
> and provide commented, minimal, self-contained, reproducible code.
>



-- 
Gavin Simpson, PhD