Re: [R] GAM model output error(?)

2014-07-16 Thread Simon Wood 
Trevor, Can I just check - were you using the new 'nb' from mgcv version 
>= 1.8 or the old (and very slow) 'negbin' family? (Not that negative 
binomial seems needed here, but just to know) best, Simon



On 16/07/14 21:16, Trevor Davies wrote:

Hi Simon,

Thank you so much for being active on this list, it really is
tremendously helpful.

Thanks you for your insights, I was wondering whether both terms were
necessary.

As for 'NB wouldn't work' it was a convergence problem (and tremendously
slow).  There were also issues with the models not giving lack of
convergence warnings but then the model saying it was not converged and
vice versa.  The QP seemed to fit the data well (and quickly) so we
switched to that.

Thanks again.
Trevor


On Wed, Jul 16, 2014 at 12:44 PM, Simon Wood mailto:s.w...@bath.ac.uk>> wrote:

Trevor,

It looks like you've added a parametric COR.YEARLY.MEAN in addition
to your s(cxe,cyn,by=COR.YEARLY.MEAN) term. Because the latter
includes a linear effect of COR.YEARLY.MEAN, then the parametric
COR.YEARLY.MEAN will not be identifiable, so gam has dropped it.

I guess from the scale parameter your data are, if anything,
under-dispersed relative to Poisson, but perhaps Poisson would have
been ok too.

Could you give a bit more detail on 'NB just wouldn't work' please?
What actually happened? (It's new functionality, so having reports
when stuff goes wrong is useful).

best,
Simon


On 16/07/14 20:25, Trevor Davies wrote:

I have run a quasipoisson spatial model via GAM (NB just
wouldn't work) and
I am getting the following output of one of my parameters
(COR.YEARLY.MEAN).  Does this suggest an error in the model fit?
The model
seems to have converged.  Apologies for the lack of reproducible
example
but it didn't really seem warranted in this case.  Thank you for the
assistance.
Trevor

Parametric coefficients:
  Estimate Std. Error t value Pr(>|t|)
(Intercept)  -3.7084 0.2185  -16.97   <2e-16 ***
COR.YEARLY.MEAN   0. 0.  NA   NA
---

Approximate significance of smooth terms:
 edf Ref.df   F
  p-value
s(PRED_cBOTTOM_TEMPERATURE.__std)  5.008  6.045  24.871
  < 2e-16 ***
s(cxe,cyn)  28.999 29.000  13.951  <
2e-16 ***
s(cDMAX.std) 3.830  4.761 243.895  <
2e-16 ***
s(cxe,cyn):male.allyear.__intensity.std8.053  9.923   4.486
  2.8e-06 ***
s(cxe,cyn):female.allyear.__intensity.std  3.733  4.334   5.478
0.000149 ***
s(cxe,cyn):COR.YEARLY.MEAN  28.979 29.886  28.502  <
2e-16 ***
s(cxe,cyn):cFISHING_INTENSITY.__std5.592  6.464  24.330
  < 2e-16 ***
-
Rank: 1/169
R-sq.(adj) =  0.697   Deviance explained = 64.1%
GCV = 0.70372  Scale est. = 0.69061   n = 4575

 [[alternative HTML version deleted]]


R-help@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.



--
Simon Wood, Mathematical Science, University of Bath BA2 7AY UK
+44 (0)1225 386603 
http://people.bath.ac.uk/sw283





--
Simon Wood, Mathematical Science, University of Bath BA2 7AY UK
+44 (0)1225 386603   http://people.bath.ac.uk/sw283

__
R-help@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.

Re: [R] GAM model output error(?)

2014-07-16 Thread Trevor Davies 
Hi Simon,

Thank you so much for being active on this list, it really is tremendously
helpful.

Thanks you for your insights, I was wondering whether both terms were
necessary.

As for 'NB wouldn't work' it was a convergence problem (and tremendously
slow).  There were also issues with the models not giving lack of
convergence warnings but then the model saying it was not converged and
vice versa.  The QP seemed to fit the data well (and quickly) so we
switched to that.

Thanks again.
Trevor


On Wed, Jul 16, 2014 at 12:44 PM, Simon Wood  wrote:

> Trevor,
>
> It looks like you've added a parametric COR.YEARLY.MEAN in addition to
> your s(cxe,cyn,by=COR.YEARLY.MEAN) term. Because the latter includes a
> linear effect of COR.YEARLY.MEAN, then the parametric COR.YEARLY.MEAN will
> not be identifiable, so gam has dropped it.
>
> I guess from the scale parameter your data are, if anything,
> under-dispersed relative to Poisson, but perhaps Poisson would have been ok
> too.
>
> Could you give a bit more detail on 'NB just wouldn't work' please? What
> actually happened? (It's new functionality, so having reports when stuff
> goes wrong is useful).
>
> best,
> Simon
>
>
> On 16/07/14 20:25, Trevor Davies wrote:
>
>> I have run a quasipoisson spatial model via GAM (NB just wouldn't work)
>> and
>> I am getting the following output of one of my parameters
>> (COR.YEARLY.MEAN).  Does this suggest an error in the model fit? The model
>> seems to have converged.  Apologies for the lack of reproducible example
>> but it didn't really seem warranted in this case.  Thank you for the
>> assistance.
>> Trevor
>>
>> Parametric coefficients:
>>  Estimate Std. Error t value Pr(>|t|)
>> (Intercept)  -3.7084 0.2185  -16.97   <2e-16 ***
>> COR.YEARLY.MEAN   0. 0.  NA   NA
>> ---
>>
>> Approximate significance of smooth terms:
>> edf Ref.df   F  p-value
>> s(PRED_cBOTTOM_TEMPERATURE.std)  5.008  6.045  24.871  < 2e-16
>> ***
>> s(cxe,cyn)  28.999 29.000  13.951  < 2e-16 ***
>> s(cDMAX.std) 3.830  4.761 243.895  < 2e-16 ***
>> s(cxe,cyn):male.allyear.intensity.std8.053  9.923   4.486  2.8e-06
>> ***
>> s(cxe,cyn):female.allyear.intensity.std  3.733  4.334   5.478 0.000149
>> ***
>> s(cxe,cyn):COR.YEARLY.MEAN  28.979 29.886  28.502  < 2e-16 ***
>> s(cxe,cyn):cFISHING_INTENSITY.std5.592  6.464  24.330  < 2e-16
>> ***
>> -
>> Rank: 1/169
>> R-sq.(adj) =  0.697   Deviance explained = 64.1%
>> GCV = 0.70372  Scale est. = 0.69061   n = 4575
>>
>> [[alternative HTML version deleted]]
>>
>> __
>> R-help@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.
>>
>>
>
> --
> Simon Wood, Mathematical Science, University of Bath BA2 7AY UK
> +44 (0)1225 386603   http://people.bath.ac.uk/sw283
>

[[alternative HTML version deleted]]

__
R-help@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.

Re: [R] GAM model output error(?)

2014-07-16 Thread Simon Wood 

Trevor,

It looks like you've added a parametric COR.YEARLY.MEAN in addition to 
your s(cxe,cyn,by=COR.YEARLY.MEAN) term. Because the latter includes a 
linear effect of COR.YEARLY.MEAN, then the parametric COR.YEARLY.MEAN 
will not be identifiable, so gam has dropped it.


I guess from the scale parameter your data are, if anything, 
under-dispersed relative to Poisson, but perhaps Poisson would have been 
ok too.


Could you give a bit more detail on 'NB just wouldn't work' please? What 
actually happened? (It's new functionality, so having reports when stuff 
goes wrong is useful).


best,
Simon

On 16/07/14 20:25, Trevor Davies wrote:

I have run a quasipoisson spatial model via GAM (NB just wouldn't work) and
I am getting the following output of one of my parameters
(COR.YEARLY.MEAN).  Does this suggest an error in the model fit? The model
seems to have converged.  Apologies for the lack of reproducible example
but it didn't really seem warranted in this case.  Thank you for the
assistance.
Trevor

Parametric coefficients:
 Estimate Std. Error t value Pr(>|t|)
(Intercept)  -3.7084 0.2185  -16.97   <2e-16 ***
COR.YEARLY.MEAN   0. 0.  NA   NA
---

Approximate significance of smooth terms:
edf Ref.df   F  p-value
s(PRED_cBOTTOM_TEMPERATURE.std)  5.008  6.045  24.871  < 2e-16 ***
s(cxe,cyn)  28.999 29.000  13.951  < 2e-16 ***
s(cDMAX.std) 3.830  4.761 243.895  < 2e-16 ***
s(cxe,cyn):male.allyear.intensity.std8.053  9.923   4.486  2.8e-06 ***
s(cxe,cyn):female.allyear.intensity.std  3.733  4.334   5.478 0.000149 ***
s(cxe,cyn):COR.YEARLY.MEAN  28.979 29.886  28.502  < 2e-16 ***
s(cxe,cyn):cFISHING_INTENSITY.std5.592  6.464  24.330  < 2e-16 ***
-
Rank: 1/169
R-sq.(adj) =  0.697   Deviance explained = 64.1%
GCV = 0.70372  Scale est. = 0.69061   n = 4575

[[alternative HTML version deleted]]

__
R-help@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.




--
Simon Wood, Mathematical Science, University of Bath BA2 7AY UK
+44 (0)1225 386603   http://people.bath.ac.uk/sw283

__
R-help@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.