Re: [r-inla] Tweedie distribution for spatial and spatiotemporal models?

[Håvard Rue]

On Tue, 2014-01-28 at 09:48 -0800, James Thorson wrote:
> Hi all,
>
>
> is there any plan for including and/or a work-around for using a
> Tweedie distribution as the likelihood of an SPDE model? Or perhaps
> some other zero-inflated model for continuous-valued positive
> numbers? The tweedie in particular is useful because it implies a
> link between the expected value of non-zero numbers, and the
> probability of encountering a zero.
>
>
> The circumstance is for modelling the weight of a target encountered
> during ecological sampling. Analysts often use
> "delta-models" (separate binomial and gamma/lognormal models for
> presence/absence, and non-zero densities respectively).

Hi,

we add likelihood models when needed... if you could send me info about
it, your proposed link function and parameterseriation, I'll have a
look. maybe we could do this off-list.

Best
H


--
Håvard Rue
Department of Mathematical Sciences
Norwegian University of Science and Technology
N-7491 Trondheim, Norway
Voice: +47-7359-3533 URL : http://www.math.ntnu.no/~hrue
Mobile: +47-9260-0021 Email: havar...@math.ntnu.no

R-INLA: www.r-inla.org

[Christopher Paciorek]

We'd be interested in this as well. As part of a large project on paleoecology and forest ecology, we're fitting Tweedie models to biomass data using gam from mgcv, and a Bayesian alternative would be of interest for comparison. We find that gam() with Tweedie likelihood has convergence issues when there are large spatial areas that are all zero, as occurs with biomass data in areas outside of the spatial range of a given species.

Havard, if this is something the INLA team finds worth pursuing, I or someone else on the project can pass along info about the Tweedie. 

-Chris

[Håvard Rue]

Finn.L discussed the Tweedie issue with S.Wood(mgcv) and although we can
extract code from mgcv, the tweedie-distr-class is a
(numerical) mess, and we will just inherit the convergence problem as
well...

H

[paradin...@gmail.com]

Hi all,

I've seen that the use of the Tweedie distribution is becoming more and more common in different R packages. I wonder whether there's been some sort of development that makes the Tweedie somehow more stable.

I agree with James Thorson and Chris.  A continuous ZIM similar to the zeroinflatedpoisson/nbinomial2 would be an invaluable addition to model many ecological processes. It allows to link the probability of presence and the abundance in zero inflated continuous datasets. Delta model or hurdle models can be particularly annoying due to their independence.

Best,

iosu

[INLA help]

The tweedie distribution is there in the package. 

Zero inflated continuous distributions does not need special implementation as a 0 is know to come from the point mass. Hence a two-likelihood approach is the easy way out. Hence a general model can be used for the zeros and also for the non zeros 

—

Haavard Rue

HelpDesk 

help@r-inla. org

--
You received this message because you are subscribed to the Google Groups "R-inla discussion group" group.
To unsubscribe from this group and stop receiving emails from it, send an email to r-inla-discussion...@googlegroups.com.
To view this discussion on the web, visit https://groups.google.com/d/msgid/r-inla-discussion-group/cda0a587-5d10-4d1a-9110-a0aa57591c3dn%40googlegroups.com.

[paradin...@gmail.com]

Indeed! Sorry I missed its implementation. It's great news!