[R-sig-ME] hypothesis testing - full model doesn't converge

[Rion Lerm]

I got one of two warnings depending on my response variable i.e.,:

Warning message:
In fitTMB(TMBStruc) :
  Model convergence problem; non-positive-definite Hessian matrix. See
vignette('troubleshooting')

or

Warning message:
In fitTMB(TMBStruc) :
  Model convergence problem; extreme or very small eigenvalues detected.
See vignette('troubleshooting')

Attached is the code I employed for scaling the predictor effects that
compare the successful and unsuccessful model runs. (Disclaimer: the steps
came from a vignette not built by me).

Run various models with different combos of predictor fixed effects and
compare AIC or BIC values and select the smallest or similar depending on
importance of predictor inclusion. (PS diversity measures also produced
large collinearity with most of my predictors; I inspected this issue by
running generliased pairs plots of the predictor fixed effects (see GGally
package's ggpairs() function)).

Regards/Groete and I hope this helps or I have addressed your query
successfully,

*Rion Lerm*

SAEON Ndlovu

C: 076 913 8381

W: 013 735 3536

SAEON / Scientific Services, Kruger National Park Gate, Phalaborwa, South
Africa

South African Environmental Observation Network

Website <https://ndlovu.saeon.ac.za/> & Location of our offices
<https://www.google.co.za/maps/place/SAEON+Ndlovu+Node/@-23.9447244,31.164957,18z/data=!4m5!3m4!1s0x1ec36ec96261cfb7:0xb89a27144ba0ca2a!8m2!3d-23.9450103!4d31.1652626>



My research profiles:
*ResearchGate* <https://www.researchgate.net/profile/Rion-Lerm>

*ORCID <https://orcid.org/0000-0002-9992-2093>*
*Google Scholar
<https://scholar.google.com/citations?user=JV5d7E4AAAAJ&hl=en>*


On Wed, 15 Mar 2023 at 10:12, Guillaume Adeux <guillaumesimon.a2 using gmail.com>
wrote:

> Hello everyone,
> I'm currently trying to analyse the relationship between crop diversity and
> pesticide use at a national scale (4k farms covering 7 climatic regions),
> while accounting for crop identity effects (proportion of each crop in the
> cropping system), in order to distinguish "true diversity" effects from
> "dilution effects" (introducing a meadow in a cropping system necessarily
> reduces pesticide use).
> The most simple model would be something like:
> *mod=glmmTMB(pesticide_use~climatic_region + crop_1 to crop_25 +
> crop_diversity + (1|climatic_region:soil)+(1|year),
> family=Tweedie(link="log"), data=pest)*
> This model converges but doesn't account for the fact that - for a given
> crop - pesticide use might be more important under certain weather
> conditions (i.e. climatic regions), which is something well known for
> ecologists and agronomists.
> Hence, a more complete model would be :
> *mod_inter=glmmTMB(pesticide_use~climatic_region * (crop_1 to crop_25 +
> crop_diversity) + (1|climatic_region:soil)+(1|year),
> family=Tweedie(link="log"), data=pest)*
> However, this "full" model doesn't converge, even when I try to boost the
> number of iterations or the fitting algorithm.
> I was hence tempted to identify the most complete model that converges (for
> example via the R buildmer package) but I understand from my reading that
> model selection and hypothesis testing don't go well together (pvalues are
> biased after model selection). I wasn't able to find whether or not it was
> possible to correct these pvalues after model selection.
> I was suggested (by Bert van der Veen) to look into variable selection
> (glmmLasso for example) but it appears less flexible (Tweedie or gamma
> families are not implemented and I will most likely have to correct for
> spatial autocorrelation in my selected model, which is possible in glmmTMB
> but not glmmLasso).
> Because of multicollinearity problems including crop diversity, I was
> thinking of identifying the most complete crop model (only interactions
> between climatic region and crop proportions) and comparing it to the same
> model including crop diversity and its interaction with crop region.
> How would you more knowledgeable folks go about tackling this problem? Is
> this procedure a little dubious ?
> Thanks a lot for your feedback.
> Have a great day.
> GA2
>
>         [[alternative HTML version deleted]]
>
> _______________________________________________
> R-sig-mixed-models using r-project.org mailing list
> https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models
>

-- 
http://www.nrf.ac.za/email-legal-notice.php 
<http://www.nrf.ac.za/email-legal-notice.php>

-------------- next part --------------
An embedded and charset-unspecified text was scrubbed...
Name: glmmTMB and scaling predictors.txt
URL: <https://stat.ethz.ch/pipermail/r-sig-mixed-models/attachments/20230317/b4c4c00b/attachment.txt>