[ADMB Users] Convergence failed:log-likelihood of gradient=

[Ben Bolker]

On 12-02-15 06:14 AM, Håkon Holand wrote:
> Dear GlmmADMB-users,
> 
> I am running models with a binary respons variable (0/1) and wanted to
> see if there was a linear trend through time (day of the season 1-117).
> I am using individual identity as a random variable in all my models. I
> am currently using R 2.14.1 (64-bit) and glmmadmb v.0.7.2.5.
> The problem is that when I run my simplest model : SYN ~ sesday +
> (1|RINGNR) family="binomial", I get the following warnings:
> 
> Warning messages:
> 1: In glmmadmb(SYN ~ sesday + (1 | RINGNR), family = "binomial", data = helgesmr) :
>   Convergence failed:log-likelihood of gradient= -0.0461339
> 2: In print.glmmadmb(list(n = 1548L, q = 1103L, formula = SYN ~ sesday +  :
>   Object has a large gradient component
> '

  I would probably start by centering your 'sesday' variable.  That may
or may not be the problem (glmmADMB tries to orthogonalize the design
matrix internally.  I would also try running with verbose=TRUE to see if
the messages (especially at the end) are any larger, and possibly
setting the maximum number of iterations larger.  As for the model fit:
the intercept is very small (on the logit scale), probably because the
zero point of 'sesday' is very far below your data range; the standard
deviation appears worryingly large.  In addition to the other
suggestions, I would (a) plot the data and (b) plot the predictions of
the model to see if they make sense.

  This model could also be tried in glmer (lme4 package).

  Ben Bolker

> 
> Summary of the model:
> 
> 
> Call:
> glmmadmb(formula = SYN ~ sesday + (1 | RINGNR), data = helgesmr, 
>     family = "binomial")
> 
> 
> Coefficients:
>              Estimate Std. Error z value Pr(>|z|)    
> (Intercept) -13.14000    0.99477  -13.21  < 2e-16 ***
> sesday        0.04878    0.00888    5.49  3.9e-08 ***
> ---
> Signif. codes:  0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1 
> 
> Number of observations: total=1548, RINGNR=1103 
> Random effect variance(s):
> Group=RINGNR
>             Variance StdDev
> (Intercept)    249.3  15.79
> 
> Log-likelihood: -566.279 
> 
>  
> 
> 
> Is this a "overly cautious" warning from glmmadmb or a more serious problem?
> I also have to say that when I add a second order term (sesday^2)
> everything works fine. I am just a bit curious why the "simpler" model
> ran into trouble.
> 
> I really like all the new functionality of glmmADMB so I would like to
> thank the
> developers!
> 
> Cheers,
> 
> Håkon
> ---------------------------------------------------------
> Håkon Holand, PhD.Student
> Centre for Conservation Biology
> Department of Biology
> Norwegian University of Science and Technology
> NO-7491 Trondheim
> Norway
> 
> 
> _______________________________________________
> Users mailing list
> Users at admb-project.org
> http://lists.admb-project.org/mailman/listinfo/users