[R-sig-ME] Likelihood drops on adding random effect

[Douglas Bates]

Is it possible to provide the data to us, Murray?  We're putting the
final touches on a new version of lme4 and would like to head off any
glitches before release.

The new version uses different optimizers and a somewhat different
representation of the model.  It also allows for evaluation of the
deviance by adaptive Gauss-Hermite quadrature (AGQ) which, as Dave
Fournier indicates, may stabilize the optimization.

On Thu, May 17, 2012 at 6:36 PM, Murray Jorgensen <maj at waikato.ac.nz> wrote:
> I guess we may have a case of a sort of spike in the loglikelihood that
> indicates that the MLE (or at least the lmer estimate) for the model without
> the 'tree' effect is spurious. (Probably the sort of thing that would be
> smoothed away by any reasonable prior.) I really was looking more closely at
> the other fixed effect parameters, regarding the intercept as a bit of a
> nuisance. I hope to say more about the data set later. Maybe I will do a
> sort of profile around the intercept value and just fit the other
> parameters.
>
> I would like to ask the list for some more help with the 'start =' parameter
> though because I don't seem to be quite using it correctly.
>
>> ip2 = c(0, 8.29931, 2.56368e-06,  0.00000)
>
>> f0128bit = lmer(fincr ~ icfac + (1|gap) + (1|box)+ (1|gap:box) + (1|tree),
>> family = binomial,
> +                  verbose = TRUE, start = ip2)
>  0:     775.67797:  0.00000  8.29931 2.56368e-06  0.00000 -3.22238 1.44743
> -2.20257  1.02515
>  1:     747.52247:  0.00000  5.04306 6.75518e-05  0.00000 -4.34098 2.12417
> -3.74289 -0.121576
>  2:     736.84946:  0.00000  4.55878 0.000215665  0.00000 -4.39655 1.52182
> -3.63273 -0.115459
> ...
>  45:     671.01076: 4.57910e-09 0.287720 0.350164 0.711054 -3.43481 1.45346
> -2.57398 0.293897
>  46:     671.01076:  0.00000 0.287721 0.350162 0.711064 -3.43483 1.45346
> -2.57397 0.293955
>
> That works well for initialising the variance parameters only.
>
>
>> f0128bit = lmer(fincr ~ icfac + (1|gap) + (1|box)+ (1|gap:box) + (1|tree),
>> family = binomial,
> +               verbose = TRUE, start = list( fixef = list( -9.28405,
> 2.81300, -4.75935,  2.91080),
> + ST = list(0, 8.29931, 2.56368e-06,  0.00000))
> + )
> Error: class(STnew[[i]]) == class(ST[[i]]) is not TRUE
> In addition: Warning message:
> In sort(names(start)) == sort(names(FL)) :
>  longer object length is not a multiple of shorter object length
>
> A failed attempt to initialise both fixed and random parameters.
>
>
>> f0128bit = lmer(fincr ~ icfac + (1|gap) + (1|box)+ (1|gap:box) + (1|tree),
>> family = binomial,
> +               verbose = TRUE, start = list( fixef = c( -9.28405, 2.81300,
> -4.75935,  2.91080),
> + ST = c(0, 8.29931, 2.56368e-06,  0.00000))
> + )
> Error: is.list(STnew) is not TRUE
> In addition: Warning message:
> In sort(names(start)) == sort(names(FL)) :
>  longer object length is not a multiple of shorter object length
>
> Another failure.
>
>
>> initpar = c(0, 8.29931, 2.56368e-06,  0.00000, -9.28405,  2.81300,
>> -4.75935,  2.91080)
>> f0128bit = lmer(fincr ~ icfac + (1|gap) + (1|box)+ (1|gap:box) + (1|tree),
>> family = binomial,
> +                  verbose = TRUE, start = initpar)
>  0:     696.35176:  1.15470 0.222222 0.182574 0.157135 -3.22238 1.44743
> -2.20257  1.02515
>  1:     687.88422: 0.646498 0.349241 0.199295 0.375671 -4.02981 1.29061
> -2.23330  1.01588
>  ...
>
>  26:     671.01076:  0.00000 0.287718 0.350162 0.711060 -3.43482 1.45346
> -2.57397 0.293937
>  27:     671.01076:  0.00000 0.287721 0.350163 0.711061 -3.43482 1.45346
> -2.57398 0.293936
>
> Runs, but does not seem to use the supplied starting parameters.
>
> Thanks, for your comments.
>
> Murray
>
>
> On 18/05/2012 5:48 a.m., Douglas Bates wrote:
>>
>> Look at the values of the coefficients and standard deviations that
>> you are "converging" to.  Your intercept is -9.28, which, with a
>> binomial family, corresponds to probabilities below 1e-4.  With icfac
>> = fem the linear predictor is -9.28 - 4.76 = -14.04 corresponding to a
>> probability of 8e-07.
>>
>> You are going to need to look at the data and the proportions of
>> positives for different levels of icfac to see what would make sense.
>> This problem will create a very ill-defined likelihood surface because
>> the fitted values will lose sensitivity to the parameters when the
>> probabilities are so extreme.  If you start extreme values you will
>> never be able to converge.
>>
>> On Thu, May 17, 2012 at 12:37 AM, Murray Jorgensen<maj at waikato.ac.nz>
>>  wrote:
>>>
>>> PS  I also tried
>>>
>>> start = list( fixef = c( -9.28405,  2.81300, -4.75935,  2.91080),
>>>        ST = c(0, 8.29931, 2.56368e-06,  0.00000))
>>>
>>> and
>>>
>>> start = list( fixef = list( -9.28405,  2.81300, -4.75935,  2.91080),
>>>       ST = list(0, 8.29931, 2.56368e-06,  0.00000))
>>>
>>> to no avail.
>>>
>>>
>>> _______________________________________________
>>> R-sig-mixed-models at r-project.org mailing list
>>> https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models
>
>
> --
> Dr Murray Jorgensen      http://www.stats.waikato.ac.nz/Staff/maj.html
> Department of Statistics, University of Waikato, Hamilton, New Zealand
> Email: maj at waikato.ac.nz      majmurr at gmail.com         Fax 7 838 4155
>
> Phone  +64 7 838 4773 wk    Home +64 7 825 0441   Mobile 021 0200 8350