[R] a warning message from lmer

[Douglas Bates]

By the way, did you notice that the levels of Emotion include both
"joy" and "joy ".  You may want to correct that.

On Sat, Jul 12, 2008 at 7:47 AM, Douglas Bates <bates at stat.wisc.edu> wrote:
> On Sat, Jul 12, 2008 at 6:23 AM, Lan Wei <lanwei at umich.edu> wrote:
>> Hi all,
>>
>> I have a problem when running lmer.
>> In my data set, Agree is a binary(0/1) response. WalkerID and ObsID is
>> the identification number of the subjects. the description of the
>> other variables are as follows:
>>>
>>> levels(regdat$Display)
>>
>> [1] "Dynamic" "Static"
>>>
>>> levels(regdat$Survey)
>>
>> [1] "HM1_A" "HM1_B" "HM1_C" "HM2_A" "HM2_B" "HM2_C" "ST_A"  "ST_B"
>> "ST_C"
>>>
>>> levels(regdat$Emotion)
>>
>> [1] "aneu" "ang"  "con"  "joy"  "joy " "sad"
>>>
>>> levels(regdat$ObsGender)
>>
>> [1] "F" "M"
>>>
>>> levels(regdat$WalkerGender)
>>
>> [1] "F" "M"
>>
>> the watning is:
>>>
>> fit1<-lmer(Agree~Display+Survey+Emotion+WalkerGender+ObsGender+(1|WalkerID)+(1|ObsID),family=binomial(link='logit'),data=regdat)
>> Warning message:
>> In mer_finalize(ans, verbose) : gr cannot be computed at initial par
>> (65)
>
>> Does anybody have some hint to solve this problem? I'd very much appreciate
>> it!
>
> In situations like this it is best to add the argument
>
> verbose = TRUE
>
> in the call to lmer so that you can see the progress of the
> iterations.  (Also, you may want to call glmer directly.  When you
> call lmer with a non-gaussian family it simply calls glmer.  You can
> avoid the extra step.)
>
> This call is returning a warning about evaluation of the gradient at
> the initial values of the parameters.  I'm not sure if it then goes on
> to optimize the approximated deviance.
>
> If the approximated deviance is not being minimized for this model you
> may want to start with a simpler model, omitting some of the terms in
> the fixed effects.
>