[R-sig-ME] singular fit

Thu Jan 2 17:02:16 CET 2020

Dear Jill,

See e.g. the help file of drop1(). The dummy example below demonstrates
that drop1() only removes main effects that are not contained in an
interaction. See also Venables (200) Exegeses on Linear Models.

m0 <- lm(Fertility ~ Agriculture * Education + Examination, data = swiss)
drop1(m0)

The LRT you ran, just compares two numbers based on both models. It does
not check whether the comparison makes sense.

I would strongly recommend that you look for help with a local statistician.

Best regards,

ir. Thierry Onkelinx
Statisticus / Statistician

Vlaamse Overheid / Government of Flanders
INSTITUUT VOOR NATUUR- EN BOSONDERZOEK / RESEARCH INSTITUTE FOR NATURE AND
FOREST
Team Biometrie & Kwaliteitszorg / Team Biometrics & Quality Assurance
thierry.onkelinx using inbo.be
Havenlaan 88 bus 73, 1000 Brussel
www.inbo.be

///////////////////////////////////////////////////////////////////////////////////////////
To call in the statistician after the experiment is done may be no more
than asking him to perform a post-mortem examination: he may be able to say
what the experiment died of. ~ Sir Ronald Aylmer Fisher
The plural of anecdote is not data. ~ Roger Brinner
The combination of some data and an aching desire for an answer does not
ensure that a reasonable answer can be extracted from a given body of data.
~ John Tukey
///////////////////////////////////////////////////////////////////////////////////////////

<https://www.inbo.be>

Op do 2 jan. 2020 om 15:20 schreef Jill Brouwer <jilbo97 using gmail.com>:

> Dear Thierry,
>
> There are 2 males and 2 females per block and 2 trials for each cross (so
> 8 trials in all) per treatment (pH). I have 16 blocks in total.
>
> Fertilisation is binomial (success ie number of fertilised eggs out of 100)
>
> I am trying to determine whether the random effects and their interactions
> are significant in the model. I don’t understand what you mean by that last
> point? Is there possibly a reference you could direct me to? Other
> likelihood ratio tests I have done when removing male but keeping the
> interaction have come up as significantly different to the full model ?
> Sorry I am still quite new to all of this.
>
> Kind regards,
> Jill
>
> On Thursday, January 2, 2020, Thierry Onkelinx <thierry.onkelinx using inbo.be>
> wrote:
>
>> Dear Jill,
>>
>> I presume you use different males and females for each block? How many
>> blocks? How many trials per block (success + failure)? Is fertilisation a
>> discrete variable?
>>
>> Removing a main (random) effect like (1|Male) while keeping interactions
>> (1|Male:Female) doesn't make sense. You'll get exactly the same model fit
>> with a different parametrisation as the interaction will model the main
>> effect.
>>
>> Best regards,
>>
>> ir. Thierry Onkelinx
>> Statisticus / Statistician
>>
>> Vlaamse Overheid / Government of Flanders
>> INSTITUUT VOOR NATUUR- EN BOSONDERZOEK / RESEARCH INSTITUTE FOR NATURE
>> AND FOREST
>> Team Biometrie & Kwaliteitszorg / Team Biometrics & Quality Assurance
>> thierry.onkelinx using inbo.be
>> Havenlaan 88
>> <https://www.google.com/maps/search/Havenlaan+88?entry=gmail&source=g>
>> bus 73, 1000 Brussel
>> www.inbo.be
>>
>>
>> ///////////////////////////////////////////////////////////////////////////////////////////
>> To call in the statistician after the experiment is done may be no more
>> than asking him to perform a post-mortem examination: he may be able to say
>> what the experiment died of. ~ Sir Ronald Aylmer Fisher
>> The plural of anecdote is not data. ~ Roger Brinner
>> The combination of some data and an aching desire for an answer does not
>> ensure that a reasonable answer can be extracted from a given body of data.
>> ~ John Tukey
>>
>> ///////////////////////////////////////////////////////////////////////////////////////////
>>
>> <https://www.inbo.be>
>>
>>
>> Op do 2 jan. 2020 om 10:09 schreef Jill Brouwer <jilbo97 using gmail.com>:
>>
>>> Sorry here is some more information:
>>>
>>> My research is looking at whether ocean acidification affects patterns
>>> of gamete compatibility between individual male/female mussels.
>>> Here I am looking at whether the ph of the fertilisation assays also
>>> influences male by female interactions.
>>>
>>> The design consists of two males and two females, crossed in every
>>> combination (so a total of four combinations) per block, with two
>>> replicates in each.
>>> There is a fixed effect of Fertilisation pH (just called Fertilisation
>>> below)
>>> Random effects are individual males and females (each assigned a unique
>>> number, but specified as a factor for the model), and block.
>>>
>>> the full model formula is this (which doesn't give the singular fit
>>> error):
>>> fertphmodel <- glmer(cbind(Success,Failure) ~ Fertilisation + (1|Block)
>>> + (1|Male) + (1|Female) + (1|Male:Female) +
>>>                       (1|Male:Fertilisation) + (1|Female:Fertilisation)
>>> + (1|Male:Female:Fertilisation),
>>>                     family = "binomial", data = fertph)
>>>
>>> I am using likelihood ratio testing to determine significance of the
>>> random effects, however when I create the reduced model with (1|Male)
>>> removed, and also the one for (1|Male:Female) removed, it spits out the
>>> singular fit error. (Formulas below). I was also reading about boundary
>>> effect problems with likelihood ratio testing, and am unsure how to account
>>> for this?
>>>
>>> fertphmodel1 <- glmer(cbind(Success,Failure) ~ Fertilisation + (1|Block)
>>> + (1|Female) + (1|Male:Female) +
>>>                         (1|Male:Fertilisation) +
>>> (1|Female:Fertilisation) + (1|Male:Female:Fertilisation),
>>>                       family = "binomial", data = fertph)
>>>
>>> fertphmodel3 <- glmer(cbind(Success,Failure) ~ Fertilisation + (1|Block)
>>> + (1|Male) + (1|Female) +
>>>                        (1|Male:Fertilisation) + (1|Female:Fertilisation)
>>> + (1|Male:Female:Fertilisation),
>>>                      family = "binomial", data = fertph)
>>>
>>> For fertphmodel1, the summary output says that the female random effect
>>> has an extremely low variance  (possibly a reason for singular fit?)
>>> var: 7.070e-10 sd: 2.659e-05
>>>
>>> And for fertphmodel3, the summary output says the Female:Fertilisation
>>> has a very low variance
>>> var 3.325e-10 sd 1.823e-05
>>>
>>> However, in the full model the all of the variances of the random
>>> effects are between 0.03 and 0.6.
>>>
>>> Hopefully this helps a bit !
>>>
>>> Thankyou,
>>> Jill
>>>
>>> On Thu, Jan 2, 2020 at 4:47 PM Thierry Onkelinx <
>>> thierry.onkelinx using inbo.be> wrote:
>>>
>>>> Dear Jill,
>>>>
>>>> Can you share the model formula and the design of your experiment? It's
>>>> hard to answer your question without such basic information.
>>>>
>>>> Best regards,
>>>>
>>>> ir. Thierry Onkelinx
>>>> Statisticus / Statistician
>>>>
>>>> Vlaamse Overheid / Government of Flanders
>>>> INSTITUUT VOOR NATUUR- EN BOSONDERZOEK / RESEARCH INSTITUTE FOR NATURE
>>>> AND FOREST
>>>> Team Biometrie & Kwaliteitszorg / Team Biometrics & Quality Assurance
>>>> thierry.onkelinx using inbo.be
>>>> Havenlaan 88
>>>> <https://www.google.com/maps/search/Havenlaan+88?entry=gmail&source=g>
>>>> bus 73, 1000 Brussel
>>>> www.inbo.be
>>>>
>>>>
>>>> ///////////////////////////////////////////////////////////////////////////////////////////
>>>> To call in the statistician after the experiment is done may be no more
>>>> than asking him to perform a post-mortem examination: he may be able to say
>>>> what the experiment died of. ~ Sir Ronald Aylmer Fisher
>>>> The plural of anecdote is not data. ~ Roger Brinner
>>>> The combination of some data and an aching desire for an answer does
>>>> not ensure that a reasonable answer can be extracted from a given body of
>>>> data. ~ John Tukey
>>>>
>>>> ///////////////////////////////////////////////////////////////////////////////////////////
>>>>
>>>> <https://www.inbo.be>
>>>>
>>>>
>>>> Op do 2 jan. 2020 om 06:47 schreef Jill Brouwer <jilbo97 using gmail.com>:
>>>>
>>>>> Hi all,
>>>>>
>>>>> I have fitted a GLMM using glmer in lme4, and when I run the model it
>>>>> comes
>>>>> out with a singular fit warning.
>>>>>
>>>>> However when I ran the isSingular command on it and changed the
>>>>> tolerance
>>>>> to 1e-05 instead of the default 1e-04 that caused the original
>>>>> warning, it
>>>>> comes out as false - no singular fit warning!
>>>>>
>>>>> Does this mean that the first warning is a false positive?
>>>>> I can't find anything that suggests what the tolerance ratio should be
>>>>> but
>>>>> in the GLMM FAQ on github, the troubleshooting example uses 1e-05.
>>>>>
>>>>> Is it fine to stay with this model - I would prefer it to include all
>>>>> the
>>>>> random effects as they are all of interest to me, and the model itself
>>>>> is
>>>>> structured based on how I ran my experiment.
>>>>>
>>>>> Sorry if this is a basic question, I am still learning!
>>>>>
>>>>> Kind regards,
>>>>> Jill
>>>>>
>>>>>         [[alternative HTML version deleted]]
>>>>>
>>>>> _______________________________________________
>>>>> R-sig-mixed-models using r-project.org mailing list
>>>>> https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models
>>>>>
>>>>

	[[alternative HTML version deleted]]