[R-sig-ME] Interaction between random and fixed effects

[Thierry Onkelinx]

No. It doesn't make sense to use an interaction term. Year + (1|Local) is
sufficient in this case.

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 ma 31 mei 2021 om 22:02 schreef Vinicius Maia <
vinicius.a.maia77 using gmail.com>:

> Thank you a lot Thierry,
>
> But besides that, is it sensical to estimate an interaction such as
> local:year if local is nested within year? I believe it does not make sense
> but when this model ran I started to question what I know.
>
> Best wishes,
>
> Vinícius
>
>
>
> Em seg., 31 de mai. de 2021 às 16:52, Thierry Onkelinx <
> thierry.onkelinx using inbo.be> escreveu:
>
>> Dear Vinicius,
>>
>> I think the problem is with your response variable. It seems like you
>> have a lot of observations with a response value a few orders of magnitude
>> smaller than the global average. This grouping is not explained by any of
>> the covariates in your model, leading to huge random effect BLUPs.
>> Splitting the BLUPs over two variables probably yields a smaller penalty.
>>
>> Fitting the model with a log transformed response leads to a singular
>> model with 0 variance for Local. This strengthens my belief that the
>> problem is with the data.
>>
>> 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 ma 31 mei 2021 om 21:21 schreef Vinicius Maia <
>> vinicius.a.maia77 using gmail.com>:
>>
>>> Dear Thierry,
>>>
>>> Thank you for your response.
>>> Local is coded with the name of the local, but I believe the nesting is
>>> implicit in the data.
>>>
>>> with(dataset, isNested(as.character(Local), as.character(Year)))
>>> returns TRUE
>>>
>>> The example is attached.
>>>
>>> Best wishes,
>>>
>>> Vinícius
>>>
>>> Em seg., 31 de mai. de 2021 às 16:10, Thierry Onkelinx <
>>> thierry.onkelinx using inbo.be> escreveu:
>>>
>>>> Dear Vinicius,
>>>>
>>>> What did you ran the interaction as a fixed effect Year:Local or a
>>>> random effect (1|Year:Local)?  How did you code Local: as a unique value
>>>> for every Local and Year combinations? Please do share output or a minimal
>>>> example so we know exactly what you did. I'm still a novice at mind reading.
>>>>
>>>> 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 ma 31 mei 2021 om 20:57 schreef Vinicius Maia <
>>>> vinicius.a.maia77 using gmail.com>:
>>>>
>>>>> Hi all,
>>>>>
>>>>> I have a subtle doubt in how to interpret the interaction between
>>>>> fixed (or
>>>>> even random) and random effects in the following case.
>>>>>
>>>>> I have a model: Y ~ Year + (1|Local)+(1|Genotype) +
>>>>> (Year:Local:Genotype)
>>>>> Year is a fixed effect because it has only 4 levels.
>>>>>
>>>>> I ran the model with the random and fixed effect interaction just to
>>>>> explore, but I was not expecting that it would work because Locals are
>>>>> completely nested within Years.
>>>>>
>>>>> To my surprise, the model ran and the variance of Year:Local:Genotype
>>>>> are
>>>>> quite big. How is it possible to have Local interacting with Year if
>>>>> they
>>>>> are nested? I also tried: Y ~ Year + (1|Local)+(1|Genotype) +
>>>>> (Year:Local)
>>>>> and the model rans too, without singular fit.
>>>>>
>>>>> I am struggling to understand if random interactions (it also extends
>>>>> to
>>>>> cases where the interactions are only between nested random effects)
>>>>> mean
>>>>> that the variance between Locals changes with Years or if the effect
>>>>> of a
>>>>> given Local changes with Years. If it is the former option I can
>>>>> understand
>>>>> why the model ran and has a high variance for the interaction, but if
>>>>> it is
>>>>> the later case (which I believe it is), how does the model estimate an
>>>>> interaction for Local:Year if they are nested?
>>>>>
>>>>> Thanks!
>>>>>
>>>>> Best wishes,
>>>>>
>>>>> Vinícius Maia
>>>>>
>>>>>         [[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]]