[R-sig-ME] correlation between random effects
Jana Dlouha
jana.dlouha at inra.fr
Thu Feb 15 13:48:29 CET 2018
Hi all again,
Thanks to Thierry and Ben for their advices and sorry to answer so late, I was out of the lab yesterday.
I looked at the 20 species with the most extreme random effects and unfortunately, it is not systematically the least represented species, one has 13 repetitions so I think it is not possible to exclude them based on the quantitative criterion. There are probably some measurement problems but not related specifically to some species or families so...is there any way to handle it? What did you mean by altering the model?
Best regards
Jana
-----Message d'origine-----
De : Thierry Onkelinx [mailto:thierry.onkelinx at inbo.be]
Envoyé : mercredi 14 février 2018 09:47
À : Ben Bolker <bbolker at gmail.com>
Cc : Jana Dlouha <jana.dlouha at inra.fr>; r-sig-mixed-models at r-project.org
Objet : Re: [R-sig-ME] correlation between random effects
I agree with Ben. Look at 20 species with the most extreme random effects. Plot those, one at a time. See what is happening for those and decide what would be a sensible way to deal with it. That could be to alter the model or to remove data which doesn't contain the information needed by the model. E.g. all species with less than x observations.
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 at 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 ///////////////////////////////////////////////////////////////////////////////////////////
2018-02-13 18:34 GMT+01:00 Ben Bolker <bbolker at gmail.com>:
> you can use as.data.frame(ranef(fitted_model)) to extract the random
> effects as a data frame, then do anything you want to look at the most
> extreme species ...
>
> On Tue, Feb 13, 2018 at 10:29 AM, Jana Dlouha <jana.dlouha at inra.fr> wrote:
>> Dear Thierry,
>>
>> You are right, some species are represented only by one or two specimens, actually we did not want to use the mixed-effect models but the reviewers of our paper asked us to do that - but if I understand well from what you say, it is maybe not very smart considering the structure of our sample?
>> I am struggling to know which species is behaving differently - is there any efficient method to visualize that? I have plotted the random effects using plot_model function but not able to change the y_axis in order to be able to read it, with 600 Species everything is overlapped...
>> Thanks in advance
>>
>> Jana
>>
>> -----Message d'origine-----
>> De : Thierry Onkelinx [mailto:thierry.onkelinx at inbo.be] Envoyé :
>> mardi 13 février 2018 11:44 À : Jana Dlouha <jana.dlouha at inra.fr> Cc
>> : r-sig-mixed-models at r-project.org Objet : Re: [R-sig-ME] correlation
>> between random effects
>>
>> Dear Jana,
>>
>> Please keep the mailing list in cc.
>>
>> I meant both centering and scaling.
>>
>> Based on the summary of the model, you have on average 3.7 observations per species, which is a bit small for a random slope model. What worries me is that the summary of the data indicates several species with > 20 observation. Hence you will have lot of species with only 1 or 2 observations. A species with only 2 observations, a small difference in dB1 and a large difference in MC will likely result in a large random slope for dB1. You'll need to investigate which species have a strong random slope and why. Most of the time that is obvious once you plotted the data for that species.
>> Tip: plot the observations, the fitted values of the model and the predictions using only the fixed effects.
>>
>> 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 at 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
>> /////////////////////////////////////////////////////////////////////
>> //////////////////////
>>
>>
>>
>>
>> 2018-02-13 11:23 GMT+01:00 Jana Dlouha <jana.dlouha at inra.fr>:
>>> Dear Thierry,
>>>
>>> Thanks a lot for your reply. Yes, I have used dB1c also in the random effect.
>>> I have just tried to scale dB1 but I have still the same problem. However, it is possible that I am not doing things well, I am not a statistician and moreover I am just discovering R...
>>> You say that I should provide more information so here is the summary of my data for the two columns I am using in this model:
>>> Species MCs dB1
>>> 327 :43 Min. 40.05 Min. :1.050
>>> 135 :35 1st Qu. 72.53 1st Qu. :1.400
>>> 307 :24 Median 89.11 Median :1.560
>>> 146 :23 Mean 99.56 Mean :1.671
>>> 328 :23 3rd Qu. 116.23 3rd Qu. :1.840
>>> 341 :22 Max. 351.49 Max. :4.220
>>> (Other):2051
>>>
>>> How I centered and scaled dB1:
>>> dB1c<-scale(data$dB1,center=TRUE)
>>> dB1s<-scale(data$dB1,center=FALSE, scale=TRUE)
>>>
>>> summary of the model without centering or scaling dB1:
>>> Linear mixed model fit by maximum likelihood t-tests use Satterthwaite approximations to
>>> degrees of freedom [lmerMod]
>>> Formula: MCs ~ dB1 + (1 + dB1 | Species)
>>> Data: data
>>>
>>> AIC BIC logLik deviance df.resid
>>> 10720.4 10754.7 -5354.2 10708.4 2215
>>>
>>> Scaled residuals:
>>> Min 1Q Median 3Q Max
>>> -8.7305 -0.3491 0.0651 0.4508 6.6068
>>>
>>> Random effects:
>>> Groups Name Variance Std.Dev. Corr
>>> Species (Intercept) 21.48 4.635
>>> dB1 11.25 3.355 -1.00
>>> Residual 6.19 2.488
>>> Number of obs: 2221, groups: Species, 598
>>>
>>> Fixed effects:
>>> Estimate Std. Error df t value Pr(>|t|)
>>> (Intercept) -64.2618 0.4249 273.4200 -151.2 <2e-16 ***
>>> dB1 98.0060 0.2800 271.1900 350.0 <2e-16 ***
>>> ---
>>> Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1
>>>
>>> Correlation of Fixed Effects:
>>> (Intr)
>>> dB1 -0.987
>>>
>>> summary of the centered model m4c:
>>>> summary(m4c)
>>> Linear mixed model fit by maximum likelihood t-tests use Satterthwaite approximations to
>>> degrees of freedom [lmerMod]
>>> Formula: MCs ~ dB1c + (1 + dB1c | Species)
>>> Data: data
>>>
>>> AIC BIC logLik deviance df.resid
>>> 10720.4 10754.7 -5354.2 10708.4 2215
>>>
>>> Scaled residuals:
>>> Min 1Q Median 3Q Max
>>> -8.7305 -0.3491 0.0651 0.4508 6.6068
>>>
>>> Random effects:
>>> Groups Name Variance Std.Dev. Corr
>>> Species (Intercept) 1.109 1.053
>>> dB1c 1.763 1.328 0.94
>>> Residual 6.190 2.488
>>> Number of obs: 2221, groups: Species, 598
>>>
>>> Fixed effects:
>>> Estimate Std. Error df t value Pr(>|t|)
>>> (Intercept) 99.54109 0.08466 290.67000 1176 <2e-16 ***
>>> dB1c 38.78838 0.11081 271.18000 350 <2e-16 ***
>>> ---
>>> Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1
>>>
>>> Correlation of Fixed Effects:
>>> (Intr)
>>> dB1c 0.575
>>>
>>> and summary of the scaled model m4s:
>>>> summary(m4s)
>>> Linear mixed model fit by maximum likelihood t-tests use Satterthwaite approximations to
>>> degrees of freedom [lmerMod]
>>> Formula: MCs ~ dB1s + (1 + dB1s | Species)
>>> Data: data
>>>
>>> AIC BIC logLik deviance df.resid
>>> 10720.4 10754.7 -5354.2 10708.4 2215
>>>
>>> Scaled residuals:
>>> Min 1Q Median 3Q Max
>>> -8.7305 -0.3491 0.0651 0.4508 6.6068
>>>
>>> Random effects:
>>> Groups Name Variance Std.Dev. Corr
>>> Species (Intercept) 21.48 4.635
>>> dB1s 33.22 5.763 -1.00
>>> Residual 6.19 2.488
>>> Number of obs: 2221, groups: Species, 598
>>>
>>> Fixed effects:
>>> Estimate Std. Error df t value Pr(>|t|)
>>> (Intercept) -64.2618 0.4249 273.4100 -151.2 <2e-16 ***
>>> dB1s 168.3687 0.4810 271.1800 350.0 <2e-16 ***
>>> ---
>>> Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1
>>>
>>> Correlation of Fixed Effects:
>>> (Intr)
>>> dB1s -0.987
>>>
>>> Thanks in advance for your help!
>>> Best regards
>>> Jana
>>>
>>> -----Message d'origine-----
>>> De : Thierry Onkelinx [mailto:thierry.onkelinx at inbo.be] Envoyé :
>>> mardi
>>> 13 février 2018 10:26 À : Jana Dlouha <jana.dlouha at inra.fr> Cc :
>>> r-sig-mixed-models at r-project.org Objet : Re: [R-sig-ME] correlation
>>> between random effects
>>>
>>> Dear Jana,
>>>
>>> I assume that you uses the centered dB1c both in the random and the fixed effects? Another thing you can try is to scale dB1c. Using sensible units is often sufficient. Don't use large units (e.g.
>>> kilometers) when you are measuring small things (e.g. millimeters).
>>>
>>> You'll need to provide more information when you need more feedback.
>>> At least the summary of the data and the model.
>>>
>>> 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 at 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
>>> ////////////////////////////////////////////////////////////////////
>>> //
>>> /////////////////////
>>>
>>>
>>>
>>>
>>> 2018-02-13 10:00 GMT+01:00 Jana Dlouha <jana.dlouha at inra.fr>:
>>>> Hi all,
>>>>
>>>> I have a problem with a correlation between random effects. I have tested several models on my data:
>>>> m0<-lm(MCs~ dB1, data)
>>>> m1<- lmer(MCs~ dB1 + (1|Species), data, REML=FALSE)
>>>> m2 <- lmer(MCs~ dB1 + (-1+dB1|Species), data, REML=FALSE)
>>>> m3<- lmer(MCs~ dB1 + (1|Species)+(0+dB1|Species), data, REML=FALSE)
>>>> m4<- lmer(MCs ~ dB1 + (1+dB1 |Species), data,REML=FALSE)
>>>>
>>>> and when I compare the AIC criterion, the lowest one is for the model m4:
>>>> m0 m1 m2 m3 m4
>>>> 11086.51 10948.72 10828.75 10830.75 10720.43
>>>>
>>>> However, in the summary I see that there is a strong correlation between random effects and associated variances are huge:
>>>> Random effects:
>>>> Groups Name Variance Std.Dev. Corr
>>>> Species (Intercept) 21.48 4.635
>>>> dB1 11.25 3.355 -1.00
>>>> Residual 6.19 2.488
>>>> Number of obs: 2221, groups: Species, 598
>>>>
>>>> For m3, random effect associated with intercept has very low variance and residual variance is only a bit higher:
>>>> Random effects:
>>>> Groups Name Variance Std.Dev.
>>>> Species (Intercept) 3.419e-14 1.849e-07
>>>> Species.1 dB1 7.968e-01 8.927e-01
>>>> Residual 6.327e+00 2.515e+00
>>>> Number of obs: 2221, groups: Species, 598
>>>>
>>>> I am tempted to take into account only the randon effect associated with the slope however I don't know if i can do this considering that the AIC is not the lowest one for this model and how to justify it in my paper?
>>>> By the way, I don't really understand why the variances associated with the random effects change so much.
>>>> I have tried to center the regressor dB1 which removed the correlation between fixed effects and changed the sign of correlation but random effects remain strongly correlated and variances large:
>>>>
>>>> Random effects:
>>>> Groups Name Variance Std.Dev. Corr
>>>> Species (Intercept) 1.109 1.053
>>>> dB1c 11.255 3.355 0.94
>>>> Residual 6.190 2.488
>>>> Number of obs: 2221, groups: Species, 598
>>>>
>>>> Could you please give me some hint to solve my problem? Thanks a
>>>> lot in advance
>>>>
>>>> Jana
>>>>
>>>>
>>>>
>>>> [[alternative HTML version deleted]]
>>>>
>>>> _______________________________________________
>>>> R-sig-mixed-models at r-project.org mailing list
>>>> https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models
>> _______________________________________________
>> R-sig-mixed-models at r-project.org mailing list
>> https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models
More information about the R-sig-mixed-models
mailing list