Model comparisons
Dear Thierry, Thank you very much for your help. The effect of "D" is not really strong but the interaction between "D" and another factor is still significant in some models. I will include it as a fixed factor and briefly report the result of the effect of "D". As for the age effect, both log and quadratic functions make sense. However, in terms of the interaction between age and other factors (e.g., testing block), the quadratic function reflects the results more than the logarithmic function. In addition, the results of the model with age in polynomial (poly(age,2)) function make more sense. So I will choose the model with quadratic function. Thank you very much again. Best wishes, Yasu On Mon, 6 Oct 2014 08:41:19 +0000
"ONKELINX, Thierry" <Thierry.ONKELINX at inbo.be> wrote:
Dear Yasu, IMHO, a model must make sense. Assume you know a priori that a certain variable "D" has an important effect on the response, but your focus is on other variables. Then you have two options: A) use a design that tests everything at the same level of "D". Then the effect of "D" is constant in the design and can be ignored. B) Measure "D" and add it to the model. You will have to report the effect of "D", although not as elaborate as the variables you're focusing on. If you don't find any literature on the relation with age, then try to thing what kind of relation could make sense. Is it monotone? Is there an optimum? Has adding 1 unit of age the same effect regardless the age? ... Bottom-line: rather choose a meaningful function, than the function which gives the lowest AIC. Best regards, ir. Thierry Onkelinx Instituut voor natuur- en bosonderzoek / Research Institute for Nature and Forest team Biometrie & Kwaliteitszorg / team Biometrics & Quality Assurance Kliniekstraat 25 1070 Anderlecht Belgium + 32 2 525 02 51 + 32 54 43 61 85 Thierry.Onkelinx at inbo.be 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 -----Oorspronkelijk bericht----- Van: r-sig-mixed-models-bounces at r-project.org [mailto:r-sig-mixed-models-bounces at r-project.org] Namens Yasuaki SHINOHARA Verzonden: zaterdag 4 oktober 2014 16:54 Aan: r-sig-mixed-models at r-project.org Onderwerp: Re: [R-sig-ME] Model comparisons Dear Thierry, Thank you very much for your help. Actually, the first and second questions were not talking about the same model. For the first question, the fixed factor of age was just an example, and the results below are not related to the first question. I was not sure whether I should include a fixed factor of which I am not reporting the results in a paper, if the model with the factor fits significantly better than the model without the factor. Do you mean that I should include it, if it has a significant effect, although I am not reporting the result of the factor? For the second question, the aim of my research is to investigate the age effects. I wanted to test how training works for speaker's production, and whether there is an age effect on it. So the factor A is "Group" (training group vs. control group), and the factor B is "Testing Block" (pre-training vs. post-training) I am trying to find some literature about how the age related to the improvement made by training, but it is really hard to find. I am not sure how I can send a figure of the relationship between age and the response through this emailing list. Sorry. Anyway, thank you very much for your help. Best wishes, Yasu about whether I should include a fixed factor in a logistic mixed effects model or not, On Fri, 3 Oct 2014 08:20:31 +0000 "ONKELINX, Thierry" <Thierry.ONKELINX at inbo.be> wrote:
Dear Yasu, It looks like your response is age dependent. Therefore you should include age into the model, so the model can take the age effect into account. I prefer to take a look at the functional relationship between age and the response (in the logit scale). There is probabily some literature on the effect of age on the response. That will give you more information on which function to choose: log(age) or poly(age, 2). Best regards, ir. Thierry Onkelinx Instituut voor natuur- en bosonderzoek / Research Institute for Nature and Forest team Biometrie & Kwaliteitszorg / team Biometrics & Quality Assurance Kliniekstraat 25 1070 Anderlecht Belgium + 32 2 525 02 51 + 32 54 43 61 85 Thierry.Onkelinx at inbo.be 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
________________________________________
Van: r-sig-mixed-models-bounces at r-project.org
[r-sig-mixed-models-bounces at r-project.org] namens Yasuaki SHINOHARA
[y.shinohara at aoni.waseda.jp]
Verzonden: vrijdag 3 oktober 2014 6:22
Aan: r-sig-mixed-models at r-project.org
Onderwerp: [R-sig-ME] Model comparisons
Dear all,
Could I ask a very basic question about glmer?
I am wondering how important using the best-fitting model is.
(1)
Please imagine I have three fixed factors "A", "B" and "C" in a
logistic mixed effects model.
I want to test these main effects and their all possible
interactions.
However, I can include another factor "D" (e.g., age) in which I am
not interested. If I include the fixed factor "D" in the model, the
model fits significantly better than the model without the factor
"D".
I know I should use the best-fitting model, and report all the
results
including the factor "D", although the results are slightly different
from the model which does not include the factor "D".
However, I also think that including unnecessary factors would
distract readers from the main point, so it may be good to analyze
data without the factor "D".
Could I ask your opinions?
(2)
Also, I do not understand why the results are so different, if I
change the relation in one of the factors.
For example, the model including the fixed factors of "A","B","C" and
"log(age)" is significantly better than another model including the
fixed factors of "A","B","C" and "poly(age,2)".
This difference (log(age) vs. poly(age,2)) affects the results of
other factors of "A", "B" and "C" as below.
Could you please explain why?
In terms of AIC value, MODEL1 is better. However, the results of
MODEL1 do not look correct.
Why is it?
MODEL1<-glmer(binomial_response~A*B*log(age)+(1|X)+(1+B|Y)+(1+B|Z),
family=binomial,
data=ALLDATA,control=glmerControl(optimizer="bobyqa"))
MODEL2<-glmer(binomial_response~A*B*poly(age,2)+(1|X)+(1+B|Y)+(1+B|Z),
family=binomial,
data=ALLDATA,control=glmerControl(optimizer="bobyqa"))
Anova(MODEL1,type=3)
Analysis of Deviance Table (Type III Wald chisquare tests)
Response: prod_corr
Chisq Df Pr(>Chisq)
(Intercept) 0.8155 1 0.366503
A 0.0059 1 0.938896
B 0.7490 1 0.386791
log(age) 8.6887 1 0.003202 **
A:B 0.0044 1 0.947053
A:log(age) 0.2471 1 0.619110
B:log(age) 2.5704 1 0.108881
A:B:log(age) 0.4881 1 0.484767
---
Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
Anova(MODEL2, type=3)
Analysis of Deviance Table (Type III Wald chisquare tests)
Response: prod_corr
Chisq Df Pr(>Chisq)
(Intercept) 41.2696 1 1.326e-10 ***
A 6.4384 1 0.0111677 *
B 13.0042 1 0.0003108 ***
poly(age, 2) 14.2490 2 0.0008051 ***
A:B 14.2547 1 0.0001597 ***
A:poly(age, 2) 1.1039 2 0.5758358
B:poly(age, 2) 3.2066 2 0.2012318
A:B:poly(age, 2) 0.3203 2 0.8520201
---
Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
Best wishes,
Yasu
_______________________________________________
R-sig-mixed-models at r-project.org mailing list
https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models
* * * * * * * * * * * * * D I S C L A I M E R * * * * * * * * * * *
* *
Dit bericht en eventuele bijlagen geven enkel de visie van de
schrijver weer en binden het INBO onder geen enkel beding, zolang dit
bericht niet bevestigd is door een geldig ondertekend document.
The views expressed in this message and any annex are purely those
of
the writer and may not be regarded as stating an official position of
INBO, as long as the message is not confirmed by a duly signed
document.
************************************ Yasuaki SHINOHARA, Ph.D. Assistant Professor Center for English Language Education (CELESE) Waseda University Faculty of Science and Engineering 3-4-1 Okubo, Shinjuku-ku, 169-8555, Tokyo JAPAN email: y.shinohara at aoni.waseda.jp
_______________________________________________ R-sig-mixed-models at r-project.org mailing list https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models * * * * * * * * * * * * * D I S C L A I M E R * * * * * * * * * * * * * Dit bericht en eventuele bijlagen geven enkel de visie van de schrijver weer en binden het INBO onder geen enkel beding, zolang dit bericht niet bevestigd is door een geldig ondertekend document. The views expressed in this message and any annex are purely those of the writer and may not be regarded as stating an official position of INBO, as long as the message is not confirmed by a duly signed document.
************************************ Yasuaki SHINOHARA, Ph.D. Assistant Professor Center for English Language Education (CELESE) Waseda University Faculty of Science and Engineering 3-4-1 Okubo, Shinjuku-ku, 169-8555, Tokyo JAPAN email: y.shinohara at aoni.waseda.jp