[R-sig-ME] Quadratic term in linear model and model over-parameterization
thierry.onkelinx at inbo.be
Wed Mar 15 10:56:23 CET 2017
This looks like a linear model and not a linear mixed model. This
mailing list is dedicated to mixed models.
I strongly recommend to find a local statistician. All your models
seem way to complex given the data. As a rule of thumb you need at
least 10 observations for each parameter in your model. Adding the
quadratic terms increases the number of parameters and only make the
ir. Thierry Onkelinx
Instituut voor natuur- en bosonderzoek / Research Institute for Nature
team Biometrie & Kwaliteitszorg / team Biometrics & Quality Assurance
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
2017-03-15 10:46 GMT+01:00 <f_fran03 op uni-muenster.de>:
> Dear all,
> I’m new to this mailing list and really hope that somebody here can help me with the following issue:
> I calculated the following linear models on a BoxCox transformed response variable with 382 data points:
> Model 1: Y~x+a+b+c+d+e+(a*b)+(a*c)+ (a*d)+…+(a*b*c)+(a*b*d)+(a*b*e)+…
> a: 'Experimental Temperature' (Temp1, Temp2)
> b: 'Host Population' (PopX, PopY)
> c: 'Parasite Population' (PopX, PopY)
> d: 'Host Gender' (male, female)
> Additionally, I included the continuous predictor variable 'Parasite Weight' (e) and all possible 2-way (10 interactions) and 3-way (10 interactions) interactions into the model.
> In model 2 I replaced the two main effects 'Host Population' and 'Parasite Population' with one variable ('Sympatry/Allopatry') that combines the two effects. Apart from this, model 2 (six 2-way interactions and four 3-way interactions) was identical to model 1.
> I am interested now in all interactions that include the continuous predictor variable 'Parasite Weight'. I got such a significant interaction ('Experimental Temperature x Parasite Population x Parasite Weight', p = 0.010) from model 1.
> We sent a manuscript containing these two models to a journal for review and got it back now with a comment from a reviewer who suggested that we look for non-linear relationships involving 'Parasite Weight'.
> Thus, I calculated model 1.2 which corresponds to model 1 but additionally added the quadratic term of 'Parasite Weight' ('Parasite Weight^2') and the respective interactions (in total 14 x 2-way interactions and 16 x 3-way interactions). I did the same for model 2, which resulted in model 2.2 with nine 2-way interactions and seven 3-way interactions.
> The significant interaction I found with model 1 was not significant anymore with model 1.2 and in model 2.2 two interactions became significant ('Host Gender x Sympatry/Allopatry x Parasite Weight', p = 0.038 and 'Host Gender x Sympatry/Allopatry x Parasite Weight^2', p = 0.044) that were not significant in model 2.
> Here are my questions:
> 1. Why is it that including the quadratic term removes some significant effects while adding others?
> 2. What does it mean when both an interaction including the linear term and the same interaction including the quadratic term become significant? Does this suggest a non-linear relationship or both a linear and a non-linear relationship?
> 3. Could it be that the disappearance of the interaction that was significant in model 1, is caused by an over-parameterization of model 1.2 and how can I prove this (with all the models we have the potential problem of many interactions and main effects)?
> 4. Are there any general arguments for when to include a quadratic term into a model and when quadratic terms should be avoided?
> 5. Which model can I trust?
> Thank you very much in advance for any advice you can give me,
> R-sig-mixed-models op r-project.org mailing list
More information about the R-sig-mixed-models