My question is about the basics of linear mixed model and it may be trivial for some of you. Basically, I followed from this tutorial http://www.bodowinter.com/tutorial/bw_LME_tutorial2.pdf on how to do linear mixed effect analyses in R. I am looking for an effect of a genotype on plant growth. I start with $3n$ seeds, $n$ of which have genotype A and other $2n$ have genotype B. I measure the height for each of these seeds by growing them in $n$ trays. Each tray has $1$ seed of genotype A and $2$ seeds of genotype B. So I have a dataframe in R with three columns. First column is plant genotype, second column is tray index and final is observed height. There are $3n$ observations. Following from Bodo Winter's tutorial, I define an alternative model in R as:

alt <- lmer(height ~ genotype + (1+genotype | tray), data=data, REML= FALSE)

I also define a null model:

nul <- lmer(height ~ genotype + (1+genotype | tray), data=data, REML= FALSE)

Then, I compare the results from these two models by doing

anova(alt, nul)

I have been interested in following two quantities:

c(coef(summary(alt))[2, "Estimate"], anova(alt, nul)$"Pr(>Chisq)"[2])

My questions are:

  1. What is the underlying matrix formulation that the alt and nul models are solving? I am good with linear algebra but I cannot find any adequately brief or self-contained description just answering my question about my models. All the theory I read is too general and I'd have to read through a lot of stuff to understand it. While I plan to do it, currently I need to finish a project and will feel very comfortable if I could get some insight in what is happening.

  2. How is anova comparing the two models to get the p value? Again I could find descriptions on how anova makes use of F-values to estimate significance, I could not relate that description to comparison of two models.

Any explanation will be highly appreciated.


I guess all your questions can be answered by this book here: http://www.springer.com/gp/book/9780387874579, chapter 5 particularly.

Question 2. Normally, you fit mixed-effects model with restricted maximum likelihood (REML), description of which is nicely provided in the above-mentioned book. Function anova() performs a likelihood ratio test and that is why you need to have REML=FALSE argument in you model. However, I can see that your nul and alt model are exactly the same models. To test a statistical significance of your fixed-effect you need to exclude it from the nul model and then run anova() function:

alt <- lmer(height ~ genotype + (1+genotype | tray), data=data, REML= FALSE)
nul <- lmer(height ~ (1+genotype | tray), data=data, REML= FALSE)

The obtained p-value would show the statistical significance of your fixed-effect.

Question 1. Do not quite understand the question, the matrix formulation for mixed-effects models is nicely described in the above-mentioned book.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.