Dear all,
I have a (slightly long) statistics / data science question concerning the applicability, implementation and concise reporting of a log-linear model.
I am also open for alternative suggestions.
Data:
We recorded activity of neurons in different brain regions, different cell types within these regions and using different stimuli. For each of these factors we get different numbers of neurons which we classify based on responsiveness criteria to belong to three groups.
So we have a multi-level contingency table with:
R1-3 = Response variable (3 level) - frequency
Several multilevel conditions:
stimulus type (S1-3), cell type (C1-2), brain region (B1-2).
My questions:
1. is a log-linear model the analysis of choice here - or am I completely off?
2. is my model selection and comparison reasonable?
3. is focusing on interaction only and disregarding main effects (which are trivial here based on random sampling over conditions) sound?
4. How should I report on model selection and interaction relevance in the most concise way possible (severe space limitations in publication)?
- - - - -
I chose to analyze this with a multifactorial log linear model and test for interactions (which may be incorrect - pls advice).
The independence model would be:
log(freq) = log(R) + log(S) + log(C) + log(B)
R:
mod0 |z|) = 0.01 **
R1:B1 - Pr(>|z|) = 0.1
Or should I report full statistics on individual model GOF (G2, df, coefficient estimates, individual p-values etc.)?
My questions (again)
1. is a log-linear model the analysis of choice here - or am I completely off?
2. is my model selection and comparison reasonable?
3. is focusing on interaction only and disregarding main effects (which are trivial here based on random sampling over conditions) sound?
4. How should I report on model selection and interaction relevance in the most concise way possible (severe space limitations in publication)?
Thanks!
T
I'm not an expert neither for log-linear models nor for R, but I found the question interesting in general, so here are a few ideas:
1. Although I'm, as mentioned, no expert, anova(mod1, mod0, test="Chi")) seems like a good idea. As a reader of your study with no deep understanding of log-linear models, I would have a hard time trying to figure out what the p-values for the interaction terms actually mean. - But I also wonder why you do not use linear regression to explain R by linear combinations of S, C and B? Probably I'm missing something here ...
2. I was wondering why you omitted some variables when testing for interactions? Why did you write mod1
- Badges
- Science topic