For an experiment, I have a large dataset of tweets from 10 different subjects. The dataset contains around 300 tweets per subject. I'd like to find out whether the subjects differ from each other in the way they use specific linguistic features, e.g. the use of nouns.

Therefore, for every tweet I have counted the occurrences of nouns and I have computed proportions by dividing each counts by the number of words in a tweet. For example: a tweet contains 10 words and 3 nouns, then the proportion of noun in this tweet is .3. For every subject, I have computed the average of the proportion of nouns.

Now, I would like to find out whether the proportions differ significantly among subjects en at this point I got stuck.

I already searched the internet and SO for an answer, but were unable to find out what test I should use. The variance of the data is not homogeneous and the data also contains lots of zeroes. Is an ANOVA appropriate? If so, what post hoc test is recommended?

Any suggestions are much appreciated.


1 Answer 1


For each subject, you have a count of number of nouns and non-nouns. That gives you a 10x2 contingency table, so you could use the chi-squared test.

One problem with this approach is that it does not take into account variability between different tweets from the same person. So to try to take that into account, for each tweet count nouns and non-nouns, and model this as a binomial mixed model. Using R lmenotation, one model can be something like

mod0 <- lme4::glmer(count ~ subject + (1 | tweet), family=binomial,

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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