1
$\begingroup$

I have the following dataset, link here.

I have three variables:

  1. Dependent Variable 1: Grammatical Category
    this is a nominal variable, and it has four categories: "verb", "adjective", "noun", "adverb".
  2. Dependent variable 2: Sentence Type
    This is a nominal variable and it has three categories: "short", "medium", "long".
  3. Third variable : Language
    This is independent variable and is nominal: "English", "Spanish".

The first column in the data sheet is the participant ID. The data is collected during individual interviews and the number of entries (observations) per participant is not the same, because each participant could have produced a different number of sentences during the interview.

My goal is to compare the combinations of the dependent variables in two languages. For example:
Is "verb-short" frequency significantly higher in English than in Spanish? and the same for other combinations: verb-medium, verb-long, adjective-short, adjective-long...

I did a general loglinear analysis, but there are a handful of expected counts which are below 5, therefore violating the test assumptions, and couple of adjusted residuals which are higher than +/-1.96.

I created a table in which I summarized the counts of the combinations per participant (111 participants in total, so 111 rows in my table) and 12 columns which corresponds to the 12 combinations of the two nominal variables. I tried a one-way anova and the assumptions of normal distribution is not met, there are a lot of zeros.

Question:

what is the best approach to analyse this data and what test should I use to compare the combinations between the two languages?

or alternatively question 2:

how can I compare only one category from variable 1 against other three categories in variable 2 within one language only? for example, in English verbs (one category from variable 1), compare "short" and "medium" and "long" with each other.

$\endgroup$
0
$\begingroup$

I am not sure exactly what you mean by "general log linear analysis" and Google did not give useful results. A log linear analysis is inappropriate here because: 1) your data are not independent and 2) You have dependent variables, log linear models don't treat variables as dependent and independent.

From your notes, you don't really have two dependent variables, you have one, with 12 levels (verb-short, verb-medium etc).

You could then use a nonlinear multilevel model, with person as an IV at one level and language at the other.

$\endgroup$
  • $\begingroup$ thanks for answer, what R package can do this test? could you please give a short explanation on this. thanks $\endgroup$ – cplus Mar 19 '17 at 13:43
  • $\begingroup$ Look into nlme in R. These are complex models and a short explanation really doesn't help much, but they are ways of removing the assumption that the errors are independent (this assumption is made by regression, log linear models, and many other models). $\endgroup$ – Peter Flom Mar 19 '17 at 14:03

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.