Hi everyone,
I'm trying to wrap my head around multilevel (or mixed-effect) models, but not quite succeeding.
I have the following data from a pilot study: 221 ratings (mean of several items on 1-5 Likert scale) for trustworthiness from 13 people on 17 internet profiles (fully crossed). I'm trying to model the following:
- profile as a random intercept (these are my items)
- rater as a random intercept (these are my subjects)
- log(word_count) as my IV that has a random slope for each rater (each person values "wordiness" differently)
There is no measurement error for word_count (it is simply calculated on the basis of the profile text).
So, I have the following in R (lme4):
Did I do this correctly? What I get is the following output:Code:model <- lmer(trust ~ log(word_count) + (1+log(word_count)|rater) + (1|profile), data=dt)
What I don't understand: why there is a correlation of 1 between the intercepts of rater and the slope of log(word_count)? Does it have something to do with the fact that I have a fully crossed design?Code:Random effects: Groups Name Variance Std.Dev. Corr profile (Intercept) 0.1298242 0.36031 rater (Intercept) 0.0009746 0.03122 log(word_count) 0.0033183 0.05760 1.00 Residual 0.1388396 0.37261 Number of obs: 221, groups: profile, 17; rater, 13 Fixed effects: Estimate Std. Error t value (Intercept) 2.7062 0.4800 5.637 log(word_count) 0.1697 0.1016 1.670 Correlation of Fixed Effects: (Intr) lg(wrd_cnt) -0.967
As I wrote in a previous thread, this a pilot study. Even though I know a multilevel approach is better, there are a number of reasons to simply aggregate the data to profile level (simplicity, methods more familiar in field, being able to easily calculate an R^2 and p, etc.). With the random effect of rater being the size that it is, would you say it makes sense to simply aggregate? (Though word count is expected to be one of the larger effects).
Last edited by Junes; 11-07-2016 at 04:27 PM.
Tweet |