Dear friends and colleagues,
my team and I are currently analyzing data from an experiment with small groups of learners. Here, we want to conduct a multilevel analysis (hierarchical linear modeling) because we assume that collaborating with others affects the knowledge of the individual group members.
We hope that you may help us model the following situation from our experiment, using a multilevel analysis.
Our question: How do we model the levels, especially: where do we enter the experimental condition? (As a predictor for the outcome variable, or as Level 3?)
Design of the study
- Participants collaborate in small groups (groups of 3 members) to solve a collaboration task
- each small group is randomly assigned to one of three experimental conditions which consist of different types of pedagogical support:
- Before the collaboration, the participants answer a pretest (individually) that assesses their knowledge about collaboration. After their collaboration, the participants answer a respective posttest (individually).
Research question: How does the experimental condition support measure influence affect participants' knowledge about collaboration?
Data set:
- 150 Participants (Level 1) in
- 50 small groups (Level 2)
The experimental conditions are a fixed (predictor) variable, not something randomly sampled. This varies at the group level. You also have the pre score. So condition_j, pre_ij, post_ij, and group_j. In R, make sure that you have defined condition_j as a factor with whatever contrasts best suit your research questions (and decide how you want to handle the pre_ij variable since that depends on a few things.
You cannot just employ multilevel modellings because you liked it. Why you specifically wanted it? Do you have a scientific justification with its evidences? Given your claim that you small groups (clsuters) in an experimental study, it is less likely that multilevel modelling techniques fit for such data. You need to first test the null hypothesis to whether employ the multilevel modelling or not. Thanks.
If you are data are in long form (that is with replicates), you simply include it as a factor (as Daniel Wright says, suitably coded, eg the control as the referent category) in the fixed part of the model. You need a multilevel analysis as you are making repeated measures of the same group which is likely to produce dependence. You only have 50 groups not 150 observations at level 2.
I would not have a level 3 as the experimental conditions are not a sample of all experimental conditions you could have analysed. That is, for me, they are fixed effects not random effects - alternative views are available!
You may want to look at the chapter on higher level variables in this:
Book Developing multilevel models for analysing contextuality, he...
I'll just add that because your Group variable has k=3 levels, you can, if you wish, use Fisher's LSD (or an analog). Contrary to what many people think, Fisher's LSD does control the FW alpha at the per-contrast alpha when there are 3 groups. See this old thread, for example:
https://www.researchgate.net/post/When_can_we_use_Fishers_Least_Significant_Difference_LSD_test
HTH.
In a multilevel analysis when analyzing data from an experiment with small groups, the correct Level 2 predictors are the group-level predictors that are not part of the intervention. The Level 3 predictors are the individual-level predictors that are not part of the intervention.
- Badges
- Science method
- Similar topics
- Mathematics
- Statistics
- Statistical Techniques
- Meta-Analysis