I suspect that if you look at the by-participant variance, you will have seen it go down. The fixed effects term for N is by-participant, so when it's left out from the model, the variation that it explains would tend to be associated with the by-participant intercept (for the main effect) or the by-participant slope (for the interaction). One and a half minor terminological points> 1. The syntax you used and the models you describe are valid for lme4 but not nlme, so did you mean lmer() instead of lme()? 2. I would avoid the L1 vs L2 distinction here. I find the terminology confusing and I would describe your design as 'crossed' and not 'nested' anyway because you could just as well think of participants as being repetitions of each item as the other way around. There's no need to think of these things as strictly hierarchical, which is why some (including me) prefer the terminology "multilevel" instead "hierarchical" model. Best, Phillip
On 18/04/2019 16:59, Thorsten Aichele wrote:
Dear List,
I am trying to specify the optimal random effect structure and I am not
sure, if there's a problem with my understanding of the random effects
structure, or with my data, or with none of these two.
Design:
- Two Levels, Repeated measures (L2 = 140 Participants)
- Measure of Personality trait 'N' on L2
- One experimental factor 'Condition' (on L1)
- The control condition contained 12 Items. The experimental
condition contained another 12 items (Item 1-12 = control group, item 13-24=
experimental group)
- Each participant answered all the items and all conditions. (Each
item was only answered once)
- The experimental comparison was: funny (experimental condition)
vs. not funny (control condition)
- I have two random factors (participants on Level 2 and items on
level 1)
- Items are nested under condition (as the items in both conditions
were not the same)
Now I want to look for a Cross-Level Interaction of the L2 variable
(personality trait 'N') with the L1 variable 'Condition' (my main
hypothesis)
I made the following Random effect structure under the assumption "include
every possible random slope"
lme(DV ~ 1 + (1 + Condition|participants) +
(1|Item)
[I excluded random slope for N on item, as the model did not converge]
Now I tried to compare this model with a
model with fixed effects + interaction for 'condition' and 'N'
lme(DV ~ 1 + Condition*N + (1 +
Condition|participants) + (1|Item)
-Fixed effects part showed
nonsignificant effect for 'condition' and 'N'
-Fixed effects part showed a significant interaction effect Condition:N
-Fixed effect for intercept was also significant
Both Models share identical residual variance (1.7450). I have no idea how
this could be possible. The interaction effect is rather small (-.0247), but
I doubt, that an interaction effect could become significant without
explaining any variance.
I would be thankful if anyone could help me with this problem
Best,
Thorsten Aichele
[[alternative HTML version deleted]]
_______________________________________________ R-sig-mixed-models at r-project.org mailing list https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models