Dear all,
I run a model in brms. The model converged without warnings, Rhats and plots look okay, but the estimates are super small.
This is the code I am running:
Fit_fullmodel_twoSLopes <- brm(formula = RT_raw ~ (Emotion*Condition*Gaze)*(QuestionnaireA+QuestionnaireB+HormoneA+HormoneB+ QuestionnaireA:HormoneA + QuestionnaireB:HormoneB)
+ (1 + Emotion*Condition*Gaze | Subject) + (1 + Emotion*Condition*Gaze | Model_ID),
data = dataset, family = 'lognormal',
warmup = 500, iter = 2000, chains = 4, cores = 4, save_all_pars = T,
control = list(adapt_delta = 0.99, max_treedepth = 12))
I used lognormal as a family, as the reaction time variable is highly skewed (also log transformations are usually done when analysing this task with ANOVAs). The categorical predictors are coded as sum-to-zero contrasts.
I am interested in how questionnaires and hormones influence task performance. The categorical predictors Emotion (happy, angry, neutral), Condition (push, pull) and Gaze (direct, averted) are within subject belonging to the task, the questionnaires and hormones are between subject.
So the interaction Emotion*Condition is included in all effects that are of interest making it such a big model. Thus, actually I am only interested ‘a few’ 3 to 4-way interactions. For example, when picking some of the effects that I am interested in:
Estimate Est.Error l-95% CI u-95% CI Eff.Sample Rhat
Intercept 6.63 0.01 6.60 6.66 423 1.01
Emotion1:Condition1 0.01 0.00 0.01 0.02 3655 1.00
Emotion2:Condition1 -0.02 0.00 -0.02 -0.01 4282 1.00
Emotion1:Condition1:QuestionnaireA -0.00 0.00 -0.01 0.00 3800 1.00
Emotion2:Condition1:QuestionnaireA 0.00 0.00 -0.00 0.01 3918 1.00
Emotion1:Condition1:QuestionnaireB -0.00 0.00 -0.00 0.00 4436 1.00
Emotion2:Condition1:QuestionnaireB -0.00 0.00 -0.00 0.00 4240 1.00
Emotion1:Condition1:HormoneA -0.00 0.00 -0.01 0.00 4470 1.00
Emotion2:Condition1:HormoneA 0.00 0.00 -0.00 0.01 4330 1.00
Emotion1:Condition1:HormoneB 0.00 0.00 -0.00 0.01 4930 1.00
Emotion2:Condition1:HormoneB -0.00 0.00 -0.01 0.00 4168 1.00
As the estimates are so small, I am not sure whether I can trust the (significant) CIs. Do you think these results are reliable?
Maybe I should use another family or another contrast coding?
I am happy about any help!
Thanks in advance,
Anna