I have performed a mixed model logistic regression using glmer and used afex::mixed to obtain the correct pvalues.
mixed(model_9, data = ReversalTaskData, method = "PB", test_intercept = T, family = "binomial")
but I receive this message
Formula (the first argument) converted to formula. Numerical variables NOT centered on 0: Stimulation, Block If in interactions, interpretation of lower order (e.g., main) effects difficult.
I was wondering if lmertest can be used as well obtain p-values in case of logistic (thus, non-linear) mixed model regression. do you have any other suggestion, otherwise?
Thank you in advance
Yes, lmerTest will do.
The summary methods will provide approximate z-Tests for the coefficients (I find it confusing that "t values" are reported. The standard errors are likely from a scaled Fisher information matrix, but the ratios of estimates and (scaled) standard errors ("t values") are taken as quantiles of the standard normal distribution. I have no Idea how well this works (at least in relatively small samples), but as long as you don't overinterpret approximate tests, it should be fine. The bigger source of possible errors is a bad assumption about the functional and stochastic model structure (is the functional relationship between predictors and the logit of the response reasonably linear? Is it reasonable to assume the same "success rate" [binomial parameter] in all subjects? If not, a beta-binomial model would be better).
Yes, lmerTest will do.
The summary methods will provide approximate z-Tests for the coefficients (I find it confusing that "t values" are reported. The standard errors are likely from a scaled Fisher information matrix, but the ratios of estimates and (scaled) standard errors ("t values") are taken as quantiles of the standard normal distribution. I have no Idea how well this works (at least in relatively small samples), but as long as you don't overinterpret approximate tests, it should be fine. The bigger source of possible errors is a bad assumption about the functional and stochastic model structure (is the functional relationship between predictors and the logit of the response reasonably linear? Is it reasonable to assume the same "success rate" [binomial parameter] in all subjects? If not, a beta-binomial model would be better).
- Badges
- Science topic
- Similar topics
- Mathematics
- Statistics
- Statistical Software
- R