I am analysing repeated blood pressure measurements using linear mixed-effects models fitted with nlme::lme. All models use the same participants, random-intercept structure and continuous autoregressive residual correlation: random = ~ 1 | id correlation = corCAR1(form = ~ age | id) The exposure group and sex are binary. I fitted the following models: mod_lin I then used: lmtest::lrtest(mod_par, mod_gam) lmtest::lrtest(mod_lin, mod_gam) lmtest::lrtest(mod_gam, mod_gam2) These comparisons returned very small p-values without warnings. However, the documentation for anova.lme states that likelihood comparisons are not meaningful for REML-fitted models with different fixed effects. Looking at the source code of lmtest::lrtest, it appears to extract logLik() from the supplied objects and does not automatically refit lme models using maximum likelihood. My questions are: Are the p-values returned by lmtest::lrtest() invalid in this situation because the models have different fixed-effects structures but were fitted by REML? Is the appropriate workflow to create ML versions of all models before the comparisons? mod_lin_ml After model comparison, is it appropriate to retain or refit the selected final model using REML for coefficient estimation, predictions and confidence intervals? All models use the same observations and the same random-effects and residual-correlation structures; only the fixed-effects formulas differ.

Full article content could not be extracted automatically. Read the original below.