nlme (WIP)
- Specify
- Model fit
- Statistical testing
- Effects
Useful links:
- https://quantdev.ssri.psu.edu/sites/qdev/files/ILD_Ch06_2017_MLMwithHeterogeneousVariance.html
Specify
| What | How | Details |
| Random intercept model with subject as grouping factor | lme(fixed = y ~ x, random = ~ 1 | subject) | |
| With random intercept | random = ~ 1 | subject | |
| With random intercept | random = list(subject = pdSymm(form = ~ 1)) | |
| With random intercept, different between-subject SD per group | random = list(subject = pdDiag(form = ~ diagnosis)) | |
| With factor-dependent within-subject error | weights = varIdent(form = ~ 1 | diagnosis) | |
| With AR-1 correlation structure | correlation = corAR1(form = ~ 1 | subject) | |
Extract info from a fitted model.
| What | How | Details |
| Fixed effects coefficients | fixef(model) | |
| Fixed effects coefficient intervals | intervals(model) | |
| Variance and SD summary | VarCorr(model) | |
| Variance-covariance matrix | vcov(model) | |
| Within-subject variance | sigma(model)^2 | |
| Within-subject SD | sigma(model) | |
Within-subject SD for group $GROUP (when using weights=varIdent) | sigma(model) * coef(model$modelStruct$varStruct, unconstrained = FALSE, allCoef = TRUE)['$GROUP'] | |
| Between-subject variance | VarCorr(model)['Residual', 'Variance'] | |
| Between-subject SD | VarCorr(model)['Residual', 'StdDev'] | |
Model fit
| What | How | Details |
| Log-likelihood | logLik(model) | |
| AIC | AIC(model) | |
| BIC | BIC(model) | |
| ICC | performance::icc(model) | Proportion of the total variance explained by the between-cluster variance. Higher is better. 0.1 is considered a meaningful threshold. |
| Marginal pseudo R-squared | r.squaredGLMM(model)[1, 'R2m'] | |
| Conditional pseudo R-squared | r.squaredGLMM(model)[1, 'R2c'] | |
| Plot residuals against predicted | plot(model) | |
| Plot random effects | plot(ranef(model)) | |
| Q-Q plot | car::qqPlot(residuals(model)) | |
Statistical testing
| What | How | Details |
| Coefficient significance | car::Anova(model) | |
| Test for nested model improvement (likelihood ratio test) | performance::test_lrt(model0, model1) | |
| Test for nested model improvement (likelihood ratio test) | anova(model0, model1) | |
Effects
| What | How | Details |
| Cohen’s d (effect size) for main effects | EMAtools::lme.dscore(model, type = 'nlme') | ? |
| Cohen’s f for main effects | effectsize::cohens_f(model) | d = 2 * f |
| Cohen’s d (effect size) for group difference | emg = emmeans::emmeans(model, ~ diagnosis)
emmeans::eff_size(emg, sigma = sigma(model), edf = data.frame(pairs(emg))$df[1]) | ? |