I am relatively inexperienced with mixed effect models and trying to build a model to fit my outcome of interest. I have read and followed along with chapter 4 of Singer & Willet (2003), as well as helpful accompanying R code from UCLA and someone's blog.

I have created a model for an unconditional means model, an unconditional growth model, and four additional models (m1-m4). I haven't found any significant interaction effects in my data, so each of these are added predictors without interaction terms.

## Making sure I'm using the exact same dataset for all models and comparisons
long.STAI.df <- long.pc.df %>% filter(., !is.na(STAI.tot) & !is.na(income) & !is.na(hours))

STAI.uncmeans <- lmer(STAI.tot ~ 1 + (1|GS_ID), long.STAI.df)
STAI.uncgrowth <- lmer(STAI.tot ~ time + (1 | GS_ID), long.STAI.df)

STAI.m1 <- lmer(STAI.tot ~ time*hours + (1|GS_ID), long.STAI.df)

STAI.m2 <- lmer(STAI.tot ~ time + hours + (1|GS_ID), long.STAI.df)

STAI.m3 <- lmer(STAI.tot ~ time + hours + income + (1|GS_ID), long.STAI.df)

STAI.m4a <- lmer(STAI.tot ~ time + hours + income + C.Access.tot + (1|GS_ID), long.STAI.df)

anova(STAI.uncgrowth, STAI.m1)
anova(STAI.m1, STAI.m2)
anova(STAI.m2, STAI.m3)
anova(STAI.m1, STAI.m4a)
anova(STAI.m2, STAI.m4a)

Can I compare model 4 to model 1 using deviance statistics from anova(STAI.m1)? Or do I have to go in a step-wise fashion, stopping if the difference in deviance statistics from model 1 to model 2 to is not significant? I ask because m1 vs. m2 is not significant.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.