0
$\begingroup$

everyone! So, I've been trying to fit a LMER for my data, which I fully described in this post . This post, therefore, is a follow-up from the previous one. There's something intriguing me, I don't get any errors when I run this ANOVA:

aov(SCORE ~ LANGUAGE * YEAR + Error(SUBJECT/LANGUAGE * YEAR), data = data)

but I'm getting the a non-converge error (model1) or a boundary error (specifically for the language variable) model2:

model1 <- lmer(SCORE ~ (YEAR * LANGUAGE | SUBJECT), data = data
Error: number of observations (=80) <= number of random effects (=80) for term (YEAR * LANGUAGE | SUBJECT); the random-effects parameters and the residual variance (or scale parameter) are probably unidentifiable

or

model2 <- lmer(SCORE ~ LANGUAGE * YEAR + (YEAR | SUBJECT) + (LANGUAGE|SUBJECT), data = data
boundary (singular) fit: see help('isSingular')

As far as I'm concerned, this lmer should resseamble the repeated measures ANOVA I've ran before, so why do I get an error for the model, but not for the Anova?

Any thoughts on it would be much appreciated. Thanks in advance!

Improve this question
$\endgroup$
4
  • $\begingroup$ Not sure about the first error, but my guess is you get singular fit from lmer because you have too few observations (or other data-related issues) to estimate the random slope, which is usually the most complex part to estimate. If I understood correctly, you only had 2 observations per participant per year - that is usually too few for random slope estimation. Repeated-measures ANOVA does not estimate the year-specific slope, so it works out. $\endgroup$
    – Sointu
    Commented Aug 26, 2022 at 12:33
  • $\begingroup$ @Sointu , thank you for your answer, so this ANOVA argument: Error(SUBJECT/LANGUAGE * YEAR) is equal to lmer(SCORE ~ Language*Year + (1|SUBJECT), data = data) ? I thought that the corresponding model to that Anova was one from the ones that I've ran before, but now I'm wondering this... $\endgroup$
    – Larissa Cury
    Commented Aug 26, 2022 at 12:38
  • $\begingroup$ To the best of my understanding, not quite as I don't think ANOVA calculates subject-specific intercepts as lmer does with the (1|subject), it just handles the correlated errors (?), but I'd say from a practical point of view a random intercept lmer and repeated measures anova usually give roughly similar results. However I might be wrong as I rarely use ANOVAs, I only wanted to comment on the lmer singular fit error as I'm familiar with it myself! I hope you'll get a better answer soon! $\endgroup$
    – Sointu
    Commented Aug 26, 2022 at 13:10
  • $\begingroup$ @Sointu , no problem! thank you very much for your insights! 😀 $\endgroup$
    – Larissa Cury
    Commented Aug 26, 2022 at 13:20

0

Reset to default

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.