Hello,
I am struggling to understand how I best use the na.action options in R's lme function for a hierarchical linear model.
I have a longitudinal data set with a two-level nested structure (occasion-level variables nested within person-level variables). My data is in long format. To test my hypotheses, I would like to perform multilevel modelling in R using lme(). Since my data contains missings, I get an error because lme() does not know what to do with these values. For my data, case wise or list wise deletion would be unfavourable.
There was a previous post on a similar topic regarding the lmer function
(https://www.researchgate.net/post/How_are_missing_data_handled_in_linear_mixed_effects_models). But I am still not sure if it is even possible to include observations with missings in the outcome variable in R's lme(). In Andy Fields' "Discovering Statistics with R", I read that na.omit and na.exclude both lead to case wise deletion and thus remove a complete observation if any of the variables used in the model contain NA.
How can I ensure to include observations with missings in my outcome variable using lme()?
Thanks in advance!
Leah
If you having missing outcome data then case wise deletion in lme will be equivalent to multiple imputation with the the MAR assumption (with the limitation that MAR depends only on variables in the model - the predictors and the outcome). If you have missing predictors then this biases the results and loses power relative to an optimal imputation strategy.
These days I'd normally switch to lme4 or brms rather than lme()from the nlme package in R. In particular brms it is particularly convenient to set up an imputation model using the mice package. However, it is Bayesian software so it is slightly more complex than lme4 or nlme. On the other hand with a little extra time you should be able to resolve any convergence issues more easily with brms.
There is a nice vignette on this for brms:
https://cran.r-project.org/web/packages/brms/vignettes/brms_missings.html
Hello Leah,
You can predict your missing observations using na.action=na.pass.
See pages 93-94 here:
https://cran.r-project.org/web/packages/lme4/lme4.pdf
You can also consider using multiple imputation to fill in your missing values before you fit the model. If you fit a linear mixed model, which I would suggest in your case, the missing data problem is dealt with the direct likelihood inference.
You can also predict your missing data using multiple imputation and fit a LMM with the complete dataset.
Good luck with your work!
If you having missing outcome data then case wise deletion in lme will be equivalent to multiple imputation with the the MAR assumption (with the limitation that MAR depends only on variables in the model - the predictors and the outcome). If you have missing predictors then this biases the results and loses power relative to an optimal imputation strategy.
These days I'd normally switch to lme4 or brms rather than lme()from the nlme package in R. In particular brms it is particularly convenient to set up an imputation model using the mice package. However, it is Bayesian software so it is slightly more complex than lme4 or nlme. On the other hand with a little extra time you should be able to resolve any convergence issues more easily with brms.
There is a nice vignette on this for brms:
https://cran.r-project.org/web/packages/brms/vignettes/brms_missings.html
- Badges
- Science topic
- Similar topics
- Mathematics
- Statistics
- Statistical Software
- R