Missing values in lmer vs. HLM
By the way, most of us don't know the acronym FIML. I have a suspicion that it is one of the many "maximum likelihood" estimators defined in the multilevel modeling literature. To a statistician these expressions are nonsensical. Once you define the probability model there is only one possible definition of likelihood and hence only one criterion for the maximum likelihood estimators to optimize. Creating a different criterion and saying that the optimizers of this criterion are the "<whatever> maximum likelihood" estimators is false advertising. Having said all this I will admit that the original sin, the "REML" criterion, was committed by statisticians. In retrospect I wish that we had not incorporated that criterion into the nlme and lme4 packages but, at the time we wrote them, our work would have been dismissed as wrong if our answers did not agree with SAS PROC MIXED, etc. So we opted for bug-for-bug compatibility with existing software.
On Sat, Jul 4, 2015 at 11:09 AM Douglas Bates <bates at stat.wisc.edu> wrote:
I think we would need to know more about the structure of the data and the models that you wish to fit to it before we could be of any assistance. To be honest, your question doesn't make sense in the context of lmer. The data for lmer must be in the "long form". That is, each observation corresponds to a row in the data frame. If one subject has 5 observations there will be 5 rows for that subject. If another has only two observations there will be two rows. To me you are describing unbalanced data, not missing data. In most cases it is more confusing than illuminating to think of data in the "wide form", with one row for each subject and multiple columns for the observations, when working with R. There is no difficulty with working with unbalanced data in lmer. On Sat, Jul 4, 2015 at 10:42 AM Carpenter, Tom <tcarpenter at spu.edu> wrote:
All,
I have a paper in which we are using a within-person model using
multi-level modeling. I ran the models in lmer in R, although we had a
substantial portion of people for whom at least one observation is still
missing. My understanding is that the default is to drop that person
entirely (e.g., na.action=na.omit) ?.is that correct? My understanding was
that the HLM software (e.g., by SSI) and most other multi-level modeling
programs can still run the models based on the remaining observations
(e.g., you may have 4 out of 5 observations per person and still be able to
run the model).
I would love to know if it is possible to do that in lmer or if some
solution is present. For example, is it possible to use FIML in lmer?
Advice for handling this situation would be appreciated, as I?m new to lmer!
Best,
Tom Carpenter, Ph.D.
Assistant Professor, Psychology
Seattle Pacific University
3307 3rd Ave W. Suite 107,
Seattle, WA, 98119
tcarpenter at spu.edu<mailto:tcarpenter at spu.edu>
Office: (206) 281-2916
Fax: (206) 281-2695
[[alternative HTML version deleted]]
_______________________________________________ R-sig-mixed-models at r-project.org mailing list https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models