Timeline for Should i exclude random effects from a model if the random effect itself has missing data?
Current License: CC BY-SA 4.0
8 events
| when toggle format | what | by | license | comment | |
|---|---|---|---|---|---|
| Oct 21 at 10:55 | answer | added | BenP | timeline score: 1 | |
| Oct 7 at 18:42 | comment | added | John Madden | Excluding the random effect would lead to a perniciously misspecified error structure. I would be quite tempted to go the Bayesian route myself here. | |
| Oct 7 at 14:46 | history | edited | kjetil b halvorsen♦ |
edited tags; edited tags
|
|
| Sep 23 at 8:15 | comment | added | Robert Long | @BenP it’s been a while since I’ve looked at the documentation but I’m pretty sure that the ‘mice’ package for R can impute grouping factors in a mixed model. | |
| Sep 21 at 11:02 | comment | added | BenP | @Robert_Long Imputation seems impossible to me here, as the id number of a bird should be "imputed". I cannot imagine how to do that. What do you think? | |
| Sep 19 at 23:02 | comment | added | Pat Taggart | Yes, a lot of missingness. I just checked my complete cases when all random effects are included in the model - the dataset is reduced in size by 73% (i.e. total dataset is 2649 observations and complete cases across variables of interest to analysis is 731 observations, with most of the missingness occurring in random effect variables). This is why my initial inclination was to just say there is too much missingness in these random effect variables for them to actually be used. However, your suggestion to run two sets of models is good. | |
| Sep 19 at 14:53 | comment | added | Robert Long |
That is quite a lot of missingness! I would suggest that you do two analyses and report the results of both; one which uses only the data with no missingness (complete case analysis, which I believe is what lme4 does), and another which handles the missing data in a statistically principled way, such a multiple imputation.
|
|
| Sep 19 at 5:06 | history | asked | Pat Taggart | CC BY-SA 4.0 |