Missing data: FIML and (multilevel) multiple imputation
Posted: Tue Aug 19, 2014 12:15 pm
Hi experts
I have >15% missing data in my continuous multilevel model; missing at independent variables at all levels but not the dependent variable on level 1. Paul Allison writes that Full-Information Maximum Likelihood has advantages compared with multiple imputation for single-level models. MLwiN apparently uses FIML (Hox, J.J. 2009 Multilevel Analysis: Techniques and Applications p138) (Twisk, Jos W.R 2003, Applied Longitudinal Data Analysis for Epidimeology, p.262) (what I gather from MLwiN manuals, forum etc) but is also not listed among 'FIML-software' in e.g. Allison 2012 (http://www.statisticalhorizons.com/wp-c ... llison.pdf). I understand that missing at cluster level results in the whole cluster being discarded. However, it does not seem to me that MLwiN uses all information available within cluster 2 (level 1) since only a listwise deletion could explain that I end up with only using 230,000 of 309,000 cases. Besides, I 'feel' MLwiN is so much quicker than MPlus (which definitely uses FIML) and results in less missing. I'm now spending 3+ days to impute level 1 and level 2 variables by level 4 clusters (countries) with Realcom Impute to see if anything improves, but I wonder how biased my model would be without the complexity of imputation in an already complex study. Note that I am forced to have 5 imputed sets if using MI.
In short, is IGLS equivalent to FIML and can MAR at level 1 be 'ignored'? Or am I missing the point (pun intended).
Thanks for any help.
EDIT: Spelling
I have >15% missing data in my continuous multilevel model; missing at independent variables at all levels but not the dependent variable on level 1. Paul Allison writes that Full-Information Maximum Likelihood has advantages compared with multiple imputation for single-level models. MLwiN apparently uses FIML (Hox, J.J. 2009 Multilevel Analysis: Techniques and Applications p138) (Twisk, Jos W.R 2003, Applied Longitudinal Data Analysis for Epidimeology, p.262) (what I gather from MLwiN manuals, forum etc) but is also not listed among 'FIML-software' in e.g. Allison 2012 (http://www.statisticalhorizons.com/wp-c ... llison.pdf). I understand that missing at cluster level results in the whole cluster being discarded. However, it does not seem to me that MLwiN uses all information available within cluster 2 (level 1) since only a listwise deletion could explain that I end up with only using 230,000 of 309,000 cases. Besides, I 'feel' MLwiN is so much quicker than MPlus (which definitely uses FIML) and results in less missing. I'm now spending 3+ days to impute level 1 and level 2 variables by level 4 clusters (countries) with Realcom Impute to see if anything improves, but I wonder how biased my model would be without the complexity of imputation in an already complex study. Note that I am forced to have 5 imputed sets if using MI.
In short, is IGLS equivalent to FIML and can MAR at level 1 be 'ignored'? Or am I missing the point (pun intended).
Thanks for any help.
EDIT: Spelling