

Missing at random and covariates 

Message/Author 

Bill Dudley posted on Thursday, November 20, 2008  4:44 pm



I am modeling change over time with 6 time points and n~800. there is considerable drop out from the study with about 70% completing the study. Age is a associated with missingness. My questions have to do with including age as a covariate in the analyses given its association with missiningess. First do I covary age using the WITH Command e.g. i s q WITH age; If the WITH command is correct, does it matter at what point in the model statement I insert the covariate command? When I use the with command my CFI drops from about .96 to .86 and the associations with age (in the with command) are not significant. Should I retain age as a coavriate given that it degrades the model. A more general question is... Are there guidelines about what to use as covariates given the presence of missingness? Thanks Bill 


These issues are discussed in my Mplus Web Talk "Missing data correlates using ML". The talk covers the new Mplus feature introduced in version 5.1: auxiliary = missing. Here, missing data "correlates" are included in the model to improve the plausibility of the MAR assumption without deteriorating fit. This follows ideas in the literature discussed by Schafer, Collins, Graham, Enders among others. Check our home page for a link to Mplus Web Talks. 

Bill Dudley posted on Friday, November 21, 2008  11:16 am



Thanks . The Web talk was very informative and led me to a problem with my growth curve analyses having to do with TYPE = Random. Myy model is: MODEL: i s q  a4wt@0 b4wt@6 c4wt@12 d4wt@18 e4wt@24 f4wt@36; If I do not specific TYPE = RANDOM I get : THE STANDARD ERRORS OF THE MODEL PARAMETER ESTIMATES COULD NOT BE COMPUTED. THE MODEL MAY NOT BE IDENTIFIED. CHECK YOUR MODEL. PROBLEM INVOLVING PARAMETER 1. If I include the TYPE = RANDOM , the model runs and I can also run a conditional model in with the TX is a binary variable indicate therapy type: MODEL: i s q  a4wt@0 b4wt@6 c4wt@12 d4wt@18 e4wt@24 f4wt@36; i s q on abtx2; However, if I try use the Auxiliary = command to increase the plausibility of the MAR, I get: The 'm' specifier in the AUXILIARY option is not available for TYPE=RANDOM. So is it valid to run the model as a RANDOM effects model and if so is there another way to specify the covariates to increase plausibility of MAR? 


Using Type = Random or not should not make a difference, except when the nonidentification condition is at the verge of the threshold. The reason for the error message without Type = Random should be explored. Please send your input, output, data and license number to support@statmodel.com. 

Back to top 

