Message/Author 


I'm trying to handle missings on covariates in a two level analysis with FIML. In order to do so, I want to include them in the model so they're not Xvariables anymore. my syntax is the following : VARIABLE: NAMES = SubjectID TIME AGE SEX SES SCORE1L1 GROUP1 GROUP2; USEVARIABLES = TIME AGE SEX SES SCORE1L1 GROUP1 GROUP2; cluster = SUBJECTID; within = TIME; between = AGE SEX SES GROUP1 GROUP2 Missing = .; MODEL: %WITHIN% B1SCORE1L1 ON TIME; %BETWEEN% SCORE1L1 ON AGE SEX SES GROUP1 GROUP2;!predictors only for intercept SCORE1L1 with B1; AGE SEX SES GROUP1 GROUP2;!mention the covariate (with missings) to add them to the model AGE SEX SES GROUP1 GROUP2 with AGE SEX SES GROUP1 GROUP2;!unrestricted structure because they do not correlate by default otherwise 1) Does it look ok ? 2) if I had another level1 covariate with missings (let's say COVL1). Should I add "TIME COVL1" & "TIME WITH COVL1" in the Within section ? Thank you very much ! Philippe 


If you say AGE SEX SES GROUP1 GROUP2; means, variances, and covariances will be estimated. It is not necessary to specify the other parameters. If you have a timevarying covariate, mention its variances on within. There is no need for an interaction with time. See Example 9.16. 


Great ! It helps me a lot and I'll give it a try. Just to be sure, although not needed, the "AGE SEX SES GROUP1 GROUP2 WITH AGE SEX SES GROUP1 GROUP2" would still estimate the exact same model? Thank you again Philippe 


Yes, it will. Do it both ways to prove this to yourself. 


Thank you. I just did and it looks like the covariance are not estimated by default. When I add the "AGE SEX SES GROUP1 GROUP2 WITH AGE SEX SES GROUP1 GROUP2" line, the number of free parameters goes from 21 to 31 (corresponding to the 10 covariances). In simple regression models however only the variances statement is needed and the covariances are estimated by default. Thanks you again Philippe 


Hello, I am using a twolevel model to account for clustering of children within families. In a simplified version of my model (i.e., using only one predictor), I get the following error message: "Data set contains cases with missing on all variables except xvariables. These cases were not included in the analysis. Number of cases with missing on all variables except xvariables: 52" However, in reality, I have no missing data on the predictor (age_C) and 52 cases with missing data on the outcome (boxcoop). Doesn't FIML allow for missingness on the outcome as long as the predictors are not missing data? Here is my input: usevar = boxcoop age_C; missing is all (999.00); cluster = famid; within= age_C; analysis: type = twolevel; model: %within% boxcoop on age_C; output: sampstat stand; Thank you, Heather 


The message says you have missing on the outcome. You need more than one outcomes for FIML. 


FIML allows for missingness on the outcome only if the predictor is modeled as well. In your setup it isn't  it is simply conditioned on as in regular regression. If you want it modeled you have to mention e.g. its variance. This way, you have a bivariate model, not a univariate one. It comes with a price, however, because you have to assume normality for the predictor which you don't in the model you show. 


Dear Dr. Muthen, I am conducting a crosslevel interaction analysis (with random slope & random intercept), however there are some missing values on the Level 1 predictor as well as on the dependent variable (both continuous variables). I thought that the FIML feature should help me to handling missing data as default.However there are two warnings: *** WARNING Data set contains cases with missing on xvariables. These cases were not included in the analysis. Number of cases with missing on xvariables: 5 *** WARNING Data set contains cases with missing on all variables except xvariables. These cases were not included in the analysis. Number of cases with missing on all variables except xvariables: 3 This is my input: variable: names = code SD1 css bos EIP; missing = all(999); usevar = code SD1 css bos EIP; cluster = code; within = css; between = SD1 EIP; analysis: type = twolevel random; model: %within% beta1j  bos on css; %between% bos on SD1 EIP; Beta1j on EIP; bos with beta1j; output: sampstat; Why does FIML not work in this case? Which estimator can I use instead? Thanks for your help!! 


Subjects with missing on x variables are deleted in regular regression, so this is not unusual. FIML is applied to DVs, not IVs. But if you worry about these cases (there are very few) you can "bring the x's into the model" by mentioning their variances. 


Thanks a lot for your quick answer! 

Back to top 