I assume that your outcomes are categorical. With categorical outcomes, missing data. and a model without covariates, pairwise present data are analyzed. You can deduce the number of observations used by looking at the covariance coverage output and the missing data patterns that are printed with the PATTERNS option of the OUTPUT command.
cindy chan posted on Friday, April 18, 2008 - 10:39 am
Eric Teman posted on Sunday, June 19, 2011 - 3:34 pm
I am not sure what this means 'The default is to estimate the model under missing data theory using all available data. To turn on listwise deletion, specify..' (Mplus manual, p.459-60). Does this suggest that all available data are included in the model and path analysis? If I have, say 80 missing values for one variable and 130 for another, how does the system deal with this inconsistent number of missing data for each variable? Does the system use pairwise deletion as a defualt to analyze the path coefficients? Thanks. pat
I am doing a correlation analysis in mplus. According to my data, I decided to use TYPE=COMPLEX and Pairwise deletion, but I didn't find any option to specify pairwise deletion, does this mean it is not provided in mplus?
If you want pairwise deletion, you will need to do it yourself. We do not have that option.
John C posted on Wednesday, May 09, 2018 - 11:39 am
I have a few questions on the default missing data mechanism for categorical outcomes using weighted least squares estimation.
In the user guide, it says "For censored and categorical outcomes using weighted least squares estimation, missingness is allowed to be a function of the observed covariates but not the observed outcomes. When there are no covariates in the model, this is analogous to pairwise present analysis."
1 - when there are no covariates, in what sense is this only "analogous," but not equivalent to, pairwise missing?
2 - when there are covariates, how is this no longer analogous to pairwise present?
3 - is there a paper I can cite for more information on this mechanism?
1. The idea is that regular pairwise implies that you estimate means/thresholds, variances, and correlation for those subjects. In our weighted least squares, thresholds are estimated for everyone having observations on that variable, leaving the correlation estimation for the subjects with complete data on both variables.
2. If the covariates predict missing on the outcomes, we have (an approximation to) ML under MAR.
3. No paper on this specifically as far as I can recall.