Hi!
We have some trouble with dealing with missing data. Children fulfilled several psychological instruments, but some of them miss questionnaires related to father or mother, if they doesn't have a parent(s). We have to conduct SEM in AMOS but I'm not sure how to deal with missing data in this case.
Do we have to exclude some participant with missing data or there is some other way we can use and save 110 participants?
Thank you! :)
It depends on what assumptions you are willing to make about the missing scores. If they are missing at random or missing completely at random (see definitions in the literature, e.g., Enders, 2010, Applied Missing Data Analysis, Guilford Press), you could use full information maximum likelihood estimation (FIML) or multiple imputation to include all available scores. If the data are missing not at random (MNAR), the situation is more difficult and you'd have to look more into the literature on missing not at random SEM models/analyses.
A good start would be to compare the groups/children with and without missing data on as many variables as possible (e.g., using t tests and effect size measures such as Cohen's d to find out on which variables the groups may differ). This would allow you to find out whether cases are systematically missing, i.e., testing the missing completely at random (MCAR) assumption.
Christian Geiser Thank you!
Maybe I didn’t describe it best. The children left out the whole test. For example, if the instrument measures the relationship with the father, and the child does not have a father, then I do not have data for that child and that is not just a couple of answers, but the whole test is skipped.
Nikola Goljović : Maybe you could describe your concrete research questions / hypotheses as well as your independent and dependent variables of interest in some detail. This would make it easier to see what a good solution / missing data handling strategy may look like for your situation.
Christian Geiser : In this particular case, we have an attachment to the parents (2 dimensions per parent), and 4 dimensions acting as an X1, X2, X3 and X4 variable to the mediation model. Mediators are some personality traits (4), and the criterion variable is presented based on the dimensions of children's emotional intelligence, so we have 6 criteria that should be found in the model.
Nikola Goljović and Christian Geiser : Why not use multilevel modeling with parents at level 1 and children at level 2? Then, all available information would be used.
HTH, Karin
Karin Schermelleh-Engel: Thank you. Great idea, but I'm still not sure if we can use 2 levels, because the data about the parents were given by the children and their self-assessment of the parents is a predictor of the variable that refers to them (the children).
Nikola Goljović : We need more information. How large is your sample size, and how many children provided data only for their mother, only for their father and for both parents?
Kind regards, Karin
Karin Schermelleh-Engel : Sure. Total - 478. Only for father - 30 and only for mother - 110.
Nikola Goljović and Karin Schermelleh-Engel Since the missing questionnaires seem to be about child-parent relationships, it may not make sense for you to include children who do not have a mother or father in the same analysis / group as children who do have both parents. The children with only one parent (or no parent at all) represent a structurally different group as they simply cannot have a relationship with the missing parent(s). I would recommend treating these children as a separate group / population and analyzing their data separately (e.g., through a multigroup SEM).
Christian Geiser and Karin Schermelleh-Engel : You are totally right. This sounds absolutely logical. Thanks a lot for the help, both of you.
- Badges
- Science method
- Similar topics
- Mathematics
- Statistics
- Statistical Techniques
- Structural Equation Modeling