More specific questions:
1. Should I estimate the missing values for all scales at the same time? (e.g. depression, satisfaction with life, positive emotions) or rather estimate missing values for each scale separately (e.g. first for depression, then for sat with life, etc.)
2. What should I do, if the estimated value for a single item is outside of the response range, e.g. SPSS suggests that a value of 0.23 should be entered for a missing data in an item that has a response range from 1 to 7. This seems intuitively wrong because it is not possible that a person enters such a value. Should I inspect completed data manually and correct them in this regard, e.g. by entering the minimal allowed value (changing to 1 values below 1, and changing to 7 values above 7)?
3. Is it correct to complete missing data for a variable that is measured with a single item? Suppose I collect satisfaction with life ratings with a single item. What I mean is that for multi-item scales (where several items ask about the same construct) it seems intuitively justified to complete a missing value because we have at least some partial direct information about the construct.
4. What do you think about completing data for categorical variables? E.g. if a person does not report their sex - should this person be excluded from the analyses or is it ok, if I complete this missing information based on EM. Why not? ;)
Thank you!
Dear Lukasz,
You may interested for this video.
https://www.youtube.com/watch?v=4yhpYKzW98M
Dear Lukasz,
Basically, the approaches are similar even though we may use another.
for example, using clustering, we may classify complete samples into some number of classes. Then, we may calculate the shortest distance
Lukasz,
Here are
Examples https://www.youtube.com/watch?v=4yhpYKzW98M https://www.youtube.com/watch?v=P57sC7sGVm8
But beware that in SPSS you are unable to run a GLM repeated using the pooled multiple imputation sample, it simple gives results for the individual sets of imputations but not the pooled.
Dear all and Dr. Kaczmarek
Its been long time since you had this discussion. I hope you have answer to your following questions mentioned below. I am having exactly similar questions regarding use of expectation maximum method in SPSS and would appreciate if you kindly advice me
1. Can I check the assumption of MCAR test for each variable separately or can I include them together and estimate MCAR Test. On one hand I am thinking it might be appropriate to look for it for all items for one measure to know if missing values on that variable (with 10 items) are MCAR. Or can I include all other variables with varying number of items and look for MCAR in the same command. I think my question relates to your question 1. "Should I estimate the missing values for all scales at the same time? or rather estimate missing values for each scale separately"
2. Is it ok to use expectation maximum for binary variables such as age group, gender?
3. If I am missing nearly 40-50% of observation in one of my dataset. Is it is advisable to use Expectation maximum (Given that data is MCAR with p>0.05)? Or due to large % of missing data on observation, it is better to use multiple imputation. I am using SPSS for conducting GLM repeated measures and it seems SPSS does not give output for pooled multiple imputation sample due to which I thought of using expectation maximum method. But wondering if I could have your advise on the same.
Thanking you all for you time.
Kind Regards
Neha
- Badges
- Science topic
- Similar topics
- Medicine
- Physiology
- Biosignals
- Biomedical Signal Processing
- Statistical Analysis