Missing Data/Imputation Discussion > multiple imputation and multivariate outliers
Hey Lisa,
I would deal with any transformations, outliers, and other adjustments to the data PRIOR to multiple imputation, so that the imputation procedures are based on the data that you will use for analysis. I hope that is helpful!
Best.
Jeremy
Jeremy Taylor
Thank you, Jeremy, that's helpful. Another question I have is if AMOS is able to import a multiply imputed dataset from SPSS (i.e., the multiple imputation procedure was run in SPSS and I would now like to run SEM on it in AMOS)? How would I go about doing this (i.e. importing the SPSS dataset)? Is there a way to define the multiply imputed SPSS data set in AMOS (so that I would not need to re-impute in AMOS...)?
Lisa
Actually, I found out that the imputations conducted in SPSS could be "grouped" in AMOS.
Lisa
That is true, great idea Lisa! I've not tried this, but it would give you an overall model fit, at least!... I don't think it would give you "summary statistics" across data sets though (such as the pooled path coefficients and/or covariances...etc). However, you could always average them by hand, which would be easy enough with a spreadsheet in excel. Great find!
Jeremy Taylor
Hi Jeremy,
I'm dealing with the same problem of how to deal with outliers when using multiple imputation. My problem is that the case I am having trouble with doesn't show up in the regression diagnostics I perform with the original dataset, because it is kicked out of the analysis due to a missing covariate. Based on running regression diagnostics on the imputed datasets, this particular datapoint appears to be problematic in most datasets, but not all. I'm not sure of how to evaluate whether to keep the case in the dataset or not.
Thanks so much for your help!
Amy
Amy
Hi Amy!
This is an interesting issue, and not one that I've seen often. If I understand you correctly, are you saying that the IMPUTED value for some cases is showing up as an outlier (in the original data they are excluded due to missing data, but they are outliers when data is imputed)? If that is true, I would check the variables that are used as predictors in the imputation model, to see if any of those variables are unusual, as they would be the variables that lead to the imputed value being abnormal. Let me know if this makes sense, and I hope I'm helpful.
Jeremy Taylor
Hi Jeremy,
In my design, I have predictor variables which are assessed for all participants. The experimental manipulation occurs before the assessment of the CRITERION, so the same participants are split into two groups.
Now for multiple imputation: I understand that I have to impute values for the criterion seperately for the two groups.
But what about the predictor variables? Do I impute values for the whole group (since the experimental manipulation occurs AFTER the assesment of the predictor variables), or seperately for the two groups? My question also concerns the outlier analyses.
My final analysis is a multiple regression.
I'm grateful for any help!!
Kristina
Hi Jeremy,
Thank you for your responses to my previous posts.
I am wondering if you would be able to provide some guidance on how best to detect multivariate outliers in a multiply imputed dataset?
Would you remove multivariate outliers from the original dataset before multiple imputation (I suppose, however, that this may not necessarily remove any multivariate outliers in each dataset that was imputed)? Thus, would you examine each dataset separately for multivariate outliers after multiple imputation?