Message/Author |
|
|
Dear Linda and Bengt, when using the Mahalanobis distance (for outliers) and the Cook's distance to identify influential cases, what critical values do you suggest: After some browsing I found that a small p-value (p<0.001)>4/n) for Cook indicates problematic cases. Would you agree with these critical values? Thanks. Hanno |
|
|
I don't have experience with such p values. I just plot the outlier score against some key variable such as the estimated intercept score in a growth model - and then visually determine if a point stands out. I tend to use the LL influence measure which is in an interpretable LL metric. |
|
|
Hello, Could you please elaborate on how to go about testing for outliers in Mplus? I am running growth models and would like to test for outliers. Are there any references for this? Thanks, Sarah |
|
|
You can find information about outlier detection in both the SAVEDATA and PLOT commands in the user's guide. |
|
|
Thank you. When I try to use the Savedata command to save the values of Influence and Cook's D, the resulting dataset that I get seems to be formatted incorrectly (i.e., the rows and columns don't all line up). I have tried specifying the format, but that doesn't seem to make a difference. Do you have any suggestions? Thanks, Sarah Dauber |
|
|
Please send your input, data, output, saved data, and license number to support@statmodel.com. |
|
|
After identifying an outlier, Is there any easy way to unselect outlier observations? Or, is it necesary to have an auxiliary variable to identify individual observations, in order to do it via a useobservations? Thanks, Fernando. |
|
|
Once identified, you would need to eliminate the outlier using the USEOBSERVATIONS option. I believe that you can obtain auxiliary variable information by holding the cursor on the outlier in the plot. |
|
Antti Kärnä posted on Thursday, October 06, 2011 - 12:02 am
|
|
|
Hello, I tried the outlier screening procedure involving the LL influence measure. One subject had a value of 39.58. When this subject was excluded from the analysis, the change in the loglikelihood was 40.37. Are the LL influence values only approximate? Thanks, Antti Kärnä |
|
|
I think that sounds very close. You should not expect an exact change. |
|
Back to top |