Limits for missing values in the multiple imputation method

[Syntax6]

I have an understanding question and I need help. I have a record with many missing values. I now wanted to use the multipile Imputation method to replace the missing values. I would now like to know how many values per subject (in percent) may be missing, so that I can use the MI method at all. Is there a documented source or guidelines? Maybe studies?

Thanks in advance

 

[hlsmith]

First question should be whether you know the mechanism causing missingness.

 

[gdaem]

If I remember correctly, it is something like anything more than 30% of your values are missing you should not use MI.

There are resources out there for handling missing data, I am sure a quick Google search of something like 'Multiple Imputation assumptions' would give meaningful results.

 

[hlsmith]

I was getting at if missing data are MAR, then MI is a decent approach. Though if data are MCAR or NMCAR then things change a bit.

 

[noetsi]

I spent a lot of time working on this issue only to find MAR assumptions were commonly not justifiable and that these approaches work poorly with non-interval data.

I have not seen suggestions about how much data can be missing and you can still use MI. Usually the question is how much data can be missing before you have to use MI.

How much data are you missing?

 

[Syntax6]

Thanks for the helpful answers. The data are MAR. I wanted to carry out a case study. There are few cases that are complete. In many cases, up to 70% of data are missing.
The data were collected with a questionnaire and some questions were not answered.
I would like to use the MI procedure to get more usable cases.

 

[hlsmith]

How many people responded to survey and what typy of data (Likert)?