How to estimate the potential effect of missing data?

[Minda Tan]

The dataset I used has a sample size of 9000. Given that I constructed an HLM model with many variables, the analytic sample size was only 6000. I checked the missing distribution, which showed that the missing happened in almost all variables. (For example, 100 people reported their economic status; however, they did not report their education level. Meanwhile, 200 people did oppositely. As a result, 300 samples were lost in this example.)

In my case, I don't know how to check the robustness and sensitivity of my results. What kind of estimation I can do to make sure my results are convincible?

Thank you!

[Carlo Lazzaro]

Minda:
first, you should investigate the missing mechanism that led to unobserved data (informative or not).
Then you should deal with missing data according to the methods covered in the literature.
Stata offers a dedicated section on this topic (-mi-) in Stata .pdf manual, along with an almost boundless literature.
Dealing with missing data is often a tricky and time-consuming job.
However, it is also a rewarding job, especially if you're planning to submit your research to a technical journal of your field: missing value issues are now considered more carefully by many reviewers (at least this is my recent experience as author and reviewer).

[Minda Tan]

Minda:
first, you should investigate the missing mechanism that led to unobserved data (informative or not).
Then you should deal with missing data according to the methods covered in the literature.
Stata offers a dedicated section on this topic (-mi-) in Stata .pdf manual, along with an almost boundless literature.
Dealing with missing data is often a tricky and time-consuming job.
However, it is also a rewarding job, especially if you're planning to submit your research to a technical journal of your field: missing value issues are now considered more carefully by many reviewers (at least this is my recent experience as author and reviewer).

Thank you so much!