Created 12-12-02

HANDOUT

• MISSING DATA EXAMPLE - showing FIML (Full information Maximum Likelihood, Mean imputation, pairwise deletion, listwise deletion)
• SIMULATION STUDY - comparing the different methods under MCAR and MAR.

EXAMPLE for thinking about different kinds of missing data.

Survey for a large medical organization to investigate violence in the workplace.

Employees were sent a survey at their homes and were asked many things, including whether they had been the victim of physical violence within the last year.

Of the 4000 surveys sent, the response rate was 55%

Different kinds of missing - are they reasonable for this data?

• MCAR - There is no relation between the reason the data is missing and what the persons response to the question about violence would be.
  • this could be reasonable if data were missing because of incorrect mailing addresses.
  • other examples?

• MAR - There might be a relation between the reason the data is missing and what the persons response would have been but this relation goes away if we adjust for certain covariates.
  • Assume there is a covariate related to the probability of responding to the questionnaire, e.g. job family - Nurses are more likely to respond to questionnaires than doctors. Now if nurses have higher rates of violence then the data are not MCAR because there is some relation between responding and experience violence. BUT, if this relation exists only because nurses tend to respond more then since we can adjust for whether the person is a nurse or not, the data is MAR.
  • other examples?

• Non-ignorable - The reason the data is missing is directly related to what the persons response would have been.
  • persons who have experienced violence are more likely to respond to the questionnaire because they have some feeling of wanting to get the message out about violence.
  • other examples?