Dear researchers.
We are doing a study for estimating the prevalence of a disease diagnosed by a blood sample test. Some participants have answered the survey questionnaire but have refused the blood sample or the test procedure has failed.
Is it possible to estimate the prevalence taking into account other personal data in the questionnaire without missing values?
Which method is the best? Multiple imputation or Bayesian estimation?
Thank you.
Do you expect the people that have refused a blood sample to have a different prevalence of the disease than those that accept to give blood? Does the disease have noticable symptoms? If not I would estimate prevalence based on the actual test, instead of having to impute a probable diagnosis. I am sure someone more experienced in imputatation will provide insight but it does not sound ideal to have to impute the actual outcome (rather than imputing dependent variables).
As Kristian is referencing, you need to grasp at the nature of the missingness: is it random, completely random, or not random? Here's a powerpoint given at an ASA webinar a few years ago.
Depending on how large your dataset is, you can validate how well an imputation method (or any method you choose) performs by taking a small sample of your complete data, delete the blood sample result, and using the rest of your data to see how well it performs by comparing the actual blood sample result with what your method would predict.
https://www.amstat.org/sections/srms/webinarfiles/ModernMethodWebinarMay2012.pdf
Great answers by Joe and Kristian. Define the missing values in detail please.
- Badges
- Science topic