I have a data set of around 1630. Some variables are missing in 20% of the sample (the later questions - survey fatigue/drop out). (I'm considering removing 250 participants who didn't answer all the relevant scales for my analysis)
Is there any literature or 'rules' regarding when I listwise delete the participant versus impute data (which isn't missing at random according to Little's MCAR test)
Thanks.
Have a look on MCMC rejection sampling algorithms : http://bjlkeng.github.io/posts/markov-chain-monte-carlo-mcmc-and-the-metropolis-hastings-algorithm/
i hvan't come across any literature rules. As a rule of thumb, it is always to keep data than deleting data. Also it depends on what type of analysis you are goin to perform with data. Probably the link below should help you to decide.
https://towardsdatascience.com/how-to-handle-missing-data-8646b18db0d4
Removing those participants from the analysis can introduce selection bias. Check the background characteristics of those complete and incomplete participants. The unequal distribution among these two groups will surely bias your result. Dr. Ali is right. You can use the MCMC method. The EM algorithm is an alternative. However, make sure the model diagnostics before making your conclusion.
My advice is based on my work with medical patient data and regulated financial institutions. I do not impute in either case the reasons are a)financial-this data is required by regulators so no data indicates a problem with the institution-I clearly say this in my reports, b)patient data-I usually am trying to say something about risk factors- as Md. Belal Hossain reports this biases my sample but i again am clear about what I do and I think that imputation could bias my result worse than otherwise. That said, in other situations where considerations such as I have mentioned are not relevant I would impute with an appropriate algorithm . Best, D. Booth
- Badges
- Science topic
- Similar topics
- Mathematics
- Statistics