1

What are your thoughts on where or how to show (non-)imputed data?

Please regard this question as a more a general question. I am in the field of medical clinical research, where missing data is very common. It is most likely caused by mixed reasons (random/not-random). I would like to use multiple imputation to assess and impute missing data.

At the moment, many medical reports do not use imputation: https://pubmed.ncbi.nlm.nih.gov/27556528/ I want to do it the correct way, but also create a manuscript with good readability (for clinicians). In addition, most journals have very strict word, table or page limits.

For example (answers don't have to be limited to this example): Is it okay to show only the imputed data in table that summarizes the main outcomes? Or should one, for example, show the raw data in a table (including numbers that have missing data) and only use the imputed data in your analyses? Or do you have to show both (which reduces readability)?

Curious for your answers!

Kim
  • 13
  • 1
    Many imputation methods don't create "imputed data" that can be reported. (For instance, multiple imputation methods create many different random versions of the dataset.) This raises the issue of what you might mean by "imputed data:" are you using some method of data replacement for your analysis? What would that method be? – whuber Nov 04 '21 at 12:01
  • 1
    Thank you for your response. I do not have much experience with imputing. I was thinking of using the mice package with predictive mean matching, which is able to create one imputed dataset by using complete(), which, as is my understanding, uses the first (of usually 5) imputed dataset. – Kim Nov 04 '21 at 13:28
  • the questions still remains! Anyone? :) – Kim Dec 02 '21 at 10:04

0 Answers0