# Missing data imputation when all variables have some missing data?

I am working with my resident survey data (n=8356) including 59 items, most of which are ordinal variables scored from 1 to 7, and others are continuous variables (e.g., age, residency length).

However, all items have some missing values, ranging from 40 to 200, which follow an arbitrarily missing pattern. This leaves me with 6001 complete observations but more than 2000 incomplete records. How can I impute the missing values in this situation?

I realize that I could use single or multiple imputation method, while it seems that both require variables without any missing value, which is not the case in my data.

Additionally, I want to generate some composite variables based on the individual items and then run my analysis model, so how to get the combined estimates if I use multiple imputation method such as chained equations?

Thanks for any help.

the method of imputation depends on your underlying assumptions and what you want to use the data for (i.e: type of regression).

The general approach I would use would be to first run the analysis on the data that you have the complete observations (n=6001) and see if your hypotheses hold. It is okay to use this for main results under the assumption that the "missingness" occurs at random and you found no pattern with it. Then after that, re-run the whole analysis with imputed data (n=8356) for robustness check, and clearly stating the assumptions you make.

There are several ways you may perform the imputation.

(1) Using a 'neutral' value of '4' for a question that asks whether "Do you like or dislike ...". This assumes that non-response indicate that the resident does not hold an opinion on this, hence skipped the question.

(2) Using the average of age/residency might be appropriate because during the regression, the mean will not have effect on the regression.

After doing the imputation, you may wish to perform follow-up investigation with a small group of participants (if possible) to test whether your assumption is appropriate.

I suggest you check your relevant research literature to see if there's an established way to deal with this. For example, some researchers may allocate "0" for firms that did not report R&D expenditure but other researchers found that this introduced a bias in the results. Instead, one should use an additional dummy to indicate that R&D expenditure is missing.

• Kudos for your points about purpose and assumptions. However, imputing means was common some years back, but nowadays, with more sophisticated methods so widely available, mean imputation is almost always a far inferior choice. It fails to take advantage of available information and introduces substantial amounts of data with zero variation, distorting multiple aspects of an analysis. Jul 28, 2019 at 17:06