First, I am new to analyzing public opinion polls and the r package "Survey". I would like some advice. I am running a regression model with weights from a Pew survey, however, I noticed that a significant portion of my data is missing because of the covariates. As a robustness check, I would like to impute my data set. I would like some advice on the best way to handle this in r.

I am most familiar with using the "mice" package in r to handle missing data. I don't believe "survey" can accommodate mice. Should I separate each imputed dataset and then perform a regression analysis (using "survey") for each dataset? Is this the most efficient method? Finally, how do you pool the estimates? You don't average the estimates from all the imputed data sets, do you?

  • 2
    $\begingroup$ for work with the survey package, you probably want mitools and not mice - mitools was written by the author of survey $\endgroup$ Commented Feb 7, 2016 at 1:34
  • $\begingroup$ As far as I know, mitools is a package to handle the imputed data. mitools does not perform imputation. $\endgroup$ Commented Mar 17, 2022 at 15:57

1 Answer 1

  1. I analyzed each imputed dataset separately with the complete command.
  2. I averaged all the coefficients. However, to calculate the SEs, I used the formula from this site (http://circoutcomes.ahajournals.org/content/suppl/2010/01/06/3.1.98.DC1/HeDS.pdf). Specifically, I averaged the variance of my coefficients across all my imputed data sets + (1+1/number of imputed datasets) x the between imputation variance. Finally, I square rooted that answer.
  • $\begingroup$ It would be great to have an updated, live link to that article. $\endgroup$
    – abalter
    Commented Aug 27, 2020 at 1:00

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.