Conducting imputation using MICE

[Njeri Muchai]

I am trying to conduct MICE but I am not as conversant with the technique. I managed to impute data using MICE, but I kindly seek some clarifications below

Are there any post-impute techniques that can be used to determine if the imputation is sound or acceptable?

How can I get a tabulated results from the pooled mi data?
mi xeq: tab varname provides the frequencies but for each imputation separately rather than from the pooled mi
mi estimate: proportion varname only gives proportions and not frequencies for pooled mi

Which command is able to give me frequencies pooled from total number of imputations done?

Thank you!

[Rich Goldstein]

for the post-impute techniques; see -midiagplots- and the associated Stata Journal article - use -search- to find and download/install; opening the help file will tell you which volume/issue of the Stata Journal and you can find it freely available at the Stata web site

I'm not sure why you would want pooled frequencies, but you should be able to get it by ignoring that the data are imputed and using whatever command you want (e.g., tabulate); however, depending on what style you used in your -mi set- command, you may have to decide first whether you want to include the non-imputed data and use the "if" qualifier to exclude if the style includes it and you don't want it

[Njeri Muchai]

Thank you for great feedback. I will read through midiagplots. Thanks again

[Felix Bittmann]

This website also has a lot of information on imputation diagnostics: https://stats.oarc.ucla.edu/stata/se...stata_pt1_new/
Not directly Stata related: https://stefvanbuuren.name/fimd/sec-diagnostics.html

[Njeri Muchai]

Thank you all for the feedback. I have encountered a challenge in conducting regression post estimation (residual diagnostic) techniques using mi set data.

For example: estat hettest to examine heteroskedasticity doesn't work under mi estimate: command
Second is how do I get studentized residual following multiple linear regression to test for normality of residuals. The predict var, rstu not available under mi predict. Any other options that are available?

Finally - How can I check if a variable improves the model, the lrtest doesn't work under mi set data.

Any assistance or support with resources that I can read on this is highly appreciated.

[Andrew Musau]

For example: estat hettest to examine heteroskedasticity doesn't work under mi estimate: command
Second is how do I get studentized residual following multiple linear regression to test for normality of residuals. The predict var, rstu not available under mi predict. Any other options that are available?

The regress command, when used with the -mi estimate:- prefix, allows for heteroskedasticity-consistent (HC) standard errors through the -robust- option. As a result, there is no need to test for heteroskedasticity. Additionally, if your sample size is sufficiently large (say 30+ observations), you can rely on the Central Limit Theorem, making it unnecessary to check for normality of residuals. In any case, normality of residuals does not necessarily imply that the assumption of normally distributed errors is met.

How can I check if a variable improves the model, the lrtest doesn't work under mi set data.

See mi test which allows you to run Wald tests following mi estimation. The likelihood-ratio (LR) test and Wald test are asymptotically equivalent.

help mi test

[Felix Bittmann]

An addition to Andrew's excellent comment: if you want the most conservative regression standard errors, use the option vce(hc3).

[Njeri Muchai]

Thank you Andrew and Felix for your invaluable input. Your support highly appreciated.