Pool results from multiply imputed data

[Sifan Cao]

Hello Statalisters,

I have a multiple imputed dataset created by previous researchers in the team (no. of imputation=20, imp=1 to 20) and I have the unimputed original data (imp=0). I have stacked them together.
My question: How should I pool the estimates of the effect size (relative risks and 95% Confidence Intervals) of each imputed dataset? I used mi xeq 0 to get the effect size for imp=0 the original data set. In SAS, this is where MIANALYZE comes in. But I'm not sure what the equivalent is in Stata.

I am running a multi-level mixed-effects log-Poisson regression.

Thanks in advance. Any thoughts would be appreciated.

Sifan

[Rich Goldstein]

what type of analysis you are doing is not clear, but the place to start is

help mi estimate

[Carlo Lazzaro]

Sifan:
see also -mi estimate- esntry, Stata .pdf manual.

[Sifan Cao]

Okay, thank you both, Rich and Carlo. So for using -mi estimate-, do we include the unimputed data as well or we simply use the imputed data with whatever number of imputations are made?

Thanks again
Sifan

[Rich Goldstein]

Stata will take care of this for you - is your data in "flong" format? if not, what format is it in?

[Sifan Cao]

Stata will take care of this for you - is your data in "flong" format? if not, what format is it in?

Yes, data in which m = 0, m = 1, . . . , m = M are all recorded in one .dta dataset.

[Carlo Lazzaro]

Sifan:
Rich's question was about a different topic (see -help mi set-).

[daniel klein]

Although it is not explicitly mentioned in the initial post, it seems the data has been multiply imputed outside of Stata (perhaps with SAS). Let's hope the multilevel (log-Poisson) structure has been incorporated correctly during imputation. Anyway, the key now is to get the data into Stata and tell Stata that it is multiply imputed data. The best way to do that is to use mi import. Depending on how the "stacked" dataset looks like, you probably want mi import flong. Read the (short) documentation and the example(s).

Do not try to workaround correctly importing your dataset into Stata. Chances are you are running into problems later.

Once you have your data mi set, you must force Stata to estimate your model because I don't think mixed effects (log-) Poisson regression is officially supported. There are three main reasons (or a combination of them) why a model might not be supported: (1) Results are invalid. (2) Results have not yet shown to be valid. (3) StataCorp. did not find the time to implement the model and certify the results. Either way, you can typically force Stata to estimate the model with the cmdok option of mi estimate.

[Sifan Cao]

Although it is not explicitly mentioned in the initial post, it seems the data has been multiply imputed outside of Stata (perhaps with SAS). Let's hope the multilevel (log-Poisson) structure has been incorporated correctly during imputation. Anyway, the key now is to get the data into Stata and tell Stata that it is multiply imputed data. The best way to do that is to use mi import. Depending on how the "stacked" dataset looks like, you probably want mi import flong. Read the (short) documentation and the example(s).

Do not try to workaround correctly importing your dataset into Stata. Chances are you are running into problems later.

Once you have your data mi set, you must force Stata to estimate your model because I don't think mixed effects (log-) Poisson regression is officially supported. There are three main reasons (or a combination of them) why a model might not be supported: (1) Results are invalid. (2) Results have not yet shown to be valid. (3) StataCorp. did not find the time to implement the model and certify the results. Either way, you can typically force Stata to estimate the model with the cmdok option of mi estimate.

Do I need to run both mi import flong & mi set flong before running the analysis?

[daniel klein]

Do I need to run both mi import flong & mi set flong before running the analysis?

No.

Type

mi describe

after you import the data.;the output will point to potential problems.

You may also want to run

mi varying

If you have not specified the respective options with mi import, you may want to mi register the relevant variables.