I don't have a licensed copy of SPSS 22 but of SPSS 17. I will try to answer your questions on the basis of SPSS 17:

1. When you run multiple imputation in SPSS the default setup runs 5 iterations and writes the data to either a new SPSS file or saves it as a new SPSS file [You have to specify the choice]. In the new file the first column gets labelled as "Imputation_"  is the iteration number starting with Original Data, then 1 [indicating iteration no 1] and so on. Let us call this file splitdata.sav

2. You can use the "splitdata.sav" file for further analysis directly and SPSS should take care of pooling. Click Data -> Split File. In the dialog box select "Imputation_" and move it to "Groups Based On:" box. Check "Compare Groups". Now any analysis that you run should get run on Original Data, Iteration No 1 data, Iteration No 2 data... and the pooled data. This is a real advantage of SPSS cos you can check the results of the original data along with the various imputations and the pooled data simultaneously.

3. If you want to physically seperate the files use Data -> Select Cases -> Set your condition [Usually the If condition] -> and output to -> A new dataset -> dataset name. In this way you can create separate data files which will contain original data and data from imputations.

4. If you have separated the datasets in to different files [as in step 3] and want to pool them into a single file use Data -> Merge File -> Add Variables. But be sure to change the names of the variables in the datafiles created in step 3. Slight differences in the variable names will do [say v01, v02, ... v05] cos SPSS will not be able to merge the variables if they have the same name. Once all the variables are in the same file you can create new pooled variables [v0P = v01 +v02 + v03 +v04 + v05]

[you may also look at the use of Data -> Aggregate function]

4. You want to bring the data from SPSS 22 to SPSS 18. Save the desired data file as SPSS portable (*.por) file. It should be readable in SPSS 18. Else you can try it by saving it as MS Excel file. Personally I prefer moving the data in CSV files

I hope this helps

Thank you for your response. However when I run the split file I only see the results of the five imputations, not the pooled variables that were predicted based on the imputations. Would I run into issues if I made the pooled values for each missing the average of the values of each imputation (i.e. Question 4 for respondent 6 is missing data and the imputted values are 2,3,2,4 and 1, can I argue confidently that I could use the average of 2.4 (2 for Likert purposes) in my analysis?)

Hi,

watch this follwoing link:

https://www.youtube.com/watch?v=ytQedMywOjQ

Bradley: Yes you can do so. You can take the average of the pooled results as the estimate of the missing value. SPSS internally pools the data when it runs any analysis. In step 4 of my reply I have mentioned the ways in which you can generate the pooled file. I hope this helps

 This is so incredibly helpful--thank you for taking the time to explain this, Dr. Kumar!

 Dear Kumar, in case of pooling multiplly imputed data sets into a single physical file for missing values that are categorical data: is it reliable to pool the five standard imputations by taking the most frequent category imputed over the five multiple imputations (if 5 are the number of imputations used, the default in SPSS).  

I am in the same proble. As I understand this, if I have a SPSS18 without license, pooled data is not being to create authomatically? But I can create it manually? Because without pooled data, I am unable to provide a final result of thr multiple imputation tehcnique

Oscar, just read read Dr Kumar replies in the previous sections to know how to do it. Daniele

Thanks all for this post and your replies.

I am using repeated measures with GLM command to analyze my data. I had missing data due to which I have planned to see if my results changes with multiple imputation. But while I analyze my imputed data, I am not able to get pooled estimates/output from 5 imputations with repeated measures. Ideally, I am expected to use pooled output while my spss gives me output for 5 different imputations and does not provide pooled output for repeated measures (as it otherwise gives for t-test or one way anova).  Can someone advise me if there is any way I could get pooled output from repeated measures. While reviewing literature and looking online, I realize that there is no direct way to get pooled estimates if multiple imputation is done in SPSS 18 or older version. I am using SPSS 24 but could not sort out any comprehensible way which I can follow to get the pooled estimates.

If it’s not possible to get pooled output for repeated measures (imputed data) in SPSS, can you please inform me: which imputed dataset (1,2,3,4,5) would be best to use for output results from repeated measures?

I would appreciate if any of you have experiences similar problem and could advise me any solution to get the pooled estimates in SPSS so that I need not deal with any other software for this. 

If there is no good way to get pooled output from imputed data in spss, can you please inform me if it is possible to do it in Stata 15. I would appreciate your advice very much. 

Hope to hear from you.

Thank you and Regards

Neha

Dear Dr. Kumar, Thanks so much for your detailed post as your post gives me some hope as I was thinking that I might need to change the software to get output from pooled estimates.

I am dealing with imputed data for the first time which could be the reason that I am having hard time for this analysis. I am trying my best to follow your steps but after splitting files into 5 imputed datasets, I could not understand step 4 (merging to get pooled dataset).  I really appreciate that you have nicely explained all the steps to follow but is it possible for you to kindly elaborate step 4. I am not clear regarding changing the variable name as I have so many variables (>1000). Do I need to change variable name in all 5 imputed datasets prior to merging to one pooled dataset? One thing is this will take a very long time and also if I do so, I am not able to understand how merging of values will happen if I am changing the variable names. Kindly advise me.

Thanking you for your time.

Kind Regards

Neha

Thank you for all.

Dear Dr. Kumar, thanks for the detailed information. I would like to run generalized linear mixed models (via the GENLINMIXED procedure) with multiply imputed data in spss. However, this procedure ignores 'split file' commands. I would like to confirm if it would be an appropriate solution to compute the average of the imputed values from my 20 datasets into a single imputed dataset. Then I could run my model the same way as I would run a complete case analysis (or one that applies listwise deletion).

Best regards,

Richard

I created a basic SPSS procedure after a SPSS Multiple Imputation (MI) in order to compress the multiply imputed dataframe into a single pooled dataframe. Hope this would solve some questions for SPSS users.

I call this the "Bar Procedure"...

The Bar Procedure is to apply the SPSS function "Output Management System" (OMS) in the "Utilities" tab and work out a "virtual" file very easily. This file will be the final reduced version containing only a single dataframe (instead of the actual multiple split files given by standard SPSS after MI.

I have a clip on you tube here to demo the entire process: https://youtu.be/zftIxv532hE

Below the key steps of the Bar Procedure:

1) perform classic MI procedure in SPSS and obtain the multiply imputed file

2) select and remove the original segment of data

3) sort the multiply imputed file by "ID(cases)" and "Imputation N"

4) Activate the OMS and:

a) select tables (Output types)

b) select frequencies (Command Identifiers)

c) select statistics (Table subtypes for...)

d) save as new dataset (give it a name)

e) in Options select "all dim in a single raw"

f) select "exclude from Viewer"

g) click "Add" button (in middle-right screen) and click OK twice to complete...

you now have generated an active "batch" file that will be your final version after the following completion steps:

h) go back to your multiply imputed file and split it by "ID(cases) choosing "organise output by group"

i) then select "Frequencies" in the "Analyze" tab and select "Mean" and "Mode".

l) click OK

At this stage close you OMS procedure by going back to the "Utilities" tab.

Select "OMS Control Panel" and click on "End All" tab in the up right part of the screen.

click OK ...

At this stage you have generated a new data file with single entry per case and each variable aggregated (across imputations) into a single mean or mode accordingly.

finish up you new SINGLY IMPUTED DATAFRAME by removing irrelevant columns and adding exact label values (per each variable) by copying them from the original file.

check out my clip to ease all the steps actions: https://youtu.be/zftIxv532hE

Is this "Bar Procedure" definitely a secure procedure because I cannot see where the pooling takes place?

Hello Gosse,

thanks for asking.

The Bar Procedure I designed is as secure as any other OMS procedure in SPSS. The pooling is performed (in a batch mode) in your PC by your processor and it stays there. (Alternatively, the pooling could be performed in the SPSS Server if you are a Client accessing it).

Have a look at my You Tube clip for a "hands-on" application of it: https://youtu.be/zftIxv532hE

NOTE: the pooling serves only the purpose to aggregate the 5 (standard) multiply imputed datasets generated by SPSS when the Multiple Imputation procedure is performed.

(the basic aggregating function: mean or mode or median...can be used but obviously you can apply SPSS Syntax or R/Python coding in SPSS and refine your aggregation function accordingly).

Hope this is helps

Daniele

Hi Daniele, I appreciate this a lot. To know sure that the pooling is correctly is for me important. Thanks for explaining. Best regards, Gosse

Hi Daniele Baranzini

I am using the youtube video to pool my multiple imputations, but in your video you seem to get 6 rows (one per participant?) and I only get one column (I have 60 participants). Could you confirm whether there should be one column with a mean per measure that I use across all participants, or whether I should have a column per participant?

Thanks

Julie

I would also like to know about that.

You should get 60 rows so I think you followed the video not completly correct.

Gosse

Julie Evans did you find the solution to your problem. I am also facing the same problem and still looking for solution.

When you have 6 participants you get 6 rows with the pooled data in one column per variable for each participant, thus after pooling the data. After imputation you get for each variable per participant more columns depending on how you set the M. M=the maximum of datasets created in the imputationprocess. I use frequently 20 datasets by imputating missing data (M=20). In pooling the data I get per participant per variable one column for the original data and I get 20 columns extra for the imputated data per variable. After pooling there is one new column per variable. I hope this helps. Gosse

Hi Julie Evans Andleeb Ismail and Charlotte Out

with 60 cases (rows) and one column (your variable) you should still get 60 cases and one column (your variable) at the end of the process.

After applying my procedure in the clip, the column would only contain (for those subjects with missing values) an imputed value (generally an average for scale variables and a mode for categorical) "filling in" the missing value for those cases having missing values.

the imputed value for each single subject is an aggregation function (e.g., averaging) of the n multiple imputed values you asked SPSS to generate ...default is 5 times,... but as you see Gosse Romkes has used 20 imputed datasets : )

Daniele

kinda a graphical representation below for 3 subjects and 1 variable

**original data frame with two missing cases

Var 1

row 1 4

row 2 .

row 3 .

** 3 multiple imputed SPSS dataframes out of the original above...

Var 1

row 1 4

row 2 (2,3)

row 3 (3,8)

Var 1

row 1 4

row 2 (2,1)

row 3 (3,9)

Var 1

row 1 4

row 2 (2)

row 3 (3,6)

**for case "row 2" and "row 3":

row 2 - > (2,3+2,1+2)/3 = imputed value = 2,13

row 3 -> (3,8+3,9+3,6)/3 = imputed value = 3,77

**FINAL dataframe

Var 1

row 1 4

row 2 2,13

row 3 3,77

**END

hope this simplified version would work out

D.

Hi Danielle, I would like to know how to reference your 'bar procedure' in my work, have you written out it in an article? my teachers and supervisor are not familiar with the procedure and they are asking me for a reference so the study can be replicated. Thanks a lot

Ana Caton , It is a technical report (no peer review) I published last year: Data SPSS Single dataframe aggregating SPSS Multiply Imputed split files

I published it here in ResearchGate

let me know

D.

Thanks for this Daniele, very helpful, and for replying so promptly! I will also try to find an article where someone has use it, maybe anyone here?

You can merge data sets, either adding cases (i.e. use the same variables, but add cases) or add variables (i.e. use the same cases but add variables). If you want to add cases, the variables must have the same names and be arranged in the same order in the files.

If your problem is only the fact that the file is split, all you need to do is to switch the "split file" off.

Thanks a lot Daniele Baranzini for sharing this 'bar procedure'. It is working very well. I am just wondering why spss outputs the message that the dataset can only be analysed as MI-Data, if the var imputation_ is available (when splitting the file). This happens with your example as well, no? Do you have an explanation/can i ignore that?