www.cmm.bristol.ac.uk/forum

Hello,
I am just getting started with multiple imputation and so as with MlwiN and RealCom.
I followed the tutorial example given in the RealCom manual, but things don't go exactly as they should (at least, my problems are not mentioned in the tutorial).
First, once I start the imputation analysis, I always get the message: "the columns below contain numerical data that cannot be precisely stored in the single precision format used by MLwiN (...)", asking if I want to convert them in categorical variables. I get this message as many times as my number of imputations, so it is a bit annoying... Is there anything I could do to avoid that?
Even though I get pretty much the same estimates than in the tutorial, another thing that I don't get is why it always says "loglikelihood not available", whereas in the tutorial you should get a value.
Thank you in advance for your help!

Unfortunately the only ways to avoid the single precision message would be one of the following:

• Reduce the precision in the output data files.
• Process the imputation output in Stata using the -RealcomImpute- (http://missingdata.lshtm.ac.uk/index.ph ... Itemid=102) command.
• Write an MLwiN macro to process the imputation and call it with mlnscript.

I believe that the reason that you don't get a likelihood for you model of interest is that these estimates are created from combining a number of different model runs (one for each imputed data set) and combining the likelihoods in the same way wouldn't make sense.

Thanks a lot for answering so quickly.
I guess in the tutorial only a single imputation was made.
One more question: do you have any idea why my imputation specification goes well with the menu, but not with the command lines? The ISAVE and IRETRIEVE commands produce error messages (Li| type).
Thanks again

I'm not sure why that would be, however the menus should generate the appropriate commands. You can check this by opening the Data Manipulation->Command Interface window and unticking "user". When you use the menus the generated commands should appear in this window, allowing you to compare them with the version that you are using.

Hi,

I am following the same Realcom Impute tutorial and have the same problem.

I believe the tutorial uses 5 imputations and I do get the same errors.
2loglikelihood is unavailable whereas in the tutorial it has a value!

My coefficient values are different from the tutorial as well. Any idea why that is the case?
Many thanks in advance!

Zahra

The screenshot in the tutorial is from an older version of MLwiN. The likelihood in the tutorial will just be the value from the last imputed model of interest, you will also notice that the "cases in use" matches the original data whereas the models run on the imputed data in fact use all cases in the data. Later versions of MLwiN remove this line for these reasons. I am not sure why you are getting different coefficient values, are they very different?

Hi Chris,

Thanks for your reply
Now I know why there is a loglikelihood value in the tutorial!
My values are not too different from the tutorial values but then again tutorial values are not so different from before imputation values, so I am not sure.
I have attached the models before imputation, after imputation ( tutorial values) and my work after imputation respectively. Do you think my values show that I have performed multiple imputation correctly?


Many thanks!
Zahra

Running through the tutorial I get the same answers as you using the most recent version of Realcom-Impute, so I suspect that the differences are just due to updates to Realcom-Impute since the tutorial document was written.

Hello,

Since the value for loglikelihood is not available after performing multiple imputation using Realcom, how can one comment on model fit of the imputed data?

Many thanks,
Zahra

Hi all,

My question is how to compare goodness of fit of nested models when log likelihood is not available (after multiple imputation using Realcom)

Many thanks

Zahra