www.cmm.bristol.ac.uk/forum

Hi. Please allow me further question.

I would like to reduce the time of imputation as my dataset is more than n=100,000 and it takes very very long (around 20 minutes) to do first iteration.
I don't know what this MCMC estimation settings, so so far I have been using default setting.

Is there a good way to reduce imputation time by changing setting?
If not, I may think about to reduce the number of imputed sheet, say, to 5. How can I change the MCMC setting to get 5 sheets?

I do appreciate your help.

Best wishes,
Ayako

In section 3.1.3, Carpenter et al. (2011) (http://missingdata.lshtm.ac.uk/preprint ... er2011.pdf) provide a description of the MCMC estimation settings and make some recommendations on what values to use. I wouldn't be able to comment on the extent to which reducing the burn-in or the gap between imputations will have an effect on your imputed values from a statistical point of view though.

Thank you for the information.

I am not familiar with statistical settings, but when I changed MCMC estimation settings to 10 (Burn-in length), 100 (Number of iterations), and 10(Scree refresh rate), it imputed quite faster than previously.
I will consult to people about the ideal settings, but for test-runs, this 10-100-10 seemed to work.

Thank you so much.

Best wishes,
Ayako

One of the issues with REALCOM is that it is slow - we are planning to do something about this with new software in the pipeline but this will not be until well into 2012. Reducing teh number of iterations will certainly shorten the time taken but with only a short burn in you may have poor mcmc chains and the imputations could be unreliable. certainly look at the chains you get. You also need imputes sufficiently far apart in the chain to ensure independence. See the multiple imputation website at the London School of Hygiene for more discussion.
Harvey Goldstein