multiple questions

Welcome to the forum for REALCOM users. Feel free to post your question about REALCOM here. The Centre for Multilevel Modelling take no responsibility for the accuracy of these posts, we are unable to monitor them closely. Do go ahead and post your question and thank you in advance if you find the time to post any answers!

Go REALCOM (Developing multilevel models for REAListically COMplex social science data) >> http://www.bristol.ac.uk/cmm/software/realcom/
Post Reply
shanekav
Posts: 36
Joined: Wed Feb 20, 2013 12:55 am

multiple questions

Post by shanekav »

Hi,
I have some questions that I hope someone can answer for me. Apologies in advance for the quantity. :oops:

I have realcom up and working with mlwin. I am running a 2 level model on cross sectional data. My exposure of interest is a level 2 variable measured using different approaches. Thus, I would like to run the model separately with each of these analogous level 2 variables. I would also like to look at the interaction between these exposure measures and another level 2 variable. My questions are as follows:

I am unclear on the process for utilising realcom in the above situation. Do I create an imputed file using all of the different exposure variables and then construct and run each model and then subsequently run the model again using the same imputed file? Or, do I need to create a separate imputed file for each model?

If I have run the analysis with the imputed file and I would like to adjust the model (to for example look for a cross level interaction) do I adjust it and run the model and then run the model with the imputed file each time I make an adjustment?

If I have set up a model that includes dummy variables, when I set up the file to be exported to realcom do I include only the ‘base’ variable that the dummys are constructed from?

I would like to ‘centre’ some of my variables. Do I need to include the ‘centred’ versions in the file that is exported to realcom, or can I use the ‘base’ variable?

I am unclear how to account for interaction terms in the imputation? Do I need to include anything specific in the file to be exported to realcom and do anything special in realcom?

Lastly, I have run the model with and without the imputed realcom file. In the case without the imputed file the mlwin output reports a number of individuals that is less than the number in the data set. This number roughly corresponds to the number I would expect if the cases with missing data are dropped. I have subsequently run the model with the imputation file from realcom. The number of individuals shown in the output does not change (though the beta co-efficients and the standard errors do). I don’t understand this. My expectation is that because the missing variables have been imputed that the number of individuals should now be the same as the full number of cases in the data set.

Many thanks in advance

Shane
Harvey Goldstein
Posts: 49
Joined: Sun Sep 06, 2009 5:30 pm

Re: multiple questions

Post by Harvey Goldstein »

Shane
A number of interesting questions!
You could include all the level 2 differently defined variables in 1 imputation run. However, MLwin is only currently set up to import the imputed datasets back into the model originally set up in mlwin so this isn't going to work very neatly - I suggest you do imputation for each one separately.
Regarding categorical variables, make sure you define these in the appropriate mlwin menu: the variable to go to realcom is then the original (coded) one and this will be correctly interpreted and new dummies created when the impvals file is read back to mlwin.
Interactions are currently a problem with all multiple imputation techniques. The best you can probbaly do is treat an interaction (or power term) that has missing data as just another variable. A new paper that gives an unbiased way to handle such variables will appear in JRSSA later this year and we shall be looking to incorporate the method.
Hope this helps.
Harvey Goldstein
shanekav
Posts: 36
Joined: Wed Feb 20, 2013 12:55 am

Re: multiple questions

Post by shanekav »

Hi Harvey,
Thanks very much for explaining that. I would like to also clarify the last issue as to why the number of individual level units in the analysis remains the same when using the imputed file. Should it not increase to the full number in the data set as no cases now having any missing variables?

Shane
ChrisCharlton
Posts: 1351
Joined: Mon Oct 19, 2009 10:34 am

Re: multiple questions

Post by ChrisCharlton »

The number of cases displayed in the equations window (and the hierarchy viewer) is based on the data currently loaded into memory. During the imputation analysis run this is replaced by each of your imputed datasets and the combined results are put into C1096...C1099 so that they appear in the equations window. At the end of the analysis the original data (including missing) is restored, so this is what will be used to display the number of cases in the equation window. Sorry if this is a bit misleading.
shanekav
Posts: 36
Joined: Wed Feb 20, 2013 12:55 am

Re: multiple questions

Post by shanekav »

Hi Chris,
Thanks for clarifying that for me. I was afraid that I was doing something wrong :?

Shane
Post Reply