I need help on the next steps after a continuous x continuous interaction with imputed data

[Jide Bamishigbin]

I imputed 20 datasets of a sample with 2,775 people and I created the interaction term prior to the imputation. After running my full model, I found a significant interaction for that imputed variable.
Here is my syntax

mi estimate: reg T3CESD Age Female T2Married T1Income40K T1College T1AdvancedCancer T1NumberComorbidPhysConditions T1POMS_DP T1FACIT_M T1FACIT_P T1FACIT_F T1FACITMxPInteraction T1FACIT_MxFInteraction T1FACIT_PxFInteraction

Here is my output:

What are my next steps to interpreting this interaction? I have recoded my variables into tertiles in an attempt to look at differences at different levels however I get constant error messages.

Here is the syntax I tried

mi estimate, cmdok: reg T3CESD i.T1FaithCat Age Female T2Married T1Income40K T1College T1AdvancedCancer T1NumberComorbidPhysConditions T1POMS_DP T1FACIT_M T1FACIT_P T1FACIT_F T1FACITMxPInteraction T1FACIT_MxFInteraction T1FACIT_PxFInteraction, if T1MeaningCat==1

But the error message I get every time is

"estimation sample varies between m=1 and m=2; click here for details"

I look forward to any help or assistance you may be able to provide.

[Clyde Schechter]

If T1MeaningCat is one of your imputed variables (or is passive), then the observations that satisfy the -if- clause of your command will vary from one imputation to the next, and that will trigger the error message you have gotten. You can suppress that error message by adding the -esampvaryok- option to your -mi estimate- command prefix. I should add that I am not deeply grounded enough in the theory of multiple imputation to be sure that MI actually achieves its intended purpose of reducing bias due to missing data when you do this. That is, I'm not certain whether Rubin's rules are applicable in the situation where membership in the estimation sample is itself conditioned on an imputed (or passive) variable.

If, however, T1MeaningCat is neither imputed nor passive, then there is some kind of error in your data, perhaps owing to incomplete imputation of all the model variables. You would have to use -mi xeq:- to run the regression on the two samples that Stata told you produced differences and see where the differences are coming from in order to know how to fix it.

[Jide Bamishigbin]

Hello Clyde,

Thank you fr your response. The T1MeaningCat variable essentially just turns the T1FACIT_M variable which is a continuous into a categorical variable with three levels (low, moderate, high). I am going to try the esampvaryok option right now!

[Clyde Schechter]

Yes, but the key issue here is whether that is a "regular" variable (i.e. it was never missing in the pre-imputation data) or whether you imputed it (or calculated it as a passive variable from T1FACIT_M, the latter being, itself, an imputed variable). If it is imputed (or passive) then using -esampvaryok- may be a solution to your problem (although, as I said in #1, I'm not sure that Rubin's rules are applicable in this circumstance). But if T1FACIT_M is a "regular" variable then something is wrong with the data, and just covering that up with -esampvaryok- is not a good idea.

[Jide Bamishigbin]

Hello Clyde,

Ah, thank you. Yes, T1FACITM is an imputed variable and esampvaryok made it work! I will check to see if Rubin's rules are applicable but I appreciate your help.