Mediation analysis after multiple imputation

#1
Hi, after multiple imputation how do i do a mediation analysis?

Do i simply obtain the product term using the pooled coefficient for X--> M (a) and M--> Y (b) and its std. error from the separate linear regression analysis to calculate using the MacKinnon (2002) method?

I was not able to find any information from the internet, and needed some assurance that it is ok to proceed...

please enlighten me, thank you
 

spunky

Can't make spagetti
#2
Do i simply obtain the product term using the pooled coefficient for X--> M (a) and M--> Y (b) and its std. error from the separate linear regression analysis to calculate using the MacKinnon (2002) method?
no, you don't. what you need to do is fit the mediation model to each multiply-imputed dataset. you save the coefficients and standard errors for each coefficient and you combine them according to Rubin's rules. if you just do what you're suggesting you're underestimating the variance of the parameter estimates and increase your type 1 error rate.

SPSS will not do this for you (and combining the standard errors can get tricky). but if you use Mplus or the semTools package in R, it will do it for you so you don't have to worry about averaging things correctly.
 
#3
Gasp!!!!!!

Thank You Spunky!
Just when i thought one problem was solved.. now another 3 problem pops up..I have problems for obtaining the cronbach alphas of the scales after doing multiple imputation.

Anyways, if i am to do it in R, do i download the semTools package?
and if import my data from SPSS into R, do i import the data set with missing data or the imputed data set.

I actually have not used R before and i will learn them now. Please be patient with me
 

spunky

Can't make spagetti
#4
Anyways, if i am to do it in R, do i download the semTools package?
and if import my data from SPSS into R, do i import the data set with missing data or the imputed data set.
you'd impute the dataset with missing data on it. semTools will take care of the multiple imputations, run the analysis, combine the results and report them back. i'm afraid now that i read you say "import the data set" as in a singular data set. you... do know you're supposed to create several data sets and you're not supposed to average across them... right?

anyhoo... semTools is based on an SEM framework. you can fit a one-factor model (Cronbach's alpha assumes one, after all) and calculate the more correct index of reliability rho (or omega.. it was many names). if you REALLY like Cronbach's alpha you can obtain it from the resulting covariance matrix after the imputations are done, since all you need to get alpha is the covariance matrix of the items.
 
#5
I C!!

thats enlightening!!

Cool. I will go figure out R for this purpose.
I thought of trying EM as it is easier, but my dataset are not MCAR for sure, so i cannot use EM. R seems like the best option.

I will devote my time this week to figure it out.
Stanley, do you have any information or readings to recommend me with regards to these two topics (Mediation and scale reliability with multiple imputed data)

and thank you very much. you are a great help. :tup::tup:
 
#8
Hi there,

I too am having to do mediation analysis with MI data. I have run MI in SPSS using fully conditional specification and get pooled coefficients and standard errors output with my linear regression. Based on the discussion above, I am not clear if it is ok to use these pooled estimates or not? If so it would be easy to implement them in the Rmediation package.
many thanks