Hi Marleen

I think the problem is "...preferably in SPSS". Multiple mediation models are really easy to specify in packages such as Mplus and they also handle missing data really (although there may be problems using bootstrapping with multiple imputation - but other benefits may outweigh this limitation).

Cheers

mark

Dear Mark, Thanks a lot for taking the time to answer my question! So if i understand you correctly, using SPSS to do a mediation analysis with imputed data is not possible.

In trying to find solutions for dealing with this problem I stumbled upon some literature/fora comments in which either Stata, SAS or the M-Plus or SemTools package in R with were used. As i have only ever worked with SPSS I find it hard to decide which program will be most user-friendly and best suited for this specific analysis. Do you have an opinion on this?

Are there any (practical) guidelines on how to use MI data to do mediation analysis in any of these statistical programs? Any help would be incredibly welcome!

Hi Marleen

> So if i understand you correctly, using SPSS to do a mediation analysis with imputed data is not possible.

If you do a single imputation then I reckon that SPSS?PROCESS will handle this fine. Multiple imputation, I'm not sure.

> As i have only ever worked with SPSS I find it hard to decide which program will be most user-friendly and best suited for this specific analysis. Do you have an opinion on this?

The answer is probably SPSS then.

You could run your data through the missing data package in SPSS, it uses the EM algorithm, to produce a nice 'complete' dataset and then run the mediation models in PROCESS. The PROCESS macro is easily downloaded and installed and uses dialogue boxes so its just a matter of clicking over your predictor, mediator(s), and dependent variable. There's loads of stuff on the web to help with this part.

Mark

Hi Mark,

Thanks again for your (fast) reply!

As i am aware of the advances multiple imputations have over a single imputation / last observation carried forward method, I am hesitant to take this 'step back' (i guess). So i do think using a programm other than SPSS, as you suggested, might benefit the quality of my results. Deciding which strategy to use however does not seem to be so straightforward...

It is good to know that PROCESS is capable of dealing with single imputation and EM, as PROCESS is the macro I've been working with for all my analysis before (ánd the macro I was planning to use in analyzing the MI datasets before I knew this would be quite impossible).

So i guess the question which remains is: if I would consider getting to know another programm to analyze my data (such as Strata/SAS/Mplus/Semtools), are there then reasons to prefer one above the other? And does anyone know of literature/guidelines which explain how to procede with this (MI datasets multiple mediation analysis) in a way which might be understandable for a non-statistician?

Marleen

I use Mplus and can vouch that it is a very flexible and powerful way to test statistical models. Indeed, you could do the mediation model using latent variables. The Mplus discussion board is incredibly helpful and there are lots of videos and online support. Mplus will provide you with different options to deal with missing data.

Mark

Dear Marleen,

one alternative would be to switch to the lavaan package in R. This paper describes how to run a mod-med model in lavaan. In addition, it is easy to run a model with the FIML estimator to treat missing data:

Jeremy N.V Miles Magdalena Kulesza Brett Ewing Regina A Shih Joan S Tucker Elizabeth J D'Amico , (2015),"Moderated

mediation analysis: an illustration using the association of gender with delinquency and mental health", Journal of Criminal

Psychology, Vol. 5 Iss 2 pp. 99 - 123

Best,

Holger

thanks for your replies! I chose to perform it in STATA and i was able to run it here, in case anyone else reading this post is interested :)

Hi Marleen van Pelt

Could you shortly describe which command you have used in STATA? I only know the way with "sem ... ,method(mlmv)". However, as far as I know this is not a multiple imputation method. Therefore, I am really interested in your solution.

Thanks in advance!