Mediation Analysis - SEM/MEDSEM vs. KHB

[Marcel Wendt]

Hello dear STATALIST forum,

I hope you have started the new year well and can assist me with a problem.
I am currently working on a project in which I am trying to observe the impact of child poverty on work values. My focus is on mediating effects through personality and parenting style.

All my variables are ordinal (quasi-metric) scaled. I have calculated a Structural Equation Model (SEM) with multiple mediators (Big 5 & Supp. Parenting Scale) and interpreted it using the MEDSEM command.

During a presentation in a team meeting, it was suggested that I should try to replicate the same relationships using the KHB method.
The results differ significantly. While occasional mediation effects (according to Baron/Kenny and Zhao, Lynch & Chen) are visible in the SEM model, this is not the case in the KHB Decomposition.

I have the following questions:

• What can account for the differences?
• Is there an error in my SEM or KHB input?
• Which results should I report? Is there a good reason to prefer SEM/MEDSEM results over KHB results?

Thank you in advance!
Best regards,
Marcel

*SEM (Bootstrap & Robust standard Errors)
*Mediation-Effects of Childhood poverty (pgarmut_1_bis_5) via Personality/Parentingstyle (gew ope extr vert neur loc m_par f_par) on Importance of Work-Life-Balance (BW_wl_bala)

bootstrap, reps(1000): sem (BW_wl_bala<-gew ope extr vert neur loc m_par f_par pgarmut_1_bis_5 mpgbilzeit fpgbilzeit fpgexpue mpgexpue migback_re dehhinc_10 bez gejobbt Berufl_Ausb sex) (gew<-pgarmut_1_bis_5) (ope<-pgarmut_1_bis_5) (extr<-pgarmut_1_bis_5) (vert<-pgarmut_1_bis_5) (neur<-pgarmut_1_bis_5) (loc<-pgarmut_1_bis_5) (m_par<-pgarmut_1_bis_5) (f_par<-pgarmut_1_bis_5), nocapslatent vce(robust) 

medsem, indep(pgarmut_1_bis_5) med(loc) dep(BW_wl_bala) mcreps(5000) rit rid zlc


*KHB 
khb ologit BW_wl_bala gew ope extr vert neur loc m_par f_par mpgbilzeit fpgbilzeit fpgexpue mpgexpue migback_re dehhinc_10 bez gejobbt Berufl_Ausb sex || pgarmut_1_bis_5

[Anna Cusimano]

Hello Marcel, have you solved this? I am facing the same issue. Thank you!

[Dirk Enzmann]

Both of you, Marcel Wendt and Anna Cusimano) didn't mention that the programs are user written and available at SSC. And both of you didn't show example data. Additionally, you (Anna Cusimano) didn't show us the syntax you used.

This may have been the reason why you did not receive a response (yet). Please, read the Stata Forum FAQ (especially #12) before posting anything to the Stata Forum, ideally completely! To better understand what medsem and khb are doing I would run a much simplerl model, perhaps using example data from either program and compare the results.

[Felix Bittmann]

To add to Prof. Enzmann's very relevant post, the KHB syntax in post 1 is probably incorrect. The usage should be as follows:

khb ologit depvar treatment || mediator1 mediator2, conc(control1 control2 control3)