Message/Author 

Jo Brown posted on Wednesday, May 16, 2012  4:47 am



Hi There, I am exploring a simple mediation model and would like to adjust it for confounders (z1  z5) which I believe may influence the c path (X> Y). I am not sure where/how to specify them. Should they be included in all regressions statements or just the Y on X one (as below)? Should they also be specified in the MODEL INDIRECT command as well? Model: Y on m ; Y on X z1 z2 z3 z4 z5; m on X; MODEL INDIRECT: Y IND X; Thanks 

Emil Coman posted on Wednesday, May 16, 2012  11:11 am



I would try to take a stab at this, if I may: Judd & Kenny (2009) list this as the most problematic assumption, that m and Y have no common causes (p. 117), so you are doing what few researcher do (great!). In your case, it seems you want to add : m on z1 z2 z3 z4 z5; And the INDIRECT command should be fine, just make it Y IND m X; INDIRECT is going to estimate the effect flowing from X through m, which will not go through the confounders too. Judd, C., & Kenny, D. (2009). Data Analysis in Social Psychology: Recent and Recurring Issues. In S. T. Fiske, D. T. Gilbert & G. Lindzey (Eds.), Handbook of Social Psychology, Volume One (pp. 113137). Hoboken, NJ: Wiley. Cheers, emil 


I would also include the confounders in the m regression: m on X z1 z2 z3 z4 z5; There is no need to change the IND statement. 

Jo Brown posted on Thursday, May 17, 2012  5:10 am



Thanks for your replies. Just to doublecheck my model should be: Model: Y on m ; Y on X z1 z2 z3 z4 z5; m on X z1 z2 z3 z4 z5; MODEL INDIRECT: Y IND X; Jo 


This is correct. 

Jo Brown posted on Friday, May 18, 2012  7:32 am



Thanks again! 

Jo Brown posted on Wednesday, May 23, 2012  4:06 am



Hi Linda, I set the model as you suggested: Model: Y on m ; Y on X z1 z2 z3 z4 z5; m on X z1 z2 z3 z4 z5; MODEL INDIRECT: Y IND X; and asked standardised estimates using output: sampstat; However, the output for the indirect and direct effect produces estimates of .00 and does not show p value unless I specify the bootstrap option; does this seem reasonable? Jo 


Please send the output and your license number to support@statmodel.com. 


Hi Linda, could you direct me towards a paper/website which explain the difference between using the two mediation commands below and what is the rational for using one over the other: Model: Y on m ; Y on X z1 z2 z3 z4 z5; m on X; and Model: Y on m ; Y on X z1 z2 z3 z4 z5; m on X z1 z2 z3 z4 z5; 


I would use the second case to avoid leaving out an important predictor of m. 

Joseph posted on Monday, June 10, 2013  12:57 pm



Dear Linda, I have a mediation model with measures collected at 3 time points: x1>m2>y3. I recently read the paper by Cole & Maxwell (2003) who suggest that to adjust a mediation model it is necessary to adjust path a for the mediator at baseline (i.e. time 1) to capture the actual change. I am not sure whether it is sufficient to do so or whether I should also adjust the other paths with m1 so that y on x m1 m2; m2 on x m2; Do you have any views on this? thanks! 


This is a good, general question for SEMNET. 

Student posted on Friday, April 10, 2015  8:36 am



I have questions about figuring out exactly which confounding variables to control for in a simple mediation model (X > M > Y): First, I believe you need to determine which variables are associated with BOTH M and Y, and control for those variables. Is this correct? Second, do you also recommend controlling for variables that are associated with X and M (or Y) as well? Third question, what is the best statistical way to determine significant confounding variables? For example, would you regress all of the potential covariates on your M latent variable, then separately on the Y latent variable? Or would you have M and Y in the same model and regress all the covariates on both? In general, is it advisable to test individual or limited groups of covariates (e.g., substance abuse variables only) in these models, instead of all at once? Thank you! 


You ask general modeling questions which may therefore be more suited for SEMNET,but here are some brief answers. Q1. That's an important issue, yes. Q2. Yes, unless X is randomized Q3. You have the term "regressed on" backwards  Y and M are regressed on confounders, not the other way around. The 2 equations can be estimated separately. Q4. No comment. 

Back to top 