Message/Author 


I apologize for the naivete of this question, but I am wondering if I could get some advice and clarity on how to use an instrumental variables approach to dealing with a confounded / endogenous X variable in a structural equations setting? The endogenous X variable is dichotomous, the instruments are continuous, and I would like to use all of the exogenous X variables as predictors of the endogenous X variable as well. Would I be correct in assuming that I can do this by simply writing the aforementioned structure into the model? If so, is this the most efficient way to go about setting up an instrumental variables analysis, or is there a more streamlined way to tell Mplus what it is I am trying to do? As always, thank you very much in advance for your assistance in this matter! 


I suggest posting this on SEMNET or contacting John Antonakis. I believe he has discussed this topic on SEMNET. 


Thank you very much  I was able to get some very helpful information! 

Jon H posted on Thursday, September 17, 2015  9:58 am



I may have asked about this earlier, but I know in many programs the default estimator for an analysis using instrumental variables is twostage least squares. I know Mplus doesn't have 2SLS, but is there an estimator in Mplus you would recommend that would accomplish the same tasks? Any papers on this topic that use Mplus and instrumental variables? 


I don't think we have an estimator that does what 2SLS does. You may want to ask on SEMNET regarding papers on it. 

Jon Heron posted on Friday, September 18, 2015  3:47 am



Hi Jon Stephen Burgess from Oxford has written about the use of SEM for MR analysis. For example: http://ije.oxfordjournals.org/content/early/2014/08/22/ije.dyu176.full best, Jon 

Jon H posted on Friday, September 18, 2015  4:16 am



Jon Thanks for the message! According to them, they used FIML as the estimator. That's news to me. Maybe I should post to SEMNET... Thanks, Jon 

Jon Heron posted on Friday, September 18, 2015  5:34 am



Yes, in a onestep approach. I have managed to replicate a simple Stata ivreg model in Mplus if that might be useful 


From the publications from Antonakis (2010/2014) I have taken, that in order to achive consistent estimates, the estimation of the residual covariance is necessary. I have a model with three latent variables. F1 by x1 x2 x3; F2 by x4 x5 x6; F3 by y1 y2 y3; F3 ON F1 F2; F1 WITH F2; For this model I found strong significant coefficients. If I include two instrumental variables (z1, z2) and the covariance between the residuals then the coefficient are insignificant. F1 F2 ON z1 z2; F3 ON F1 F2; F1 WITH F2; F3 WITH F1 F2; How can this be explained? 


You may want to discuss this on SEMNET where Antonakis participates. 

Back to top 