Dear fellow members of Statalist,
I hope every one of you has been well and healthy during this disruptive time and that we will soon be at the other end of this.
My question relates to defining two residuals to correlate (allowing the correlation to be a free parameter) in an SEM model which seems to lead to a convergence failure when I try to run multi-group analysis (MGA) on this model. I am looking at the effect of two exposure (stigma, discrimination) on physical health and the mediating effect of mental health on their effects. Stigma and physical health are the 2 latent variables, both having 3 indicators. The other factors (discrimination, mental health) are themselves observed variables.
Model 1 is the proposed model without a residual correlation:
Modification indices for this model suggested that adding a correlation between the residuals of health1 and health2 will greatly improves model fit. I re-specified accordingly to create model 2:
Model 2 fits much better and is thus retained.
The problem is, when I proceeded to do multi-group analysis for model 2 by gender (mcoef and mcons set to be equal across groups), it did not converge. Removing the correlation between the said two residuals makes MGA successfully made the model to converge to a solution. If I remove the said correlation add other residual correlations, the model still converged. So it seems that it is this particular specification of a correlation between the residuals of health1 and health2 that STATA does not like. I am having trouble understanding why and would be very grateful if anyone is willing to offer some insight?
All the bests,
I hope every one of you has been well and healthy during this disruptive time and that we will soon be at the other end of this.
My question relates to defining two residuals to correlate (allowing the correlation to be a free parameter) in an SEM model which seems to lead to a convergence failure when I try to run multi-group analysis (MGA) on this model. I am looking at the effect of two exposure (stigma, discrimination) on physical health and the mediating effect of mental health on their effects. Stigma and physical health are the 2 latent variables, both having 3 indicators. The other factors (discrimination, mental health) are themselves observed variables.
Model 1 is the proposed model without a residual correlation:
Code:
sem (Stigma -> stigma1, ) (Stigma -> stigma2, ) (Stigma -> stigma3, ) (Stigma -> Mental, ) (Stigma -> Physical, ) (Discriminate -> Mental, ) (Discriminate -> Physical, ) (Mental -> Physical, ) (Physical -> helath1, ) (Physical -> health2, ) (Physical -> health3, ), covstruct(_lexogenous, diagonal) cov(_lexogenous*_oexogenous@0) standardized latent(Stigma Physical) cov( Discriminate*Stigmal) nocapslatent
Code:
sem (Stigma -> stigma1, ) (Stigma -> stigma2, ) (Stigma -> stigma3, ) (Stigma -> Mental, ) (Stigma -> Physical, ) (Discriminate -> Mental, ) (Discriminate -> Physical, ) (Mental -> Physical, ) (Physical -> helath1, ) (Physical -> health2, ) (Physical -> health3, ), covstruct(_lexogenous, diagonal) cov(_lexogenous*_oexogenous@0) standardized latent(Stigma Physical) cov( Discriminate*Stigma e.health1*e.health2) nocapslatent
The problem is, when I proceeded to do multi-group analysis for model 2 by gender (mcoef and mcons set to be equal across groups), it did not converge. Removing the correlation between the said two residuals makes MGA successfully made the model to converge to a solution. If I remove the said correlation add other residual correlations, the model still converged. So it seems that it is this particular specification of a correlation between the residuals of health1 and health2 that STATA does not like. I am having trouble understanding why and would be very grateful if anyone is willing to offer some insight?
All the bests,
