Tom,

In the ideal world of no missing data and perfectly true models, it makes no difference.

That said, including both endogenous variables in a single model can reduce bias in estimation, as well making it possible to detect flaws in the measurement model involving the endogenous variable indicators that can not be seen in separate models.

The bias points are twofold. At least for maximum likelihood estimation, information from all parts of the model can affect all parameter estimates. So including only one endogenous variable at a time can lead to differences in your estimates involving the exogenous variables. Also, with missing data, including more variables can only help the missing-at-random assumption.

Hope this helps,

Pat

Tom,

In the ideal world of no missing data and perfectly true models, it makes no difference.

That said, including both endogenous variables in a single model can reduce bias in estimation, as well making it possible to detect flaws in the measurement model involving the endogenous variable indicators that can not be seen in separate models.

The bias points are twofold. At least for maximum likelihood estimation, information from all parts of the model can affect all parameter estimates. So including only one endogenous variable at a time can lead to differences in your estimates involving the exogenous variables. Also, with missing data, including more variables can only help the missing-at-random assumption.

Hope this helps,

Pat

Thanks Patrick (and apologies for the belated reply)

In the end, I had two run two separate models as it was the only means I could justify to make the larger, overall model more parsimonious (I had quite a lot of indicators as well, even after parcelling). However, for future (better!) data sets I'm sure your advice will serve me well.

Cheers

Tom

Good to hear!