PDA

View Full Version : comparing path coefficients

yhchang
11-17-2011, 08:09 PM
Hi all,
I've got a question when conducting a path analysis.

I've used one sample group and created three different path analysis models (which are mostly identical, except for one mediator.)

To make my questions clear, the three models are illustrated below:

1)

a---> b1--->c (dependent variable=continuous)

2)

a---> b2--->c

3)
a---> b3--->c

Now, my question is, after running the the 3 models respectively, how can I test the significance of the differences in coefficients of b1-->c, b2-->c, and b3-->c?

best,
Clara

noetsi
11-18-2011, 08:31 AM
There are no "dependent" variables in path analysis. There are endogeous and exogenous variables.

You would not normally determine if the direct effects in three models were statistically different. As far as I know there is no way to do this. Instead you evaluate paramaters by looking at the model fit indices such as Chi square RMSEA, TLI etc. If one model fits the data and one does not, or if one signficantly improves the indicators that suggest that what you specified in that model is better. But you can't be sure any one element did this, unless this is the only change you made.

You can compare the standardized direct effect to see if one is larger, although this is not a statistical test.

yhchang
11-20-2011, 10:03 PM
There are no "dependent" variables in path analysis. There are endogeous and exogenous variables.

You would not normally determine if the direct effects in three models were statistically different. As far as I know there is no way to do this. Instead you evaluate paramaters by looking at the model fit indices such as Chi square RMSEA, TLI etc. If one model fits the data and one does not, or if one signficantly improves the indicators that suggest that what you specified in that model is better. But you can't be sure any one element did this, unless this is the only change you made.

You can compare the standardized direct effect to see if one is larger, although this is not a statistical test.

Lazar
11-20-2011, 10:17 PM
This is somewhat problematic if the models are not nested. I think that comparing the AIC and BIC would also be reasonable given that both are used to compare the fit of non-nested models.

In passing, why not include all paths in a single model if you have the sample size to warrant that?