I am trying to understand how to use Lisrel to produce the best model for my survey data. I have read through the guides provided by the website but still do not seem to understand the following:
1. How to tell if a path is significant
2. How to interpret the modification indices (I was hoping for a table that would tell me that removing or adding a path or variable will increase my model fit by X like in regression models)
Is there a good free or relatively cheap reference that could help me understand this better?
FYI: I have 16 entry characteristics, 2 scaled scores used to predict two outcome variables. Sample size 212 (yes I know that's a little low for the number of variables I have at the moment)
Thanks for any advice!
Hi,
Your question: How to conduct measured path analysis using Lisrel?
http://userwww.sfsu.edu/efc/classes/biol710/path/SEMwebpage.pdf
http://www3.nd.edu/~rwilliam/stats2/l95b.pdf
http://hcd.illinois.edu/people/faculty/lleras_christy/publications/Path_Analysis.pdf
http://www.mvstats.com/downloads/supplements/sem_basics.pdf
With due respect, I don't think any of the sources you have provided give any information about how to determine whether or not to remove variables or paths from a model in LISREL nor how to determine which paths are significant. Those were just sources that gave an overview of what Path Analysis is.
It has been some time since I used LISREL, lately I have been using AMOS and lavaan in R. But I try to answer your questions nonetheless.
1. Significant paths: similar to regression, LISREL will show standardized or nonstandardized estimates (coefficients) for each path. P-values (statistical significancy) are always given alongside those estimates. You can use p-value to determine statistical significance, like in most inferential statistics techniques. With large sample, you can find that many paths are just significant.
2. Modification indices: This is quite a difficult topic. The safest way to treat this topic is not to do any modification indices at all. But if you have to, then you need to have very good theoretical or empirical justification for the modification indices. One of the most common modification you can do is to correlate residual errors of observed variables within a latent variable. For example: Suppose that you have a latent variable called auditory hallucination, which is measured with 3 items/observed variables: seeing things that don´t exist (A), hearing voices (B), and hearing voices that converse with one another (C). It is probably reasonable to correlate the residuals/error of the two hearing voices items. Adding this correlation paths between the residuals or error means that you assumed that whatever is the error of B is related with C, but not with A. Do you see how problematic and difficult is modification indices? It is because you are arguing for an assumption that something is related even though you do not know what that thing is. This is often acceptable for methodological reasons (e.g. similar measurement method).
Although there are many decisions that are debatable concerning modification indices, there are some decisions that everyone agrees to be bad. For one, it is generally not acceptable to correlate errors (residuals) of observed variables outside its latent variable (e.g. do not correlate errors of items on depression questionnaire with items on anxiety questionnaire). Another one would be to correlate residuals of exogenous latent variable with residuals of endogenous latent variable.
Lastly, a simulation analysis has shown that model fit due to modification indices from paths is not replicable. Maybe you can randomly choose 50% of your dataset (with SPSS) and try to fit your model in that dataset, and then test your model in the other 50% dataset, and lastly in the complete dataset. This technique is called "cross-validation".
Hope it helps! Good luck!
Oh, I also find David Kenny´s website on SEM to be practically helpful: http://davidakenny.net/cm/respec.htm
Hi Bradely
more than a decade since used LISREL having moved over the Mplus
you get the standard error for each path estimate
you calculate the 95% confidence interval, Beta +/- 1.96 std err. If it contains 0 drop the path
the CI, based on standardized parameters also allows you to compare the relative strengths of different paths
hope this helps
- Badges
- Science topic