SEM with Cross Sectional-Time Series ...
Message/Author
 EZalberta posted on Tuesday, November 23, 2010 - 5:55 pm
Hi,
I am pretty new to Mplus and to structural equation modeling (SEM) in general. I have a dataset with cross sectional-time series data structure: a list of variables for 170 countries measured across 20 years.

These variables include: dependent variable, which is a count variable; and variables which are supposed to be indicators of 4 latent variables; and finally a set of control variables.

I want to estimate both the independent and interaction effects of the 4 latent variables on the count variable, controlling for the control variables. I wonder whether this type of analysis could be done using structural equation modeling, which type of SEM I should use, and whether MPLUS can achieve this.

 Linda K. Muthen posted on Wednesday, November 24, 2010 - 9:24 am
It sounds like you want to regress a count variable on a set of observed and latent exogenous variables and their interactions. That can be done in Mplus. Each interaction between two latent variables or a latent variable and an observed variable requires one dimension of integration. A model with more than four dimensions of integration is not recommended.
 EZalberta posted on Wednesday, November 24, 2010 - 11:09 am
Thanks Linda. This is very helpful!

The interactions in my estimation will be mostly two-way interactions among the 4 latent variables, which means theoretically there can be as many as 6 two-way interactions. Can MPLUS accommodate this many interactions?

Also panel data regression can be estimated by fix-effects and random effects models. If I want to estimate three models: country-level fixed effect model, country and year two-way fixed effects model, and random effect model; can MPLUS do all three? And can you refer me to some example syntax?

Thank you!
 EZalberta posted on Wednesday, November 24, 2010 - 11:23 am
And both my independent variables and dependent variables are measured on 20 time points. I understand for panel data with only a few waves, where for example, independent variable is measured in wave 1 and dependent variables measured in wave 2, SEM can be easily estimated.

But for my case, both independent variables and dependent variables are measured in 20 waves, where I hope to use independent variables measured in time t-1 to estimate dependent variable in time t, how shall I specify the model in MPLUS? If there is no latent variables, I can simply use xtpoisson in stata to do the job, the existence of latent variables just complicates the estimation. Hope MPLUS is more powerful on this type of analysis.

Thanks!
 Linda K. Muthen posted on Thursday, November 25, 2010 - 9:29 am
A model with more than four dimensions of integration is not recommended. This would be four latent variable interactions. It is not likely that all are signficant so I would investigate that and include on the ones that are significant.

See the examples chapters of the user's guide.

You can use the ON option as follows:

y2-y20 PON x1-x19;
 Thomas Eagle posted on Tuesday, October 14, 2014 - 4:11 pm
Linda and Bengt,

I have run across a very interesting paper that purports to model time series cross sectional data as a multi-level SEM. What makes this stand out is that the data for each time period is cross sectional. That is, a unique sample of respondents is collected for each time period without any overlap of Rs. The link is below.

Can I do anything similar to this in MPLUS?

Thanks,

Tom
 Tihomir Asparouhov posted on Wednesday, October 15, 2014 - 9:11 am
Tom

This estimation method is not available yet in Mplus. Some time series possibilities are discussed in

http://mplus.fss.uu.nl/files/2012/09/Mplus-Users-Meeting-2012-Tihomir-Asparouhov.pdf
 Thomas Eagle posted on Wednesday, October 15, 2014 - 5:15 pm
Thank you, Tihomir. I'll look the slides over.

Tom