Mediation model with longitudinal data

[Yule Hai]

Hi, I am trying to analyze longitudinal data using mediation model. The unit of observation is province. There are 34 provinces. The data runs for 30 years. I have one dependent variable, two mediation variables, and one independent variable, along with some covariables. I want to run a partial mediation analysis, so that the impact of my independent variable on the dependent variable tends to be reduced through the mediation variables. I am using STATA13. I understand that "gsem" is the command language. But I am not clear whether I simply should treat year as the second level, province as the first level. Viewed in this way, the data is two level. But I feel somewhat uncomforable because two level usually refers to clustered data, such as pupils are clustered among schools. So I wonder if there is anything more specfic to deal with the time dimension in longitudinal data as in my case for mediation models. Your advice would be highly appreciated. Thank you very much.

[wbuchanan]

You have several different issues that you are trying to deal with here. In terms of the description of the structure of your data, you have 30 (assuming annual) observations of > 4 variables clustered in 34 provinces (this is the last part that you were talking about). You don't describe the level at which the variables are measured and/or vary (e.g., are all your data time varying or are some time invariant?). Usually, longitudinal mediation models are fitted to data in wide format (e.g., one variable per variable * time pair). This, however, is not feasible with the data you've described.

The clarify a few other things in your post:
(1) It is Stata NOT STATA (note the spelling/capitalization in the banner at the top of the page "The Stata Forum").
(2) gsem is a command in Stata; it is not a command language.
(3) In the context of a hierarchical model, time is typically clustered within the observational units.
(4) To give you any advice beyond this, you would need to provide a bit more information about your data, model, and what you are trying to test.

[Yule Hai]

Thank you for getting back to me. I appreciate it. The data is indeed annual information of provinces' GDP growth. This is the dependent variable. The independent variable and mediation variables are all measured at the province level on an annual basis. If it were a one year data, then there would be 34 observations, in which I would like to understand how GDP growth is mediated by the two mediation variables from the independent variable. But my data runs 3 decades, so I want to, at a minimum, control for time specific effects. I am thinking of splitting the sample into three periods, where the country (it is China's data) had three different leaders (the three leaders had different ideologies, which influence their economic policies). This is one (not very good) way that I may be able to show potential time specfic impact, if there is any. But in the end, I still wish to integrate the time dimension into the model. I hope this clarifies the question a bit better. Thanks a lot for taking your time helping me. If you have any questions, please let me know. Thanks very much.