Panel Data + Multiple responses from survey (last wave only) - Question on analysis

[Matthew Berg]

Dear Statalist,

I am kindly seeking your advice on how to best analyze a survey (multiple responses) that follows panel data.

I have 290 individuals and their investment portfolios over three years, so panel data on 870 observations and a survey (regarding their investment behavior) following only the third year.

Due to my reshape to long format command, the survey responses have been triplicated and appended to every line of the three years of portfolio information.

My question now centers on how to best analyze these multiple responses.

Some answers to the survey have already been incorporated in the main xt regression models, for example, whether having an investment adviser (asked in the survey) leads to more purchases (ie portfolio inflows) (analyzed through panel data).

However, other answers to the survey have not been incorporated, such as how participants analyze risk in their portfolios, and I feel they yield essential insights into investment behavior.

I frequently saw mrtab (https://ideas.repec.org/c/boc/bocode/s437201.html) (I'm sorry I can't find the ado file - as requested per FAQ for user-written programs) as well as https://www.stata.com/support/faqs/d...ple-responses/ as helpful resources, however, my question is:

Can I analyze the survey completely separately from the panel data since every individual filled out the survey exactly once (at the end of the three years), or will this lead to errors (like an omission of clustered errors at the individual level)? I don't know whether this is important but I verify that their survey responses have been stable for the past three years - e.g., how they analyze risk has not changed over the last 3 years).

Is it a problem that I've already included some answers to the survey in the regression?

I am most grateful for any help or advice!

Thank you very much, and kind regards,

Matt Berg

[Clyde Schechter]

I'm confused by your description of the data. At first, I had the impression that each investor was surveyed only once, at the end, and that you just have three identical copies of the data in your panel data set, the result perhaps of merging the survey data set with the portfolio data set. But then you speak of verifying that the survey response have been stable over three years--which doesn't make any sense unless they were actually surveyed in all three years. So which is it? Were they surveyed once only, or all three times?

Assuming they were only surveyed once, you can certainly analyze the panel data separately. To do that, put the survey variables and the investor id variable(s) into a separate data set and then keep only one observation per investor. Forget about using panel data analysis commands: this is no longer panel data. Forget also about clustered standard errors--with one observation per investor there is no possibility of within panel heteroskedasticity or autocorrelation, as there are no panels! Just use ordinary commands that you would use to analyze a single cross section, because a single cross section is what you have.

[Matthew Berg]

Dear Clyde,

Thank you very much for your comments. Yes, every investor was only surveyed once at the end of the third year. We have portfolio data on years 1 and 2 but no accompanying surveys for those two years. For the third year, we have portfolio data (same individuals, same variables) and, in addition, the qualitative survey regarding their investment behavior - therefore the questions asking if their behavior has changed over the past three years.

I will create a separate dataset to only analyze the qualitative survey using cross-sectional commands.

I appreciate all your help!

Kind regards,
Matt