Hello all,
I am hoping someone can advise me on the preferred approach for testing for a mediating variable in panel data, if any?
I know that mediation analysis is hotly-debated for cross-sectional data, so I am circumspect about potentially wading into this fray with a data structure that might be even more problematic. But, thought I would post here to see if anyone has any insights to share.
I have balanced panel data for 110 weeks. Using VAR, Granger Causality tests, and IRFs, it appears as though there MIGHT be a mediating role played by one of my variables in the relationship between two other variables. Separate VAR tests reveal that x Granger causes m (2 lags optimal) and m Granger causes y (2 lags optimal). A third test reveals that x granger causes y (4 lags optimal). Fwiw, this relationship is also supported by the theoretical framework.
Super-imposing the IRfs for these three relationships, for purely illustration's sake, further supports (at least in temporal terms) a potential mediating role of m between x and y. (Please see hastily made Excel plot of IRFs for these three separate VARs.)
But, obviously, if it is possible to conduct an actual mediation test, that would be helpful. I have done some looking around, and one of the issues I am encountering is how to best incorporate the optimal lags into such an analysis. Moreover, since mediation analysis is already so debated, I am not sure wading into this with panel data and lags is really a good idea. If not, maybe an illustration of a potential mediating role of m between x and y is sufficient for the purpose of my study?
I also want to note that I am well aware that Granger tests (and IRFs) do not test for causality in the theoretical sense, nor is that my intention. Predictive power, forecasting, and temporal correlations are sufficient for my needs. But, even with this lower threshold, what does mediation in a forecasting context really look like? How does one "best" test for it, and what assumptions/conditions about the data necessarily accompany these tests? (And, what relevant Stata packages are available, if any?)
Thank you in advance for anyone who has insight to share on this!
I am hoping someone can advise me on the preferred approach for testing for a mediating variable in panel data, if any?
I know that mediation analysis is hotly-debated for cross-sectional data, so I am circumspect about potentially wading into this fray with a data structure that might be even more problematic. But, thought I would post here to see if anyone has any insights to share.
I have balanced panel data for 110 weeks. Using VAR, Granger Causality tests, and IRFs, it appears as though there MIGHT be a mediating role played by one of my variables in the relationship between two other variables. Separate VAR tests reveal that x Granger causes m (2 lags optimal) and m Granger causes y (2 lags optimal). A third test reveals that x granger causes y (4 lags optimal). Fwiw, this relationship is also supported by the theoretical framework.
Super-imposing the IRfs for these three relationships, for purely illustration's sake, further supports (at least in temporal terms) a potential mediating role of m between x and y. (Please see hastily made Excel plot of IRFs for these three separate VARs.)
But, obviously, if it is possible to conduct an actual mediation test, that would be helpful. I have done some looking around, and one of the issues I am encountering is how to best incorporate the optimal lags into such an analysis. Moreover, since mediation analysis is already so debated, I am not sure wading into this with panel data and lags is really a good idea. If not, maybe an illustration of a potential mediating role of m between x and y is sufficient for the purpose of my study?
I also want to note that I am well aware that Granger tests (and IRFs) do not test for causality in the theoretical sense, nor is that my intention. Predictive power, forecasting, and temporal correlations are sufficient for my needs. But, even with this lower threshold, what does mediation in a forecasting context really look like? How does one "best" test for it, and what assumptions/conditions about the data necessarily accompany these tests? (And, what relevant Stata packages are available, if any?)
Thank you in advance for anyone who has insight to share on this!
Attached Files
