I'm working with project-level data (one observation per project) spanning roughly 20 years. Each project has a specific announcement date, which I aggregate to the month of announcement. The dependent variable is the (logged) capital efficiency of the project. The independent variables are macroeconomic conditions (e.g., the ECB refinancing rate), which vary monthly and are shared across projects announced in the same month.
Augmented Dickey-Fuller tests suggest that several of these macro variables are non-stationary (I(1)), so I plan to include them in first differences. However, the dependent variable is not a time series — it is observed only once per project and does not track the same units over time.
My question is:
Is it methodologically valid to regress a project-level outcome (in levels) on first-differenced macroeconomic variables, given that the dependent variable is not a time series and therefore not subject to unit root behavior in the traditional sense?
I do not wish to collapse the data into a time series and estimate a full first-difference model on monthly averages, as that would change the level of analysis and the research question. However, I have encountered conflicting opinions on whether such a “hybrid” model is valid or whether it violates assumptions by producing non-stationary residuals. I would be very grateful for your guidance on how to correctly justify this structure.
Thank you!
Augmented Dickey-Fuller tests suggest that several of these macro variables are non-stationary (I(1)), so I plan to include them in first differences. However, the dependent variable is not a time series — it is observed only once per project and does not track the same units over time.
My question is:
Is it methodologically valid to regress a project-level outcome (in levels) on first-differenced macroeconomic variables, given that the dependent variable is not a time series and therefore not subject to unit root behavior in the traditional sense?
I do not wish to collapse the data into a time series and estimate a full first-difference model on monthly averages, as that would change the level of analysis and the research question. However, I have encountered conflicting opinions on whether such a “hybrid” model is valid or whether it violates assumptions by producing non-stationary residuals. I would be very grateful for your guidance on how to correctly justify this structure.
Thank you!
