Forecasting or regression?

[Teaa]

Hello!

I am currently writing a research paper on energy tax impact on individual sector production and I've got quite a problem with my statistical research. Maybe someone will be able to help me?

So I have annual data (2001 - 2012) on production, energy tax tariffs and energy use. My first step was to run a regression - everything worked in this case. Y in my model is production, explanatory variables - energy tax tariffs*energy use.

Now here comes my problem. In 2016 there might be a change in tax tariffs. So I need to model two scenarios:
1) Tax tariffs remain the same as in 2012. So production and energy use will change, but tariffs remain the same.
2) Tax tariffs change in 2016. Again, production and energy use will change and tariffs change from 2016 as well.

And I am really lost in here! How can it be possible to find out what effect each of scenarios will have on production until 2020? To forecast changes in energy use and production, then leave tariffs as they are suposed to be and run a regression on forecasted data (2001 - 2020)? I don't think that it is reasonable.. or is it?

Thanks in advance to anyone who can help as I am REALLY lost in this situation!

 

[noetsi]

Time series is not well designed for cases when there is a fundamental break from the past (which researchers call structural breaks). Hays et el suggested a form of ARIMA and interrupted time series do deal with interventions, but you have to have actual data past the intervention for this to work. If tarrifs stay the same and energy and production behave as they have in the past, which can include a trend, you can model this with exponential smoothing. This assumes of course that the forces that occured in the past continue into the future (so things can change, but they have to change as they have in the recent past).

Regression commonly will generate wrong t tests because of autocorrelation. You can try regression with autoregressive error to adjust for this. That still does not deal with structural breaks, however. You can try some type of sensitivity analysis where you plug in various values for tarrifs (assuming you have a general sense of what the impact will be and what the change in tarrifs will be) in a regression model and see what you get. Note that I have not actually seen this done, it is just a suggestion that makes sense to me