I am new to the CausalImpact package in R and wanted to test the effect of a marketing intervention.
Now, if we assume that a specific intervention only took place in 2013, is it possible to use time series from 2012, 2011 and 2010 to serve as the control time series? Or would it be better to use specific portions of the time series (let's say +/- 8 weeks from the intervention) and to control for any differences during other years?
I believe the control time series needs to be contemporaneous, but not affected by the intervention. Examples are clicks in non-affected markets or clicks on other sites (where there's no intervention).
Quoting from p. 254 of the paper:
Control time series that received no treatment are critical to our method for obtaining accurate counterfactual predictions since they account for variance components that are shared by the series, including, in particular, the effects of other unobserved causes otherwise unaccounted for by the model.... All covariates are assumed to be contemporaneous
This means that you cannot use other periods of your outcome time series as a control for two reasons. First, doing that cuts into the data that you have to learn how the control time series maps into your target outcome time series. More importantly, once you know that mapping, you use the control time series to capture the unobserved shocks that are shared by the target time series to form your counterfactual. Using previous years' data will not accomplish this, unless the shocks are the same every year.
