CausalImpact - Should I use more than one control? In the intro document (https://google.github.io/CausalImpact/CausalImpact.html) it suggests that using one predictor is not ideal. Am I current in understand that they mean one control? If so, should I look into using multiple time series as controls? I was using this method, but summing up all of the control streams into 1 since I am not sure whether using multiple controls is possible. Are there pros and cons of both approachs?
 A: Why there are controls
The CausalImpact package estimates the causal effect of an intervention in terms of the variability seen in the response variable during the post-period that cannot be explained away by other means. This is why CausalImpact rests so critically on control time series (= predictors). In practice, CausalImpact analyses often contain between 3 and 50 predictors.
Each control may help explain away, or regress out, some of the variability seen in your response time series. For example, one control might share the same overall trend as your response variable; another control might share the same day-of-week effect. You can see how well your controls explain your response variable by looking at the pre-period in the plot.
Why you usually don't want to sum your controls
It's generally not a good idea to 'sum' your controls before running CausalImpact. Instead, you want to include each control as an individual time series in your model. You can do this by providing a data frame where the first column contains the time points, the second column contains the response variable, and all further columns contain controls.
Supplying many controls means the model can estimate individual coefficients for each. Summing up controls beforehand would restrict the model to solutions where each predictor is scaled by the same coefficient.
One situation where it may make sense to pre-aggregate controls is when there are too many of them. For example, if you had hundreds or thousands of controls, you might want to reduce their dimensionality using a PCA.
A: Having multiple controls or single controls depends on a specific situation. I would recommend doing both i.e., use single control and as well as multiple controls. CausalImpact is nothing more than a extrapolation/forecasting method that projects data just before a known intervention. So you could infact use any standard forecasting methods and project multiple controls to comapre test.
An alternative method that automatically handles multiple control is Synthetic control model and it has an associated R/Matlab/Stata package.
Hope this helps
