# CausalImpact Vs Synth

Can CausalImpact package be used in lieu of the Synth package to create a synthetic control ? The R implementation of the Synth package is very confusing compared to the Stata demo for the Synth function.

• I think it would be better if the question would be rephrased in terms of methods which are implemented by the two mentioned packages. As it stands now it looks like a software question. – mpiktas Oct 28 '15 at 6:43
• They are two different and distint methodologies that can be used to model causality in observational data. yes both can be used interchangeably. – forecaster Oct 28 '15 at 11:28
• Does anyone have an example code for reference on CausalImpact ? I somehow concluded that you do need a control series to use the package, Isn't the purpose of the package to avoid having to have a control series ? Alternatively, may I consider the pre period time series as the control aka x1 ? – user3386377 Oct 28 '15 at 13:50
• The control time series in the CausalImpact package are really regressors: variables that are correlated with your outcome variable and that were not themselves affected by the intervention. Thus, the control time series don't need to reflect controls in the experimental sense. For example, you could estimate the causal effect of an advertising campaign on sales using Google searches for your competitors as control time series. – Kay Brodersen Nov 4 '15 at 4:12

The authors of the Synth package may have much more to add, but in principle, CausalImpact and Synth address the same fundamental problem: How can we draw causal inferences in settings where there is only one treated unit, and where treatment was not necessarily assigned at random? Typical examples include a new policy that's instituted in a particular state, or an advertising campaign that targets a particular market.
• Synth uses pre-treatment variables for matching. For example, you might consider the overall state of other markets prior to the beginning of the advertising campaign to construct a synthetic market resembling the treated market.
• CausalImpact, in contrast, uses the full pre- and post-treatment time series of predictor variables for matching. For example, you could consider using demand, sales, or even general GDP time series in other markets to construct a synthetic control for the market that was treated. For more details from the perspective of CausalImpact, see the paragraph starting at the bottom of p. 250 in the manuscript.