1
$\begingroup$

I want to use to Google's causal impact function to impute the effect of an intervetion. However, my data is structured as: pre-period=1991-1995, intervetion occurs, post-period=1996-2017. To clarify, the intervention is the entrance of a new buyer to the market. I want to test if the demand of the new buyer changed production. So I want impute the counterfactual for 1991-1995 as if the buyer was always in the market.

Because I have so many more years of post intervetnion data, I think it would be safer to train on the post-period data and try to backcast the pre-period data.

Does anyone know if this is possible using the CausalImpact function? Is anyone aware of a better function or resource?

Thanks.

$\endgroup$
  • $\begingroup$ Causal effect of the future on the past? Could you clarify what exactly do you want to estimate? $\endgroup$ – Tim Nov 2 '18 at 14:23
  • $\begingroup$ Yes, I updated the post. To clarify, the intervention is the entrance of a new buyer to the market. I want to test if the demand of the new buyer changed production. So I want impute the counterfactual for 1991-1995 as if the buyer was always in the market. Does that make sense? I am wandering If I can simply reverse the data? @Tim $\endgroup$ – Anna Nov 2 '18 at 14:31
0
$\begingroup$

This is technically possible by reverting the time series, and makes sense given the short actual pre-period (which will most likely not give sufficient power to fit a reasonable model).

The interpretation of the fitted model could be tricky though. It should at least not contain any autoregressive components (as this would mean explaining past production based on future production), i.e. only rely on covariates for the prediction. Furthermore, check whether reversibility seems to be a plausible assumption for the given data set in the first place:

  • Could the production time series also have been observed in reverse order?
  • Do we expect that the disappearance of a buyer of the market has the reverse effect as the appearance of a new buyer?
$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.