0
$\begingroup$

I am running a CausalImpact analysis on a time series and my pre.period goes from 01.01.15 to 30.03.15. I want my post period to be from 15.04.15 to 17.04.15.

Is it ok if I create a time series that ignore the gap and just goes from 01.01.05 to 30.03.15 + from 15.04.15 to 17.04.15 (for a total of 3 months and 3 days)?

I am afraid I might obtain wrong results if CausalImpact take the data as continuous while in reality there is a gap of 15 days between them.

Thanks!

$\endgroup$
  • $\begingroup$ I have never hear of CausualImpact, and I imagine that many reviewers are in the same camp. What kind of time series model are you fitting? $\endgroup$ – HStamper Dec 28 '18 at 1:39
0
$\begingroup$

CausalImpact just performs a cross-correlation of residuals from ARIMA models, with some super beefy (unwarranted) cause-y language thrown on top. Depending on the shape or form of the intervention (like in the case of constant additive effect), you probably won't have a powerful analysis. The approach is just not smart enough to find a lag that best aligns the residuals. To fix this, you can just shift the time series yourself, but be warned this kind of post-processing compromises the validity of statistical analysis. For instance, how could I or any external reviewer trust that the number of lagged days that you tweaked was on the basis of prespecified hypotheses and not "what made the analysis most statistically significant"?

$\endgroup$
  • $\begingroup$ Thanks AdamO, can I ask you some clarifications? In my case, I want the system to learn during a pre.period (1.01.15 to 30.03.15) and analyse the impact of events that occur only on single days (14.04.15 in the example) on a 3 day-post.period (15.04.15 to 17.04.15). To do so, I run CausalImpact as many times as the events I want to analyse, starting from the event on the 31.03.15 (the day immediately after the pre.period). I create a different dataset each time, made of the unchanged first 90 days of pre.period and of the 3 days of the post.period after single event Does this make sense? $\endgroup$ – Alfredo Dec 28 '18 at 11:07
  • $\begingroup$ @Alfredo hard to say. You may post another question. Add graphs, sample data, anything to make your points clearer. Sounds like you should just a simple t-test. $\endgroup$ – AdamO Jan 3 at 15:00

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.