Model - Causal Impact


Causal Impact test allows to assess the significance of a variation pre-post an intervention on a test group.


This can be very useful to assess the impact of an action in wide variety of business, such as SEO experiments, marketing and advertising campaigns or Stock Market analysis.

For instance, you are a large website and you want to test a new strategy in SEO (Search Engine Optimization). Say that you want to include the price of a product in the meta-description seen by users in Google search results, and see the impact of that change on the click through rate. You cannot properly AB test such a thing: Google allows you to refer to only one page per product, and you can’t split Google users between A and B variants. So you end up creating a set of product pages that will make that change on a specific day (your variant group) and you keep a set of product pages that won’t change at all (your control group) and will be used as a base line to assess the real impact of the change.

Here is an example of a compatible dataset. Note that CausalImpact is a time series approach so requires a “date” column.

Once uploaded in Datama Impact, you then need to define what you are comparing (in that case, pre-post the change, on the 30th of Jan 2022)


assess - causal model

You also need to optin for the Causal Impact test in the “settings” section. Then, define where is the Control/ Variant group column and what’s the segment that is impacted by the change

assess - causal model

Once settings are properly defined, you can hit “Compute” and results will show you

assess - causal model

In the example above, the Click/ Impression for the test group has increased quite a bit after the change (30th of Jan 2022), but the control group also increased by the same range, hence the causal impact is not significant: We don’t think that the observed variation of the KPI was actually due to the applied change, but to another reason.


Note: Datama Causal Impact is based on the work from Kay H. Brodersen, Alain Hauser at Google released through the CausalImpact R package

See also: Causal Impact

Test data: Gsheet