What to do when the observation unit is different from the randomization unit
10 hours ago
A/B testing is the gold standard for causal inference because it allows valid causal statements to be made under minimal assumptions. randomization. In fact, by randomly assigning process (medicines, advertisements, products, etc.). result Interests around the world (diseases, corporate profits, customer satisfaction, etc.) Subject (patients, users, customers, etc.) and attribute the mean differences in outcomes to treatment causality.
Sometimes something like this happens: Treatment unit assignment differs from observation unit. In other words, rather than deciding whether to treat all observations independently, we decide whether to treat them as a group. For example, you might decide to process all customers in a particular region while observing results at the customer level, or all products of a particular brand while observing results at the product level. This usually occurs due to practical constraints.In the first example, the so-called geography experimentthis happens because users cannot be tracked due to the deprecation of cookies.
In this case, the therapeutic effect not independent Beyond observation. In fact, when a customer in one area is treated, other customers in the same area are also treated. If a product of a certain brand is not carried, other products of the same brand will also not be carried. This dependency must be taken into account when making inferences. Standard errors, confidence intervals, and p-values need to be adjusted. In this article, we’ll explore how to do that using: Cluster-robust standard error.
Imagine you are an online platform and you are interested in increasing your sales. You have come up with a great idea. Carousel of related articles Prompt customers to add other items to their basket during checkout. To understand whether carousels increase sales, you decide to do an AB test. In principle, you can randomly decide whether to display the carousel for each order. However, this…
