AB Testing
"To" docs
Design Considerations¶
Sam
Vid: A/B testing and design - AB tests can compare:
-
Between groups (ie, differences between conditions)
-
Within groups (ie, trend between measures)
Vid: Considerations in A/B testing - Only use A/B when:
-
Subjects/traffic are meaningful
-
Time available for design and tests
-
Clear hypothesis.
A/B test conditions:
-
Data fluctuations: Is there anything outside of the test causing subjects to change?
-
Number of variables: Only test 1 variable (see family wise error rate)
-
Regression to the mean
Metric Considerations¶
Sam
Steps to defining a metric
-
Purpose: What are we using this metric for?
-
Invariant checking (Sanity checks): Metrics that should NOT change in test vs control
-
Evaluation
-
High level: Revenue, market share, users
-
Detailed: Time on page, etc
-
-
-
High level concept: How many active users?
-
Detailed: How do we define active? Which events count as active?
-
Summarize: Convert these details into 1 single metric
The key is to choose summary statistics that match our data distribution (Poisson, Pareto, etc).