Quasi Experimentation at Netflix

Netflix Technology Blog
6 min readintermediate
--
View Original

Overview

The article discusses the implementation of quasi-experimentation at Netflix, highlighting the importance of experimentation in decision-making processes. It explains the limitations of A/B testing and introduces quasi-experiments as a method to analyze the impact of marketing strategies and product changes.

What You'll Learn

1

How to design quasi-experiments to measure marketing impact

2

Why quasi-experiments are necessary when A/B testing is not feasible

3

How to leverage location-based randomization for testing

Key Questions Answered

What are the limitations of A/B testing at Netflix?
A/B testing has limitations such as the inability to randomize individual users and the potential for interference between different experiences, violating the stable unit treatment value assumption (SUTVA). This makes it difficult to accurately measure the effects of certain changes.
How does Netflix implement quasi-experiments?
Netflix implements quasi-experiments by randomly selecting cities for marketing campaigns instead of individual users. This allows them to compare changes in test regions against control regions over time, providing insights into the effectiveness of their marketing strategies.
What is the purpose of the Quasimodo product at Netflix?
The Quasimodo product aims to automate aspects of the quasi-experimentation workflow, enabling Netflix scientists to run more experiments in parallel and focus on hypothesis generation and results interpretation without getting bogged down by operational details.

Technologies & Tools

Backend
Open Connect
Open Connect is used by Netflix to stream content to users, and improvements to its delivery systems are tested through quasi-experiments.

Key Actionable Insights

1
Utilize quasi-experiments for marketing strategies when A/B testing is impractical.
This approach allows for effective measurement of marketing impact in scenarios where individual randomization is not possible, such as out-of-home advertising.
2
Focus on improving statistical power by increasing the number of comparisons in experiments.
By measuring impacts before and after marketing interventions, Netflix can better understand the effectiveness of their strategies and make more informed decisions.
3
Leverage cross-functional collaboration to enhance quasi-experiment designs.
Collaboration among scientists can lead to innovative designs and analyses, which is crucial in a rapidly evolving field like quasi-experimentation.

Common Pitfalls

1
Assuming that quasi-experiments will yield results as precise as A/B tests.
Quasi-experiments are inherently less precise due to the lack of individual randomization, which can lead to imbalances and skewed results. It's important to set realistic expectations when interpreting the outcomes.