How to run A/B Tests as a Data Scientist!

Let's learn about how & why you should use Bayesian Testing. And some advantages of the Bayesian approach over frequentist approach with REAL data/code. Note: Bayesian Appraoch isn't necessarily better in every way - it is another perspective of looking at data.

TIMESTAMPS
0:00 Introduction
0:56 Define the Experiment
4:25 Data Collection
6:20 Data Processing
8:43 Experiment: The Frequentist Approach
12:50 Experiment: The Bayesian Approach
14:40 Bayesian: Generating Priors
20:06 Bayesian: Generating Posteriors
22:00 Interpreting results
26:00 Bayesian Vs Frequentist

[CodeEmporium]

Hope you all enjoy the video. Just note that the Bayesian approach isn't necessarily better than the frequentist in every way - it is another perspective of looking at data. In fact, @Cassie Kozyrkov (head of decision intelligence at Google) has an amazing video comparing the 2 approaches with a simple and fun example. Thought I'd give it a shout out since its super cool

[enrices]

Great video. I just wanted to add a few points :
- Frequentist approach also allows you to monitor the evolution of p-value as week goes by
- Frequentist approach (with 1 tailed t-test for example) also allowed for checking hypothesis like "Is the uplift at least 2%?"
- Small correction on your explanation of Bayesian approach : You don't get the proba of having a lift of -0.14% or something. You get the probability of having a negative lift, and then if we have to estimate the lift, then it would be -0.14%. but that's independent

[sololife]

your explanations and your voice are crystal clear. Thank you:)

[geoffreyanderson]

Thank you for clearly demonstrating this technique. It will be useful in a lot of experiments because you can peek while still collecting outcomes and it won't mess up the decision making, making it more practical to use than the classical p-value based decision making.

[yangsun]

A/B testing for conversion rate should use Z test instead of chi-square test. It's never about independence of two random variables. It's about the difference of the two groups is by chance or by design.

[yenliknurasheva]

Wow! You are simply the best, thank you so much for this video!

[TheBjjninja]

To sum it up in business terms, you can use this approach to potentially save on testing costs. That is the benefit of sampling.

[nanlinr]

Thank you this is very clear in terms of how you executed the Bayesian approach with priors and updating your beliefs. However I feel like your interpretation of the result is pretty subjective - it's true that you get more confident when you obtain more weeks of data (30% -> 19% ->15% probability treatment is better than control), so likely we will not roll out with new page based on conversion alone, but how are you "super confident" at 85% likelihood of control winning? With frequentist approach, often times we need p value <= 5% to be confident (statistically significant difference in C vs. T) meaning 95% of the time we would observe the same directional difference we are currently observing. So 85% likelihood from a statistical standpoint seems not confident enough for me actually. But the direction in which it's becoming more confident week over week is at least reassuring.

[heliyahasani]

Amazing content please upload more videos.💛

[mohamedgaal]

Thank you so much for simplifying this! Really enjoyed your tutorial.

[arianvc]

You make the things I look for.
Good job BTW 👍 thanks!

[kasun]

Thank you so much for this. Super helpful!!!

[pallavibothra]

Thank you for awesome videos. I have a question: could you please explain line 255. I'm not sure why did you use means of prior but not directly prior? It is similar to bootstrapping sampling but you already have enough sample size, then what is the need of bootstrapping? I am new to this field, so sometimes concepts become really confusing :(

[lucasgarces]

How do you know which prior distribution to choose? And how do you choose the parameters of the chosen prior distribution?

[orjihvy]

Can you do more bayesian approaches for future videos, assuming there are others also interested.

[mehdizoghinia]

great video! learned a lot, thank you.

[ciafacil]

what happens if at the beginning, I don't want to eliminate the 1.3% of users that have seen both and want to make sure that doest happens again, like, trying to separate them??

[orjihvy]

I'm not sure if 'probability that we are seeing a 2% lift' measure you computed toward the end is actually a probability. Because what's to keep it bounded between 0 and 1.

[MntAll]

Thanks four your great explanation! Just want to check whether I understood one aspect correctly: in a chi-square test, dont we need to state null hypothesis like: "conversion and group are independent" instead of "control and treatment are independent"? The second option is very confusing...

[orjihvy]

How does one learn all these panda tricks? I keep forgetting them if I don't end up using them.