zapier/testr

Name: testr

Owner: Zapier

Description: Tools for Bayesian A/B testing

Forked from: ayakubovich/testr

Created: 2017-02-09 20:39:13.0

Updated: 2017-02-09 20:39:14.0

Pushed: 2016-04-09 22:20:13.0

Homepage:

Size: 91

Language: R

GitHub Committers

User	Most Recent Commit	# Commits

Other Committers

User	Email	Most Recent Commit	# Commits

README

testr: Tools for Bayesian analysis of A/B tests

testr is an R package for Bayesian design and analysis of randomized experiments (A/B tests). It uses an adaptive design (i.e. optional stopping rules), allowing tests to end as early as possible ^[5](5).

Two probability models are currently supported:

1. Beta binomial model for testing retention/conversion ^{[1](1), [7](7)}
2. Zero-inflated log-normal model for testing revenue. This model has two stages: modelling spenders vs. non-spenders with a beta-binomial model, and modelling revenue amongst spenders with a lognormal model ^[3](3).

testr provides a convenient API to these models, and extends them in several ways:

• Generalized the early stopping rule ^[5](5) to the Zero-inflated log-normal model
• Uses Monte Carlo integration of the posterior to evaluate stopping rules
• Implements methods for visualizing the prior/posterior distributions

Example Usage

Install the package using devtools:

stall.packages('devtools')
ools::install_github('ayakubovich/testr')
ary(testr)
tr

To run a simple test for conversion with a Beta-Binomial model:

_binomial_ab_test(y = c(100,115), n = c(1000, 1000)) # flat prior
_conversion_prior(expected_conversion_rate = .1, alpha0 = 2) # try tweaking alpha0 to decrase variance
_binomial_ab_test(y = c(100,115), n = c(1000, 1000), expected_conversion_rate = .1, alpha0 = 2) # informative prior

To run a test for revenue with a zero-inflated lognormal model:

mulate some data from an A/B/C test:
 10000
ta <- rbinom(n, 1, conversion) * rlnorm(n, meanlog = 0, sdlog = 0.2)
ta <- rbinom(n, 1, conversion) * rlnorm(n, meanlog = 0.08, sdlog = 0.2)
ta <- rbinom(n, 1, conversion) * rlnorm(n, meanlog = 0.15, sdlog = 0.2)
 <- data.frame(ab_group = rep(c('A', 'B', 'C'), each = n), ltv = c(A_data, B_data, C_data))

ot the priors and run the test:
_revenue_prior(expected_conversion_rate = 0.12, alpha0 = 15, expected_revenue_converted_users = 1.5, v0 = 73, k0 = 100, s_sq0 = 1.2) # specify prior
  lognormal_ab_test(data, expected_conversion_rate = 0.65, alpha0 = 15, expected_revenue_converted_users = 1.5, v0 = 73, k0 = 100, s_sq0 = 1.2)
(l)

To do

• [ ] A/B/C…N tests: Handle tests with more than two groups using Hierarchical Models. [4](4), [10](10), [11](11) are good references on the topic. testr currently allows for more than two groups in a test, but does not take multiple comparisons into account.
• [ ] Model checking: Add methods for verifying that the probability model good fit for the data. Gelman et al. describe some methods for doing this in [4](4).
• [ ] Priors: Add more ways to specify priors.

References

1. Feldman, Sergey. Bayesian A/B Tests
2. Feldman, Sergey. Bayesian Analysis of Normal Distributions with Python
3. Feldman, Sergey. Bayesian A/B Testing with a Log-Normal Model
4. Gelman, Andrew, et al. Bayesian data analysis. Vol. 2. London: Chapman & Hall/CRC, 2014
5. Stucchio, Chris. Agile A/B testing with Bayesian Statistics and Python
6. Davidson-Pilon, Cameron. Bayesian Methods for Hackers on GitHub
7. Kula, Maciej. Bayesian AB Testing
8. Wulczyn, Ellery. How Naive A/B Testing Goes Wrong
9. Stucchio, Chris. The Metrics Manifesto - Why you need an objective function
10. Nandi, Manojit. A/B Testing with Hierarchical Models in Python
11. Stich, Slater. Use a Hierarchical Model
12. Robinson, David. Is Bayesian A/B Testing Immune to Peeking? Not Exactly