Skip to content

A/A Tests show false positive rate exceeds alpha.  #7

@sjoelee

Description

@sjoelee

I'm interested in using this library so I tested to see if the False Positive Rate is controlled within alpha that I set. However, the results are showing that the number of tests that the portion of tests rejected far exceeds alpha. For 100 trials, where each trial is comparing two Bernoulli proportions (of 0.01 conversion rate) and alpha=0.05, beta=0.10, and the total number of visitors is set to 100K, I get the following percent of reject and accept:

% rejected:  43
% accepted:  56.99999999999999

Here's the link to my code: https://gist.github.com/sjoelee/d4fed8b80e1af1d2e0cf7aac37d09a90. Once each visitor arrives and has been bucketed to treatment/control, I simulate a biased coin flip (based on conversion rate for variation) for their conversion. I add each individual data through addData and then look at the results to see if they finished, determining whether the result from addData returned true (accept null) or false (reject null). Could you provide more documentation on how the thresholds are calculated? And have any tests been done to see if A/A tests are still controlled under alpha? Thanks!

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions