Model-based multiple testing correction for large-scale experiments at YouTube

Speaker Name: 
Jacopo Soriano
Speaker Organization: 
Youtube
Start Time: 
Monday, October 19, 2015 - 4:00pm
End Time: 
Monday, October 19, 2015 - 5:00pm
Location: 
Engineering 2, Room 180

Randomized experiments are widely used at YouTube to inform decisions.
For each experiment, thousands of metrics are tested for differences
between the treatment group and the control group. Given the large
number of hypotheses, multiple testing correction is crucial. In this
talk I discuss approaches to directly model the dependence structure
across hypotheses to improve the performance of the multiple testing
procedure.