article thumbnail

Towards optimal experimentation in online systems

The Unofficial Google Data Science Blog

If $Y$ at that point is (statistically and practically) significantly better than our current operating point, and that point is deemed acceptable, we update the system parameters to this better value. And we can keep repeating this approach, relying on intuition and luck. Why experiment with several parameters concurrently?

article thumbnail

To Balance or Not to Balance?

The Unofficial Google Data Science Blog

In an ideal world, experimentation through randomization of the treatment assignment allows the identification and consistent estimation of causal effects. Identification We now discuss formally the statistical problem of causal inference. We start by describing the problem using standard statistical notation.

Insiders

Sign Up for our Newsletter

This site is protected by reCAPTCHA and the Google Privacy Policy and Terms of Service apply.

article thumbnail

Knowledge

Occam's Razor

The Awesome Power of Visualization 2 -> Death and Taxes 2007. Build A Great Web Experimentation & Testing Program. Experimentation and Testing: A Primer. Tip #9: Leverage Statistical Control Limits. Tip#1: Statistical Significance. Web Analytics Career Advice: Statistics, Business, IT & Mushrooms.

KPI 124
article thumbnail

Changing assignment weights with time-based confounders

The Unofficial Google Data Science Blog

For example, imagine a fantasy football site is considering displaying advanced player statistics. A ramp-up strategy may mitigate the risk of upsetting the site’s loyal users who perhaps have strong preferences for the current statistics that are shown. One reason to do ramp-up is to mitigate the risk of never before seen arms.

article thumbnail

Measuring Incrementality: Controlled Experiments to the Rescue!

Occam's Razor

You need people with deep skills in Scientific Method , Design of Experiments , and Statistical Analysis. The team did the normal modeling to ensure that the results were statistically significant (large enough sample set, sufficient number of conversions in each variation). * ask for a raise. It is that simple. Okay, it is not simple.

article thumbnail

The Lean Analytics Cycle: Metrics > Hypothesis > Experiment > Act

Occam's Razor

Remember that the raw number is not the only important part, we would also measure statistical significance. Circle of Friends was a social community built atop Facebook that launched in 2007. The result? The properties with professional photography had 2-3 times the number of bookings! The graph is impressive, right?

Metrics 156
article thumbnail

Estimating causal effects using geo experiments

The Unofficial Google Data Science Blog

A geo experiment is an experiment where the experimental units are defined by geographic regions. Statistical power is traditionally given in terms of a probability function, but often a more intuitive way of describing power is by stating the expected precision of our estimates. They are non-overlapping geo-targetable regions.