Ronny Kohavi, Microsoft

Slides available at http://exp-platform.com

Add an item to your shopping cart at a website

Most sites show the cart
At Amazon, Greg Linden had the idea of showing
recommendations based on cart items
Evaluation
Pro: cross-sell more items (increase average basket size)
Con: distract people from checking out (reduce conversion)
HiPPO (Highest Paid Person’s Opinion) was: stop the project
Simple experiment was run,
wildly successful
From Greg Linden’s Blog: http://glinden.blogspot.com/2006/04/early-amazon-shopping-cart.html

The less data, the stronger the opinions

Whenever you feel stressed that a decision is made without

data, squeeze the Stress-HiPPO
Put one in your office to show others you believe in data-
driven decisions based on experiments
Hippos kill more humans than any
other (non-human) mammal (really)
Don’t let HiPPOs in your org
kill innovative ideas. ExPeriment!

1
 4

Controlled Experiments in one slide

Examples: you’re the decision maker
Cultural evolution: hubris, insight through measurement,
Semmelweis reflex, fundamental understanding
Controlled Experiments: deeper dive

Two key messages to remember

It is hard to assess the value of ideas.
Get the data by experimenting because data trumps intuition
Make sure the org agrees what you are optimizing

Concept is trivial
Randomly split traffic between
two (or more) versions
A/Control
B/Treatment
Collect metrics of interest
Analyze

Must run statistical tests to confirm differences are not due to chance
Best scientific way to prove causality, i.e., the changes in metrics are
caused by changes introduced in the treatment(s)

Three experiments that ran at Microsoft recently

All had enough users for statistical validity
Game: see how many you get right
Everyone please stand up
Three choices are:
A wins (the difference is statistically significant)
A and B are approximately the same (no stat sig diff)
B wins
If you guess randomly
1/3 left standing after first question
1/9 after the second question

2
 7

“Find a house” widget variations

Overall Evaluation Criterion: Revenue to Microsoft
generated every time a user clicks search/find button

A B
• Raise your right hand if you think A Wins
• Raise your left hand if you think B Wins
• Don’t raise your hand if you think they’re about the same

If you did not raise a hand, please sit down

If you raised your left hand, please sit down
A was 8.5% better
Since this is the #1 monetization, it effectively raised
revenues significantly.

Actual experiment had 6 variants.

There was a “throwdown” (vote for the winning variant) and nobody
from MSN Real Estate or Zaaz (the company that did the creative)
voted for the winning widget

OEC: Clickthrough rate for Search box and popular searches

Differences: A has taller search box (overall size is the same), has magnifying glass icon,
“popular searches”
B has big search button

• Raise your right hand if you think A Wins

• Raise your left hand if you think B Wins
• Don’t raise your hand if they are the about the same

3
 10

If you raised any hand, please sit down

Insight
Stop debating, it’s easier to get the data

11

A later test showed that changing the magnifying glass to an

actionable word (search, go, explore) was highly beneficial.
This:

is better than

OEC: Clicks on revenue generating links (red below)

A B

• Raise your right hand if you think A Wins

• Raise your left hand if you think B Wins
• Don’t raise your hand if they are the about the same

4
 13

If you did not raise a hand, please sit down

If you raised your left hand, please sit down
B was 64% worse

What % of the audience is still standing?

Humbling!

14

Any figure that looks interesting or different is

usually wrong

If something is “amazing,” find the flaw!

Examples

If you have a mandatory birth date field and people think it’s unnecessary,
you’ll find lots of 11/11/11 or 01/01/01

If you have an optional drop down, do not default to the first alphabetical entry,
or you’ll have lots jobs = Astronaut

Traffic to web sites doubled between 1-2AM November 6, 2011 for many sites,
relative to the same hour a week prior. Why?

The previous Office example assumes click maps to revenue.
Seemed reasonable, but when the results look so extreme, find the
flaw

15

Features are built because teams believe they are useful.

But most experiments show that features fail to move the
metrics they were designed to improve
We joke that our job is to tell clients that their new baby is ugly
In the recently published book Uncontrolled, Jim Manzi writes
Google ran approximately 12,000 randomized experiments in 2009, with [only]
about 10 percent of these leading to business changes.
In an Experimentation and Testing Primer by Avinash Kaushik,
authors of Web Analytics: An Hour a Day, he wrote
80% of the time you/we are wrong about what a customer wants

5
 16

QualPro tested 150,000 ideas over 22 years

75 percent of important business decisions and
business improvement ideas either have no impact on
performance or actually hurt performance…
Based on experiments at Microsoft (paper)
1/3 of ideas were positive ideas and statistically significant
1/3 of ideas were flat: no statistically significant difference
1/3 of ideas were negative and statistically significant
Our intuition is poor: 60-90% of ideas do not improve the
metric(s) they were designed to improve (domain dependent).
Humbling!

17

One page article by Stefan Thomke, May 2006

Mistakes refer to the wrong actions that result from poor
judgments or inattention; they should be avoided because
they produce little new or useful information.
A poorly planned or badly conducted experiment that
results in ambiguous data, forcing researchers to repeat the
experiment, is a mistake.
Another common mistake is repeating a prior failure or
learning nothing from the experience

18

Story about Tom Watson Sr., IBM's founder

A promising young executive involved in a risky new venture,
managed to lose more than $10 million while trying to make the
venture work
When the nervous man was called into Watson's office, he offered
to accept the logical consequence of losing the company such a
large amount of money: "I guess you want my resignation, Mr.
Watson.“
Much to his surprise, Watson countered: 'You can't be serious!
We've just spent 10 million [dollars] educating you."

6
Avoid the temptation to try and build optimal features
through extensive planning without early testing of ideas
Experiment often
To have a great idea, have a lot of them -- Thomas Edison
If you have to kiss a lot of frogs to find a prince,
find more frogs and kiss them faster and faster
-- Mike Moran, Do it Wrong Quickly
Try radical ideas. You may be surprised
Doubly true if it’s cheap to implement (e.g., shopping cart
recommendations and Behavior-Based search at Amazon)
If you're not prepared to be wrong, you'll never come up
with anything original – Sir Ken Robinson, TED 2006

20

If you remember one thing from this talk, remember this point
OEC = Overall Evaluation Criterion
Agree early on what you are optimizing
Getting agreement on the OEC in the org is a huge step forward
Suggestion: optimize for customer lifetime value, not immediate
short-term revenue
Criterion could be weighted sum of factors, such as
Time on site (per time period, say week or month)
Visit frequency
Report many other metrics for diagnostics, i.e., to
understand why the OEC changed and raise new hypotheses

21

Controlled Experiments in one slide

Examples: you’re the decision maker
Cultural evolution: hubris, insight through measurement,
Semmelweis reflex, fundamental understanding
Controlled Experiments: deeper dive

Two key messages to remember

It is hard to assess the value of ideas.
Get the data by experimenting because data trumps intuition
Make sure the org agrees what you are optimizing

7
 22

It is difficult to get a man to understand something when his

salary depends upon his not understanding it.
-- Upton Sinclair

Why people/orgs avoid controlled experiments

Some believe it threatens their job as decision makers
At Microsoft, program managers select the next set of features to
develop. Proposing several alternatives and admitting you don’t
know which is best is hard
Editors and designers get paid to select a great design
Failures of ideas may hurt image and professional standing.
It’s easier to declare success when the feature launches
We’ve heard: “we know what to do. It’s in our DNA,” and
“why don’t we just do the right thing?”

The org goes through stages in its cultural evolution

Stage 1: we know what to do and we’re sure of it
True story from 1849
John Snow claimed that Cholera was caused by polluted water
A landlord dismissed his tenants’ complaints that their water stank
Even when Cholera was frequent among the tenants
One day he drank a glass of his tenants’ water to show there was
nothing wrong with it
He died three days later
That’s hubris.
Even if we’re sure of our ideas, evaluate them
Controlled experiments are a powerful tool to evaluate ideas

Measurement
• Semmelweis worked at Vienna’s General Hospital, an
important teaching/research hospital, in the 1830s-40s
• In 19th-century Europe, childbed fever killed more than a
million women
• Measurement: the mortality rate for women giving birth was
• 15% in his ward, staffed by doctors and students
• 2% in the ward at the hospital, attended by midwives

8
 Insight
Control
• He tries to control all differences
• Birthing positions, ventilation, diet, even the way laundry was done
• He was away for 4 months and death rate fell significantly
when he was away. Could it be related to him?
• Insight:
• Doctors were performing autopsies each morning on cadavers
• Conjecture: particles (called germs today) were being transmitted
to healthy patients on the hands of the physicians
He experiments with cleansing agents
• Chlorine and lime was effective: death rate fell from 18% to 1%

Success? No! Disbelief. Where/what are these particles?

Semmelweis was dropped from his post at the hospital
He goes to Hungary and reduced mortality rate in obstetrics to 0.85%
His student published a paper about the success. The editor wrote
We believe that this chlorine-washing theory has long outlived its
usefulness… It is time we are no longer to be deceived by this theory
In 1865, he suffered a nervous breakdown and was beaten at a
mental hospital, where he died
Semmelweis Reflex is a reflex-like rejection of new knowledge
because it contradicts entrenched norms, beliefs or paradigms
Only in 1800s? No! A 2005 study: inadequate hand washing is one
of the prime contributors to the 2 million health-care-associated
infections and 90,000 related deaths annually in the United States

In 1879, Louis Pasteur showed the presence of

Streptococcus in the blood of women with child fever
2008, 143 years after he died, there is a 50 Euro coin
commemorating Semmelweis

9
 Accept Results
Measure and avoid Fundamental
Hubris Semmelweis
Control Understanding
Reflex

In many areas we’re in the 1800s in terms of our

understanding, so controlled experiments can help
First in doing the right thing, even if we don’t understand the
fundamentals
Then developing the underlying fundamental theories

29

Controlled Experiments in one slide

Examples: you’re the decision maker
Cultural evolution: hubris, insight through measurement,
Semmelweis reflex, fundamental understanding
Controlled Experiments: deeper dive

Two key messages to remember

It is hard to assess the value of ideas.
Get the data by experimenting because data trumps intuition
Make sure the org agrees what you are optimizing

30

With data mining, we find patterns, but most are

correlational, providing hypotheses for possible causes
Here is one a real example
of two highly correlated
variables

10
 31

• Data for the city of Oldenburg,

Germany
• X-axis: stork population
• Y-axis: human population

What your mother told you about babies

and storks when you were three is still
not right, despite the strong
correlational “evidence”

Ornitholigische Monatsberichte 1936;44(2)

True statement (but not well known):

Palm size correlates with your life expectancy
The larger your palm, the less you will live, on average.
Try it out - look at your neighbors and you’ll see who is
expected to live longer

But…don’t try to bandage your hands

Women have smaller palms and live 6 years longer
on average

33

Controlled experiments test for causal relationships, not

simply correlations
When the variants run concurrently, only two things could
explain a change in metrics:
1. The “feature(s)” (A vs. B)
2. Random chance
Everything else happening affects both the variants
For #2, we conduct statistical tests for significance
The gold standard in science and the only way to prove
efficacy of drugs in FDA drug tests

11
 34

If you don't know where you are going, any road will take you there
—Lewis Carroll

Scope: Experimentation is not applicable everywhere

Necessary ingredients for experimentation to be useful detailed in
separate paper (see http://exp-platform.com/expMicrosoft.aspx)
Sweet spot: websites and services that practice agile
development (Scrum, XP, etc)
Org has to agree on OEC (Overall Evaluation Criterion).
This is hard, but it provides a clear direction and alignment
Quantitative metrics, not always explanations of “why”
A treatment may lose because page-load time is slower.
At Amazon, we slowed pages by 100-250msec and lost 1% of revenue
A treatment may have JavaScript that fails on certain browsers, causing
users to abandon.

35

Primacy/novelty effect
Primacy: Changing navigation in a website may degrade the customer
experience (temporarily), even if the new navigation is better
Novelty: new flashing icon catches the eye and everyone clicks once
Evaluation may need to focus on new users, or run for a long period
Multiple experiments
Even though the methodology shields an experiment from other changes,
statistical variance increases making it harder to get significant results.
There can also be strong interactions (rarer than most people think)
Consistency/contamination
On the web, assignment is usually cookie-based, but people may use
multiple computers, erase cookies, etc. Typically a small issue
Launch events / media announcements sometimes
preclude controlled experiments
The journalists need to be shown the “new” version

Run A/A tests

Run an experiment where the Treatment and Control variants
are coded identically and validate the following:
1. Are users split according to the planned percentages?
2. Is the data collected matching the system of record?
3. Are the results showing non-significant results 95% of the time?
This is a powerful technique for finding bugs and other
integration issues before teams try to make data-driven
decisions
Generating some numbers is easy
Getting correct numbers you trust is much harder!

12
 37

A very common mistake is to declare a winner when the difference

could be due to random variations
Compute 95% confidence intervals on the metrics to determine if the
difference is due to chance or whether it is statistically significant
Increase percentage if you do multiple tests
(e.g., use 99%)
Idea: run an A/A test in concurrent to your A/B test to make sure the
overall system doesn’t declare it as significant more than 5% of the
time (great QA)

38

Ramp-up
Start an experiment at 0.1%
Do some simple analyses to make sure no egregious problems can be
detected
Ramp-up to a larger percentage, and repeat until 50%
Big differences are easy to detect because the min sample size is
quadratic in the effect we want to detect
Detecting 10% difference requires a small sample and serious problems can
be detected during ramp-up
Detecting 0.1% requires a population 100^2 = 10,000 times bigger
Abort the experiment if treatment is significantly worse on OEC or
other key metrics (e.g., time to generate page)

39

Unless you’re running an experiment on one of the top 10

sites in the world, run experiments at 50/50%
Novice experimenters run 1% experiments
To detect an effect, you need to expose a certain number of
users to the treatment (based on power calculations)
Fastest way to achieve that exposure is to run equal-
probability variants (e.g., 50/50% for A/B)
If you perceive risk, don’t start an experiment at 50/50%
from the beginning: Ramp-up over a short period

13
 40

Good randomization is critical.

It’s unbelievable what mistakes developers will make in favor of efficiency
Properties of user assignment
Consistent assignment. User should see the same variant on
successive visits
Independent assignment. Assignment to one experiment should
have no effect on assignment to others (e.g., Eric Peterson’s code
in his book gets this wrong)
Monotonic ramp-up. As experiments are ramped-up to larger
percentages, users who were exposed to treatments must stay in
those treatments (population from control shifts)

41

Run concurrent univariate experiments

Vendors make you think that MVTs and Fractional Factorial designs are
critical---they are not. The same claim can be made that polynomial models
are better than linear models: true in theory, less useful in practice
Let teams launch multiple experiments when they are ready, and do the
analysis to detect and model interactions when relevant (less often than you
think)
Backend integration (server-side) is a better long-term approach to
integrate experimentation than Javascipt
Javascript suffers from performance delays, especially when running multiple
experiments
Javascript is easy to kickoff, but harder to integrate with dynamic systems
Hard to experiment with backend algorithms (e.g., recommendations)

42

The less data, the stronger the opinions

1. It is hard to assess the value of ideas
Listen to your customers
Get the data by experimenting because data trumps intuition
Examples are humbling. More at http://exp-platform.com
2. Empower the HiPPO with data-driven decisions
OEC: make sure the org agrees what you are optimizing (long term lifetime value)
3. Compute the statistics carefully
Power, 95% confidence, ramp-up
Stats/details described in DMKD journal paper: http://exp-platform.com/hippo_long.aspx
4. Experiment often
Triple your experiment rate and you triple your success (and failure) rate.
Fail fast & often in order to succeed
Accelerate innovation by lowering the cost of experimenting

14
 43

44

Support.microsoft.com shows “top issues”

OEC = click-through rate
A shows top issues
B filters top issues to
OS & Browser used to
visit site (useragent)

Personalization rarely hurts, but does it help?

• Raise your right hand if you think B Wins by over 30%
• Raise your left hand if you think B Wins by under 30%
• Don’t raise your hand if you think they’re about the same

45

If you did not raise a hand, please sit down

If you raised your left hand, please sit down
B was >50% better

Personalization helps more than people think!

15
 A: When user clicks on email
hotmail opens in same window
B: Open hotmail in separate window
Trigger: only users that click in the
module are in experiment
(no diff otherwise)
OEC: clicks on home page (after trigger)

• Raise your right hand if you think A Wins

• Raise your left hand if you think B Wins
• Don’t raise your hand if they are the about the same

48

If you didn’t raise a hand, please sit down

If you raised your right hand, please sit down
For those in the experiment, clicks on MSN Home Page
increased +8.9%
<0.001% of users in B wrote negative feedback about the
new window

16
 49

The experiment report was sent by the BI/CI team to all multiple
teams across the world

Someone who saw the report wrote
This report came along at a really good time and was VERY useful.

I argued this point to my team (open Live services in new window from
HP) just some days ago.
They all turned me down.

Funny, now they have all changed their minds.

50

“Find a house” widget variations

Overall Evaluation Criterion: Revenue to Microsoft
generated every time a user clicks

Think which one should win…

51

The widget that performed the best was…

the simplest

Revenue increase over control: +9.7%

Interesting note: nobody from MSN Real Estate or Zaaz (the
company that did the creative) thought this widget would win

17
Proposal: New Offers module below Shopping

Control

Treatment

Value proposition
The Offers module appears below the fold
Sales estimated the three ads would sell for several millions of
dollars a year
Concern
Do more ads degrade the user experience?
How do we trade the two off?
Experiment!

Ran experiment for 12 days on 5% of traffic

Clickthrough rate per user (CTR) decreased 0.31%
(p-value =0.044). This result is statistically significant
Clicks per user decreased 0.72% (p-value=0.015)
Value of click from home page: talk to finance and the SEM team (how much are
you paying to drive traffic from search engines)

The net result: losing idea

18
 Determine impact of 2 factors for video ads.
1) Factor A: pre-roll vs. post-roll ads
2) Factor B: time between ads (90, 120, 180, 300, 900 seconds)
OEC: revenue from ad starts

56

Eliminating the initial pre-roll ad

Increased repeat visits by 1-3%, but
Reduced overall ad views by 50-60%
The amount of time between ad plays had no statistically significant impact on
repeat visits
Decreasing the length of time between ad streams increased the total ad streams
without impacting loyalty
(as measured by return visits)
Reducing the time to 90 seconds would improve annual revenue by millions
Client wrote
There is a preponderance of opinion driven design…
The results of the experiment were in some respect counterintuitive.
They completely changed our feature prioritization. It dispelled long held assumptions about
video advertising. Very, very useful.

This experiment ran in Windows Marketplace / Game Downloads

Which image has the higher clickthrough? By how much?

A: Solitaire game
in hero position

B: Poker game
in hero position

A is 61% better

19
 For many years, the prevailing
conception of illness was that the sick
were contaminated by some toxin
Opening a vein and letting the sickness
run out – bloodletting.
One British medical text recommended bloodletting for
acne, asthma, cancer, cholera, coma, convulsions, diabetes, epilepsy, gangrene, gout, herpes,
indigestion, insanity, jaundice, leprosy, ophthalmia, plague, pneumonia, scurvy, smallpox, stroke,
tetanus, tuberculosis, and for some one hundred other diseases
Physicians often reported the simultaneous use of fifty or more leeches
on a given patient.
Through the 1830s the French imported about forty million leeches a
year for medical purposes

Lancet

President George Washington had a sore throat and

doctors extracted 82 ounces of blood over 10 hours
(35% of his total blood), causing anemia and hypotension.
He died that night.
Pierre Louis did an experiment in 1836 that is now recognized as one of
the first clinical trials, or randomized controlled experiment. He treated
people with pneumonia either with
early, aggressive bloodletting, or
less aggressive measures
At the end of the experiment, Dr. Louis counted the bodies. They were
stacked higher over by the bloodletting sink.

Tight integration with other systems (e.g., content management)

allowing “codeless experiments”
Accurate results in near real-time
Trust is important
Quickly detect and abort poorly performing experiments
High-performance data pipeline with built-in data loss detection
Minimal risk for experimenting applications
Encourage bold innovations with reduced QA cycles
Auto-abort catches bugs in experimental code
Client library insulates app from platform bugs
Experimentation should be easy
Client library exposes simple interface
Web UI enables self-service
Service layer enables platform integration
60

20
 61

Searches for “24” are underspecified, yet

most humans are probably searching for the
TV program

Prior to Behavior-based search, here is what

you would get (you can get this today by
adding an advanced modifier like –foo to
exclude foo)

Mostly irrelevant stuff:

• 24 Italian songs
• Toddler clothing suitable for 24 month olds
• 24” towel bar
• Opus 24 by Strauss
• 24- lb stuff, cases of 24, etc

Ronny Kohavi

62

Amazon/P13N had an engine for X to Y

People who bought item X bought Y
Feed searches, i.e.,
People who searched for X bought Y
Prototype looked great
Integration with Search would take a long time: different team in A9 in
the Bay Area
They also highlighted a flaw:
The results don’t always have the search terms.
Example: search for “duracell charger” and you will see a best selling Sony
charger, which does not have the word Duracell.
But people choose to buy it after searching with this phrase!
(*) Based on UW iEdge Seminar talk by Amazon, 4/2006

63

Ran experiment with very

thin integration
Strong correlations shown
at the top of the page,
pushing search results
down
Implemented simple de-
duping of results
Result : +3% increase to
revenue.
3% of $12B is $360M Ronny Kohavi

21
 64

I’m a Netflix user since 1/2000

Great example of a company tweaking things

(*) TIMITI acronym by Jim Sterne Ronny Kohavi, Microsoft Confidential

65

Ronny Kohavi, Microsoft Confidential

66

Details in Business 2.0 Apr 21, 2006.

The evolution of the NetFlix envelope
Ronny Kohavi, Microsoft Confidential

22
 67

Article in Organization Science Management can support

experimentation and highlight it as a value (normative
influence)
However, inconsistent reward systems that punish failure
lead to aversion
Especially in organizations that are under constant
evaluation for perfect execution

68

Tiger Woods comes to you for advice on how to spend his

time: improving golf, or improving ad revenue (most revenue
comes from ads)
Short term, he could improve his ad revenue
by focusing on ads…
But to optimize lifetime financial value
(and immortality as a great golf player),
he needs to focus on the game

While the example seems obvious,

organizations commonly make the mistake
of focusing on the short term

69

HBR article by Stefan Thomke (Feb, 2001)

Experimentation lies at the heart of every company's ability
to innovate
In the past, testing was relatively expensive, so companies
had to be parsimonious with the number of experimental
iterations
The electric light bulb, required more than 1,000 complex
experiments with filament materials and shapes,
electromechanical regulators, and vacuum technologies

23
 70

Essentials for enlightened experimentation

1. Organize for rapid experimentation
a) Encourage rapid experimentation
b) Use small development groups with key people to iterate rapidly
c) Perform experiments in parallel
2. Fail early and often, but avoid mistakes
a) Embrace failure to advance knowledge
b) Failure can expose important gaps in knowledge
c) Don't forget the basics: well designed tests with clear objectives
d) IDEO's “fail often to succeed sooner”
e) Don't develop expensive sleek prototypes, as you become committed to
them before you know if they work.

71

Essentials (CONT)
3. Anticipate and exploit early information
a) Front-load to identify problems and provide guidance when it's cheap
b) Acknowledge trade-off between cost and fidelity.
Low-fidelity experiments (costing less) are suited in early exploratory stages
4. Combine new and traditional technologies
a. Today's new technology might eventually replace its traditional counterpart,
but it could then be challenged by tomorrow's new technology

Beware of measuring what is easy instead of what’s

important. For example:
Clicks to the beginning of the purchase pipeline rather than actual purchases (you may
change the probability of purchase given someone enters the pipeline).
McNamara’s Fallacy:
The first step is to measure what can easily be measured.
This is OK as far as it goes.
The second step is to disregard that which can't easily be measured or give it an arbitrary
quantitative value.
This is artificial and misleading.
The third step is to presume that which can't be measured easily really isn't important.
This is blindness.
The fourth step is to say that what can't be easily measured really does not exist.
This is suicide.

24
Roger Longbotham,
Principal Statistician, Microsoft

What to measure
How to compare Treatment to Control
How long to run test
Start up options
Good test design
Data validation and cleansing
Before your first experiment
Common errors
MultiVariable Tests
Advanced Topics

Start with objective

Of the site (content, ecommerce, marketing, help/support,…)
Of the experiment
What can you measure to tell you if you met your objective?
Content site: clicks/user, pageviews/user, time on site
Ecommerce: rev/visitor, units purchased/visitor, cart-adds/visitor
Marketing: referrals/visitor, time on site
Help/support: Pct of users engaged, Pct of users who print,
email or download content, time on site

1
Measures of user behavior
Number of events (clicks, pageviews, visits, downloads, etc)
Time (minutes per session, total time on site, time to load page)
Value (revenue, units purchased, ads clicked)

Analysis units
Per user (e.g. clicks per user)
Per session (e.g. minutes per session)
Per user-day (e.g. pageviews per day)
Per experiment (e.g. clicks per pageview)

It is very helpful to have a single metric that summarizes

whether the Treatment is successful or not – the Overall
Evaluation Criterion, or OEC
Examples:
Content site: OEC could be clicks/user or time on site
Ecommerce: rev/user or lifetime value
Help/support site: Survey responses
OEC could also capture monetary value of the site, aka ROI
(return on investment)

Experimentation Principle: all non-test factors should have

the same effect on the Treatment as on the Control

Online experiments don’t have stationarity

We would expect some metrics to vary by hour and day – such
as number of visitors and any metric related to that (number of
pageviews, clicks, etc.)
However, even those metrics that we think could be relatively
constant (stationary) are not

2
Metric plotted is number of clicks in an hour divided by the number of pageviews

Experimentation Principle: all non-test factors should have

the same effect on the Treatment as on the Control
Online experiments don’t have stationarity
Must run Treatment and Control concurrently
Days: 1 2 3 4 5 6 7 8
NOT C C C C T T T T
and
NOT C T C T C T C T

Randomly assign one set of users to T and other to C and keep

them in those groups throughout the experiment (ideal)

Took a valid experiment that ran for two weeks (50% in T,

50% in C)
Treatment effect in week 1: +1.93%
Treatment effect in week 2: +2.09%
For same experiment
Compare Treatment week 2 to Control week 1: +11.38%
Difference was due to uncontrolled differences between
weeks 1 and 2

3
Single Treatment
Two-sample t test works well
Large samples sizes => Normal distribution for means
Calculate 95% Confidence Interval (CI) for difference in two means

(ܺതܶ − ܺത‫ ) ܥ‬± 1.96 ∗ ‫ܺݏ‬തܶ −ܺത‫ܥ‬

if zero not in the interval conclude Treatment mean different from Control
May have many tests, OEC critical
Multiple Treatments
Multiple applications of two-sample t test
Analysis of Variance

Included:
Averages for
both variants

P-values

Percent change

Significance
Confidence
Intervals

103 metrics

P-value is the probability of getting a difference farther from

zero than observed under assumption of no difference
CI for percent effect must use special formula (Fieller’s
theorem)
Care must be taken in calculating standard deviations
When randomization is by user, any metric that is not per user
must take into account non-independence in calculating
standard deviation
We routinely use bootstrap or delta method to estimate standard
deviations for metrics that are not by user

4
 May need to consider
Sample size needed for normality of means

Sample size needed for sufficient power

Sample size needed for representative sample

By metric Normality of mean (criterion:

Sample size needed to assume normality max(LHS,RHS)<20%)
holds for confidence interval of the mean |Skewness| Min sample size
This table uses the |skewness| of the 0.05 8
distribution (sample) 0.1 12
The criterion used: that the probability of 0.2 18
being in the left or right tail of the CI (LHS, 0.4 60
RHS) is no more than 20% different from 0.7 175
nominal (.025), i.e. no larger than .03 and 1.0 350
no smaller than .02. 1.5 800
Most online metrics need 1000s for each 2.0 1400
group 3.0 3200
5.0 9000

16 * r *σ 2
n=
∆2

The power of a test is the probability of detecting a

difference (∆) of a given size i.e., it is 1-Prob(Type II error)

Power depends on
• The size of effect you want to be able to detect, ∆
• Variability of the metric
• Number of users in each group (T/C)

It is typical to determine the sample size needed to

achieve 80% power

5
Example: Number of users needed for each variant (group)
to achieve 80% power, with equal number of users in
Treatment and Control and with standard deviation s is
32 * s 2
N=
∆2

Often good practice is to start with small percent in

Treatment and increase when you have confidence
Treatment is bug-free
Sample ramp up schedule:
1% in Treatment for 4 hours
5% in Treatment for 4 hours
20% in Treatment for 4 hours
50 % in Treatment for 14 days

In an online experiment that ran for 35 days, the Treatment

had a click-through rate (CTR) of 29.6% and the Control had a
CTR of 30.9% for a -4.3% drop (Control did better)
When CTR effect is calculated by day => Treatment better
Treatment % Effect by Day
4.00%
Percent Difference in Click
Through Rates (T-C)/C

3.00%

2.00%

1.00%

0.00%

-1.00%

-2.00%
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35

What’s going on? Which is correct?

6
Can happen when numerators and denominators are accumulated
over groups of unequal sizes
Famous example: UC Berkeley gender bias lawsuit. The graduate
school was sued in 1973. Admission rates:
Applicants Admitted So, a larger percentage
Men 8442 44% of men admitted than
Women 4321 35% women

But for two large departments:

Two Departments
A C
Applicants/
Men Admitted
825 / 62% 325 / 37%
Applicants/
Women Admitted
108 / 82% 593 / 34%

In these two departments, admission rates did not seem to favor men, but when
combined, the admission rate for men was 55% and for women it was 41%

Why did the overall comparison show Treatment was 4.3%

worse than Control when 34 of 35 days Treatment was better?
Treatment and Control means and % in Treatment

50%

40%

30%

20%

10%

0%
1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35
Control Treatment % in Tmt

Example: Real Estate widget design

Test five alternatives to the current design
OEC: clicks to links weighted by revenue per click

Control T1 T2

T3 T4 T5

7
 The widget that performed the best was the simplest

Revenue increase over control: +9.7%

Note Ronny’s example earlier compared the best Treatment to another Treatment, not the Control

Triggering
Blocking
Measuring non-test factors
Randomization

Only allow users into your experiment if they “trigger” the

experiment. i.e. a user’s data should only be used in the
analysis of the experiment if they saw one of the variants
Example: MSN Hotmail experiment
Control: When user clicks on email
hotmail opens in same window
Treatment: Open hotmail in
separate window
Business Question: Will this increase user
engagement on MSN homepage?
Which users do you want to track as part of your
experiment?

8
Triggering
Blocking
Measuring non-test factors
Randomization

Factor is controlled such that it affects both treatment and

control equally, hence not affecting the estimate of the
effect
Blocking on a factor is more common than keeping it fixed
(keeping it constant throughout the experiment)
Advantages to blocking
Can estimate the effect of the different levels of the factor, e.g. what is
the effect on weekends/weekdays
Can make inference to a broader population

Time (time of day, day of week, etc.)

Always run treatment and control concurrently in online
experiments to block on all factors that vary with time

Content
Ex: If content of a site changes during the experiment it must be the
same for both Treatment and Control at all times

User
May use before/after for users or a cohort group

Updates to site

9
The Treatment and Control groups should be as alike as
possible except for application of the treatment
Who is in the experiment
What is done during the experiment
etc.

Example: One partner was conducting an A/A test (same as

an A/B test but no real change is made) What would you
expect?
Results: T very significant (much more than it should be)
Why?
Found out another group was using their Treatment group
to test something so there really was a difference between
T and C

Ex: A site was testing a change to the layout of their page

Content to T and C was not the same for a 7 hour period
Hourly Clickthrough Rate for Treatment and Control for Module
1.2%

1.0%
Clickthrough Rate

0.8%

0.6%

CTR_Control
0.4% CTR_Tmt

0.2%

0.0%
10/15/07 14:00

10/15/07 18:00

10/15/07 22:00

10/16/07 2:00

10/16/07 6:00

10/16/07 10:00

10/16/07 14:00

10/16/07 18:00

10/16/07 22:00

10/17/07 2:00

10/17/07 6:00

10/17/07 10:00

10/17/07 14:00

10/17/07 18:00

10/17/07 22:00

10/18/07 2:00

10/18/07 6:00

10/18/07 10:00

10/18/07 14:00

10/18/07 18:00

10/18/07 22:00

10/19/07 2:00

10/19/07 6:00

10/19/07 10:00

10/19/07 14:00

10
 Triggering
Blocking
Measuring non-test factors
Randomization

Measuring the value of non-test factors allows you to

• Delve into why the treatment had the effect it did (e.g. more PVs
are correlated with faster load time which explains almost all the
effect of T)
• Determine if subpopulations behave the same (e.g. did the
treatment have the same effect for new users as for experienced
users?)
Example: Add a couple of ads below the fold
• Expected no change
• T had significant negative effect on click-through rate – why?
• Suspect it made the page load slower – no evidence
• Later we added capability to measure time to load by PV.

Triggering
Blocking
Measuring non-test factors
Randomization

11
Why randomize?

So that those factors you can’t control (or don’t know

about) don’t bias your results
Unknown
Factors

“Randomization is too important to be left to chance”

Robert Coveyou, ORNL

How to randomize? (online tests)

Randomly assign T or C to user (alternately could use user-session,
search query, page view or product/SKU)
Usually best by user (store UserID in cookie)
How persistent is the UID?
Ideally user always gets same treatment group
Limitations:
Clearing cookies => can change treatment
Different browser => may get different treatment
Can’t allow opt-in or opt-out

Make sure users and conditions are as representative of

launch environment as possible
Time period: not holiday (unless holiday factor), pre-holiday,
complete cycle (day, week)
Users: all users who would see T in the future,
not robots, not internal testers, outliers(?)
Not during special events

12
Remove robots (web crawlers, spiders, etc.) from analysis
They can generate many pageviews or clicks in Treatment or
Control skewing the results
Remove robots with known identifiers (found in the user agent)
Develop heuristics to identify robots with many clicks or
pageviews in short period of time
Other patterns may be used to identify robots as well, such as
very regular activity

Carry out checks to make sure data is not affected by some

unknown factor
Check that percentage of users in each variant is not different
from planned (statistical test)
Check that number of users in the experiment is approximately
what was expected (and doesn’t change too much during
experiment)
Check that the Treatment effect does not change too much
during experiment
Check that means for primary metrics do not change
unexpectedly

Conduct logging audit

Compare data collected for experiment to system of record
Should have approximately same number of users, clicks,
pageviews, orders, etc.

Conduct A/A test

Split users into two groups that get same experience
Should have about 5% of tests significant
p-values should have U(0,1) distribution
No p-values should be extremely small (say <.001)

13
 Not conducting logging or A/A tests
Find caching issues, UID reassignment
Not keeping all factors constant or blocking
Content changes to site
Redirect for Treatment but not for Control
Sample size too small
Caution with using surrogates for OEC!
Measure clicks to buy button (instead of revenue)
Clicks to download button (instead of completed downloads)

Several factors/variables, each of which has two or more

levels (C/T1/T2/…)
Main effects: Comparison of Treatments to Control for each
variable (i.e. compare means for T and C same as before)
Interactions: Determine if combinations of variables have
different effect than adding main effects

Factors/variables

F1: Size of Right col ad

C = current size
T1 = 10% larger
T2 = 10% smaller
F1
F2: MSNBC news stories
C = Top international
F2 T = Specific to country ID’d

F3: Sports/Money placement

C = Sports above Money
T = Money above Sports
OEC: Clicks per User
F3 Other metrics: PVs, CTR

(This is for illustration purposes only, it does not reflect any previous or planned test on MSN HP)

14
Advantages:
– Can test many things at once, accelerating innovation
– Can estimate interactions between factors
Disadvantages
– Some combinations of factors may give negative customer
experience
– Analysis and interpretation is more difficult
– May take longer to set up test

Off-line experimental designs use specific subsets of all

combinations (fractional factorials, Plackett-Burman,…)*
Example: Test 7 factors each at 2 levels
27 = 128 combinations (full factorial)
27-4 fractional factorial has 8 combinations
can estimate main effects, but in order to
estimate two-factor interactions need
27-1 fractional factorial with 64 combinations

*sometimes known as Taguchi designs

On-line experiments can simply run overlapping, concurrent,

independently randomized experiments
Example: Test 7 factors each at 2 levels
Set up 7 separate experiments to run at the same time with
the same users. Get all 128 combinations in the results.
Advantages:
– Easier to implement
– Can turn off one experiment if negative
– Get all interactions

15
 Procedure for analyzing an MVT for interactions
1. Since there are potentially a vary large number of interactions
among the variables being tested, restrict the ones you will
look at to a few you suspect may be present. (If 7 factors, 21
two-factor interactions, 35 three-factor interactions, etc.)
2. Conduct the test to determine if the interaction between two
factors is present or not
3. If interaction is not significant, stop!
If the interaction IS significant, look at the graphical output to
interpret.

Example: Factors from MSN HP illustration

F2: MSNBC news stories F3 Sports/Money placement
C = Top international C = same order every day
T = Specific to country ID’d T = Sports higher on wkends
and Money higher wkdays

Hypothesis tests for interactions similar to main effects

(details omitted)

Factors/variables

F2: MSNBC news stories

C = Top international
T = Specific to country ID’d

F3: Sports/Money placement

C = Sports above Money
F2 T = Money above Sports

OEC: Clicks per User

Other metrics: PVs, CTR
F3

(This is for illustration purposes only, it does not reflect any previous or planned test on MSN HP)

16
 If hypothesis test for interaction is not significant
Assume no interaction present
Interaction graph would show lines approximately parallel
If interaction is statistically significant
Plot interaction to interpret

Case 1: No Interaction (parallel lines)

Data Table Main Effects Results

F2 - C F2 - T Pct Effect p-value

F3 - C 4.06 4.10 Effect(F2) 0.98% <.001
F3 - T 4.08 4.12 Effect(F3) 0.49% 0.032
F2xF3 Interaction
4.13
No Interaction
4.12
Average Clicks per User

4.11
4.10
4.09 F3 - C
4.08 F3 - T
4.07
4.06
4.05
F2 - C F2 - T

When interaction is statistically significant

Two types of interactions:
Synergistic – when the presence of both is more than the sum of
the individual treatments

Antagonistic – when the presence of both is less than the sum of

the individuals

17
Case 2: Synergistic Interaction
Data Table Main Effects Results

F2 - C F2 - T Pct Effect p-value

F3 - C 4.08 4.09 Effect(F2) 0.74% 0.008
F3 - T 4.08 4.13 Effect(F3) 0.49% 0.032

F2xF3
F2xF3Interaction
Interaction
4.14
4.14
Synergistic
Synergistic Interaction
Interaction
4.13
4.13
UserHP
MSN

4.12
4.12
per
Visit

4.11
4.11
Clicks

F3 - C
Days

4.10
4.10 F3 - C
Average

F3 - T
Number

4.09
4.09

4.08
4.08

4.07
4.07
Microsoft Confidential
F2
F2 -- C
C F2
F2--TT

Case 3: Antagonistic Interaction

Data Table Main Effects Results
F2 - C F2 - T
Pct Effect p-value
F3 - C 4.08 4.11
Effect(F2) 0.18% 0.396
F3 - T 4.12 4.11
Effect(F3) 0.55% 0.028
F2xF3
F2xF3 Interaction
Interaction
4.14
Antagonistic Interaction
4.13
HP
User
MSN

4.12
per
Visit

4.11
Clicks
Days

4.10 F3F3
-C-C
Average
Number

4.09 F3 - T

4.08

4.07
F2 - C Confidential
Microsoft F2
F2-- TT

Variance calculations for metrics

Non-parametric alternatives to t-test, ANOVA
Robot detection
Automatic detection of interesting population segments
Experimentation with exploration/exploitation schemes

18
For metrics that are not “per user” (i.e. not the same as the
randomization unit) – cannot use usual standard dev
formula
Can use bootstrap or delta method to estimate variance
Delta method uses a formula to take into account correlation of
experimental units
Example: Clickthrough rate (CTR) per experiment

True variance is much larger than that from Binomial distribution

Permutation or Mann-Whitney tests are natural

Pros
Can get a p-value
May have better power for some metrics
Works better for small sample sizes
Cons
Understandability by business managers
Not a test of a shift of a change in the mean
Confidence intervals for effect not straight-forward

What is “best” way to develop heuristics to detect robots?

What is “best” way to assess how well heuristics are doing?
How to adjust robot detection parameters based on the
specific site in the test?
For example
Sites with low traffic may need more aggressive robot filtering
Sites that have more robot traffic may need more aggressive
robot filtering
Sites that expect active users (e.g. many clicks per hour) need
less aggressive robot filtering
Sites that require users to sign-in should have fewer robots so
less aggressive robot filtering required

19
A population segment is interesting if their response to the
Treatment is different from the overall response
Segments can be defined by a number of variables
Browser or operating system
Referrer (e.g. from search engine, etc.)
Signed-in status
Loyalty
Demographics
Location – country, state, size of city (use IP lookup)
Bandwidth

Want to automatically display best content based on

exploration/exploitation strategy
Is this strategy better than editor-placed content?
What are the optimal parameter values?
Percent in exploration group?
How long to test content in exploration group?
What level of significance is needed?

BREAK

20
 Ronny Kohavi, Microsoft

Slides available at http://exp-platform.com

Wrap-up section based on KDD 2012 paper, co-authored by

Ronny Kohavi, Alex Deng, Brian Frasca, Roger
Longbotham, Toby Walker, and Ya Xu
How does one determine the OEC for a search engine?
What are some of the most surprising results we faced, and
how did we resolve them

An OEC is the Overall Evaluation Criterion

It is a metric (or set of metrics) that guides the org as to
whether A is better than B in an A/B test
In prior work, we emphasized long-term focus and thinking
about customer lifetime value, but operationalizing it is hard
Search engines (Bing, Google) are evaluated on query share
(distinct queries) and revenue as long-term goals
Puzzle
A ranking bug in an experiment resulted in very poor search
results
Distinct queries went up over 10%, and revenue went up over 30%
What metrics should be in the OEC for a search engine?

1
 4

Degraded (algorithmic) search results cause users to search more

to complete their task, and ads appear more relevant
Analyzing queries per month, we have


   
 
   
where a session begins with a query and ends with 30-minutes of inactivity.
(Ideally, we would look at tasks, not sessions).
Key observation: we want users to find answers and complete tasks
quickly, so queries/session should be smaller
In a controlled experiment, the variants get (approximately) the
same number of users by design, so the last term is about equal
The OEC should therefore include the middle term: sessions/user

A piece of code was added, such that when a user clicked

on a search result, additional JavaScript was executed
(a session-cookie was updated with the destination)
before navigating to the destination page
This slowed down the user experience slightly, so we
expected a slightly negative experiment.
Results showed that users were clicking more!

Why?

User clicks (and form submits) are instrumented and form

the basis for many metrics
Instrumentation is typically done by having the web browser
request a web beacon (1x1 pixel image)
Classical tradeoff here
Waiting for the beacon to return slows the action (typically
navigating away)
Making the call asynchronous is known to cause click-loss, as
the browsers can kill the request (classical browser optimization
because the result can’t possibly matter for the new page)
Small delays, on-mouse-down, or redirect are used

2
 7

Click-loss varies dramatically by browser

Chrome, Firefox, Safari are aggressive at terminating such
reqeuests. Safari’s click loss > 50%.
IE respects image requests for backward compatibility
reasons
White paper available on this issue here
Other cases where this impacts experiments
Opening link in new tab/window will overestimate the click delta
Because the main window remains open, browsers can’t
optimize and kill the beacon request, so there is less click-loss
Using HTML5 to update components of the page instead of
refreshing the whole page has the overestimation problem

Primacy effect occurs when you change the navigation on a

web site
Experienced users may be less efficient until they get used to the
new navigation
Control has a short-term advantage
Novelty effect happens when a new design is introduced
Users investigate the new feature, click everywhere, and introduce
a “novelty” bias that dies quickly if the feature is not truly useful
Treatments have a short-term advantage

Given the high failure rate of ideas, new experiments are

followed closely to determine if new idea is a winner
Multiple graphs of effect look like this
Cumulative Effect
Negative on day 1: -0.55% 0.40%

Less negative on day 2: -0.38% 0.00%

-0.40%
Less negative on day 3: -0.21% -0.80%

Less negative on day 4: -0.13% -1.20%

8/30/2011

8/31/2011

9/1/2011

9/2/2011

9/3/2011

9/4/2011

The experimenter extrapolates linearly

and says: primacy effect.
This will be positive in a couple of days, right?
Wrong! This is expected

3
 10

For many metrics, the standard deviation of the mean is

proportional to 1 ⁄ √ , where is the number of users
As we run an experiment longer, more users are admitted
into the experiment, so grows and the conf interval
shrinks 0.80%

The first days are highly variable 0.60%

0.40%

The first day has a 67% chance 0.20%

Effect

0.00%
of falling outside the 95% CI -0.20%
0 5 10 15 20

at the end of the experiment -0.40%

-0.60%

The second day has a 55% chance -0.80%

Experiment Days

of falling outside this bound. 95% bound 21-day bound

11

The longer graph

Cumulative Effect
0.60%

0.20%

-0.20%

-0.60%

-1.00%

-1.40%

This was an A/A test, so the true effect is 0

12

We expect the standard deviation of the mean (and thus the

confidence interval) to be proportional to 1 ⁄ √ ,
where is the number of users
So as the experiment runs longer and more users are
admitted, the confidence interval should shrink
Here is the graph for sessions/user
X-axis: Treatment size 2
Confidence Inerval
Width (percent)

Y-axis: conf interval

1
Three lines: 1,2,3 weeks
Overlapping lines? 0
That’s the problem! 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
Size of Treatment (relative factor)
1 week 2 weeks 3 weeks

4
 13

The distribution changes

Users churn, so they contribute zero visits
New users join with fresh count of one
We have a mixture
Empirically, the coefficient of variation (ratio of the standard
deviation to the mean) grows at the same rate as √
Running an experiment longer does not increase statistical
power for some metrics; you must increase the variant size

14

Experiment is run, results are surprising.

(This by itself is fine, as our intuition is poor.)
Rerun the experiment, and the effects disappear
Reason: bucket system recycles users, and the prior
experiment had carryover effects
These can last for months!
Must run A/A tests, or re-randomize

15

OEC: evaluate long-term goals through short-term metrics

The difference between theory and practice is greater in
practice than in theory
Instrumentation issues (e.g., click-tracking) must be understood
Carryover effects impact “bucket systems” used by Bing, Google,
and Yahoo require rehashing and A/A tests
Experimentation insight:
Effect trends are expected
Longer experiments do not increase power for some metrics.
Fortunately, we have a lot of users