MLCONF SEATTLE — MAY 1, 2015
A large scale online natural experiment
Measuring causal impact of display ads 
Robert Moakler — rmoakler@stern.nyu.edu | robert@integralads.com
@ MLconf Seattle 2015
MLCONF SEATTLE — MAY 1, 2015
The $100+ billion question!

Does online advertising really work? 
$104.57
$120.05
$140.15
$160.18
$178.45
$196.05
$213.89Digital ad spending!
% change!
2012 2013 2014 2015 2016 2017 2018!
Source: www.emarketer.com, “Global Ad Spending Growth to Double This Year”
20.4%
14.8% 16.7% 14.3%
11.4% 9.9% 9.1%
MLCONF SEATTLE — MAY 1, 2015
The $100+ billion question!

Does online advertising really work?
MLCONF SEATTLE — MAY 1, 2015
The $100+ billion question!

Does online advertising really work? 

Do online ads cause you to take some action?
MLCONF SEATTLE — MAY 1, 2015
Measuring causal impact!
Option 1: Randomized A/B test

•  Pros
–  If setup correctly, gives unbiased
causal estimates 
•  Cons
–  Control ads cost as much as real
ones
–  Planned before campaign starts
–  Coordination of multiple media
partners
–  Too many levers to test them all

Campaign AdPSA
MLCONF SEATTLE — MAY 1, 2015
Measuring causal impact!
Option 2: Observational study

•  Pros
–  Cheap
–  Flexible
•  Cons
–  Enormous amount of selection bias
MLCONF SEATTLE — MAY 1, 2015
Confounding in digital advertising campaigns!
•  Why is there selection bias in observational techniques?
–  Online ads are targeted to specific segments of the population based on
particular demographics, user interests and behaviors, etc.
–  Targeting ads to specific populations makes comparing users that have
received an ad to those that did not very problematic; estimates of causal
impact will be overestimated.
MLCONF SEATTLE — MAY 1, 2015
Confounding in digital advertising campaigns!
•  Why is there selection bias in observational techniques?
W
User
features
A
Served
ads
Y
Convert
MLCONF SEATTLE — MAY 1, 2015
Confounding in digital advertising campaigns!
•  Why is there selection bias in observational techniques?
W
User
features
A
Served
ads
Y
Convert
Unless we know what information
targeters are using, we will never be
able to fully adjust for selection bias.
MLCONF SEATTLE — MAY 1, 2015
Viewability!
•  Web page layout, ad placement details, and user browsing behavior
and setup can all impact the way in which ads are seen online.
–  Some ads are served far down on the page (below the fold)
–  Ads can be loaded in hidden tabs or windows
–  Users may not stay on a page long enough for it to finish loading
MLCONF SEATTLE — MAY 1, 2015
Viewability!
MLCONF SEATTLE — MAY 1, 2015
Viewability!
MLCONF SEATTLE — MAY 1, 2015
Viewability!
MLCONF SEATTLE — MAY 1, 2015
Viewability as a natural experiment!
Introduce a mediating variable — viewability
W
User
features
A
Served
ads
Y
Convert
V
Viewable
ad
MLCONF SEATTLE — MAY 1, 2015
Methodology!
Conversion (Y=1)
Web page visit
Effect window
T0
Untreated user
Treated user
Viewable ad (V=1)
Unviewable ad (V=0)
MLCONF SEATTLE — MAY 1, 2015
Methodology!
Conversion (Y=1)
Web page visit
Effect window
T0
Untreated user
Treated user
Viewable ad (V=1)
Unviewable ad (V=0)
Parameter of interest
MLCONF SEATTLE — MAY 1, 2015
Data!
•  Seven display advertising campaigns run during the 4th quarter of
2014
–  Diverse industries such as auto insurance, beauty products, finance, and
online marketing 
–  3 million - 29 million impressions
–  2,000 - 2 million conversions
MLCONF SEATTLE — MAY 1, 2015
Using viewability as a natural experiment!
Compared to the naïve analysis of comparing users that were served and
not served ads, we find a drastic decrease in estimated lift when utilizing
viewability.
MLCONF SEATTLE — MAY 1, 2015
Validation!
•  How do we know a reduction in lift means our new estimates are
correct?
•  Use negative control tests
–  Use the impressions of one campaign to predict an unrelated conversion
MLCONF SEATTLE — MAY 1, 2015
Validation!
•  How do we know a reduction in lift means our new estimates are
correct?
•  Use negative control tests
–  Use the impressions of one campaign to predict an unrelated conversion 
W
User
features
A
Served
ads
Y
Convert
V
Viewable
ad
Y-
Unrelated
outcome
MLCONF SEATTLE — MAY 1, 2015
Validation!
Focusing on Campaign B from the previous example, we measure the
ads’ impact on unrelated outcomes
MLCONF SEATTLE — MAY 1, 2015
Bias in the natural experiment!
•  We don’t see zero effect on many of our negative controls
–  There can be other factors that affect viewability and conversion that we don’t
account for
MLCONF SEATTLE — MAY 1, 2015
Bias in the natural experiment!
•  We don’t see zero effect on many of our negative controls
–  There can be other factors that affect viewability and conversion that we don’t
account for
W
User
features
A
Served
ads
Y
Convert
V
Viewable
ad
W’
User
features
MLCONF SEATTLE — MAY 1, 2015
Bias in the natural experiment!
•  We don’t see zero effect on many of our negative controls
–  There can be other factors that affect viewability and conversion that we don’t
account for
W
User
features
A
Served
ads
Y
Convert
V
Viewable
ad
W’
User
features
Parameter of interest
MLCONF SEATTLE — MAY 1, 2015
Validation!
Returning to Campaign B from the previous example, we measure the
ads’ impact on irrelevant outcomes
MLCONF SEATTLE — MAY 1, 2015
Summary!
•  Viewability enables a natural experiment
–  Combines the benefits of A/B tests and observational analysis
–  Adjustment for viewability features is easier than adjusting for targeting features
–  Results in a large reduction in bias
•  Negative controls allow for validation of models when the true value being
estimated is unknown 
–  As the true effect of a natural experiment is usually unknown, negative controls provide a
method for validation
MLCONF SEATTLE — MAY 1, 2015
Versatility!
•  Features that can be used in a natural experiment can be found in data sets from a
wide array of industries
–  Viewability of stories in a user’s news feed
–  Listening to songs on shuffle
–  Winning bids in online advertising real-time bidding systems
•  Valid negative controls naturally exist in many industries
–  Purchasing unrelated products
–  Clicking unrelated links
MLCONF SEATTLE — MAY 1, 2015
Acknowledgments!
Integral Ad Science
Daniel Hill
Ekaterina Eliseeva
Gijs Joost Brouwer
Kiril Tsemekhman
NYU Stern
Foster Provost

UC Berkley
Alan Hubbard
MLCONF SEATTLE — MAY 1, 2015
Thanks!
Robert Moakler — rmoakler@stern.nyu.edu | robert@integralads.com

Robert Moakler, Data Science Intern, Integral Ad Science at MLconf SEA - 5/01/15

  • 1.
    MLCONF SEATTLE —MAY 1, 2015 A large scale online natural experiment Measuring causal impact of display ads Robert Moakler — rmoakler@stern.nyu.edu | robert@integralads.com @ MLconf Seattle 2015
  • 2.
    MLCONF SEATTLE —MAY 1, 2015 The $100+ billion question! Does online advertising really work? $104.57 $120.05 $140.15 $160.18 $178.45 $196.05 $213.89Digital ad spending! % change! 2012 2013 2014 2015 2016 2017 2018! Source: www.emarketer.com, “Global Ad Spending Growth to Double This Year” 20.4% 14.8% 16.7% 14.3% 11.4% 9.9% 9.1%
  • 3.
    MLCONF SEATTLE —MAY 1, 2015 The $100+ billion question! Does online advertising really work?
  • 4.
    MLCONF SEATTLE —MAY 1, 2015 The $100+ billion question! Does online advertising really work? Do online ads cause you to take some action?
  • 5.
    MLCONF SEATTLE —MAY 1, 2015 Measuring causal impact! Option 1: Randomized A/B test •  Pros –  If setup correctly, gives unbiased causal estimates •  Cons –  Control ads cost as much as real ones –  Planned before campaign starts –  Coordination of multiple media partners –  Too many levers to test them all Campaign AdPSA
  • 6.
    MLCONF SEATTLE —MAY 1, 2015 Measuring causal impact! Option 2: Observational study •  Pros –  Cheap –  Flexible •  Cons –  Enormous amount of selection bias
  • 7.
    MLCONF SEATTLE —MAY 1, 2015 Confounding in digital advertising campaigns! •  Why is there selection bias in observational techniques? –  Online ads are targeted to specific segments of the population based on particular demographics, user interests and behaviors, etc. –  Targeting ads to specific populations makes comparing users that have received an ad to those that did not very problematic; estimates of causal impact will be overestimated.
  • 8.
    MLCONF SEATTLE —MAY 1, 2015 Confounding in digital advertising campaigns! •  Why is there selection bias in observational techniques? W User features A Served ads Y Convert
  • 9.
    MLCONF SEATTLE —MAY 1, 2015 Confounding in digital advertising campaigns! •  Why is there selection bias in observational techniques? W User features A Served ads Y Convert Unless we know what information targeters are using, we will never be able to fully adjust for selection bias.
  • 10.
    MLCONF SEATTLE —MAY 1, 2015 Viewability! •  Web page layout, ad placement details, and user browsing behavior and setup can all impact the way in which ads are seen online. –  Some ads are served far down on the page (below the fold) –  Ads can be loaded in hidden tabs or windows –  Users may not stay on a page long enough for it to finish loading
  • 11.
    MLCONF SEATTLE —MAY 1, 2015 Viewability!
  • 12.
    MLCONF SEATTLE —MAY 1, 2015 Viewability!
  • 13.
    MLCONF SEATTLE —MAY 1, 2015 Viewability!
  • 14.
    MLCONF SEATTLE —MAY 1, 2015 Viewability as a natural experiment! Introduce a mediating variable — viewability W User features A Served ads Y Convert V Viewable ad
  • 15.
    MLCONF SEATTLE —MAY 1, 2015 Methodology! Conversion (Y=1) Web page visit Effect window T0 Untreated user Treated user Viewable ad (V=1) Unviewable ad (V=0)
  • 16.
    MLCONF SEATTLE —MAY 1, 2015 Methodology! Conversion (Y=1) Web page visit Effect window T0 Untreated user Treated user Viewable ad (V=1) Unviewable ad (V=0) Parameter of interest
  • 17.
    MLCONF SEATTLE —MAY 1, 2015 Data! •  Seven display advertising campaigns run during the 4th quarter of 2014 –  Diverse industries such as auto insurance, beauty products, finance, and online marketing –  3 million - 29 million impressions –  2,000 - 2 million conversions
  • 18.
    MLCONF SEATTLE —MAY 1, 2015 Using viewability as a natural experiment! Compared to the naïve analysis of comparing users that were served and not served ads, we find a drastic decrease in estimated lift when utilizing viewability.
  • 19.
    MLCONF SEATTLE —MAY 1, 2015 Validation! •  How do we know a reduction in lift means our new estimates are correct? •  Use negative control tests –  Use the impressions of one campaign to predict an unrelated conversion
  • 20.
    MLCONF SEATTLE —MAY 1, 2015 Validation! •  How do we know a reduction in lift means our new estimates are correct? •  Use negative control tests –  Use the impressions of one campaign to predict an unrelated conversion W User features A Served ads Y Convert V Viewable ad Y- Unrelated outcome
  • 21.
    MLCONF SEATTLE —MAY 1, 2015 Validation! Focusing on Campaign B from the previous example, we measure the ads’ impact on unrelated outcomes
  • 22.
    MLCONF SEATTLE —MAY 1, 2015 Bias in the natural experiment! •  We don’t see zero effect on many of our negative controls –  There can be other factors that affect viewability and conversion that we don’t account for
  • 23.
    MLCONF SEATTLE —MAY 1, 2015 Bias in the natural experiment! •  We don’t see zero effect on many of our negative controls –  There can be other factors that affect viewability and conversion that we don’t account for W User features A Served ads Y Convert V Viewable ad W’ User features
  • 24.
    MLCONF SEATTLE —MAY 1, 2015 Bias in the natural experiment! •  We don’t see zero effect on many of our negative controls –  There can be other factors that affect viewability and conversion that we don’t account for W User features A Served ads Y Convert V Viewable ad W’ User features Parameter of interest
  • 25.
    MLCONF SEATTLE —MAY 1, 2015 Validation! Returning to Campaign B from the previous example, we measure the ads’ impact on irrelevant outcomes
  • 26.
    MLCONF SEATTLE —MAY 1, 2015 Summary! •  Viewability enables a natural experiment –  Combines the benefits of A/B tests and observational analysis –  Adjustment for viewability features is easier than adjusting for targeting features –  Results in a large reduction in bias •  Negative controls allow for validation of models when the true value being estimated is unknown –  As the true effect of a natural experiment is usually unknown, negative controls provide a method for validation
  • 27.
    MLCONF SEATTLE —MAY 1, 2015 Versatility! •  Features that can be used in a natural experiment can be found in data sets from a wide array of industries –  Viewability of stories in a user’s news feed –  Listening to songs on shuffle –  Winning bids in online advertising real-time bidding systems •  Valid negative controls naturally exist in many industries –  Purchasing unrelated products –  Clicking unrelated links
  • 28.
    MLCONF SEATTLE —MAY 1, 2015 Acknowledgments! Integral Ad Science Daniel Hill Ekaterina Eliseeva Gijs Joost Brouwer Kiril Tsemekhman NYU Stern Foster Provost UC Berkley Alan Hubbard
  • 29.
    MLCONF SEATTLE —MAY 1, 2015 Thanks! Robert Moakler — rmoakler@stern.nyu.edu | robert@integralads.com