In the last decade predictive modeling has changed American political campaigns, especially at the presidential level. Long before Election Day 2012, Obama campaign staffers were confident that President Obama would be re-elected because they had sophisticated modeling predicting wins in many important states. More importantly, modeling helps political campaigns learn which voters to target with particular messages. This session will summarize predictive modeling in American politics, with an eye toward the way it might be developed for international applications.
3. That was mostly all true!
The
Obama
campaign
used
predic8ve
analy8cs
to:
◦ Contact
voters
more
efficiently
◦ Track
our
real
status
vs.
Romney
◦ BeCer
media
targe8ng
◦ Raise
more
money
But
we
didn’t
invent
modeling
There
is
a
long
history
of
analy8cs
in
U.S.
poli8cs
5. The American voter file
First,
there
was
the
voter
file
◦ For
a
very
long
8me,
each
U.S.
state
has
kept
a
semi-‐public
data
file
of
all
registered
voters
◦ This
file
contains
each
voter’s
name,
address,
age,
gender,
some8mes
race,
some8mes
party
registra8on
◦ Also
shows
each
person’s
vote
history
–
which
elec8ons
did
they
vote
in?
6. The American voter file
In
the
1960s,
70s,
80s,
and
90s,
poli8cal
campaigns
started
to
use
the
voter
file
to
iden8fy
broad
groups
to
target
◦ Example:
Send
a
piece
of
mail
to
all
women
over
40
who
have
voted
in
at
least
three
of
the
last
four
elec8ons,
convincing
them
to
vote
for
your
candidate
At
some
point,
people
figured
out
you
could
enhance
the
voter
file
◦ Example:
Census
block
à
average
income
in
the
neighborhood
◦ Example:
Commercial
data
matches,
public
records
matches
7. Predic;ve analy;cs in American poli;cs
In
the
2000s,
poli8cal
opera8ves
started
developing
models
The
steps
of
building
a
model:
◦ Conduct
a
massive
voter
survey
(5000+)
◦ Ask
about
candidates
or
issues
◦ Use
voter
file
informa8on
to
make
models
◦ Age,
gender,
vote
history,
Census
variables,
etc
◦ Decision
tree
models,
regression
models,
etc
◦ Validate
on
a
held-‐out
subsample
◦ Assign
a
model
score
to
en8re
voter
file
???
8. Micro-‐targe;ng
We
use
models
to
“micro-‐target”
voters
to
receive
different
types
of
contact
◦ Encouragement
to
vote
◦ Persuasion
to
vote
for
your
candidate
◦ Recrui8ng
volunteers
◦ Voter
suppression
(joking!)
Likelihood
of
vo8ng
Support
for
your
candidate
GOTV
Persuasion
Volunteer
recruitment
9. Issue
modeling
and
other
innova8ons
We
can
model
almost
anything!
◦ Environmentalism
◦ Women’s
rights
◦ Religiosity
We
can
even
model
who
is
easy
to
persuade
11. Nate Silver and the 2008 elec;on
Campaigns
were
not
the
only
ones
using
predic8ve
analy8cs
In
2008
a
guy
named
Nate
Silver
(and
other
nerds)
started
using
public
polls
to
run
Monte
Carlo
simula8ons
of
the
presiden8al
elec8on,
making
predic8ons
that
were
quite
accurate
◦ Dozens
of
polls,
each
with
n=400-‐1000
◦ Simula8on
accounts
for
each
poll’s
MOE
◦ Also
accounts
for
each
pollster’s
quality/accuracy
12. More uses for Monte Carlo simula;ons
I
built
models
to
predict
likely
outcomes
in
state
legisla8ve
elec8ons
in
Oregon
and
Alabama
◦ Linear
regression
model
at
the
precinct-‐level,
using
past
elec8on
results
and
other
variables
These
results
were
used
to
channel
money
toward
districts
where
it
would
make
the
biggest
impact
14. Data and modeling, to the max!
We
did
all
of
that,
and
more
The
2012
Obama
campaign
had
a
huge
data
and
analy8cs
team
◦ Analy8cs
department:
50+
people
◦ Data
team:
20+
people
◦ Digital
analy8cs:
15
people
◦ Tech
team:
30+
people
15. Data and modeling, to the max!
Television
targe8ng
◦ What
TV
programs
are
best?
What
geographical
zones?
Fundraising
(online
and
offline)
Models
for
persuasion,
turnout,
issues,
etc
◦ Direct
mail
◦ Online
adver8sing
◦ Phone
calls
and
door
knocking
by
volunteers
16. Data-‐driven volunteers
From
2012
Campaign
Manager
Jim
Messina:
“My
favorite
story
is
from
a
volunteer
in
Wisconsin
10
days
out
[from
Elec8on
Day].
She
was
knocking
on
doors
on
one
side
of
the
street
and
the
Romney
campaign
was
knocking
on
doors
on
the
other
side
of
the
street…”
17. Data-‐driven conversa;ons
“…
[The
Obama
volunteer]
was
asked
to
hit
two
doors.
One
was
an
undecided
voter
and
she
knew
exactly
what
to
say.
The
other
was
an
absentee
ballot
and
she
was
told
to
make
sure
they
filled
it
out
and
returned
it.
On
the
other
side
of
the
street,
the
Romney
campaign
was
knocking
on
every
single
door.
Most
of
the
people
weren’t
home,
and
most
of
the
people
that
were
home
were
already
suppor8ng
Barack
Obama.
She
looked
at
me
and
said,
‘You’re
using
my
8me
wisely.’
That’s
what
data
can
do.”
-‐
Obama
2012
Campaign
Manager
Jim
Messina
18. Our own internal Nate Silver-‐style modeling
On
the
day
of
the
elec8on
in
2012,
we
knew
we
would
win
◦ Our
internal
modeling
bounced
around
less
than
Nate
Silver’s
22. Email
tes8ng:
subject
lines
version
Subject
line
v1s1
Hey
v1s2
Two
things:
v1s3
Your
turn
v2s1
Hey
v2s2
My
opponent
v2s3
You
decide
v3s1
Hey
v3s2
Last
night
v3s3
Stand
with
me
today
v4s1
Hey
v4s2
This
is
my
last
campaign
v4s3
[NAME]
v5s1
Hey
v5s2
There
won't
be
many
more
of
these
deadlines
v5s3
What
you
saw
this
week
v6s1
Hey
v6s2
Let's
win.
v6s3
Midnight
deadline
Test sends
6 drafts x 3 subject lines
=
18 possible versions
23. Email
tes8ng:
gexng
results
version
Subject
line
donors
money
v1s1
Hey
263
$17,646
v1s2
Two
things:
268
$18,830
v1s3
Your
turn
276
$22,380
v2s1
Hey
300
$17,644
v2s2
My
opponent
246
$13,795
v2s3
You
decide
222
$27,185
v3s1
Hey
370
$29,976
v3s2
Last
night
307
$16,945
v3s3
Stand
with
me
today
381
$25,881
v4s1
Hey
444
$25,643
v4s2
This
is
my
last
campaign
369
$24,759
v4s3
[NAME]
514
$34,308
v5s1
Hey
353
$22,190
v5s2
There
won't
be
many
more
of
these
deadlines
273
$22,405
v5s3
What
you
saw
this
week
263
$21,014
v6s1
Hey
363
$25,689
v6s2
Let's
win.
237
$17,154
v6s3
Midnight
deadline
352
$23,244
$0
$1
$2
$3
$4
ACTUAL
($3.7m)
IF
SENDING
AVG
IF
SENDING
WORST
Full send (in millions)
¨ $2.2
million
addi8onal
revenue
from
sending
best
draz
vs.
worst,
or
$1.5
million
addi8onal
from
sending
best
vs.
average
Test sends
24. Results of the online campaign
Campaign
raised
over
one
billion
dollars,
half
of
which
was
raised
online
◦ Over
4
million
Americans
donated
Recruited
tens
of
thousands
of
volunteers,
publicized
thousands
of
events
and
rallies
Did
I
men8on
raising
>$500
million
online?
◦ Conserva8vely,
tes8ng
probably
resulted
in
~$200
million
in
addi8onal
revenue
27. The 2016 U.S. Presiden;al Elec;on
The
Democrats
◦ Hillary
Clinton
will
probably
be
the
Democra8c
nominee
◦ Clinton
will
have
a
huge
analy8cs
team,
with
many
Obama
alums
The
Republicans
◦ Whoever
wins
the
Republican
nomina8on
will
make
a
strong
effort
to
build
a
data
and
analy8cs
team
(well,
maybe
not
Trump)
◦ In
2012
the
Romney
campaign’s
analysts
and
pollsters
failed
spectacularly,
and
the
Republicans
do
not
want
that
to
happen
again
28. Opportuni;es for enterprises
Poli8cs
and
social
movements
are
huge
opportuni8es
for
the
data
and
technology
industries
◦ US
poli8cal
analy8cs
industry
growing
◦ Other
countries
are
learning
from
the
U.S.
example
29. Opportuni;es for enterprises
Supply
beCer
data
◦ In
the
US
and
everywhere
else,
good
models
require
good
data
Supply
the
first
voter
file
◦ In
countries
where
voter
files
are
not
common,
the
first
par8es
or
advocacy
organiza8ons
to
get
them
will
have
a
huge
advantage
Supply
the
first
micro-‐targe8ng
model!