aapc deck intro to modeling draft 6.6.14

24
BlueLabs www.bluelabs.com @Blue_Labs

Upload: bluelabskat

Post on 06-Jul-2015

55 views

Category:

Technology


0 download

TRANSCRIPT

Page 1: Aapc deck intro to modeling draft 6.6.14

BlueLabs

www.bluelabs.com@Blue_Labs

Page 2: Aapc deck intro to modeling draft 6.6.14

Predictive Analytics:

Impact Field Programs:

Determine priority targets for volunteer contact and Election day

turn out.

Direct Mail:

Determine priority households to receive issue-specific mail.

Social Media:

Target online persuasion efforts toward the most persuadable

voters.

Television Advertising

Produce a list of efficient buys using the TV optimizer based on what

persuadable voters are watching.

Predictive

Analytics

Page 3: Aapc deck intro to modeling draft 6.6.14

Support

Turnout

Volunteer / Donate

Persuasion

Message / Issue

Channel

• Likelihood to support Democrat

• Likelihood to vote in the election

• Likelihood to volunteer or donate

• Likelihood to switch vote to Dem after exposure to campaign message

• Identify the most effective message

• Identify most effective channel or tactic to influence voter behavior

Who to target

What to say

How to contact

Core campaign data models What the data models predict Applications

Experiments• Test real-world interventions to

evaluate the impact of programsInformed using real world tests

Modeling Opinions and Behaviors

Page 4: Aapc deck intro to modeling draft 6.6.14

Example: Modeling Support

1. Collect Data

Select a random sample of voters from the population.

Page 5: Aapc deck intro to modeling draft 6.6.14

Example: Modeling Support

1. Collect Data

Select a random sample of voters from the population.

Page 6: Aapc deck intro to modeling draft 6.6.14

Example: Modeling Support

1. Collect Data

Collect relevant data on a sample voters.

“Who are you voting for?”

Page 7: Aapc deck intro to modeling draft 6.6.14

Example: Modeling Support

1. Collect Data

D RDR

DDR R

Collect relevant data on a sample voters.

“Who are you voting for?”

Page 8: Aapc deck intro to modeling draft 6.6.14

Example: Modeling Support

1. Collect Data

D RDR

DDR R

Collect relevant data on a sample voters.

“Who are you voting for?”

2. Build Model

Under 30

Union Member

Hispanic

Build statistical model to identify significant data

points

Hunter

40-49 years old

Registered Republican

Using data from a voter file, appended to additional data sources, we identify characteristics that are correlated with support of the Democratic candidate.

Page 9: Aapc deck intro to modeling draft 6.6.14

Example: Modeling Support

1. Collect Data

D RDR

DDR R

Collect relevant data on a sample voters.

“Who are you voting for?”

2. Build Model

Under 30

Union Member

Hispanic

Build statistical model to identify significant data

points

Hunter

40-49 years old

Registered Republican

3. Predict Outcome

In the original universe, predict likelihood of support

for each individual.

Page 10: Aapc deck intro to modeling draft 6.6.14

Example: Modeling Support

1. Collect Data

D RDR

DDR R

Collect relevant data on a sample voters.

“Who are you voting for?”

2. Build Model

Under 30

Union Member

Hispanic

Build statistical model to identify significant data

points

Hunter

40-49 years old

Registered Republican

3. Predict Outcome

In the original universe, predict likelihood of support

for each individual.

Page 11: Aapc deck intro to modeling draft 6.6.14

Using in cycle testing and experiments, we build experimentally-informed models that predict who is most likely to change their

vote

Next Generation: Combining Modeling & Testing

1. Conduct Experiment 2. Build Model

D

D

Treatment

Control

3. Predict Outcome

R D R

DR R R

Page 12: Aapc deck intro to modeling draft 6.6.14

Persuasion Scores ID Targets at

individual level

Strong TargetsAvoid Weak Targets

Example Likely Voter Universe by Persuasion Score

Persuasion Score

Lift

Page 13: Aapc deck intro to modeling draft 6.6.14

X%

Example: Targeting with the persuasion score in VA

Independents

Random Voters

High Persuasion Scores

316 thousandlikely voter targets

1,683persuaded voters

0.5%

1.2%

3.9%

316 thousandlikely voter targets

316 thousandlikely voter targets

3,738persuaded voters

12,193persuaded voters

• Modeling increased the efficiency of the persuasion program by a factor of 7

• The number of voters persuaded represents nearly a 25,000 vote swing

Targets Call Capacity % Impact Votes won

Page 14: Aapc deck intro to modeling draft 6.6.14

Case Studies

Page 15: Aapc deck intro to modeling draft 6.6.14

Wendy Davis: Texas 2014 Gubernatorial

Problems faced by Texas• Big state with sparse targets

• Requires balance of – Registration

– Persuasion

– GOTV

Page 16: Aapc deck intro to modeling draft 6.6.14

Terry McAuliffe - Virginia 2013

Gubernatorial

30 40 50 60 70McAuliffe 2−Way %

VA Expected McAuliffe Support

Page 17: Aapc deck intro to modeling draft 6.6.14

Fully Integrated Analytics

Program

Analytics Program

Support Models

Turnout Models

GOTV Model

Persuasion Model

Undecided Model

Media Optimization

Direct Mail EIP Models

Tracking Polls

Race/Ethnicity Models

Analytics Tech

Embedded Analytics

Staff

Campaign

Strategy

Field Program

TV Advertising

Direct Mail

Polling

Volunteer

Recruitment

Resource Allocation

McAuliffe

Win

Page 18: Aapc deck intro to modeling draft 6.6.14

Optimizing Field

ContactsOur modeling in VA in 2013 improved field program’s GOTV targeting in October GOTV by over 20% compared to the most recent election, helping volunteers reach more strong Democrats and fewer Republicans and undecideds.

-25%

-20%

-15%

-10%

-5%

0%

5%

10%

15%

20%

25%

Strong Democrat Lean Democrat Undecided / Republucan

Change in targets reached in late October as compared to 2009

Page 19: Aapc deck intro to modeling draft 6.6.14

Data Driven Decision Making

Where we put offices Where we sent canvassers

Page 20: Aapc deck intro to modeling draft 6.6.14

1. Build Model 2. Match Targets 3. Media Optimizer

Match list of targets with set-top box or online data + cost per program estimates

Optimizer produces list of efficient ad buys, with detail for specific programs/sites in key markets on specific days

Use models to identify the individuals we want to reach

with television ad buys

$ $ $ $ $

Reach 18% more targets or spend 18% less money

Optimizing Ad Buys – Maximizing Impact

Page 21: Aapc deck intro to modeling draft 6.6.14

Persuasion and GOTV targets allow specific targeting on social media and ads

Individual Level Targeting & Social Media/Online Ads

Page 22: Aapc deck intro to modeling draft 6.6.14

Changing Minds:

Persuasion Case Study:

In VA in 2013, our field persuasion program

alone succeeded in reaching an estimated 4x

persuasion effect compared to traditional

targeting.

The program persuaded about 12,500

additional voters to support Terry McAuliffe,

netting an estimated 25,000.

Since many of these voters would have likely

voted for the Republican candidate, the actual

effect on the vote margin was much larger—

which is especially significant given that the

entire winning margin for McAuliffe was

56,435.

7x

Page 23: Aapc deck intro to modeling draft 6.6.14

Scaling to State Campaigns: Virginia

2014In 2014, we helped 2 Democrats win special election races for state senate in Virginia by helping identify and mobilize their strongest targets with high accuracy.

Holding these two seats was the difference between Democrats maintaining or losing the majority.

SD-6: Lynwood Lewis SD-3: Jennifer Wexton

Our modeled turnout: 21.9% Actual turnout: 22.5%

Our modeled turnout: 20.2% Actual turnout: 20.4%

Page 24: Aapc deck intro to modeling draft 6.6.14

Scaling to State Campaigns: NC-12 in

2014In NC-12, Alma Adams faced a crowded primary election in an electorate that had not tuned in yet and were still largely undecided.

Partnering with EMILY’s List and Diane Feldman, and using a new analytical approach requiring smaller scale data collection efforts, we constructed a universe of voters receptive to Alma’s message. This universe allowed EMILY’s List to construct an effective mail program that was cost effective and targeted at the voters most open to Alma’s message.

Alma won the primary with over 40% of the vote, securing her the Democratic nomination.

1. Messaging2. Small scale test3. Create targeted universe4. Mail5. Win