Michelangelo D’Agostino: Physics and presidential politics
Background
- harvard undergrad, berkeley PhD, worked on Icecube (neutrino detection)
Approach
- neural networks and support vector machines to score liklihood of particular outcomes (neutrinos or voter stats)
- Data collection — mass profile data from Facebook / Twitter.
Campaign organization
Tech department
- voter calling tool from web interface
- dashboard – online field office to match up activists
- database maintenance
digital team
- email fundraising
- social media pages for the president
- digital advertising
Analytics
- cave “learned from physics how to be in a closed windowless room with 40 people”
- modeling / polling / experiments
Building models
- using historical data make models for predictions.
- how likely is someone to donate / change vote in response to door knock / volunteer?
- use time and money to target persuadable people and infrequent voters.
- all individuals in battle ground states get support, turnout, persuasion, and getout-the-vote.
Information
- states provide pubically for all registered voters how much they vote, age, etc (not who they voted for)
- political parties also collect similar data.
- Match this data to 5-10K people you’ve called on the phone. Use to train models.
- have statistical models of liklihood of effect, rather than gut feeling and attack ads.
- cable TV collects data on what channels people are watching. correlate with data files. Intermediary controls so that no individual has all the data on a given individual (what person X watches and how X votes).
- can advertize to persuadable voters on cheap niche channels rather than big public channels — make money go further.
Fund raising
- 504 million of over 1 billion raised came from digital: data driven analysis of email campaign.
- randomized control experiments. Starts as a dozen different version sent out to small random control group. Analyze returns (donations/clicks/unsubscription). Pick winning email send to list.
- difference in returns millions of dollars per email, 4-10 fold difference in return rate.
- How much to ask for? different emails for new vs. repeat contacts?
- people respond better when you spam them then if you show restraint.
- optimize form appearance: fast loading helps. ask people to enter 1 thing at a time rather than a whole form.
Monte Carlo simulations
- How big should a test group be? (optimization of selecting the best and not burning through too much of the list)
social media
- 50% of 18-29 yr olds are unreachable on phone: no number or refuse to pick up. Only getting worse.
- 85% of them are on Facebook.
- 24 million Twitter followers
- Targeted sharing on facebook. Authorize campaign’s app: gives access to email/name/etc + photos + friends birthdays, likes and locations.
- write to you to reach out to your friends in battleground states with a personal message to ask them to vote, (in particular if model predicts they are unlikely to get out to vote).
- high response rate on people sending messages (10% clickback). 1 million users authorized app, 100 million people get access.
- Twitter
- sent targeted direct messages from the president’s account / first lady’s account, asking people to contact their followers.
- people really liked getting messages directly from VP/pres etc.
- use peoples tweet words + SVM to ID if individual is conservative or progressive.
- conservatives tweeted more about Obama than progressives. Both tweeted about Romney at similar rates.
- don’t use known info about contacts to assign political connections.
Data Science
Questions
- predictions – well predicted but well sealed, don’t let predictions alter response.
This entry was posted in
Seminars. Bookmark the
permalink.