Giving is Caring: Understanding Donation Behavior through Email
-
Upload
yelena-mejova -
Category
Technology
-
view
507 -
download
1
Transcript of Giving is Caring: Understanding Donation Behavior through Email
Giving is Caring Understanding Dona1on Behavior through Email
Yelena Mejova, Qatar Compu1ng Research Ins1tute Ingmar Weber, Qatar Compu1ng Research Ins1tute Venkata Rama Kiran Garimella, Aalto University Michael C. Dougal, University of California, Berkeley
circa Fall 2012 @
@ Computer Supported Coopera1ve Work and Social Compu1ng (CSCW14) February 19, 2014
Mo1va1on
hTp://www.opensecrets.org/pres12/
Impact real life!
1. Can we detect dona1ons in email? 2. Can we verify sociological theories on charitable giving?
Mo1va1on
What could affect dona1on behavior?
• Demographics – “individual capacity”: educa1on & income1 – belonging to a social group2
• Interest – seeing as more needy and deserving3
• Solicita1ons – influence of the email deluge4
• Social network – influence / homophily, poli1cal affilia1on5
• External influence – events triggering a new interest or awareness6
1. Shier, M., and Handy, F. Understanding online donor behavior: the role of donor characteris1cs, percep1ons of the internet, website and program, and influence from social networks. Interna'onal Journal of Nonprofit and Voluntary Sector Marke'ng (2012).
2. Tajfel, H., and Turner, J. An integra1ve theory of intergroup conflict. The social psychology of intergroup rela'ons 33 (1979), 47.
3. Bekkers, R. Who gives what and when? a scenario study of inten1ons to give 1me and money. Social Science Research 39, 3 (2010), 369–381.
4. Chan, M. The impact of email on collec1ve ac1on: a field applica1on of the side model. New Media & Society 12, 8 (2010), 1313–1330.
5. Bekkers, R., and Wiepking, P. A literature review of empirical studies of philanthropy. Nonprofit and Voluntary Sector Quarterly 40, 5 (2011), 924–973.
6. Olson, M. The logic of collec've ac'on: Public goods and the theory of groups, vol. 124. Harvard Univ Pr, 1965.
Data • Collec1ng “chari1es” (total: 480)
– Scraping Forbes and US News top chari1es lists – Top 100 US poli1cal campaign organiza1ons – Chari1es relevant to the prominent news stories in the 1me period (Wikipedia Current Events)
• Anonymized Yahoo! Mail – user agrees to research – email addresses replaced with hashes – fields: from, to, 1tle
• July 19 – September 19, 2012
hTp://www.forbes.com/lists/2011/14/200-‐largest-‐us-‐chari1es-‐11_rank.html hTp://www.usnews.com/usnews/biztech/chari1es/lists/intl_deve-‐lopment.htm hTp://www.fec.gov/data/CommiTeeSummary.do?format=html&elec1on_yr=2012 hTp://en.wikipedia.org/wiki/Portal:Current_events
• Emails from charity: matching from field • Emails thanking for dona1on: manually tuned regex (86% assessed
precision) – from: [email protected] subject: Thank you for your dona'on!
• Manually categorized chari1es which have at least 100 emails in dataset: – Medical, Humanitarian, Poli1cs, Environmental, Chris1an/Religious,
Military, Children, Public Broadcas1ng, Animals, Internet
Data
• Donors (≈100,000) – charity thanked them for a dona1on • Interested (≈100,000) – got email from a charity but did not
donate • General (≈10,000) – a random sample of the rest ≈ 1 billion emails total
• Is the email treated as bulk or spam? <10% labeled as spam: cancer.org, lls.org, redcross.org >50% labeled as bulk: stjude, dscc.org >50% labeled as spam: wikimedia.org
• In the analysis, we pay aTen1on only to emails which reach the inbox
Data
Demographics
Demographics • Self-‐reported from user profiles
– age, gender, zip code • US Census to es1mate
– % of Bachelor degrees, median household income
age gender (male=1)
% bachelors median HHI
In agreement with the US Presiden1al Elec1on exit polls:
– Younger – Female – Less affluent … voters favor Obama
age gender (male=1)
% bachelors median HHI
Demographics
hTp://elec1ons.msnbc.msn.com/ns/poli1cs/2012/all/president/#exitPoll
Interest
Interest • Classify email 1tles into topical categories using manually-‐compiled
keyword regexes (avg precision: 86.2%):
incoming outgoing
for par1cular topic for par1cular topic PBS WGBH
Solicita1on
Solicita1on • Does increase in solicita1on prompt more dona1ons?
1. Compute number of dona1ons per day for each charity 2. Divide into three terciles: low, medium, high 3. Compute Cohen’s kappa with incoming mail from charity
43% of organiza1ons have Cohen’s kappa > 0.3 That is, there is a moderate to high rela1onship between solicita1ons and dona1on
Solicita1on • Can we detect this on a personal scale?
– Compute probability that, given the user will donate to an organiza1on, that he or she donates within some number of days of receiving a solicita1on
– Compare to a uniform baseline
On average, cumula1ve probability of a dona1on a{er a solicita1on is higher by 8%
na1onalmssociety.org Medical 48.99 wycliffe.org Chris1an 37.98 lls.org Medical 33.82 opera1onsmile.org Medical 29.41 intervarsity.org Chris1an 18.78 feedthechildren.org Humanitarian 18.26 dscc.org Poli1cs 15.39 worldvision.org Humanitarian 13.88 tbn.org Chris1an 13.82 wikimedia.org Internet 12.99 marchofdimes.com Medical 11.82 mdausa.org Medical 11.45 ronpaul2012.com Poli1cal 8.92 greenpeace.org Environmental 7.89 irusa.org Humanitarian 7.5
increase in dona1on probability
Social network
Social network • Is there a rela1onship between your dona1ons and those of your friends?
1. For each donor, find other Yahoo users they exchanged emails with 2. Normalize number of friends by sampling with replacement to get 100
friends 3. Within those friends we can find other donors and interested users 4. Measure strength of rela1onship by minimum emails sent or received
Spike in 7th bucket (closest friends) is greater than buckets 1-‐6 at p < 0.01
Donors to poli1cal causes have almost no interac1on with others who donated to the opposite side
each bucket contains the same number of users
External Influence
External Influence
0 0.2 0.4 0.6 0.8 1 1.2
0 0.2 0.4 0.6 0.8 1
1.2
7/19/1
7/21/1
7/23/1
7/25/1
7/27/1
7/29/1
7/31/1
8/2/12
8/4/12
8/6/12
8/8/12
8/10/1
8/12/1
8/14/1
8/16/1
8/18/1
8/20/1
8/22/1
8/24/1
8/26/1
8/28/1
8/30/1
9/1/12
9/3/12
9/5/12
9/7/12
9/9/12
9/11/1
9/13/1
9/15/1
9/17/1
9/19/1
Solicita
(on Vo
lume
Dona
(on Vo
lume
Dona1ons Solicita1ons
0
0.2
0.4
0.6
0.8
1
1.2
0
0.2
0.4
0.6
0.8
1
1.2
7/19/1
7/21/1
7/23/1
7/25/1
7/27/1
7/29/1
7/31/1
8/2/12
8/4/12
8/6/12
8/8/12
8/10/1
8/12/1
8/14/1
8/16/1
8/18/1
8/20/1
8/22/1
8/24/1
8/26/1
8/28/1
8/30/1
9/1/12
9/3/12
9/5/12
9/7/12
9/9/12
9/11/1
9/13/1
9/15/1
9/17/1
9/19/1
Solicita
(on Vo
lume
Dona
(on Vo
lume
Dona1ons Solicita1ons
barackobama.com
miTromney.com
“Photo going around on Facebook”
Paul Ryan announced as VP candidate
T-‐shirt promo “I built this”
Republican Na1onal Conven1on Democra1c Na1onal Conven1on
Pu|ng it all together • Sample data to mi1gate influence of chari1es with more
representa1on • Logis1c regression predic1ng whether user donates to a charity
all coefficients are significant at p < 0.01 except I4 and S4 For each friend who donates to a charity, the likelihood for
the user to also donate to that charity goes up by 24%
Conclusions
• Want to run a campaign? – Email campaigns can be effec1ve – solicita1on works! – Tailor to your campaign to your audience – Leverage social network – Don’t get stuck in spam/bulk folders
• What’s next? – Social influence or homophily? – How to tailor solicita1ons?
The Language that Gets People to Give: Phrases that Predict Success on Kickstarter Tanushree Mitra | Eric Gilbert
[images by Wikimedia]