Thailand

Mobile Operator Call Records:

Potential and Pitfalls

Linus Bengtsson MD, Ph.D. linus.bengtsson@flowminder.org

Content

1. Flowminder

2. Mobile Operator Data: What is it?

3. Application areas

1. Challenges

Non-profit foundation working with data providers and international/government agencies to operationalize

and scale applications in support of vulnerable populations and sustainable development.

Key partners and donors:

Pioneered Anonymized Mobile Network Data for Infectious Disease: (2008 Zanzibar, Kenya, 2012 Haiti, 2013- Namibia, Indonesia) and Crisis Response (Haiti 2010 earthquake and cholera. Nepal 2015)

Ouest

Sud

Centre

Artibonite

Nord-Est

Grande-Anse

Nord

Sud-Est

Nord-Ouest

Nippes

Average daily numbers of sims that moved out from the communal sections surroundingSaint-Marc, Oct 15 to Oct 23, 9:00 am, 2010.

Outbreak area

250

500

100

10

All Methods are Open and Published in Peer Reviewed Journals for

Validation and Transparency

Mobile Operator Data: What Is It?

User makes a call

from location X

User travels to

Y and makes a

call

X

Y

Call routed through

nearest tower Network operator records time

and tower of call for billing

8

FLOWMINDER.ORG

Call Data Records (CDR) is billing information

(A) number – caller: anonymized

Cell_ID: location

Type: call, SMS, data, etc

Date & time

(B) number – receiving party: anonymized

Data volume

Mobile network data: Call detail records (CDR)

Operators Call Detail Records’ (CDRs)

including low-resolution location data

(nearest tower location) anonymized on

separate server hosted by operator.

We conduct analyses under operator supervision, anonymous raw data always behind operator firewall

GSMA data integrity guidelines: Data never leaves mobile operator’s system

Aggregated mobility estimates are exported and made open access - can be used with other mobility estimates, epidemiological data

Mobile operator firewall

Preserving User Privacy

Application Areas:

1. Population Data and Migration

Nigeria 2015

Refs: See www.worldpop.org

Dynamic Facility Catchment Populations

Ref: Zu Erbach et. al. 2016. Under review. Deville et al. PNAS. 2014

High-Resolution Maps of Population Characteristics

2010-11 Nigeria LSMS:

Consumption-based metric

FLOWMINDER.ORG

Total Top-ups

High

Low

% Nocturnal Calls High

Low

Radius of Gyration

High

Low

Enhanced Vegetation Index Lush Vegetation

Bare Ground

Temperature

High

Low

Distance to Roads

Near Far

Precipitation

High

Low

Nightlight Intensity

High

Low

Ref: Steele et. al. Under review

Combining Mobile and Remote Sensing Data for Poverty Mapping

Ref: Steele et. al. Under review

Ref: Lu et al. Global Environmental Change, 2016

Migration Patterns

Application Areas:

2. Disaster response and Preparedness

Disasters Cause Large-Scale Population Movements

Post-earthquake movements from Port-au-Prince Mobile phone movements from cholera outbreak area

Haiti Earthquake 2010

Reference: Bengtsson et al. PLoS Medicine 2011

Predictive Modeling

20

Wilson et al. PLoS Current 2016.

Nepal Earthquake 2015: Setup & First Insights Within 14 Days

2. Kathmandu Valley

Kathmandu Valley is here defined as the districts

Kathmandu, Bhaktapur and Lalitpur. Kathmandu

Valley is one of the most densely populated areas

in Nepal and home to ca 2.8 m people [1].

Key f indings:

➔ An estimated 390,000 people more than

normal had left the Kathmandu valley -

comparing May 1 with the day before the

earthquake April 24 (ratio to the

population: 14%).

➔ An estimated 247,000 persons less than

normal had come into the area during the

same period(ratio to the population: 8.8%)

➔ People leaving Kathmandu Valley went to

a large number of areas, notably the

populous areas in the south and the

Central and West Development Regions.

Contacts:

linus.bengtsson@flowminder.org +41 78 964 88 28

erik.wetter@flowminder.org +46 70 893 88 37

andy.tatem@flowminder.org +44 7703 392 192

+390,000(246,000~540,000)

Nepal Populat i on Est im at es

as of M ay 1, 2015

4

2.8m

Above normal f lows f rom Kathmandu Valley t o other dist r ict s

Above normal flows from Kathmandu Valley to other

districts (comparing pre-earthquake April 24 and May 1).

Produced 8 May 2015

[1] www.worldpop.org

Populat ion out f low

(above normal)

Pre-earthquake populat ion

Flowminder.org is a non-profit organization registered in Stockholm, Sweden. Ncell is a mobile operator in Nepal

and part of the TeliaSonera group. Analyses are based on de-identified mobile network data and conducted in

accordance with mobile industry (GSMA) Guidelines on the protection of privacy in the use of mobile phone data for

responding to the Ebola outbreak, published October 2014.

Populat ion inf low

(above normal)

- 247,000( - 155,000~- 339,000)

Ref: Bengtsson et al. Sci Rep. 2015; Wesolowski et al. PLoS Currents 2014; Wilson et al. PLoS Current 2016.

“During the Nepal earthquake, the Office of the UN Resident Coordinator established an assessment unit providing evidence-based information and situation awareness of key humanitarian priorities to a number of different humanitarian actors.

Flowminder mobile analysis really helped OCHA and other agencies in doing better coordination and operational planning.”

Luca Dell’Oro, UN OCHA

“When the first Flowminder-Ncell report came out after the Nepal earthquake we used it right away in our national assessment of food security. Displaced people are often the most food insecure. Getting national and district level numbers on displaced populations was thus an important component in our assessment of where to focus support after the earthquake.”

Kurt Burgar, UN WFP

Operational Impact

Key Issues and Challenges

• Agree on indicators

• Quantifying uncertainty. Reducing bias. Who is left out?

• Importance of combing multiple data sources

Application Areas:

3. Infectious Disease Modelling

Ouest

Sud

Centre

Artibonite

Nord-Est

Grande-Anse

Nord

Sud-Est

Nord-Ouest

Nippes

Average daily numbers of sims that moved out from the communal sections surroundingSaint-Marc, Oct 15 to Oct 23, 9:00 am, 2010.

Outbreak area

250

500

100

10

Population Dynamics: Disease Elimination

How does mobility impact

transmission?

Where should resources

be focussed?

Ouest

Sud

Centre

Artibonite

Nord-Est

Grande-Anse

Nord

Sud-Est

Nord-Ouest

Nippes

Average daily numbers of sims that moved out from the communal sections surroundingSaint-Marc, Oct 15 to Oct 23, 9:00 am, 2010.

Outbreak area

250

500

100

10

Key Issues and Challenges

• Infectious diseases are fundamentally different and require different

types of modeling

• Limited number of robust studies

• Operator can be very important but sometimes not the biggest

limitation

The Star Trek Fallacy

1. Data is the tool, not the solution – issue-driven vs. data-driven problem solving.

2. Remote sensing data and analytics can augment but not replace field data (”ground truth”), eg. surveys.

3. Very few studies of bias

4. Mobile network data is not a magic potion: • Mobile data is heterogenous – market/operators.

• Representativeness – what does a SIM card represent?

• Fundamental characteristics (subscribers) constantly changing.

• Realtime mobile data without validation = realtime mistakes

Summary and Policy Implications

• Evaluating use of apps is different from understanding the

population as a whole

• Importance of aggregating and combing with other data sources

Need for preparation. Disasters not suitable for experimentation

• Different data needed for different applications

• Need for continuous research and development

• Importance of legal frameworks

www.worldpop.org

www.flowminder.org

@WorldPopProject

@Flowminder