PowerPoint Presentation
1
Professor Andy TatemPopulation statistics using new forms of
data
This presentation provides an overview of geospatial methods,
drawing on big data and other sources, for complimenting
traditional sources of demographic data in building evidence bases
for monitoring progress towards the SDGs at subnational scales, and
designing strategies. The examples come from work being undertaken
by the WorldPop project (www.worldpop.org) and the Flowminder
Foundation (www.flowminder.org), which both focus on improving the
spatial demographic evidence bases for low and middle income
countries, partnering with national statistical agencies, UN
organizations, mobile network operators and data providers to add
value to traditional datasources and integrate them with novel
datasources to fill gaps. 1
Key partners and donors
WorldPop: Research program in Dept Geography & Environment
focused on methods for improving the demographic evidence base in
low/middle income countries
Flowminder: Non-profit foundation working with data providers
and international/government agencies to operationalize and scale
research in support of vulnerable populations and sustainable
development in low/middle income countries
50+ staff; focus on data science and integration in fields of
spatial demography and mobility to support disease, disaster and
development applications
DFID, DLR, ORNL, UNICEF, ICF2
Population statistics: what do we need?Mapping the
denominator
Governments are reliant on accurate and up-to-date data
onpopulation numbers anddistributions for planning
-What is the catchment population of my health facility?-How
many people will be affected by development x?Targeting vulnerable
populations
To improve the health of the poorest and most vulnerable and
target interventions to lift them out of poverty, we need to know
where they are
-How much vaccine is needed for this ward?-Do the poorest have
access to financial support here?
Designing policies around mobility
Data on migration, seasonal movements, commuting are valuable
For effective policy andservice delivery strategy design
-How many people commute from x to y?-Where should services for
mobile populations be best placed?
Emergency response
Effective response requires rapid and ongoing assessments of
numbers of vulnerable peopleaffected and future risks
-How many people were impacted by the flood?-Which people were
displaced, and where to?
OECD!3
Data
Data
The challenge20102020203020152025CensusCensusCensus??DataCensus
data are valuable, but expensive and collected once a
decadeIncreasing need for more timely and detailed dataRegistry,
administration data can help fill gapsNot available in many
low/middle-income settingsChallenge of tracking progress towards
development goalsIn some countries even the most basic census data
are lacking????
Years since last census
Here the challenges of relying solely on census data are laid
out. How do we measure and monitor in the inter-censal years? Plus
it can often take a year+ for census data to appear, plus many
countries have not had reliable censuses for many years how do we
define a baseline and measure change subnationally and regularly
from this baseline?4
What do we have to help us?
Data
Data
20102020203020152025CensusCensusCensusData
SurveySurveySurveySurveySurveySurveySatelliteMobile phone
Geolocated household surveysSatellite and GIS dataMobile phone
data
New geospatial datasources can help compliment existing
traditional sources to fill gaps and keep a subnational focus.
Three types of data will be introduced in the following slides,
each of which have the valuable feature of being available more
regularly than census data, in inter-censal periods, and each of
which is already finding operational use in low and middle income
settings. While these datasources are valuable in providing
regularly measurements and subnational detail, they do all contain
biases and gaps, hence the integration with each other and with
more traditional datasources is required to draw on the strengths
of each and overcome some of these biases.5
Geolocated Surveys
53.1% of women literate
Low
HighProportion of women who are literate
Increasingly survey clusters in national surveys are using
Global Positioning Systems (GPS) to record the precise locations
where household surveys were undertaken, and the coordinates are
made available e.g. Demographic and Health Surveys (DHS), Living
Standards Measurement Study (LSMS). This enables moving from
national or broad subnational regions in SDG target metrics (as in
the Nigeria national and provincial female literacy example) to a
much more fine-grained picture that is provided by the GPS-located
survey cluster data, highlighting heterogeneities that are masked
in aggregate statistics. The close up for Nigeria shows how
integrating with satellite maps of urban areas immediately
highlights how literacy is higher in urban areas and decreases with
greater remoteness. This links to next slide on satellite/GIS
layers that show how such factors can be measured in detail. The
global image shows all the GPS-located clusters that are publicly
available through the DHS program, highlighting the potential to
measure SDG target metrics at fine spatial scales for most of the
low/middle income world (when integrated with other datasets in
appropriate statistical models to account for small sample sizes,
uncertainties etc see later slides).6
Satellite and GIS data
Massive increases in geospatial data collection, processing and
availability have occurred over the past decade, with volumes,
coverage, variety and quality all on the rise. Examples here of
settlements mapped from 0.5m spatial resolution satellite imagery
using computer-automated methods, open streetmap buildings
constructed through volunteers mapping from detailed satellite
imagery, refugee camps mapped digitally with estimated population
sizes integrated, topography and landuse data derived from
satellite image processing, the density of geolocated tweets in
Indonesia providing an indication of population densities, and
health facilities/roads mapped in Kenya. These all provide
relatively complete and consistent coverage across countries, with
often regular updates they also cover factors associated with
poverty, demographics etc, and so offer opportunities for
integration with geolocated survey data in models to produce
predictive maps of key SDG target metrics.7
Mobile phone call detail records
User makes a call from location X
User travels to Y and makes a call
XYCall routed through nearest towerNetwork operator records time
and tower of call for billing
Mobility: Changing densities, flows, seasonal/permanent
migration
Social networks: Number of contacts, calling patterns
Consumption: Credit purchase frequencies, top-up
amountsProtecting confidentiality -Aggregate summaries-Regulator
approval-Raw data never leaves operator
Finally, mobile phone ownership has rocketted in the past
decade, with the vast majority of the population of even the lowest
income countries owning and regularly using cellphones. The last 5
years has seen a rise in the use of call detail records (CDRs).
These are records, collected anonymously, of the time, type of
communication sent/received, and cell tower that the communication
was routed through. The example shows how a user making a call in
one location has this routed through the nearest tower and the
event recorded. The user later moves to a different location and
makes another call this is routed through a different tower and
this communication event is again recorded, showing that the user
moved from the vicinity of one tower to the vicinity of another
evidence of movement. Across the millions of customers, and
billions of communications, detailed aggregate patterns of changing
population densities, mobility patterns, social networks and
spending on credit can be observed without being able to identify
individuals or compomise confidentiality. CDRs provide valuable
complimentary data on population mobility, social networks and
consumption in near real time.8
Data integration
Massive increases in geospatial data collection, processing and
availability have occurred over the past decade, with volumes,
coverage, variety and quality all on the rise. Examples here of
settlements mapped from 0.5m spatial resolution satellite imagery
using computer-automated methods, open streetmap buildings
constructed through volunteers mapping from detailed satellite
imagery, refugee camps mapped digitally with estimated population
sizes integrated, topography and landuse data derived from
satellite image processing, the density of geolocated tweets in
Indonesia providing an indication of population densities, and
health facilities/roads mapped in Kenya. These all provide
relatively complete and consistent coverage across countries, with
often regular updates they also cover factors associated with
poverty, demographics etc, and so offer opportunities for
integration with geolocated survey data in models to produce
predictive maps of key SDG target metrics.9
Thinking in grids
Grids: consistent and comparable format and a framework for
integrating differing data types
100 x 100m gridded estimates of population density
Administrative unit population totalsGrids: flexibility in
summarization to any administrative unit level
Comprehensive Emergency Obstetric and Neonatal Care (CEmONC)
Facilities overlaid on grid of women of childbearing agePercentage
of women of childbearing age per woreda within 50km of a
CEmONCGrids: flexibility in analysis and data integration
10
Census data disaggregation
Aggregate census counts100x100m gridded counts
Census counts generally aggregated at coarse, irregular
administrative unit level, making integration and comparisons with
other data challengingWomen of childbearing age per 1x1km
2015Integration with satellite/GIS data related to human population
distribution patterns to disaggregate counts to regular grids
Mapping populations per grid square provides a detailed and
consistent dataset for integration with other datasources e.g.
population-environment interactions, healthcare access. Here, we
show how aggregated census counts can be disaggregated to a grid
square representation through exploiting the relationships between
population densities and the range of satellite and GIS datasets
presented earlier. The outputs are 100x100m estimates of total
numbers of residents and their breakdowns by age and sex, as well
as full metadata describing methods, input datasets and accuracy
assessments, and are freely available from www.worldpop.org, while
the computer code used to produce the outputs is freely available.
As the approach is disaggregating counts from census data linked to
administrative boundaries, the importance of using reliable census
numbers enumerated at a high administrative unit level remain as a
key determinant of outputs mapping accuracies. The output datasets
have been used in multiple international reports by UN agencies and
others, as well as substantial government usage within countries.
This is being scaled globally to produce freely available gridded
age-sex structured population data.11
www.worldpop.org
12
But...Years since last census
Population estimation in the absence of national census
Mapping buildings and settlements from satellite imagery
150 MicrocensusTraining Clustersand Validation Sets in Kano
State
Population estimates per 90x90m grid square
However, the census data disaggregation approach is not ideal
where there is a lack of reliable census data. In some countries,
the last census was a decade or more ago with substantial changes,
so the disaggregation of census counts will still not provide an
accurate evidence base for contemporary population distributions.
Here, very high resolution satellite imagery (50x50cm grid squares)
is processed to extract building and settlements. In urban areas,
different types of neighbourhoods that correspond to different
population densities can be automatically recognised by computer
algorithms and mapped too. Through integration of these
building/settlement extractions with small area sample
microcensuses, the prediction of population numbers across entire
countries can be undertaken with quantification of uncertainty in
outputs. Work done in collaboration with the Gates Foundation and
Oak Ridge National Labs, is now being used to direct vaccination
efforts in Nigeria and similar work is being started in Afghanistan
with the government and UNFPA.14
People dont stay still.
15
Modelling changing populations through integration of census and
mobile network data
The previous slides present approaches that still just provide a
static picture of population distributions, and can be laborious to
update. Cellphone data offers solutions here the integration of
mobile data enables dynamics to be measured. Through calibration
against census data to translate densities of SIM card
communications to population totals, mobile data can be used to
produce population distribution maps and measure changes
dynamically across the course of days, weeks, months or years. This
slide shows a short video that shows the results of doing this for
France - See this case study for more details:
http://flowminder.herokuapp.com/case-studies/dynamic-population-mapping-using-mobile-phone-data16
Pop density change per square km
Namibia Pop: 2.3 millMTC active subscriptions: 2.1 mill
How is this useful?
The France example provided a good validation test, but within
WorldPop/Flowminder, we are more focussed on low/middle income
countries where data gaps and limitations are more likely to exist.
Here the same approach was adapted to Namibia in collaboration with
the National Vector Borne Disease Control Program, who were
particularly interested in understanding how health facility
catchment populations change for estimating fluctuations in
resource demands. The slides show pop density change per square
kilometer ie huge changes at many times of year are evident, with
substantial seasonal variability. WorldPop/Flowminder are working
on extending these methods to many other low+middle income
countries in collaboration with mobile operators, governments and
international agencies.17
Dynamic facility catchment populations
18
Measuring migration
R2 = 0.92Annual migration flow from census dataAnnual migration
flow from cellphone data
The value of mobile phone data is further shown in measuring
migration. Using Namibia as an example again here, the population
fluxes measured using the change of residence census question was
compared over the same period as the national census to mobile
phone user changes of residences shows strong correlations. This
highlights strong potential for ongoing migration statistics
production using mobile data.19
Population dynamics across scales
Responding to outbreaks, disasters etc and building resilient
systems to mitigate against their impacts requires understanding
and measuring the dynamics of populations and how they respond to
events. Mobile data can provide unprecedented insights into this
across a range of spatial and temporal scales. The maps show
temporal changes in population densities across the course of a
week (for Accra and surrounding area in Ghana), for Kenyan regions
across seasons, and for Chinese New year.20
Resilience, Risk reductionCrisis response
Here, some recent examples of mobile data usage for responding
to outbreaks (Ebola W Africa, Cholera Haiti), and natural disasters
(Nepal quake displacements). As seen in the previous slides, mobile
data can be used to measure population connectivity through
movements and communications. Many examples in the literature have
shown the value of incorporating human mobility data from mobile
data in predicting disease spread, and the West Africa and Haiti
examples shown are just two of these. In Nepal, ongoing data feeds
enabled monitoring and rapid update of population displacements
after the recent earthquakes, with reports delivered to agencies to
guide response efforts. Ongoing data feeds are being set up in
these regions and many others to ensure that resilient response
systems are in place and that rapid responses can be provided just
like was done in Nepal.21
-Greater mobility = lower poverty-Wider social network = lower
poverty-Large, regular credit top up = lower poverty
-Increasing distance from major roads = increasing
poverty-Increasing urbanicity = decreasing poverty
High
LowPoverty rate
Data integration for mapping population characteristics
Population characteristics measured in household surveys can be
strongly related to features we can measure everywhere
22
Improving mapping of socioeconomic indicatorsGPS-located survey
cluster data
Cellphone and satellite data are collected 24/7 = Potential for
ongoing monitoring
Here us an example for Bangladesh it shows the types of data we
have available through GPS-located surveys, satellite/GIS and
cellphone-derived indicators for exploring relationships with
cluster level socieconomic metrics and predicting in unsampled
locations.23
Maybe these are crazy ivory-tower academic ideas that will never
find use?
24
Datasets, methods already in use
Children under 5yrsSyria population distribution Malawi live
births
Malaria prevalence 2000-15
Are all of these new datasets and methods just crazy academic
ideas? No, theyre already being used to produce health metrics,
situation reports, provide development indicators. A small subset
of these examples are provided here, with census disaggregation
gridded population datasets used to provide the numbers presented
in the World Malaria report, humanitarian needs assessments and
World Bank reports. Maps of births and pregnancies were used to
provide statistics in recent UNFPA midwifery reports, and mapping
from GPS-located survey data was used to measure and map changing
malaria prevalence in a recent UNICEF-WHO report.25
Datasets, methods already in use
Hurricane responseEarthquake-induced displacements
These data and methods are also already being used to guide
vaccination planning, measure financial access, support disaster
response. The bottom-up population mapping approaches have been
used to underpin polio vaccination planning in northern Nigeria,
while the detailed poverty maps have formed part of the Gates
Foundations efforts in measuring and mapping access to financial
services for the poor. Finally, the Nepal earthquake population
displacement analyses fed into reports distributed to all aid and
response agencies.26
Datasets, methods already in use
Population living under $2 a daywww.fspmaps.com
Children under 5yrs
National population estimation for election planning
These data and methods are also already being used to guide
vaccination planning, measure financial access, support disaster
response. The bottom-up population mapping approaches have been
used to underpin polio vaccination planning in northern Nigeria,
while the detailed poverty maps have formed part of the Gates
Foundations efforts in measuring and mapping access to financial
services for the poor. Finally, the Nepal earthquake population
displacement analyses fed into reports distributed to all aid and
response agencies.27
Building expertise
-Data science skills and capacity are increasing everywhere
-Open data and software are accelerating uptake
-Making full use of existing traditional data and complimenting
these with new dataset integration is more cost-effective than new
data collection
-Local ownership and analysis are key to sustainable
implementation
Key messagesIncreasing pressure to produce statistics at finer
spatial scales and more regularlyStatisticians and traditional
datasources are more valuable than everBut to produce estimates for
different geographical scales and time periods, the integration of
multiple types of data to compliment traditional sources is often
requiredNovel datasets (e.g. phones, satellite) are prone to
biases, but each has advantages over census data in terms of the
frequency of measurement and spatial precisionMethods to account
for biases, reporting uncertainties and providing clear
metadata/documentation to inform users are all importantGreat
potential in complimenting traditional sources to build strong
demographic databases
29
Further information
www.worldpop.orgwww.flowminder.org
@WorldPopProject
@FlowminderE-mail: [email protected]
30
