Using Facebook and LinkedIn Data to

Study International Mobility

Carolina Coimbra Vieira1 , Masoomali Fatehkia2 ,
Kiran Garimella3 , Ingmar Weber2 and Emilio Zagheni1
1
Max Planck Institute for Demographic Research, Germany
2
Qatar Computing Research Institute,
Hamad Bin Khalifa University, Qatar
3
Massachusetts Institute of Technology, Cambridge, USA

For the final version of this work, check the book: ”Data Science for Migra-
tion and Mobility”, Edited by Albert Ali Salah, Emre Eren Korkmaz, and Tuba
Bircan. (https://global.oup.com/academic/product/data-science-for-
migration-and-mobility-9780197267103?view=Grid&lang=en&cc=lb#)

Abstract
Migration and mobility present major societal challenges, while also
representing key opportunities. New data sources from social media, such
as Facebook and LinkedIn, offer valuable insights that can help us measure
and understand patterns of long-term migration and short-term mobility,
as well as immigrant integration along various dimensions. In this chapter,
we describe how Facebook data for advertisers can be used to quantify
migration, and to study levels of assimilation of migrants based on the
interests they express online. We use the term assimilation in a technical
or neutral sense to express a measure of the distance between groups. In
LinkedIn, aggregate-level information can be harvested to discover where
migrants have studied, and where they live and work. In addition, all
social media advertising data sources provide information about the lan-
guages that people speak (e.g., Polish speakers on Snapchat in London).
Across all case studies, we discuss the limitations and risks of using these
data, including privacy and legal concerns. Given the difficulty of col-
lecting timely migration data, particularly for traits related to cultural
assimilation, the methods we develop and the results we provide in this
study open up new lines of research that computational social scientists
are well-positioned to address.
Keywords: migration, advertising data, social media, Facebook, LinkedIn

1
1 Introduction to Advertising Data
Online social networks are well known for promoting social interaction between
people by connecting them even when they are physically separated, by a few
miles or across continents. To deliver their free service to their users, most
online social networks rely on targeted advertisements as their business model.
Facebook, for example, hosts over a quarter of the world’s population, and its
advertising platform alone, Facebook Ads, is responsible for more than 98%
of total revenues. Similarly, Twitter, Instagram, and TikTok are social media
platforms whose business model is based on advertising. Even social networks
that offer a freemium model such as LinkedIn, where many features are available
for free, but some require a paid membership, rely on targeted advertisements
for additional revenues.
The key appeal to advertisers to use these platforms to promote their prod-
ucts lies in the targeting capabilities offered by the social networks. Highly
targeted social media ads can potentially deliver the right message to the right
consumer, offering a good return on investment. Social networks are able to
offer these targeting capabilities due to the rich user data they collect, which
includes detailed demographic information, such as age, gender, home location,
income level, or education level, but also topical interests, and certain behaviors.
Some of these attributes are explicitly self-declared by the user, such as their
age and gender, others are derived from meta-information on how they access
the social network platform, such as their likely home location or their device
type, and some are inferred through machine learning models, using their likes,
social interactions, or status updates as input.
In other words, we can see social media advertising as the process of matching
social network users by their profile data available on the social network to target
groups specified by the advertiser to deliver advertisements. Thanks to the
personalized targeting and scalability of online advertising, advertisers benefit
from huge increases in conversions and sales with a lower cost of acquisition.
Social media advertising has thus become a popular source of advertising, with
a projected $43 billion spent on it in 2020.
The richness of the aggregated data provided by online social media adver-
tising platforms has also been explored by the academic community using the
same tools the platforms provide to advertisers. These tools allow advertisers
to obtain the estimated number of users on a platform that matches given input
criteria based on demographic attributes before the ad is launched. Using these
tools, researchers obtain data to study demographics across several research ar-
eas, including mobility and migrant assimilation. Facebook and LinkedIn host
two of the main advertising platforms explored by researchers to study migra-
tion, especially due to their features which provide the easiest way to target
migrants.1
Facebook is the most popular online social network and its purpose is very
generic since its users utilize the platform for different purposes of their daily
1 Though Twitter also provides similar features, to the best of our knowledge, there has

not been any research done using Twitter data.

2
lives, such as contacting friends, sharing emotions, and reading the news. Sim-
ilarly, users on LinkedIn can use the social network to connect and build a pro-
fessional network. By using Facebook’s advertising manager, for example, it is
possible to target the users by their current location as well as their hometown,
thus allowing us to identify potential migrants. Similarly, LinkedIn provides
the previous and the current place of work or study which can be used to study
migration. The users’ information about the native language and the spoken
languages can also be used as a proxy of nationality or home country and to
study migration by using this network.
In terms of research involving the Facebook Advertising Platform, for ex-
ample, it is possible to study in high-level the process of migration as well as
the levels of assimilation of migrants, as seen through the interests they express
online. For instance, in Section 2, we show how the Facebook Ads platform can
be used to collect the audience of women aged 18+ living in Colombia who used
to live in Venezuela and who primarily use a 4G connection to access the social
network site. Since it is possible to collect information regarding the institution
of study or work by using LinkedIn data, the LinkedIn Advertising Platform can
be used to study professional migration. The LinkedIn platform is used to har-
vest information about where migrants have studied, where they live and work
as well as their professional connections. As an example, in Section 3 we show
how the LinkedIn Ads platform can be used to estimate the number of female
LinkedIn members who hold a Ph.D., studied at the University of Cambridge,
and are now living in Germany.
The data is publicly accessible (Facebook/LinkedIn account needed) for ev-
ery user without any special access requirements. The advertising platform
webpage provides the most intuitive and user-friendly interaction, the adver-
tiser can select the desired demographic attributes while the platform provides
the expected number of users which match the criteria. However, this approach
is relatively slow and does not scale. To automate the data collection, we can
use dedicated application programming interfaces (APIs) provided by the plat-
forms. This will require a certain level of programming and web-scraping skills.
These APIs are not publicly advertised and are used for serving data to the
website frontend and can be easily identified using network monitoring in any
modern web browser.
Despite the advantage of having access to the data for free, both platforms
provide an incredible ability to segment the advertising, allowing target specific
audiences (without actually placing an ad). When compared with other data
sources, such as surveys and census data, the data collection through online
social media platforms requires less time, effort, and cost to deliver the data
requested. The amount of data available, scalability, and speed of updates are
also advantages of using these low-cost real-time platforms, especially because
of the popularity in terms of the number of users providing their information
on social networks and the existence of APIs. Although there are a lot of
advantages, there are ethical issues, and the platforms are structured as a black
box which imposes limits on the use of such data for research. We discuss such
limits in Section 4 after discussing examples from two major platforms providing

3
free advertising data: Facebook and LinkedIn.

2 Facebook example(s) - what has been done

Figure 1: Screenshot of Facebook’s advertising platform for an ad targeting

women aged 18+ living in Colombia who used to live in Venezuela and who
primarily use a 4G connection to access the social network site. The advertis-
ing platform displays an estimated audience size of 150k users matching these
targeting criteria.

As for most social networks, Facebook’s main revenue stream is targeted

online advertising, where advertisers are offered a rich set of targeting options
to reach the desired audience. In addition to the frequently supported targeting
around age, gender, device type, or topical interests, Facebook’s ad platform

4
provides the ability to target users based on the country of the previous residence
(“users who used to live in country X”). For a specified set of targeting criteria,
the platform then provides an estimate of the audience size called the Monthly
Active Users (MAU). For example, Figure 1 shows the specifications for an ad
targeting women aged 18+ living in Colombia who used to live in Venezuela
and who primarily use a 4G connection to access the platform. For this target
group, the ad platform displays an estimated 150k users matching this criteria2 .
These estimates can be collected programmatically through an API for various
combinations of targeting options providing aggregated and anonymized data
on the distribution of Facebook users in a given location. Such data can provide
opportunities for social science research including studies on migration. This
section reviews some works that use this data for migration research and the
methods they employ to analyze the data and address bias.

Figure 2: Facebook and American Community Survey (ACS) profiles of stocks

of migrants by age and sex for Mexicans in California and in Texas. Source:
Zagheni et al. (2017).

Zagheni et al. (2017) and Spyratos et al. (2019) study how Facebook’s
audience estimates of users from different countries of previous residence can
be used to estimate the size of migrant populations to complement migration
statistics. Zagheni et al. (2017) observe strong correlations between the fraction
of Facebook users from various countries of the previous residence across US
states and the size of foreign-born populations from the American Community
Survey (ACS) and at the country level with the size of migrant stocks from the
2 Accessed on November 10, 2020.

5
World Bank data. Comparing the ACS data to Facebook’s audience estimates,
the authors observe patterns of bias across age and country of previous residence
(Figure 2) which they correct for by including country and age fixed effects
in a regression model that predicts the survey data from the Facebook user
estimates. They observe improved model accuracy in estimating the size of
migrant populations when this bias correction is done. Spyratos et al. (2019)
use these data to compute estimates of migrant populations by country of origin
globally in different destination countries. They address the selection biases in
the sample of Facebook users by dividing the audience estimates for each migrant
group by the Facebook penetration so as to correct for differences in Facebook
usage across different countries of origin and destination and across different age
and gender groups. They develop a model to compute the Facebook penetration
for each migrant group as a weighted average of the Facebook penetration in
the countries of current and previous residences. These weights parameters were
then estimated to give the best predictive fit between the corrected Facebook
estimates and survey data.
Alexander et al. (2019) study the utility of such data for estimating short-
term population movements such as in the aftermath of natural disasters, fo-
cusing on the case of out-migration of Puerto Ricans to the continental US after
Hurricane Maria. They estimated the percentage change in the population of
Puerto Ricans by comparing the number of Facebook users in the USA who
previously lived in Puerto Rico in the periods before and after Hurricane Maria.
In order to account for changes not due to the Hurricane, such as an update to
Facebook’s estimation algorithm which may have affected user estimates for a
broad range of users, they adopt a difference-in-differences approach whereby
the observed change in the user estimates for other migrant groups was deducted
from the observed change in the user estimate for Puerto Ricans. Given that
changes in the estimates of Facebook users for other migrant groups generally
moved in lock-step with that for Puerto Ricans in the periods prior to the Hurri-
cane, this approach was used to isolate changes in the population of the Puerto
Ricans resulting from the Hurricane. The authors then estimate the number of
Puerto Ricans who moved to the continental US by multiplying the estimated
percentage increase in the population of Puerto Rican Facebook users by the
number of Puerto Ricans in the US prior to the Hurricane based on the survey
data. The authors observed an increase in the number of Puerto Ricans with a
marked increase for the younger, working-age population.
Palotti et al. (2020) study how the estimates of Facebook users from Venezuela
could be used to estimate the size of Venezuelan refugee populations at national
and subnational levels for various countries in Latin America. To estimate the
size of Venezuelan refugees, they correct Facebook’s estimates of users who pre-
viously lived in Venezuela by the overall Facebook penetration in the destination
country. The authors find a strong correlation between these estimates and the
available survey data at national and subnational geographic resolutions. In
addition to estimating the size of refugee populations, the authors also look at
insights that can be gained about the educational and economic situation of
these populations. They develop regression models to estimate income levels

6
based on the device types used on the platform and use the information on
self-declared education levels to look at the educational qualifications of the
Venezuelan refugees across countries and across regions in a country.
Beyond quantitative estimates of the size of migrant populations, Dubois
et al. (2018) and Stewart et al. (2019) study the qualitative aspect of cultural as-
similation by using the interest-based targeting options of the platform. Dubois
et al. (2018) collect the audience estimates of Facebook users for a list of inter-
ests in both the destination country and the home country of the target migrant
group under study. They compare the number of users with these different in-
terests in the home and destination countries in order to filter the interests to
those that are more popular in the destination countries. Using the filtered list
of interests, they then compute an assimilation score that compares the relative
popularity of these interests in the destination country and among the targeted
migrant groups in the destination country. Dubois et al. (2018) apply their ap-
proach to study migrant assimilation for Arabic-speaking migrants in Germany
across different demographic groups. Stewart et al. (2019) use the methodol-
ogy of Dubois et al. (2018) to study the assimilation of Mexican migrants to the
Anglo and African-American populations in the USA. For their study, Stew-
art et al. (2019) focus on interests in musical genres to study the assimilation
of Mexican migrants along demographic dimensions by age, gender, education
levels, and language.
The wide range of targeting criteria provided on the Facebook advertis-
ing platform makes it possible to study other aspects of migration and mobil-
ity. Spyratos et al. (2020) use data on estimates of users by travel frequency
(using the option to target users who are “frequent travelers” or “frequent in-
ternational travelers”) to study the traveling behavior of migrant groups by
country of previous and current residence. As a validation step, the authors
observe strong correlations between the estimates of Facebook users by country
of the previous residence who are frequent travelers or frequent international
travelers and the per capita number of international travels by nationality in
the UK and per capita income by nationality in the US respectively. To model
the traveling behavior of migrants in this sample of Facebook users, the authors
fit various regression models to explain the fraction of users who are frequent
(international) travelers as a function of other variables such as user demograph-
ics, characteristics of the country of previous or current residence such as income
levels and gender inequalities and Facebook penetration rates.
As the studies reviewed here have shown, the advertising audience estimates
data from the Facebook marketing platform provides a useful data source for
studies of migration and mobility. Applications can range from estimating the
size of migrant populations to complement traditional migrant statistics, esti-
mating short-term migration movements as a result of natural disasters or in the
context of a refugee crisis where official data may be lacking or appear with a
time-lag and gaining further insights on other aspects of migration such as edu-
cation, socioeconomics, cultural assimilation, and traveling behavior. However,
the advertising audience estimates data has limitations. Section 4 discusses the
limitations of using social media data sources for studies of migration.

7
3 LinkedIn example(s) - what has been done
Unlike Facebook, LinkedIn’s advertising platform does not provide a mecha-
nism to explicitly target users based on countries they have lived in before.
However, LinkedIn supports targeting users by “Member Schools” to “reach
members who completed a course at a specific school, college, university or
other learning institution”. This targeting is available for schools that have
a dedicated page on LinkedIn, such as https://www.linkedin.com/school/
university-of-cambridge/ for the University of Cambridge. Using the ap-
propriate search API endpoint3 , one can look up the corresponding ID, in this
case urn:li:school:12691. This ID is needed for API calls to obtain audience
estimates through the corresponding endpoint 4 .

Figure 3: A screenshot of LinkedIn’s advertising platform, showing a target

audience size of 420 members for the selection of female LinkedIn users who
studied at the University of Cambridge, hold a PhD, and are living in Germany.

To approximate the criterion of “having lived in country X”, we compiled

a list of higher education institutions in country X, so that we could con-
struct a query for LinkedIn users who “studied at a higher education institu-
tion in country X”. Concretely, for each European country, including territories
3 https://www.linkedin.com/campaign-manager-api/campaignManagerAdTargetingEntities
4 https://www.linkedin.com/campaign-manager-api/campaignManagerAudienceCounts

8
such as the Isle of Man, we (i) obtained a list of universities from uniRank
(https://www.4icu.org/) to search for, (ii) searched for academic institutions
for each city, region, or country targetable by LinkedIn, and (iii) later filtered
the returned results to remove kindergartens and high schools, as well as false
positives, such as universities with ambiguous names from non-European coun-
tries.
Figure 4 shows a circular plot of the estimate of LinkedIn users who (i)
studied in country X (origin of the arrow), and (ii) now5 live in country Y,
different from X (target of the arrow). The estimates are in millions and small
countries are not shown to avoid clutter. Furthermore, the estimates are rescaled
to correct for LinkedIn penetration in the target country: for a host country
with a population of P and a total number of resident LinkedIn users L, all
estimates are scaled upwards by P/L. Assuming that migrants are more likely
than non-migrants to be on LinkedIn, these re-scaled estimates would be upper
bounds on the true numbers.
Whereas the fairly large country-to-country mobility shown in Figure 4 might
not surprise given the European context and programs such as Erasmus6 , the
same methodology can be applied to shine a light on otherwise overlooked high-
skilled migrants. Figure 5 shows similarly obtained estimates for LinkedIn users
who (i) studied at a university in Syria, but (ii) are currently7 living in a Eu-
ropean country. These estimates, which were generated by Michele Vespe and
Spyridon Spyratos at the European Commission, are obtained by scaling up
the raw estimates by the LinkedIn penetration among college graduates. Con-
cretely, for each host country separately, they computed the ratio of “LinkedIn
users with a self-declared BA, MA, or Ph.D.” / “population with a univer-
sity degree according to Eurostat”. Assuming the LinkedIn penetration among
university-educated Syrians is similar to the LinkedIn penetration among the
host population, this correction factor appropriately corrects for the undersam-
pling. This case study and the corresponding figure were shared by the authors
of Vespe and Spyratos (2019) and are being used with permission.
Note that one challenge with this type of data collection is that LinkedIn
users who studied in several countries are counted multiple times. So a LinkedIn
user who studied both in Germany and in the United Kingdom, while currently
living in Spain, would be counted both as having moved to Spain from the
UK, as well a having moved to Spain from Germany. This double-count could
conceptually be addressed by iterating over pairs or triples of origin countries.
Concretely, one could obtain a separate estimate for the number of people who
have lived in two countries, such as Germany and the UK, and then subtract
these estimates from the bigger count based on single countries of origins. While
theoretically possible, this would increase the number of queries that would need
to be issued to the appropriate API by roughly a factor of n=28, the number
of EU member countries in 2018.
Another practical challenge lies in the fact that, unlike Facebook, LinkedIn
5 As of the time of data collection, i.e. October 2018.
6 https://ec.europa.eu/programmes/erasmus-plus/node_en
7 As of 2018.

9
 Belgium

mark
50 0 5

Den
45 0
40 5
35

Kin
Un om
0
30

gd
5

ite
25

d
10

ce
an
20 15

Fr
15 20

10

25
5
0

0
5
10

10
Tur
key
5

15
0

20
y
German

25
10

Switzerland
5

30
0

35
40
5

Sweden

45 50
0
25

0
Irela
nd

5
20
15

ain
Sp

0
10

5
5

Ita

10
0 ly
15
20
10
rat n
de sia

5
ion

0
Fe Rus

0 5
Ne

5 10
gal

0 15
the

0
tu

Norway
Poland

10 5 0
rlan
Por

ds

Figure 4: A circular plot showing estimates for how many people studied in
one EU country, but are currently living in another. Numbers are in thousands
(1000’s). The estimates are obtained through LinkedIn’s advertising platform
and are corrected for LinkedIn penetration rates in the host country. Assum-
ing migrants are over-represented on LinkedIn, compared to non-migrants, the
estimates would be upper bounds for the true numbers.

does not offer documentation for the existing free-of-charge APIs that are some-
what hidden. Documentation offered online by LinkedIn instead refers to a pre-
mium API that is aimed at bigger advertisers. As such, we advise researchers
who want to automate data collection from LinkedIn to study the hidden API
using network monitoring tools that are built into the browser8 . These net-
work monitoring tools log the requests sent to the LinkedIn server by the web
8 See https://developer.mozilla.org/en-US/docs/Tools/Network_Monitor for Firefox,

or https://developers.google.com/web/tools/chrome-devtools/network for Chrome

10
Figure 5: Estimates of the number of college-educated people from Syria living
in the EU. The estimates are derived from LinkedIn advertising data with an
attempt to correct for selection bias (see text). Source: Vespe and Spyratos
(2019).

browser. These requests are being made, usually unnoticed by the user, while
the LinkedIn Campaign Manager is being used normally. Through these tools,
it is easy to understand, say, which API is being called to search for a school
of a particular name, or which API provides the actual audience estimates. Us-
ing the appropriate structure of the API calls, together with required cookies
obtained after signing in, it is relatively easy to automate data collection using
the APIs. In Section 4 we present the common limitations and challenges when
working with advertising platforms.

11
4 General limitations/challenges
Despite all the advantages of using these novel methods to get data for de-
mographic research, it is important to be aware of their limitations, which we
should be careful about.
Self-selection bias: The first limitation is part of any social network re-
search: Self-selection bias. Most social network users self-select to be on the
platform, hence they do not create a representative sample of the society. Cer-
tain affordances provided by platforms also make them amenable to certain
types of user sets. e.g. Highly educated users are overrepresented on LinkedIn,
urban users are overrepresented on Twitter, etc. Perrin and Anderson (2019).
Self-report bias: Even if we correct for such biases, there are issues with
using observational data, which relate to self-report bias. Users might not al-
ways provide the right information. For instance, in many cases, we rely on
self-reported data such as education level on Facebook or schools attended on
LinkedIn. The current assumption with such datasets is to take what users
provide at face value, which might not always be the best case.
Algorithmic bias: A more serious issue particular to advertising data is
that they are not devised to be used for the purpose we illustrate in this chapter.
Firstly, these datasets were generated for advertising purposes and not to study
migration, and, hence might have different objectives. Secondly, the data is a
result of black-box algorithms, which are in most cases, are company secrets that
will not be revealed. So any biases going into these black-box algorithms will
be replicated in the data, and eventually, our analysis. Not just the biases, but
any errors in the inference process will also be replicated in our analysis. Third,
given the proprietary nature of the data, and the process of generation, it might
undergo constant updates, which are not be reflected to the researchers, who
mistakenly assume that the data collection process is consistent. For instance,
Facebook keeps changing its definition of ‘users who lived in country X’ every
year. As a result, even though we see the ease of getting new data as a positive
(e.g. to measure how trends changed after an event, such as a natural disaster),
we can not be sure if the change was due to the changes in Facebook algorithms
or due to the new users who came into the platform.
Brittleness of the APIs: On the practical side, the APIs to get the data
change all the time or disappear altogether, leading to issues with replicable
research. For instance, in our own experience, LinkedIn advertising API changed
from HTTP GET to HTTP POST unannounced, thus leading us to rewrite all
our code.
Privacy concerns: The next big concern with such data is the privacy
risks posed. The datasets we propose to use in this chapter are aggregated
anonymized data. They only use counts of users who satisfy certain criteria.
However, in the past, there have been studies that showed how such aggregated
data and other tools that Facebook provides could be misused Speicher et al.
(2018). Though these risks have now been fixed, it shows the potential that
aggregated data is not always anonymized. To preserve individual anonymity,
these advertising APIs do not return any user counts less than a certain mini-

12
mum threshold (1,000 users on Facebook in November 2020). A practical side
effect of the aggregation on research studies is that such aggregation could lead
to biases. The counts of small populations can not be obtained accurately. One
technique to reduce these biases was recently proposed in Rama et al. (2020).
Even more important to note is that anonymized, aggregated data does
not necessarily mitigate all privacy concerns. Privacy is not just of concern
to individuals. There are group-level harms that can be possible, even with
aggregate data, which can be misused by bad actors. For instance, authoritarian
governments could use such data to obtain aggregate statistics of regions in a
country with high immigrant populations or people of a certain faith.
Ethical and legal concerns: Finally, there are ethical and legal challenges
that are not completely clear in using such data. It is important to keep in
mind that some of the data, particularly related to inferring interests of users
are collected by the platforms without the consent of the users. The use of data
brokers Venkatadri et al. (2019), might involve data collected without consent
and for users to opt-out. There are some interesting ongoing legal arguments
about how compliant the computation and targeting with such interests is with
GDPR.9
In summary, though advertising data might appear to be a viable source of
data for demographic research, it has its own caveats and limitations which we
should be aware of while using it in our research. This is especially important
since the output of such research could be used to inform policy.

5 The next frontier: Combining data and fully

leveraging the infrastructure of the digital age
This last section of the chapter discusses expected future research developments
in this area. More specifically, it focuses on the new possibilities offered by
combining different types of data sources with those provided by the advertise-
ment platforms for Facebook and LinkedIn. The first main way of combining
probabilistic samples and passively collected data involves developing statistical
models, often using Bayesian approaches. The second main way of combining
sources and tools involves using advertisement platforms to recruit participants
for surveys.
An emerging line of literature that we expect to develop further in years to
come is about combining different types of data sources using Bayesian statisti-
cal methods. The underlying idea is that different types of data have different
types of biases and imperfections. These issues that affect data quality can be
explicitly modeled by incorporating additional information. This can be done
in a formal and reproducible way using Bayesian methods. More specifically,
researchers can incorporate beliefs either via the use of so-called priors or by
structuring models in a way that certain types of information are pooled across
9 https://techcrunch.com/2018/05/16/facebook-faces-fresh-criticism-over-ad-

targeting-of-sensitive-interests/

13
space and time.
In a recent example centered around nowcasting stocks of migrants in the US,
Alexander et al. (2020) show that a Bayesian hierarchical model that combines
data from the Facebook advertisement platform and data from the American
Community Survey produces more accurate short-term forecasts than either a
model that relies only on Facebook data or a model that relies only on time
series from the American Community Survey.

Figure 6: Schematic illustration of the approach used to combine data from

the Facebook advertisement platform and data from the American Community
Survey to produce more accurate short-term forecasts of migrant stocks in the
US. Source: Alexander et al. (2020).

Figure 6 shows a schematic illustration of the approach used to combine

data from the Facebook advertisement platform and data from the American
Community Survey to produce more accurate short-term forecasts of migrant
stocks in the US. While the figure refers to the paper by Alexander et al. (2020),
the approach is quite general and a number of studies are testing Bayesian
models for combining social media and survey data to estimate stocks or flows
of migrants (e.g., Hsiao et al. (2020)) The second main area of development is
the use of the advertisement infrastructure to collect data via surveys. Migrants
are a hard-to-reach population. Targeted advertisements may provide the tools
necessary to reach these populations and to invite them to participate in a
survey. Initial approaches showed that this is a cost-effective way of collecting
data (Pötzschke and Braun (2017)). A rapidly expanding related literature in
public health has also shown that, when appropriate post-stratification weights
are used, surveys run via advertisement platforms can generate estimates that
are close approximations of the ones obtained from probabilistic samples Grow
et al. (2020).

14
 In sum, we expect that the next frontier is about blending approaches from
classic data collections, survey methods, and Bayesian statistics, with the in-
frastructure of the digital age, which offers not only digital trace data but also
new opportunities for data collection and survey experiments.

6 Acknowledgements
We thank Carlos Callejo Peñalba who in 2018-2019 was a Master’s student at
Aalto University in Finland. As part of his course project, Carlos collected and
visualized LinkedIn data on the number of users who studied in one EU country
but now live in another EU country.
We also thank Michele Vespe and Spyros Spyratos at the Joint Research
Center of the European Commission in Ispra, Italy. Michele and Spyros let us
use their analysis of the number of LinkedIn users who studied in Syria but now
live in an EU country.

References
Alexander, M., Polimis, K., and Zagheni, E. (2019). The Impact of Hur-
ricane Maria on Out-migration from Puerto Rico: Evidence from Face-
book Data. Population and Development Review, 45(3):617–630. eprint:
https://onlinelibrary.wiley.com/doi/pdf/10.1111/padr.12289.

Alexander, M., Polimis, K., and Zagheni, E. (2020). Combining social media
and survey data to nowcast migrant stocks in the united states. Population
Research and Policy Review.
Dubois, A., Zagheni, E., Garimella, K., and Weber, I. (2018). Studying migrant
assimilation through facebook interests. In International Conference on Social
Informatics, pages 51–60. Springer.
Grow, A., Perrotta, D., Del Fava, E., Cimentada, J., Rampazzo, F., Gil-Clavel,
S., and Zagheni, E. (2020). Addressing public health emergencies via facebook
surveys: Advantages, challenges, and practical considerations.

Hsiao, Y., Fiorio, L., Wakefield, J., Zagheni, E., et al. (2020). Modeling the bias
of digital data: an approach to combining digital and survey data to estimate
and predict migration trends. Technical report, Max Planck Institute for
Demographic Research, Rostock, Germany.
Palotti, J., Adler, N., Morales-Guzman, A., Villaveces, J., Sekara, V., Herranz,
M. G., Al-Asad, M., and Weber, I. (2020). Monitoring of the Venezuelan ex-
odus through Facebook’s advertising platform. PLOS ONE, 15(2):e0229175.
Publisher: Public Library of Science.

15
Perrin, A. and Anderson, M. (2019). Share of us adults using social media,
including facebook, is mostly unchanged since 2018. Pew Research Center,
10.
Pötzschke, S. and Braun, M. (2017). Migrant sampling using facebook adver-
tisements: A case study of polish migrants in four european countries. Social
Science Computer Review, 35(5):633–653.
Rama, D., Mejova, Y., Tizzoni, M., Kalimeri, K., and Weber, I. (2020). Face-
book ads as a demographic tool to measure the urban-rural divide. In Pro-
ceedings of The Web Conference 2020, pages 327–338.

Speicher, T., Ali, M., Venkatadri, G., Ribeiro, F., Arvanitakis, G., Benevenuto,
F., Gummadi, K., Loiseau, P., and Mislove, A. (2018). Potential for discrim-
ination in online targeted advertising.
Spyratos, S., Vespe, M., Natale, F., Iacus, S. M., and Santamaria, C. (2020).
Explaining the travelling behaviour of migrants using Facebook audience es-
timates. PLOS ONE, 15(9):e0238947. Publisher: Public Library of Science.
Spyratos, S., Vespe, M., Natale, F., Weber, I., Zagheni, E., and Rango, M.
(2019). Quantifying international human mobility patterns using Facebook
Network data. PLOS ONE, 14(10):e0224134. Publisher: Public Library of
Science.

Stewart, I., Flores, R. D., Riffe, T., Weber, I., and Zagheni, E. (2019). Rock,
Rap, or Reggaeton?: Assessing Mexican Immigrants’ Cultural Assimilation
Using Facebook Data,. In The World Wide Web Conference, WWW ’19,
pages 3258–3264, San Francisco, CA, USA. Association for Computing Ma-
chinery.

Venkatadri, G., Sapiezynski, P., Redmiles, E. M., Mislove, A., Goga, O.,
Mazurek, M., and Gummadi, K. P. (2019). Auditing offline data brokers
via facebook’s advertising platform. In The World Wide Web Conference,
WWW ’19, pages 1920–1930, New York, NY, USA. ACM.
Vespe, M. and Spyratos, S. (2019). A changing migration data land-
scape? Global Working Group on Big Data for Official Statis-
tics, International Meeting on Measuring Human Mobility. https:
//unstats.un.org/bigdata/events/2019/tbilisi/presentations/
Session%202/Tbilisi%20Vespe%20-%20final.pdf.
Zagheni, E., Weber, I., and Gummadi, K. (2017). Leveraging
Facebook’s Advertising Platform to Monitor Stocks of Migrants.
Population and Development Review, 43(4):721–734. eprint:
https://onlinelibrary.wiley.com/doi/pdf/10.1111/padr.12102.

16