Understanding the Demographics of Twitter Users
Alan Mislove†Sune Lehmann∗Yon g- Yeo l A h n†Jukka-Pekka Onnela‡J. Niels Rosenquist‡
†Northeastern University ∗Tec h n i c a l U nive r s i ty of D e n m ark ‡Harvard Medical School
Every second, the thoughts and feelings of millions of people
across the world are recorded in the form of 140-character
tweets using Twitter. However, despite the enormous poten-
tial presented by this remarkable data source, we still do not
have an understanding of the Twitter population itself: Who
are the Twitter users? How representative of the overall pop-
ulation are they? In this paper, we take the ﬁrst steps towards
answering these questions by analyzing data on a set of Twit-
ter users representing over 1% of the U.S. population. We
develop techniques that allow us to compare the Twitter pop-
ulation to the U.S. population along three axes (geography,
gender, and race/ethnicity), and ﬁnd that the Twitter popula-
tion is a highly non-uniform sample of the population.
Online social networks are now a popular way for users to
connect, communicate, and share content; many serve as the
de-facto Internet portal for millions of users. Because of the
massive popularity of these sites, data about the users and
their communication offers unprecedented opportunities to
examine how human society functions at scale. However,
concerns over user privacy often force service providers to
keep such data private. Twitter represents an exception: Over
91% of Twitter users choose to make their proﬁle and com-
munication history publicly visible, allowing researchers
access to the vast majority of the site. Twitter, therefore,
presents a unique opportunity to examine the public com-
munication of a large fraction of the population.
In fact, researchers have recently begun to use the con-
tent of Twitter messages to measure and predict real-world
phenomena, including movie box ofﬁce returns (Asur and
Huberman 2010), elections (O’Connor et al. 2010), and the
stock market (Bollen, Mao, and Zeng 2010). While these
studies show remarkable promise, one heretofore unan-
swered question is: Are Twitter users a representative sam-
ple of society? If not, which demographics are over- or un-
derrepresented in the Twitter population? Because existing
studies generally treat Twitter as a “black box,” shedding
light on the characteristics of the Twitter population is likely
to lead to improvements in existing prediction and measure-
ment methods. Moreover, understanding the characteristics
!2011, Association for the Advancement of Artiﬁcial
Intelligence (www.aaai.org). All rights reserved.
of the Twitter population is crucial to move towards more
advanced observations and predictions, since such an un-
derstanding will help us determine what predictions can be
made and what other data is necessary to correct for any bi-
In this paper, we take a ﬁrst look at the demographics of
the Twitter users, aiming to answer these questions. To do
so, we use a data set of over 1,755,925,520 Twitter messages
sent by 54,981,152 users between March 2006 and August
2009 (Cha et al. 2010). We focus on users whose identiﬁed
location is within the United States, because the plurality of
users at the time of the data collection are in U.S., and be-
cause we have the detailed demographic data for U.S. popu-
lation. Even with the location constraint, our dataset covers
over three million users, representing more than 1% of the
entire U.S. population.
Ideally, when comparing the Twitter population to society
as a whole, we would like to compare properties including
socio-economic status, education level, and type of employ-
ment. However, we are restricted to only using the data that
is (optionally) self-reported and made visible by the Twitter
users, including their name, location, and the text of their
tweets. We develop techniques to examine the properties
of the Twitter population along three separate but interre-
lated axes, based on the feasibility of comparison. First, we
compare the geographic distribution of users to the popu-
lation as a whole using U.S. Census data. We demonstrate
that Twitter users are more likely to live within populous
counties than would be expected from the Census data, and
that sparsely populated regions of the U.S. are signiﬁcantly
underrepresented. Second, we infer the gender of Twitter
users and demonstrate that a signiﬁcant male bias exists,
although the bias is becoming less pronounced over time.
Third, we examine the race/ethnicity of Twitter users and
demonstrate that the distribution of race/ethnicity is highly
Detection location using self-reported data
To d e t e r m i ne geog r a p h ic i n f o rm a ti o n a b ou t users, w e use
the self-reported location ﬁeld in the user proﬁle. The loca-
tion is an optional self-reported string; we found that 75.3%
of the publicly visible users listed a location. In order to turn
the user-provided string into a mappable location, we use
the Google Maps API. Beginning with the most popular lo-
cation strings (i.e, the strings provided by the most users),
we query Google Maps with each location string. If Google
Maps is able to interpret a string as a location, we receive a
latitude and longitude as a response. We restrict our scope to
users in the U.S. by only considering response latitudes and
longitudes that are within the U.S.. In total, we ﬁnd map-
pings to a U.S. longitude and latitude for 246,015 unique
strings, covering 3,279,425 users (representing 8.8% of the
users who list a location).
To com p a r e o u r Tw i tt e r d at a t o t h e 2 00 0 U . S. C e n su s , i t
is necessary to aggregate the users into U.S. counties. Using
data from the U.S. National Atlas and the U.S. Geological
Survey, we map each of the 246,015 latitudes and longitudes
into their respective U.S. county. Unless otherwise stated,
our analysis for the remainder of this paper is at the U.S.
Limitations We now b rieﬂ y d i sc u ss po te nt ia l li m it at io ns
of our location inference methodology. First, it is worth not-
ing that Google Maps will also interpret locations that are
at a granularity coarser than a U.S. county (e.g., “Texas”).
We m a n ua ll y re mo v ed t h es e , incl u d in g th e map p i n gs o f all
50 states, as well as “United States” and “Earth.” Second,
users may lie about their location, or may list an out-of-date
location. Third, since the location is per-user (rather than
per-tweet), a user who moves from one city to another (and
updates his location) will have all of his tweets considered
as being from the latter location.
Geographic distribution of Twitter users
We b e g i n b y ex am in in g th e ge ogra p h ic d i s tr ib u t io n of Twit -
ter users, and comparing it to the entire U.S. population.
Overall, the 3,279,425 Twitter users who we are able to geo-
locate represent 1.15% of the entire population (at the time
of the 2000 Census). However, if we examine the distribu-
tion of Twitter users per county, we observe a highly non-
Figure 1 presents this analysis, with the county popula-
tion along the xaxis and the fraction of this population we
observe in Twitter along the yaxis. We see that, as the popu-
lation of the county increases, the Tw i t t e r repre s e n t a t i o n rate
(simply the number of Twitter users in that county divided
by the number of people in that county in the 2000 U.S.
Census) increases as well. For example, consider the median
per-county Twitter representation rate of 0.324%. We ob-
serve that 93.5% of the counties with over 100,000 residents
have a higher Twitter representation rate than the median,
compared to only 40.8% of the counties with fewer than
100,000 residents (were Twitter users a truly random pop-
ulation sample, we would expect these percentages to both
be 50%). Thus, the Twitter users signiﬁcantly overrepresent
populous counties, a fact underscored by the difference be-
tween the median (0.324%) per-county Twitter representa-
tion rates and the overall population sample of 1.15%.
The overrepresentation of populous counties in and of it-
self may not come as a surprise, due to the patterns of so-
cial media adoption across different regions. However, the
Twitter Representation Rate
Figure 1: Scatterplot of US county population versus Twitter
representation rate in that county. The dark line represents
the aggregated median, and the dashed black line represents
the overall median (0.324%). There is a clear overrepresen-
tation of more populous counties.
magnitude of the difference is striking: We observe an or-
der of magnitude difference in median per-county Twitter
representation rate between counties with 1,000 people and
counties with 1,000,000 people. This indicates a bias in the
Twitter p o pu l a ti o n ( re l a t ive t o th e U. S . po p ul a t i o n) a n d sug-
gests that entire regions of the U.S. may be signiﬁcantly un-
Distribution across counties We n o w e x am in e which re-
gions of the U.S. contain these over- and underrepresented
counties. To do so, we plot a map of the U.S. based on the
Twitter re p r e s e n ta t i o n ra t e , r e l a t ive to t h e m e d ia n r a t e o f
0.324%. Figure 2 presents this data, using both a normal rep-
resentation and an area cartogram representation (Gastner
and Newman 2004). In this ﬁgure, the counties are colored
according to the level of over- or underrepresentation, with
blue colors representing underrepresentation and red colors
representing overrepresentation, relative to the median rate
of 0.324%. Thus, the same number of counties will be col-
ored red as blue.
These two maps lead to a number of interesting conclu-
sions: First, as evident in the normal representation, much of
the mid-west is signiﬁcantly underrepresented in the Twit-
ter user base in this time period. Second, as evident in the
signiﬁcantly red hue of the area cartogram, more populous
counties are consistently oversampled. However, the level of
oversampling does not appear to be dependent upon geogra-
phy: Both east coast and west coast cities are clearly visible
(e.g., San Francisco and Boston), as well as mid-west and
southern cities (e.g, Dallas, Chicago, and Atlanta).
Detecting gender using ﬁrst names
As we have very limited information available on each user,
we rely on using the self-reported name available in each
user’s proﬁle in order to detect gender. To do so, we ﬁrst ob-
tain the most popular 1,000 male and female names for ba-
bies born in the U.S. for each year 1900–2009, as reported
by the U.S. Social Security Administration (Social Secu-
rity Administration 2010). We then aggregate the names to-
gether, calculating the total frequency of each of the result-
ing 3,034 male and 3,643 female names. As certain names
occurred in both lists, we remove the 241 names that were
(a) Normal representation (b) Area cartogram representation
Figure 2: Per-county over- and underrepresentation of U.S. population in Twitter, relative to the median per-county represen-
tation rate of 0.324%, presented in both (a) a normal layout and (b) an area cartogram based on the 2000 Census population.
Blue colors indicate underrepresentation, while red colorsrepresentoverrepresentation.Theintensityofthecolorcorresponds
to the log of the over- or underrepresentation rate. Clear trends are visible, such as the underrepresentation of mid-west and
overrepresentation of populous counties.
less than 95% predictive (e.g., the name Avery was observed
to correspond to male babies only 56.8% of the time; it was
therefore removed). The result is a list of 5,836 names that
we use to infer gender.
Limitations Clearly, this approach to detecting gender is
subject to a number of potential limitations. First, users may
misrepresent their name, leading to an incorrect gender in-
ference. Second, there may be differences in choosing to re-
veal one’s name between genders, leading us to believe that
fewer users of one gender are present. Third, the name lists
above may cover different fractions of the male and female
Gender of Twitter users
We ﬁr s t d et e rm in e th e n um ber of th e 3 , 27 9, 425 U. S. -b as e d
users who we could infer a gender for, based on their name
and the list previously described. We do so by comparing
the ﬁrst word of their self-reported name to the gender list.
We o b s e rv e th a t the r e e x is t s a mat c h f o r 64 .2% of t h e u se rs .
Moreover, we ﬁnd a strong bias towards male users: Fully
71.8% of the the users who we ﬁnd a name match for had a
2007-01 2007-07 2008-01 2008-07 2009-01 2009-07
Fraction of Joining Users
who are Male
Figure 3: Gender of joining users over time, binned into
groups of 10,000 joining users (note that the join rate in-
creases substantially). The bias towards male users is ob-
served to be decreasing over time.
To fur t h e r e x p l or e t h is t r e nd , w e ex am i n e th e h is t o r ic g e n -
der bias. To do so, we use the join date of each user (avail-
able in the user’s proﬁle). Figure 3 plots the average fraction
of joining users who are male over time. From this plot, it
is clear that while the male gender bias was signiﬁcantly
stronger among the early Twitter adopters, the bias is be-
coming reduced over time.
Detecting race/ethnicity using last names
Again, since we have very limited information available
on each Twitter user, we resort to inferring race/ethnicity
using self-reported last name. We examine the last name
of users, and correlate the last name with data from the
U.S. 2000 Census (U.S. Census 2000). In more detail, for
each last name with over 100 individuals in the U.S. dur-
ing the 2000 Census, the Census releases the distribution of
race/ethnicity for that last name. For example, the last name
“Myers” was observed to correspond to Caucasians 86% of
the time, African-Americans 9.7%, Asians 0.4%, and His-
Race/ethnicity distribution of Twitter users
We ﬁr s t d e t er mi ne d th e num b e r o f U.S . - b as ed u s e rs f or
whom we could infer the race/ethnicity by comparing the
last word of their self-reported name to the U.S. Census
last name list. We observed that we found a match for
71.8% of the users. We the determined the distribution of
race/ethnicity in each county by taking the race/ethnicity
distribution in the Census list, weighted by the frequency
of each name occurring in Twitter users in that county.1
Due to the large amount of ambiguity in the last name-to-
race/ethnicity list (in particular, the last name list is more
than 95% predictive for only 18.5% of the users), we are un-
able to directly compare the Twitter race/ethnicity distribu-
1This is effectively the census.model approach discussed in
prior work (Chang et al. 2010).
(a) Caucasian (non-hispanic) (b) African-American (c) Asian or Pacific Islander (d) Hispanic
Figure 4: Per-county area cartograms of Twitter over- and undersampling rates of Caucasian, African-American, Asian, and
Hispanic users, relative to the 2000 U.S. Census. Only counties with more than 500 Twitter users with inferred race/ethnicity
are shown. Blue regions correspond to undersampling; red regions to oversampling.
tion directly to race/ethnicity distribution in the U.S. Census.
However, we are able to make relat iv e comparisons between
Twitter u s e r s in differen t g e o g ra p hi c r e g io n s , al l o w i n g us to
explore geographic trends in the race/ethnicity distribution.
Thus, we examine the per-county race/ethnicity distribution
of Twitter users.
In order to account for the uneven distribution of
race/ethnicity across the U.S., we examine the per-county
race/ethnicity distribution relative to the distribution from
the overall U.S. Census. For example, if we observed that
25% of Twitter users in a county were predicted to be His-
panic, and the 2000 U.S. counted 23% of people in that
county as being Hispanic, we would consider Twitter to be
oversampling the Hispanic users in that county. Figure 4
plots the per-county race/ethnicity distribution, relative to
the 2000 U.S. Census, per all counties in which we observed
more than 500 Twitter users with identiﬁable last names.
dersampling of Hispanic users in the southwest; the under-
samping of African-American users in the south and mid-
west; and the oversampling of Caucasian users in many ma-
cial network users. For example, recent studies have exam-
ined the ethnicity of Facebook users (Chang et al. 2010),
general demographics of Facebook users (Corbett 2010),
and differences in online behavior on Facebook and MyS-
pace by gender (Strayhorn 2009). However, studies of gen-
eral social networking sites are able to leverage the broad
nature of the proﬁles available; in contrast, on Twitter, users
self-report only a minimal set of information, making calcu-
lating demographics signiﬁcantly more difﬁcult.
Twitter h a s r e c eived s i g n i ﬁ c an t r e s e ar c h i n te r e st l a t e l y asa
means for understanding, monitoring, and even predicting
real-world phenomena. However, most existing work does
not address the sampling bias, simply applying machine
learning and data mining algorithms without an understand-
ing of the Twitter user population. In this paper, we took
ined the population along the axes of geography, gender, and
race/ethnicity. Overall, we found that Twitter users signif-
icantly overrepresent the densely population regions of the
U.S., are predominantly male, and represent a highly non-
random sample of the overall race/ethnicity distribution.
Going forward, our study sets the foundation for future
work upon Twitter data. Existing approaches could imme-
diately use our analysis to improve predictions or measure-
ments. By enabling post-hoc corrections, our work is a ﬁrst
step towards turning Twitter into a tool that can make infer-
ences about the population as a whole. More nuanced anal-
yses on the biases in the Twitter population will enhance
the ability for Twitter to be used as a sophisticated inference
We th a n k F a br ic io B e nev e nto and M e e y ou ng Ch a for t h e i r
assistance in gathering the Twitter data used in this study.
We als o t h an k Ji m Ba gr ow f o r va l u ab le d is c ussio n s a n d his
collection of geographic data from Google Maps. This re-
search was supported in part by NSF grant IIS-0964465 and
an Amazon Web Services in Education Grant.
Asur, S., and Huberman, B. 2010. Predicing the future with social
Bollen, J.; Mao, H.; and Zeng, X.-J. 2010. Twitter mood predicts
the stock market. In ICWSM.
Cha, M.; Haddadi, H.; Benevenuto, F.; and Gummadi, K. 2010.
Measuring user inﬂuence in twitter: The million follower fallacy.
Chang, J.; Rosenn, I.; Backstrom, L.; and Marlow, C. 2010.
epluribus: Ethnicity on social networks. In ICWSM.
Corbett, P. 2010. Facebook demographics and statistics re-
port 2010. http://www.istrategylabs.com/2010/01/
Gastner, M. T., and Newman, M. E. J. 2004. Diffusion-based
method for producing density-equalizing maps. PNAS 101.
O’Connor, B.; Balasubramanyan, R.; Routledge, B.; and Smith, N.
2010. From tweets to polls: Linking text sentiment to public opin-
ion time series. In ICWSM.
Social Security Administration. 2010. Most popular baby names.
Strayhorn, T. 2009. Sex differences in use of facebook and mys-
pace among ﬁrst-year college students. Stud. Affairs 10(2).
U.S. Census. 2000. Genealogy data: Frequently occur-
ring surnames from census. http://www.census.gov/