Proceedings of the Thirteenth International AAAI Conference on Web and Social Media (ICWSM 2019)

Environmental Changes and the Dynamics of Musical Identity

Samuel F. Way,1,2 Santiago Gil,2 Ian Anderson,2 Aaron Clauset1,3

1
University of Colorado at Boulder, 2 Spotify, 3 Santa Fe Institute
samuel.way@colorado.edu, santiago@spotify.com, iananderson@spotify.com, aaron.clauset@colorado.edu

Abstract shapes and is shaped by musical tastes (Schwartz and Fouts

2003; Schäfer and Mehlhorn 2017). Studies have found re-
Musical tastes reflect our unique values and experiences, our peatedly that mood regulation is among the most common
relationships with others, and the places where we live. But and important reasons for why people listen to music (Slo-
as each of these things changes, do our tastes also change
to reflect the present, or remain fixed, reflecting our past?
boda and O’neill 2001; Saarikallio and Erkkilä 2007): music
Here, we investigate how where a person lives shapes their helps listeners relax, improve their mood, or simply relate to
musical preferences, using geographic relocation to construct others through the emotions of music and its lyrics (Wells
quasi-natural experiments that measure short- and long-term and Hakanen 1991). Music also plays a crucial role as
effects. Analyzing comprehensive data on over 16 million a social currency, helping initiate and strengthen relation-
users on Spotify, we show that relocation within the United ships (Erickson 1996), for example, through the exchange
States has only a small impact on individuals’ tastes, which of new music or shared experiences at live performances.
remain more similar to those of their past environments. We In these ways, music brings together individuals, forming
then show that the age gap between a person and the music communities or “scenes” around particular genres, artists,
they consume indicates that adolescence, and likely their en- or the lifestyles they personify (Bennett and Peterson 2004;
vironment during these years, shapes their lifelong musical
tastes. Our results demonstrate the robustness of individuals’
Lena 2012; Cohen 1991).
musical identity, and shed new light on the development of When a community forms around some kind of mu-
preferences. sic, the surrounding environment takes on an identity of
its own. Cultural geographers have investigated this in-
Music is the soundtrack of our lives. It reflects our mood teraction between place and musical style (Hudson 2006;
and personality, as well as the important people, places, and Nash and Carney 1996), treating music as primary source
times in our past (DeNora 2000). In this way, a person’s mu- material for understanding what places are or used to be
sical identity—the set of musical tastes or preferences that like (Kong 1995). Research in this direction has investi-
they hold, as well as anything that might modulate those gated, for example, the evolution of music styles in space
preferences1 (MacDonald, Hargreaves, and Miell 2002)— and time (Carney 1974), the impact of tourism on shaping
represents an ever-evolving depiction of their cumulative local musical culture (Hebdige 2003; Gibson and Connell
experiences and values. Understandably then, various sci- 2003) and, the effects of migration on altering the musical
entific communities have devoted much attention to resolv- landscape of places (Carney 1998; Baily and Collyer 2006).
ing what determines a person’s musical tastes and, inversely, In much the same way that a person’s musical identity re-
what can be inferred or predicted about someone based on flects important elements of their past and present experi-
their musical tastes. Progress in either direction broadens our ences and values, the musical identity of a place tells the
understanding of the development of individual identity and history of its people.
culture, their rigidity and transmissibility, and the many roles
These studies highlight just a few of the broader cat-
that music plays in shaping our personal and social lives.
egories of research on musical identity. Despite spanning
A common theme in musicology research explores indi- a wide range of ideas and disciplines, a common theme
viduals’ use of music to modulate or express their mood, emerges: musical identity—of individuals and places—is in-
particularly among adolescents (North, Hargreaves, and herently dynamic and ever-changing. These changes happen
O’Neill 2000), and the extent to which personality both both quickly, on the time-scale of our moods, and slowly,
Copyright c 2019, Association for the Advancement of Artificial as the cultural landscape of our environments and music it-
Intelligence (www.aaai.org). All rights reserved. self shifts gradually. However, many studies of individuals’
1
As an example, the experience of being a classically trained musical identity analyze musical taste and its correlates at a
musician affects the music that person is exposed to, how they eval- single point in time. This limitation stems in large part from
uate it, and, ultimately, whether or not they like it. the difficulty of characterizing people’s musical tastes, and

527
tracking their changes over time. In recent years, though, sical tastes during these periods and analytical tools to mea-
more and more people listen to music online, providing a sure the impact of geographic relocation on musical tastes.
detailed digital record of how individuals’ music consump- Changes in listener environment. In 2017–2018, over
tion and tastes evolve over time, all over the world. 32 million Americans relocated to a new residence, with
In this study, we analyze music consumption patterns approximately 4.8 million of those moves crossing state
on Spotify, a popular music streaming platform2 . We focus boundaries (US Census Bureau 2018). Past research has ex-
specifically on the United States, the world’s largest market plored the regional subcultures of individual states, driven
for music (IFPI 2018), and one of the earliest and largest in large part by historical differences in the ethnoreligious
adopters of online music streaming. Coincidentally, the U.S. identities, cultural preferences, and ways of life unique to the
is also one of the most studied locations in musicology re- various groups who settled the United States (Fischer 1991;
search, providing rich context to guide our analyses and the Lieske 1993). In light of these regional differences, we con-
interpretation of their results. We focus on understanding a sider the effects of environment on musical tastes, defining a
key determinant in the development of individuals’ musi- listener’s environment as their state of residence, which we
cal identity: the role of environment in shaping a person’s infer from the person’s most frequent streaming location.
tastes. Specifically, we measure environments’ effects on in- Based on reports from U.S. moving companies (Allied
dividuals’ preferences by treating geographic relocation as Van Lines, Inc. 2017), the majority of state-to-state reloca-
the basis for constructing quasi-natural experiments, using tions happen during the summer months, when the weather
a matched pairs experimental design to mitigate the effects is generally more convenient and most American education
of confounding variables and natural variation. In addition, systems are on break. For this reason, we recorded state-
we investigate the relationships between the age of a listener level relocations between May and September 2017. Corre-
and the music they consume, informing the likely timing of spondingly, we defined a trio of three-month sample peri-
when and where musical identity takes shape. ods: one just before the moving months (P1 : March to May
We begin by describing the primary sources of data used 2017); another immediately following the moving months
in our analyses, most notably individual music consumption (P2 : September to November 2017); and a third, several
histories and location summaries during three sample peri- months later (P3 : December 2017 to February 2018). For
ods. We then outline our approach for characterizing musical each period, we aggregated users’ artist streams and stream-
taste profiles, built up from data-derived music genres, and ing locations over nine randomly selected days. We then
for measuring whether changes in a listener’s environment identified individuals who relocated by noting changes in
induce changes in their musical tastes. We conclude with a their most frequent streaming location between periods P1
discussion of our results and an outlook on the future of mu- and P2 . In later sections, we will compare individuals’ pro-
sical identity research. files during these periods to assess short-term effects of re-
location.
Data and Methods To assess longer-term effects of relocation, we build on
Our study analyzes the music consumption of Spotify users cultural norms in the United States specifying Thanksgiv-
in the United States between December 2016 and February ing and Christmas as travel holidays that are tradition-
2018. After excluding individuals with low activity, miss- ally spent at home with family (Benney et al. 1959). In
ing or invalid demographic information, or unreliable lo- 2017, an estimated 107 million Americans traveled in late-
cation data, our dataset spans the consumption histories of December alone, with about half of those trips exceeding
N =16,445,318 users, called “listeners” throughout. Con- fifty miles (American Automobile Association 2017). Our
sumption histories include, for each listener: (1) daily stream sample frame spans three such “home holidays”: Christ-
totals for each artist, (2) daily stream totals for each song re- mas 2016, Thanksgiving 2017, and Christmas 2017. Using
lease year or vintage, and (3) state-level location data, esti- streaming locations during these holidays (i.e., the five-day
mated from the listener’s streaming IP address. In addition, window centered around the holiday), we inferred plausible
these histories provide limited demographic information, in- past locations for listeners. The results presented here con-
cluding listeners’ self-reported gender (coded as “M”, “F”, sider listeners who spent two or more of these three holidays
and “X”) and an estimate of their self-reported age, aggre- in a state other than their location during P1 and P2 , sug-
gated into 5-year windows (e.g., birth years between 2000– gesting a past move. Qualitatively, our results are unchanged
2004 and birth years between 1975–1979. These two exam- for individuals who traveled for a single holiday. Naturally,
ples represent the youngest and oldest age groups in our this heuristic restricts our analyses to listeners who both ob-
analyses). Aggregating daily histories, we constructed sta- serve and have the means to travel for these holidays (Mallett
tistical profiles that summarize listeners’ musical tastes and 2001). We discuss this limitation further in our conclusions
locations during several sample periods. We begin by de- but proceed with this important caveat in mind.
scribing our motivation behind selecting these time periods Characterizing musical tastes. People tend naturally to
and, from them, formulating quasi-natural experiments. We describe their musical tastes in terms of genres (Rentfrow
then outline our method for characterizing individuals’ mu- and Gosling 2003). This coarse-grained description high-
lights individuals’ general tastes and masks details about
2
In mid-2018, Spotify reported having over 190 million ac- regionally-specific artists within a particular genre. Given
tive users worldwide, including more than 75 million users in the our goal of assessing changes in musical tastes, not predict-
United States alone (Spotify Technology S.A. 2018). ing location, genres provide a suitably abstract characteriza-

528
 0.5 0.7 50

Number of unique genres

Adj. Mutual Information

40

Completeness
0.6
30
0.4
0.5 20
Agglomerative
K-Means 10
0.4
Spectral
0.3 0
100 200 300 400 100 200 300 400 0 100 200 300 400 500
Number of clusters, k Number of clusters, k Number of streams

Figure 1: Clustering metrics suggest natural groupings of Figure 2: Rarefaction curves suggest 200 streams cap-
around 200 artists. Average adjusted mutual information ture the range of a listener’s musical tastes over our
(left) and completeness (right) scores are shown, varying data-derived genres. Rarefaction curves, shown here for a
the number of clusters in three clustering methods. Adjusted sample of 500 listeners (average shown in black), indicate
mutual information peaks, and completeness begins to level that the number of unique genres spanned by each person’s
out at around 200 artists, suggesting a reasonable number of streaming history begins to level out after 200 streams. This
genres for our analyses. Other approaches, including silhou- analysis informed our sampling depth for constructing lis-
ette analysis and information criterion methods (not shown) teners’ taste profiles.
further support this number.

listeners’ musical tastes using K = 200 genres, derived

tion of tastes. But, genres can vary in size and specificity. from the agglomerative clustering results. Past studies
Some studies suggest that there are as few as five dimen- suggest that this level of abstraction may provide more
sions to musical preferences (Rentfrow, Goldberg, and Lev- detail than is often described or even perceived by typical
itin 2011). In contrast, Spotify characterizes music using a listeners (Rentfrow and Gosling 2003). Our characterization
growing list of over 1700 genres and subgenres (Johnston of musical tastes thus implicitly assumes that listeners are
2018), ranging from broad categories like “rock” and “jazz” attuned to differences between these 200 genres, making
to narrow subgenres that distinguish, for example, “metal- our analyses perhaps more sensitive to change than listeners
core” from “power metal” and “bebop” from “hard bop.” themselves. Finally, to name these genres, we selected the
These differences in definitions present a challenge in choos- most common Spotify genre label among the artists in each
ing an appropriate level of categorization. cluster. In few cases, clusters were comprised of artists with
Here, we adopt a data-driven approach, defining genres no associated genre labels (these appear as “UNKNOWN”
as clusters of artists, whose similarity is derived from the in Figure 3).
frequency with which listeners stream two artists in succes- In sum, we characterize an individual’s musical tastes dur-
sion. We focus on the N = 10, 000 most-streamed artists in ing each sample period (e.g., P1 ) by summing together their
the U.S., who collectively account for the vast majority of stream counts for artists in each of the 200 data-derived gen-
all streams in the country. For these artists, we construct a res. This process constructs musical taste profiles as 200-
transition matrix T whose entries Ti,j denote the probability dimensional vectors that we analyze using the methods out-
that a listener streamed a song by artist i then artist j dur- lined below. To ensure that these vectors are representative
ing our sample frame. We then converted T into an N ×N of users’ tastes, we analyzed rarefaction curves (Figure 2) to
distance matrix, D, by computing the pairwise correlation determine the minimum number of streams required to con-
distances between each pair of artist vectors Ti and Tj : struct reliable taste profiles. In the worst case, in which all
genres are equally distinct, our analyses suggest a minimum
(Ti − T̄i ) · (Tj − T¯j )
Di,j = 1 − , (1) of around 200 streams. This limit informed our sampling
||Ti − T̄i ||2 ||Tj − T¯j ||2 depth for each period, ensuring sufficient depth to character-
where T̄i is the mean of the elements of vector Ti , and || · ||2 ize the tastes of nearly all listeners. Repeating this analysis
is the Euclidean norm. using the diversity measures introduced below suggests that
We then evaluated several unsupervised cluster- the range of most users’ tastes can typically be inferred from
ing algorithms—agglomerative, k-means, and spectral many fewer streams.
clustering—to obtain data-driven clusters of similar artists, Measuring how tastes change. Our goal is to quantify
or genres. To determine an appropriate number of clusters, the extent to which an individual’s musical tastes shift in re-
we calculated cluster purity metrics over a varying number sponse to a change in their environment, namely their state
of clusters. Qualitatively, the three clustering techniques of residence. As described above, our construction of musi-
produce similar-scoring partitions of the data, and suggest cal taste profiles characterizes each person’s preferences as a
a natural number of between 150 and 250 genres of mu- distribution of counts over 200 data-derived genres. Changes
sic (Figure 1). Based on this analysis, we characterized in these profiles from one time period to the next can be

529
characterized in two primary directions: (i) changes in how

e
skate ssive post−hardcor
diverse a person’s tastes are, and (ii) changes in what genres

alternative emo

progre e emo (1)

alternativmetalcore
UNKNOW
child rens msleep l(2
child
a person consumes. Measuring changes of this manner is a

stoner metal
power metal

e nime l pop (2)

rens usic )

emo

(1)
anthempunk
desi usic (2 )

classifypop

folk punk

s
indianp hop

lue
rock
taiw

arab pop

melodic

o
clas mizra

celticamo
m

antitiviral p
N (11)

metal
hi

lb
UN
central focus in ecological research, which frequently char-

g ou ec ctro
s dn ele e
ane k− pop5)

a
sica hpi

re tro av

s N oit l sou
kor OW bea(2)p

scre
(2)

an t
KN

ir
vapicore

re eggpic en rn g N ( p
(2)

)
djen

gg a al es os 7)

1
alt U

re porw
v

r ro k e W ho

no l
se pop

pa pe
l(
e

deospther k
(3

U tr e n
er NK

an ( ts

(5)

t oc th O ip
po

(1 )
na N te

w (2
a
gli chilip h (3) p

r ou KN h
acterizes the diversity within (called “α-diversity”) or be-

edch hlhop p

)
tiv O en o

flo w
t
m o

n flo
e W p pm
h N o
a ario

to n
an

ae eto
th a ioc it
pr em wcap ar ers

o
tween (“β-diversity”) ecosystems (Whittaker 1972), based og
re
ss w orspell
ive orsh hip a
a
at to k c niv
ch na un u tive 2)
ba alleep f nejo rna an ( (1)
v e r ta lte isti an
po tr ac ouccmip
h d e a hr ti
on counts of species’ abundances and measures of their c
dir tyhristia
ly
eh po e
all p
e
da nesia tra stepse
nc n nc
s tin c ris 2)
la tin ch a (
la tin er
la rup bia onid
era
(2)
g m ia s an
sou hip soc(2)
n cuumb o mexic an (1) p (2)
phylogenetic similarity. Here, we draw inspiration from the th ra ho a
clas otacp (1)
sic ore
p c jan nal exic n po
te gio al m xica
re gion al me nol (2) p (1)
comal (1) re gion espa an po
ecology literature for measuring changes within and be- texas counetrdy
new am
tradition
coun y
ericana try
re ck en mexic
ro gional )
re era (1 )
grup pop (2
al coun latin hera
tween listeners’ profiles. latin poplatin
underground
try
(1)
rap
ranc
deep lat
in hip ho
p
grupera (3)bia sonidera
deep cum e
reggae duranguens
To measure the range or diversity of a person’s musi- reggae rock (2)
broadway
adult standards (1)
deep norteno
banda
focus (3)
focus (1)
classic rock
cal tastes within a given time period, we used Rao-Stirling motopop
new wavee (2)
ng
post−gru rd rock
wn christian
focus (4) relaxative (1)
compos
focus itional ambie
ha focu (5) nt (1)
metal
divergence (Rao 1982; Stirling 2007). This technique is a al te rn ativenge (1))
ru
post−gae rock
(1
r rap
co s (6)
UNmpositio
UNKKNOW nal ambi
m NO N (4 ent (2
reggangsteip hop heaeditatioWN (6 ) )
popular measure of biodiversity and is closely related to g

o ra
h onicy
tr tr
eleccoun (1)
ry edm (3)
v lin
c iral p g
fo ham op
h lk− ber
n (2 )
)

tem
p m ic
ed musary c aw pop pop
other approaches used throughout ecological research (Mar- con p r )
tra mpo b (1se
nte r& ou op
co die l h p a
fochristiaiian
c u a
o hil s n
s pe dr (2) relax
c co ra en ati
an in pica ronic(1)p s la re tic s m
tin 2002; Lozupone and Knight 2008). It has also recently urb tr o a
iet WNp r (1)
indNO po op
sc cor ssic cor po usic
or ec al e ( p (1)
ve
(2)

ne ke sif al e ( )
rdcdan p fr rn ie foock k
ec or (2 2)

tra od ind lk r pun

fa las ssicativ l (3
w
K h or e )

c la x ica
ce an ro lk
e (1)

ag
UN

c ela ss 1) (2) )
ou c re ha ca ck
hip
been applied to the study of musical diversity by Park et al.

fo p
(3

nd hic hip ll is

r la e ( tive n (1
po

e
dir west hip ano ho(1)

c n xa tio
d )

c ho r p

toela ita )
n

y (4 1)
UN southoast p (2ap

r ed (2 enta
u

KN ra tra )

m e nm

(1 )
ro

rou azo OWNp (2 p

tonnviro (1) ental (1)

rg

)
nd nto (2 )

e leep nm ntal )
hip be )

s nviro nme (10

de

e viro OWN (8)

UNK die r&op (3ts
(Park et al. 2015), who highlighted the measure’s advantage

NOW b (2 )

enNKN WN
N )

U
un

th ja(9)

UNvi
zz

en vironmen
en
vocarul mba
bossa jazz

jam ba−rock
o

nova

blues
cool jaz z

indie po
rock−and−ro z

indie garage
ll

funk
(2)

new wave
adult standards (3)
indie garage rock (2) trip hop
jaz

K
h

ronm tal (4
N
ha

adult standards

O
over existing approaches, namely that some genres can be

smoo

nd
ty

l (5
rgr

in

en
p
de

)
(3)
ta l (2)
un
very similar to others, biasing approaches that count unique

erg

rock (1)

)
und
genres or otherwise ignore their similarity. Given a taste pro-
file p, constructed as a probability distribution over our data-
derived genres, Rao-Stirling divergence is calculated as
Figure 3: Dendrogram showing relatedness of the 200
X data-derived genres. Constructed using the UPGMA al-
dRS (p) = pi × pj × d(i, j), (2) gorithm and genre-genre correlation distances, this dendro-
i,j∈K gram serves as the basis for our UniFrac-based compar-
where pi and pj denote the fraction of streams from gen- isons of musical taste profiles. This clustering captures many
res i and j, respectively, and d(i, j) denotes the dissimi- known relationships between genres. Notably, the largest
larity of the two genres. To quantify the dissimilarity be- distinction in the tree, shown near 2 o’clock, splits musical
tween our data-derived genres, we measured the number of genres by the language of their lyrics. Genres titled “UN-
times listeners consumed genres i and j in P1 , forming a KNOWN” represent groups of artists with no specified genre
co-consumption matrix of genres. We then computed d(i, j) labels on Spotify. Numbered genres differentiate individual
as correlation distances, comparing the rows of the resulting clusters having the same most-common Spotify genre label.
matrix (similar to Equation 1).
To measure the difference between two taste profiles, we
used UniFrac, another approach adopted from the ecology vergence. Given two probability distributions a and b, dKL
literature. UniFrac is a family of distance metrics used to and dJS are defined as
assess the dissimilarity of two ecosystems that, like Rao-
Stirling, takes into account the similarity of the counted ele- X a(x)
ments (Lozupone and Knight 2005). These metrics are con- dKL (a||b) = a(x)log2
b(x)
structed using a phylogenetic tree that summarizes the evo- x∈X
lutionary distances separating the species or, in our case, the 1 1
dJS (a||b) = dKL (a||c) + dKL (b||c),
distances between data-derived genres. We used the same 2 2
genre correlation distances (d(i, j)) as in our Rao-Stirling 1
calculations to construct such a tree of genres, using hierar- where c = (a + b).
2
chical clustering (UPGMA algorithm (Sokal 1958); result-
ing tree shown in Figure 3). Distances between taste profiles
were then calculated based on the the amount of distinct ver- Qualitatively, our findings were not sensitive to the choice of
sus shared branch length spanned by the two profiles. In our weighted UniFrac or Jensen-Shannon divergence. As such,
analyses, we used weighted UniFrac (dW U ), a variant that we present just the weighted UniFrac-based results below.
considers not just which genres are consumed but in what Inferring state-level musical identities. Finally, in sev-
proportions (i.e., the abundance of each lineage). eral of our analyses, we test whether individuals’ musical
For completeness, we repeated our UniFrac-based analy- preferences change in response to relocating from one state
ses using Jensen-Shannon divergence, a more common dis- to another. To ground these measurements, we constructed
similarity measure that incorporates no information about state-level musical taste profiles by summing together the
the relatedness of genres. Jensen-Shannon divergence (dJS ) streams from all listeners from each state over our 200 data-
is related to the popular Kullback–Leibler (dKL ) divergence derived genres. These state-level profiles enable more con-
in that dJS is an averaged, symmetrized of measure dKL di- crete definitions of change for individuals. For instance, if a

530
 A Texas Country B Gospel, Soul C West Coast Trap, Hyphy

D Ranchera, Mariachi E Hawaiian, Ukulele F Jazz, Bebop -2 -1 0 +1 +2

Streams from cluster (z-score)

Figure 4: Data-derived genres encode regional information. These heatmaps show the fraction of each state’s streams coming
from six data-derived genres, displayed using z-scores. The maps highlight different elements of the states’ unique musical
identities and provide a useful check to ensure that the genres capture patterns that should be expected historically.

person moves from state a to state b, do their tastes become does not correlate with the entropy of U.S. Census-reported
more similar to the aggregate profile for state b? Or less sim- distributions for race/ethnicity. However, there is a moderate
ilar to that of a? Additionally, these state-level characteriza- correlation (Pearson’s ρ = 0.49, p < 0.001) between Rao-
tions provide a valuable sanity check to verify that the states Stirling diversity and states’ Hispanic composition (Kaiser
do, in fact, possess distinct musical tastes that might influ- Family Foundation 2018). This correlation is likely due in
ence the tastes of individuals. Qualitatively, we found that part to large disparities in the co-consumption of genres that
consumption patterns for many data-derived genres matched differ in the primary languages of their lyrics (see Figure 3),
intuitions based on the history of the states and their ethnic, which contributes to higher levels of measured diversity.
religious, and cultural compositions (Figure 4). While states vary in the diversity of their aggregate com-
Analyzing the general diversity of these state-level taste positions, states exhibit similar distributions of diversity cal-
profiles using Rao-Stirling divergence, we found that states culated at the level of individuals (p > 0.05, Conover post-
vary in their aggregate diversity (Figure 5). This quantity hoc test for multiple comparisons). Together, these two ob-
servations suggest that the higher diversity of some state
taste profiles is driven by diverse compositions of individ-
uals, not because the individuals in those states have more
Distributions of diversity MA
diverse tastes themselves.
Distribution of diversity
in individual taste profiles
in state taste profiles
for each state
Results
Density

WV We devised two sets of matched pair analyses to study the

CO
short- and long-term effects of relocation on individuals’
IL
musical tastes. We begin by analyzing short- then long-term
TX
effects, followed by an examination of the relationship be-
0.00 0.02 0.04 0.06 0.08 0.10 tween the ages of listeners and the music they consume.
Musical diversity (Rao-Stirling) Short-term effects of relocation. We tested whether
moving from one state (a) to another (b) induces short-
term changes in individuals’ musical tastes by construct-
Figure 5: Musical diversity of individuals follows a sim- ing matched pairs of listeners. Each mover m in our sample
ilar distribution across the states, even though state- frame was paired to a non-mover n by exact matching under
level diversity varies. Multiple distributions are depicted. the following criteria:
In green, we show the distributions of musical taste diver-
1. Both m and n lived in state a during P1
sity (Rao-Stirling) for individuals within all fifty states. In
black, we show the distributions of musical taste diversity of 2. m moved to state b between P1 and P2
the aggregate state profiles. 3. n continued living in a during P2

531
4. m and n share the same reported gender and age group

Fraction of matched pairs

These criteria formed a total of N = 592, 716 matched 0.15
pairs of listeners. First, we tested whether movers exhibit
a systematic change in the overall diversity of their musical Mover becomes Mover becomes
0.10 more similar to less similar to
tastes, measured using Rao-Stirling divergence. Specifically, past home past home
we measured m’s change in diversity from before (P1 ) to (49.9%) (50.1%)
after their relocation (P2 ) as dRS (mP2 ) − dRS (mP1 ). We 0.05
then compared this change in diversity to the same quan-
tity calculated for m’s matched pair individual, n. Sampling
1000 matched pairs from each state, we found no significant 0.00
change in movers’ overall taste diversity compared to non- -3 -2 -1 0 1 2 3
Relative difference in dissimilarity (weighted UniFrac)
movers, neither between P1 and P2 (p = 0.72, matched-pair
t-test) nor P1 and P3 (p=0.24).
Next, we tested whether movers’ taste profiles shift in re- Figure 6: Short-term effects of relocation on musi-
sponse to relocation, possibly becoming less similar to their cal tastes are insignificant, within the range of typical
former home and more similar to their new home. First, we month-to-month variability. The distribution of changes in
calculated the difference in dissimilarities between m and dissimilarity between movers (m) and their previous home
a’s taste profiles, during P1 and P2 , state (a) compared to non-movers. Changes here are scaled
D(m, a|P2 , P1 ) = dW U (mP2 , aP2 )−dW U (mP1 , aP1 ). (3) by the amount of expected variability from within-person,
month-to-month fluctuations in musical tastes (see main
This quantity captures whether m is more similar to a during text). As shown, no systematic changes were found across
time period P1 or P2 . Next, we calculated the same quantity our sample, and individuals’ changes were predominantly
for n to compare, within the range of typical variability.
D(n, a|P2 , P1 ) = dW U (nP2 , aP2 )−dW U (nP1 , aP1 ). (4)
The difference between these two quantities,
measured changes relative to another non-mover, rather than
D(m, a|P2 , P1 ) and D(n, a|P2 , P1 ), captures whether
aggregate state taste profiles, and found similar outcomes.
m’s dissimilarity to their former home state a increases or
Lastly, we added another matching criterion to the list above,
decreases after moving to state b, compared to their matched
requiring that matched pairs share the same “favorite” (i.e.,
pair, who remained in state a. Sampling 1000 matched pairs
most consumed) genre during P1 . Each of these modifica-
from each state, we found no significant differences in the
tions only served to corroborate the observations above.
similarity of states’ musical taste profiles to individuals
Our results here indicate that relocation has little effect
who live in versus moved away from that state (p = 0.70,
on individuals’ taste profiles in the months following their
matched-pair t-test). That is, moving to a new state does
move. While no detectable changes were found, our mea-
not, in the short term, appear to have a significant effect on
surements cannot rule out the possibility that differences
individuals’ musical tastes, neither making them more nor
may become significant over longer periods of time. Accord-
less like their former home.
ingly, in the next section, we outline a similar matched pair
Dividing these differences by the month-to-month vari-
analysis to reason about how relocation affects taste profiles
ability of all non-movers (i.e., standard deviation of
in the long term.
D(n, a|P2 , P1 ) − D(n, a|P3 , P2 )), we note that, despite
some amount of heterogeneity in the magnitude and size Long-term effects of relocation. We tested whether re-
of individuals’ changes, the observed differences generally location induces long-term changes in individuals’ musical
fall within one standard deviation of typical month-to-month tastes by constructing matched pairs of listeners based on
fluctuations in a person’s tastes (Figure 6). their locations during the three home holidays spanned by
Next, replacing a with b in Equations 3 and 4, we tested our sample frame. Listeners who traveled to a different state
whether individuals’ tastes become more or less similar to for at least two of these holidays were assumed to have for-
their new state (b) after relocating there. As before, we sam- merly resided in that state, having since moved to their cur-
pled 1000 matched pairs from each state and found no sig- rent state. As in the analysis above, we paired each mover,
nificant differences between movers and non-movers (p = m, with a non-mover n—someone who spent the home hol-
0.77, matched-pair t-test). In the short term, moving to a idays in their current home state—by applying the following
new state does not seem to move an individual’s musical criteria:
taste profile closer to their new state’s. 1. m traveled to and is assumed to have lived in state a
The results of these comparisons hold over slightly longer
periods of time (i.e. comparing taste profiles between P1 and 2. n lives in and spent the holidays in state a
P3 , rather than P1 and P2 ), despite a trend towards signifi- 3. m now lives in state b
cance. They are also robust across different age groups and 4. m and n share the same reported gender and age group
genders, as well as isolating the effects for particular choices
of a or b, and using more narrowly-defined state taste pro- These criteria formed a total of N = 469, 935 matched
files, constructed for each age-gender group. In addition, we pairs. First, we tested whether movers differ from non-

532
 18 (y=2000)

28 (y=1990)
38 (y=1980)
48 (y=1970)
0.20
Fraction of matched pairs

10-1

Fraction of all streams

0.15 Mover is Mover is
more similar to more similar to
10-2 Age group
present home past home
0.10 (57.5%) (42.5%) 1975–1979
1980–1984
0.05 10-3 1985–1989
1990–1994
1995–1999
0.00 10-4 2000–2004
-3 -2 -1 0 1 2 3
Relative difference in dissimilarity (weighted UniFrac) 100 101 102
Song age in 2018

Figure 7: Long-term effects of relocation on musical

tastes indicate that listeners’ tastes shift towards their Figure 8: Listeners of all ages predominantly stream cur-
current environments’, if only somewhat. Here, negative rent music. Distributions of song age (i.e., 2018 minus re-
values indicate that m is more similar to their present home lease year) consumed by our six age groups. Shown on log-
state (b) than their past home state (a), relative to n. Changes log axes, two patterns emerge: (1) regardless of their age,
are normalized by the amount of expected variability in non- listeners generally consume new or recently-released music,
movers’ taste profiles. and (2) listener age correlates with song age.

movers in the general diversity of their musical tastes, mea- shift only slightly, if at all, towards their present home state.
sured by Rao-Stirling divergence for aggregate profiles of Unlike our first analysis, here we lack information about
listeners’ streams from periods P1 through P3 . Sampling when listeners may have moved away from their previous
1000 matched pairs from each state, we found that movers home state and therefore how these effects may develop over
and non-movers had similarly diverse musical taste profiles time. Nevertheless, we find that both past and present envi-
(p = 0.41, Mann-Whitney). This result is unintuitive given ronments do appear to shape listener preferences in the long
that exposure to a wider variety of environments could plau- term. To gain a better understanding of when listener en-
sibly instill more varied musical tastes. To ensure that this re- vironments likely impact musical tastes, we now shift our
sult was not driven by moves to neighboring states with sim- focus towards the relationship between the ages of listeners
ilar cultures, we repeated this test, omitting moves between and the music they consume.
states that share a border. Under this restriction, movers Timing of environmental influence. In the previous sec-
exhibit only marginally higher diversity than non-movers tion, we found that both past and present environments play
(0.1% higher; p < 0.001, matched-pair t-test). a role in shaping listeners’ musical tastes. Here, we test
Next, using these same aggregate profiles, we determined when and therefore which environments are likely to affect
whether movers’ tastes are more similar to their inferred past these tastes by analyzing how listener age predicts the age
(a) or present (b) home states, relative to non-movers. Said of the music they consume. Specifically, we look for indica-
differently: do movers’ tastes shift to reflect their new envi- tions of when musical tastes are formed in order to inform
ronment? To test this possibility, we used weighted UniFrac when environment is most likely to have an impact.
to compute the difference in dissimilarity between m and the First, we analyzed the general relationship between the
two states as age of listeners and their music (Figure 8). We found that
D(m, a, b) = dW U (m, b) − dW U (m, a). (5) listeners of all ages consume predominantly current music,
with 28% of all streams coming from songs that are less than
Again drawing a sample of 1000 matched pairs from each a year old. This observation applies to all age groups, though
state, we found that the distribution of D(m, a, b) skews pos- older listeners are more likely to consume older music.
itive (p < 0.001, t-test), with 64% of movers being more sim- While all listeners tend to consume music that has been
ilar to their past home a than their current home b. Next, we released recently, music of a given age may be more or less
calculated a similar quantity (D(n, a, b)) for m’s matched likely to be consumed by listeners of different ages. We an-
pair, n, and considered the difference between the result- alyzed the relationship between a song’s release year and
ing quantities. We found that movers are significantly more the age of its current listeners when the song was released
similar to their present home state than their past home state, by calculating the distribution of streams by listener age for
compared to their matched pair (p < 0.001, matched-pair t- each song release year. We then calculated z-scores for the
test; see Figure 7). Here, 57.5% of movers were more similar fractions of streams from each listener age. Applying this
to their current state. transformation highlights an affinity in listeners for music
Together these findings paint a nuanced picture of the that was released when they were 10–20 years old (Fig-
long-term effects of relocation: on average, listeners’ mu- ure 9).
sical tastes continue to resemble their past home states’, and This pattern corroborates past studies (Stephens-

533
 2.0 the popularity distribution are what make listeners’ tastes
1955 unique (Anderson 2006; Goel et al. 2010). Our decision to

Streams by age (z-score)

1.5
not re-weight taste profiles was made consciously, under the
Song release year

1965 1.0
1975 0.5 assumption that listeners themselves would define their mu-
0.0 sical tastes based on how often they listen to each genre,
1985
-0.5
not how unique those genres are. This assumption should
1995 itself be explored by future studies in order to better under-
-1.0
2005 stand how people describe their interests and how descrip-
-1.5
2015 tions vary depending on social context and the curated ver-
-2.0
-30 -20 -10 0 10 20 30 40 50 sion of self the person wishes to portray.
Age of listener when song was released Second, our classification of environment as states of resi-
dence masks a large amount of cultural variation within most
states, adding noise to our study and potentially hiding sub-
Figure 9: The distribution of song vs. listener age high- tle patterns. This limitation is particularly true for states with
lights the importance of adolescence in musical identity both sizable urban and rural populations. For example, the
formation. For each song release year, the distribution of cultural differences between moving to Manhattan versus a
current listener ages when the song was released. Negative small town in upstate New York can be substantial. We at-
ages indicate listeners streaming music that was released be- tempted to mitigate this effect in our matched pair analyses
fore they were born. Two age regions were excluded due to by incorporating additional matching criteria (e.g., requir-
low representation of older listeners (upper right) and young ing that matches share favorite genres), but future studies
or unborn listeners (bottom left). should consider investigating these sub-state cultural varia-
tions more directly. For instance, what is the impact of mov-
ing from an urban to a rural environment, or vice versa? And,
Davidowitz 2018; Schwartz and Fouts 2003) that similarly are there personal attributes that predict the malleability of
found adolescence to be a crucial period in the development someone’s tastes?
of musical taste and identity. In the context of our other
Lastly, individuals who use Spotify or similar services
results—that musical tastes are generally robust to change
may have more robust musical tastes than others. For one,
but reflect our past locations—this pattern implies that
these listeners may have a higher baseline awareness of or
it is both the timing and geographic location of person’s
interest in music. Perhaps more importantly, however, the
adolescence that casts their musical identity.
success of music streaming platforms stems in part from
their ability to give listeners access their music at any time,
Discussion anywhere. Paired with increasingly prevalent mobile phone
In this study, we used a comprehensive data set on music technology, these platforms have ushered in a new era of ac-
consumption in the United States to measure the impact of cessibility in music and have accelerated a transition from
geographic relocation on individuals’ musical taste profiles. what has traditionally been a “push” model of music con-
Analyzing short-term effects, we found that listeners’ mu- sumption (i.e., radio stations decide what gets played) to
sical tastes are robust to changes in environment, both in more of a “pull” model (i.e., individuals decide for them-
terms of their overall diversity as well as in composition. selves what to play), giving listeners more control over what
Over longer periods of time, relocation does appear to shift they listen to, when, and how. This increased level of in-
individuals’ tastes marginally towards those of their new en- dividualized play control may contribute to the portability
vironment. The size of this effect, however, is small, and lis- and thus robustness of tastes observed here. That is, had this
teners’ tend to more strongly resemble their past rather than analysis been somehow possible 30 years ago, we might ex-
present environments. Finally, listeners’ affinities for music pect to see greater malleability of musical tastes because lis-
released during their adolescence suggests that a person’s teners’ choice was more directly constrained by what was lo-
musical environment during this period ultimately shapes cally available after a move. Moreover, the effect of this shift
their musical identity. in control on the formation of musical tastes represents an-
Our results indicate that musical tastes, characterized here other interesting direction for future research, to make sense
as distributions over 200 data-derived genres, are largely ro- of preferences in light of nearly limitless choice and con-
bust to relocation from one state to another in the U.S. There trol (Vanderbilt 2016).
are at least three factors that might help explain this obser- Our study focuses on Spotify users in the U.S. in particu-
vation. First, our analysis studies the changes in listeners’ lar, a population that skews towards younger and, by virtue
consumption of general styles of music, as captured by the of having access to the Internet and services like Spotify,
10,000 most popular artists in the United States. Naturally, likely wealthier individuals. This limitation is compounded
the popularity of these artists (and the genres to which they in our analysis of the long-term effects of relocation, in
belong) varies tremendously, with the most popular artists which we assume past environments are suggested by holi-
receiving orders of magnitude more streams than the least day travel patterns. We acknowledge that these assumptions
popular artists. We made no effort to re-weight or otherwise likely exclude people of lower socioeconomic status, and
adjust for common tastes so as to amplify any differences, may introduce some noise, for example, from individuals
yet it may well be the case that artists on the long tail of who regularly travel during the holidays but to a location

534
other than their former home. Other, more precise methods Carney, G. O. 1974. Bluegrass grows all around: The spatial
for inferring location histories may be possible and would dimensions of a country music style. Journal of Geography
enable more nuanced analyses in the future. Nevertheless, 73(4):34–55.
listeners of higher socioeconomic status are generally re- Carney, G. 1998. Music geography. Journal of Cultural
garded as “cultural omnivores” (Peterson and Kern 1996; Geography 18(1):1–10.
Park et al. 2015), which may make their tastes more mal-
leable than others and lead us to underestimate the rigidity Cohen, S. 1991. Rock culture in Liverpool: Popular music
of musical preferences in the larger population. in the making. Oxford University Press on Demand.
Selection biases also complicate the idea of carrying out DeNora, T. 2000. Music in everyday life. Cambridge Uni-
similar analyses for international migration: relocating be- versity Press.
tween countries is expensive, both financially and socially, Erickson, B. H. 1996. Culture, class, and connections.
and people who do may not be wholly representative of American Journal of Sociology 102(1):217–251.
those with more restricted mobility. However, overcoming
Fischer, D. H. 1991. Albion’s seed: Four British folkways in
this limitation would offer valuable insight into the robust-
America. Oxford University Press.
ness and transmissibility of international culture. For exam-
ple, what aspects of a culture (e.g., Hofstede’s cultural di- Gibson, C., and Connell, J. 2003. Bongo Fury: tourism, mu-
mensions (Hofstede 1984)) make it more or less resistant sic and cultural economy at Byron Bay, Australia. Tijdschrift
to change? To what extent do migrants adopt the culture of voor Economische en Sociale Geografie 94(2):164–187.
their new environment, and at what rate? And, could any of Goel, S.; Broder, A.; Gabrilovich, E.; and Pang, B. 2010.
these variables be predicted beforehand? As access to ser- Anatomy of the long tail: ordinary people with extraordi-
vices like Spotify increases in the future, answering such nary tastes. In Proceedings of the third ACM international
questions may become possible. conference on Web search and data mining, 201–210. ACM.
As people increasingly discover, consume, and share mu- Hebdige, D. 2003. Cutn’Mix: Culture, Identity and
sic through online platforms, the field of musicology is Caribbean Music. Routledge.
uniquely poised to produce new insights on the development
of tastes and identity, their determinants, and their interac- Hofstede, G. 1984. Cultural dimensions in management and
tion with surrounding communities and cultures. In addition planning. Asia Pacific Journal of Management 1(2):81–99.
to providing insights into these topics, further research in Hudson, R. 2006. Regions and place: music, identity and
this direction may enhance the way people experience mu- place. Progress in Human Geography 30(5):626.
sic through these platforms, while also ensuring that services IFPI. 2018. International federation of the phonographic
are mindful of their potential impacts on the development industry digital music report 2018.
of individual identity and on the evolution of culture more
Johnston, M. 2018. How spotify discovers the genres of
broadly.
tomorrow. Spotify for Artists Blog.
Acknowledgments Kaiser Family Foundation. 2018. Population distri-
bution by race/ethnicity. https://www.kff.org/other/state-
We thank Manish Nag, Nathan Stein, Scott Wolf, Rozmin indicator/distribution-by-raceethnicity/.
Daya, Clay Gibson, Will Shapiro, Glenn McDonald, Ri-
cardo Monti, Laura Norris, Daniel Larremore, Abigail Ja- Kong, L. 1995. Popular music in geographical analyses.
cobs, Allison Morgan, and Herrissa Lamothe for helpful Progress in Human Geography 19(2):183–198.
conversations. Lena, J. C. 2012. Banding together: How communities cre-
ate genres in popular music. Princeton University Press.
References Lieske, J. 1993. Regional subcultures of the united states.
Allied Van Lines, Inc. 2017. Why summer is the The Journal of Politics 55(4):888–913.
peak moving season. https://www.allied.com/blog/view/all- Lozupone, C., and Knight, R. 2005. UniFrac: a new phylo-
blogs/2017/06/20/why-summer-is-the-peak-moving-season. genetic method for comparing microbial communities. Ap-
American Automobile Association. 2017. 2017 year-end plied and Environmental Microbiology 71(12):8228–8235.
holiday travel forecast review. Lozupone, C. A., and Knight, R. 2008. Species divergence
Anderson, C. 2006. The long tail: Why the future of business and the measurement of microbial diversity. FEMS Micro-
is selling more for less. Hyperion. biology Reviews 32(4):557–578.
Baily, J., and Collyer, M. 2006. Introduction: Music and mi- MacDonald, R.; Hargreaves, D.; and Miell, D., eds. 2002.
gration. Journal of Ethnic and Migration Studies 32(2):167– Musical identities, volume 2. Oxford University Press.
182. Mallett, W. J. 2001. Long-distance travel by low-income
Bennett, A., and Peterson, R. A. 2004. Music scenes: local, households. TRB Transportation Research Circular E-
translocal and virtual. Vanderbilt University Press. C026—Personal Travel: The Long and Short of It 169–177.
Benney, M.; Weiss, R.; Meyersohn, R.; and Riesman, D. Martin, A. P. 2002. Phylogenetic approaches for describ-
1959. Christmas in an apartment hotel. American Journal ing and comparing the diversity of microbial communities.
of Sociology 65(3):233–240. Applied and Environmental Microbiology 68(8):3673–3682.

535
Nash, P. H., and Carney, G. O. 1996. The seven themes of Schwartz, K. D., and Fouts, G. T. 2003. Music Preferences,
music geography. Canadian Geographer 40(1):69–74. Personality Style, and Developmental Issues of Adolescents.
North, A. C.; Hargreaves, D. J.; and O’Neill, S. A. 2000. Journal of Youth and Adolescence 32(3):205–213.
The importance of music to adolescents. British Journal of Sloboda, J. A., and O’neill, S. A. 2001. Emotions in ev-
Educational Psychology 70(2):255–272. eryday listening to music. Music and emotion: Theory and
Park, M.; Weber, I.; Naaman, M.; and Vieweg, S. 2015. research 415–429.
Understanding musical diversity via online social media. In Sokal, R. R. 1958. A statistical method for evaluating sys-
Ninth International Conference on Web and Social Media. tematic relationship. University of Kansas science bulletin
Peterson, R. A., and Kern, R. M. 1996. Changing high- 28:1409–1438.
brow taste: From snob to omnivore. American Sociological Spotify Technology S.A. 2018. Financial results for thrid
Review 900–907. quarter 2018.
Rao, C. R. 1982. Diversity and dissimilarity coefficients: a Stephens-Davidowitz, S. 2018. The songs that bind. https://
unified approach. Theoretical Population Biology 21(1):24– www.nytimes.com/2018/02/10/opinion/sunday/favorite-
43. songs.html.
Rentfrow, P. J., and Gosling, S. D. 2003. The do re mi’s Stirling, A. 2007. A general framework for analysing diver-
of everyday life: the structure and personality correlates of sity in science, technology and society. Journal of the Royal
music preferences. Journal of Personality and Social Psy- Society Interface 4(15):707–719.
chology 84(6):1236. US Census Bureau. 2018. CPS historical migra-
Rentfrow, P. J.; Goldberg, L. R.; and Levitin, D. J. 2011. tion/geographic mobility tables.
The structure of musical preferences: a five-factor model. Vanderbilt, T. 2016. You May Also Like: Taste in an Age of
Journal of Personality and Social Psychology 100(6):1139– Endless Choice. Simon and Schuster.
1157.
Wells, A., and Hakanen, E. 1991. The Emotional Use
Saarikallio, S., and Erkkilä, J. 2007. The role of mu- of Popular-Music by Adolescents. Journalism Quarterly
sic in adolescents’ mood regulation. Psychology of Music 68(3):445–454.
35(1):88–109.
Whittaker, R. H. 1972. Evolution and measurement of
Schäfer, T., and Mehlhorn, C. 2017. Can personality traits species diversity. Taxon 213–251.
predict musical style preferences? A meta-analysis. Person-
ality and Individual Differences 116:265 – 273.

536