ARTICLE OPEN
The associa ion be ween u ban land use and dep essi e
symp oms in young adul hood: a FinnTwin12 coho s udy
Zhiyang Wang
1
, Alyce M. Whipp
1,2
, Ma ja Heinonen-Guzeje
2
, Ma ia Fo as e
3,4,5,6
, Jo di Júl ez
4,7
and Jaakko Kap io
1,2
✉
© The Au ho (s) 2023
BACKGROUND: Dep essi e symp oms lead o a se ious public heal h bu den and a e conside ably a ec ed by he en i onmen .
Land use, desc ibing he u ban li ing en i onmen , influences men al heal h, bu complex ela ionship assessmen is a e.
OBJECTIVE: We aimed o examine he complica ed associa ion be ween u ban land use and dep essi e symp oms among young
adul s wi h di e en ial land use en i onmen s, by applying mul iple models.
METHODS: We included 1804 indi idual wins om he FinnTwin12 coho , li ing in u ban a eas in 2012. The e we e eigh ypes o
land use exposu es in h ee bu e adii. The dep essi e symp oms we e assessed h ough he Gene al Beha io In en o y (GBI) in
young adul hood (mean age: 24.1). Fi s , K-means clus e ing was pe o med o dis inguish pa icipan s wi h di e en ial land use
en i onmen s. Then, linea elas ic ne penalized eg ession and eX eme G adien Boos ing (XGBoos ) we e used o educe
dimensions o p io i ize o impo ance and examine he linea and nonlinea ela ionships.
RESULTS: Two clus e s we e iden ified: one is mo e ypical o ci y cen e s and ano he o subu ban a eas. A he e ogeneous pa e n
in esul s was de ec ed om he linea elas ic ne penalized eg ession model among he o e all sample and he wo sepa a ed
clus e s. Ag icul u al esiden ial land use in a 100 m bu e con ibu ed o GBI mos (coe ficien : 0.097) in he “subu ban”clus e
among 11 selec ed exposu es a e adjus men wi h demog aphic co a ia es. In he “ci y cen e ”clus e , none o he land use
exposu es was associa ed wi h GBI, e en a e u he adjus men wi h social indica o s. F om he XGBoos models, we obse ed
ha anks o he impo ance o land use exposu es on GBI and hei nonlinea ela ionships a e also he e ogeneous in he wo
clus e s.
IMPACT:
●This s udy examined he complex ela ionship be ween u ban land use and dep essi e symp oms among young adul s in
Finland. Based on he FinnTwin12 coho , wo dis inc clus e s o pa icipan s we e iden ified wi h di e en u ban land use
en i onmen s a fi s . We hen employed wo plu alis ic models, elas ic ne penalized eg ession and XGBoos , and e ealed
bo h linea and nonlinea ela ionships be ween u ban land use and dep essi e symp oms, which also a ied in he wo
clus e s. The findings sugges ha analyses, in ol ing land use and he b oade en i onmen al p ofile, should conside aspec s
such as popula ion he e ogenei y and linea i y o comp ehensi e assessmen in he u u e.
Keywo ds: Land use; Dep essi e symp oms
Jou nal o Exposu e Science & En i onmen al Epidemiology (2024) 34:770–779; h ps://doi.o g/10.1038/s41370-023-00619-w
INTRODUCTION
Dep essi e symp oms a e e y common and eflec a ch onic,
complex, and mul i ac o ial men al heal h condi ion. The bu den
o dep essi e symp oms is g owing, especially among younge
people. The e has been a la ge ise in he incidence o dep essi e
episodes o diso de s among young adul s ac oss mul iple
coun ies [1–3]. The COVID-19 pandemic induced a nega i e
men al heal h impac and inc eased he p e alence o dep essi e
symp oms among young adul s [4,5]. Mo eo e , dep essi e
symp oms we e associa ed wi h a highe odds o isky beha io
such as subs ance use and sel -ha m, which esul ed in u he
psychological and physical heal h p oblems [6]. Al hough he e is
a gene ic p edisposi ion o occu mo e dep essi e symp oms,
which a me a-analysis in 2020 es ima ed a he i abili y o 37% [7],
se e al win s udies ac oss coun ies ha e iden ified he i al ole
o en i onmen al influences on men al heal h, including dep es-
si e symp oms among young adul s, inspi ing e iological con-
side a ion o a ious en i onmen s [8,9].
Recei ed: 27 Ap il 2023 Re ised: 20 No embe 2023 Accep ed: 22 No embe 2023
Published online: 11 Decembe 2023
1
Ins i u e o Molecula Medicine Finland, Helsinki Ins i u e o Li e Science, Uni e si y o Helsinki, Helsinki, Finland.
2
Depa men o Public Heal h, Uni e si y o Helsinki, Helsinki,
Finland.
3
PHAGEX Resea ch G oup, Blanque na School o Heal h Science, Uni e si a Ramon Llull (URL), Ba celona, Spain.
4
ISGlobal-Ins i u o de Salud Global de Ba celona Campus
MAR, Pa c de Rece ca Biomèdica de Ba celona (PRBB), Ba celona, Spain.
5
Uni e si a Pompeu Fab a (UPF), Ba celona, Spain.
6
CIBER Epidemiología y Salud Pública (CIBEREsp),
Mad id, Spain.
7
Clinical and Epidemiological Neu oscience (Neu oÈpia), Ins i u d’In es igació Sani à ia Pe e Vi gili (IISPV), Reus, Spain. ✉email: jaakko.kap io@helsinki.fi
www.na u e.com/jes Jou nal o Exposu e Science & En i onmen al Epidemiology
1234567890();,:
Land use desc ibes he human u iliza ion o land, in ol ing he
ans o ma ion om unde eloped a eas in o esiden ial and li ing
en i onmen s. U baniza ion is a pi o al d i ing o ce o he
change o cu en land use sys ems [10], and u ban planne s
conside mul iple concep s such as sui abili y, compe i i eness,
need di e si y, o esou ce sca ci y o e alua e land use [11]. A
ecen UK biobank s udy iden ified specific u ban en i onmen al
p ofiles including u ban land use densi y ha a ec men al heal h
h ough he egional b ain olume and pe inen biological
pa hways [12]. A Finnish s udy ound ha a iables e e ed o
he u ban en i onmen including land use ela ed o a low
incidence o se ious men al illnesses [13]. The e o e, ad ancing
li eable ini ia i es and shaping di e se land use is able o p omo e
heal hy li es yles, u ban ameni ies, and na u e conse a ion o
ul ima ely imp o e human heal h [14,15]. Some s udies ha e
specifically add essed he ela ionship be ween land use, ia
di e en indeces, and men al heal h/s a us, bu hei esul s we e
inconsis en [16–18]. Exis ing indices ha e some limi a ions, such
as insensi i eness o cap u e he in e ac ion be ween di e en
ypes o land use [19]. Inconsis en e idence eflec s he
complexi y o he land use e ec , demanding u he sophis i-
ca ed analysis, while we will encoun e di ficul ies such as high-
dimensionali y and small e ec sizes [20]. Ins ead o con en ional
eg ession models wi h a single index, in e p e able and obus
mul i-exposu e models a e ecommended. Ohanyan and collea-
gues ha e applied some machine lea ning models, illus a ed hei
cha ac e is ics, and employed hem in a s udy on a wide ange o
u ban exposu es and ype-2 diabe es [21,22]. Some simula ion
and e iew s udies ha e compa ed s a is ical app oaches and
assessed model pe o mance [23–25]. Howe e , his ype o
esea ch is ela i ely a e on men al heal h.
To ulfill he cu en esea ch gap, we hypo hesized he e is a
complex ela ionship be ween land use, unable o be quan ified
by con en ional indices, and dep essi e symp oms wi h h ee
objec i es: a) o clus e pa icipan s who sha ed a simila pa e n
o u ban land use; b) o assess bo h he linea and nonlinea
ela ionships be ween hem in young adul hood; and c) o
obse e he possible di e ences in hese ela ionships be ween
clus e s.
SUBJECTS AND METHODS
S udy pa icipan s
The pa icipan s we e om he FinnTwin12 coho , which is a popula ion-
based p ospec i e coho among all Finnish wins bo n be ween 1983 and
1987, and hei pa en s. A baseline, 5522 wins we e in i ed and 5184
wins eplied o ou ques ionnai e (age 11–12, wa e one), and hey
compose he o e all coho . All wins we e in i ed o pa icipa e in he fi s
ollow-up su ey wi h 92% e en ion a age 14 (wa e wo). Mo eo e , a
age 14, 1035 amilies we e in i ed o ake pa in an in ensi e subs udy
wi h psychia ic in e iews, some biological samples, and addi ional
ques ionnai es, and 1854 wins pa icipa ed in hese in e iews. They
we e also in i ed o a second in ensi e su ey as young adul s, wi h a
pa icipa ion a e o 73% (n=1347 indi idual wins), and comple ed he
de ailed young adul hood ques ionnai es and in e iews (pa o wa e
ou ). In addi ion, all o he wins in he o e all coho comple ed gene al
age 17 ques ionnai es (wa e h ee) and wins om he non-in ensi e s udy
comple ed young adul ques ionnai es (wa e ou ). Wa e ou was
conduc ed om 2004 o 2012, in which o e all 4824 indi idual wins
we e in i ed and 3404 eplied. In his s udy, we included wins who
pa icipa ed in wa e ou . An upda ed e iew o his coho was published
ecen ly [26].
Measu es
Dep essi e symp oms. In his s udy, he sho - e sion Gene al Beha io
In en o y (GBI) was used o e alua e dep essi e symp oms among wins in
young adul hood [27]. I is a sel - epo ed in en o y designed o iden i y
mood- ela ed beha io s, which is composed o 10 ques ions wi h a 4-poin
Like scale om 0 (ne e ) o 3 ( e y o en) o que y he occu ence o
dep essi e symp oms [28]. The o al sco e anges om 0 o 30, and a
highe sco e implies mo e dep essi e symp oms exis . To alida e he GBI,
we compa ed i o a Diagnos ic and S a is ical Manual o Men al Diso de s-
IV diagnosis o majo dep essi e diso de (MDD) assessed by he Semi-
S uc u ed Assessmen o he Gene ics o Alcoholism (SSAGA) in e iew
om he in ensi e s udy [29]. In a logis ic eg ession model, he GBI sco e
in young adul hood s ongly p edic ed MDD, wi h he a ea unde he
ecei e ope a ing cha ac e is ic cu e (AUC) o 0.8328 (among wins
included in his s udy’s analysis).
Land use. The EUREF-FIN geocodes o wins om bi h o 2021 we e
de i ed om he Digi al and Popula ion Da a Se ices Agency, Finland. We
used geocodes in 2012 o me ge he land use exposu es, de i ed om
U ban A las (UA) 2012, o he win da a. UA is a pa o land moni o ing
se ices o p o ide eliable, in e -compa able, high- esolu ion land use
maps in he Eu opean Union and Eu opean F ee T ade Associa ion
coun ies in 2006, 2012, and 2018 [30]. We used UA 2012 because i co e s
mo e a eas, o e 700 la ge unc ional u ban a eas, and con ains mo e
de ailed ca ego ies o land use in o ma ion, compa ed o UA 2006. Land
use exposu es included he pe cen age o 8 ypes o land use (high-densi y
esiden ial, low-densi y esiden ial, indus ial and comme cial, in as uc-
u e, u ban g een, ag icul u al esiden ial, na u al, and wa e ) in an a ea o
100, 300, and 500 m adius bu e zones o each geocode in u ban Finland
( o al o 24 exposu es).
Addi ionally, we also calcula ed he land use mix index in di e en
bu e s, which desc ibed he di e si y o land uses h ough he Shannon’s
E enness Index. I p o ides in o ma ion on a ea composi ion and ichness,
co e ing di e en land use ypes and hei ela i e abundances. The
equa ion is defined as ollows [31]:
land use mix index ¼
X
n
i¼1
Pi´ln Pi
!
=ln n
P
i
is he pe cen age o each ype o land use in zone i;nis he numbe o
land use ypes. I anges om 0 o 1, and a highe alue indica es a mo e
balanced dis ibu ion o land be ween he di e en ypes o land use.
Co a ia es. Se en co a ia es (demog aphic) we e defined a p io i: sex
(male, emale), zygosi y (monozygo ic (MZ), dizygo ic (DZ), unknown),
pa en al educa ion (limi ed, in e media e, high), smoking (ne e , o me ,
occasional, cu en ), wo k s a us ( ull- ime, pa - ime, i egula , no wo k-
ing), seconda y le el school ( oca ional, senio high school, none), and age.
The la e ou a iables came om he young adul hood su ey. Pa en al
educa ion was based on ma e nal and pa e nal epo s, while zygosi y was
based on DNA polymo phisms and/o a alida ed zygosi y ques ionnai e
[32].
Ano he ou social indica o s: age s uc u e (p opo ion o people o e
age 18 in he o al popula ion), educa ion le el (bachelo ´s/equi alen o
abo e o he popula ion o e age 16 (%)), unemploymen (unemploymen
a es among people who we e be ween 25 and 54 yea s old (%)), and
income le el (p opo ion o households in highes income qua ile in he
coun y) we e in oduced o accoun o socioeconomic s a us seg ega ion.
We de i ed social indica o s in 2012 a he pos al code le el o he wins’s
esidence a ha ime om S a is ics Finland.
Analysis
P epa a ion and desc ip ion. We only included hose wins who had
a ailable land use exposu es in 2012 in u ban a eas (as defined abo e),
indica ing ha hey li ed in he u ban a eas in Finland, and p o ided GBI
assessmen in young adul hood, in o de o ha e a la ge sample size and
ha e he wo measu emen s be as close as possible on he ime scale. A
o al o 1804 indi idual wins (589 win pai s and 626 indi idual wins)
we e included and he mean age in p o iding GBI assessmen was 24.07
yea s (a ound 2007–2011). Due o he skewness o he GBI sco e, we added
one o he GBI sco e and log- ans o med i o he ollowing analysis. A
co ela ion ma ix was d awn be ween land use exposu es. Then, we
p oposed se e al app oaches o assess he ela ionship be ween land use
exposu es and dep essi e symp oms.
Unsupe ised clus e ing. To g oup win indi iduals who ha e simila land
use in an explo a o y way, we used unsupe ised K-means clus e ing. The
K-means clus e ing me hod employs a non-hie a chical pa i ional algo-
i hm. I calcula s he o al wi hin-clus e a ia ion as he sum o he
Z. Wang e al.
771
Jou nal o Exposu e Science & En i onmen al Epidemiology (2024) 34:770 – 779
squa ed Euclidean dis ance be ween each sample and he co esponding
K-numbe andom-assigned cen oid in each clus e (k). X
ik
is he i
h
obse a ion belonging o clus e (k=1, 2, …., K) and n
K
is he numbe o
obse a ions in clus e k. The o e all wi hin-clus e a ia ion is defined as
ollows [33]:
X
K
k¼1X
nk
i¼1
Xik 1
nKX
nk
i¼1
Xik
!
2
The p ocess s ops when he c i e ion is me (smalles o e all wi hin-
clus e a ia ion) [33]. I is one o he simples and as es clus e ing
me hods, and is also able o handle ou lie s o inapp op ia e a iables
[34,35]. Only he 24 land use exposu es we e included in he clus e ing
algo i hm. We used he Silhoue e me hod o es ima e he op imal numbe
o p e-specified clus e [36], and wo clus e s we e iden ified (Supple-
men al Fig. 1). The R package “Fac oex a”was used [35].
Plu alis ic analysis. We spli he win pa icipan s in o aining and
es ing subse s. In ull win pai s, we pe o med a 1:1 andom spli wi hin
he pai . The emaining indi idual wins all wen in o he aining subse .
The aining sample size was 1215 and he es ing sample size was 589,
and he size in each clus e a ied (Supplemen al Table 1). By he
spli ing p ocess, we do no need o conside he s a is ical e ec o
complex sampling clus e e ec s by win pai s a us since all indi iduals
in bo h samples a e un ela ed. We chose wo ypes o models and
adjus ed co a ia es o e alua e he isk es ima ion o 24 land use
exposu es (j).
Fi s , he linea elas ic ne penalized eg ession model was applied o
ea u e selec ion, which uses a hyb id o he lasso and idge penalized
me hods o fi he gene alized linea model [37]. I encou ages he
g ouping e ec ha co ela ed a iables end o be in o ou o he
model oge he wi h simila coe ficien s, and hen a iables a e selec ed
based on hei p edic i e powe in he con ex o penal y [38].
Coe ficien s a e sh unk, e en o ze o, o p omo e spa si y and educe
mul icollinea i y [39]. I is e y use ul in da ase s wi h highly co ela ed
a iables. A ypical linea eg ession model based on N pa icipan s wi h
he combined penalized e m is defined as ollows [39]:
min
β0;β
1
2NX
N
i¼1
yiβ0xT
iβ
2þλX
p
j¼1
1α
2
β2
jþαβ
j
!
y
i
is he dependen esponse and x
i
is he independen ac o a
obse a ion i.λis a posi i e egula iza ion pa ame e . β
0
and βa e scala
and p- ec o coe ficien s, espec i ely. We se he α, anging om 0.1 o
1.0, as a uning pa ame e , o he penal y. The final models we e selec ed
by 10- old c oss- alida ion wi h minimal c i e ia o de e mine he op imal
deg ee o penaliza ion [37]. The e we e wo adjus men plans: 1)
demog aphic co a ia es (minimal), and 2) demog aphic co a ia es and
social indica o s ( u he ). We o ced he demog aphic co a ia es and social
indica o s in o he models, wi hou penal y, o ully adjus hem. S a a
package “elas icne ”was used.
Fu he , o assess he nonlinea ela ionship, he supe ised machine
lea ning model eX eme G adien Boos ing (XGBoos ) was used. I is a ee-
based g adien boos ing echnique, u ilizing he weigh s o ees, which is
good a p edic ing and educes he isk o o e fi ing [40,41]. The
objec i e unc ion o XGBoos s a s wi h wo pa s: a loss unc ion and a
egula iza ion e m, and we aim o ob ain he op imal ou pu alue (O
alue
)
o minimize he unc ion, defined as ollows:
X
n
i¼1
Ly
i;p 1
iþO alue
þγTþ1
2λO2
alue
p 1
iis he p e ious p edici on o ee a obse a ion i.Tis he numbe
o lea nodes in a ee, and γand λa e he definable penal y ac o s o
a oid o e fi ing. Then, we ew i e he loss unc ion acco ding o he 2nd
Taylo App oxima ion:
Ly
i;P 1
iþO alue
Ly;pi
ðÞþ
d
dpi
Ly;pi
ðÞ
O alue þ1
2
d2
dp2
i
Ly;pi
ðÞ
O2
alue
¼Ly;pi
ðÞþgO alue þ1
2hO2
alue
Ly;pi
ðÞis he loss unc ion o he p e ious p edic ion, and i s fi s and
second de i a i e a e labeled as gand h, espec i ely. The op imum ou pu
alue could hen be de i ed wi h Gand H(sum o gand h) as:
O aluej ¼1
2X
j¼1
G2
j
HjþλþγT
The de ailed ma hema ical model and algo i hm a e desc ibed in
p e ious li e a u e [42]. This model is able o cha ac e ize in e ac ions and
nonlinea i y [21]. The uning hype pa ame e s we e calib a ed by
pa allelizable Bayesian op imiza ion based on se en ini ializa ion e alua-
ions and mul iple epochs, using he R package “Pa BayesianOp imiza ion”
[43,44]. We an aining XGboos models wi h 3000 ounds a fi s , hen
he op imal numbe o ounds (n) was selec ed by mean-squa ed e o
(MSE) as he ollowing equa ion:
MSEn<0:99 1
20 MSEn1þ¼þMSEn21
ðÞ
The final XGBoos analysis was conduc ed wi h all hype pa ame e s
using he R package “xgboos ”[40]. Finally, we used he Shapley (SHAP)
alue o in e p e and isualize he esul s om he XGboos machine
lea ning model wi h highe anspa ency by he R package “SHAP o xg-
boos ”[45,46], and i was commonly used in p e ious s udies [21,22,47].
The SHAP alue unifily measu es he impo ance o each land use exposu e
on GBI om he XGBoos model based on he coope a i e game heo y
[45]. The di ec ion o SHAP alue indica es whe he each land use exposu e
impac s posi i ely o nega i ely he p edic ion o GBI. The XGboos model
was conduc ed wice. Fi s , we pu all land use exposu es and
demog aphic co a ia es in o he model, hen social indica o s we e added.
Models we e pe o med among o e all pa icipan s and in he wo
clus e s. We used oo -mean-squa ed e o (RMSE) o measu e model
pe o mance in he aining and es ing subse s, which is a weigh ed
measu e calcula ed be ween o ecas and obse ed alues.
Sensi i i y analysis. To con ol he po en ial gene ic e ec , we u he
pe o med he linea mixed model, in which he win pai was assigned as
he fixed e m in he model. This model was o speci y ha he land use
exposu es did no a y be ween co wins and o compu e hei wi hin-pai
e ec . Two adjus men plans we e employed, excluding zygosi y and
pa en al educa ion which do no a y wi hin pai s. Then, we conduc ed a
pos -hoc linea eg ession be ween he land use mix index and log-
ans o med GBI sco e, which aims o compa e wi h ou no el findings.
Two adjus men plans we e employed and he clus e e ec o sampling
based on amilies o win pai s was con olled by he obus s anda d e o .
Ap alue less han 0.05 was conside ed s a is ically significan and 95%
confidence in e als (CI) a e epo ed.
RESULTS
K-means clus e ing and desc ip i e s a is ics
Figu e 1depic s he dis ibu ion o each land use ca ego y o e all
and in he wo clus e s. Clus e 2 had a highe pe cen age o high-
densi y esiden ial land use, while Clus e 1 had a highe
pe cen age o low-densi y esiden ial land use ega dless o he
bu e adii o he wins’loca ion. Supplemen a y Fig. 2 shows he
wins’loca ion in he g ea e Helsinki a eas (as an example), and
wins om Clus e 2 li ed in mo e u banized a eas (o en close o
ci y o own cen e s), while wins om Clus e 1 we e mo e
subu ban. Va iable names and de ails a e shown in Supplemen a y
Table 2. We also calcula ed he simple a ios o means be ween
he wo clus e s and ound low-densi y esiden ial, ag icul u al
esiden ial, and na u al land use in a 100 m bu e ha e no ably
“ ela i e”di e ences be ween he wo clus e s ( a io>10).
Acco ding o he co ela ion ma ix based on he aining subse
(Supplemen al Fig. 3), he same land use wi h di e en adii o he
bu e zone was highly co ela ed. High-densi y and low-densi y
esiden ial land use we e nega i ely co ela ed. No ably, he e was
a highe numbe o co wins om MZ pai s who bo h li ed in
Clus e 1 han li ed disco dan ly, compa ed o DZ pai s
(Supplemen a y Table 3).
Table 1shows he dis ibu ion o cha ac e is ics o e all and in
he wo clus e s. O e all, he majo i y o wins a e emale (58.7%),
dizygo ic (61.3%), and epo ed ne e smoking (55.1%) in he
Z. Wang e al.
772
Jou nal o Exposu e Science & En i onmen al Epidemiology (2024) 34:770 – 779
young adul hood ques ionnai e. Addi ionally, 48.8% and 67.7% o
wins epo ed ha hey we e in ull- ime wo k and had a ended
senio high school, espec i ely. The majo i y (51.1%) o wins’
pa en s had limi ed educa ion le els (less han senio high school).
The means o GBI sco e we e 4.4, 4.1, and 4.7 among o e all
pa icipan s, hose in Clus e 1 (subu ban), and in Clus e 2 (ci y
cen e ), espec i ely, and hei dis ibu ions a e p esen ed in
Supplemen a y Fig. 4. Unsupe ised K-means clus e ing did no
ake in o accoun hese demog aphic co a ia es. We obse ed
significan di e ences in smoking, wo king s a us, seconda y le el
school, and pa en al educa ion be ween he wo clus e s by Chi-
squa ed es o uni a iable linea eg ession accoun ing o win
sampling. The e we e mo e wins who cu en ly smo ked, wo ked
ull ime, and a ended oca ional schools in Clus e 1 han in
Clus e 2, bu pa en s in Clus e 2 had a lowe pe cen age o
ecei ing limi ed educa ion. Add ionally, in all ou social
indica o s, he e we e significan di e ences be ween clus e s.
Linea elas ic ne eg ession model
A e minimal adjus men o demog aphic co a ia es, in Clus e 1
(subu ban), 11 land use exposu es we e significan enough o be
cap u ed by he linea elas ic ne eg ession model in assessing
hei ela ionship wi h GBI (Table 2). The ag icul u al esiden ial
land use in a 100 m bu e inc eased log- ans o med GBI sco es
wi h he la ges penalized coe ficien (coe ficien : 0.097). A e
u he adjus men wi h he social indica o s, he numbe o
selec ed land use exposu es inc eased o 17, and he new
exposu es we e: u ban g een and na u al land use in bo h 100 and
500 m bu e s, and high-densi y esiden ial and wa e land use in a
300 m bu e . The penalized coe ficien o he ag icul u al
esiden ial land use in a 100 m bu e was a enua ed (coe ficien :
0.067), while i s ill had he la ges e ec size and was posi i ely
co ela ed wi h GBI. Su p isingly, he e we e no land use
exposu es emaining in he Clus e 2 (ci y cen e ) model in
nei he adjus men phase. Supplemen al Table 4 p esen s he
esul s in he o e all model, and a e u he adjus men , he e
we e also mo e land use exposu es selec ed. The pa e n o
coe ficien s including he e ec size and di ec ion was ela i ely
he e ogeneous wi h Clus e 1. The coe ficien s o low-densi y
esiden ial land use in a 100 m bu e we e he same (coe ficien :
−0.011) be ween he o e all and Clus e 1 models a e minimal
adjus men .
Refi ing o linea mixed model
Acco ding o he selec ed land use exposu es om he a o emen-
ioned elas ic ne eg ession, we efi ed hem in o linea mixed
models o assess hei wi hin-pai e ec on log- ans o med GBI
sco es (Supplemen a y Table 5). In Clus e 1, a e minimal
adjus men , comme cial and indus ial land use in a 300 m bu e
we e significan ly and posi i ely associa ed wi h GBI, while he e ec
a enua ed a e u he adjus men . In he o e all model, a e bo h
minimal and u he adjus men , highe low-densi y esiden ial land
use in a 100 m bu e significan ly educed he GBI.
XGBoos model
We lis ed he op fi e mos impo an ac o s wi h SHAP alues in
each clus e ’s XGBoos model. A e minimal adjus men (Fig. 2A), in
Clus e 1 (subu ban), he mos impo an land use exposu e was
na u al land use in a 100 m bu e , and he second was comme cial
and indus ial land use in a 300 m bu e . A e u he adjus men ,
na u al land use in a 100 m bu e became he mos impo an
(Fig. 2B). In Clus e 2 (ci y cen e ), he mos impo an land use
exposu e was always in as uc u e land use in a 300 m bu e a e
minimal (Fig. 2C) and u he adjus men (Fig. 2D). Co a ia es we e
no lis ed and a e no shown in he figu e. The cu e o SHAP alues
sugges ed nonlinea a ibu ion o each land use exposu e on GBI.
No ablely, he cu es o in as uc u e land use in a 300 m bu e wi h
SHAP alues a e also simila a e minimal o u he adjus men .
Fig. 1 His og am o pe cen age o land use exposu e among o e all pa icipan s, hose in Clus e 1, and in Clus e 2.
Z. Wang e al.
773
Jou nal o Exposu e Science & En i onmen al Epidemiology (2024) 34:770 – 779
The e was a fla incline o SHAP alue be ween 0 and ~10%. Then,
he alue sha ply inc eased when i s pe cen age passed ~10% and
he impac o in as uc u e land use in a 300 m bu e on he
p edic ion o GBI swi ched om nega i e o posi i e. A e he
pe cen age was g ea e han ~20%, he cu e slowly inc eased. The
esul s o o e all XGBoos models a e p esen ed in Supplemen al
Fig. 5. A e minimal adjus men , same as Clus e 2, he mos
impo an land use exposu e is in as uc u e land use, bu , in a
100 m bu e (Supplemen a y Fig. 5A). A e u he adjus men , he
mos impo an becames na u al land use in a 100 m bu e
(Supplemen a y Fig. 5B).
Model pe o mance and compa ison
The s anda d de ia ions (SD) o he log- ans o med GBI sco e
we e 0.8825, 0.8851, and 0.8774 among o e all, Clus e 1’s and
Clus e 2’s wins, espec i ely. The aining and es ing RMSE a e
shown in Supplemen a y Table 6. The e a e no majo di e ences
be ween he wo ypes o models and clus e s, and hey mos ly
ha e lowe SDs han hose o he log- ans o med GBI sco e,
implying good model pe o mance.
Linea eg ession wi h he land use mix index
The esul s o linea eg ession in he o e all and he wo
sepa a ed clus e models a e p esen ed in Table 3. In he c ude
Clus e 1 (subu ban) model, a highe land use mix index in a
300 m bu e was significan ly associa ed wi h highe log-
ans o med GBI sco es (be a: 0.51, 95% CI: 0.02, 1.01). A e ei he
minimal o u he adjus men , he e was no significan associa-
ion, which implies he need o complex assessmen s be ween
land use and GBI.
Table 1. Cha ac e is ics o all included wins o e all and in he wo clus e s. The p alues a e o di e ences be ween Clus e s 1 and 2 by Chi-squa ed
es o uni a iable linea eg ession accoun ing o win sampling.
Cha ac e is ic N(%) / Mean (SD) P alue (be ween
clus e s)
O e all (indi idual win
n=1804)
Clus e 1 (indi idual win
n=736)
Clus e 2 (indi idual win
n=1068)
GBI in young adul hood 4.42 (4.7) 4.05 (4.4) 4.67 (4.8) 0.01
Demog aphic co a ia es
Sex 0.16
Male 745 (41.3) 289 (39.3) 456 (42.7)
Female 1059 (58.7) 447 (60.7) 612 (57.3)
Zygosi y 0.92
Monozygo ic 615 (34.1) 252 (34.2) 363 (34.0)
Dizygo ic 1105 (61.3) 448 (60.9) 657 (61.5)
Unknown 84 (4.7) 36 (4.9) 48 (4.5)
Smoking
Ne e 994 (55.1) 405 (55) 589 (55.2) 0.03
Fo me 191 (10.6) 78 (10.6) 113 (10.6)
Occasional 205 (11.4) 66 (9.0) 139 (13.0)
Cu en 414 (23.0) 187 (25.4) 227 (21.3)
Wo k <0.0001
Full- ime wo k 880 (48.8) 409 (55.6) 471 (44.1)
Pa - ime wo k 280 (15.5) 94 (12.8) 186 (17.4)
I egula wo k 239 (13.3) 76 (10.3) 163 (15.3)
No wo king 405 (22.5) 157 (21.3) 248 (23.2)
Seconda y le el school <0.0001
Voca ional 486 (26.9) 262 (35.6) 224 (21.0)
Senio high school 1222 (67.7) 439 (59.7) 783 (73.3)
None 96 (5.3) 35 (4.8) 61 (5.7)
Pa en al educa ion <0.0001
Limi ed 922 (51.1) 429 (58.3) 493 (46.2)
In e media e 410 (22.7) 155 (21.1) 255 (23.9)
High 472 (26.2) 152 (20.7) 320 (30.0)
Age 24.07 (1.7) 24.15 (1.7) 24.01 (1.7) 0.10
Social indica o s
a
Age s uc u e (%) 82.7 (7.2) 78.2 (5.8) 85.8 (6.4) <0.0001
Educa ion le el (%) 25.8 (9.0) 21.8 (8.0) 28.5 (8.6) <0.0001
Unemploymen (%) 9.6 (4.0) 8.9 (4.1) 10.0 (3.9) <0.0001
Income le el (%) 25.5 (10.0) 26.3 (10.3) 24.9 (9.8) 0.01
a
The de ailed desc ip ion o social indica o s was in oduced in he “Subjec s and Me hods”sec ion.
Z. Wang e al.
774
Jou nal o Exposu e Science & En i onmen al Epidemiology (2024) 34:770 – 779
DISCUSSION
Based on 1804 wins om he FinnTwin12 s udy wi h in o ma ion
on esiden ial geocodes linked o land use cha ac e is ics, we
iden ified wo clus e s o he land use en i onmen he wins li ed.
S eng hened by mul iple s a is ical app oaches, bo h linea and
nonlinea ela ionships be ween land use and dep essi e symp-
oms we e disco e ed o exis . In he linea elas ic ne penalized
eg ession model, among o e all wins and Clus e 1 (subu ban)’s
wins, he e was a he e ogeneous pa e n in selec ed ea u es,
e ec sizes, and e ec di ec ions. In he Clus e 1 model,
ag icul u al esiden ial land use in a 100 m bu e was associa ed
wi h dep essi e symp oms wi h he la ges ela i e e ec size. A e
con olling o he influence o he social en i onmen , mo e land
use exposu es we e ound o be associa ed wi h dep essi e
symp oms. Wi h u he con ol o he gene ic e ec , based on he
efi ing mixed models, no land use exposu e was s ongly
associa ed wi h dep essi e symp oms, implying a po en ial
inhe i able e ec behind. In con as , no land use exposu es we e
significan enough o be a ibu ed o dep essi e symp oms in
Clus e 2, no ma e he adjus men o he social en i onmen ,
which was ypical o ci y o own cen e s. The XGBoos model
o e ed a p o ound unde s anding o he mul i ace ed ela ionships
ega ding he in ica e in e play be ween a ious land use
measu es and hei ela i e impo ance on dep essi e symp oms.
The impo ance anks and nonlinea i y o land use exposu es on
dep essi e symp oms we e he e ogeneous be ween he o e all,
Clus e 1, and Clus e 2 models. The mos impo an we e
comme cial and indus ial land use in a 300 m bu e in Clus e 1
and in as uc u e land use in a 300 m bu e in Clus e 2, a e
adding social indica o s in. As a hypo hesis-gene a ing s udy,
elemen s such as popula ion he e ogenei y, en i onmen al in e -
ac ion, and cha ac e is ics o he e ec (such as linea i y) should be
conside ed mo e in u u e analyses be ween land use, as well as
he b oad u ban en i onmen , and dep essi e symp oms.
Table 2. Mul iple-exposu e elas ic ne penalized eg ession o associa ions be ween land use and GBI in Clus e s 1 and 2. The emaining coe ficien s
we e significan enough o be selec ed.
Land use (Bu e ) uni : % S anda dized elas ic ne coe ficien
Clus e 1 Clus e 2
Minimally adjus ed
a
Fu he adjus ed
b
Minimally adjus ed
a
Fu he adjus ed
b
High-densi y esiden ial (100 m) 0.089 0.056
Low-densi y esiden ial (100 m) −0.011 −0.043
Comme cial and indus ial
(100 m)
In as uc u es (100 m)
U ban g een (100 m) 0.001
Ag icul u al esiden ial (100 m) 0.097 0.067
Na u al (100 m) −0.003
Wa e (100 m)
High-densi y esiden ial (300 m) 0.002
Low-densi y esiden ial (300 m)
Comme cial and indus ial
(300 m)
0.084 0.065
In as uc u es (300 m) −0.031 −0.029
U ban g een (300 m) 0.081 0.058
Ag icul u al esiden ial (300 m)
Na u al (300 m) −0.014
Wa e (300 m) 0.002
High-densi y esiden ial (500 m) 0.046 0.026
Low-densi y esiden ial (500 m) 0.035 0.036
Comme cial and indus ial
(500 m)
In as uc u es (500 m) −0.012 −0.005
U ban g een (500 m) 0.010
Ag icul u al esiden ial (500 m) −0.067 −0.019
Na u al (500 m)
Wa e (500 m) 0.020 0.012
Model ea u e (10- old CV
selec ion)
α=1.00, λ=0.01,
Ou -o -sample
R
2
=0.06,
CV p edic ion
e o =0.74
α=0.10, λ=0.05,
Ou -o -sample
R
2
=0.06,
CV p edic ion
e o =0.74
α=1.00, λ=0.04,
Ou -o -sample
R
2
=0.10,
CV p edic ion
e o =0.70
α=1.00, λ=0.04,
Ou -o -sample
R
2
=0.09,
CV p edic ion
e o =0.70
a
Adjus ed o sex, zygosi y, smoking, wo k s a us, seconda y le el school, pa en al educa ion, and age when wins p o ided he GBI assessmen in young
adul hood.
b
Adjus ed o sex, zygosi y, smoking, wo k s a us, seconda y le el school, pa en al educa ion, age when wins p o ided he GBI assessmen in young
adul hood, as well as age s uc u e, educa ion le el, unemploymen , and income le el.
Z. Wang e al.
775
Jou nal o Exposu e Science & En i onmen al Epidemiology (2024) 34:770 – 779
Fi s , he clus e ing analysis e ealed a specific pa e n in
u baniza ion, and wins om Clus e s 1 and 2 mos ly li ed in he
“subu bs”and “ci y o own cen e s”, espec i ely. The land use
exposu es a e less impo an o dep essi e symp oms among
people li ing in ci y o own cen e s. The possible mechanisms
may be h ough di e en ial heal hca e se ice access, social
needs, anspo a ion connec edness, o neighbo hood en i on-
men [17,48,49]. Fo example, li ing in he subu bs usually
equi es longe house- o-job commu ing dis ances, which has
been ound o be associa ed wi h poo e men al heal h [48].
Longe job commu es impliy g ea e need o anspo a ion
in as uc i e, and, simila o ou linea elas ic ne eg ession
model, he highe pe cen age o in as uc u e land use was
ela ed o ewe dep essi e symp oms in Clus e 1 (subu ban).
Ne e heless, Pelg ims e al. de ec ed no significan associa ion,
a e ull adjus men , be ween g een su ounding, s ee co ido
and canyon e ec s, and dep essi e diso de among pa icipan s
li ing in he highly u banized B ussels, Belgium [50]. Fu he mo e,
he impac o he social en i onmen on he ela ionship be ween
land use exposu es and dep essi e symp oms is mo e p o-
nounced in subu ban a eas compa ed o ci y cen e s. In China, he
media ing ole o neighbo hood-le el social capi al was shown o
be e iden in he connec ion be ween u baniza ion and
dep essi e symp oms [51]. Since his is a single-coun y s udy,
Finland, compa ed o o he de eloped coun ies, has quie e and
g eene u ban spaces ha need o be conside ed in he
in e p e a ion. We did no in end o dis inguish people wi h an
a bi a y bina y classifica ion, ins ead, we p omo e he hypo hesis
ha he ela ionship be ween land use and dep essi e symp oms
exis s in he specific land use con ex .
Mo e b oadly, land use exposu es, ha signaled u baniza ion,
we e ei he selec ed by he penalized model o we e among he
op fi e in he XGBoos model, indica ing hem as good
candida es o explain dep essi e symp oms. Niu e al. de eloped
a amewo k o he coupling coo dina ion ela ionship be ween
u baniza ion and land use ansi ion in China and sugges ed a
con e gence phenomenon be ween hem [52]. Ne e heless,
p e ious e idence on he e ec o u baniza ion on dep ession is
no consis en . A 2020 e iew ound a p o ec i e e ec o
u baniza ion on dep ession in h ee Chinese s udies, while ou
o he coun ies’s udies had opposi e findings due o di e en
geog aphic egions and income le els [53]. An inc easing end in
dep ession p e alence among young adul s and hose who li ed
in u al a eas wi h low popula ion densi y was obse ed in a
longi udinal Ge many na ionwide su ey [54]. Howe e , Mo ozo
indica ed ha u baniza ion ad e sely a ec ed men al heal h ia
se e al ac o s including noise and isual agg essi eness o he
en i onmen in Russia [55]. The e may be conjunc o nonaddi i e
ela ionships wi hin land use o b oad u ban li ing en i onmen s.
The en i onMENTAL Conso ium has ske ched he mul iple
mechanisms be ween u ban li ing en i onmen al p ofiles wi h
mo e han a hund ed a iables and psychia ic symp oms [12]. A
ypical example o complexi y is he u ban hea island e ec , a
highe egional empe a u e in u ban a eas han in su ounding
u al a eas. I is di e en ially influenced by many land use ac o s,
in which expansion o buil -up a ea inc eased bu wa e a eas
educed he egional empe a u e [56], and mo eo e , he u ban
hea island inc eases he isk o dep ession [53]. Addi ionally, he
di ec ions o wo nega i ely co ela ed land use exposu es’
influence we e no always consis en ac oss a ying bu e s, hus
a single exposu e canno be in e ed as a isk o p o ec i e ac o .
Bu e s p o ide a conside a ion o con ex ual e ec , which
inca na e he spa ial scale o di e en pa hways linking u ban
en i onmen s o heal h [57]. Thus, o he implica ion o u ban
planning and imp o emen , we ad oca e ha policymake s
ecognize he in ica e na u e o ou u ban en i onmen and
adop a pe spec i e ha encompasses i as a holis ic in eg a ion,
ins ead o a limi ed se o indices o indica o s.
Including mul iple land use exposu es in a single analy ic pla o m
allows us o disen angle he indi idual e ec s and assess he
complex ela ionships. To some ex en , machine lea ning models
allow us o adjus o conside he mu ual e ec be ween di e en
land use exposu es ins ead o epea ed single eg ession models. The
linea elas ic ne penalized eg ession models selec ed a subse o
he mos impo an land use exposu es and educed he isk o
co ela ing and o e fi ing, wi h be e pe o mance [38]. Because we
aim o e eal ela ionships ins ead o p edic ion, we did no efill he
land use exposu es o he no mal eg ession model and he
in e p e a ion o e ec size was weakened. Len e s e al. ha e applied
his app oach o p ena al chemical exposu es o sol e he
in e connec ed e ec s o mix u es [37]. We also obse ed he
nonlinea ela ionship ia he in e p e able SHAP isualiza ion om
XGBoos , bu , like Ohanyan and colleagues’s udies, we did no
s aigh o wa dly assess he in e ac ion due o modes e ec sizes
and o he ac o s [21,58]. P e ious applica ions o his machine
lea ning me hod imp o ed he p edic ion and o ecas o ai quali y
Fig. 2 Shapley (SHAP) dependence plo s o he op fi e mos influen ial exposu es in XGBoos models. The dependence plo shows he
ela ionship be ween he SHAP alue and land use exposu es in ou models. Clus e 1 wi h minimal adjus men (A), Clus e 1 wi h u he
adjus men (B), Clus e 2 wi h minimal adjus men (C), Clus e 2 wi h u he adjus men (D). Demog aphic co a ia es and social indica o s
we e included in he models bu supp essed in plo s o highligh land use exposu es.
Z. Wang e al.
776
Jou nal o Exposu e Science & En i onmen al Epidemiology (2024) 34:770 – 779
in China [41,59]. Ma e al. also compa ed he p edic ion accu acy
be ween XGBoos and Lasso penalized eg ession models [59], while,
in ou s udy, we wished o obse e he in ica e e ec s ins ead o
compa ing accu acy, so we used RMSE, no AUC, o e alua e model
pe o mance. Ano he Chinese s udy also explo ed he nonlinea
e ec be ween he buil and social en i onmen s and bus use
among olde adul s [42]. In ad ancing he con en ional eg ession
model wi h limi ed exposu es, he u ili y o mul iple machine
lea ning algo i hms p o ides a p elimina y ske ch o he laby in hine
ela ionship be ween u ban land use and dep ession symp oms.
Clus e ing analysis ocused on mul iple land use exposu es and
acili a es he segmen a ion o esiden s o ailo ed epidemiolo-
gical assessmen o he e ec o land use on dep essi e symp oms
and cus omizes u he imp o emen and in e en ion. The
di e en ial pa e n o u ban land use en i onmen was e y
ob ious in ou findings. Me hodologically, clus e ing analysis has
gained inc easing a en ion in he field o exposu e science.
Tognola and colleagues clus e ed child en in F ance by exposu e
o ex emely low- equency magne ic fields [60], and ano he
s udy de eloped a no el wo kflow in clus e ing wi h mul iple
ea u es including specific and gene al ex e nal exposomes and
iden ified sub-popula ions in ype-2 diabe es pa ien s [61].
The e a e some limi a ions in ou s udies. Fi s , he in o ma ion on
dep ession symp oms was ob ained be o e 2012, so he po en ial
causali y and di ec ion a e unable o be confi med due o
empo ali y. Addi ionally, empo ali y also leads o he ques ion o
he leng h and s abili y o exposu es, so a li ecou se s udy is needed.
Second, compa ed o p e ious simila s udies, he sample size is
ela i ely small. Al hough he wo machine lea ning me hods a e
able o sh ink he o e fi ing due o he small sample size, we s ill
need o be cau ious abou he findings. Thi d, we did no “ ully”
le e age he win s uc u e o quan i y he po en ial gene ic
influence, al hough conco dance and disco dance in clus e s
di e ed be ween monozygo ic and dizygo ic wins. Ins ead, we
used a mixed model o u he explo e he wi hin-pai e ec o
p ope ly con ol he unde lying gene ic e ec . Inco po a ion o a
win design could guide he in es iga ion o unde lying gene ic
influence in he high-dimensional en i onmen al s udy in he u u e.
Fou h, he e a e po en ial con ounding e ec s s emming om
o he physical exposu es such as ai pollu ion and noise. Al hough
he land use exposu es al eady ca y some in o ma ion abou hese
exposu es [62], ou o hcoming endea o s will employ ad anced
echniques and models o measu e hese. Finally, he in e p e abili y
o he machine lea ning model is a significan challenge ha
equi ed mo e endea o in he field o da a science. We ound he
nonlinea i y pa e n, bu i is di ficul o elabo a e on. This s udy is a
pilo s udy o explo a ion,and u he ollow-ups udiesa e
welcome o s eng hen he e idence.
CONCLUSION
This s udy is he fi s , o ou knowledge, o in es iga e he complex
ela ionship be ween mul iple u ban land use exposu es and
dep essi e symp oms in young adul hood. The plu alis ic mul i-
model in e ences selec ed o p io i ized he mo e impo an u ban
land use exposu es o dep essi e symp oms and e ealed linea and
nonlinea ela ionships, which ad ances he con en ional assess-
men wi h a single index. Clus e ing analysis showed a no able
he e ogeneous pa e n in hese ela ionships be ween pa icipan s
wi h di e en land use en i onmen s, implying he e ec s a e unde
aspecific con ex . Due o sample size, model cha ac e is ics, and
empo ali y, ou finding in e p e a ion is cau ious a p esen , and
mo e e o s a e wa an ed o co obo a e.
DATA AVAILABILITY
The FinnTwin12 da a is no publicly a ailable due o he es ic ions o in o med
consen . Howe e , he FinnTwin12 da a is a ailable h ough he Ins i u e o
Molecula Medicine Finland (FIMM) Da a Access Commi ee (DAC) (fimm-dac@hel-
sinki.fi) o au ho ized esea che s who ha e IRB/e hics app o al and an ins i u ionally
app o ed s udy plan. To ensu e he p o ec ion o p i acy and compliance wi h
na ional da a p o ec ion legisla ion, a da a use/ ans e ag eemen is needed, he
con en and specific clauses o which will depend on he na u e o he eques ed
da a.
REFERENCES
1. Twenge JM, Coope AB, Joine TE, Du y ME, Binau SG. Age, pe iod, and coho
ends in mood diso de indica o s and suicide- ela ed ou comes in a na ionally
ep esen a i e da ase , 2005-2017. J Abno m Psychol. 2019;128:185–99.
2. Fila o a S, Upadhyaya S, K ons öm K, Suominen A, Chudal R, Lun amo T, e al.
Time ends in he incidence o diagnosed dep ession among people aged 5–25
yea s li ing in Finland 1995–2012. No d J Psychia y. 2019;73:475–81.
Table 3. Linea eg ession be ween land use mix index and GBI in young adul hood.
Land use mix index Log- ans o med GBI sco es in young adul hood
Mean (SD) Unadjus ed be a (95% CI) Minimally adjus ed be a (95% CI)
a
Fu he adjus ed be a (95% CI)
b
O e all (indi idual win n =1804)
In a 100 m bu e 0.38 (0.14) 0.11 (−0.17, 0.39) 0.17 (−0.10, 0.44) 0.17 (−0.10, 0.44)
In a 300 m bu e 0.60 (0.12) 0.17 (−0.18, 0.52) 0.14 (−0.19, 0.48) 0.12 (−0.22, 0.45)
In a 500 m bu e 0.67 (0.11) 0.30 (−0.08, 0.68) 0.25 (−0.11, 0.62) 0.19 (−0.18, 0.56)
Clus e 1 (indi idual win n =736)
In a 100 m bu e 0.38 (0.15) 0.02 (−0.39, 0.44) 0.12 (−0.28, 0.52) 0.11 (−0.29, 0.50)
In a 300 m bu e 0.59 (0.13) 0.51 (0.02, 1.01)* 0.46 (−0.01, 0.92) 0.26 (−0.21, 0.72)
In a 500 m bu e 0.66 (0.13) 0.50 (−0.00, 1.01) 0.42 (−0.07, 0.91) 0.16 (−0.35, 0.67)
Clus e 2 (indi idual win n =1068)
In a 100 m bu e 0.38 (0.14) 0.16 (−0.21, 0.53) 0.18 (−0.18, 0.54) 0.16 (−0.21, 0.52)
In a 300 m bu e 0.60 (0.11) −0.25 (−0.73, 0.23) −0.27 (−0.73. 0.19) −0.30 (−0.77, 0.18)
In a 500 m bu e 0.68 (0.10) −0.08 (−0.63, 0.48) −0.14 (−0.67, 0.40) −0.19 (−0.76, 0.38)
a
Adjus ed o sex, zygosi y, smoking, wo k s a us, seconda y le el school, pa en al educa ion, and age when wins p o ided he GBI assessmen in young
adul hood.
b
Adjus ed o sex, zygosi y, smoking, wo k s a us, seconda y le el school, pa en al educa ion, age when wins p o ided he GBI assessmen in young
adul hood, as well as age s uc u e, educa ion le el, unemploymen , and income le el.
*P alue < 0.05.
Z. Wang e al.
777
Jou nal o Exposu e Science & En i onmen al Epidemiology (2024) 34:770 – 779
3. Sho ey S, Ng ED, Wong CHJ. Global p e alence o dep ession and ele a ed
dep essi e symp oms among adolescen s: a sys ema ic e iew and me a-
analysis. B J Clin Psychol. 2022;61:287–305.
4. Wang C, Wen W, Zhang H, Ni J, Jiang J, Cheng Y, e al. Anxie y, dep ession, and
s ess p e alence among college s uden s du ing he COVID-19 pandemic: a
sys ema ic e iew and me a-analysis. J Am Coll Heal. 2021;71:2123–30.
5. Hawes MT, Szenczy AK, Klein DN, Hajcak G, Nelson BD. Inc eases in dep ession
and anxie y symp oms in adolescen s and young adul s du ing he COVID-19
pandemic. Psychol Med. 2022;52:3222–30.
6. Pozuelo JR, Desbo ough L, S ein A, Cip iani A. Sys ema ic e iew and me a-
analysis: dep essi e symp oms and isky beha io s among adolescen s in low-
and middle-income coun ies. J Am Acad Child Adolesc Psychia y.
2022;61:255–76.
7. Sulli an PF, Neale MC, Kendle KS. Gene ic epidemiology o majo dep ession:
e iew and me a-analysis. Am J Psychia y. 2000;157:1552–62.
8. Lau JYF, Eley TC. Changes in gene ic and en i onmen al influences on dep essi e
symp oms ac oss adolescence and young adul hood. B J Psychia y. 2006;189:422–7.
9. Hu Y-M. Sex di e ences in gene ic and en i onmen al con ibu ions o
dep ession symp oms in Sou h Ko ean adolescen and young adul wins. Twin
Res Hum Gene . 2008;11:306–13.
10. Nuissl H, Sieden op S. U banisa ion and land use change BT - sus ainable land
managemen in a Eu opean con ex : a co-design app oach. In: Wei h T, Ba k-
mann T, Gaasch N, Rogga S, S auß C, Zscheischle J (eds). Sp inge In e na ional
Publishing: Cham, 2021, pp 75–99.
11. Dong G, Ge Y, Jia H, Sun C, Pan S. Land use mul i-sui abili y, land esou ce sca ci y
and di e si y o human needs: a new amewo k o land use conflic iden ifi-
ca ion. Land. 2021;10. h ps://doi.o g/10.3390/land10101003.
12. Xu J, Liu, Polemi i N, Ga cia-Mond agon E, Tang J L, Liu X, e al. E ec s o u ban
li ing en i onmen s on men al heal h in adul s. Na Med. 2023;29:1456–67.
13. Vou ilainen A, Ha ikainen S, She wood PR, Taipale H, Tolppanen A-M,
Veh iläinen-Julkunen K. Associa ions ac oss spa ial pa e ns o disease inci-
dences, socio-demog aphics, and land use in Finland 1991–2010. Scand J Public
Heal h. 2015;43:356–63.
14. Sambell CE, Holland GJ, Haslem A, Benne AF. Di e se land-uses shape new bi d
communi ies in a changing u al egion. Biodi e s Conse . 2019;28:3479–96.
15. B own BB, Yamada I, Smi h KR, Zick CD, Kowaleski-Jones L, Fan JX. Mixed land use
and walkabili y: a ia ions in land use measu es and ela ionships wi h BMI,
o e weigh , and obesi y. Heal h Place. 2009;15:1130–41.
16. Miles R, Cou s C, Mohamadi A. Neighbo hood u ban o m, social en i onmen ,
and dep ession. J U ban Heal. 2012;89:1–18.
17. Melis G, Gelo mino E, Ma a G, Fe acin E, Cos a G. The e ec s o he u ban buil
en i onmen on men al heal h: a coho s udy in a la ge No he n I alian ci y. In J
En i on Res Public Heal h. 2015;12. h ps://doi.o g/10.3390/ije ph121114898.
18. Wu W, Chen WY, Yun Y, Wang F, Gong Z. U ban g eenness, mixed land-use, and
li e sa is ac ion: e idence om esiden ial loca ions and wo kplace se ings in
Beijing. Landsc U ban Plan. 2022;224:104428.
19. Bo doloi R, Mo e A, Sa ka PP, Mallika juna C. Quan ifica ion o land use di e si y
in he con ex o mixed land use. P ocedia - Soc Beha Sci. 2013;104:563–72.
20. Guloksuz S, an Os J, Ru en BPF. The exposome pa adigm and he complexi ies
o en i onmen al esea ch in psychia y. JAMA Psychia y. 2018;75:985–6.
21. Ohanyan H, Po engen L, Huss A, T aini E, Beulens JWJ, Hoek G, e al. Machine
lea ning app oaches o cha ac e ize he obesogenic u ban exposome. En i on
In . 2022;158:107015.
22. Ohanyan H, Po engen L, Kaplani O, Huss A, Hoek G, Beulens JWJ, e al. Asso-
cia ions be ween he u ban exposome and ype 2 diabe es: Resul s om pena-
lised eg ession by leas absolu e sh inkage and selec ion ope a o and andom
o es models. En i on In . 2022;170:107592.
23. Lydiane A, Lü zen P, Ma c C-H, Xa ie B, Lise G-A, Valé ie S, e al. A sys ema ic
compa ison o linea eg ession–based s a is ical me hods o assess exposome-
heal h associa ions. En i on Heal h Pe spec . 2016;124:1848–56.
24. Mai e L, Guimbaud J-B, Wa embou g C, Güil-Oum ai N, Pe one PM, Chadeau-
Hyam M, e al. S a e-o - he-a me hods o exposu e-heal h s udies: esul s om
he exposome da a challenge e en . En i on In . 2022;168:107422.
25. Hosko ec L, Benka-Coke W, Se e son R, Magzamen S, Wilson A. Model choice o
es ima ing he associa ion be ween exposu e o chemical mix u es and heal h
ou comes: a simula ion s udy. PLoS One. 2021;16:e0249236.
26. Rose RJ, Sal a o e JE, Aal onen S, Ba PB, Bogl LH, Bye s HA, e al. FinnTwin12
coho : an upda ed e iew. Twin Res Hum Gene . 2019;22:302–11.
27. Kokko K, Pulkkinen L. Unemploymen and psychological dis ess: media o
e ec s. J Adul De . 1998;5:205–17.
28. Depue RA, Sla e JF, Wol s e e -Kausch H, Klein D, Gople ud E, Fa D. A beha-
io al pa adigm o iden i ying pe sons a isk o bipola dep essi e diso de : A
concep ual amewo k and fi e alida ion s udies. J Abno m Psychol.
1981;90:381–437.
29. Bucholz KK, Cado e R, Cloninge CR, Dinwiddie SH, Hesselb ock VM, Nu nbe ge
JI, e al. A new, semi-s uc u ed psychia ic in e iew o use in gene ic linkage
s udies: a epo on he eliabili y o he SSAGA. J S ud Alcohol. 1994;55:149–58.
30. U ban A las LCLU 2012. 2021. h p://land.cope nicus.eu/local/u ban-a las/u ban-
a las-2012/ iew (accessed 7 Sep 2022).
31. Lo Papa G, Pale mo V, Dazzi C. Is land-use change a cause o loss o pedodi-
e si y? The case o he Mazza one s udy a ea, Sicily. Geomo phology.
2011;135:332–42.
32. Huppe z C, Ba els M, de Geus EJC, an Beijs e eld CEM, Rose RJ, Kap io J, e al.
The e ec s o pa en al educa ion on exe cise beha io in childhood and you h: a
s udy in Du ch and Finnish wins. Scand J Med Sci Spo s. 2017;27:1143–56.
33. Kan a dzic M. Da a mining: concep s, models, me hods, and algo i hms. John
Wiley & Sons, Inc., 2019 h ps://www.wiley.com/en-us/Da a+Mining%3A
+Concep s%2C+Models%2C+Me hods%2C+and+Algo i hms%2C+3 d
+Edi ion-p-9781119516071.
34. Liao M, Li Y, Kiani a d F, Obi E, A cona S. Clus e analysis and i s applica ion o
heal hca e claims da a: a s udy o end-s age enal disease pa ien s who ini ia ed
hemodialysis. BMC Neph ol. 2016;17:25.
35. Kassamba a A. P ac ical guide o clus e analysis in R. 1s ed. STHDA, 2017.
36. Rousseeuw PJ. Silhoue es: a g aphical aid o he in e p e a ion and alida ion o
clus e analysis. J Compu Appl Ma h. 1987;20:53–65.
37. Len e s V, Po engen L, Rignell-Hydbom A, Jönsson BAG, Lindh CH, Pie sma AH,
e al. P ena al ph hala e, pe fluo oalkyl acid, and o ganochlo ine exposu es and
e m bi h weigh in h ee bi h coho s: mul i-pollu an models based on elas ic
ne eg ession. En i on Heal h Pe spec . 2016;124:365–72.
38. Zou H, Has ie T.Regula iza ion and a iable selec ion ia he elas ic ne .J R S a
Soc Se B (S a is ical Me hodol.). 2005;67:301–20.
39. F iedman JH, Has ie T, Tibshi ani R. Regula iza ion pa hs o gene alized linea
models ia coo dina e descen . J S a So w. 2010;33:1–22.
40. Chen T, Gues in C. XGBoos : a scalable ee boos ing sys em. P oc 22nd ACM
SIGKDD In Con Knowl Disco Da a Min 2016. h ps://doi.o g/10.1145/2939672.
41. Chen Z-Y, Zhang T-H, Zhang R, Zhu Z-M, Yang J, Chen P-Y, e al. Ex eme g adien
boos ing model o es ima e PM2.5 concen a ions wi h missing-filled sa elli e
da a in China. A mos En i on. 2019;202:180–9.
42. Wang L, Zhao C, Liu X, Chen X, Li C, Wang T, e al. Non-linea e ec s o he buil
en i onmen and social en i onmen on bus use among olde adul s in China: an
applica ion o he XGBoos model. In J En i on Res Public Heal h. 2021;18. h ps://
doi.o g/10.3390/ije ph18189592.
43. Wilson S. Pa allel Bayesian op imiza ion o hype pa ame e s. 2022 h ps://c an. -
p ojec .o g/web/packages/Pa BayesianOp imiza ion/Pa BayesianOp imiza ion.pd .
44. Rincou S-L, Michiels S, D ubay D. Complex disease indi idual molecula cha -
ac e iza ion using infini e spa se g aphical independen componen analysis.
Cance In . 2022;21:11769351221105776.
45. Lundbe g S, Lee S-I. A unified app oach o in e p e ing model p edic ions. 2017.
h ps://doi.o g/10.48550/a xi .1705.07874.
46. Liu Y, Jus A, Maye M. SHAP o xgboos : SHAP plo s o “XGBoos ”.R Packag
e sion 010 2020.
47. Chen Y, Zhang X, G ekousis G, Huang Y, Hua F, Pan Z, e al. Examining he
impo ance o buil and na u al en i onmen ac o s in p edic ing sel - a ed
heal h in olde adul s: an ex eme g adien boos ing (XGBoos ) app oach. J Clean
P od. 2023;413:137432.
48. Shen Y, Ta N, Liu Z. Job-housing dis ance, neighbo hood en i onmen , and
men al heal h in subu ban Shanghai: a gende di e ence pe spec i e. Ci ies.
2021;115:103214.
49. McLoughlin C, McLoughlin A, Jain S, Abdalla A, Cooney J, MacHale S. The
subu ban-ci y di ide: an e alua ion o eme gency depa men men al heal h
p esen a ions ac oss wo cen es. I J Med Sci. 2021;190:1523–8.
50. Pelg ims I, De leesschauwe B, Guyo M, Keune H, Naw o TS, Remmen R, e al.
Associa ion be ween u ban en i onmen and men al heal h in B ussels, Belgium.
BMC Public Heal h. 2021;21:635.
51. Wang R, Xue D, Liu Y, Chen H, Qiu Y. The ela ionship be ween u baniza ion and
dep ession in China: he media ing ole o neighbo hood social capi al. In J
Equi y Heal h. 2018;17:105.
52. Niu B, Ge D, Yan R, Ma Y, Sun D, Lu M, e al. The e olu ion o he in e ac i e
ela ionship be ween u baniza ion and land-use ansi ion: a case s udy o he
Yang ze Ri e Del a. Land. 2021;10. h ps://doi.o g/10.3390/land10080804.
53. Sampson L, E man CK, Galea S. U baniza ion, u banici y, and dep ession: a
e iew o he ecen global li e a u e. Cu Opin Psychia y. 2020;33. h ps://
doi.o g/10.1097/YCO.0000000000000588.
54. S e en A, Thom J, Jacobi F, Hols iege J, Bä zing J. T ends in p e alence o
dep ession in Ge many be ween 2009 and 2017 based on na ionwide ambula-
o y claims da a. J A ec Diso d. 2020;271:239–47.
55. Mo ozo PV. Men al heal h and u baniza ion: a Russian pe spec i e. Cu Opin
Psychia y. 2018;31. h ps://doi.o g/10.1097/YCO.0000000000000415.
Z. Wang e al.
778
Jou nal o Exposu e Science & En i onmen al Epidemiology (2024) 34:770 – 779
