Adolescent Loneliness Trends and Contextual Correlates Across 38 Countries from 2000 to 2022

Abstract

Purpose.

This study estimates multinational trends and correlates of adolescent loneliness.

Methods.

Data from the Programme for International Student Assessment administered to 1,267,476 adolescents aged 15–16 across 38 Organization for Economic Cooperation and Development-member countries in 2000, 2003, 2012, 2015, 2018, and 2022 were used. Surveys included six questions on a four-point Likert-type scale about loneliness in the school context (range: 1–4). Multilevel piecewise regression was used to estimate trends in loneliness scores during three periods: 2000–2012, 2012–2015, and 2015–2022. We estimated associations between loneliness and the following country-level factors: COVID-19 mortality, public health distancing measures, internet use, unemployment rate, income inequality (Gini index), and peacefulness. Models of COVID-19-related factors were restricted to 2022. The remaining models included three levels (individuals, country-years, and countries), allowing estimation of time-invariant between-country and longitudinal within-country effects.

Results.

We observed a 0.007-point decrease in mean loneliness scores per year from 2000–2012 (95%CI: −0.022, 0.008), though this was not statistically significant. Loneliness scores increased by 0.045-points per year from 2012–2015 (95%CI: 0.035, 0.054) and by 0.017-points per year from 2015–2022 (95%CI: 0.008, 0.025). In within-country models, a one-hour higher country-level weekly internet use and one percentage-point higher unemployment rate were associated with 0.012 (95%CI: −0.021, −0.003) and 0.006 (95%CI: −0.011, −0.001) lower scores per year, respectively. Remaining contextual measures were not associated with loneliness.

Conclusions.

Loneliness substantially increased from 2012–2015, followed by smaller annual increases until 2022. Within countries, greater increases in unemployment rates and internet use over time are associated with lower loneliness.

Loneliness—a subjective feeling that arises when “belonginess needs are insufficiently met”¹—is particularly impactful during adolescence.² Loneliness among adolescents may precede worse health behaviors,³ physiological stress responses⁴, and psychological sequelae including depression and other mental distress.^3,5 Loneliness among adolescents may also harm cardiovascular health, self-reported health, and risk of depression longitudinally.²

Researchers have documented substantial and rising levels of loneliness among adolescents in the past two decades.^6,7 Twenge et al. identified a near doubling of loneliness prevalence from 2012 to 2018 among adolescents from 36 countries that participated in the Programme for International Student Assessment (PISA).⁶ Updated estimates of loneliness trends are needed, especially given the rise of social connection as a public health priority⁸ and the scarcity of multinational studies using data that were collected after the onset of COVID-19.^6,7,9

Widespread shifts in loneliness may reflect broader societal influences. Prior research primarily focuses on individual predictors of loneliness—including personality differences, peer victimization, and sociodemographic factors.¹⁰ These studies often frame social connection as resting within individual agency or fixed traits rather than societal conditions. Yet, group-level cultural values, economic conditions, and social norms can influence key drivers of social connection, such as trust in others and cooperation.^11–14 Therefore, considering both individual predictors and societal factors is critical for the comprehensive understanding of factors that contribute to loneliness.

The COVID-19 pandemic and contemporaneous economic and sociopolitical shifts present a unique opportunity to study recent trends and societal correlates of loneliness.¹⁵ There is some evidence that the pandemic impacted loneliness—either through increased mortality rates,¹⁶ or through social distancing measures that limited in-person contact¹⁵ and exacerbated mental health challenges such as social anxiety.¹⁷ The digital world of adolescents has also transformed in the past two decades. Researchers have speculated that widespread internet use has had mixed effects on loneliness by replacing some in-person interactions with online ones, yet offering platforms for connection.^6,18,19 Other potential societal contributors of loneliness include economic and sociopolitical conditions such as income inequality, unemployment, and peacefulness. Higher income inequality and unemployment rates may erode catalysts of social connection such as well-being and trust by worsening economic and health inequities.^11–14,20 Peacefulness may correlate with loneliness by shaping social embeddedness and, in times of conflict, contributing to trauma or loss.^14,21

We aimed to 1) characterize temporal trends in adolescent loneliness from 2000 to 2022 and 2) examine country-level correlates of loneliness by estimating associations (during selected years) between loneliness and the following group-level indicators: COVID-19 mortality, public health distancing laws, internet use, unemployment, income inequality, and peace. Our findings may clarify the need to address adolescent loneliness and generate hypotheses regarding its societal determinants.

METHODS

Data source and study population

The study population included 15- and 16-year-old students who completed the Organization for Economic Cooperation and Development (OECD) PISA survey. The survey uses a two-stage stratified sampling strategy to obtain nationally representative data on scholastic achievement and well-being. Schools within a country were grouped into mutually exclusive strata based on country-specific factors (e.g. region, urbanicity), and selected from these strata according to probabilities proportional to the number of 15-year-olds they enrolled.²² PISA requires a weighted school response rate of 85%, and a student response rate of 80% across responding schools. If a school does not respond, it is replaced with a predetermined replacement school. Students within OECD-member countries who were administered the sense of belonging questionnaire in 2000, 2003, 2012, 2015, 2018, or 2022 were considered (N=1,319,544 students, 38 countries). We excluded students who skipped any loneliness questions (N=52,068), leaving a final population of 1,267,476 students. As data are publicly available and anonymized, this study was considered non-human subjects research and did not require Institutional Review Board approval.

Measures

Loneliness

Each survey included six questions about loneliness at school: “I feel like an outsider (or left out of things),” “I make friends easily” (reverse-coded), “I feel like I belong” (reverse-coded), “I feel awkward and out of place,” “Other students seem to like me” (reverse-coded), and “I feel lonely.” Responses were on a 4-point Likert-type scale ranging from “strongly disagree” (1) to “strongly agree” (4), with higher scores indicating higher loneliness. In 2022, students were randomly assigned five of the six loneliness questions. For each survey year, we computed a mean loneliness score (range: 1–4) by averaging the scores of all loneliness questions that were administered. The six-item scale has a Cronbach’s alpha of 0.82, while the five-item scales have values ranging from 0.78 to 0.81, depending on which item is excluded. The scale correlates with negative affect (r: 0.18 to 0.46), positive affect (r: −0.29 to −0.47), and life satisfaction (r: −0.20 to −0.45).⁶

Twenge and colleagues conducted measurement invariance analyses of the PISA loneliness scale within countries over time, finding full invariance in 12 countries, partial in 12 countries, and failure to meet configural invariance in 13 countries—five of which were included in the present analysis (France, Iceland, Japan, Mexico, South Korea).⁶ To account for variability in measurement invariance, we will present loneliness trends within each country. To determine whether the scale measures similar constructs of loneliness across cultures, we also analyzed measurement invariance across countries.

Country-level Correlates of Loneliness

We included the following country-level correlates: COVID-19 cumulative mortality rates, public health distancing laws, frequency of internet use, unemployment rate, income inequality, and peacefulness. COVID-19 cumulative mortality was defined as the country-level total deaths per 10,000 population from the start of the pandemic to December 31, 2022. Data were from the World Health Organization.²³ Public health distancing was defined using the Oxford Stringency Index (hereafter, “stringency index”), a 1–100 scale composed of nine domains of public health responses to COVID-19.²⁴ To account for policy differentiation by vaccination status, we used a weighted stringency index, which combines indices for vaccinated and non-vaccinated individuals according to the proportion of the country’s population that is vaccinated. We averaged each country’s daily weighted stringency index values across 2022 to produce a country-specific annual index.

Weekly frequency of internet use was obtained from two questions about weekday and weekend use: “On a typical weekday [weekend], for how long do you use the internet outside of school?”. We converted response options from the original question into approximate hours (“No time”=0, “1–30 minutes”=0.25, “31–60 minutes”=0.75, “1–2 hours”=1.5, “2–4 hours”=3, “4–6 hours”=5, “>6 hours”=7). Each students’ typical internet use on the weekdays and weekends was multiplied by five and two, respectively, and then summed together. Students’ weekly internet use frequencies were then averaged within each country-year. Because internet use questions were unavailable prior to 2012, and changed in 2022, only 2012 through 2018 surveys were included.

We obtained unemployment rate, income inequality (Gini index), and per capita gross domestic product (GDP) for each country and year from the World Bank.²⁵ The unemployment rate is defined as the average yearly percentage of the total labor force of a country that is without work and is actively seeking work. The Gini index measures the extent to which the distribution of income among individuals within an economy deviates from a scenario of perfectly equal distribution. The Gini index ranges from 0 (perfect equality) to 100 (complete inequality) and was available from 2000 to 2018.

Peace was measured using the Global Peace Index (GPI), a measure of the peacefulness of countries developed by the Institute for Economics and Peace.²⁶ The index is based on 23 qualitative and quantitative indicators spanning three domains: societal safety and security, domestic and international conflict, and militarization. The GPI ranges from 1 to 5, with higher scores indicating lower levels of peacefulness. The GPI was available from 2012 to 2022.

Covariates

Sex, maternal education, and paternal education were included as covariates. Sex was ascertained using the question: “Are you male or female?”. PISA does not provide additional response categories or ask about gender identity. Maternal and paternal education were drawn from the question: “What is the highest level of schooling completed by your mother [father]?”. The question contained five response options, ranging from less than the International Standard Classification of Education (ISCED) level 1 to completion of ISCED 3A. ISCED level 1 corresponds to primary education. ISCED level 3A corresponds to upper secondary education that is designed to prepare students for tertiary education.

Weights

To account for variability in the population size of the included countries, weights were applied using methods described by Luyten (2024).²⁷. Briefly, the student weights provided by PISA—which indicate how many students a respondent represents—were summed within each country and year. Standardized weights were then computed using the following equation:

$$w_j=\frac{{\mathit{\text{population}}}_j}{{\mathit{\text{population}}}_t}\times 38$$

Where:

w_j: Standardized weight for country j

population_j: Average size across survey years of the populations represented by the samples of country j

population_t: Sum of the populations represented by all 38 countries

Thus, each country’s contribution to prevalence estimates and analytic models is scaled according to the size of its student population.

Statistical analysis

Time trends of loneliness

Weighted mean loneliness scores were described for each survey year, overall and by country and world region. We used multilevel, piecewise regression to estimate trends of loneliness across survey years within distinct time periods. The multilevel model contains nested random effects to model the variation across students (level-1), country-years (level-2), and countries (level-3). The piecewise regression technique recognizes that a single linear regression model may not capture changing trends at different time points. We estimated an intercept and three linear slopes for calendar year, one for each of the following periods: 2000–2012, 2012–2015, and 2015–2022. We selected these inflection points based on prior work describing substantial increases in loneliness from 2012–2015, following and preceding smaller changes in loneliness, using PISA data from several OECD-member countries.⁶ Hosozawa et. al similarly observed “a marked escalation [in loneliness] beginning in 2012.”⁷ Our model included weights and covariates for sex, maternal education, and paternal education.

To account for the possibility that the school loneliness scale combines distinct constructs of loneliness, we conducted a sensitivity analysis using alternative combinations of items to represent the following constructs: perceived sociability,²⁸ perceived lack of belonging, positively-worded items, negatively-worded items, and a single-item direct measure of loneliness (details in supplement).²⁹

Country-level correlates of loneliness

Analyses of COVID-19-related variables were restricted to 2022 as this was the only post-pandemic year in which PISA surveys were administered. For this analysis, participants (level-1) were nested within countries (level-2).

In analyses of country-level correlates for which we had multiple years of data available, the model used the same structure as the trends analysis, with students (level-1), country-years (level-2), and countries (level-3). We applied Fairbrother’s extension of multilevel modeling to separately estimate between-country and within-country associations between each correlate and loneliness.³⁰ The between-country effects reflect variations in loneliness that are associated with absolute, cross-national differences in a country-level factor. To include between-country effects, country-level correlates were grand-mean centered. For example, a country’s unemployment rate was included as a level-3 covariate for the deviation of the country’s mean unemployment rate across survey years from the mean unemployment rate in the full sample. The within-country effects reflect how yearly deviations in a country-level factor from its average are associated with corresponding changes in loneliness. To include within-country effects, country-level correlates were group-mean centered. For example, a country’s unemployment rate for a given year was included as the level-2 deviation from the country’s mean unemployment rate across all years. We also included a variable for year, centered on the first time point of observation, to account for simultaneous time trends. The model took the following form:

$${\text{Y}}_{\text{itj}}={\beta }_{\text{0}}+{\beta }_{\text{1}}{\text{c}}_{\text{itj}}+{\beta }_{\text{2}}{\overline{\text{x}}}_{\text{j}}+{\beta }_{\text{3}}{\text{x}}_{\text{tjM}}+{\beta }_{\text{4}}{\text{time}}_{\text{tj}}+{\text{u}}_{\text{j}}+{\text{u}}_{\text{tj}}+{\text{e}}_{\text{itj}}$$

Where:

Y_itj: mean loneliness score for student i in year t in country j

c_itj: covariates for student i in year t in country j

${\overline{\text{x}}}_{\text{j}}$: (mean unemployment across all years in country j)–(overall unemployment)

x_tjM: (unemployment in year t in country j)–(mean unemployment across all years in country j)

time_tj: year

u_j, u_tj: random intercepts for country and country-year

e_itj: individual error

All models included weights and covariates for sex, maternal education, and paternal education. Models for unemployment and Gini index were additionally adjusted for per capita GDP. We modeled year linearly to simplify interpretations. To determine how using a linear parameterization of time affected results, we conducted a sensitivity analysis with time modeled in segments. Multicollinearity was examined using variance inflation factors, with values <5 considered acceptable.³¹ Models used the “lme4” package in RStudio version 4.3.2.³²

RESULTS

The school loneliness measure demonstrated configural and metric but not scalar invariance across countries (Supplementary Table 1). Descriptive statistics for country-level correlates of loneliness are available in Supplementary Table 2.

Time trends of loneliness

School loneliness increased from 2000–2022, with the largest increase from 2012–2015 (Table 1). This trend was consistent across world regions except for East Asia (Figure 1). In 2022, loneliness was the lowest in Western Europe and East Asia. In adjusted models, mean loneliness scores showed a slight and non-statistically significant decrease of 0.007 points each year from 2000–2012 (95%CI: −0.022, 0.008) (Table 2). From 2012–2015, mean loneliness scores increased by 0.045 points each year (95%CI: 0.035, 0.054). Loneliness also increased from 2015–2022, but to a lesser extent (B: 0.017, 95%CI: 0.008, 0.025). These trends were similar when examining alternative operationalizations of loneliness (Supplementary Figure 1, Supplementary Table 3).

Table 1.

Mean Loneliness Score Over Time Within OECD-Member Countries¹

Country	Region	2000 Mean (SD)	2003 Mean (SD) [Percent difference from prior survey]	2012 Mean (SD) [Percent difference from prior survey]	2015 Mean (SD) [Percent difference from prior survey]	2018 Mean (SD) [Percent difference from prior survey]	2022 Mean (SD) [Percent difference from prior survey]
Worldwide		1.82 (0.003)	1.77 (0.001) [−2.38]	1.77 (0.002) [0.03]	1.99 (0.002) [12.26]	1.97 (0.002) [−0.84]	1.98 (0.002) [0.34]
Australia	Australia and New Zealand	1.82 (0.009)	1.76 (0.004) [−3.49]	1.91 (0.005) [8.3]	2.03 (0.005) [6.6]	2.09 (0.005) [2.77]	2.07 (0.005) [−0.84]
New Zealand		1.8 (0.011)	1.77 (0.007) [−1.7]	1.88 (0.01) [6.16]	2.02 (0.008) [7.71]	2.08 (0.007) [2.79]	2.09 (0.016) [0.63]
Japan	East Asia	1.97 (0.01)	2.04 (0.007) [3.5]	1.84 (0.008) [−9.81]	1.96 (0.007) [6.74]	1.96 (0.007) [−0.04]	1.81 (0.007) [−7.48]
South Korea		1.99 (0.009)	1.96 (0.006) [−1.68]	1.87 (0.008) [−4.54]	1.84 (0.007) [−1.39]	1.82 (0.007) [−1.01]	1.81 (0.007) [−0.54]
Czech Republic	Eastern Europe	1.91 (0.008)	1.87 (0.005) [−1.77]	1.89 (0.008) [1.1]	2.06 (0.007) [8.72]	2.1 (0.007) [2.01]	2.09 (0.006) [−0.24]
Estonia		--	--	1.88 (0.009)	1.97 (0.008) [4.78]	2.04 (0.008) [3.64]	2.02 (0.007) [−1.03]
Hungary		1.72 (0.009)	1.73 (0.007) [0.77]	1.74 (0.009) [0.64]	1.91 (0.008) [9.32]	1.91 (0.008) [0.08]	1.84 (0.007) [−3.43]
Latvia		1.96 (0.01)	1.86 (0.006) [−5.39]	1.85 (0.009) [−0.16]	2.06 (0.008) [11.29]	2.12 (0.008) [2.78]	2.1 (0.008) [−0.98]
Lithuania		--	--	1.76 (0.01)	2.19 (0.01) [24.03]	2.09 (0.008) [−4.42]	2 (0.008) [−4.11]
Poland		2 (0.011)	1.84 (0.007) [−7.97]	1.86 (0.009) [0.98]	2.11 (0.01) [13.42]	2.11 (0.008) [0.09]	2.14 (0.007) [1.3]
Slovak Republic		--	1.84 (0.005)	1.95 (0.009) [6.08]	2.09 (0.007) [7.23]	2.11 (0.008) [1]	2.05 (0.008) [−3.01]
Slovenia		--	--	1.78 (0.008)	2 (0.008) [12.75]	2.02 (0.007) [0.54]	1.92 (0.007) [−4.84]
Chile	Latin America	1.71 (0.01)	--	1.71 (0.008) [−0.09]	1.93 (0.008) [13.05]	1.99 (0.008) [2.95]	2.06 (0.009) [3.78]
Colombia		--	--	1.77 (0.006)	2.14 (0.007) [21.26]	2.02 (0.007) [−5.69]	2 (0.007) [−1.05]
Costa Rica		--	--	1.66 (0.005)	2.07 (0.01) [24.32]	1.94 (0.008) [−6.24]	2.01 (0.009) [3.41]
Mexico		1.75 (0.009)	1.72 (0.003) [−1.73]	1.74 (0.003) [1.34]	2.05 (0.008) [17.65]	1.97 (0.008) [−3.91]	2.08 (0.009) [5.7]
Israel	Middle East and North Africa	1.61 (0.011)	--	1.62 (0.009) [0.51]	--	--	--
Turkey		--	1.97 (0.007)	1.8 (0.01) [−8.68]	2.28 (0.01) [26.44]	2.06 (0.008) [−9.41]	2.16 (0.008) [4.82]
Canada	United States and Canada	1.76 (0.004)	1.78 (0.003) [0.87]	1.83 (0.005) [3.2]	2.04 (0.004) [11.3]	2.08 (0.004) [2.05]	2.06 (0.004) [−1.24]
United States		1.85 (0.012)	--	1.85 (0.009) [−0.08]	1.99 (0.008) [7.71]	2.11 (0.009) [5.93]	2.1 (0.009) [−0.43]
Austria	Western Europe	1.69 (0.011)	1.6 (0.008) [−5.36]	1.51 (0.009) [−5.63]	1.77 (0.009) [17.1]	1.78 (0.008) [0.55]	1.73 (0.008) [−2.68]
Belgium		1.9 (0.008)	1.88 (0.005) [−0.55]	1.82 (0.007) [−3.57]	1.92 (0.006) [5.55]	1.94 (0.006) [0.94]	1.91 (0.006) [−1.48]
Denmark		1.74 (0.01)	1.76 (0.007) [1.57]	1.74 (0.008) [−1.46]	1.88 (0.008) [8.37]	1.86 (0.007) [−1.29]	1.89 (0.008) [1.94]
Finland		1.77 (0.009)	1.77 (0.006) [−0.37]	1.80 (0.007) [1.93]	1.89 (0.008) [5.12]	1.96 (0.008) [3.7]	1.92 (0.006) [−2.17]
France		1.87 (0.01)	1.85 (0.007) [−0.93]	1.89 (0.009) [2.05]	1.96 (0.007) [4.18]	2.06 (0.007) [4.74]	1.96 (0.008) [−4.94]
Germany		1.73 (0.01)	1.68 (0.008) [−2.77]	1.61 (0.01) [−4.27]	1.81 (0.009) [12.24]	1.8 (0.009) [−0.41]	1.81 (0.008) [0.29]
Greece		1.81 (0.009)	1.74 (0.006) [−3.6]	1.79 (0.009) [2.65]	1.86 (0.008) [3.91]	1.93 (0.007) [3.73]	1.97 (0.007) [2.11]
Iceland		1.75 (0.012)	1.72 (0.009) [−1.58]	1.66 (0.012) [−3.6]	1.89 (0.013) [13.6]	1.96 (0.013) [4.09]	1.88 (0.011) [−4.15]
Ireland		1.76 (0.011)	1.73 (0.007) [−1.53]	1.79 (0.009) [3.41]	1.94 (0.008) [8.29]	2.04 (0.007) [4.9]	2 (0.007) [−1.76]
Italy		1.79 (0.009)	1.74 (0.004) [−2.38]	1.77 (0.003) [1.73]	1.90 (0.005) [7.07]	1.92 (0.005) [1.06]	1.95 (0.006) [1.77]
Luxembourg		1.79 (0.013)	1.68 (0.008) [−6.08]	1.71 (0.01) [1.28]	1.9 (0.009) [11.23]	1.93 (0.009) [1.82]	--
Netherlands		1.76 (0.011)	1.77 (0.006) [0.71]	1.76 (0.008) [−0.3]	1.82 (0.007) [3.42]	1.82 (0.008) [−0.13]	1.87 (0.008) [2.76]
Norway		1.74 (0.011)	1.67 (0.008) [−3.67]	1.75 (0.01) [4.61]	1.85 (0.009) [5.79]	1.81 (0.008) [−2.42]	1.84 (0.008) [1.87]
Portugal		1.76 (0.008)	1.72 (0.006) [−2.29]	1.73 (0.008) [0.44]	1.89 (0.007) [9.34]	1.87 (0.007) [−0.88]	1.88 (0.007) [0.23]
Spain		1.79 (0.008)	1.68 (0.004) [−6]	1.61 (0.004) [−4.17]	1.7 (0.008) [5.55]	1.7 (0.003) [−0.51]	1.78 (0.004) [5.19]
Sweden		1.67 (0.01)	1.65 (0.007) [−1.31]	1.74 (0.01) [5.42]	1.97 (0.01) [13.26]	1.97 (0.009) [−0.16]	1.92 (0.009) [−2.42]
Switzerland		1.7 (0.009)	1.7 (0.006) [0.21]	1.63 (0.006) [−4.42]	1.78 (0.008) [9.43]	1.83 (0.008) [2.41]	1.78 (0.007) [−2.6]
United Kingdom		1.74 (0.007)	1.74 (0.005) [−0.28]	1.82 (0.006) [4.96]	1.99 (0.005) [9.14]	2.09 (0.005) [4.9]	2.07 (0.005) [−0.92]

Abbreviations: OECD, Organization for Economic Cooperation and Development; SD, Standard deviation

^1.Mean loneliness scores are computed using weights to account for country population.

Figure 1. Mean Loneliness Score from 2000 to 2022, Overall and by World Region.

Mean loneliness scores are computed using weights to account for country population. Some regions only included two countries and thus sample sizes were small. For example, East Asia only includes Japan and South Korea.

Table 2.

Piecewise multilevel regression of mean loneliness score associated with survey year

Time period	Crude model Beta estimate (95% CI)1	Adjusted Model Beta estimate (95% CI)1,2
2000–2012	−0.009 (−0.024, 0.006)	−0.007 (−0.022, 0.008)
2012–2015	0.043 (0.034, 0.052)	0.045 (0.035, 0.054)
2015–2022	0.016 (0.007, 0.024)	0.017 (0.008, 0.025)

^1.Models use piecewise multilevel linear regression with knots at the years 2012 and 2015, a random intercept for country, and weights to account for differences in country population. Model includes 1,169,308 individuals clustered within 38 countries.

^2.Estimates are adjusted for sex, maternal education, and paternal education. Maternal and paternal education are based on the International Standard Classification of Education. Levels of education are coded continuously with a range of 1–5, with higher numbers indicating higher levels of education.

Country-level correlates of loneliness

Estimates for associations of country-level COVID-19 mortality and stringency index with loneliness in 2022 were close to zero and not statistically significant (Table 3).

Table 3.

Country-level, COVID-19-related correlates of mean school loneliness, 2022

	Crude model Beta estimate (95% CI)1	Adjusted Model Beta estimate (95% CI)1,2
COVID-19 cumulative mortality rate1 (n=235,779, c=35)	0.002 (−0.001, 0.006)	0.002 (−0.001, 0.006)
Oxford Stringency Index1 (n=235,779, c=35)	−0.001 (−0.006, 0.005)	−0.001 (−0.007, 0.005)

^1.Variable is grand-mean centered. Cumulative mortality rate is calculated as the total deaths per 10,000 population that occurred from the beginning of the COVID-19 pandemic to December 31, 2022.

^2.Estimates are adjusted for sex, maternal education, and paternal education.

Detailed results of three-level multilevel models, including variance components at the country, period, and individual level, are in Supplementary Tables 4–7. Most variance in loneliness was explained at the individual level, followed by period and country levels (Supplementary Tables 4–7). Within countries, greater increases in internet use over time were associated with larger decreases in loneliness; each additional hour of weekly internet use within a country was associated with a 0.012-point reduction in loneliness scores per year (95%CI: −0.021, −0.003) (Table 4). There was no association between cross-country differences in internet use and loneliness. A higher unemployment rate was associated with lower loneliness at within-country (B: −0.006, 95%CI: −0.011, −0.001) but not between-country (B: 0.001, 95%CI: −0.009, 0.011) levels. The Gini index and GPI were not associated with loneliness at between-country (Gini: B: 0.006, 95%CI: −0.004, 0.017; GPI: B: 0.080, 95%CI: −0.018, 0.177) or within-country (Gini: B: 0.006, 95%CI: −0.021, 0.033); GPI: B: 0.132, 95%CI: −0.080, 0.343) levels. In sensitivity analyses modeling time in segments (pre-2012, 2012–2015, post-2015), estimates were slightly attenuated compared to main models (Supplementary Table 8).

Table 4.

Country-level correlates of mean school loneliness across multiple survey years

	Crude model beta estimate (95% CI)1	Main model beta estimate (95% CI)1	Expanded model beta estimate (95% CI)1
Internet Use (weekly hours) (n=446,623, c-y=76, c=26)2 2012–2018
Period Variables
Change in Internet use3	−0.019 (−0.022, −0.015)	−0.012 (−0.021, −0.003)
Year		0.013 (−0.002, 0.028)
Context variables
Mean Internet use4	0.010 (−0.002, 0.022)	0.009 (−0.004, 0.022)
Unemployment Rate (n=1,169,308, c-y=208, c=38)2 2000–2022
Period Variables
Change in unemployment5	−0.011 (−0.017, −0.005)	−0.006 (−0.011, −0.001)	−0.006 (−0.011, −0.001)
Change in gross domestic product3			0.000 (0.000, 0.000)
Year		0.009 (0.008, 0.011)	0.010 (0.007, 0.013)
Context variables
Mean unemployment rate4	0.002 (−0.008, 0.012)	0.001 (−0.009, 0.011)	−0.003 (−0.013, 0.007)
Mean gross domestic product4			0.000 (0.000, 0.000)
Gini Index (n=317,243, c-y=60, c=14)2 2000–2018
Period Variables
Change in Gini index3	−0.032 (−0.065, 0.001)	0.006 (−0.021, 0.033)	0.008 (−0.019, 0.035)
Change in gross domestic product3			0.000 (0.000, 0.000)
Year		0.013 (0.009, 0.017)	0.007 (−0.006, 0.020)
Country variables
Mean Gini index4	0.007 (−0.005, 0.018)	0.006 (−0.004, 0.017)	0.006 (−0.007, 0.019)
Mean gross domestic product4			0.000 (0.000, 0.000)
Global Peace Index (n=874,680, c-y=140, c=35)2 2012–2022
Period Variables
Change in GPI3	0.033 (−0.232, 0.297)	0.132 (−0.080, 0.343)
Year		0.017 (0.013, 0.021)
Context variables
Mean GPI4	0.098 (0.007, 0.189)	0.080 (−0.018, 0.177)

^1.All models contain nested random effects to model the variation across students (level-1), country-years (level-2), and countries (level-3). All models are weighted to account for differences in country population. Main and expanded models are adjusted for sex, maternal education, and paternal education. Maternal and paternal education are based on the International Standard Classification of Education. Levels of education are coded continuously with a range of 1–5, with higher numbers indicating higher levels of education.

^2.Within each country-level factor analysis, n’s represent the number of individuals (Level-1), c-y represents the number of country-years (Level-2), and c represents the number of countries (Level-3).

^3.Variable is group-mean centered.

^4.Variable is grand-mean centered.

DISCUSSION.

School loneliness increased worldwide from 2000 to 2022, with most notable increases from 2012 to 2015. When examining within-country associations, on years when a country’s internet use and unemployment rates were higher than its average across the study period, loneliness scores were lower. There was no association between income inequality or peace with loneliness at between- or within-country levels. COVID-19-related indicators (mortality, stringency index) were not associated with loneliness.

Our trends analysis builds on the seminal work of Twenge et al. and Hosozawa et al. by incorporating data through 2022 and analyzing trends within distinct time periods.^6,7 Despite including a different set of countries than analyzed in either of these studies, we similarly observed that loneliness increased from 2000 to 2018. Although the increase in loneliness scores was small relative to the scale’s range, even modest changes in loneliness have been shown to impact health outcomes such as depression and self-rated health.³ Given loneliness is highly prevalent, small shifts in social connection at the population level may have meaningful impacts on mental health and other loneliness-related health outcomes.

Our approach of decomposing societal correlates of loneliness into between-country and within-country effects limits the comparability of our findings with previous studies that used standard multilevel models.³⁰ For instance, we identified a positive and non-significant relationship between cross-national differences in internet use and loneliness (i.e., the between-country effect), while within-country increases in internet use were linked to lower loneliness. In contrast, previous multilevel studies have identified a positive relationship between internet use and loneliness.⁶ When applying these multilevel models in sensitivity analyses, we similarly observed a positive and significant relationship between internet use and loneliness (Supplementary Table 9). These differences emphasize the need for careful methodological consideration when interpreting associations between societal factors and loneliness.

The mixed evidence on associations between internet use and loneliness have contributed to ongoing debates about how digitalization impacts social well-being. Apart from differences across studies related to modeling choices and exposure/outcome characterization that may have contributed to different findings, several conflicting thoughts have been put forward. Some argue that more expansive internet use hinders social connection by impacting social skills and shifting cultural norms, such as through the digitalization of interactions that would have otherwise been in-person.^6,18 These cultural shifts may be particularly impactful during adolescence, a developmental period marked by building social-emotional regulation and peer relationships.² In contrast to those studies, our within-country findings support the view that increased internet use enables more people to pursue in-person and virtual connections.^19,33 Higher country-level internet penetration rates have been linked to lower levels of loneliness,³⁴ and one study estimated that the average number of adult friendships grew from 2002 to 2007—a period marked by substantial increases in internet use.³⁵ Further research on specific internet use-related behaviors (e.g. social media use, negative content immersion) may offer deeper insights into how group-level digital behaviors shape social connection.

A higher unemployment rate than a country’s average was associated with lower loneliness scores, which is supported by other multilevel analyses of unemployment and adolescent loneliness.⁶ However, the relationship between unemployment and loneliness was small, and null after adjusting for GDP. Drawing from the sociological framework of time as a “network good,” higher unemployment rates may coincide with greater social connection by increasing the collective availability of free time.³⁶ The isolating effects of unemployment may also be mitigated when it is more common; unemployed individuals have reported lower levels of loneliness when country-level unemployment was high and thus their experience was more normative.³⁷

We found no within- or between-country associations between income inequality or peacefulness with loneliness. While the null findings for income inequality are supported by Twenge et al.,⁶ other studies have linked higher income inequality to loneliness among adults,³⁸ reductions in trust among adolescents and adults,²⁰ and increased bullying among adolescents.³⁹ We expected a negative relationship between peace and loneliness, given prior reports that experiencing violence and worrying about crime increases loneliness.^21,40 Overall, few studies have examined associations between group-level income inequality or peacefulness and loneliness. More research is needed to clarify inconsistencies among existing reports.

Our 2022-only analyses found no relationship between COVID-19 mortality or stringency index values with loneliness. A 2024 review of the longitudinal effects of COVID-19 on adolescent loneliness identified negative effects among the majority, but not all, of the included studies.¹⁷ The effects of COVID-19 mortality and public health stringency on loneliness, if any, may have diminished by the time PISA administered its 2022 survey. Moreover, our analyses of COVID-19-related variables were limited to only between-country effects and lack the temporal variability that would allow for a deeper understanding of these relationships.

This study has several strengths and limitations. Loneliness was estimated using large, nationally representative surveys with high response rates, thus enhancing the generalizability of the findings to participating countries and minimizing nonresponse bias. Second, PISA’s loneliness scale included consistent question wording across survey years and has demonstrated validity with other measures of psychological well-being. Third, we separately estimated associations between country-level correlates and loneliness that arise from time-invariant cross-national differences and within-country temporal variability, which are often erroneously conflated in multilevel studies.³⁰ This study is limited by the possibility that cultural norms shape how individuals respond to loneliness items, which may bias cross-national comparisons of loneliness. To address this limitation, we tested measurement invariance of the school loneliness scale across countries. The results showed metric invariance, implying that each loneliness item contributed to the same latent construct to a similar degree across countries. Second, examining correlations of societal indicators with loneliness cannot prove causation. However, as experimental studies on cultural change are impossible, this study still provides useful observations for generating hypotheses related to the societal determinants of loneliness. Finally, our population was limited to adolescents receiving formal education in predominantly higher-income countries. We restricted analyses to OECD countries because membership requires countries to share a certain set of democratic and economic principals, thus improving comparability on many characteristics that are challenging to account for in analyses. As a result, our findings may not be generalizable to low-to-middle-income countries and several countries located in Middle Eastern, Latin American, African, and Asian regions.

In conclusion, in OECD member countries, loneliness substantially increased from 2012 to 2015 and continued to increase or remain elevated thereafter. Within countries, lower unemployment rates and lower frequency of internet use in a given year are associated with higher levels of loneliness. Our findings underscore the urgency of addressing loneliness among adolescents. Identification of societal-level factors that co-vary with loneliness may generate hypotheses about large-scale interventions to promote social connection.

IMPLICATIONS AND CONTRIBUTION.

This study provides recent multinational estimates of trends in adolescent loneliness, which increased between 2012 and 2015, and to a lesser extent between 2015 and 2022. Further, it identifies country-level internet use and unemployment rates as correlates of loneliness among adolescents. Findings underscore the urgency of addressing loneliness among adolescents.

Abbreviations:

OECD

Organization for Economic Cooperation and Development

PISA

Programme for International Student Assessment

GDP

Gross Domestic Product

GPI

Global Peace Index