1 of 42

�

Annecy Behavioral Science Lab,

239 Route De Talloires , 74290 Menthon Saint Bernard

SIRET Number: 892741430 00011

Account Number: FR76 3000 4006 1700 0108 1140 480

2 of 42

Historical Time Trends in Loneliness in US Young Adults and Old Adults:��A Cross-Temporal Meta-Analysis�

Hans Rocha IJzerman

Director – Annecy Behavioral Science Lab

hans@absl.io

@hansijzerman

3 of 42

Conclusions

No increase of loneliness between 1998 and 2020

Results largely consistent with earlier meta-analysis

The term “loneliness epidemic” is likely not justified.

4 of 42

Loneliness Matters

5 of 42

Defining Loneliness

6 of 42

Defining Loneliness

Subjective social isolation

7 of 42

Loneliness and Public Health

Anxiety and depression

Poor sleep

Suicidal ideation and behavior

Premature mortality

8 of 42

Contradictions in prior research

Study	Age group	Time period	Loneliness measure	Results
Trzesniewski & Donnellan (2010)	Adolescents	1976 to 2006	Ad-hoc loneliness scale	Decrease
Twenge et al. (2019)	Adolescents	1976 to 2017	Ad-hoc loneliness scale	Increase
Twenge et al. (2021)	Adolescents	2012 to 2018	Ad-hoc loneliness scale	Increase
Clark et al. (2015)	Young adults	1978 to 2009	UCLA loneliness scale (Revised version)	Decrease
Buecker et al. (2021)	Young adults	1976 to 2019	UCLA loneliness scale (all versions)	Increase
Hawkley et al. (2019)	Old adults	2005 to 2016	UCLA loneliness scale (3-item version)	Stable

9 of 42

Contradictions in prior research

Topic Title 2

Variations in age group, time period, and loneliness measure across studies make any comparisons difficult

Study	Age group	Time period	Loneliness measure	Results
Trzesniewski & Donnellan (2010)	Adolescents	1976 to 2006	Ad-hoc loneliness scale	Decrease
Twenge et al. (2019)	Adolescents	1976 to 2017	Ad-hoc loneliness scale	Increase
Twenge et al. (2021)	Adolescents	2012 to 2018	Ad-hoc loneliness scale	Increase
Clark et al. (2015)	Young adults	1978 to 2009	UCLA loneliness scale (Revised version)	Decrease
Buecker et al. (2021)	Young adults	1976 to 2019	UCLA loneliness scale (all versions)	Increase
Hawkley et al. (2019)	Old adults	2005 to 2016	UCLA loneliness scale (3-item version)	Stable

10 of 42

The present study

Goal: Estimating the historical time trends in loneliness in US young adults and old adults

11 of 42

Present Study

Design: Cross-temporal meta-analysis

No strong priors (given contradictory data)

12 of 42

Contradictions in prior research

Topic Title 2

Study	Age group	Time period	Loneliness measure	Results
Trzesniewski & Donnellan (2010)	Adolescents	1976 to 2006	Ad-hoc loneliness scale	Decrease
Twenge et al. (2019)	Adolescents	1976 to 2017	Ad-hoc loneliness scale	Increase
Twenge et al. (2021)	Adolescents	2012 to 2018	Ad-hoc loneliness scale	Increase
Clark et al. (2015)	Young adults	1978 to 2009	UCLA loneliness scale (Revised version)	Decrease
Buecker et al. (2021)	Young adults	1976 to 2019	UCLA loneliness scale (all versions)	Increase
Hawkley et al. (2019)	Old adults	2005 to 2016	UCLA loneliness scale (3-item version)	Stable

Even in last three studies, where measures were comparable, contradictory results

13 of 42

Focus

The UCLA loneliness scale

Acceptable psychometric properties (converging and diverging validity; internal coherence; test-retest reliability)

14 of 42

Literature Search & Inclusion Criteria

Literature search
Citation track of Russel (1996) on Google Scholar

Inclusion criteria
Studies were written in English or French
Studies assessed loneliness with the UCLA 3 loneliness scale
Studies sampled US young adults (18 to 29 years old) or US old adults (60 years old and above)
Studies reported their sample size, mean loneliness score, and standard deviation
Studies included a sample not used in another study
Studies did not preselect their participants based on their loneliness scores
Studies did not collect data over multiple years
Studies were not case studies

15 of 42

Studies identified through database search

(k = 4,273)

Full-texts articles coded in step 1a

(k = 4,273)

Full-texts articles coded in step 2

(k= 258)

Studies included

(k_studies = 267;

k_means = 370;

n_participants = 69,340)

US young adults

(k_studies = 199;

k_means = 257;

n_participants = 50,167)

Search

Step 1b coding

Step 2 coding

Included

Step 1a coding

Articles excluded

(k = 2,891)

Articles excluded

(k = 1,124)

US old adults

(k_studies = 75;

k_means = 113;

n_participants = 19,173)

kappa_MEAN = .98

kappa_RANGE = [.85, 1]

kappa = .96

kappa = .93

Full-texts articles coded in step 1b

(k = 1,382)

16 of 42

16

Coding procedure

Studies identified through database search

(k = 4,273)

Full-texts articles coded in step 1a

(k = 4,273)

Full-texts articles coded in step 1b

(k = 1,382)

Full-texts articles coded in step 2

(k= 258)

Studies included

(k_studies = 267;

k_means = 370;

n_participants = 69,340)

US young adults

(k_studies = 199;

k_means = 257;

n_participants = 50,167)

Search

Step 1b coding

Step 2 coding

Included

Step 1a coding

Articles excluded

(k = 2,891)

Articles excluded

(k = 1,124)

US old adults

(k_studies = 75;

k_means = 113;

n_participants = 19,173)

kappa_MEAN = .98

kappa_RANGE = [.85, 1]

kappa = .96

kappa = .93

17 of 42

17

Coding procedure

Studies identified through database search

(k = 4,273)

Full-texts articles coded in step 1a

(k = 4,273)

Full-texts articles coded in step 1b

(k = 1,382)

Full-texts articles coded in step 2

(k= 258)

Studies included

(k_studies = 267;

k_means = 370;

n_participants = 69,340)

US young adults

(k_studies = 199;

k_means = 257;

n_participants = 50,167)

Search

Step 1b coding

Step 2 coding

Included

Step 1a coding

Articles excluded

(k = 2,891)

Articles excluded

(k = 1,124)

US old adults

(k_studies = 75;

k_means = 113;

n_participants = 19,173)

kappa_MEAN = .98

kappa_RANGE = [.85, 1]

kappa = .96

kappa = .93

18 of 42

18

Coding procedure

Studies identified through database search

(k = 4,273)

Full-texts articles coded in step 1a

(k = 4,273)

Full-texts articles coded in step 1b

(k = 1,382)

Full-texts articles coded in step 2

(k= 258)

Studies included

(k_studies = 267;

k_means = 370;

n_participants = 69,340)

US young adults

(k_studies = 199;

k_means = 257;

n_participants = 50,167)

Search

Step 1b coding

Included

Step 1a coding

Articles excluded

(k = 2,891)

Articles excluded

(k = 1,124)

US old adults

(k_studies = 75;

k_means = 113;

n_participants = 19,173)

Step 2 coding

kappa_MEAN = .98

kappa_RANGE = [.85, 1]

kappa = .96

kappa = .93

Secondary variables extracted
Sample type (specific population sampled)
Sample mean age
Sample female percentage
Scale internal consistency (Cronbach alpha)
Scale administration mode (whether it was in a written or in an oral manner)
Labels of response options of the scale
Number of response options on the scale
Scale completeness (whether the authors used the 20 items or not)

Main variables extracted
Year of data collection
Sample mean loneliness score and standard deviation
Sample size

19 of 42

19

Coding procedure

Studies identified through database search

(k = 4,273)

Full-texts articles coded in step 1a

(k = 4,273)

Full-texts articles coded in step 1b

(k = 1,382)

Full-texts articles coded in step 2

(k= 258)

Studies included

(k_studies = 267;

k_means = 370;

n_participants = 69,340)

US young adults

(k_studies = 199;

k_means = 257;

n_participants = 50,167)

Search

Step 1b coding

Step 2 coding

Included

Step 1a coding

Articles excluded

(k = 2,891)

Articles excluded

(k = 1,124)

US old adults

(k_studies = 75;

k_means = 113;

n_participants = 19,173)

kappa_MEAN = .98

kappa_RANGE = [.85, 1]

kappa = .96

kappa = .93

20 of 42

20

Data imputation and transformation

Year of data collection = year of publication MINUS 2 (if not reported)

Conversion of mean scores to sum scores

Score reversion when required

Data imputation

Data transformation

Conversion of scores to their equivalent on the original scale when required*

21 of 42

21

Meta-analytic procedure

Random effects meta-regression with cluster robust variance estimates

Accounts for the dependency in mean loneliness scores that occurs at the year of data collection level due to:

Assumes that the true mean loneliness scores estimated in each sample are not identical

Quantifies heterogeneity in mean loneliness scores (Q-test; TAU²; I²)

Metafor R package (version 3.4.0)

ClubSandwich R package (version 0.5.6)

Predictor: Year of data collection

Dependent variable: Mean loneliness score

Mean loneliness scores have different weights in the regression (inverse variance weighting)

Nesting within year of data collection

Studies reporting multiple sample means

22 of 42

22

Meta-analytic procedure

Random effects meta-regression with cluster robust variance estimates

Accounts for the dependency in mean loneliness scores that occurs at the year of data collection level due to:

Assumes that the true mean loneliness scores estimated in each sample are not identical

Quantifies heterogeneity in mean loneliness scores (Q-test; TAU²; I²)

Metafor R package (version 3.4.0)

ClubSandwich R package (version 0.5.6)

Predictor: Year of data collection

Dependent variable: Mean loneliness score

Mean loneliness scores have different weights in the regression (inverse variance weighting)

Nesting within year of data collection

Studies reporting multiple sample means

23 of 42

23

Meta-analytic procedure

Random effects meta-regression with cluster robust variance estimates

Accounts for the dependency in mean loneliness scores that occurs at the year of data collection level due to:

Assumes that the true mean loneliness scores estimated in each sample are not identical

Quantifies heterogeneity in mean loneliness scores (Q-test; TAU²; I²)

Metafor R package (version 3.4.0)

ClubSandwich R package (version 0.5.6)

Predictor: Year of data collection

Dependent variable: Mean loneliness score

Mean loneliness scores have different weights in the regression (inverse variance weighting)

Nesting within year of data collection

Studies reporting multiple sample means

24 of 42

24

Meta-analytic procedure

Random effects meta-regression with cluster robust variance estimates

Accounts for the dependency in mean loneliness scores that occurs at the year of data collection level due to:

Assumes that the true mean loneliness scores estimated in each sample are not identical

Quantifies heterogeneity in mean loneliness scores (Q-test; TAU²; I²)

Metafor R package (version 3.4.0)

ClubSandwich R package (version 0.5.6)

Predictor: Year of data collection

Dependent variable: Mean loneliness score

Mean loneliness scores have different weights in the regression (inverse variance weighting)

Nesting within year of data collection

Studies reporting multiple sample means

25 of 42

25

Description of the samples included

k_studies = 199

k_means = 257

n_participants = 50,167

US young adults

k_studies = 75

k_means = 113

n_participants = 19,173

US old adults

M_AGE = 20.82

M_CRONBACH = .91

Samples are relatively homogeneous (90.27% of university students samples)

M_AGE = 74.07

M_CRONBACH = .88

Samples are relatively heterogeneous (even at within study level)

k_studies = 267

k_means = 370

n_participants = 69,340

All data

151 journal articles

107 dissertations and theses

Year of data collection span 1998 to 2020

26 of 42

26

Description of the samples included

k_studies = 199

k_means = 257

n_participants = 50,167

US young adults

k_studies = 75

k_means = 113

n_participants = 19,173

US old adults

M_AGE = 20.82

M_CRONBACH = .91

Samples are relatively homogeneous (90.27% of university students samples)

M_AGE = 74.07

M_CRONBACH = .88

Samples are relatively heterogeneous (even at within study level)

k_studies = 267

k_means = 370

n_participants = 69,340

All data

151 journal articles

107 dissertations and theses

Year of data collection span 1998 to 2020

27 of 42

27

Results for US young adults (main analysis)

b₁ = 0.04, 95% CI [-0.09, 0.16] NS

Q(255) = 20285.23, p < .001 S

τ² = 22.34

I² = 98.03%

Inconsistent with Buecker et al. (2021), even on matched populations and time periods

Parameter testing

Heterogeneity

28 of 42

28

Results for US young adults (controlling for study characteristics)

Mean-centered (continuous covariates)

Dummy-coded (dichotomic covariates)

Mean-centered

Covariate	b₁	b₂
Sample type (university students vs. other)	b₁ = 0.03, 95% CI [-0.08, 0.15] NS	b₂ = 4.13, 95% CI [-0.48, 8.73] NS
Sample mean age (continuous)	b₁ = 0.05, 95% CI [-0.05, 0.15] NS	b₂ = 0.22, 95% CI [-0.27, 0.70] NS
Sample female percentage (continuous)	b₁ = 0.06, 95% CI [-0.06, 0.18] NS	b₂ = -0.02, 95% CI [-2.18, 2.14] NS
Scale internal consistency (continuous)	b₁ = 0.08, 95% CI [-0.05, 0.21] NS	b₂ = -3.08, 95% CI [-19.63, 13.47] NS
Scale administration mode (written vs. oral)	Not enough samples to run the model
Labels of response options of the scale (original vs. alternative)	b₁ = 0.03, 95% CI [-0.09, 0.16] NS	b₂ = 0.40, 95% CI [-1.37, 2.17] NS
Number of response options on the scale (original vs. alternative)	b₁ = 0.04, 95% CI [-0.09, 0.16] NS	b₂ = 0.52, 95% CI [-2.13, 3.16] NS
Scale completeness (complete vs. incomplete)	b₁ = 0.04, 95% CI [-0.09, 0.16] NS	b₂ = -1.46, 95% CI [-5.08, 2.15] NS

Italicized categories are reference categories

29 of 42

29

Results for US young adults (moderator analyses)

University students (reference category) vs. other

b₃ = -0.18, 95% CI [-1.35, 0.98] NS

Parameter testing

Historical time trends in loneliness are the same across the different populations of the young adults age group studied here

30 of 42

30

Results for US old adults (main analysis)

b₁ = 0.14, 95% CI [-0.17, 0.44] NS

Q(111) = 7135.64, p < .001 S

τ² = 21.78

I² = 97.77%

Parameter testing

Heterogeneity

Consistent with Hawkley et al. (2019) on matched time periods

31 of 42

31

Comparing historical time trends in loneliness between US young adults and old adults

Young adults (reference category) vs. old adults

b₃ = 0.10, 95% CI [-0.20, 0.40] NS

Parameter testing

The historical time trends in loneliness in US young adults and old adults between 1998 and 2020 don’t differ from each other

32 of 42

32

Assessing publication bias

Journal articles (reference category) vs. dissertations and theses

The historical time trends found in the present work don’t seem to vary across manuscript types, for both young adults and old adults

b₃ = -0.07, 95% CI [-0.33, 0.19] NS

Parameter testing (young adults)

b₃ = -0.37, 95% CI [-0.76, 0.02] NS

Parameter testing (old adults)

33 of 42

33

Results are not that inconsistent if we look at the effect sizes and their 95% confidence intervals

Contradictory Findings?

34 of 42

Conclusions

We observe loneliness trends as being stable for young adults and older populations
Results are consistent with Hawkley et al.
Results are inconsistent Buecker et al and Clark et al

35 of 42

Conclusions

Designation of pandemic (or epidemic) – based on these observations – not justified

36 of 42

Conclusions

Designation of pandemic or epidemic – based on these observations – not justified

Epidemic: “an unexpected increase in the number of disease cases in a specific geographical area”

“World Health Organization (WHO)(link is external and opens in a new window) declares a pandemic when a disease’s growth is exponential.”

37 of 42

Conclusions

Designation of pandemic (or epidemic) – based on these observations – not justified
Measurement matters – may misdiagnose the cause if we misdiagnose the numbers
If not a pandemic – is loneliness not important?

Other public health matters (e.g., cancer)

38 of 42

Conclusions

We don’t find changes, but retrospective design makes causal inferences difficult, if not impossible
UCLA Loneliness Scale – no known tests of longitudinal invariance
We are unable to observe distributional changes

39 of 42

Maybe we are wrong

40 of 42

Maybe we are wrong

Out of the 267 studies included in the study, only one had representative samples (one sample for young adults, one sample for old adults)
Samples for young adults were mostly university students (90.27%).
For older adults, quite heterogeneous

41 of 42

Conclusions

Cause of the differences:

Measurement error?
Sampling?
Narrow focus on loneliness?

So it is absolutely important to figure out the cause of the differences between studies. Are we dealing with measurement error, sampling differences, or conceptual problems – a lack of theoretical coherence – in how we measure loneliness? I think it is all three, and we have to tackle them one by one.

I will already say that I think there is absolutely a problem with measurement error in this work, as you will also see in the next talk – even different loneliness measures only correlate with one another to a limited degree.

It is also very likely that differences in sampling can partly explain differences between studies, while a narrow focus on loneliness in our study – and a lack of agreement on what constitutes social connection – is another key cause of the discrepancies between studies.

42 of 42

�

Thank you