An effective short form of the 20-item University of California Los Angeles Loneliness Scale version 3: item response theory and network psychometrics

Abstract

Loneliness is a complex and usually unpleasant emotional response to isolation, which has been considered the latest global health epidemic exacerbated by the coronavirus disease 2019 pandemic, affecting nearly two-thirds of older adults. Some profound health implications carried by loneliness include depression, cognitive impairment, hypertension and frailty. Across the world, there is no consensus definition of loneliness, and its measure is based on the phenomenological perspective of the individual. The 20-item University of California Los Angeles Loneliness Scale version 3 (UCLA-20) is the most common measure. This scale demonstrates acceptable psychometric properties but is too long and complex for a phone interview. This paper addresses the increasing need to shorten this scale by adopting classical item response theory and network psychometrics to advance scale development. Through an item reduction analysis, we trimmed the original scale into an effective short form, which is as valid as the original one. With respondents’ time at a premium in most research nowadays, this short-form scale is an efficient and practical alternative to the original UCLA-20.

Introduction

Scales, or psychometric instruments, are essential for capturing complex constructs—such as behaviours, emotions or perceptions—that cannot be directly measured by a single item. Using multiple items improves measurement reliability by accounting for item-specific errors, thereby enhancing the validity of research findings.¹

Loneliness and social isolation have increasingly been recognised as a global public health concern. Before the coronavirus disease 2019 pandemic,² loneliness had a baseline prevalence of 20%–35% among U.S. adults, which increased substantially thereafter.3,5 Chronic loneliness is associated with severe mental and physical health risks, including depression, anxiety, cognitive impairment, dementia, hypertension and even premature mortality.⁶ Notably, the impact of social disconnection on mortality is comparable to smoking 15 cigarettes per day.⁷

As an inherently subjective experience, loneliness cannot be measured directly, necessitating reliable and valid scales for its assessment. However, no universally accepted definition of loneliness exists, leading to heterogeneous measurement approaches across studies. Among these, the 20-item University of California Los Angeles Loneliness Scale version 3 (UCLA-20) is one of the most widely used and psychometrically validated instruments, demonstrating strong test-retest reliability, internal consistency and construct validity.⁸ It is particularly prevalent in ageing research; for instance, a PubMed search on “loneliness and ageing” yields 112 full-text articles, 39.83% of which employ the UCLA-20, nearly twice the usage of the 11-item De Jong Gierveld Loneliness Scale (19.49%) or single-item measures (19.49%).

Despite its widespread use, the UCLA-20 presents a substantial response burden, particularly in survey-based and telephone-administered studies.⁹ This has led to efforts to develop short-form versions, including a 4-item version by Russell¹⁰ and a 6-item version by Neto.¹¹ However, few studies have systematically validated these short forms or directly compared their psychometric properties, leaving uncertainty about their effectiveness.

In this study, we leverage classical item response theory (IRT) and emerging psychometric network approaches to derive and validate an optimised short form of the UCLA-20. Through rigorous item reduction analysis and validation procedures, we aim to retain only the most psychometrically robust items while preserving the scale’s reliability and validity. By offering a more efficient measure with minimal response burden, our proposed short-form scale has the potential to improve loneliness assessment in both research and clinical settings.

Method

Study participants

This study analysed a combined sample of three cohorts (n=353) with similar inclusion criteria: (1) community-dwelling adults, (2) written informed consent, (3) fluency in English, (4) physical and mental ability to complete assessments, (5) no dementia diagnosis and (6) completion of the UCLA-20 loneliness scale.

Participants (mean (standard deviation (SD)) age: 62 (18) years; range: 26–101 years) were 50% female, 78.8% were Caucasian, 54.1% had some college or a bachelor’s degree, 33.2% had a postgraduate degree and 58.8% were married. An analysis of variance test (p=0.263) showed no significant differences in UCLA-20 scores across cohorts, supporting their reasonable combination for analysis and a potential to reflect a wider range of loneliness.

Measure

Loneliness measure

The UCLA-20 assesses subjective loneliness and social isolation. Participants rate 20 items on a 4-point Likert scale, with nine reverse-scored items. The total score (0–60) categorises loneliness as low (<28), moderate (28–43), or high (>43).¹²

Sociodemographic and clinical characteristics

Demographic variables included age, gender, education, race/ethnicity, marital status, living situation and income. Clinical measures assessed psychological traits/states: depression (Patient Health Questionnaire-9),¹³ anxiety (Brief Symptom Inventory (BSI)-Anxiety subscale), perceived stress (Perceived Stress Scale),¹⁴ resilience (Connor-Davidson Resilience Scale),¹⁵ optimism (Life Orientation Test-Revised),¹⁶ mental well-being (36-Item Short Form Health Survey (SF-36)),¹⁷ and wisdom (San Diego Wisdom Scale).¹⁸ Physical health was evaluated via SF-36 physical functioning scores.

Statistical and psychometric analyses

Exploratory factor analysis

While the UCLA-20 has demonstrated strong internal consistency, some studies suggest potential multidimensionality. To assess its latent structure, we conducted an exploratory factor analysis (EFA), regressing standardised observed variables onto latent factors to evaluate internal consistency. The optimal number of factors was determined using three criteria: parallel analysis,¹⁹ minimum average partial correlation (MAP) procedure²⁰ and very simple structure (VSS) criterion. This decision guided subsequent item reduction, ensuring selections were made within each identified factor.

Item reduction analysis

IRT was applied to evaluate individual item characteristics, including item-total correlations and item discrimination. Both parametric graded response models (GRM)²¹ and non-parametric IRT models²² were used for a comprehensive assessment.

The parametric GRM is suitable for our context. Let $k(=1,...,K)$ index the ordinal level in a scale (eg, Likert scale), the probability of the response for the $m$-th subject on the $i$-th item follows:

$${\displaystyle \mathrm{P}\mathrm{r}(x_{im}\ge k|z_m,{\alpha }_i,{\beta }_{ik})=\frac{exp\left\{{\alpha }_i(z_m-{\beta }_{ik})\right\}}{1+exp\left\{{\alpha }_i(z_m-{\beta }_{ik})\right\}}=g({\eta }_{ik})}$$ (1)

and

$${\displaystyle \mathrm{P}\mathrm{r}(x_{im}\ge k+1|z_m,{\alpha }_i,{\beta }_{i,k+1})=\frac{exp\left\{{\alpha }_i(z_m-{\beta }_{i,k+1})\right\}}{1+exp\left\{{\alpha }_i(z_m-{\beta }_{i,k+1})\right\}}=g({\eta }_{i,k+1})}$$

yielding

$${\displaystyle \mathrm{P}\mathrm{r}(x_{im}=k|z_m,{\alpha }_i,{\beta }_{ik},{\beta }_{i,k+1})=g({\eta }_{ik})-g({\eta }_{i,k+1}),}$$ (2)

In the parametric IRT model, ${\alpha }_i$ is the discrimination parameter for item $i$, while ${\beta }_{i1}\S amp;lt;{\beta }_{i2}\S amp;lt;\cdots \S amp;lt;{\beta }_{iK}$ are the item-specific difficulty parameters, and $z_m$ denotes the trait level for subject $m$. Estimation requires numerical integration and optimisation (eg, Expectation-Maximization algorithm, Newton-Raphson). The primary focus is on the discrimination parameter (${\alpha }_i$,), which indicates how well an item differentiates between respondents.

In contrast, the non-parametric IRT model uses kernel smoothing to generate item option-characteristic curves, visually assessing discrimination. Greater separation between curves indicates stronger discrimination.

We selected items for the shortened scale based on three criteria:

1. High item-total correlation (≥0.55, excluding the item itself).

2. High discrimination (parametric GRM slope ≥1.50).

3. Better separation in option-characteristic curves (non-parametric IRT).²²

The selected items formed a candidate short-form scale, which was further evaluated for reliability and validity.

Reliability and validity

Internal consistency was assessed using Cronbach’s alpha, a common reliability index and Revelle’s omega,²³ which is more robust as it allows items to contribute uniquely to the latent factor.

Validity was first examined through Spearman’s correlations with external clinical measures, followed by replicating results from other statistical models. Specifically, we conducted bivariate correlations between sociodemographic and clinical variables using both the UCLA-20 and the shortened scale, with significance tested via the Fisher r-to-z transformation.

To assess structural validity, we replicated a prior finding by Lee et al, who identified an unusual ‘W’-shaped relationship between UCLA-20 scores and age using a linear spline model with a change point at age 53 years.²⁴ We applied the same analysis to the shortened scale for comparison.

Factor invariance

To assess whether the shortened loneliness scale functions similarly across sexes, we tested factor invariance using confirmatory factor analysis. Given prior findings of higher loneliness severity in men,²⁵ we examined construct validity across sex groups. Additionally, we applied group-stratified item response models to compare item characteristic curves, evaluating differences in response probabilities between men and women.

Psychometric networks

To further assess the shortened scale’s ability to retain information, we employed a novel psychometric network approach, which conceptualises items as mutually reinforcing nodes rather than assuming a latent construct.^{26 27} Networks were built using partial correlations (inverse of variance-covariance matrix), with edge weights (−1 to 1) indicating the strength of item connections after adjusting for all others. Regularisation was applied to reduce overfitting and eliminate spurious edges.28,30

Two network comparisons were conducted:

1. Item-level networks of the 20-item and shortened scales to compare structural integrity.

2. External variable networks, incorporating clinical characteristics to examine differences in item associations.

Regularised networks were estimated and visualised using R package qgraph, where stronger partial correlations appeared as thicker edges, with positive and negative connections coloured blue and red, respectively.

Results

The items in the scale had moderate item-total correlations (mean (SD): 0.70 (0.10); range: 0.36–0.82), and moderate item-dropping correlations after correcting for item overlap (mean (SD): 0.67 (0.09); range: 0.35–0.79).

Dimensionality and shortened scale

Unidimensionality of the UCLA-20

For the optimal number of factors, multiple criteria yielded different results. Horn’s parallel analysis suggested four factors, while Velicer’s MAP criterion with both oblimin and promax rotations indicated two factors. In contrast, the VSS criterion suggested a single factor. Given these discrepancies, a closer examination of factor diagrams and fit indices was necessary.

The one-factor model produced a Tucker-Lewis Index of 0.77 and a root mean square error of approximation of 0.15. In comparison, the two-factor model showed a slightly better fit with a Tucker-Lewis Index of 0.89 and a root mean square error of approximation of 0.09. However, the factor diagram revealed that the nine reverse-coded items clustered separately from the 11 non-reversed items (see online supplemental figure S1). The extremely high correlation of 0.80 between these two factors suggested that this distinction was primarily an artefact of response patterns rather than a meaningful difference in latent constructs. Additionally, the omega-hierarchical coefficient of UCLA-20 was 0.80, indicating that the variance common to a single trait was sufficient to justify a unidimensional structure.

Given these findings, the UCLA-20 was determined to be unidimensional. In the EFA with oblimin rotation, after correcting all items in the same theoretical direction, most demonstrated moderate to large loadings on the primary factor, with a mean loading of 0.74, a median loading of 0.75 and a range from 0.41 to 0.86 (see online supplemental figure S2). Among these, 10 items—3, 5, 6, 7, 10, 13, 14, 16, 19 and 20—exhibited particularly strong loadings on the latent factor.

Establish a shortened version of the UCLA-20

The process of establishing a shortened version of the UCLA-20 scale involved selecting items that demonstrated strong psychometric properties across multiple criteria. Items that showed high correlations with the total score included 2, 3, 4, 6, 7, 10, 11, 13, 14, 19 and 20, with correlations exceeding 0.70 (see online supplemental table S1). The parametric graded response model identified items 2, 3, 6, 7, 10, 13, 14, 16, 19 and 20 as having the highest discrimination, with values exceeding 2 (see online supplemental table S2). Meanwhile, the item characteristic curves indicated that items 1, 4, 5, 6, 8, 9, 10, 11, 12, 15, 17 and 18 exhibited better separations (not shown).

The final selection of items was based on these individual properties and on factor loadings and item content. The selected items—1, 6, 7, 10, 11, 16, 18, 19 and 20—were chosen to balance strong psychometric performance with theoretical considerations. Each of these items had a factor loading >0.60 in the one-factor solution, ensuring that they contributed meaningfully to the latent construct of loneliness. Furthermore, the item information curves demonstrated that these nine items provided broad coverage of the loneliness dimension (see online supplemental figure S3).

This newly established nine-item scale, referred to as UCLA-9, was reassessed using factor analysis, which confirmed its unidimensionality, consistent with the structure of the original UCLA-20. Descriptive statistics for the UCLA-9 scale are presented in table 1 and compared with those of the UCLA-20 in online supplemental table S3, with both scales transformed to a 0–100 scale for comparability. The loneliness scores from both measures were moderately positively skewed, with skewness values of 0.79 for UCLA-20 and 0.86 for UCLA-9, indicating that more participants scored below the midpoint. Other key statistics, including mean, median and range, were highly similar between the two versions, suggesting that the UCLA-9 effectively preserves the characteristics of the original scale while offering a more concise measure of loneliness.

Table 1. Descriptive statistics of the shortened UCLA (nine items) score.

Items	N	Description	Mean (SD)
uclals01r	350	I am unhappy doing so many things alone.	1.64 (0.60)
uclals06r	349	I find myself waiting for people to call or write.	1.75 (0.67)
uclals07	348	There is no one I can turn to.	1.66 (0.86)
uclals10r	349	I feel left out.	1.68 (0.67)
uclals11	348	I feel completely alone.	1.99 (0.74)
uclals16r	349	I feel isolated from others.	1.77 (0.72)
uclals18	350	It is difficult for me to make friends.	2.12 (0.82)
uclals19r	350	I feel shut out and excluded by others.	1.45 (0.68)
uclals20r	349	People are around me but not with me.	1.42 (0.66)

Reliability and validity

Reliability

The original UCLA-20 has a high internal consistency with 0.95 in Cronbach’s alpha (95% confidence interval (CI) 0.94 to 0.96). For Revelle’s omega,²³ in addition to the omega total of 0.95, the omega hierarchical of 0.80 confirmed that the amount of variance common to a single factor underlying all items is sufficient. In comparison, Cronbach’s alpha for the UCLA-9 scale was still as high as 0.90 (95% CI 0.88 to 0.91). The narrower CI highlighted its internal consistency after the item reduction. Besides, the robust omega total was 0.93 on the reduced nine-item scale, and the omega hierarchical was quite close to the original scale (0.77 vs 0.80), indicating its ability to capture most of the information. In the omega analysis, the eigenvalue of the main factor (3.89) far exceeded the factor loadings of residuals of values (0.58, 0.70 and 0.57, see online supplemental figure S3), suggesting a clear structure and reasonable model fit. Therefore, the reduced UCLA-9 is as reliable as the original scale.

Validity

Our validity evaluation was conducted at several levels. The first-level comparison using Spearman’s correlations is presented in online supplemental table S4, where the UCLA-20 had moderately positive correlations with negative psychological states and traits, and negative correlations with positive psychological states and traits. It also had a small but significant negative correlation with physical health. After item reduction, the UCLA-9 maintained a high correlation with UCLA-20 (0.95, p<0.001). More importantly, it preserved the correlation with all external variables in direction and significance, with only minimal variation in magnitude. For example, the correlation between anxiety (BSI) and UCLA-20 was 0.32 (p<0.001) and 0.29 (p<0.001) with UCLA-9. Some correlations even experienced an increase, such as with physical health (−0.15 for UCLA-9 vs −0.12 for UCLA-20), presenting a more significant p value (0.005 for UCLA-9 vs 0.033 for UCLA-20). These results confirmed the content validity of the UCLA-9 scale.

In a related study, Lee et al²⁴ first identified an unusual ‘W’ shape of UCLA-20 across age, where loneliness was found to peak in individuals in their late 20s, mid 50s and late 80s. To further assess the validity of the UCLA-9, we reproduced the data analysis using the UCLA-9 in place of the UCLA-20 as the outcome. The original and new model parameters from a linear spline model with a change point at the age of 53 years are presented in online supplemental table S5. Remarkably, the results demonstrated that using the UCLA-9 retained the W shape, as shown in online supplemental figure S5, which was the key finding of the previous study, but the parameter estimates and significances were also very similar across the two models. By its capacity for detecting major findings of the published work, this UCLA-9 scale shows great potential for future research.

Invariance of UCLA-9 across sex

For the factor invariance of UCLA-9, the confirmatory factor analysis results were similar for models assessed separately within each sex. Furthermore, no apparent sex differences were detected in the item characteristic curves for any item (not shown). Lastly, the density of the overall UCLA-9 scores for each sex was similar in shape, as shown in online supplemental figure S6.

Psychometric networks

Exploratory findings in psychometric networks

Our novel application of the psychometric networks showed some interesting results. First, the network of UCLA-20 demonstrated a strong and positive connection between items 19 and 20, which were both selected and reserved such a strong link in the UCLA-9 (see figure 1). Item 19 evaluates whether the person feels ‘shut out and excluded by others’. In contrast, item 20 assesses if one feels “people are around me but not with me”. Both items are related to responses to external connections, but with nuances. Since “being excluded” and “not aligning with other people” capture different aspects of loneliness, it is worth keeping both in the reduced scale.

Figure 1. Exploratory psychometric networks comparing item connections in (a) original UCLA (20 items) and (b) shortened UCLA (nine items).

Another remarkable finding was that the UCLA-9 presented a three-cluster structure, with items 7, 11 and 18 forming a more independent cluster, while items 1, 6, 10 and 16 formed another cluster, which was more connected with items 19 and 20. Pre-eminently, this structure resembled that of UCLA-20. For example, items 1, 5, 6, 9, 10 and 16 were more closely connected, further linked to items 15, 19 and 20 through item 16. On the other hand, items 7, 11 and 18 were found at the bottom of the UCLA-20 network, and this larger cluster also presented negative connections with most items on the top.

Finally, we depicted the two networks involving the previous clinical characteristics and loneliness items to identify their relationship. Item 11 (“I feel completely alone”) was found to be closely linked to stress (Perceived Stress Scale Total) in both networks. Also, the direct link between item 6 (“I find myself waiting for people to call or write”) and the mental health component score in SF-36 was much clearer in the UCLA-9 network (see online supplemental figure 7). Similarly, the link between item 1 (“I am unhappy doing so many things alone”) and physical health was depicted more distinctly in the UCLA-9 network. Although exploratory, those results can help generate hypotheses for further investigation.

Discussion

Loneliness prevalence can vary depending on the measurement tool used, as different scales capture distinct aspects of loneliness and may be sensitive to reporting biases. We developed the shortened UCLA-9 scale using advanced psychometric techniques, demonstrating a high correlation with the UCLA-20 (0.95, p<0.001), strong reliability (α=0.90, Ω_total=0.93) and comparable validity. The UCLA-9 also maintained invariance across sex and replicated key findings related to age. Our exploratory network analysis uncovered novel associations between loneliness and clinical characteristics.

Hence, the shortened UCLA-9 scale is a reasonable representation of the UCLA-20. Given the increasing time constraints in research, the UCLA-9 offers a practical alternative, reducing respondent burden by approximately 55%. Self-administration of the UCLA-9 takes only 1–2 min (about five items per minute), a reduction that could increase the quality of data collected by allowing participants to respond more quickly and with greater attention.

Our UCLA-9 scale contains more informative items than some other shortened forms, such as the six-item scale.¹¹ With a considerable reduction, it is unclear whether they can fully capture the complexity of loneliness. For example, despite some overlap with our core items, the six-item scale lacks items 19 and 20, essential for depicting one’s relationship with the external world, including feelings of being shut out, excluded or misaligned with others. These elements are crucial in understanding the social dimensions of loneliness, which our UCLA-9 addresses.

This study’s strengths include using a well-characterised cohort from across the adult lifespan and considering both psychological traits and states. These strengths enhance the external validity of our findings and support the robustness of the UCLA-9 scale. Additionally, combining various psychometric tools allowed us to design a more effective and efficient shortened scale.

However, there are limitations. The cross-sectional design precluded assessment of test-retest reliability, and self-reported measures may have introduced social desirability bias. Additionally, the sample was predominantly educated and Caucasian, which may limit generalisability to other populations. Future validation studies will evaluate the UCLA-9 in more diverse groups and assess its test-retest reliability.

In conclusion, the UCLA-9 is an efficient, reliable and valid tool for measuring loneliness. Its brevity and strong psychometric properties make it timely for research that demands quick, high-quality data collection.

Learning Points

• Loneliness is a complex and distressing emotional response to social isolation, affecting nearly two-thirds of older adults. The UCLA Loneliness Scale Version 3 (UCLA-3) is among the most widely used instruments for assessing loneliness and has demonstrated acceptable psychometric properties.

• The 20-item UCLA Loneliness Scale Version 3 is often too lengthy for phone-based or time-limited research settings. This study introduces the UCLA-9, a psychometrically robust short-form scale that retains the reliability and validity of the full version. Its brevity reduces respondent burden by over 50%, making it a practical and efficient alternative for high-quality loneliness assessment in modern research contexts.

• This study provides a brief, reliable, and valid tool to assess loneliness, which can enhance large-scale data collection in both clinical and population-based studies. By reducing the response burden, the UCLA-9 facilitates broader implementation in public health surveillance, policy evaluation, and intervention research. It also supports more accurate and consistent measurement of loneliness across diverse settings, critical for tracking trends and informing strategies to address this growing global health concern.

Biography

Jinyuan Liu obtained her PhD in Biostatistics in 2022 from UC San Diego in the USA. Since then, she has joined Vanderbilt University Medical Center in the USA and is currently working as an Assistant Professor in the Department of Biostatistics with a secondary appointment in the Department of Psychiatry. Her main research interests include semi-parametric efficiency, causal inference, psychometrics and building effective dimension-reduction and efficient integrative modelling of high-dimensional data from various disciplines, including genomics, imaging and wearables.