Each team approached the task differently: some wrote their code from scratch, others modified code that Nick Brown shared, and all documented ambiguities and discrepancies they encountered along the way. This post summarises the most important findings from the reproduction teams. A more complete report on the reproduction effort can be found on our OSF page (in the file “Reproduction summary.pdf”), along with all of the teams’ code and graphical output.

What did we find? While some aspects of the original study were reproducible, our teams discovered significant issues that call into question the narrative that has emerged after the publication of the original article regarding a worsening “epidemic” of social isolation.

The Original Study: Methods and Main Findings

Kannan & Veazie (K&V) examined American social connectedness using the American Time Usage Survey (ATUS) dataset. The ATUS is a nationally representative survey in which Americans report how they spend their time throughout a single, randomly selected day, including whom they spend it with. It is conducted by the Bureau of Labor Statistics and contains detailed time diaries from over 210,000 Americans between 2003 and 2020.

The authors tracked three key dimensions of social connectedness:

1. Social isolation – Time spent alone (when no one else was present)

2. Social engagement – Broken down into four relationship categories:

• Time with household family (people living in the same home)
• Time with non-household family (relatives living elsewhere)
• Time with friends
• Time with “others” (neighbors, coworkers(1), acquaintances, etc.)

3. Companionship – Time spent socializing or in leisure activities with others

K&V’s findings painted a concerning picture of Americans’ social lives:

K&V also conducted “joinpoint” analyses to identify years when significant changes in trends occurred. For social isolation, they identified 2018 as a critical turning point where isolation began increasing more rapidly—a finding that suggested a worsening crisis even before the pandemic.

What Our Reproducing Teams Found

1. Basic Statistics Were Mostly Reproducible—With Some Exceptions

All six teams attempted to reproduce K&V’s descriptive statistics from the ATUS data. The results reveal notable discrepancies:

For income, only the teams that limited their analysis to 2010–2020 using the HEFAMINC variable (Brown and Zaneva) reproduced K&V’s exact percentages. Other teams, using different approaches which appear to us to be equally justifiable, found substantially different distributions.

For hours worked, the situation was even more problematic. Brown eventually discovered that by merging “variable hours” workers (TEHRUSLT code −4) into the 1–25 hour category, he could reproduce K&V’s exact percentages. However, this required testing multiple variables and making non-obvious decisions about how to handle special cases. Without explicit methodological descriptions, other researchers cannot know if this approach actually matches what K&V did.

These inconsistencies demonstrate that even basic descriptive statistics were impossible to reproduce with certainty without access to the original code. They also raise questions about the extent to which K&V’s results were influenced by their potentially somewhat arbitrary decisions regarding variable selection, which we have not found reported anywhere.

2. Joinpoint Findings Were Inconsistent and Potentially Misleading

K&V used joinpoint regression to identify years where significant changes occurred in time-use trends. For example, they reported a significant joinpoint in social isolation at 2018, suggesting a sudden acceleration of isolation just before the pandemic—a finding that helped shape the “epidemic” narrative in subsequent reporting and wider discussions.

Some teams attempted multiple analytical strategies and in what follows, we report replication results pooled from all attempts. Our teams found it essentially impossible to reproduce their joinpoints analyses consistently:

(Note: Matches with K&V are in bold. Brown and Urschler did not attempt joinpoint analyses.)

No single team reproduced all six sets of joinpoints reported by K&V. For the critical “social isolation” category—where K&V’s finding of a 2018 joinpoint helped support the epidemic narrative—only two teams identified 2018 as a turning point. The other teams found completely different patterns with joinpoints at various years between 2006 and 2019.

These discrepancies highlight multiple concerns:

1. Even with the same dataset, different analysts reached different conclusions about when significant changes occurred
2. The joinpoint method itself appears highly sensitive to small analytical choices
3. Without access to K&V’s exact statistical procedures, reproducing their findings becomes a guessing game
4. Perhaps most critically, the joinpoint method involves multiple statistical tests without a priori hypotheses about when changes should occur, dramatically increasing the probability of false positives

This last point is particularly important. When researchers conduct multiple tests looking for “significant” changes without specific hypotheses about when those changes should occur, they inevitably find “significant” results by chance alone. With 18 years of data, there are 16 possible joinpoints that could be identified, substantially inflating the risk of false positives. Using the standard α = 0.05 significance level, this creates approximately a 56% chance of falsely identifying at least one joinpoint—essentially a coin flip (and considering they looked at 6 measures with 16% each, this chance rises to 99.2%). Without proper correction for multiple testing, which K&V did not report doing, the joinpoint analysis becomes highly prone to detecting “significant” changes that may be mere statistical artifacts.

The inconsistency of these results, combined with the problem of multiple testing, fundamentally undermines confidence in one of the paper’s claims: that social isolation suddenly accelerated around 2018. Without theoretical grounding for why 2018 specifically should show a change—rather than, for instance, 2008 (when there was a financial crisis) or any other year—the identification of 2018 as a critical turning point likely represents a statistical artifact rather than a meaningful social trend.

3. The “Epidemic” Narrative Was Amplified by Plot Design

Perhaps the most striking revelation came from examining the visual presentations in K&V’s paper. Their graphs used truncated Y-axes that visually magnified relatively modest changes.

For example, in their social isolation plot, the Y-axis started at around 280 minutes per day rather than 0, making the increase from 285 to 309 minutes (an 8.4% change over 17 years) appear dramatic. When our teams recreated these plots with Y-axes starting at zero (the standard practice for ratio data, which has a true zero, such as minutes), the apparent crisis largely disappeared:

The same pattern applies across all K&V’s visualizations. For “Friends” social engagement, their graph suggests a catastrophic decline, but on a properly scaled axis, the reduction (while still substantial at about 43% from 2003–2019) appears as part of a broader pattern of slower, continued annual decrease rather than a crisis.

This shows how graphing choices can transform modest changes into apparent crises. When visualized according to standard scientific practices, the alarming trends in the original paper appear very much less severe.

4. Removing 2020 Data Further Flattens the Trend

The COVID-19 pandemic was an exceptional period of mandated social distancing, lockdowns, and remote work. It affected the year 2020, which was included in the original analysis. Including this anomalous year in trendlines (as K&V did) creates the impression of accelerating isolation.

When our teams removed 2020 from their analyses and focused on the pre-pandemic period (2003–2019), many of the apparent trends became even less pronounced:

These findings suggest that the “epidemic” narrative was partly driven by the inclusion of an extraordinary year that was unlike any that came before it. By including 2020 in their trendlines and discussions—despite acknowledging its exceptional nature—K&V’s paper may have created an inflated sense of crisis that doesn’t accurately reflect the more modest pre-pandemic changes.

5. Modest Effect Sizes Never Reported in the Original

K&V did not report effect sizes, standard errors, or confidence intervals for their findings—statistical information that would have helped readers judge the magnitude and reliability of the reported changes. When our reproduction teams calculated these statistics, they found that(2):

• The largest increases in time spent alone were typically 5–10 minutes per day over a 15–17 year span
• Over the entire period from 2003 to 2019, social isolation increased by about 24 minutes per day, or approximately 146 hours per year
• Expressed as standardized effect sizes, this change reflects a small effect: Cohen’s d = 0.10.
• For social engagement with friends, the effect size was similarly small: Cohen’s d = –0.19.
• Other domains showed even smaller changes, with Cohen’s d ranging from –0.05 (non-household family) to –0.11 (companionship, household family).
• The 95% confidence intervals for most of these changes overlapped substantially from year to year, meaning that many of the apparent year-to-year changes cannot be distinguished from random measurement error and sampling variation.(3)

While these changes are not trivial, they represent a relatively modest shift over nearly two decades—hardly the dramatic surge implied by the “epidemic” framing. Crucially, these data only measure time spent with others, not the quality of those interactions. Studies measuring loneliness—which reflects the quality rather than quantity of social connections—show remarkably stable trends over similar time periods. Research by Hawkley et al., Surkalim et al., and Dahlberg et al. (as documented in our companion blog post) demonstrates that subjective loneliness has remained stable or even decreased slightly among various populations, including older adults.(4) This suggests that while the pattern of social interactions may be changing, the quality of social connections people maintain has not significantly deteriorated. What we are likely observing is an evolution in how people interact rather than a crisis of social disconnection.

6. Between-Group Differences Eclipsed Time Trends

Our most striking finding is that demographic differences in social connectedness were consistently larger—often by an order of magnitude—than changes over time. This table compares the largest between-group differences with the overall temporal trend:

These findings suggest that who you are matters far more for social connectedness than when you lived—at least within this dataset. The differences between young and older adults, between those who work and those who don’t, and between income brackets dwarf the changes that occurred over nearly two decades. However, we must be cautious about drawing causal conclusions from these demographic associations. The relationships between age, work status, income, and social isolation likely involve complex interactions and potential confounding factors that cannot be disentangled through simple descriptive statistics. 

More sophisticated analytical approaches—such as conditional random forests, causal inference methods, or longitudinal analyses that follow individuals over time—would be needed to identify which factors most directly influence social connectedness. Despite these causal uncertainties, the magnitude of these demographic differences strongly suggests that structural and social factors are far more consequential for understanding isolation than mere temporal trends in the socio-economic environment.

7. Groups Most Affected by Social Isolation

Given that demographic differences are so substantial, we identified which groups appear most isolated or least socially engaged. This information could help target interventions more effectively than trying to address a general “epidemic”:

These findings suggest that social connection disparities might be better addressed through targeted interventions for specific at-risk groups—starting, perhaps, with those experiencing the greatest differences from the mean—rather than broad societal campaigns against an “epidemic.”

While our analysis cannot establish causal relationships between demographic factors and social isolation, it clearly identifies where the largest disparities exist. Practitioners and policymakers should focus their attention on these high-disparity groups—older adults, Black Americans, and/or those with limited income(5)—while researchers work to untangle the complex causal mechanisms at play. 

Even without definitive causal evidence, these substantial between-group differences provide valuable guidance on where to direct resources and which populations may benefit most from tailored support. Understanding why these disparities exist will require more sophisticated research, but knowing where to look is an essential first step toward developing effective interventions.

8. Adjusted Analyses Raised Serious Concerns

We found some other concerning aspects that were not as central to their analyses. For their income and hours-worked analyses, K&V reported using statistical adjustments for variables like age, sex, race, and ethnicity. However, they did not explain how these adjustments were performed. Here’s what happened: 

• In K&V’s plot of Non-Household Family social engagement by income groups (Fig. 5), the value for the lowest income group in 2019 appears to be approximately −10 minutes per day, which is mathematically impossible. Without access to the original code, we cannot determine whether these issues reflect conceptual misunderstandings or simple computational errors. Either way, they severely undermine confidence in a substantial portion of the paper’s analyses and conclusions.
• All eight adjusted lines for “Others Social Engagement” (in both Fig. 5 and Fig. 6) showed a strange synchronized dip in 2016 that appears in none of the unadjusted data. As above, without access to the original code, we cannot determine the cause of these issues (e.g., error or model artifact). These patterns suggest fundamental errors in the adjustment methodology rather than legitimate statistical controls.
• The adjusted plots often showed dramatic shifts that cannot be explained by any standard statistical adjustment procedure. For example, the mean values for time spent with “Others” in the overall data was consistently around 45 minutes per day from 2007 to 2017. However, in K&V’s income-adjusted and hours-worked-adjusted plots for the same period, all four group lines suddenly jumped to around 75 minutes per day—a 67% increase that should be mathematically impossible through standard adjustment techniques.

What Does This Mean for the “Epidemic of Loneliness”?

Our findings do not deny that loneliness and isolation are serious concerns worth addressing. However, they do show that the most prominent recent claims about trends in social isolation were built on:

• Ambiguous methods that proved difficult to reproduce
• Visual representations that magnified modest changes
• Statistical techniques whose results varied substantially between analysts
• Failure to report effect sizes or uncertainties
• Computational issues that produced impossible values in some analyses

The most defensible conclusion from the ATUS data appears to be that there was a modest increase in time spent alone between 2003 and 2019 (about 24 minutes per day, or 8.4%), with varying impacts across demographic groups.(6) This trend was then dramatically accelerated during the exceptional circumstances of the COVID-19 pandemic.

However, it’s crucial to recognize that time spent alone is not equivalent to loneliness.(7) While the U.S. Surgeon General used social isolation data to support claims about an “epidemic of loneliness,” the best available evidence on actual loneliness trends tells a different story. Studies by Hawkley et al., Surkalim et al., and Dahlberg et al. show that subjective experiences of loneliness have remained relatively stable over similar time periods, with some populations even showing slight decreases.

This disconnect between objective time use and subjective experience is important. People may be spending slightly more time alone, but they don’t necessarily feel more lonely. This suggests that the quality of social connections may be maintained even as the quantity shifts, possibly through:

• More meaningful but less frequent in-person interactions
• Supplemental digital connections not captured in the ATUS data
• Changes in expectations and preferences for social engagement

Additionally, the modest nature of these changes—just 24 minutes per day in time spent alone over nearly two decades—suggests an evolution in social patterns rather than a crisis. When properly contextualized alongside the much larger demographic disparities we found, these temporal changes appear far less alarming than the “epidemic” framing suggests.

What Should We Do Next? Implications for Research and Policy

Based on our findings, we recommend several critical steps for researchers, policymakers, and practitioners:

1. Focus on structural inequality: The largest disparities in social connection are between demographic groups, not across time periods. Addressing persistent inequalities by targeting isolated groups like older adults, Black Americans, and those with limited income may be more important than reversing modest temporal trends. While our analysis cannot establish causal relationships between demographic factors and social isolation, it clearly identifies where the largest disparities exist.
2. Examine work conditions as structural barriers to connection: Our analysis shows that hours worked significantly impact social engagement. Policy initiatives could address excessive work hours or inadequate wages that require multiple jobs, which limit time available for meaningful relationships.
3. Integrate quantity and quality measures: Moving forward, researchers and policymakers should focus on understanding both the quantity and quality of social connections. The ATUS data only capture time spent with others, not whether those interactions are meaningful or satisfying. True progress will require nuanced measurements that distinguish between mere co-presence and genuine connection.
4. Develop targeted interventions: Resources that could be directed toward addressing documented structural drivers of isolation among specific vulnerable groups might be squandered on programs designed to combat a general “epidemic” that the data do not actually support. Practitioners should focus their attention on high-disparity groups while researchers work to untangle the complex causal mechanisms at play.
5. Demand transparency and reproducibility: When studies influence national narratives or policy, we must insist on complete methodological transparency, including code sharing. Statistical errors, questionable visualization choices, and non-reproducible findings should not shape major policy decisions.

Conclusion

This collaborative reanalysis demonstrates why scientific transparency matters. Without being able to examine exactly how researchers arrived at their conclusions, we risk building policy on shaky foundations.

The work of the reproduction teams has shown that the narrative of rapidly accelerating social isolation is not as clear-cut as previously presented. While Americans may indeed be spending somewhat more time alone than they did two decades ago, the magnitude of this change is modest, and its meaning requires more careful interpretation than it has received.

If we genuinely care about addressing social isolation and loneliness, we must begin with accurate diagnoses based on transparent, reproducible science. The solutions we develop must be grounded in evidence that withstands scrutiny, not in narratives that collapse under careful examination.

All data, code, and detailed findings from this reanalysis project are publicly available. We welcome further feedback, replication, and use of these results as we work toward a more nuanced understanding of social connection in contemporary society.

Note: This is the first phase of our reanalysis project. In Phase 2, independent teams will develop their own approaches to analyzing the ATUS data, potentially revealing new insights about social connectedness beyond what was examined in the original study.

Footnotes:

(1) Note that ATUS does not measure whether time was spent with others or alone when working in the participant’s principal occupation. So to the extent that time spent with co-workers is included here, this principally reflects time spent in a social context, or for example commuting in a shared vehicle (“Traveling for work” is included as an activity category for measuring time spent with others or alone). Indeed, Kannan & Veazie’s methodology appears to assume that time spent at work is neither isolation nor companionship, which seems like a substantial concession given that it accounts for about one-third of the waking hours of most adults under retirement age, and that many people presumably have both positive and negative experience of interactions with co-workers or of time spent working alone.

(2) The analysis scripts for the calculations of these Cohen’s ds can be found on the OSF. Note that there are different approaches to calculating Cohen’s d, and they can yield slightly different results depending on the assumptions about variance. In this analysis, we follow a conservative approach by using the pooled weighted standard deviation across all years as the denominator. This method aligns with techniques used in mixed models and population-level surveillance studies. 

(3) Again, interpreting year-to-year changes in population trends is the risk of ecological fallacies and inflated correlations. Ecological correlations often overestimate effect sizes and can even misrepresent the direction of associations. Moreover, even when changes are statistically significant, small standardized effects (e.g., Cohen’s d ≈ 0.10) may fall within the range of what Meehl (1990) and others have referred to as “crud”—variance driven by unknown or trivial factors rather than meaningful psychological processes. Thus, both the limited precision of year-to-year comparisons and the possibility of overinterpreting small effects warrant a cautious reading of these findings.

(4) K&V do refer to an article that suggests that youth loneliness could have increased between 2000-2018. We believe this conclusion to be mistaken, due to elementary mistakes in their analyses. 

(5) Even if there were an overall trend in the population, one would need more careful examination of each subgroup’s contribution to the overall trend. For instance, it is likely to be true that – with the postwar boom – that the population is aging. As we see in our data, older adults tend to be more socially isolated, so the relative contribution of a subgroup would need to be better understood. K&V’s methodology—which only examines time usage trends within ATUS categories, rather than the numbers of people who make up each category—would not be able to discern such a trend.

(6) One important critique we have not discussed yet is the problem of ecological fallacies and inflated correlations in population studies. Population studies often analyze aggregated data (e.g., average scores per year), not individual-level data. That can lead to ecological correlations, which might be much larger—or even in a different direction—than correlations at the individual level (see Trzesniewski & Brent Donnellan, 2010, for a comprehensive critique).

(7) This point deserves far more nuanced exploration than a footnote can provide, but in brief: As people navigate life transitions, their social patterns often shift qualitatively rather than merely quantitatively. They may strategically reduce time with peripheral contacts while deepening engagement with close family and friends—potentially improving their overall social well-being despite appearing more “isolated” in time-use data. Research by Carstensen and colleagues on socioemotional selectivity theory suggests this process may be adaptive rather than concerning. Furthermore, recent research indicates that solitude itself isn’t inherently problematic; moderate time alone can be restorative and contribute positively to psychological well-being when balanced with meaningful social connection (see, for instance, Nguyen & Rodriguez, 2024; Weinstein et al., 2022). The complex interplay between quantity and quality of social interaction across different life stages remains an important area for future research.