Misery loves company: Predictors of treatment response to a loneliness intervention

Abstract Objective The last 10 years have seen a surge of interest in loneliness and interventions to reduce it. However, there is little evidence regarding differential treatment effectiveness and predictors of treatment outcome. This paper aimed to investigate possible predictors of treatment response. Methods We analysed data from two clinical trials of an evidence-based loneliness intervention: Groups 4 Health (G4H). Study 1 had 163 observations across two timepoints, n = 94; Study 2 had 297 observations across four timepoints; n = 84. Theorised predictors—symptom severity at baseline, program engagement, and demographic characteristics—were assessed for their effect on the primary outcome: loneliness. Results Across both trials, participants with more severe baseline loneliness or social anxiety, or who attended more sessions, experienced greater improvement in loneliness. In Study 2, those with diagnosed mental illness or more severe baseline depression also tended to have better outcomes. There was no evidence that age, gender, or ethnicity predicted program efficacy. Conclusion Overall, those with greater need—reflected in either severity of loneliness or psychological distress—tended to show greater improvement over time. This was due, in part, to greater engagement with the program among those who were lonelier. We discuss how loneliness interventions can be deployed most effectively to combat this profound public health challenge.

Loneliness is a major public health concern, not only because of its inherently aversive quality, but also because of its robust causal links to poor mental and physical health (Holt-Lunstad et al., 2015;Ingram et al., 2020;Valtorta et al., 2016). Although loneliness has been known to be a malady of modern society for some decades, the COVID-19 pandemic and associated restrictions on social interaction have increased its profile, both as a research topic and in the minds of the general public. Yet despite this growing profile and interest, previous research has had more success in establishing the detrimental consequences of loneliness than in designing and evaluating effective interventions to reduce it. A recent meta-analysis found that the average effect size of loneliness interventions in RCTs (relative to a control group) was relatively modest (Hedges' g = 0.32; Eccles & Qualter, 2020). This stands in contrast to psychological interventions for depression (where 0.62 < g < 0.92 relative to waitlist; Barth et al., 2013) and anxiety (where g = 1.22 pre-post; Bandelow et al., 2015).
Despite this slow start in the push to reduce loneliness, some new interventions are showing promise. In particular, interventions that prioritize opportunities for group-based interaction seem to be more effective than one-on-one or educational interventions (Cattan et al., 2005). There is also evidence of heterogeneity in benefit based on the type of intervention. In their meta-analysis, Hickin et al. (2021) found a borderline significant difference (p = .06) between intervention type, with (1) social identity or (2) reminiscence approaches having the largest effect sizes. In particular, GROUPS 4 HEALTH (G4H), which is a manualised social identity based intervention designed to boost group-based belonging to counter isolation, has been found to reduce loneliness in three clinical trials (across three different populations, pre-post effect sizes ranged from 0.86 < d < 1.07; Haslam et al., 2016;Cruwys et al., 2022).
Although these developments are promising, several questions remain unanswered. First, it is not clear whether the benefits of completing a loneliness intervention are experienced uniformly by all patients in the program. It is possible that the average level of improvement masks the fact that some people experience considerable improvement while others experience very little, if any at all. Second, if benefits are not uniform, it is important that predictors of treatment response (and/or non-response) are identified. This should allow for more effective targeting of intervention and also inform efforts to improve interventions by ensuring that they addresses the needs of treatment-resistant groups (see also Bessaha et al., 2020).

Predictors of Treatment Response
Although new to the domain of loneliness research, in the broader psychotherapy literature there has been a concerted effort to establish predictors of treatment response (e.g., Kraemer, 2016;Kraemer et al., 2006). This has been championed in the interest of accurately predicting prognosis, personalizing patient care, and refining interventions so that they benefit a higher proportion of patients. The existing evidence mostly concerns common mental disorders. For example, there is some evidence that differential treatment response is predicted by the severity of a person's symptoms when commencing treatment. More specifically, people with more severe depression symptoms at baseline tend to obtain more benefit from antidepressant medication and interpersonal psychotherapy (Elkin et al., 1995). Echoing this finding in a different population, a high-quality evaluation of predictors of treatment response in an eating disorder prevention group program also found that severity of eating disorder symptoms at baseline predicted greater response to treatment (Butryn et al., 2014). Indeed, this study found that initial severity was a better predictor of treatment response than facilitator or group factors. Another larger evaluation (n = 977) of the same program replicated this and additionally found that the presence of a diagnosable clinical eating disorder at baseline predicted better treatment response (Müller & Stice, 2013). Nevertheless, some research has shown the opposite, with poorer outcomes of depression treatment for those with more severe or recurrent symptoms (Ezquiaga et al., 1999).
One finding that is particularly relevant to loneliness intervention is that high quality social support and social relationships have consistently emerged as predictors of better treatment response in depression (Ezquiaga et al., 1999;Hallgren et al., 2017;Trivedi et al., 2005;Woods et al., 2021). And while a systematic review found that this relationship is particularly well established for depression treatment, it has also been observed for schizophrenia, bipolar, and anxiety disorders (Wang et al., 2018). However, it does not seem to be straightforwardly the case that social connectedness predicts better treatment outcomes across conditions. For example, one study that evaluated a behavioural intervention for shyness found that the participants who benefitted most from the program were those with more overt symptoms and social impairment (Alden & Cappe, 1988). One important distinction may be that the primary target of this intervention was social in nature, rather than relating to clinical symptomatology. Although loneliness is not a clinical disorder, it may be the case that when it is the target of treatment, initial severity is positively associated with the benefit a person will derive from the program.
Nevertheless, very little existing evidence speaks specifically to the predictors of effectiveness in the case of interventions designed to tackle loneliness. Most explorations have focused on the difference between intervention types (e.g., between group vs. individual delivery; Cattan et al., 2005;Mann et al., 2017;Masi et al., 2011), rather than on the characteristics of participants that enhance or undermine treatment response. However, two recent reviews provide some initial insights into this question. First, a recent meta-analysis of 31 trials of loneliness interventions found no evidence that age or gender predicted treatment effectiveness (Hickin et al., 2021). Second, another meta-analysis that focused specifically on young people (39 studies), found no evidence that age, gender, or clinical at-risk status (present vs. absent) predicted treatment outcome (Eccles & Qualter, 2020). Indeed, to our knowledge, the literature has not identified any predictors of treatment response in loneliness interventions. Accordingly, the question remains whether there might be other factors that do predict treatment response. Certainly, were this the case, this would help with the task of improving loneliness interventions and targeting them more effectively.
With this in mind, the goal of the present study was to identify predictors of treatment response to GROUPS 4 HEALTH (G4H). G4H is a psychotherapeutic program that seeks to reduce loneliness (and future risk and recurrence of loneliness) by supporting participants to build skills that help them form group-based connections. Furthermore, G4H seeks to provide opportunities for participants to experientially develop group-based connection within the therapy group, as a scaffold to supporting enduring connections after the program ends. G4H consists of five modules each of 60-90 min duration, and therapy groups range in size from 5-8 participants plus two facilitators. The modules involve (1) psychoeducation about the link between social connection and health, (2) mapping one's social world and identifying areas for growth, (3) developing strategies to make the most out of one's social groups, (4) developing strategies to form new social ties, and (5) a booster session to troubleshoot challenges and celebrate successes. Modules are held weekly except for the booster, which is scheduled one month later. As an active psychotherapy program founded on behavioural principles, G4H is interactive. Participants practice strategies in a collaborative fashion, reflect on learning, and are assigned homework activities to complete between sessions in their own time.
Previous research of the form reviewed above guided our choice of potential predictors of treatment response to investigate, and, on this basis, we identified eight potential predictors across three categories of variables. These were (1) severity of initial presentation, operationalized both in terms of severity of loneliness, and in terms of severity of clinical symptoms of depression, social anxiety, and mental health diagnostic status (present vs. absent); (2) program engagement, operationalized as record of attendance; and (3) demographic characteristics, specifically ethnicity, age, and gender. We focused on these three characteristics for two reasons: first, because research and practitioners who have sought to tailor psychotherapy to specific subpopulations have tended to focus on these variables (e.g., Huey & Tilley, 2018;Whiston et al., 2019), and second, because these variables were available in both datasets.
Previous literature led us to propose two competing a priori hypotheses for the relationship between severity of presentation at baseline and treatment response. First, severity of initial presentation, especially when conceptualized in terms of (impaired) social relationships, tends to predict poorer treatment response (e.g., Wang et al., 2018), and so we predicted (H1 A ) that severity at baseline might be negatively associated with change in loneliness over time. Conversely, on the basis of prior evidence that patients with more severe symptoms, especially on the primary target of treatment, tend to have a better treatment response (e.g., Elkin et al., 1995), we predicted (H1 B ) that baseline severity might be positively associated with change in loneliness over time. We further hypothesized (H2) that greater engagement with the G4H program, as indexed by attendance record, would be associated with a more positive treatment response. Demographic characteristics were examined as potential predictors for completeness; however, no specific hypotheses were made regarding these variables given limited evidence in the literature of any systematic differences in treatment response on the basis of gender, age, or ethnicity (Hamilton & Dobson, 2002;Nilsen et al., 2013;Parker et al., 2011).

Study 1
Study 1 was conducted with a clinical sample of 95 community members who completed G4H as part of a Phase II randomized controlled trial that sought to establish the program's efficacy among people with mental ill-health and loneliness (trial procedures and efficacy data are reported elsewhere; Haslam et al., 2019). 1 Participants were recruited via advertising at mental health, primary care, and allied health services throughout the region. G4H was facilitated in accordance with the manual by provisional psychologists under supervision of a clinical psychologist to ensure adherence to the protocol.

Method
Participants ranged in age from 17 to 68 years (M = 30.84; SD = 26.50). Full demographic details are provided in Table I. Inclusion criteria were that participants subjectively reported loneliness and psychological distress and reported either (a) diagnosis of a mental health condition by a health professional, or (b) symptoms of depression corresponding to at least the mild clinical range on the PHQ-9 (> = 5). We prioritized depression symptomatology because the evidence for the protective role of social identity processes is strongest for depression, relative to other clinical presentations (Cruwys, Haslam, Dingle, Haslam, & Jetten, 2014;Postmes, Wichmann, van Valkengoed, & van der Hoef, 2019). The sample were, on average, in the moderate clinical range for depression at baseline on the PHQ-9 (M = 12.89, SD = 5.99). Ethical approval was provided by the University of Queensland (#2014000731) and participants provided informed written consent.

Measures
Predictors were measured at baseline (prior to program commencement) and treatment outcome was measured at 4-month follow up (i.e., 2 months after program completion). Of the 95 participants, 68 (72%) had follow-up data available.

Loneliness
The widely used and extensively validated eight-item Roberts UCLA scale (Roberts et al., 1993) was used to measure loneliness at baseline (α = .70) and follow-up. Items such as "I feel left out" were rated on a four-point scale from 1 (never) to 4 (always). Compared to the population distribution for this measure of loneliness found in validation studies (M = 7.06; SD = 4.32; Roberts et al., 1993), almost the entire sample (98%) were >1 standard deviation higher than the mean, and 66% were >2 standard deviations higher, M = 17.08, SD = 2.92.

Depression
The seven-item depression subscale from the Depression, Anxiety, and Stress Scale-21 (Lovibond & Lovibond, 1995) was used to measure depression at baseline (M = 20.25; SD = 9.57). Consistent with prior research, we found that this widely-used measure was reliable in our sample (α baseline = .85). Items such as "I felt that I had nothing to look forward to" were rated on a four-point scale from 0 (did not apply to me at all) to 3 (applied to me very much, or most of the time).

Social Anxiety
The validated short form of the Social Phobia Inventory (the mini-SPIN; Connor et al., 2001) was used to measure severity of social anxiety symptoms at baseline (M = 3.39; SD = 1.04; α = .81). Items such as "I avoid activities in which I am the centre of attention" were rated on a five-point scale from 1 (not at all) to 5 (extremely).

Diagnostic Status
Participants were asked "have you ever been diagnosed with a mental illness by a health professional?" (emphasis in original). Due to its sensitive nature, participants were reminded that this question was optional and so could choose to skip it. Participants who answered yes were asked to indicate the type of health professional who provided the diagnosis, the calendar year in which it was made, and what the diagnosis was. Previous research has found strong correspondence between symptom severity and diagnostic history using this measure (Cruwys & Gunaseelan, 2016). Half of the participants (49.5%) reported having a formal mental health diagnosis, however, this is likely an underestimate because the optional nature of this question meant that 13.7% chose not to respond.

Attendance
Participant attendance at each of the five G4H sessions was recorded by facilitators. When a participant missed a session, they were offered the opportunity to attend a catch-up session with their facilitator and those who took this opportunity were recorded as having attended. Total attendance across the program was calculated as a total score ranging from 0 to 5, and treated as a continuous predictor in the analyses. The modal number of sessions attended was 5, the median was 4 sessions, and the mean was 3.55 (SD = 1.77). Overall, 26.3% of people did not attend at least 60% of the sessions (established in previous trials as the threshold for protocol adherence).

Demographics
Age, gender identity, and ethnicity were collected at baseline. Based on participants' free-responses, a numeric ethnicity variable was coded as either White (1) or ethnic minority (2). 2

Analysis Plan
All consenting participants who provided baseline data are included in this sample, even if they subsequently withdrew from the program or did not provide follow-up data. We used full information maximum likelihood (FIML) to manage missing data, which meant that while not all participants had data available for all measures at all timepoints, all available observations (163 in this case) could be included in the analyses even for people whose data were incomplete (Larsen, 2011). The hypotheses were tested using mixed-effects models in R (packages lme4, emmeans, and sjplot), with a p-value < .05 classified as significant. Such modelling had several advantages for analysing these data compared to a simpler regression approach. First, mixed-effects modelling allows the dependencies in the data to be modelled appropriately, which in this case included random intercepts for participants (n = 95) and for which therapy group that those participants were embedded within (n = 13). Second, FIML is readily implemented in mixed-effects modelling to account for missing data (restricted estimated maximum likelihood is the default in lme4, but FIML can be specified). This approach has been widely recommended for analysis of clinical trial data with missing observations (Beunckens et al., 2005;Molenberghs et al., 2004;Siddiqui, 2011). Timepoint was a categorical fixed effect and each model included a fixed effect for one of the hypothesized predictors (e.g., age; baseline depression severity; attendance). The interaction between these two fixed effects was the key term for the test of the hypotheses (e.g., whether baseline depression severity predicted a different rate of change in loneliness across time). Full details of the final models are available in the Supplementary Appendix.
A potential weakness of this analytic method (specific to analyses with loneliness as a predictor) is that it was possible for it to be affected by two statistical artefacts. Specifically, a floor effect, where the scale is limited in its capacity to detect improvement for those with low initial scores, and regression to the mean, whereby those with the most severe loneliness at baseline were most likely to experience improvement regardless of the intervention (Morton & Torgerson, 2003). To correct for these, two sensitivity analyses were conducted. The first, to address the potential floor effect, used multiple regression including a modified dependent variable which represented each participant's improvement in loneliness as a proportion of their possible improvement on the scale. The few participants whose loneliness symptoms increased during treatment were coded as "0" improvement for this purpose, and those with missing data were excluded using listwise deletion. This analysis corrects for a floor effect by weighting participant change scores, such that those participants whose scores are lowest at baseline (and so can improve the least) are given a greater weighting than those participants whose scores are highest at baseline (and so can improve the most; following Gelman & Hill, 2007; see also Šimkovic & Träuble, 2019). The second sensitivity analysis, to address the potential regression to the mean effect, used a hybrid structural equation model with FIML for missing data to calculate the latent decrease in loneliness over time (akin to a residualised change score; following McArdle, 2009; see also Webb et al., 2017;Seymour-Smith et al., 2021). This is defined, literally, as that part of the follow-up loneliness score that is not identical to the baseline loneliness score. This allows an assessment of the relationship between baseline loneliness and loneliness change that relates to true initial value dependence, rather than regression to the mean (McArdle, 2009).

H1:
The role of baseline symptom severity in predicting outcomes Loneliness Unsurprisingly, baseline loneliness and average loneliness across timepoints were strongly associated, β = .82, p < .001. The main effect of timepoint was also significant, β = -.77, p < .001. There was a significant timepoint * baseline loneliness interaction, β = -.47, p < .001. This reflected the fact that (consistent with H1 B but contrary to H1 A ), participants with more severe loneliness at baseline experienced a steeper decline in their loneliness over time. The estimated marginal means are presented in Figure 1.
The first sensitivity analysis assessed what proportion of possible improvement in loneliness participants achieved. The results were identical, with baseline loneliness severity positively predicting improvement in loneliness, F(1,66) = 5.53, p = .022. The second sensitivity analysis assessed the relationship between loneliness at baseline and latent decrease in loneliness at follow-up. Again, we found that baseline loneliness severity positively predicted improvement in loneliness, β = .55, p < .001.

Depression
Depression severity was significant and positively associated with loneliness across timepoints, β = .33, p < .001. The main effect of timepoint was also significant, β = -.77, p < .001. However, depression severity at baseline did not interact with time to predict the degree of improvement in loneliness, β = -.20, p = .110.

Social Anxiety
Social anxiety severity at baseline was not associated with loneliness across timepoints, β = .17, p = .060. The main effect of timepoint was significant, β = -.76, p < .001. The interaction between timepoint and baseline social anxiety severity was also significant, β = -.28, p = .030. This reflected the fact that, aligning with H1 B (but contrary to H1 A ), G4H participants with more severe social anxiety at baseline experienced a greater reduction in loneliness over time (see Figure 2).

Diagnostic Status
Mental health diagnostic status was not significantly associated with loneliness across timepoints, β = .15, p = .459. The main effect of timepoint was also nonsignificant, β = -.22, p = .639. There was also no interaction between timepoint and diagnostic status Attendance was not associated with loneliness on average across time, β = −.01, p = .798. The main effect of timepoint was also non-significant, β = .32, p = .535. However, consistent with H2, attendance interacted with time to predict degree of decline in loneliness, β = -.25, p = .037. As can be seen in Figure 3, this interaction arose from the fact that participants who attended less than 3 of 5 the G4H sessions experienced no significant change in loneliness, while the loneliness scores of those who attended all the sessions reduced by approximately 18%, or a full standard deviation from baseline.

Demographic Predictors of Treatment Outcome
Neither gender, age, nor ethnicity interacted significantly with time to predict the degree of change in loneliness (all ps > .130). Of these three variables, only age significantly predicted loneliness, β = .07, p = .012, with older participants reporting lower levels of loneliness across timepoints.

Discussion
Study 1 identified three significant predictors of treatment response to loneliness intervention. Consistent with H1 B , participants with a more severe presentation at baseline-in terms of either loneliness or social anxiety-tended to derive the greatest benefit from the G4H program. In addition, and consistent with H2, those who engaged more with the program, as indexed through attendance, tended to derive greater benefit. Finally, there was no evidence of demographic predictors of treatment response, with gender, age, and ethnicity all unrelated to degree of improvement in loneliness.

Study 2
Study 2 sought to replicate the above findings in a second clinical trial that had several strengths in comparison to Study 1. First, the follow-up period was longer, with the primary follow-up timepoint being at 12 months after baseline. Additional follow-up assessments were also available at program completion and 6 months after baseline, and so all four timepoints were utilized to model change in loneliness over time. Second, more comprehensive and reliable measures were available in Study 2, both for the primary outcome variable (loneliness) as well as for one of the focal predictors (social anxiety). Finally, Study 2 addressed a limitation of our assessment of H2 in Study 1 relating to the fact that a significant proportion of participants whose attendance was low were lost to follow-up-thereby reducing confidence in our capacity to estimate their rate of change. In contrast, Study 2 had extremely high retention of participants at the follow-up assessment points (>90% completed at least one). This meant that follow-up data were available for most of those participants who attended few sessions.  Table I for demographics). Inclusion criteria were that participants fell within the target age range (15-25 years), subjectively reported loneliness and low mood, and reported either (a) diagnosis of a mental health condition by a health professional, (b) symptoms of depression corresponding to at least the mild clinical range on the PHQ-9 (> = 5), and/or (c) loneliness corresponding to >40 on the UCLA scale. The sample were in the moderate clinical range for depression on average at baseline on the PHQ-9 (M = 14.01, SD = 5.21). Less than a third (27.4%) of the sample reported a formal mental health diagnosis, however, this is likely to be an underestimate because participants were reminded that this question was optional and 11.9% chose not to respond.

Measures
The following variables used the same measures as in Study 1: Depression (M = 19.76; SD = 8.54), attendance, diagnostic status, and demographics. Attendance was high (mode = 5, median = 5; M = 4.17; SD = 1.40) with only 14.3% of participants not completing at least 60% of the program. Completion of the follow-up timepoints was also high, with 83.3% of participants completing the postprogram measures, 85.7% completing the 6-month follow-up, and 84.5% completing the 12-month follow-up.

Loneliness
The 20-item UCLA-3 scale (Hays & DiMatteo, 1987) was used to measure loneliness at baseline (α = .90), program completion, and at 6-month and 12-month follow up. Items such as "I felt isolated from others" were rated on a four-point scale from 1 (never) to 4 (often). Compared to the population distribution for this measure of loneliness found in validation studies (M = 32.82, SD = 9.43; Shevlin, Murphy, & Murphy, 2015), almost the entire sample (98%) were higher than the population mean, and 84.5% were >1 standard deviation higher (M = 51.95, SD = 9.94).

Social Anxiety
The 17-item Social Phobia Inventory (SPIN; Connor et al., 2000) was used to measure severity of social anxiety symptoms at baseline (M = 36.63; Psychotherapy Research 615 SD = 12.95; α = .90). Items such as "Talking to strangers scares me" were rated on a five-point scale from 1 (not at all) to 5 (extremely).

Analysis Plan
As in Study 1, all consenting participants who provided baseline data are included in this sample, even if they subsequently withdrew from the program or did not provide follow-up data. FIML was used to manage missing data and all available observations (297 in this case) were used for analyses. Mixed-effects models in R were again used to assess the hypotheses, with observations (297) nested within participants (84) who were nested within therapy groups (13). Timepoint was treated as a categorical predictor, and so the test of the interaction effect was the χ 2 comparing the model including the three vectors representing the interaction (each other timepoint vs. baseline) to the model which did not include the interaction. The package emmeans was used to additionally fit the interaction where timepoint was treated as a linear, quadratic, cubic effect, to assess which of these models provided the best fit for the trend. Full details of the final models are provided in the supplementary appendix. As in Study 1, two sensitivity analyses were conducted to address floor effects (using a weighted version of the dependent variable) and regression to the mean (using a latent change score).
ResultsH1: Baseline severity as a predictor of treatment outcome Loneliness. Unsurprisingly, baseline loneliness was associated with average loneliness over time, β = .57, p < .001. The addition of timepoint significantly improved the model over one including only baseline loneliness, χ 2 (3) = 101.87, p < .001. Consistent with H1 B , adding the interaction between time and baseline loneliness significantly improved the model, χ 2 (3) = 40.45, p < .001. The linear estimation of the interaction effect was the best fit for the data, t(240) = −5.99, p < .001, such that participants with more severe loneliness at baseline experienced a steeper reduction in their loneliness across time. In other words, more severe loneliness at baseline predicted a steeper trajectory of improvement in loneliness (see Figure 4). The first sensitivity analysis yielded identical results, with baseline loneliness severity positively predicting improvement in loneliness, F(1,69) = 5.16, β = .26, p = .026. The second sensitivity analysis also provided consistent evidence, with baseline loneliness severity positively predicting improvement in loneliness, β = .43, p < .001.
Social Anxiety. Severity of social anxiety at baseline was not significantly associated with average loneliness over time, β = .11, p = .166. The addition of timepoint significantly improved the model over one including only baseline social anxiety, χ 2 (3) = 95.69, p < .001. Adding the interaction between time and baseline social anxiety significantly improved the model further, χ 2 (3) = 10.07, p = .016. The linear estimation of the interaction effect was the best fit for the data, t(225) = −2.85, p = .005. This reflected the fact that social anxiety severity at baseline predicted a steeper decline in loneliness over time (see Figure 5).

Depression
Depression symptom severity at baseline was significantly positively associated with loneliness across timepoints, β = .24, p = .003. The addition of timepoint significantly improved the model over one including only baseline depression, χ 2 (3) = 95.42, p < .001. Adding the interaction between time and baseline depression significantly improved the model further, χ 2 (3) = 13.77, p = .003. The linear estimation of the interaction effect was the best fit for the data, t(231) = −3.72, p < .001. This arose from the fact that while depression severity at baseline was positively associated with loneliness across both timepoints, it also predicted a steeper reduction in loneliness over time (see Figure 6).

Diagnostic Status
Finally, mental health diagnostic status was not significantly associated with loneliness on average across timepoints, β = -.10, p = .580. The addition of timepoint significantly improved the model over one including only diagnostic status, χ 2 (3) = 75.05, p < .001. Adding the interaction between time and diagnostic status significantly improved the model further, χ 2 (3) = 11.38, p = .010. The linear estimation of the interaction effect was the best fit for the data, t (196) = −3.17, p = .002. Again, consistent with the pattern for other predictors and with H1 B , having a formal mental health diagnosis at baseline was associated with a steeper reduction in loneliness, with the difference between conditions widening between 6 and 12 months after treatment completion (see Figure 7). H2: Program engagement as a predictor of treatment outcome Attendance was not significantly associated with loneliness on average across timepoints, β = -.05, p = .450. The addition of timepoint significantly improved the model over one including only attendance, χ 2 (3) = 95.08, p < .001. Adding the interaction between time and attendance did not significantly improve the model when timepoint was treated as categorical, χ 2 (3) = 5.19, p = .159. However, when timepoint was treated as a linear variable, the interaction effect was significant and the best fit for the data, t(236) = −2.05, p = .041. The pattern of means was such that participants who were lonelier at baseline were more likely to attend groups regularly. They subsequently experienced a steeper decrease in their loneliness symptoms that was sustained across the follow-up period. By contrast, participants with poorer attendance experienced less improvement in loneliness and this tended to trend upwards across the follow-up timepoints (see Figure 8).

Demographic Predictors of Treatment Outcome
Gender, age, and ethnicity were each assessed either as main effects or in interaction with time to predict degree of improvement in loneliness. None of these effects were significant (all ps > .068).

Discussion
Study 2 had more available observations than Study 1 as a result of outcome data being collected across four timepoints and having very little attrition. It also used a more comprehensive and reliable measure of loneliness than Study 1. Yet, like Study 1, it provided robust support for H1 B .
Here, though, all four indicators of baseline symptom severity (loneliness, depression, social anxiety, and mental health diagnosis) were significant predictors of a greater decrease in loneliness over time. That said, the measure of mental health diagnostic history was self-report and thus we can be less confident of its validity. However, given that all four of indicators of baseline symptom severity showed the same pattern of results in Study 2, this concern does not affect the overall conclusions. Furthermore, Study 2 found a dynamic relationship between loneliness and program engagement. People with more severe loneliness at baseline were more likely to engage with the program by attending a greater number of sessions. This is especially notable given the overall high adherence in this trial, as there was less variability in attendance than would likely be seen in usual care (Swift & Greenberg, 2012). Greater engagement, in turn, predicted a steeper decline in loneliness over time, and this was maintained at follow-up. Indeed, there was a trend whereby participants with lower attendance saw a trend toward loneliness recurrence over the follow-up period. The much longer followup period in Study 2 meant that, unlike Study 1, it was able to assess trends of this form, long after the active treatment phase of the study was completed. Study 2 found no evidence for any demographic predictors of treatment response (in terms of age, gender, or ethnicity). In particular, unlike Study 1, there was no evidence of a main effect of age on loneliness. It is possible that these findings may be explained by the very different age profile of the two studies. In Study 2 the age range was very narrow (due to the targeting of this trial toward young people specifically) and as a result it was less powered to detect age differences in loneliness than might be the case where participants represented the full age range in the population.

General Discussion
This project investigated predictors of treatment outcome in a loneliness intervention, using data from two clinical trials of GROUPS 4 HEALTH. Across two studies the primary consistent finding was that more vulnerable people (especially on social dimensions, i.e., more severe loneliness and social anxiety) tended to derive more benefit from the program (defined as degree of positive change over time). While some previous research has found similar patterns (e.g., such that people with more severe eating pathology tend to get greater benefit from an eating disorder prevention program (Butryn et al., 2014;Stice et al., 2008)), findings have been inconsistent in this regard. Indeed, lacking strong supportive social networks has been found to predict poorer prognosis for several conditions, especially depression (Cientanni et al., 2019;Ezquiaga et al., 1999). However, where the present research differs from previous investigations is that G4H specifically targets and seeks to remedy loneliness. In this context, it appears that participants with a stronger need for the program were exactly those who experienced greatest benefit. Furthermore, most previous studies have not attended to the risk of regression to the mean and floor effects, and how these may confound investigations of treatment response. Our sensitivity analyses suggested that lonelier participants did not only experience greater improvement, but also achieved a greater proportion of the total possible improvement that our measures were capable of detecting.

Psychotherapy Research 619
The second consistent finding across the two studies was that degree of engagement in the intervention (indexed as frequency of attendance) was robustly related to better outcomes. While this is to be expected, it increases confidence that the effectiveness of the G4H is not confounded with other factors but is indeed derived from program content and/or the group-based learning environment in which it is conducted. Less intuitive here, though, was the finding that, in Study 2, attendance was positively associated with loneliness severity at baseline. This suggests that unmet social needs motivated the loneliest participants to engage most with treatment (as also seen with primary care attendance; see Cruwys et al., 2018). Moreover, given that this attendance led, in turn, to better outcomes, this supports the conclusion that G4H is an appropriate and effective strategy for people experiencing loneliness to fulfil their need for social connection, both in the short and long term.

Implications
The finding that psychotherapeutic interventions can benefit those in greatest need is a positive outcome from a clinical practice perspective. It also has deeper theoretical implications, suggesting that we need to utilize theoretical frameworks that do not conceptualize mental ill-health and loneliness as intractable, or clinical improvement as inevitably slow, expensive, and hard-won (Deacon, 2013;Lilienfeld et al., 2013). By contrast, major improvements can sometimes be achieved with only brief interventions (see also Catanzano et al., 2020;McCabe et al., 2018). However, such radical improvements are most likely to occur if interventions correctly target the underlying maintaining factors for a person's problems. Accordingly, researchers have called for more investment in "wise" interventions (Walton, 2014) that invest scarce resources in targeting those leverage points that will achieve the greatest benefit with the lightest touch intervention possible. The evidence presented here suggests that GROUPS 4 HEALTH may be one such brief intervention, achieving large improvements in loneliness among those who need it the most by specifically targeting the lack or loss of social identities.
A further implication relates to the conceptual separation of prevention from cure in intervention research. The distinction between prevention, early-intervention, and treatment in the continuum of care may be more arbitrary in practice than has sometimes been asserted in theory. Although our studies focused on samples with mild-to-moderate clinical symptoms, its findings accord with previous research on interventions designed for prevention or early intervention, which actually showed the most pronounced benefits among those with the most severe symptoms (see also Butryn et al., 2014b). While our studies similarly found that people with severe symptoms obtained greater benefit, it was not the case that there was no benefit for those with milder symptoms. Taken in the context of the broader literature, this suggests that it is possible for the same intervention to succeed in spanning both early-intervention and treatment domains. This is exciting because it suggests that it might be possible to manage scarce healthcare resources by reducing the separation between public health, primary care, and specialist care .
Finally, this study has practical implications for mental health practice. In particular, recent years have witnessed a growing emphasis on personalized healthcare which requires clinicians to "match" patients to the most suitable intervention. The evidence from the present study is that visible characteristics of patients, such as gender, age, and ethnicity, were not important in predicting the effectiveness of a loneliness intervention. This is largely consistent with previous meta-analytic findings both in the context of loneliness interventions and for psychotherapeutic treatments more broadly (Dingle et al., 2021;Hickin et al., 2021;Nilsen et al., 2013). By contrast, patient differences on treatment-relevant dimensions (notably symptom severity and diagnosis) are important in predicting not only their likely benefit from treatment, but also whether they will engage with treatment at all. However, the direction of this effect may go in the opposite direction to clinical intuition in some cases. Anecdotally, for example, clinicians have suggested that a group program like G4H is less likely to benefit patients with severe social anxiety, or that brief interventions will be insufficient for people with diagnosable mental illness. However, in contrast to this view, the findings presented here suggest that such people are precisely those who are most likely to experience the greatest overall improvement from the program. Where resources are limited, it thus appears that directing loneliness interventions toward the most vulnerable populations (by assessing severity of loneliness and mental ill-health) is likely to yield the greatest benefit.

Strengths, Limitations, and Future Research
Strengths of this research included its use of two separate, rigorously screened clinical samples. The results were largely in agreement, despite diversity in composition of the participants on dimensions including age, cultural background, and presenting problems. However, like all research, the current project had limitations. Most obviously, it only evaluated one form of loneliness intervention-G4H. Although this is an intervention with a strong evidence base (from Phase I, II and III trials), it has some unique features (notably, the focus on groupbased social connection) which may limit the generalisability of these findings to other types of interventions. The use of secondary data to test our hypotheses also means that we were reliant on the measures and samples available, and so caution is warranted in generalizing beyond this context. Another weakness is that the study design was not able to shed light on whether there were some participants whose symptom severity (in terms of social anxiety in particular) may have been a barrier to consenting to participate in the trial at all. If selection effects of this form were at play, then the actual relationship between baseline severity and treatment response would likely be curvilinear in the population, with benefit greatest for those with high (but not the very highest) symptom severity.

Conclusions
The current project provides the first evidence to date of factors that predict people's treatment response to loneliness intervention-with a specific focus on G4H. Here prior research provided a contradictory picture of whether baseline symptom severity would be positively or negatively associated with treatment response. However, in data drawn from two clinical trials, we found consistent evidence that symptom severity (especially on the dimensions of loneliness and social anxiety) predicted a more positive trajectory in loneliness over time.
For GROUPS 4 HEALTH specifically, this analysis helps to flesh out a burgeoning evidence base of its efficacy by providing the first answers to the question of who is most likely to benefit from the program. Here our findings suggest that the program is beneficial for diverse demographic groups, but that it may lead to the largest changes among those participants who have the greatest need (as indexed by baseline severity of social and health impairment). Importantly too, additional analysis gives us reason to believe that this is unlikely to be simply a reflection of statistical artefacts-notably floor effects or regression to the mean.
There are reasons for concluding, then, that misery loves company. Typically, this proverb is understood to imply that sorrow is socially toxic-with those who are sad seeking out, and gaining comfort from, the sadness of others. However, our research supports a more positive invocation of the proverb-one which speaks to the fact that when we are down, we can be lifted by the community of others in ways that might not otherwise be possible. Indeed, in these terms, the value of GROUPS 4 HEALTH as a remedy for loneliness is precisely that it helps to create that company where it is most needed.
Notes 1 Our sample for the present analysis (n = 95) included both those who were either randomly assigned to G4H and those who were assigned to the waitlist and chose to complete G4H after their waitlist period was completed. 2 We recognise that this approach does not represent the diversity of cultural and racial backgrounds in the ethnic minorities group. However, given the small numbers in each ethnic minority category, it was necessary to provide enough statistical power to detect an effect, and does allow for comparisons between the culturally dominant group in the country of data collection (Australia) versus groups that face some degree of marginalisation or exclusion (see Connelly et al., 2016 for discussion).