Loneliness is a growing public health concern in aging societies and is associated with higher risks of morbidity and mortality [1]. Psychosocial interventions can reduce loneliness [2]; among these, information and communication technologies, including social robots, have attracted attention as scalable options [3,4]. In a previous randomized controlled trial (RCT), we found that a 4-week intervention using an operator-mediated humanoid communication robot that enabled conversations with operators and family members reduced loneliness and improved psychological well-being among community-dwelling older adults living alone in Japan [5].

Notably, previous studies have primarily reported average intervention effects [4], which may mask variation in individual responses. Existing subgroup analyses have focused on intervention characteristics [2,4]. Thus, little is known about which participant characteristics are associated with greater changes in loneliness during social robot interventions. Identifying such characteristics may generate hypotheses for future trials examining how social robot interventions could be better targeted under resource-constrained scenarios.

In this secondary analysis of our previously published RCT [5], we used intervention-phase data to explore which participant characteristics were associated with greater reductions in loneliness during the intervention. We explored which participant characteristics were associated with greater reductions in loneliness during the intervention, hypothesizing that less frequent preintervention social contact would be associated with larger decreases in loneliness.

Of the 67 participants, 94% (n=64) were female, and the mean age was 82.3 (SD 6.5) years. Loneliness decreased by a mean of 2.6 (SD 6.4, range −18 to 13) points.

Table 1 presents the factors related to improvements in loneliness. In model 1, less frequent social contact was associated with a greater reduction in loneliness (b −2.14 per category toward less frequent contact, 95% CI −4.31 to −0.04), as was high preintervention loneliness (b −2.52, 95% CI −5.11 to −0.07). Participants assigned to the immediate sequence exhibited less improvement than those assigned to the delayed sequence (b 3.16, 95% CI 0.47-5.86). No other characteristic was associated with changes in loneliness. In model 2, the interaction between less frequent social contact and high loneliness was not statistically significant (b –0.02, 95% CI –5.30 to 5.25).

To our knowledge, this is the first study to examine which participant characteristics were associated with greater reductions in loneliness during a social robot intervention. In model 1, less frequent social contact and higher loneliness before the intervention were each associated with greater reductions in loneliness. The association with less frequent social contact may reflect greater unmet needs for interaction and conversation among participants with little social contact. Social isolation and subjective loneliness are related but distinct [8], and the interaction between them was not statistically significant, although the wide CI renders this finding inconclusive.

Likewise, higher preintervention loneliness was associated with greater reductions in loneliness. This finding should be interpreted with caution. As preintervention loneliness was both a predictor in the model and the initial value used to calculate the change-score outcome, the observed association may partly reflect regression to the mean [9] and mathematical coupling between preintervention loneliness and the change score. We also pooled the intervention-phase data rather than comparing the on-robot and off-robot periods for the same person; therefore, we cannot attribute this association to the robot. Accordingly, the association with preintervention loneliness should be regarded as exploratory.

Overall, these findings should be viewed as hypothesis-generating rather than as evidence of differential treatment benefit. In particular, preintervention social contact warrants testing as a potential effect modifier in future trials.

This study has several limitations. The sample was small, predominantly female, very old, drawn from 1 region, and restricted to individuals who reported loneliness; therefore, the findings may not be generalizable. Moreover, the analysis was exploratory, included multiple predictors, and was not corrected for multiple testing. It identified characteristics associated with loneliness reduction rather than treatment-effect modification. The small sample also limited the power to detect an interaction between preintervention social contact and preintervention loneliness. The smaller reduction observed in the immediate sequence remains difficult to explain but may reflect period-related or other contextual differences because the intervention phases occurred during different study periods for the 2 sequences.

In conclusion, less frequent preintervention social contact was associated with greater reductions in loneliness during the social robot intervention. The association with higher preintervention loneliness should be interpreted cautiously. Future trials should examine whether social contact modifies treatment effects.