Conditional pathways to loneliness in older adults with vision impairment: a decision-tree analysis Authors: Ya-Han Chang, Silvia Sörensen, Michelle Putnam Published: 2026 ABSTRACT AbstractBackground and ObjectivesThe prevalence of vision impairment continues to increase globally due to population growth and aging. Among older adults with vision impairment, loneliness is common. Prior studies assessing loneliness risk have relied on regression-based approaches, which may not capture complex, heterogeneous conditional pathways. We apply a theoretically guided machine learning approach to examine loneliness risk among older adults with age-related macular degeneration (AMD).Research Design and MethodsUsing baseline data from 210 adults with AMD (Agemean = 80.1 years) enrolled in a randomized controlled trial, we applied a decision tree method using measures of legal blindness, emotional responses to vision loss, social life satisfaction, anxiety, depression, optimism, future planning, and sociodemographic factors to investigate heterogeneous pathways to loneliness.ResultsDecision tree models demonstrated stable discrimination (AUCtrain = .88, AUCtest = .78, DeLong 95% confidence interval: 0.68–0.88) and robust classification performance across thresholds. Social life satisfaction emerged as the primary branching condition differentiating loneliness outcomes. Distinct pathways suggested that high social life satisfaction was strongly associated with lower loneliness; optimism served as a compensatory resource when social satisfaction was lower. Additional pathways highlighted the effects of vision loss severity, anxiety, and emotional responses to vision loss on loneliness. Depressive symptoms differentiated loneliness outcomes within a lower-anxiety branch, suggesting depression may overshadow protective factors.Discussion and ImplicationsThese findings suggest that loneliness arises through heterogeneous psychological pathways rather than single risk factors. By identifying conditional configurations of psychosocial and health-related factors, interpretable decision-tree models may inform more targeted screening, evaluation, and intervention strategies.Keywords: Machine learning, Vision impairment, Depression, Optimism, Social life satisfaction This study used a theory-guided machine-learning approach with an interpretable decision-tree model to examine loneliness in older adults with age-related macular degeneration. The model captures complex, nonlinear, and conditional pathways linking psychosocial, emotional, and vision-related factors to loneliness, revealing patterns that are often difficult to capture with traditional regression methods. It includes psychosocial resources and vulnerability information that can support the development of targeted screening, risk stratification, and more personalized interventions. Findings provide a foundation for geropsychologists, social workers, and interventionists working with older adults who have sensory and social challenges. Loneliness has become a major global public health issue among older adults. According to the World Health Organization (2024), worldwide, approximately 16% of people experience loneliness. A growing body of evidence shows that loneliness is associated with poorer mental health (Bekhet & Zauszniewski, 2012; Hawkley & Cacioppo, 2010; Raymo & Wang, 2022 ), and worse physical and functional health outcomes (Bekhet & Zauszniewski, 2012; Dykstra et al., 2005; Hawkley & Cacioppo, 2010; Raymo & Wang, 2022) in later life. For instance, Bekhet and Zauszniewski (2012) found that older adults with higher loneliness reported higher levels of depression and anxiety symptoms. Furthermore, longitudinal studies indicate that loneliness is not static in later life but varies in response to changes in health, social networks, and life events (Bekhet & Zauszniewski, 2012; Dykstra et al., 2005). For example, Dykstra et al.’s longitudinal study (2005) found, first, stability in loneliness for the majority of participants but increases for 18% and decreases for 13% over the 7-year study period. Second, increases in loneliness were greater among those with better functional capacity and partnered status at baseline; third, participants whose functional status declined and networks shrank over time reported an increase in loneliness over a 7-year period, and those with improved functional capacity and increasing network size reported slight declines in loneliness. Thus, loneliness is shaped by multiple interacting factors that evolve across the aging process, with substantial heterogeneity in how older adults develop and experience loneliness. Although loneliness and social isolation are related constructs, they are conceptually distinct. Social isolation refers to the objective absence or limited frequency of social contacts, whereas loneliness reflects the subjective perception that one’s social relationships are inadequate in quantity or quality. This study focuses on loneliness, which individuals may experience without being socially isolated, and vice versa (Hawkley & Cacioppo, 2010; Raymo & Wang, 2022). Older adults with vision impairment are particularly vulnerable to loneliness (Alma et al., 2011; Kotwal et al., 2021), as vision loss not only limits daily functioning but also disrupts social life and increases psychological distress (Hodge & Eccles, 2013). Age-related macular degeneration (AMD) is the leading cause of irreversible vision loss among adults aged 65 and older in developed nations (Mitchell et al., 2018; Rein et al., 2022). AMD is a progressive retinal disorder affecting central vision. Individuals with AMD often experience difficulties with reading, recognizing faces, driving, and engaging in social and leisure activities (Stanford et al., 2009) and recognizing visual social cues (Klauke et al., 2023). These functional limitations may reduce social participation and threaten independence, thereby increasing vulnerability to loneliness. According to the Centers for Disease Control and Prevention (2024), AMD prevalence increases sharply with age, rising from approximately 11.6% among adults aged 60–64 to more than 60% among those aged 95–99. As the prevalence of vision impairment continues to increase globally (Chen et al., 2024; Swenor et al., 2020), understanding the mechanisms contributing to loneliness in this population has become an important research priority. Understanding the mechanisms underlying loneliness among older adults with AMD requires analytical approaches capable of capturing heterogeneous and conditional relationships among contributing factors. Such approaches may also provide a valuable framework for guiding future psychosocial screening, evaluation, and intervention planning. Traditional regression-based approaches have substantially advanced our understanding of loneliness by identifying significant associations among emotional outcomes, functional limitations, and social characteristics (Mick et al., 2018; Miyata et al., 2021; Mojon-Azzi et al., 2008; Salive et al., 1994; Silverstein et al., 2021). However, these approaches are less suited for identifying hierarchical conditional pathways where future planning, psychosocial resources, and vulnerabilities differentiate loneliness risk. Throughout this study, psychosocial resources and vulnerabilities are conceptualized as characteristics that may function as protective factors by buffering loneliness risk and as risk factors by increasing loneliness risk, respectively. Future planning, a multidimensional construct reflecting preparation for later life, is one of the key psychosocial characteristics. Its theoretical foundation is introduced in the following section. Psychosocial adjustment to AMD varies substantially depending on the severity of vision loss, coping resources, and social support, highlighting the heterogeneous nature of the condition (Stanford et al., 2009). Decision-tree pathway-oriented information extends beyond identifying individual risk factors averaged across a sample, as, for example, in epidemiological research; rather, it reveals how specific factors become more or less influential depending on preceding conditions, thereby potentially informing more targeted screening and intervention strategies. Accordingly, the present study applies a theory-informed decision-tree approach to identify distinct pathways to loneliness among older adults with AMD. By examining how psychological resources, emotional responses to vision loss, and future planning combine within different subgroups, this study aims to advance understanding of loneliness risk and inform individualized intervention efforts. Building on this rationale, the innovative analyses are organized around several key foci: (a) the protective role of psychosocial resources, including optimism and older adults’ satisfaction with their social life; (b) the contextual influence of vision loss severity; (c) the role of anxiety and emotional responses to vision loss, and (d) the role of depressive symptoms within lower-anxiety conditions. These foci illustrate how loneliness among older adults with vision impairment may unfold through hierarchical psychological pathways shaped by interacting psychosocial resources and vulnerabilities. Theoretical framework To guide the conceptualization of psychological and behavioral predictors of loneliness, the present study draws on two complementary theoretical perspectives. First, the Broaden-and-Build Theory of positive emotions (Fredrickson, 2001) suggests that positive emotional experiences broaden individuals’ thought–action repertoires and help build enduring psychological and social resources over time. Conversely, persistent negative emotional states may contribute to downward emotional spirals characterized by withdrawal, reduced coping capacity, and increased vulnerability to loneliness (Garland et al., 2010; Huxhold & Suanet, 2023; Segel-Karpas & Ayalon, 2020). This perspective guides the present study to examine how both psychosocial resources and vulnerabilities jointly shape loneliness among older adults with vision impairment. Second, the aging-related preparation (ARP; Sörensen et al., 2012) framework emphasizes proactive coping processes (Aspinwall & Taylor, 1997) that help individuals anticipate and adapt to age-related challenges. Preparation for Future Care describes cognitive and behavioral processes through which older adults anticipate potential care needs and plan accordingly (cf. Sörensen et al., 2011, 2017). Engagement in proactive planning behaviors has been associated longitudinally with improved psychological outcomes, including lower depression (Sörensen et al., 2012). In the context of progressive conditions such as vision loss, proactive planning may help older adults maintain social engagement and perceived control despite increasing functional limitations. Building on the broader ARP framework, future planning is conceptualized in this study as a multidimensional construct encompassing anticipatory beliefs, attitudes, psychological processes (i.e., awareness, avoidance, and deciding preference), and planning across multiple domains of later-life preparation (i.e., financial, aging in place, and care needs). Building on these perspectives, this study focuses on psychological and behavioral resources, including optimism, emotional responses to vision loss, and proactive future planning—that may shape loneliness risk among older adults with vision impairment. Individuals with similar levels of vision impairment may experience different psychosocial outcomes, suggesting that loneliness may arise through heterogeneous pathways, not a single uniform process. The study applies a theory-guided decision-tree approach to identify conditional pathways linking psychological resources, future planning, contextual conditions, and loneliness among older adults with AMD. Method Machine learning decision tree rationale A classification tree approach was selected because it allows the identification of heterogeneous conditional pathways through which combinations of predictors jointly influence an outcome. Unlike regression models that estimate independent effects averaged across the sample, decision trees recursively partition the data to reveal hierarchical structures among predictors. In contrast to latent class or latent profile analysis, which identify profiles at a single level, decision-tree analysis identifies conditional branching structures that reflect hierarchical relationships among predictors. These hierarchical pathway structures may provide insights that inform the future development of more individualized screening and intervention strategies, which traditional regression approaches are less well-suited to identify. To facilitate interpretation, the resulting decision-tree pathways were organized using a theory-guided interpretation to highlight broader configurations of psychological resources and vulnerabilities that contribute to higher or lower loneliness. Data and participants This study used data from the Macular Degeneration and Aging Study (Chang, 2026; Sörensen et al., 2015), a community-based randomized controlled trial focused on preventing depression and enhancing preparation for future care. Detailed descriptions of sampling procedures, inclusion and exclusion criteria, and survey administration are available elsewhere (Chang, 2026; Sörensen et al., 2015). We selected baseline data from 224 adults to explore the study’s aim, from which we excluded respondents with over 50% missing data on any multi-item measure while retaining cases with missing values on single-item indicators. After applying this criterion, the final analytic sample consisted of 210 participants. As shown in Table 1, Participants ranged in age from 60 to 96, with a mean age of 80.1. Of the participants, 65.2% identified as women, while 31% reported being legally blind. According to U.S. criteria, legal blindness is defined as best-corrected visual acuity of 20/200 or worse in the better-seeing eye and/or a visual field of 20 degrees or less (Social Security Administration, n.d.). More than two-thirds (69%) of participants were visually impaired from AMD but did not meet the legal blindness threshold. Regarding marital status, 38.1% were married or had a partner, whereas 61.9% were widowed, divorced, separated, or never married. Regarding living status, 48.6% lived alone, while 51.4% lived with someone. Table 1Demographic characteristics of participants.Characteristics N % M SD Age21080.18.1 Sex Male13765.2 Female7334.8 Marital status Married or had a partner8038.4 Others13061.9 Living status Lived alone10248.6 Lived with someone10851.4 Demographic characteristics of participants. Materials and measures A set of 30 predicting variables was included in the machine learning model, spanning sociodemographic and vision-related variables, psychological condition variables, and future planning variables. Table 2 provides detailed measurement information. Table 2Measurement characteristics and coding of predictor and outcome variables used in the classification tree analysis.VariablesCodes and informationNo. of itemsCronbach’s α Sociodemographic and vision-related variables Age1– Gender0 = Female, 1 = Male1– Marital status 1 = Married or living with partner2 = All other 1– Live alone0 = Yes, 1 = No1– Financial status 1 = Not at all adequate5 = More than adequate 1– Are you considered legally blind?0 = No, 1 = Yes1– Emotional response to vision loss 1 (recoded into 0) = all of the time5 (recoded into 100) = None of the time (better) 12.90 Future planning variables Becoming aware 1 = Not at all true of me5 = Completely true of me 6.78 Avoiding awareness 1 = Not at all true of me5 = Completely true of me 4.70 Gathering information 1 = Not at all true of me5 = Completely true of me 6.76 Deciding preferences 1 = Not at all true of me5 = Completely true of me 6.82 Concrete plans 1 = Not at all true of me5 = Completely true of me 6.71 Planfulness 1 = Not at all true of me5 = Completely true of me 5.75 Belief that planning is not useful 1 = Not at all true of me5 = Completely true of me 7.67 Expect to need care 1 = Not at all true of me5 = Completely true of me 5.66 Number of preference options considered for aging locationsCount of items endorsed7– Concreteness value of plan with highest concretenessHighest score among all preference items7– Worry about future care need 1 = Not at all true of me5 = Completely true of me 4.83 Compared to people your own age, how much do you prepare for the future?1 = Not at all, 5 = Some, 7 = A lot/Completely1– Compared to people your own age, to what extent would you say you have achieved your personal goals?1 = Not at all, 5 = Some, 7 = A lot/Completely1– Feel secure as a result of planning 1 = Very insecure5= Very secure 2.71 Accepting help if needed 1 = Disagree strongly4 = Agree strongly 6.66 Financial plan 1 = Disagree strongly4 = Agree strongly 5.81 Thinking about plans 1 = Disagree strongly4 = Agree strongly 6.78 Satisfied with plans 1 = Very dissatisfied5 = Very satisfied 3.88 Psychological resources and vulnerabilities State-Trait Anxiety Inventory (STAI) anxiety score1 = Low, 5 = High6.82 Depression symptoms0 = Not at all or less than 1 day, 1 = 1–2 days, 2 = 3–4 days, 3 = 5–7 days20.83 Perceived Stress Scale0 = never, 4 = very often10.86 How satisfied are you with your social life? 1 = Very dissatisfied5 = Very satisfied 1 Optimism 1 = I disagree a lot5 = I agree a lot 6.73 Target variable University of California, Los Angeles (UCLA) Loneliness Scale 1 = Below average2 = Above average 8.79 Measurement characteristics and coding of predictor and outcome variables used in the classification tree analysis. 1 = Married or living with partner 1 = Married or living with partner 2 = All other 2 = All other 1 = Not at all adequate 1 = Not at all adequate 5 = More than adequate 5 = More than adequate 1 (recoded into 0) = all of the time 1 (recoded into 0) = all of the time 5 (recoded into 100) = None of the time (better) 5 (recoded into 100) = None of the time (better) 1 = Not at all true of me 1 = Not at all true of me 5 = Completely true of me 5 = Completely true of me 1 = Not at all true of me 1 = Not at all true of me 5 = Completely true of me 5 = Completely true of me 1 = Not at all true of me 1 = Not at all true of me 5 = Completely true of me 5 = Completely true of me 1 = Not at all true of me 1 = Not at all true of me 5 = Completely true of me 5 = Completely true of me 1 = Not at all true of me 1 = Not at all true of me 5 = Completely true of me 5 = Completely true of me 1 = Not at all true of me 1 = Not at all true of me 5 = Completely true of me 5 = Completely true of me 1 = Not at all true of me 1 = Not at all true of me 5 = Completely true of me 5 = Completely true of me 1 = Not at all true of me 1 = Not at all true of me 5 = Completely true of me 5 = Completely true of me 1 = Not at all true of me 1 = Not at all true of me 5 = Completely true of me 5 = Completely true of me 1 = Very insecure 1 = Very insecure 5= Very secure 5= Very secure 1 = Disagree strongly 1 = Disagree strongly 4 = Agree strongly 4 = Agree strongly 1 = Disagree strongly 1 = Disagree strongly 4 = Agree strongly 4 = Agree strongly 1 = Disagree strongly 1 = Disagree strongly 4 = Agree strongly 4 = Agree strongly 1 = Very dissatisfied 1 = Very dissatisfied 5 = Very satisfied 5 = Very satisfied 1 = Very dissatisfied 1 = Very dissatisfied 5 = Very satisfied 5 = Very satisfied 1 = I disagree a lot 1 = I disagree a lot 5 = I agree a lot 5 = I agree a lot 1 = Below average 1 = Below average 2 = Above average 2 = Above average The assessment included five subscales (awareness, avoidance, gathering information, deciding preferences, and concrete planning) from a 29-item Preparation for Future Care Needs measure (Sörensen & Pinquart, 2001), two subscales from the Long-Term Care Planning (LTCP) scale (Friedemann et al., 2004), 12 selected items from the National Eye Institute Visual Function Questionnaire-25 (NEI-VFQ-25; Mangione et al., 2001) indicating emotional response to vision loss, State-Trait Anxiety Inventory (Spielberger, 1983), the Center for Epidemiologic Studies Depression Scale Revised (CESD-R; Eaton et al., 2004), Perceived Stress Scale (PSS; Cohen et al., 1983), UCLA Loneliness Scale (Version 3) (Russell, 1996), Life Orientation Test-R to assess optimism (LOT-R; Scheier et al., 1994) and a social life satisfaction question. We evaluated internal consistency for all multi-item scales using Cronbach’s α based on the original (non-imputed) raw data (Table 2). Unlike some clinical mental disorders’ screening instruments, the UCLA Loneliness Scale does not have a universally accepted clinical cutoff. In the present study, loneliness scores were dichotomized using the sample mean scores. Participants scoring above the mean were classified as having higher loneliness (48.6%), whereas those scoring at or below the mean (M = 1.83/4) were classified as having lower loneliness (51.4%). This approach was consistent with the study objective of identifying conditional pathways associated with relatively higher versus lower loneliness using classification tree analysis. Data analysis For machine learning analyses, we addressed missing data by performing median imputation for variables with less than 15% missingness, following established recommendations to maintain consistency across the analytic dataset (Acuña & Rodriguez, 2004). We removed cases if missing data exceeded 50% across multi-item scales; however, we did not apply this criterion to single-item measures. SPSS V.30 was used for data cleaning and descriptive analysis. A decision tree approach, specifically a classification tree, was employed to identify conditional pathways predicting binary loneliness among older adults with vision-threatening eye disease. Given the modest sample size of this retrospective dataset, we implemented several procedures to reduce overfitting and improve the robustness of model evaluation. We used an outcome-stratified train-test split to ensure the proportion of participants with high and low loneliness remained similar across training and testing datasets, thereby reducing the risk that model performance would be influenced by an imbalanced distribution of the outcome. We used repeated 10-fold cross-validation to evaluate model performance across multiple resampled subsets of the data, providing more stable performance estimates than a single partition. Hyperparameter tuning was performed to identify an appropriately complex tree structure and reduce the risk of overfitting. In addition, all categorical predictors were reasonably represented, with the smallest category comprising about 31% of the sample, reducing concerns related to sparse predictor categories. The dataset was partitioned into training (60%) (n = 127) and testing (40%) (n = 83) subsets, with stratification by the binary outcome. This stratified sampling was employed using the R package to preserve the proportion of high loneliness versus low loneliness cases across training and testing sets, thereby ensuring that class imbalance did not bias model estimation or evaluation. Within the training set, we applied a 10-fold cross-validation repeated 10 times procedure to optimize model hyperparameters. Repeating the procedure 10 times ensures stability of results. Specifically, the complexity parameter (cp), minimum split size (minsplit), and minimum bucket size (minbucket) were tuned during cross-validation to balance model complexity and predictive accuracy. The complexity parameter (cp) was tuned using a grid search from 0.0001 to 0.01. In addition to cross-validation and hyperparameter tuning, we evaluated model performance using multiple complementary metrics, including the area under the receiver operating characteristic curve (AUC) with DeLong confidence intervals and confusion matrix indices. Confusion matrix metrics were reported at both the conventional 0.50 threshold and the Youden-optimal threshold to evaluate model performance under a standard classification criterion and an empirically optimal classification threshold. The confusion Matrix measures included classification accuracy, sensitivity, specificity, precision, and F1 score, which evaluate the model’s performance relative to the data. Confidence intervals (CIs) for AUC were estimated using DeLong’s method. DeLong 95% CIs did not include 0.50, indicating discriminatory performance above chance levels. We reported multiple performance metrics because each captures a different aspect of classification performance. Feature importance (FI) values were examined to quantify the relative contribution of each predictor variable to model performance. In classification tree models, FI is derived from the total reduction in node impurity (e.g., Gini index) attributable to each predictor across all splits, with larger reductions indicating stronger predictive influence. Results From the analytic sample of 210 participants, the dataset was partitioned into a 60% training set (n = 127) and a 40% hold-out test set (n = 83) with stratification by the binary outcome. The classification tree model demonstrated strong discriminatory ability in the training set (AUC = 0.88) and good discriminatory ability in the independent test set (AUC = 0.78; DeLong 95% CI: 0.68–0.88) (Figure 1). Using a 0.50 probability cutoff, test set classification accuracy was 72.3%, with sensitivity = 79.1%, specificity = 65.0%, precision = 70.8%, and F1 = 74.7%. With Youden’s J index, the optimal classification threshold was 0.784, yielding a sensitivity of 0.80, a specificity of 0.65, and an overall accuracy of 0.72. The consistency of performance metrics across both the conventional 0.50 cutoff and the Youden-optimized threshold indicates that the model demonstrated stable classification performance. Confidence intervals for classification metrics are provided in Table S1. Feature importance analysis identified the top five predictors of loneliness as social life satisfaction, depressive symptoms, anxiety, perceived stress, and optimism (Figure 2). Receiver operating characteristic (ROC) curves demonstrating the classification performance of the final classification tree for loneliness among older adults with age-related macular degeneration. AUC = area under the receiver operating characteristic curve; CI = confidence interval. Relative importance of predictors contributing to the classification of loneliness among older adults with age-related macular degeneration. Feature importance values represent the total reduction in node impurity (Gini or deviance) attributed to each predictor across all splits in the classification tree, rescaled from 0 to 100, with the highest variable set to 100. The final decision tree (Figure 3) illustrates the hierarchical sequence of predictor splits that classify loneliness outcomes among older adults with AMD. The tree structure provides an interpretable representation of the conditional pathways through which individual differences in emotional, psychological, and vision health factors contribute to loneliness classification. In the training sample (n = 127), social life satisfaction emerged as the root split, indicating its central role in distinguishing between higher and lower loneliness outcomes. Final classification tree for predicting loneliness. The model was trained on 127 older adults with age-related macular degeneration. CESD-R = Center for Epidemiologic Studies Depression Scale – Revised. aTerminal nodes are numbered 1 to 8 from left to right and are detailed in the Results section. bEach terminal node label indicates the group with the majority predicted class, either high or low loneliness. cRatios ([minority class cases]/[total cases]) show the number of cases from the non-majority class relative to the total node size. dThe percentages show the proportion of the training set in each terminal node category, but this does not mean that all participants within a node belonged to the majority class. The final tree contained eight terminal nodes representing distinct conditional pathways from the root to each outcome classification. Rather than interpreting each pathway as an isolated decision rule, these pathways can be organized into broader patterns that reflect how combinations of emotional resources, vision-related conditions, and psychological factors shape loneliness risk. The following sections describe these pathways according to four broader conceptual patterns. Protective psychosocial resources: social life satisfaction and optimism The first pathway highlights the protective role of psychosocial resources in shaping loneliness outcomes. Social life satisfaction emerged as the root split of the decision tree, indicating its central importance in distinguishing higher from lower loneliness among older adults with vision impairment. Participants reporting the highest level of social life satisfaction (≥5 on a 0–5 scale) were predominantly classified into the low-loneliness group (33% of the training set; Figure 3), suggesting that positive evaluations of one’s social life may function as a strong protective factor against loneliness. A complementary pathway (terminal nodes 7 and 8) emerged among individuals with lower social life satisfaction (<4). Within this branch of the decision tree (30% of the training set), optimism differentiated loneliness outcomes: individuals with higher optimism were more frequently classified into the low-loneliness group (6%), whereas those with both lower social life satisfaction and lower optimism were predominantly classified into the high-loneliness category (24%). Together, these pathways illustrate two protective configurations associated with lower loneliness classification: high social life satisfaction and, among individuals with lower social satisfaction, higher optimism. Notably, these pathways emerged without further differentiation by vision loss severity within these branches of the tree. The decision tree structure further indicated that additional conditional pathways emerged when considering the severity of vision loss, as described in the next section. Vision loss severity context: legal blindness A distinct pathway emerged among participants with moderate social life satisfaction (score = 4) who were classified as legally blind (terminal node 2). Within this subgroup, individuals were predominantly classified into the low-loneliness category (9% of the training set; Figure 3), showing that moderate social satisfaction differentiated loneliness outcomes under conditions of severe vision loss. This pathway highlights the contextual role of legal blindness in shaping loneliness classification within the decision tree structure. Anxiety and emotional responses to vision loss Additional conditional pathways emerged among individuals with vision impairment who were not classified as legally blind (28% of the training set). Within this subgroup, participants reporting moderate social life satisfaction (score = 4) were further differentiated by anxiety levels. Among individuals with lower anxiety (<1.9; M = 1.41), an intermediate branch representing 24% of the training set was observed. Within this branch, loneliness classification was further differentiated by emotional responses to vision loss. Participants reporting more frequent negative emotional responses to vision loss (ERVL <64; M = 67.61) were predominantly classified into the low-loneliness group (terminal node 3; 6% of the training set). In contrast, participants with higher anxiety (≥1.9) were predominantly classified into the high-loneliness group (terminal node 6; 5% of the training set). These results illustrate how loneliness classification within this subgroup was first differentiated by anxiety and subsequently by emotional responses to vision loss. Their complex combination effect will be further explained in the Discussion section. Depressive symptoms as a differentiating condition A further conditional pathway emerged among individuals reporting moderate social life satisfaction who were not classified as legally blind and reported lower anxiety levels. Within this subgroup, participants with relatively better ERVL (≥64) were further differentiated by depressive symptoms. Individuals with lower depressive symptoms (CESD-R <.28) were predominantly classified into the low-loneliness group (terminal node 4; 9% of the training set), whereas those with higher depressive symptoms (CESD-R ≥.28) were classified into the high-loneliness group (terminal node 5; 9%). Together, these pathways demonstrate how loneliness classification among older adults with vision impairment emerged through a sequence of conditional psychological factors, including social satisfaction, anxiety, emotional responses to vision loss, and depressive symptoms. Discussion This study identified several conditional psychological pathways associated with loneliness among older adults with vision impairment. Social life satisfaction emerged as the primary branching condition, followed by contextual differences in vision loss severity, social life satisfaction, and emotional regulation processes, including optimism, anxiety, emotional responses to vision loss, and depressive symptoms. Together, these findings suggest that loneliness among older adults with vision impairment may arise through heterogeneous psychological conditional pathways rather than a single linear or nonlinear mechanism, as revealed by the interpretable machine-learning decision tree approach used in this study. Protective role of social life satisfaction and optimism in loneliness pathways The decision tree identified social life satisfaction as a primary condition distinguishing higher and lower loneliness among older adults with vision impairment. Individuals reporting high social life satisfaction were predominantly classified into the low-loneliness group, suggesting that positive evaluations of one’s social life may serve as an important psychological resource that buffers against loneliness. This finding is consistent with prior research highlighting the negative relationship between social well-being and loneliness. For example, Vespa et al. (2023) found that social well-being was negatively linked to loneliness among older adults. High social satisfaction may therefore function as a resilience factor that helps individuals maintain psychological well-being despite challenges with distressing and painful disruptions of relationships and reduced social interaction often experienced by people with vision impairment (e.g., Dunlop et al., 2025). The tree structure also revealed a complementary pathway among individuals reporting lower social life satisfaction. Lower social life satisfaction was further differentiated into a moderate level of social life satisfaction (=4) and lower (<4) social life satisfaction. Within this lower social life satisfaction subgroup, optimism differentiated loneliness outcomes: individuals with higher optimism were more likely to be classified into the low-loneliness group, whereas those with both low social satisfaction and low optimism were predominantly classified into the high-loneliness category. Although reduced social satisfaction is commonly associated with greater loneliness, this conditional pathway is consistent with Broaden-and-Build Theory (Fredrickson, 2001) and further suggests that optimism may function as a compensatory psychological resource specifically when social life satisfaction is limited. Longitudinal evidence further supports the role of optimism in shaping loneliness trajectories. A 10-year study in Zurich found that higher optimism was associated with lower levels of overall, emotional, and social loneliness across multiple follow-ups among older adults (Rius-Ottenheim et al., 2012). Additional research has linked optimism to a stronger sense of community and social connectedness (Kagan et al., 2024). These findings are consistent with theoretical perspectives such as the Broaden-and-Build Theory (Fredrickson, 2001; Garland et al., 2010; Huxhold & Suanet, 2023), which suggest that positive emotional orientations broaden individuals’ cognitive and behavioral repertoires, facilitating adaptive coping and social engagement. Conversely, limited optimism may contribute to negative expectations, reduced engagement in rewarding activities, and diminished confidence in managing age- or vision-related challenges. These cognitive–emotional processes may amplify the perceived impact of reduced social satisfaction and increase vulnerability to loneliness among older adults with AMD. Rather than implying causal relationships, the present classification structure highlights social life satisfaction and optimism as key psychological resources that differentiate loneliness outcomes among older adults with vision impairment. Together, these findings suggest that emotional resources may play a central role in shaping heterogeneous pathways to loneliness in later life. Vision loss severity as a contextual condition for loneliness Terminal node 2 indicates that individuals reporting a moderate level of social life satisfaction (=4) and experiencing legal blindness tended to fall into the low loneliness category. This finding is somewhat counter-intuitive, as increasing severity of visual impairment is associated with worse mental health outcomes (Suzuki et al., 2026). This may, however, be explained by higher levels of assistance older adults with legal blindness require, which may expose them to more—though perhaps not as satisfying—social contact through informal care and assistance (Marques et al., 2018). Prior research has rarely examined the joint relationship among social satisfaction, legal blindness, and loneliness. A longitudinal study in New Zealand found that older adults with vision impairment reported significantly higher loneliness compared with sighted older adults (Grow et al., 2015), but it did not assess the severity of vision impairment. In contrast, Chu & Chan (2022), using ANOVA, reported no significant association between vision impairment and loneliness among Taiwanese adults aged 45–70. These mixed findings suggest that the relationship between vision loss and loneliness may depend on additional contextual and psychological conditions. The present results extend those previous findings by identifying a hierarchical conditional pattern in which the contribution of moderate social life satisfaction to loneliness depends on the severity of vision impairment. Among individuals with vision impairment who were not classified as legally blind, we observed different conditional pathways (terminal nodes 3–6). Rather than suggesting that vision loss severity directly determines loneliness outcomes, these results highlight how contextual conditions and psychological resources interact in shaping loneliness risk. The decision-tree approach allows such hierarchical conditional relationships to emerge, illustrating that the role of vision loss severity may depend on accompanying social and emotional resources rather than acting as an isolated predictor. This finding is consistent with Broaden-and-Build Theory (Fredrickson, 2001), which proposes that positive emotions and positive social experiences broaden individuals’ thought-action repertoires and strengthen enduring psychological and social resources. Anxiety and emotional responses to vision loss in loneliness pathways The decision tree identified anxiety as a key differentiating condition in loneliness pathways among individuals with AMD who were not classified as legally blind and who reported moderate social life satisfaction. Within this subgroup, elevated anxiety was associated with classification into the high-loneliness group, whereas lower anxiety defined a distinct branch in which loneliness outcomes were further differentiated by emotional responses to vision loss. Anxiety symptoms are common in people diagnosed with vision-threatening conditions, regardless of actual vision impairment, since concern about having to live with vision loss in the future can be unsettling (Demmin & Silverstein, 2020). Our findings suggest that, for individuals with moderate social satisfaction, elevated anxiety poses an additional risk for loneliness. At the same time, the lower-anxiety branch revealed an even more complex conditional pattern. Among individuals with lower anxiety, those reporting more frequent negative emotional responses to vision loss (lower scores indicated more distress) were nevertheless classified into the low-loneliness group. Rather than indicating that distress (negative emotional responses to vision loss) itself is protective, this pattern should be interpreted within the broader conditional structure of the pathway: the classification occurred in the context of moderate social life satisfaction, absence of legal blindness, and lower anxiety. In this context, emotional responses to vision loss further differentiate individuals within a relatively lower-risk branch rather than being as an isolated predictor of less loneliness. Alternatively, those individuals with more negative emotional responses to their vision loss are more able to activate support mechanisms, either because their diagnosis is more recent and/or because they are able to verbalize their frustrations more effectively. This is consistent with the ARP framework (Sörensen et al., 2021), which suggests that engaging with initially stressful considerations regarding one’s aging-related needs can have positive effects in activating resources. Prior research suggests a link between loneliness and anxiety. For example, Steiner et al. (2022) found that older adults who became less lonely also experienced decreases in anxiety one year later. The present findings extend this literature by showing that anxiety does not operate uniformly across the sample but instead differentiates loneliness classification within a specific contextual and psychosocial subgroup. This hierarchical structure illustrates how decision-tree modeling can identify sequential conditional dependencies that may remain obscured in conventional linear or interaction-based models. This pattern further demonstrates that psychological factors operate differently depending on surrounding contextual and emotional conditions, rather than as independent risk factors. Depressive symptoms as a dominant classifier within the low-anxiety branch Finally, within the low-anxiety branch of the decision tree, depressive symptoms emerged as a dominant classifier distinguishing loneliness outcomes. Among individuals with moderate social life satisfaction who were not classified as legally blind and who had relatively low anxiety and not overly negative responses to vision loss, depressive symptoms further differentiated classification into the low- versus high-loneliness groups. This pattern suggests that even within a relatively lower-risk psychological context, depressive symptoms remain a key vulnerability factor shaping loneliness outcomes. This is consistent with previous research that indicates strong associations between depressive symptoms and loneliness in later life. Longitudinal studies have shown that these associations may remain significant over extended follow-up periods, including up to 12 years (e.g., Lee et al., 2021). The present findings align with this literature but further extend it by demonstrating that depressive symptoms emerged as an important differentiating condition within specific conditional contexts in addition to more general associations of depression and loneliness. In the current pathway, depressive symptoms differentiated loneliness outcomes only after accounting for social satisfaction, vision-loss status, anxiety levels, and emotional responses to vision loss. This suggests that mediators, moderators, and reciprocal relationships need to be investigated more thoroughly for therapies and programs to determine best points of intervention for older adults with vision loss. In addition, these findings suggest that interventions addressing depressive symptoms may be especially important for older adults with vision impairment who otherwise appear relatively protected by moderate social satisfaction and lower anxiety levels. Although future planning variables were included among the candidate predictors based on the ARP framework (Sörensen et al., 2021), these variables were not selected in the final decision-tree model. This surprising finding does not necessarily indicate that future planning is unrelated to loneliness outcomes. Instead, it may suggest that within the present sample, emotional resources and psychological states (e.g., social life satisfaction, anxiety, and depressive symptoms) served as more immediate differentiating conditions in the classification pathways. Proactive planning processes may operate indirectly through these psychological resources or may become more salient under different modeling conditions (Chang & Sörensen, 2024), with longitudinal data, or in different contextual circumstances. Future research could further examine the potential role of proactive planning in shaping loneliness trajectories, particularly in studies designed to explore alternative model specifications or longitudinal dynamics. Lastly, these findings do not contradict the ARP framework. Rather, they suggest that the influence of future planning may be contingent upon broader psychosocial strengths, such as optimism, which creates conditions under which preparation-related processes are more likely to contribute to successful adaptation (Sörensen et al., 2014). Limitations Although classification tree models provide interpretable conditional pathways, they do not establish causal relationships. For example, the pathway in which participants with low social life satisfaction but high optimism were predominantly classified into the low loneliness group should not be interpreted as evidence that optimism directly prevents loneliness. Instead, this finding reflects patterns of conditional co-occurrence that require further investigation. Moreover, the model was trained in a modest sample size, and the node-specific distributions (e.g., minority cases within terminal nodes) may be sensitive to sampling variation. The sample consisted predominantly of White older adults with macular degeneration, limiting generalizability to more diverse racial/ethnic groups, younger individuals, and those with other vision-threatening conditions. These limitations highlight the need for future research using larger and more diverse samples to validate the conditional structures identified in the present study. Although the findings suggest that depressive symptoms may exert a dominant influence on loneliness even in the presence of protective factors such as low anxiety, the directionality of these variables is not identified in our study. Future studies should examine how depression interacts with other psychological resources across time and whether it consistently overrides their buffering effects, or whether increases in loneliness may instead contribute to rising depressive symptoms. Similarly, the possible compensatory role of optimism, whereby high optimism appeared to buffer against loneliness even under conditions of low social life satisfaction, needs to be further tested across cultural contexts and age to assess its generalizability. Despite these limitations, the findings underscore the importance of examining both psychological and social resources together. By integrating longitudinal designs, diverse samples, and socioecological perspectives, future research can clarify the mechanisms through which protective and risk factors jointly shape loneliness outcomes. Such efforts will inform the classification of patients with vision loss for more targeted interventions to support well-being among older adults facing health and social challenges. Contributions and implications The methodological approach used in this study offers several advantages for advancing the study of complex psychosocial processes in aging. By applying a decision-tree machine learning approach, this study addresses an important methodological gap in gerontological research by capturing complex, nonlinear, and conditional pathways to loneliness among older adults with AMD. The classification-tree model identifies interactive combinations of psychosocial, emotional, and functional factors, offering a more nuanced understanding of how loneliness may develop in this population. By capturing how psychosocial resilience factors and emotional processes interact with sensory impairment, this study provides new insights into the psychological pathways associated with late-life emotional health that extend beyond single-domain explanations. The decision-tree approach also has important practical implications. Because decision trees produce hierarchical and interpretable conditional pathways, they provide a step-by-step framework that may be more readily translated into assessment and intervention planning than traditional regression-based approaches, which primarily estimate independent associations among variables. For example, decision trees might provide guidelines for a Sequential, Multiple Assignment, Randomized Trial (SMART), in which participants are randomized at critical decision points to different intervention options based on screening questions, early intervention response or individual subgroup characteristics (Lei et al., 2012). Rather than requiring practitioners to consider all risk factors simultaneously, the tree structure highlights sequential combinations of conditions that may be associated with elevated loneliness risk. It could also carry information on potential protective resources. These features may facilitate the identification of meaningful intervention targets while acknowledging the heterogeneous nature of loneliness experiences among older adults with AMD. From an applied perspective, the identified pathways may have relevance across multiple service settings that support older adults with AMD. Primary care providers, ophthalmology or community-based aging organizations may be well-positioned to incorporate brief psychosocial screening procedures informed by the hierarchical risk and preventive factors identified in the present study. For instance, such tools could be incorporated into the Patient-Reported Outcomes Measurement Information System measures used in routine medical care, which can facilitate easy access to Patient Health Questionnaire-2 scores, and other indicators of distress (Bass et al., 2021). Such approaches may facilitate earlier identification of individuals who may be at elevated risk of loneliness and support more targeted referral and intervention efforts. In addition, the interpretable structure of decision-tree models may inform the future development of accessible digital screening or self-assessment tools that help individuals recognize psychosocial risk factors and engage with supportive resources. However, because the present study did not directly evaluate implementation strategies or digital applications, these possibilities warrant consideration as promising directions for future translational research. To summarize, unlike many machine-learning approaches that operate as “black boxes,” decision-tree models produce interpretable pathways that may inform screening strategies and intervention targets. This enhances the practical applicability of the findings for clinicians and practitioners working with visually impaired older adults, allowing interventions to be tailored according to different conditional pathways rather than relying on a one-size-fits-all approach. Conclusions and future directions This study applied a classification tree approach to identify conditional pathways linking social life satisfaction, psychological characteristics, and vision impairment to loneliness in later life. The results demonstrated that combinations of social life satisfaction, vision loss severity, emotional response to vision loss, anxiety, optimism, and depression formed distinct pathways associated with higher or lower loneliness risk. These findings extend prior research by illustrating how multiple psychosocial resources and vulnerabilities may combine in nonlinear and conditional ways, offering a more nuanced understanding of loneliness risk than traditional linear regression models typically provide. Although replication in larger and more diverse samples, and with longitudinal data, is needed, the findings provide a foundation for developing more targeted screening strategies and individualized interventions to address loneliness among older adults living with vision impairment. Supplementary Material Contributor Information Ya-Han Chang, Department of Counseling and Human Development, Warner School of Education and Human Development, University of Rochester, Rochester, New York, United States. Silvia Sörensen, Department of Counseling and Human Development, Warner School of Education and Human Development, University of Rochester, Rochester, New York, United States. Supplementary material Supplementary material is available at Innovation in Aging online. Funding This research was supported by a National Institute on Aging research grant (R01 AG032032, S. Sörensen). Conflicts of interest None declared. Data availability This study was preregistered with ClinicalTrials.gov (NCT02224963) and approved by the Institutional Review Board of the University of Rochester (RSRB 27150). The protocol was also published before analyses were conducted (Sörensen et al., 2015). The de-identified data, analysis code, and supporting materials are available from the corresponding author upon reasonable request. Acknowledgments We presented the preliminary findings based on selected variables from this study at the 2025 Gerontological Society of America Meeting (Chang & Sörensen, 2025). References Associated Data Supplementary Materials Data Availability Statement This study was preregistered with ClinicalTrials.gov (NCT02224963) and approved by the Institutional Review Board of the University of Rochester (RSRB 27150). The protocol was also published before analyses were conducted (Sörensen et al., 2015). The de-identified data, analysis code, and supporting materials are available from the corresponding author upon reasonable request. Acknowledgments We presented the preliminary findings based on selected variables from this study at the 2025 Gerontological Society of America Meeting (Chang & Sörensen, 2025).