Review Reports - Curiosity in Later Life: Identifying Psychosocial Predictors Using Random Forest Modeling

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

The paper addresses an important topic. Curiosity is plausibly relevant for healthy aging, and the combination of age-group comparison plus a predictive modeling approach is promising. I have a few comments to improve the manuscript.

In the paper, you define curiosity as “joyous exploration” and use the Joyous Exploration subscale as the outcome. That is coherent. However, several top predictors look very close to curiosity in content (e.g., “new and interesting prospects”). I suggest adding a short subsection stating which scales and items were eligible predictors, and explicitly confirm that no curiosity items (including other 5D curiosity subscales, if present) were used as predictors. If any were included, please remove them and re-run the model, or justify clearly why inclusion is not circular.
I think that turning a continuous curiosity score into “high” vs “low” (upper quartile vs the rest) would sacrifice information and create class imbalance. I suggest using a Random Forest regressor predicting continuous curiosity (plus interpretable effect plots). If classification is retained, consider comparing top vs bottom quartile (more separation) or run sensitivity analyses across multiple thresholds (e.g., median split, tertiles, top 20 percent), and report stability of the “top predictors.”
There are contradictions between the abstract, Results, and Table 3. Also, the confusion matrix counts imply different accuracies than reported. Example: using the counts shown, “all variables” appears to classify 37/41 correctly (about 0.90), not 0.95. For the top 20 model, 35/41 is about 0.85, not 0.95. Please check metric calculations, ensure you are consistently reporting performance on the same split (test set vs training vs cross-validation).
In the current manuscript, some key details are missing for readers to assess generalizability. I suggest that in the paper, adding exact predictor set size (number of items used). How you handle missing data? You removed 11 cases, but please clarify item-level missing data handling. How does train-test split procedure with random seed? What’re the hyperparameters for the final RF (n_estimators, max_depth, max_features, min_samples_leaf, class_weight, etc.), and whether tuning was performed. I also suggest use cross-validationing. A single 80/20 split can be unstable with a small high-curiosity test sample. Please consider repeated stratified k-fold CV and report mean plus SD for AUC, balanced accuracy, and recall for the high-curiosity class.
With a minority “high curiosity” class, accuracy can be misleading. I suggest reporting balanced accuracy, macro-averaged F1, PR-AUC (precision-recall AUC).
In the paper, several formulas appear incorrect or confusing, and the paper does not truly need them at this level of detail. First, Gini impurity is typically defined as 1 minus the sum of squared class proportions, but the manuscript shows only the sum of squares. Second, the RF “formula 1” is not informative as written for most readers. I suggest removing most formulas and replacing with precise narrative descriptions and citations, or correct them with standardize notation.
Gini-based feature importance can be biased and is not directly interpretable as an effect size. Please adding permutation importance, and ideally SHAP values, plus partial dependence or accumulated local effects plots for the top predictors.
In Appendix Table A1, some correlations are negative for constructs that are described as “optimism,” which suggests the item is reverse-scaled (e.g., 1 = agree a lot to 5 = disagree a lot). Similar issues may apply to “energy” items. I suggest reverse-code so that higher scores always mean more of the construct, then re-run correlations and update interpretation. If you intentionally kept raw coding, state this clearly and interpret coefficients accordingly.
This is a cross-sectional study; however, the language in the current manuscript sometimes implies sustaining or preventing (e.g., “help sustain,” “reducing loneliness,” “preventing loneliness”) which can read causally. Please revise to associational language throughout.
Table 2: the younger group is labeled “>60” but I think it should be “<60”?
Figure 1 label says “Anova T-Test,” but the analysis is an independent-samples t-test.
Methods: “matlib” should likely be “matplotlib.”
Random Forest, random forest, Random-Forest. Please standardize to make them consistent.
For the abstract, please ensure the performance numbers match the final Results table and that the listed predictors match the final model output exactly.
In the Method, I suggest providing the exact curiosity scale scoring range and how the “cumulative score” was computed. Also, please provide sample characteristics for the ≥60 subgroup, not only the full sample. Please also clarify whether predictors were item-level or scale-level. If item-level, justify and discuss implications for interpretability and multiple comparisons.
When reporting the t-test, consider adding the mean difference with CI in the text for readability. Table A1 title says “R2 values” but the column shows correlations. Please fix the label and define the asterisks.
The Discussion is generally coherent, but it should acknowledge the modeling limitations (single split, small minority-class test size, importance bias) and treat predictor rankings as provisional until validated. I also suggest considering adding a short “Implications” subsection that translates top predictors into testable intervention.

Author Response

Comment 1: In the paper, you define curiosity as “joyous exploration” and use the Joyous Exploration subscale as the outcome. That is coherent. However, several top predictors look very close to curiosity in content (e.g., “new and interesting prospects”). I suggest adding a short subsection stating which scales and items were eligible predictors, and explicitly confirm that no curiosity items (including other 5D curiosity subscales, if present) were used as predictors. If any were included, please remove them and re-run the model, or justify clearly why inclusion is not circular.

Response:

Thank you for this helpful comment. In the revised manuscript, we clarify the eligibility criteria for predictor inclusion, an explicitly confirm that no curiosity items were used as predictors. Eligible predictors consisted of ordinal Likert-type items assessing personality, psychological, behavioral, and temporal constructs related to perceptions of and preparations for aging. Curiosity was operationalized exclusively using the cumulative score of the Joyous Exploration subscale as the outcome variable (Section 2.2.4). No curiosity items or subscales, including other dimensions of the Five-Dimensional Curiosity Scale or the individual Joyous Exploration items, were included as predictors in any analysis. To further reduce conceptual overlap, all items related to openness to learning from the personality inventory were removed prior to modeling. These clarifications are provided in Section 2.2 and Section 2.2.1 (lines 102–115) of the Measures.

Comment 2: I think that turning a continuous curiosity score into “high” vs “low” (upper quartile vs the rest) would sacrifice information and create class imbalance. I suggest using a Random Forest regressor predicting continuous curiosity (plus interpretable effect plots). If classification is retained, consider comparing top vs bottom quartile (more separation) or run sensitivity analyses across multiple thresholds (e.g., median split, tertiles, top 20 percent), and report stability of the “top predictors.”

Response:

Thank you for highlighting this methodological consideration and for the constructive suggestions. In alignment with these recommendations, we expanded the analyses in two ways:

First, we conducted a Random Forest regression predicting continuous curiosity across the full sample. This analysis was included as a benchmark to avoid information loss associated with dichotomization and to identify stable predictors of curiosity. Using repeated cross-validation, the regression model demonstrated good performance (mean R² = .64, SD = .09; RMSE = 0.95). Feature importance analyses indicated that future time perspective items were the strongest predictors of continuous curiosity, followed by engagement in leisure and meaning-oriented activities. Feature importance plot is reported in Figure 2, and PDP plots illustrating the marginal effects of the most influential predictors is presented in the Supplementary Materials Figure A1.

Second, to further address concerns regarding class imbalance and threshold dependence in the ≥60 cohort, we retained the classification framework and conducted sensitivity analyses using alternative extreme-group definitions (top vs. bottom 35%, 40%, and median - 50%). As class definitions were broadened, classification performance metrics (accuracy, precision, and recall) declined modestly but remained stable overall. Across all sensitivity analyses, several predictors consistently emerged as influential, including items related to seeking new prospects, meaning in life, leisure engagement, and generativity.

Comment 3: There are contradictions between the abstract, Results, and Table 3. Also, the confusion matrix counts imply different accuracies than reported. Example: using the counts shown, “all variables” appears to classify 37/41 correctly (about 0.90), not 0.95. For the top 20 model, 35/41 is about 0.85, not 0.95. Please check metric calculations, ensure you are consistently reporting performance on the same split (test set vs training vs cross-validation).

Response:

Thank you for this careful observation. All classification performance metrics reported in the Results and tables are now derived consistently from held-out test data only, using a single stratified 80/20 train-test split for point estimates. We have now corrected these inconsistencies and clarified the reporting of model performance throughout the manuscript.

Comment 4: In the current manuscript, some key details are missing for readers to assess generalizability. I suggest that in the paper, adding exact predictor set size (number of items used). How you handle missing data? You removed 11 cases, but please clarify item-level missing data handling. How does train-test split procedure with random seed? What’re the hyperparameters for the final RF (n_estimators, max_depth, max_features, min_samples_leaf, class_weight, etc.), and whether tuning was performed. I also suggest use cross-validationing. A single 80/20 split can be unstable with a small high-curiosity test sample. Please consider repeated stratified k-fold CV and report mean plus SD for AUC, balanced accuracy, and recall for the high-curiosity class.

Response:

Thank you for this thorough and helpful comment. We have revised the Methods section to provide the requested details. Specifically, we now report the exact size of the predictor set, which consisted of item-level variables derived from all the scales mentioned, after exclusion of curiosity-related items. Missing data handling is now clarified: participants with any missing item-level data on predictors or the curiosity outcome were excluded prior to analysis, resulting in the removal of 11 cases; the final analytic sample contained complete data, and no imputation was required.

We also explicitly describe the train-test split procedure, including the use of stratified 80/20 splits for classification analyses and a fixed random seed (random_state = 42) to ensure reproducibility. All Random Forest hyperparameters are now fully reported (n_estimators = 500; class_weight = “balanced” for classification; remaining parameters left at scikit-learn defaults, including max_depth, max_features, min_samples_split, and min_samples_leaf), and we clarify that no additional hyperparameter tuning was performed.

To address concerns regarding instability from a single split and a small high-curiosity class, we additionally implemented repeated stratified k-fold cross-validation (5 folds × 10 repetitions). Mean and standard deviation values are now reported for ROC-AUC, balanced accuracy, and recall for the high-curiosity class. These revisions are detailed in Sections 2.3 and 2.4 of the Methods.

Comment 5: With a minority “high curiosity” class, accuracy can be misleading. I suggest reporting balanced accuracy, macro-averaged F1, PR-AUC (precision-recall AUC).

Response:

We have expanded the reported evaluation metrics to include balanced accuracy, macro-averaged F1, and precision-recall AUC (PR-AUC), in addition to ROC-AUC. These metrics are now emphasized in both the Results and cross-validation summaries in the supplementary section.

Comment 6: In the paper, several formulas appear incorrect or confusing, and the paper does not truly need them at this level of detail. First, Gini impurity is typically defined as 1 minus the sum of squared class proportions, but the manuscript shows only the sum of squares. Second, the RF “formula 1” is not informative as written for most readers. I suggest removing most formulas and replacing with precise narrative descriptions and citations, or correct them with standardize notation.

Response:

Thank you for noting this. In the revised Methods, we have removed the non-informative Random Forest “Formula 1” and the Gini impurity expression, and replaced these with concise narrative descriptions of the Random Forest procedure (bootstrap aggregation and random feature subsampling).

Comment 7: Gini-based feature importance can be biased and is not directly interpretable as an effect size. Please adding permutation importance, and ideally SHAP values, plus partial dependence or accumulated local effects plots for the top predictors.

Response:

Thank you for this important methodological suggestion. In response, we have extended the analysis to include permutation importance (Figure A1), quantified as the mean decrease in model performance (ROC-AUC) under random permutation of each predictor. Partial dependence plots for the top predictors have also been added to illustrate their marginal associations with curiosity (Figure 7). Gini-based impurity importance is now used only for preliminary screening and is not interpreted as an effect size. These additions are described in the revised Methods and reported in the Results and Supplementary Materials.

Comment 8: In Appendix Table A1, some correlations are negative for constructs that are described as “optimism,” which suggests the item is reverse-scaled (e.g., 1 = agree a lot to 5 = disagree a lot). Similar issues may apply to “energy” items. I suggest reverse-code so that higher scores always mean more of the construct, then re-run correlations and update interpretation. If you intentionally kept raw coding, state this clearly and interpret coefficients accordingly.

Response:

Thank you for this helpful suggestion. In light of the expanded analyses conducted during revision, Appendix Table A1 has been removed. Detailed tables reporting feature importance and correlation coefficients are now provided for each classification specification (see Tables A3, A5, A7, and A9).

In these revised tables, all reverse-worded items (e.g., optimism and energy-related measures) were explicitly reverse-coded prior to analysis so that higher values consistently reflect greater levels of the underlying construct.

Comment 9: This is a cross-sectional study; however, the language in the current manuscript sometimes implies sustaining or preventing (e.g., “help sustain,” “reducing loneliness,” “preventing loneliness”) which can read causally. Please revise to associational language throughout.

Response:

Thank you for this important clarification. In response, we have carefully revised the manuscript to replace causal or intervention-oriented language (e.g., “sustain,” “prevent,” “reduce,” “protect”) with strictly associational phrasing (e.g., “associated with,” “linked to,” “co-exist”).

Comment 10: Table 2: the younger group is labeled “>60” but I think it should be “<60”?

Response:

Amended. Thank you for the correction.

Comment 11: Figure 1 label says “Anova T-Test,” but the analysis is an independent-samples t-test.

Response:

Amended. Thank you for the correction.

Comment 12: Methods: “matlib” should likely be “matplotlib.”

Response:

Amended. Thank you for the correction.

Comment 13: Random Forest, random forest, Random-Forest. Please standardize to make them consistent.

Response:

All terms have been standardized to Random-Forest and RF as acronym throughout the remainder of the script.

Comment 14: For the abstract, please ensure the performance numbers match the final Results table and that the listed predictors match the final model output exactly.

Response:

Thank you for this suggestion. We have revised the abstract to ensure full consistency with the finalized Results tables.

Comment 15: In the Method, I suggest providing the exact curiosity scale scoring range and how the “cumulative score” was computed. Also, please provide sample characteristics for the ≥60 subgroup, not only the full sample. Please also clarify whether predictors were item-level or scale-level. If item-level, justify and discuss implications for interpretability and multiple comparisons.

Response:

Thank you for these helpful suggestions.

We have revised the Methods section to clarify the scoring of the Joyous Exploration (JE) subscale, including its item composition, 7-point response scale, and the computation of the cumulative score (Section 2.2.4).
Descriptive characteristics for the ≥60 subgroup have been added to Table 1 to complement the full-sample statistics.
We further clarify that all predictors entered into the Random Forest models were measured at the item level (Section 2.2).

Comment 16: When reporting the t-test, consider adding the mean difference with CI in the text for readability. Table A1 title says “R2 values” but the column shows correlations. Please fix the label and define the asterisks.

Response:

Thank you for this helpful suggestion. The reporting of the independent-samples t-test has been revised to include the mean difference and its 95% confidence interval for improved readability. Specifically, Section 3.2 now reports a mean difference of 0.78 (95% CI [0.49, 1.06]), alongside the test statistic and effect size.

In addition, Appendix Table A1 has been removed to reflect the expanded analyses conducted during revision. Correlation results are now reported in updated appendix tables corresponding to each classification specification (Tables A3, A5, A7, and A9). In these tables, the column label has been corrected from “R² values” to “Correlation (r)”, and asterisks denoting statistical significance are explicitly defined in the table notes.

Comment 17: The Discussion is generally coherent, but it should acknowledge the modeling limitations (single split, small minority-class test size, importance bias) and treat predictor rankings as provisional until validated. I also suggest considering adding a short “Implications” subsection that translates top predictors into testable intervention.

Response:

We have revised the Discussion to more explicitly acknowledge modeling limitations, including reliance on a single train-test split for point estimates, the relatively small minority-class test sample, and the potential bias of impurity-based feature importance. Predictor rankings are now explicitly described as provisional and requiring replication in independent samples (lines 448-454). In addition, we have added a brief “Implications” subsection translating the most consistent predictors into concrete, testable directions for future community-based and intervention research (lines 427-436).

Reviewer 2 Report

Comments and Suggestions for Authors

This manuscript examines age-related differences in trait curiosity and applies a Random Forest model to identify psychosocial predictors of high curiosity among older adults in a Taiwanese cohort. The integration of machine learning to explore multidimensional predictors is potentially valuable; however, several conceptual, methodological, and reporting issues need to be addressed before the contribution can be fully supported.

Major Comments

The authors equates “curiosity” with the Joyous Exploration subscale of the Five-Dimensional Curiosity Scale. This is concerning because the conceptual implications of equating this specific subcomponent with curiosity as a broader construct are not sufficiently discussed. Joyous Exploration reflects a positively valenced, trait-like orientation toward novelty seeking and self-expansion, but does not capture other theoretically distinct dimensions of curiosity, for example, deprivation sensitivity/thrill seeking, which may follow different developmental trajectories in later life.

As a result, statements throughout the manuscript risk overgeneralizing the findings to “curiosity in later life,” when the empirical results more narrowly pertain to age-related differences and predictors of joyous exploration. Given the multi-dimensional nature of curiosity, it is plausible that some dimensions may decline with age while others remain stable or even increase in older adulthood.

To strengthen conceptual clarity and alignment between measurement and interpretation, the authors are encouraged to more explicitly delimit their conclusions to the joyous exploration dimension of trait curiosity. In addition, the discussion would benefit from a clearer differentiation between the dimension examined in the present study and other forms of curiosity that were not assessed, along with an acknowledgment that the reported age-related patterns may not generalize to curiosity as a whole.

The reviewer is concerned about the approach to dichotomize age, since age is fundamentally a continuous variable. Dichotomization of age can lead to loss of information, reduced statistical power, and an increased risk of masking gradual age-related patterns. In the context of lifespan research, collapsing age into broad categories may obscure meaningful within-group variability and oversimplify developmental processes that are likely to unfold continuously rather than abruptly at a single cutoff.

The manuscript would benefit from a more explicit discussion of these limitations and from additional analyses treating age as a continuous predictor. Sensitivity analyses examining whether the observed age-related effects hold when age is modeled continuously, or when alternative cutoffs are considered, would strengthen the robustness and interpretability of the findings.

The same comment also applies to dichotomize curiosity into high versus low groups using the 75th percentile.

Many of the top predictors reflect closely related constructs (e.g., multiple items assessing meaning, purpose, and optimism). Since Random Forest feature importance tens to be inflated when predictors are correlated, the authors should discuss this limitation and clarify that importance rankings do not imply independent or causal effects.

At several points, the statements implies that certain behaviors or psychosocial factors “sustain” or “support” curiosity. For example, in the abstract, the authors stated "…older adults who maintain high curiosity do so through emotionally meaningful, socially connected activities…". The phrasing “maintain high curiosity” implicitly assumes that curiosity would otherwise decline and that these activities play a role in preserving it. In addition, the expression “do so through” is a strongly causal formulation, suggesting that the identified behaviors actively contribute to sustaining curiosity rather than merely being associated with higher curiosity levels. Similar statements listed below:

The paragraph in discussion section explaining predictors: “…highlight the importance of adaptive behavioral engagement in sustaining curiosity during older adulthood.”
The paragraph explaining meaning/purpose/future orientation:

“…factors may help sustain curiosity even in later life.”

“…providing psychological resources that sustain curiosity…”

The paragraph concluding the finding: "…offers valuable insights for policymakers and geriatric institutions seeking to cultivate curiosity and its drivers…"

Given the cross-sectional design, these statements should be softened to reflect ASSOCIATION rather than directionality, or the limitations of causal inference should be emphasized more strongly.

Minor Comments

The authors may remove the random forest algorithm details or move them to the supplementary material. Given that Random Forest is a well-established and widely used machine learning technique, much of this methodological detail appears unnecessary for the main text and interrupts the overall flow of the Methods section.

Figure 3 (Confusion Matrix) shows misaligned grid lines. It is unclear to the reviewer whether these issues originate from the authors’ submitted figure or were introduced during the journal’s preparation of the review version. In either case, the figure would benefit from careful formatting checks to ensure proper alignment and visual clarity.

The Discussion section could be tightened by reducing repetition of literature already covered extensively in the Introduction.

This study addresses an important question about curiosity in later life and offers an interesting application of machine learning to psychosocial data. However, concerns regarding outcome operationalization, dichotomization strategies, and interpretation of model performance warrant substantial revision before the conclusions can be fully supported.

Comments on the Quality of English Language

At several points, the statements implies that certain behaviors or psychosocial factors “sustain” or “support” curiosity. For example, in the abstract, the authors stated "…older adults who maintain high curiosity do so through emotionally meaningful, socially connected activities…". The phrasing “maintain high curiosity” implicitly assumes that curiosity would otherwise decline and that these activities play a role in preserving it. In addition, the expression “do so through” is a strongly causal formulation, suggesting that the identified behaviors actively contribute to sustaining curiosity rather than merely being associated with higher curiosity levels. Similar statements listed below:

- The paragraph in discussion section explaining predictors: “…highlight the importance of adaptive behavioral engagement in sustaining curiosity during older adulthood.”
- The paragraph explaining meaning/purpose/future orientation:

“…factors may help sustain curiosity even in later life.”

“…providing psychological resources that sustain curiosity…”

The paragraph concluding the finding: "…offers valuable insights for policymakers and geriatric institutions seeking to cultivate curiosity and its drivers…"

Given the cross-sectional design, these statements should be softened to reflect ASSOCIATION rather than directionality, or the limitations of causal inference should be emphasized more strongly.

Author Response

Comment 1:

Major Comments

The authors equates “curiosity” with the Joyous Exploration subscale of the Five-Dimensional Curiosity Scale. This is concerning because the conceptual implications of equating this specific subcomponent with curiosity as a broader construct are not sufficiently discussed. Joyous Exploration reflects a positively valenced, trait-like orientation toward novelty seeking and self-expansion, but does not capture other theoretically distinct dimensions of curiosity, for example, deprivation sensitivity/thrill seeking, which may follow different developmental trajectories in later life.

Response:

Thank you for this important conceptual clarification.

We have revised the manuscript to explicitly delimit all interpretations to the Joyous Exploration (JE) dimension of trait curiosity and avoid generalizing findings to curiosity as a broader construct.
Throughout the manuscript, references to “curiosity” have been clarified as referring specifically to JE, and the Discussion now more clearly differentiates JE from other dimensions of curiosity (e.g., deprivation sensitivity, thrill seeking, social curiosity) that were not assessed in the present study.
Finally, we further acknowledge in the limitations (lines 442-447; 455-460) that different curiosity dimensions may follow distinct developmental trajectories in later life and that the observed age-related patterns may not generalize to curiosity as a whole.

Comment 2: The reviewer is concerned about the approach to dichotomize age, since age is fundamentally a continuous variable. Dichotomization of age can lead to loss of information, reduced statistical power, and an increased risk of masking gradual age-related patterns. In the context of lifespan research, collapsing age into broad categories may obscure meaningful within-group variability and oversimplify developmental processes that are likely to unfold continuously rather than abruptly at a single cutoff.

The same comment also applies to dichotomize curiosity into high versus low groups using the 75th percentile.

Response:

We thank the reviewer for highlighting this important methodological consideration. We now explicitly acknowledge in the Methods (lines 170-180) and Limitations (Discussion section lines 445- 458) sections that dichotomizing age and curiosity (JE) entails information loss and may obscure continuous developmental patterns. Importantly, to address this concern, age was additionally modeled as a continuous predictor using linear regression (section 3.1) across the full sample, yielding results consistent with the categorical comparison. Likewise, JE was analyzed as a continuous outcome using Random Forest regression (section 3.5) prior to any dichotomization, and sensitivity analyses using alternative JE cutoffs (35/35, 40/40, and median splits) demonstrated robust patterns of predictors.

Comment 3: Many of the top predictors reflect closely related constructs (e.g., multiple items assessing meaning, purpose, and optimism). Since Random Forest feature importance tens to be inflated when predictors are correlated, the authors should discuss this limitation and clarify that importance rankings do not imply independent or causal effects.

Response:

We thank the reviewer for this important point. We have revised the Limitations section to explicitly note that several top predictors reflect closely related psychosocial constructs and that Random Forest impurity-based feature importance may be influenced by correlated predictors (lines 448-454). We now clarify that importance rankings should be interpreted as reflecting shared predictive contributions rather than independent or causal effects of individual variables. This limitation is discussed alongside the use of partial dependence plots to aid interpretability (Figure 7; Appendix A1).

Comment 4: At several points, the statements implies that certain behaviors or psychosocial factors “sustain” or “support” curiosity. For example, in the abstract, the authors stated "…older adults who maintain high curiosity do so through emotionally meaningful, socially connected activities…". The phrasing “maintain high curiosity” implicitly assumes that curiosity would otherwise decline and that these activities play a role in preserving it. In addition, the expression “do so through” is a strongly causal formulation, suggesting that the identified behaviors actively contribute to sustaining curiosity rather than merely being associated with higher curiosity levels. Similar statements listed below:

- The paragraph in discussion section explaining predictors: “…highlight the importance of adaptive behavioral engagement in sustaining curiosity during older adulthood.”
- The paragraph explaining meaning/purpose/future orientation:

“…factors may help sustain curiosity even in later life.”

“…providing psychological resources that sustain curiosity…”

The paragraph concluding the finding: "…offers valuable insights for policymakers and geriatric institutions seeking to cultivate curiosity and its drivers…"

Given the cross-sectional design, these statements should be softened to reflect ASSOCIATION rather than directionality, or the limitations of causal inference should be emphasized more strongly.

Response:

Thank you for highlighting this important issue regarding causal language. We have systematically revised the Abstract, Discussion, and concluding paragraphs to replace causal or directional phrasing (e.g., “sustain,” “support,” “do so through,” “cultivate”) with language that reflects association and co-existence. In addition, we have strengthened the Limitations section to explicitly emphasize the cross-sectional nature of the data and the inability to infer directionality or causal effects (lines 455-459).

Minor Comments

Comment 1: The authors may remove the random forest algorithm details or move them to the supplementary material. Given that Random Forest is a well-established and widely used machine learning technique, much of this methodological detail appears unnecessary for the main text and interrupts the overall flow of the Methods section.

Response:

Thank you for this suggestion. To improve clarity and flow in the Methods section, we have removed detailed mathematical formulations of the Random Forest algorithm and replaced them with concise narrative descriptions of the procedure, focusing on key elements such as bootstrap aggregation and random feature subsampling.

Comment 2: Figure 3 (Confusion Matrix) shows misaligned grid lines. It is unclear to the reviewer whether these issues originate from the authors’ submitted figure or were introduced during the journal’s preparation of the review version. In either case, the figure would benefit from careful formatting checks to ensure proper alignment and visual clarity.

Response:

Thank you for noting this. We carefully re-checked the original source file for the confusion matrix (now figure 4) and confirmed that the grid lines and cell boundaries are properly aligned. It is possible that the misalignment observed in the review version was introduced during the manuscript conversion or proofing process. To ensure maximum visual clarity, we have re-exported Figure 4 removing the grid lines.

Comment 3: The Discussion section could be tightened by reducing repetition of literature already covered extensively in the Introduction.

Response:

We have revised the Discussion to reduce repetition of background literature already covered in the Introduction and to focus more explicitly on interpretation, integration of findings, and implications.

Comment 4: This study addresses an important question about curiosity in later life and offers an interesting application of machine learning to psychosocial data. However, concerns regarding outcome operationalization, dichotomization strategies, and interpretation of model performance warrant substantial revision before the conclusions can be fully supported.

Response:

Thank you for this and constructive assessment. We agree that greater clarity regarding outcome operationalization, dichotomization strategies, and interpretation of model performance is essential. In response, we have made several substantial revisions to strengthen the conceptual precision and methodological robustness of the study:

First, regarding outcome operationalization, we now explicitly delimit our findings to the joyous exploration subcomponent of curiosity. This refinement allows for a more targeted and theoretically coherent interpretation of curiosity in later life, thus reducing heterogeneity inherent in broader or composite curiosity measures.
Second, concerning dichotomization strategies, we have expanded the analytic framework to include a Random Forest regression treating curiosity as a continuous outcome across the full age range (see section 3.3). This benchmark analysis mitigates potential information loss associated with dichotomization and provides a complementary, distribution-sensitive perspective on how key psychosocial predictors relate to curiosity across the lifespan. Most notably, the highest-ranking predictors identified in the regression largely overlapped with those emerging from the classification analyses, supporting the robustness of the feature set across outcome specifications.
Third, to temper interpretation of model performance, we have revised the manuscript to emphasize predictive stability rather than point estimates of accuracy. Specifically, we report cross-validated performance metrics and conduct sensitivity analyses across multiple dichotomization thresholds, including extreme splits (top/bottom 35% and 40%) as well as a median (50/50) split (See sections 3.4 and 3.5). These analyses demonstrate that the relative importance and directionality of the strongest predictors remain consistent across classification specifications, while we explicitly avoid causal or deterministic interpretations of model performance.

Reviewer 3 Report

Comments and Suggestions for Authors

Curiosity! This manuscript addresses an important and timely topic, curiosity later in life. It is an interesting study that combines traditional statistical comparison and machine learning. The paper is well written and clearly structured. However the following methodological and reporting issues need to be addressed.

The manuscript reports a very high model performance and an accuracy around .98 and UAC around .85. However, different sections of the paper report different accuracy values. It is not clear which metrics are from training data and test data. Since it is a relatively small sample size, such high accuracy raises concerns about possible overfitting of data leakage. Please make sure to explain how was the data split into training data. Which performance metrics are based on test data only and whether feature selection was done before or after the split.
Curiosity is divided into 'high and 'low groups using the 75th percentile. This creates a very small 'high curiosity' group and therefore may reduce statistical power and model stability. Consider modeling curiosity as a continuous variable or providing robustness checks showing that results are stable throughout different points.
Several top predictors, such as 'meaning of life', and 'optimism' appear conceptually very close to 'curiosity' itself. This raises the possibility that the model is predicting curiosity using the same variables that partially measure the same construct. Please clarify how these constructs differ from curiosity.
There are small technical and topographical issues in the methods section.
Please add more details on the sampling and recruitment so to realize generalizability.
An age comparison could be strengthened by modeling age as a continuous variable rather than using a single cut-off at 60 years old.

Author Response

Comment 1: The manuscript reports a very high model performance and an accuracy around .98 and UAC around .85. However, different sections of the paper report different accuracy values. It is not clear which metrics are from training data and test data. Since it is a relatively small sample size, such high accuracy raises concerns about possible overfitting of data leakage. Please make sure to explain how was the data split into training data. Which performance metrics are based on test data only and whether feature selection was done before or after the split.

Response 1:

Thank you for raising this important concern. We have revised the manuscript to clarify that all performance metrics reported in the Results and tables are based exclusively on held-out test data, obtained using a stratified 80/20 train/test split. Training-set performance is no longer reported. To address potential overfitting and instability due to the relatively small sample size, we additionally report repeated stratified 5-fold cross-validation (10 repetitions) results as mean ± SD for ROC-AUC, balanced accuracy, and recall. Feature importance and feature selection were conducted after the train-test split, using training data only, and models were then evaluated on the independent test set. These revisions are detailed in Sections 2.3 and 2.4 of the Methods, and all reported performance values have been harmonized across the manuscript.

Comment 2: Curiosity is divided into 'high and 'low groups using the 75th percentile. This creates a very small 'high curiosity' group and therefore may reduce statistical power and model stability. Consider modeling curiosity as a continuous variable or providing robustness checks showing that results are stable throughout different points.

Response 2:

Thank you for this helpful comment. To address concerns regarding reduced statistical power and model stability arising from dichotomizing curiosity at the 75th percentile, we conducted complementary analyses modeling curiosity as a continuous variable using Random Forest regression across the full sample. In addition, we performed robustness checks using alternative extreme-group definitions (top vs. bottom 35%, 40%, and median 50%) within the ≥60 cohort. Across these specifications, overall model performance and the ranking of key predictors remained largely stable, indicating that the findings are not driven by a single cutoff choice. These results are reported in Sections 3.3-3.5, with additional details provided in the Supplementary Materials.

Comment 3: Several top predictors, such as 'meaning of life', and 'optimism' appear conceptually very close to 'curiosity' itself. This raises the possibility that the model is predicting curiosity using the same variables that partially measure the same construct. Please clarify how these constructs differ from curiosity.

Response 3:

Thank you for raising this important point regarding potential conceptual overlap. In our manuscript, curiosity was operationalized exclusively using the Joyous Exploration subscale, which captures an intrinsic, approach-oriented motivation to seek new information and experiences. In contrast, constructs such as meaning in life and optimism assess evaluative beliefs about life coherence, purpose, and expected future outcomes, rather than exploratory motivation or interest in novelty. Although these constructs are theoretically related and may reasonably co-occur, they are conceptually and psychometrically distinct and do not constitute alternative measures of curiosity.

Comment 4: There are small technical and topographical issues in the methods section.

Response 4:

Thank you for noting this. We have carefully proofread the Methods section and corrected minor technical and typographical issues, including inconsistencies in terminology, notation, and formatting, to improve clarity and readability.

Comment 5: Please add more details on the sampling and recruitment so to realize generalizability.

Response 5:

Thank you for this suggestion. We have expanded the Participants section to provide additional detail on the sampling and recruitment procedures for the Taiwan cohort, including the use of an online survey format and recruitment through online channels as part of the Ageing as Future project.

Comment 6: An age comparison could be strengthened by modeling age as a continuous variable rather than using a single cut-off at 60 years old.

Response 6:

Thank you for this helpful suggestion. In response, we expanded the analyses to model age as a continuous predictor of curiosity using linear regression across the full sample. These results are now reported in the Results (section 3.1) and indicate a gradual age-related pattern consistent with, rather than driven by, the categorical comparison. The age-group analysis (<60 vs. ≥60) was retained for comparability with prior lifespan research and to support interpretation focused on later-life processes.

Reviewer 4 Report

Comments and Suggestions for Authors

- While the Random Forest (RF) model is effective, the paper does not explain why RF was selected over other classifiers or regression models. No comparison is made with simpler or more interpretable methods (e.g., logistic regression, decision trees), which could question the necessity of RF in a relatively small dataset (N = 191 older adults). Justify the model choice and, if possible, provide a comparison with at least one traditional statistical approach.

- The method used to label "high curiosity" in older adults is unclear. Was a median split used? A clinical cutoff? The validity of the machine learning model hinges on how the binary outcome was defined. Clearly describe the operationalization of “high curiosity” and discuss the implications of dichotomizing a continuous variable.

- The reported performance metrics (Accuracy = 0.95; AUC = 0.85) suggest potential overfitting, especially given the relatively small sample and large number of predictors. No mention is made of cross-validation strategy or external validation. Include k-fold cross-validation or bootstrapping to validate model stability and report confidence intervals for AUC.

- Although RF can handle correlated features, the paper does not examine or discuss how closely related predictors (“leisure-time hobbies” vs. “interests outside work”) may influence each other’s importance or be conceptually redundant. Provide variable clustering or partial dependence plots to enhance interpretability and reduce feature overlap.

- While the paper identifies psychosocial predictors, it stops short of translating these into actionable insights for interventions. How should practitioners use this model to design or prioritize programs that foster curiosity in older adults? Expand the Discussion with a concrete proposal for community-based or clinical applications based on top predictors.

Author Response

Comment 1: While the Random Forest (RF) model is effective, the paper does not explain why RF was selected over other classifiers or regression models. No comparison is made with simpler or more interpretable methods (e.g., logistic regression, decision trees), which could question the necessity of RF in a relatively small dataset (N = 191 older adults). Justify the model choice and, if possible, provide a comparison with at least one traditional statistical approach.

Response 1:

Thank you for this constructive suggestion. To address this concern, we have expanded the Methods section to clarify the rationale for selecting Random Forest (RF) and have added a formal model comparison with simpler and more interpretable approaches. Specifically, RF performance was benchmarked against logistic regression and a single decision tree classifier using the primary ≥60, 75/25 JE split and repeated stratified 5-fold cross-validation with 10 repetitions. As shown in Table A1, RF demonstrated the strongest overall discriminative performance, particularly in terms of ROC-AUC and PR-AUC, while logistic regression provided a competitive and transparent baseline.

Comment 2: The method used to label "high curiosity" in older adults is unclear. Was a median split used? A clinical cutoff? The validity of the machine learning model hinges on how the binary outcome was defined. Clearly describe the operationalization of “high curiosity” and discuss the implications of dichotomizing a continuous variable.

Response 2:

Thank you for this important clarification.

We have now explicitly described the operationalization of “high curiosity” in Section 3.4 of the Results. In the ≥60 age group, high curiosity was defined using an upper-quartile split, with the 75th percentile (JE = 5.20) serving as the primary cutoff to distinguish high curiosity (top 25%) from lower JE (remaining 75%). This definition was selected to focus on individuals with particularly elevated levels of curiosity rather than relying on an arbitrary clinical threshold.
To assess the robustness of this operationalization, additional sensitivity analyses were conducted using alternative split definitions, including exact bottom-top splits at 35% and 40%, as well as a median split (curiosity = 4.0), which divided the sample into approximately equal high and low curiosity groups. Results were consistent across these alternative definitions, supporting the stability of the findings. A box-and-whisker plot illustrating the interquartile range and the primary cutoff is presented in Figure 3.

We now explicitly discuss the implications and limitations of dichotomizing a continuous construct in the Limitations section, noting the potential loss of information while emphasizing that sensitivity analyses were conducted to mitigate this concern (442-454).

Comment 3: The reported performance metrics (Accuracy = 0.95; AUC = 0.85) suggest potential overfitting, especially given the relatively small sample and large number of predictors. No mention is made of cross-validation strategy or external validation. Include k-fold cross-validation or bootstrapping to validate model stability and report confidence intervals for AUC.

Response 3:

Thank you for raising this important concern.

External validation was not feasible due to data availability constraints. However, to address the potential risk of overfitting, we have now explicitly incorporated and reported a repeated stratified k-fold cross-validation strategy. Specifically, model performance was evaluated using repeated stratified 5-fold cross-validation with 10 repetitions, which allowed us to assess model stability across multiple train-test partitions while preserving class balance.
We have updated the Methods section to clearly describe this cross-validation procedure and now report mean and standard deviation values for key performance metrics, including ROC-AUC, balanced accuracy, and recall for the high-JE class. These cross-validated results are presented in Appendix A and demonstrate that model performance remains stable across folds, mitigating concerns of overfitting arising from single-split estimates.

Comment 4: Although RF can handle correlated features, the paper does not examine or discuss how closely related predictors (“leisure-time hobbies” vs. “interests outside work”) may influence each other’s importance or be conceptually redundant. Provide variable clustering or partial dependence plots to enhance interpretability and reduce feature overlap.

Response 4:

Thank you for this insightful comment regarding correlated and conceptually related predictors. To address this concern, we have expanded the interpretability analyses in two ways:

First, we now explicitly acknowledge in the Discussion that several top-ranked predictors reflect overlapping behavioral and motivational domains (e.g., leisure-time hobbies and engagement in interests outside work), and that their relative importance should be interpreted as complementary rather than independent effects (lines 442-448).
Second, to enhance interpretability beyond impurity-based feature importance, we have added partial dependence plots (PDPs) for the most influential predictors identified by the Random Forest models. These plots illustrate the marginal association between individual predictors and Joyous Exploration while averaging over other features, thereby clarifying how conceptually related variables contribute to model predictions. The PDPs are presented in Figure 7 and A1.

Comment 5: While the paper identifies psychosocial predictors, it stops short of translating these into actionable insights for interventions. How should practitioners use this model to design or prioritize programs that foster curiosity in older adults? Expand the Discussion with a concrete proposal for community-based or clinical applications based on top predictors.

Response 5:

Thank you for this suggestion. We have expanded the Discussion (lines 421-430) to include a dedicated applied implications paragraph that translates the most consistent predictors into concrete, testable directions for community-based and geriatric programming. Specifically, we outline how programs emphasizing meaningful engagement, use of accumulated skills and experience, and intergenerational contribution (e.g., hobby-based groups, mentoring, and community education initiatives) may align with higher Joyous Exploration in later life. We also explicitly clarify that these implications are associative rather than causal, consistent with the cross-sectional design.

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

Thank you for the revised submission and detailed response to prior reviewer comments. The manuscript has improved substantially. I have several remaining comments that I consider important for the validity and interpretability of the modeling results.

The manuscript mentions selecting the “top 20 predictors” and then retraining/evaluating a reduced model. Please clarify whether the top 20 were selected using training data only. If feature selection used the full dataset, performance estimates and predictor rankings may be optimistic. I strongly recommend nested cross-validation (inner loop for feature selection, outer loop for performance estimation) or an equivalent approach that prevents leakage.
The manuscript states Random Forest had the strongest overall performance. The benchmark table indicates logistic regression outperforms RF on several imbalance-relevant metrics (balanced accuracy, minority-class recall, and macro F1), while RF leads on ROC-AUC. Please revise the text to match the table and justify the chosen primary model based on a pre-specified primary metric. Consider presenting logistic regression alongside RF as a co-primary comparator, given its stronger performance on key metrics for class imbalance.
The Results note that minority-class performance remained unchanged, but the confusion matrix/classification report indicate changes in recall and error counts across models. Please ensure all narrative statements align with the reported confusion matrices and metrics.
Methods state that mean and SD will be reported for ROC-AUC, balanced accuracy, and minority-class recall under repeated stratified k-fold CV. In the main Results, CV reporting appears incomplete (e.g., recall and/or PR-AUC and macro F1 are not consistently reported alongside ROC-AUC and balanced accuracy). Please present a consistent set of CV metrics (at least ROC-AUC, PR-AUC, balanced accuracy, macro F1, and minority-class recall) with mean and SD.
In the manuscript, predictors are item-level Likert responses. Please specify how these ordinal items were encoded (treated as numeric, one-hot, or otherwise), whether scaling was applied.
Because recruitment occurred via online channels, the ≥60 sample may over-represent healthier, more educated, and more digitally engaged older adults. I recommend acknowledging this potential bias more directly and discussing how it may affect generalizability of predictor rankings and performance.
Please proofread for grammatical errors.

Overall, I think the paper is close. Addressing the concerns above would substantially strengthen the manuscript.

Author Response

Dear Reviewer,

We sincerely thank you for the time and care invested in reviewing our manuscript and for the constructive feedback provided. Your thoughtful and detailed comments, which have substantially strengthened both the statistical rigor and conceptual clarity of the manuscript.

In response, we have revised the manuscript carefully and addressed each comment point by point, as detailed below and in the annotated response document. All revisions in the manuscript are highlighted in its corresponding colors to facilitate review.

Comment 1:The manuscript mentions selecting the “top 20 predictors” and then retraining/evaluating a reduced model. Please clarify whether the top 20 were selected using training data only. If feature selection used the full dataset, performance estimates and predictor rankings may be optimistic. I strongly recommend nested cross-validation (inner loop for feature selection, outer loop for performance estimation) or an equivalent approach that prevents leakage.

Response 1: Thank you for this clarification. We have now explicitly clarified in the Methods (Lines 190-196) that feature selection was conducted within a nested repeated stratified cross-validation framework. Specifically, within each outer fold (5 folds x 10 repeats), the top 20 predictors were identified using the training partition only, after which the reduced Random Forest model was refit on the same training data and evaluated on the held-out test fold.

Nested cross-validated performance estimates for the reduced (top 20) model under the primary 75/25 split, and the other sensitivity analyses (35/35, 40/40, and median split 50/50) listing the mean ± SD for ROC-AUC, PR-AUC, balanced accuracy, macro-F1, and recall for the high-JE class are now reported in lines 335-342, Tables 5 and A10, respectively. – All changes for this response are highlighted in blue.

Comment 2: The manuscript states Random Forest had the strongest overall performance. The benchmark table indicates logistic regression outperforms RF on several imbalance-relevant metrics (balanced accuracy, minority-class recall, and macro F1), while RF leads on ROC-AUC. Please revise the text to match the table and justify the chosen primary model based on a pre-specified primary metric. Consider presenting logistic regression alongside RF as a co-primary comparator, given its stronger performance on key metrics for class imbalance.

Response 2: Thank you for this important correction and clarification. We have revised the manuscript to align the text explicitly with the benchmark results in Table A1 and to clarify the rationale for model selection.

We now explicitly state in section 2.3 (lines 175-187) that while logistic regression demonstrated stronger performance on several threshold-dependent, imbalance-sensitive metrics (balanced accuracy, recall for the high-JE class, and macro-F1), the RF model showed superior discrimination performance as indexed by ROC-AUC and PR-AUC (see Table A1). Because the primary aim of the classification analysis was to identify and rank psychosocial correlates of high JE, rather than to optimize sensitivity at a single decision threshold, ROC-AUC was specified a priori as the primary evaluation metric.

Therefore, RF models were retained as the primary models for subsequent classification, feature selection, and interpretive analyses. Logistic regression is presented as a benchmarking comparator illustrating alternative performance trade-offs under class imbalance. – All changes for this response are highlighted in grey.

Comment 3: The Results note that minority-class performance remained unchanged, but the confusion matrix/classification report indicate changes in recall and error counts across models. Please ensure all narrative statements align with the reported confusion matrices and metrics.

Response 3: Thank you for noting this inconsistency. We have revised the Results text (lines 363-366) to accurately reflect the observed changes in minority-class performance. The updated wording now states that recall for the high-JE group improved modestly in the reduced (top 20) model (from 0.73 to 0.82), consistent with the confusion matrices and classification metrics reported in Table 4 and Figure 4. – All changes for this response are highlighted in green.

Comment 4: Methods state that mean and SD will be reported for ROC-AUC, balanced accuracy, and minority-class recall under repeated stratified k-fold CV. In the main Results, CV reporting appears incomplete (e.g., recall and/or PR-AUC and macro F1 are not consistently reported alongside ROC-AUC and balanced accuracy). Please present a consistent set of CV metrics (at least ROC-AUC, PR-AUC, balanced accuracy, macro F1, and minority-class recall) with mean and SD.

Response 4: Thank you for noting the incomplete cross-validation (CV) reporting. The Results to present a consistent set of repeated stratified 5-fold CV metrics (10 repetitions), reported as mean ± SD, as requested. Specifically, we now report ROC-AUC, PR-AUC, balanced accuracy, macro F1, and minority-class recall (high-curiosity class) for the primary 75/25 split in Table 4, and for all sensitivity analyses (35/35, 40/40, and 50/50 splits) in Tables A4, A6, and A8, respectively. We have revised the Methods (lines 234-235) to reflect these changes. All changes for this response are highlighted in yellow.

Comment 5: In the manuscript, predictors are item-level Likert responses. Please specify how these ordinal items were encoded (treated as numeric, one-hot, or otherwise), whether scaling was applied.

Response 5: Thank you for this clarification request. We have revised the manuscript to explicitly describe the encoding and preprocessing of item-level Likert predictors. In the Measures section, we now specify that ordinal Likert-type items were treated as numeric ordinal variables reflecting their original response scales (lines 107–109). In the Procedure and Analysis section, we clarify that no feature scaling or normalization was applied, as the analytic models employed do not rely on distance-based assumptions or scale-sensitive splitting criteria (lines 178–179; Hastie et al., 2009). – All changes for this response are highlighted in teal.

Comment 6: Because recruitment occurred via online channels, the ≥60 sample may over-represent healthier, more educated, and more digitally engaged older adults. I recommend acknowledging this potential bias more directly and discussing how it may affect generalizability of predictor rankings and performance.

Response 6: Thank you for this thoughtful suggestion. We have revised the Limitations section (lines 494-502) to more explicitly acknowledge potential selection bias introduced by online recruitment, particularly within the ≥60 subgroup. Specifically, we now note that the sample may over-represent healthier, more educated, and more digitally engaged older adults, and we discuss how this may limit the generalizability of both predictor rankings and model performance to the broader older population. We also emphasize the need for replication using offline or mixed-mode recruitment strategies. – All changes for this response are highlighted in pink/purple.

Comment 7: Please proofread for grammatical errors.

Response 7: Thank you for this suggestion. We have carefully revised the text throughout to correct grammatical and typographical issues, improve sentence clarity, and ensure consistent terminology (i.e., standardizing references to the outcome as “high-JE”). All revisions are highlighted in grey, with no substantive content removed.

We hope that the revised version is now suitable for publication and would be happy to address any additional questions or concerns.

Thank you again for your thoughtful review and guidance.

Sincerely,

Authors.

Reviewer 2 Report

Comments and Suggestions for Authors

I thank the authors for the revision and do not have further comments. Wishing the authors best of luck in their future research endeavors.

Author Response

Dear Reviewer,

We sincerely thank you for your careful evaluation of the revised manuscript and for your thoughtful and constructive guidance throughout the review process. Your detailed feedback was tremendously helpful and played a key role in significantly improving the statistical rigor, conceptual clarity, and overall quality of the manuscript. We also wish you the very best in your future research endeavors.

Sincerely,

Authors.

Reviewer 4 Report

Comments and Suggestions for Authors

I thank the authors for their thoughtful and thorough responses to the previous review, which have significantly improved the manuscript. The methodological clarifications, particularly regarding model comparisons, the operationalization of high curiosity, and cross-validation strategies, were clear and well executed.

As a secondary suggestion, I believe the manuscript would benefit from referencing recent work that has addressed similar challenges related to small, unbalanced datasets and feature-rich behavioral models. For example, Trabassi et al. (Optimizing rare disease gait classification through data balancing and generative AI: insights from hereditary cerebellar ataxia, 2024) demonstrate how data balancing and generative AI can improve predictive modeling in the context of gait classification for rare neurodegenerative diseases. Although the clinical context differs, the methodological overlap, particularly in handling limited, high-level observations and maximizing statistical power, makes the study a valuable reference in the current context.

Author Response

Dear Reviewer,

In response, we have revised the manuscript carefully and addressed your secondary suggestion:

Comment 1: As a secondary suggestion, I believe the manuscript would benefit from referencing recent work that has addressed similar challenges related to small, unbalanced datasets and feature-rich behavioral models. For example, Trabassi et al. (Optimizing rare disease gait classification through data balancing and generative AI: insights from hereditary cerebellar ataxia, 2024) demonstrate how data balancing and generative AI can improve predictive modeling in the context of gait classification for rare neurodegenerative diseases. Although the clinical context differs, the methodological overlap, particularly in handling limited, high-level observations and maximizing statistical power, makes the study a valuable reference in the current context.

Response 1: Thank you for this helpful suggestion. In response, we have incorporated references to prior work addressing small, class-imbalanced, and feature-rich datasets to contextualize our analytic approach. Specifically, we added a sentence in the Limitations section (lines 493-498) noting that similar data conditions are common in applied predictive modeling and can yield meaningful and interpretable results when appropriate robustness checks are applied, while emphasizing the need for replication. The added text and citations are highlighted in red.

We hope that the revised version is now suitable for publication and would be happy to address any additional questions or concerns.

Thank you again for your thoughtful review and guidance.

Sincerely,

Authors.