Analyzing the Temporal Factors for Anxiety and Depression Symptoms with the Rashomon Perspective

This paper introduces a new modeling perspective in the public mental health domain to provide a robust interpretation of the relations between anxiety and depression, and the demographic and temporal factors. This perspective particularly leverages the Rashomon Effect, where multiple models exhibit similar predictive performance but rely on diverse internal structures. Instead of considering these multiple models, choosing a single best model risks masking alternative narratives embedded in the data. To address this, we employed this perspective in the interpretation of a large-scale psychological dataset, specifically focusing on the Patient Health Questionnaire-4. We use a random forest model combined with partial dependence profiles to rigorously assess the robustness and stability of predictive relationships across the resulting Rashomon set, which consists of multiple models that exhibit similar predictive performance. Our findings confirm that demographic variables \texttt{age}, \texttt{sex}, and \texttt{education} lead to consistent structural shifts in anxiety and depression risk. Crucially, we identify significant temporal effects: risk probability demonstrates clear diurnal and circaseptan fluctuations, peaking during early morning hours. This work demonstrates the necessity of moving beyond the best model to analyze the entire Rashomon set. Our results highlight that the observed variability, particularly due to circadian and circaseptan rhythms, must be meticulously considered for robust interpretation in psychological screening. We advocate for a multiplicity-aware approach to enhance the stability and generalizability of ML-based conclusions in mental health research.

💡 Research Summary

This paper introduces a “Rashomon perspective” for interpreting machine‑learning models in public mental‑health research, emphasizing that many near‑optimal models can exist with similar predictive performance yet divergent internal structures. Using a large PHQ‑4 dataset (over 34,000 adult respondents), the authors operationalize the Rashomon set through bootstrap resampling: they generate 500 bootstrap samples of the data, fit a random‑forest model to each, and collect all models whose error lies within a small ε of the best‑performing model. This collection constitutes the Rashomon set, representing a space of plausible explanations rather than a single chosen model.

To assess the stability of predictor effects across this set, the authors compute Partial Dependence Profiles (PDPs) for each model and overlay bootstrap‑based confidence intervals. The variables examined are demographic (age, sex, education) and temporal (hour of day, day of week). The analysis reveals that age, sex, and education consistently shift the predicted risk of anxiety and depression across all models: younger adults (18‑34) show markedly higher risk, risk declines with increasing age, females exhibit slightly higher risk than males, and higher education is associated with lower risk.

Temporal analysis uncovers robust diurnal and circaseptan patterns. Risk peaks in the early morning hours (approximately 3‑5 am) and reaches a trough during mid‑day. This daily rhythm persists across all days of the week, indicating a circaseptan (weekly) rhythm that is relatively stable, with only minor variations on Mondays and Fridays. The consistency of these PDP shapes and their confidence intervals across the entire Rashomon set suggests that the observed temporal effects are genuine signals in the data rather than artifacts of a particular model specification.

Methodologically, the paper makes three contributions: (1) it demonstrates how to construct a Rashomon set via bootstrap resampling, thereby quantifying model‑selection uncertainty; (2) it integrates PDPs with bootstrap confidence intervals to evaluate the robustness of predictor effects across near‑optimal models; and (3) it distinguishes the Rashomon approach from traditional ensemble learning, arguing that the former focuses on explanatory stability rather than purely predictive gain.

The authors acknowledge limitations: reliance on a single model family (random forests), absence of comparisons with other algorithms (e.g., gradient boosting, neural networks), the use of self‑reported PHQ‑4 scores which may contain measurement error, and the fact that temporal variables are limited to survey response timestamps, which may not fully capture underlying biological rhythms. They suggest future work should include external validation cohorts, broader model families, causal inference techniques, and integration of physiological data (e.g., actigraphy) to deepen understanding of circadian influences.

In conclusion, by “rashomonizing” the analysis—examining the collective behavior of many equally good models—the study provides a more nuanced and reliable picture of how demographic and temporal factors shape anxiety and depression risk. It highlights the necessity of moving beyond a single best model in mental‑health screening, especially when interpreting time‑dependent risk patterns that have clear implications for public‑health monitoring and intervention timing.