This study analysed data from 20,303 respondents aged 15 years and older nested within 312 communities drawn from the Indonesia Family Life Survey Wave 5 (IFLS-5). Depressive symptoms were assessed using the 10-item Centre for Epidemiologic Studies Depression Scale (CES-D-10). To examine and calibrate the measure, the authors applied a Rasch Partial Credit Model to assess the CES-D-10 measurement properties within the sample.
Flood exposure was considered as a community-level shock of interest and community membership provided the multilevel structure for analysis. The dataset permitted linking individual-level depressive symptom scores to community-level covariates representing observable resources.
The authors implemented Bayesian multilevel models to quantify heterogeneity in depressive symptoms across communities and to evaluate the extent to which observable community resources explain between-community differences. Two complementary constructs of community resilience were operationalised:
Structural resilience: an index based on observable socioeconomic indicators and measures of social capital intended to capture resource-based capacity at the community level.
Latent Community Protective Capacity (LCPC): a model-derived residual proxy intended to represent remaining contextual variation in depressive-symptom risk not captured by the structural indicators.
The modelling approach allowed partitioning variance between individual and community levels and estimating associations while accounting for clustering and uncertainty using Bayesian estimation.
The multilevel analysis found that about 6 percent of the variation in depressive symptoms was attributable to between-community differences. Observable structural resources accounted for only part of this heterogeneity, indicating that resource-based indices did not fully explain community-level differences in depressive-symptom risk.
This partial explanatory role suggests that additional contextual or unmeasured factors contribute to between-community variation in mental health following flood exposure.
The study reported a weak correlation between structural resilience and Latent Community Protective Capacity (LCPC) (r = 0.155). This low correlation indicates that the two constructs capture distinct aspects of community-level variation: structural resilience reflects observed socioeconomic and social-capital resources, while LCPC represents residual contextual features inferred from the multilevel model.
Because LCPC is derived from model residuals, it should be interpreted as a proxy for unobserved contextual protective or risk factors rather than a directly measured construct.
Moderation analyses tested whether community-level resilience indices modified the association between flood exposure and depressive symptoms. Results showed:
No clear evidence that structural resilience altered the flood–depression association.
LCPC exhibited a directionally consistent buffering pattern (i.e., its estimates suggested a mitigating effect), but this pattern was statistically uncertain and not definitive.
Overall, the moderation findings did not provide strong, unambiguous support for either index as a robust moderator of flood-related depressive-symptom risk in this sample.
Predictive models that incorporated community-level information improved discrimination for depressive-symptom outcomes compared with models lacking such information. The best-performing model in the analysis reached an ROC-AUC of approximately 0.71, indicating modest-to-moderate discriminative ability when community-level predictors and latent contextual measures were included.
While inclusion of community-level predictors enhanced prediction, the results also underscore limits to predictive accuracy and the need for additional or better-measured contextual variables.
Key interpretive points from the analysis are:
Observable resource-based indices provide an incomplete account of community-level vulnerability to depressive symptoms following flood exposure.
A model-derived residual measure (LCPC) may capture complementary contextual information that structural indicators miss, but LCPC is a proxy and requires cautious interpretation.
The weak correlation between structural resilience and LCPC signals that each index reflects different dimensions of community context; neither alone fully explains between-community heterogeneity.
Moderation evidence was inconclusive: structural resilience did not clearly buffer flood-related depressive symptoms, and LCPC's potential buffering effect was uncertain.
The authors emphasise the need for independent validation of residual contextual measures and careful interpretation before applying such proxies in policy or intervention design.
The Coventry University Ethics Committee waived the requirement for additional ethical approval for this secondary analysis of de-identified IFLS-5 data. The original IFLS-5 data collection received approvals from the RAND Corporation and Universitas Gadjah Mada institutional review boards, and participant informed consent was obtained in the primary data collection.
The IFLS-5 public-use data are available from RAND following free user registration. The analysis code used in this study is available from the corresponding author on reasonable request. The study declared funding from the UK Department for Environment, Food and Rural Affairs (Defra) and the authors reported no competing interests.
Limitations reported or implied by the analytic design include reliance on available IFLS-5 measures for structural resources and use of a model-derived residual proxy (LCPC) whose interpretation and external validity require additional work and replication in independent datasets.