Vulnerability to Heat Effects and Regional Inequalities Among Older Adults in the State of São Paulo, Brazil

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

This study presents a novel and valuable contribution to the understanding of health outcomes related to heatwaves, utilizing long-term data from a large sample size. However, the rationale behind the weighting procedure used is not clearly explained, and it would be beneficial to provide justification for this methodology. Additionally, the analysis does not account for other confounding variables, such as humidity and air pollutants, which are also strongly associated with mortality alongside heat stress. the discussion of the same will further aid the in the utility of the paper.

Author Response

Comments 1:

However, the rationale behind the weighting procedure used is not clearly explained, and it would be beneficial to provide justification for this methodology.

Response 1

Thank you for this important comment. We agree that the rationale underlying the weighting procedure used in the construction of the Climate Vulnerability Index required further clarification. In response, we have revised the Methods section to explicitly justify both the relative contribution of each domain and the weights assigned to individual indicators within domains.

A greater total weight was assigned to the sensitivity domain (0.65) than to adaptive capacity (0.35) to reflect the primary aim of the index, which was to capture differential health susceptibility to heat-wave exposure among older adults under conditions of relatively homogeneous climatic exposure across Regional Health Departments. Under these conditions, regional variation in heat-related outcomes is more likely to arise from differences in population-level physiological and epidemiological sensitivity than from differences in exposure itself.

The weighting structure was defined according to the conceptual and temporal proximity of each indicator to the risk of acute cardiorespiratory decompensation associated with heat-wave exposure. Because the index was designed to capture short-term vulnerability to heat-related health events among older adults, greater weights were assigned to empirically estimated outcome indicators derived from the distributed lag non-linear models—namely, the cumulative relative risks of cardiorespiratory hospitalizations (0.25) and mortality (0.25). These indicators directly reflect observed population-level responses to heat exposure. Equal weighting was adopted to avoid biasing the index toward either service availability (which may influence hospitalization rates) or fatal outcomes (which may be affected by access to care), thereby capturing different levels of clinical severity without privileging health system performance.

The proportion of older adults (0.15) was assigned a lower weight as it represents a demographic predisposing factor rather than a direct measure of heat-related outcome manifestation.

Within the adaptive capacity domain, sanitation coverage (0.10) and primary healthcare coverage (0.09) were assigned relatively higher weights due to their more immediate relevance to hydration, behavioral adaptation, early detection of decompensation, and community-based monitoring during heat events. In contrast, HDI (0.08), health workforce density (0.05), and public health financing indicators (0.03) were weighted more conservatively, reflecting their more distal relationship with short-term physiological responses to thermal stress and their indirect influence on acute heat-related morbidity and mortality.

To evaluate the potential influence of normative weighting choices, a sensitivity analysis using equal weights across all indicators was conducted. The resulting index showed high concordance with the theoretically weighted version, with only minor reclassification between adjacent vulnerability categories across four RHDs and no changes in the overall spatial pattern. This supports the robustness of the adopted weighting strategy.

These clarifications have been incorporated into the revised manuscript (Section 2.8, Page 9-10, Lines 332-351).

Comments 2: Additionally, the analysis does not account for other confounding variables, such as humidity and air pollutants, which are also strongly associated with mortality alongside heat stress. The discussion of the same will further aid the utility of the paper.

Response 2: We agree that environmental co-exposures such as relative humidity and ambient air pollutants may influence heat-related health outcomes and represent important determinants of cardiorespiratory morbidity and mortality. As the present analysis was designed to estimate the short-term health impacts of synoptic-scale heat-wave events across the state of São Paulo, rather than to model combined environmental exposures at the microclimatic level, these co-exposures were not explicitly incorporated into the distributed lag non-linear modelling framework. This decision also considered the limited availability of consistent daily monitoring data for relative humidity and air pollutants across all 17 Regional Health Departments throughout the study period. Inclusion of these variables would have resulted in substantial spatial and temporal data loss, potentially compromising the comparability of regional estimates. Nevertheless, we acknowledge that high relative humidity may impair evaporative heat loss and increase thermal strain, while exposure to air pollutants such as particulate matter and ozone may exacerbate cardiovascular and respiratory vulnerability during heat-wave events. These factors may therefore interact with thermal stress and contribute to residual heterogeneity in the estimated associations between heat waves and cardiorespiratory outcomes. In response to the reviewer’s comment, we have expanded the Discussion section to acknowledge the potential role of these factors as effect modifiers of heat-related health risks among older adults. These limitations have now been explicitly addressed in Discussion (pages 25, Lines 598-613), and Limitations sections of the revised manuscript (page 28, Lines 749-755).

 

Reviewer 2 Report

Comments and Suggestions for Authors

I have carefully read the manuscript. The topic has been studied with great importance, which has become even more important in recent years, and in the future, the importance of mortality due to extreme heat caused by heat waves is likely to increase. As I am not a native English speaker, I cannot comment on the writing and English structure of the manuscript, and I recommend that a native English speaker check it. However, I did not see any problems with the writing and structure of the manuscript. The only change that could help the quality of the manuscript is to add the two words bioclimatology at the beginning and São Paulo at the end of the keywords. After this small change, the manuscript is acceptable in my opinion.

Author Response

Comments 1: The only change that could help the quality of the manuscript is to add the two words bioclimatology at the beginning and São Paulo at the end of the keywords.

Response 1: Thank you for pointing this out. In response, we have revised the list of keywords to include the term bioclimatology at the beginning and São Paulo at the end, as recommended. This change has been incorporated into the revised manuscript (Page 1, Line 34).

 

Reviewer 3 Report

Comments and Suggestions for Authors

Thank you for submitting your manuscript "Vulnerability to Heat Effects and Regional Inequalities among Older Adults in the State of São Paulo, Brazil" to the Journal of Ageing and Longevity. The manuscript has now been reviewed, including the supplementary material, however substantive methodological concerns must be addressed before the manuscript can be considered for publication.

A major revision is required. Please address the following points in full.

 

MAJOR CONCERNS

1. Heat wave definition and spatial homogeneity. The binary operationalization of heat wave exposure collapses a continuous variable into a single state-level trigger that is, by the authors' own admission in Supplementary Table S2, invariant across RHDs (weight = 0, as differentiation by RHD was not possible). When exposure does not vary across the analytic units, the DLNM produces RR heterogeneity driven entirely by outcome variation. This has direct implications for the validity of using those RRs as index inputs and for framing the DLNM as a regionally differentiated exposure-response analysis. The authors should explicitly acknowledge this limitation in the main text and verify that the citation used to justify this methodological choice (Estoque et al. 2023) specifically supports its application in subnational DLNM contexts.

 

2. DLNM specification. The manuscript does not report degrees of freedom for seasonality/trend smooth functions, cross-basis knot placement, or basis function type. Since exposure is binary, the non-linear component of the DLNM requires justification. The authors should clarify whether a simpler distributed lag model would be equivalent and provide overdispersion checks and residual autocorrelation diagnostics.

 

3. Confidence interval width. Several RRs in Table 3 present extremely narrow confidence intervals (e.g., RHD I mortality 1.18 [1.18–1.19]). This requires explanation. The authors should confirm whether quasi-Poisson or negative binomial specifications were tested and whether variance estimation accounts for autocorrelation.

 

4. Vulnerability index weighting, circularity, and transparency of sensitivity analysis. The weighting justifications provided in Supplementary Table S2 are noted but do not resolve two outstanding issues. First, the two RR components together carry 0.50 of the total index weight, meaning the index is substantially a rescaled function of the DLNM output from the same dataset. This circularity is not acknowledged in the main text and should be. Second, the manuscript states that "only four regions changed vulnerability category" under equal weights, but does not disclose that two of those changes involve RHDs currently classified as High vulnerability (Araçatuba and Araraquara) being reclassified as Moderate — a directionally meaningful finding for a tool explicitly designed for policy prioritization. This should be reported transparently in the main text, not only in the supplementary table.

 

5. Asymmetric outcome periods. Supplementary Table S4 confirms that heat wave data for 2024 exists (9 events recorded), yet mortality data ends at 2023 while hospitalization data runs through 2023–2024. The reason for this asymmetry is not stated. Since both RRs feed directly into the vulnerability index, the difference in temporal coverage affects their comparability and should be explicitly justified or resolved.

 

6. ERA5 validation. The sole validation reference cited for ERA5 in Brazil (Araújo et al. 2025) covers the Caatinga biome, which is climatically distinct from São Paulo state. A geographically relevant validation reference or direct station-comparison should be provided.

 

7. Confounding. The models do not appear to control for relative humidity, co-pollutants (PM₂.₅, ozone), or day-of-week effects — standard inclusions in environmental time-series studies. Their exclusion should be explicitly justified.

 

8. Harvesting. The 0–15 day lag window may conflate net excess mortality with mortality displacement. The authors should discuss this limitation or provide a harvesting-specific sensitivity analysis.

 

9. Separate cardiovascular and respiratory results. The rationale for combining outcomes is accepted, but separate results for cardiovascular and respiratory causes should be provided in supplementary material.

 

10. Variable description. The primary care coverage variable in Table 1 appears to be a composite index (range 1.87–3.23), not a percentage, despite its label. Units and scales for all composite indicators should be defined in the main text, not deferred to supplementary material.

 

MINOR CONCERNS

- The heat wave threshold is described as "3–5°C" above the seasonal mean. The specific threshold applied should be stated precisely in both the abstract and methods.

- The GitHub link in Section 2.9 points to the microdatasus package repository, not to study-specific analytical code. Please provide the correct repository link or clarify.

- Figure 1 caption attributes the DLNM results to "DATASUS 2024" as source; this should read as the authors' own analysis of DATASUS data.

- The SUS background in the Introduction (approximately 300 words across three paragraphs) is disproportionate to its contextual function and should be condensed.

- Please verify the dual citation of reference 25 on page 19 — Santos et al. and Fernandez-Medina et al. appear to be conflated.

Author Response

Comments 1: Heat wave definition and spatial homogeneity. The binary operationalization of heat wave exposure collapses a continuous variable into a single state-level trigger that is, by the authors' own admission in Supplementary Table S2, invariant across RHDs (weight = 0, as differentiation by RHD was not possible). When exposure does not vary across the analytic units, the DLNM produces RR heterogeneity driven entirely by outcome variation. This has direct implications for the validity of using those RRs as index inputs and for framing the DLNM as a regionally differentiated exposure-response analysis. The authors should explicitly acknowledge this limitation in the main text and verify that the citation used to justify this methodological choice (Estoque et al. 2023) specifically supports its application in subnational DLNM contexts.

Response 1: We thank the reviewer for this important observation. We agree that the binary operationalization of heat-wave exposure, derived from ERA5 temperature data at the state level, results in limited spatial variability across Regional Health Departments (RHDs), as acknowledged in Supplementary Table S2. Because heat-wave events occurred synchronously across the state during the study period, all RHDs shared the same sequence of heat-wave and non–heat-wave days, precluding meaningful spatial differentiation in exposure. In this context, the DLNM framework does not estimate spatially differentiated exposure–response relationships in the conventional sense. Rather, the observed heterogeneity in cumulative relative risks across RHDs reflects regional differences in population response to a common climatic stressor rather than spatial variation in exposure intensity itself. To address this limitation transparently, we have revised the Methods, Results, Discussion, and Limitations sections to clarify that DLNM-derived cumulative relative risks were incorporated into the vulnerability index as empirical indicators of regional sensitivity to heat-wave events under relatively homogeneous exposure conditions, and should not be interpreted as reflecting spatial variation in exposure–response relationships. We further acknowledge that this modelling assumption may affect the validity of incorporating these RRs as sensitivity components within the vulnerability index. These clarifications have been incorporated in Sections 2.7 (page 8, lines 288–289), 2.8 (page 8-9, lines 301-313), 3 (page 11, lines 399–401), 4 (page 26, lines 654–660), and 5 (page 28, lines 765-771) of the revised manuscript. Furthermore, we have verified that the reference to Estoque et al. (2023) is used in accordance with the revised IPCC vulnerability framework, in which exposure is conceptualized as a contextual hazard distinct from spatially varying components of sensitivity and adaptive capacity (page 8-9, lines 301–313).

Comments 2: DLNM specification. The manuscript does not report degrees of freedom for seasonality/trend smooth functions, cross-basis knot placement, or basis function type. Since exposure is binary, the non-linear component of the DLNM requires justification. The authors should clarify whether a simpler distributed lag model would be equivalent and provide overdispersion checks and residual autocorrelation diagnostics.

Response 2: We thank the reviewer for this important methodological comment. Additional details regarding the DLNM specification have now been incorporated into the Methods section. Heatwave exposure was modelled as a binary variable (1 = heatwave day; 0 = non–heatwave day). Accordingly, the exposure–response dimension of the cross-basis was specified using a linear function, while flexibility was applied exclusively to the lag dimension through a natural cubic spline with 3 degrees of freedom over a lag period of 0–15 days. This approach allows for the estimation of delayed effects of heatwave exposure on health outcomes while avoiding unnecessary non-linearity in the exposure–response relationship. Long-term trends and seasonality were controlled using a penalized smooth function of time, with approximately 8 degrees of freedom per year of observation (k = 88), and day-of-week effects were included as a categorical variable. Overdispersion was assessed by comparing residual deviance to model degrees of freedom, and sensitivity analyses using standard Poisson and negative binomial specifications were conducted to verify the stability of cumulative relative risk estimates. Residual autocorrelation was evaluated through inspection of autocorrelation function (ACF) plots. Given the linear specification of the exposure–response dimension, the adopted DLNM framework is functionally equivalent to a distributed lag model with a flexible lag structure. These clarifications have been incorporated in Section 2.6 (page 7-8, lines 264–278).     Comments 3: Confidence interval width. Several RRs in Table 3 present extremely narrow confidence intervals (e.g., RHD I mortality 1.18 [1.18–1.19]). This requires explanation. The authors should confirm whether quasi-Poisson or negative binomial specifications were tested and whether variance estimation accounts for autocorrelation. Response 3: The narrow confidence intervals observed in some Regional Health Departments (e.g., RHD I – Greater São Paulo) primarily reflect the large volume of daily cardiorespiratory events accumulated over the 11-year observation period, particularly in highly populated regions. To ensure that this apparent precision was not due to variance underestimation, overdispersion was formally assessed by comparing the residual deviance with the model degrees of freedom. When appropriate, models were specified using a quasi-Poisson distribution to account for extra-Poisson variability. In addition, alternative model specifications assuming standard Poisson and negative binomial distributions were fitted as sensitivity analyses. The cumulative relative risk estimates remained stable across these alternative specifications. Residual autocorrelation was also evaluated through inspection of autocorrelation function (ACF) plots, with no meaningful remaining temporal structure detected. These results indicate that the narrow confidence intervals reflect high statistical power in regions with large event counts rather than inadequate modelling of variance. A brief clarification regarding the interpretation of narrow confidence intervals in highly populated regions has also been added to the Discussion section of the revised manuscript (page 24, lines 548-551).   Comments 4: Vulnerability index weighting, circularity, and transparency of sensitivity analysis. The weighting justifications provided in Supplementary Table S2 are noted but do not resolve two outstanding issues. First, the two RR components together carry 0.50 of the total index weight, meaning the index is substantially a rescaled function of the DLNM output from the same dataset. This circularity is not acknowledged in the main text and should be. Second, the manuscript states that "only four regions changed vulnerability category" under equal weights, but does not disclose that two of those changes involve RHDs currently classified as High vulnerability (Araçatuba and Araraquara) being reclassified as Moderate — a directionally meaningful finding for a tool explicitly designed for policy prioritization. This should be reported transparently in the main text, not only in the supplementary table. Response 4: We acknowledge that the inclusion of DLNM-derived cumulative relative risks as weighted components of the vulnerability index may introduce partial circularity, as these estimates were derived from the same outcome data used in index construction. To address this concern, the Limitations section has been revised to explicitly recognize this potential overlap between model-derived sensitivity indicators and outcome-based vulnerability measures (page 28-29, lines 772-774). In addition, the Methods section has been revised to remove interpretative statements regarding the results of the equal-weight sensitivity analysis (page 10, lines 372-377). The Results section has been updated to transparently report the findings of this analysis in the main text. Under equal weighting, four Regional Health Departments (RHDs) changed vulnerability category compared to the primary weighted specification. Notably, the RHDs of Araçatuba and Araraquara were reclassified from High to Moderate vulnerability, whereas São José do Rio Preto and Sorocaba shifted from Moderate to High vulnerability under equal weighting (Page 19-20, lines 497-503). These findings indicate that regional vulnerability rankings, and therefore index-based policy prioritization, may be sensitive to the weighting scheme applied. This consideration has also been incorporated into the Limitations section (page 29, lines 775–777). We agree that transparency regarding weighting sensitivity is essential for policy-oriented tools, and the revised manuscript now explicitly acknowledges the implications of weighting choices for regional prioritization.

Comments 5:

Asymmetric outcome periods. Supplementary Table S4 confirms that heat wave data for 2024 exists (9 events recorded), yet mortality data ends at 2023 while hospitalization data runs through 2023–2024. The reason for this asymmetry is not stated. Since both RRs feed directly into the vulnerability index, the difference in temporal coverage affects their comparability and should be explicitly justified or resolved.

 Response 5:

Mortality data from the Mortality Information System (SIM) were available through 2023 at the time of analysis, as data for 2024 had not yet been fully consolidated in DATASUS. To avoid potential bias arising from incomplete mortality records, mortality analyses were restricted to the period 2010–2023. In contrast, hospitalization data from the Hospital Information System (SIH-SUS) were available through 2024 and were analysed accordingly.

We acknowledge that the use of different temporal coverage for hospitalization and mortality outcomes may affect the direct comparability of DLNM-derived relative risks incorporated into the vulnerability index. This limitation has now been explicitly stated in the Methods (page 7, lines 244-249) and Limitations sections (page 29, lines 778-782) of the revised manuscript.

 

Comments 6:

ERA5 validation. The sole validation reference cited for ERA5 in Brazil (Araújo et al. 2025) covers the Caatinga biome, which is climatically distinct from São Paulo state. A geographically relevant validation reference or direct station-comparison should be provided.

Response 6:

We agree that the previously cited validation study for ERA5 temperature data was conducted in the Caatinga biome, which presents climatic characteristics distinct from those of São Paulo State.

To address this concern, we have revised the Methods section to include a geographically relevant validation study evaluating ERA5 temperature data against ground-based meteorological observations in the state of São Paulo (Santos Junior et al., 2022). In this study, ERA5-derived climatic variables were compared with station data from CIIAGRO and INMET, demonstrating moderate agreement in temperature-based indices under local climatic conditions.

We have also retained the global ERA5 evaluation by Hersbach et al. (2020) as supporting evidence of dataset performance across diverse climatic environments. These revisions have been incorporated in Section 2.4 of the revised manuscript (page 6, Lines 195-202).

 

Comments 7:

Confounding. The models do not appear to control for relative humidity, co-pollutants (PM₂.₅, ozone), or day-of-week effects — standard inclusions in environmental time-series studies. Their exclusion should be explicitly justified.

 

Response 7:

 We agree that environmental co-exposures such as relative humidity and ambient air pollutants (e.g., PM₂.₅ and ozone) may influence cardiorespiratory morbidity and mortality and are commonly considered in environmental time-series analyses.

However, consistent daily monitoring data for relative humidity and air pollutants were not available across all Regional Health Departments for the full study period. Inclusion of these variables would have resulted in substantial spatial and temporal data loss, potentially compromising the comparability of regional estimates across the 17 RHDs. Limitations sections (page 28, lines 749-755)

We note that day-of-week effects were included a priori in all primary models as a categorical variable (factor(dow)) to control for systematic intra-week variation in health outcomes.

In response to this comment, we have clarified in the Methods section (page 8, lines 272-273) that day-of-week was included in all models and have explicitly acknowledged the absence of humidity and air pollution variables as a limitation in the revised manuscript.

 

 

Comments 8:

Harvesting. The 0–15 day lag window may conflate net excess mortality with mortality displacement. The authors should discuss this limitation or provide a harvesting-specific sensitivity analysis.

Response 8:

We acknowledge that short-term increases in mortality associated with extreme heat may partially reflect mortality displacement (harvesting) among highly vulnerable individuals. The lag window of 0–15 days was selected to capture both immediate and delayed effects of heat-wave exposure, consistent with distributed lag structures commonly applied in environmental time-series studies of heat-related mortality. This window allows for short-term displacement effects to be partially absorbed within the cumulative risk estimation.

However, we recognize that this modelling framework was not designed to formally disentangle mortality displacement from net excess mortality beyond the selected lag period. In response to this comment, this limitation has now been explicitly acknowledged in the Discussion section of the revised manuscript (page 24, lines 576–580).

 

 

Comments 9:

Separate cardiovascular and respiratory results. The rationale for combining outcomes is accepted, but separate results for cardiovascular and respiratory causes should be provided in supplementary material.

Response 9:

 In response, cause-specific cumulative relative risk estimates for cardiovascular and respiratory outcomes have now been provided separately in the Supplementary Material (Supplementary Tables S5 and S6).

As described in the revised Methods section (Section 2.6), analyses were initially conducted separately for cardiovascular and respiratory causes for both hospitalizations and mortality. The cause-specific results showed directionally consistent and statistically significant associations across most Regional Health Departments, with magnitudes comparable to those observed for the combined cardiorespiratory outcome. Given the shared pathophysiological mechanisms underlying heat-related cardiorespiratory decompensation and to enhance statistical precision, cardiovascular and respiratory causes were combined into a single cardiorespiratory category for the primary inferential models.

A note has been included in the main text directing readers to the supplementary cause-specific results (page 7, line 252 and page 17, lines 559–462).

The revised Supplementary Material, including Tables S5 and S6, has been updated accordingly and resubmitted as a ZIP file.

 

 

Comments 10:

Variable description. The primary care coverage variable in Table 1 appears to be a composite index (range 1.87–3.23), not a percentage, despite its label. Units and scales for all composite indicators should be defined in the main text, not deferred to supplementary material.

Response 10:

The variable previously labelled as “Primary care coverage (%)” corresponds to a composite index derived from the “Coverage” thematic dimension of the Regional Synthetic Health Indicator of São Paulo State (ISRS/SP) [10], which combines estimated population coverage by primary care teams and monitoring of health conditionalities under the Bolsa Família Program. The index was standardized using z-scores and rescaled by adding a constant (+3), resulting in a dimensionless score ranging approximately from 0 to 6. The label in Table 1 has now been revised accordingly, and a detailed description of this variable has been included in the Methods section (page 9, lines 323-331).

 

MINOR CONCERNS

- The heat wave threshold is described as "3–5°C" above the seasonal mean. The specific threshold applied should be stated precisely in both the abstract and methods.

Response:
Heatwaves were defined as periods of at least three consecutive days with daily mean temperature exceeding the seasonal climatological mean by ≥ 3 °C. This definition has now been specified consistently in both the Abstract and Methods sections (page 6, line 216) of the revised manuscript.

 

- The GitHub link in Section 2.9 points to the microdatasus package repository, not to study-specific analytical code. Please provide the correct repository link or clarify.

Response:

 The GitHub link previously cited refers to the source code of the microdatasus R package used for mortality data extraction from the DATASUS platform, and not to study-specific analytical scripts. This has now been clarified in Section 2.9. Study-specific data processing and statistical modelling scripts are available from the corresponding author upon reasonable request.

 

 

- Figure 1 caption attributes the DLNM results to "DATASUS 2024" as source; this should read as the authors' own analysis of DATASUS data.

Response:

 The caption of Figure 1 has been revised to clarify that the DLNM results are based on the authors’ own analysis using mortality and hospitalization data obtained from SIM/SUS (DATASUS) (page 19, lines 479-480).

 

- The SUS background in the Introduction (approximately 300 words across three paragraphs) is disproportionate to its contextual function and should be condensed.

Response:

We thank the reviewer for this suggestion. The background description of the Brazilian Unified Health System (SUS) in the Introduction has been condensed to improve focus and proportionality, while retaining the contextual information relevant to the regional organization of healthcare services in the state of São Paulo. (Page 3, Lines 114-122)

 

- Please verify the dual citation of reference 25 on page 19 — Santos et al. and Fernandez-Medina et al. appear to be conflated.

Response:
We thank the reviewer for bringing this to our attention. The in-text citation on page 19 has been revised to ensure that Santos et al. and Fernandez-Medina et al. are correctly cited as separate references in accordance with the reference list.

 

 

 

 

 

Reviewer 4 Report

Comments and Suggestions for Authors

This manuscript addresses an important and timely public health issue by integrating heat-wave exposure, cardiorespiratory outcomes, and regional structural determinants into a composite Climate Vulnerability Index for older adults. The use of DLNM models combined with an IPCC-based vulnerability framework represents a methodologically sound and policy-relevant approach.

The study is well structured, analytically rigorous, and clearly situated within both the international heat–health literature and the Brazilian public health system context. The decision to analyze Regional Health Departments (RHDs) as meso-level governance units is particularly valuable from a planning and adaptation perspective.

The statistical modelling strategy is appropriate, and the results are clearly presented. The integration of cumulative relative risks into the vulnerability index is conceptually consistent with the sensitivity dimension of the IPCC framework. The sensitivity analysis strengthens the robustness of the index.

I have only minor suggestions for clarification:

1. The rationale for excluding the exposure dimension from the composite index could be slightly expanded, particularly by clarifying how spatial homogeneity of heat-wave exposure was statistically assessed.
2. A brief clarification on potential ecological fallacy limitations in the Discussion could further strengthen the methodological transparency.
3. While the manuscript discusses policy implications for SUS, a short paragraph outlining how this index could be operationalized within existing regional planning instruments would enhance practical applicability.

Overall, these are minor clarifications and do not affect the validity of the findings.

Author Response

Comments 1: The rationale for excluding the exposure dimension from the composite index could be slightly expanded, particularly by clarifying how spatial homogeneity of heat-wave exposure was statistically assessed.

Response 1: We thank the reviewer for this important suggestion. Heat-wave exposure was initially assessed at the regional level using daily mean temperature data from the ERA5 reanalysis dataset, extracted according to the geographic coordinates of the 17 Regional Health Departments (RHDs) in the State of São Paulo. Heat-wave events were defined as periods of at least three consecutive days with daily mean temperature exceeding 3°C above the 1991–2020 climatological reference. This preliminary assessment indicated that the temporal occurrence and frequency of heat-wave events were largely synchronous across RHDs, reflecting the synoptic-scale nature of extreme heat episodes in southeastern Brazil. Given the limited inter-regional variability in exposure, the climatic time series was subsequently treated as a common statewide hazard, and heat-wave events were identified using the spatially aggregated temperature series for the entire state. As a result, the binary exposure variable used in the models (heat wave: yes/no) did not exhibit inter-regional variability. Consequently, formal statistical tests of spatial homogeneity (e.g., ANOVA or non-parametric equivalents) were not applicable, as there was no regional variability in exposure to be compared. The analytical focus of the study was therefore on how differences in regional sensitivity and adaptive capacity modify vulnerability to a shared climatic stressor, rather than on spatial variation in exposure itself. This clarification has been included in page 6, lines 206-212.

Comments 2: A brief clarification on potential ecological fallacy limitations in the Discussion could further strengthen the methodological transparency.

Response 2:  A statement addressing the potential for ecological fallacy inherent to aggregated regional-level analyses has been added to the Limitations section (page 29, lines 783-787).   Comments 3: While the manuscript discusses policy implications for SUS, a short paragraph outlining how this index could be operationalized within existing regional planning instruments would enhance practical applicability. Response 3:  A paragraph has been added to the Discussion section to clarify how the proposed Climate Vulnerability Index could be operationalized within existing regional planning instruments of the Brazilian Unified Health System (SUS). Specifically, the revised text outlines how the index may be incorporated into the situational analysis phase of Regional Health Plans and used at the Regional Health Department level to inform preparedness planning within Intermanagerial Regional Commissions. This clarification has been included in the Discussion section (page 26, lines 677-684).

 

 

Round 2

Reviewer 3 Report

Comments and Suggestions for Authors

1. Confidence interval width. The explanation offered in lines 491–494 — that narrow confidence intervals in highly populated RHDs reflect large daily event counts — is plausible for RHD I (Greater São Paulo) but does not account for similarly narrow intervals observed in RHDs with substantially lower event counts. For example, RHD VIII (Franca) reports a mortality RR of 1.29 (1.28–1.30), yet its daily event counts are considerably smaller than those of Greater São Paulo. The current explanation is therefore selective and insufficient as a general account. The authors should either extend the explanation to cover all affected RHDs, or report the model-estimated dispersion parameters (e.g., quasi-Poisson dispersion estimates or negative binomial theta values) for each RHD so that readers can assess whether variance inflation was adequate.

2. Confounding by humidity and air pollution. The authors have added appropriate discussion (lines 541–556) and a Limitations paragraph (lines 692–698) acknowledging that relative humidity and co-pollutants were not modelled, citing the absence of consistent daily monitoring data across all RHDs for the full study period. This justification is stated as a general assertion rather than a verified one. The authors should clarify whether these data were genuinely unavailable for all 17 RHDs across the full study period, or whether availability was partial. If consistent data exist for even a subset of RHDs, a sensitivity analysis incorporating humidity or a co-pollutant proxy in those regions would substantially strengthen the data-availability claim. If data were entirely absent across all regions, this should be stated explicitly with reference to the relevant monitoring infrastructure.

3. Mortality displacement (harvesting). Lines 519–523 correctly acknowledge that the 0–15 day lag window does not allow formal distinction between short-term displacement and net excess mortality. However, the assertion that the modelling framework does not permit such a test is not accurate. Extending the cumulative lag window to 28–30 days and examining whether RRs attenuate or reverse is a standard approach within the DLNM framework already implemented in this study. The authors should either conduct this sensitivity analysis and report the results, or provide a specific methodological justification for why it is not feasible or informative in this context.

No further concerns are raised beyond those listed above. 

Author Response

Comments 1:

Confidence interval width. The explanation offered in lines 491–494 — that narrow confidence intervals in highly populated RHDs reflect large daily event counts — is plausible for RHD I (Greater São Paulo) but does not account for similarly narrow intervals observed in RHDs with substantially lower event counts. For example, RHD VIII (Franca) reports a mortality RR of 1.29 (1.28–1.30), yet its daily event counts are considerably smaller than those of Greater São Paulo. The current explanation is therefore selective and insufficient as a general account. The authors should either extend the explanation to cover all affected RHDs, or report the model-estimated dispersion parameters (e.g., quasi-Poisson dispersion estimates or negative binomial theta values) for each RHD so that readers can assess whether variance inflation was adequate.

Response 1:

We thank the reviewer for this important observation. We agree that the width of the confidence intervals cannot be explained solely by differences in average daily event counts across Regional Health Departments.

In time-series analyses using distributed lag non-linear models (DLNM), the precision of cumulative relative risk estimates depends not only on the average number of daily events but also on the total number of observations across the full time series, the cumulative number of events contributing to the estimation, and the smoothing structure applied to the lag–response function. Because the DLNM estimates cumulative effects across the lag dimension using the entire temporal series, regions with relatively low daily counts may still yield precise estimates when the study period spans many years and the total number of events is sufficiently large.

In the present study, even regions with relatively low daily mortality counts accumulated several thousand cardiorespiratory deaths over the study period, allowing stable estimation of cumulative relative risks. In addition, all RHD-specific models were estimated using identical temporal spans and model specifications, which contributes to comparable statistical precision across regions.

Overdispersion was explicitly addressed by fitting the models using a quasi-Poisson specification, and sensitivity analyses using standard Poisson and negative binomial models yielded similar cumulative relative risk estimates. These results suggest that variance inflation was adequately accounted for in the modelling framework.

We have revised the corresponding explanation in the manuscript to clarify that the width of the confidence intervals reflects both event counts and the modelling structure of the DLNM applied to the full time series (page 23, lines 508-513).

Comments 2:

Confounding by humidity and air pollution. The authors have added appropriate discussion (lines 541–556) and a Limitations paragraph (lines 692–698) acknowledging that relative humidity and co-pollutants were not modelled, citing the absence of consistent daily monitoring data across all RHDs for the full study period. This justification is stated as a general assertion rather than a verified one. The authors should clarify whether these data were genuinely unavailable for all 17 RHDs across the full study period, or whether availability was partial. If consistent data exist for even a subset of RHDs, a sensitivity analysis incorporating humidity or a co-pollutant proxy in those regions would substantially strengthen the data-availability claim. If data were entirely absent across all regions, this should be stated explicitly with reference to the relevant monitoring infrastructure.

Response 2:

 Environmental co-exposures such as relative humidity and air pollutants (e.g., PM₂.₅ and ozone) are indeed important factors in environmental epidemiology and may influence cardiorespiratory outcomes.

In Brazil, however, the ground-based monitoring network for air pollutants remains spatially heterogeneous and limited to a subset of urban centers. Several regions of the country lack continuous monitoring infrastructure, resulting in substantial data gaps for epidemiological analyses [27].

In addition, although model-derived estimates of particulate matter based on satellite observations or atmospheric reanalysis systems are increasingly available, previous validation studies conducted in Brazil have demonstrated systematic discrepancies between these modeled estimates and ground-monitoring measurements, requiring local calibration before their use in epidemiological analyses [37].

Because the present study aimed to maintain methodological comparability across all 17 Regional Health Departments and consistent daily ground-monitoring data were not uniformly available for the full study period across these regions, humidity and air pollution variables were not incorporated into the primary models.

We have clarified this point in the revised Discussion section (page 24, lines 573-576) and explicitly acknowledged the absence of these environmental covariates as a limitation of the study( page 27-28, lines 729-734).

Comments 3- 

Mortality displacement (harvesting). Lines 519–523 correctly acknowledge that the 0–15 day lag window does not allow formal distinction between short-term displacement and net excess mortality. However, the assertion that the modelling framework does not permit such a test is not accurate. Extending the cumulative lag window to 28–30 days and examining whether RRs attenuate or reverse is a standard approach within the DLNM framework already implemented in this study. The authors should either conduct this sensitivity analysis and report the results, or provide a specific methodological justification for why it is not feasible or informative in this context.

Respond 3

We thank the reviewer for this pertinent observation. Indeed, the investigation of potential mortality displacement (harvesting) can be explored within the DLNM framework by extending the cumulative lag window. Following the reviewer’s suggestion, we conducted a sensitivity analysis by extending the cumulative lag from 0–15 days to 0–30 days to evaluate whether the estimated relative risks (RR) would attenuate or reverse over longer lag periods.

The results of this analysis are presented in Supplementary Table S7. Overall, the cumulative RR estimates for the 0–30 day lag were consistently lower than those observed for the 0–15 day lag, indicating attenuation of the estimated effects when longer lag windows are considered. However, the RR values remained above unity across all Regional Health Departments, and the 95% confidence intervals continued to indicate a positive association between heatwave exposure and mortality.

For example, in DRS VI – Bauru the cumulative RR decreased from 1.75 (0–15 days) to 1.58 (0–30 days), while in DRS I – Greater São Paulo the RR declined from 1.18 to 1.04. Similar patterns of attenuation without risk reversal were observed across the remaining regions.

These findings suggest that although part of the short-term increase in mortality may reflect the advancement of deaths among highly vulnerable individuals, the observed excess mortality associated with heatwaves cannot be explained solely by short-term mortality displacement.

The sensitivity analysis using a cumulative lag of 0–30 days has been included in the Supplementary Material (Table S7) and referenced in the revised manuscript. A description of this sensitivity analysis has been added to the Methods section (Section 2.6, page 7, lines 244–247), the results are reported in the Results section (page 18, lines 442–443), and their interpretation has been incorporated into the Discussion section (page 23, lines 542–546). The previous statement referring to mortality displacement as a limitation has also been revised accordingly.