Continental Scale Regional Flood Frequency Analysis: Combining Enhanced Datasets and a Bayesian Framework
Aldo Gardini
Round 1
Reviewer 1 Report
Comments and Suggestions for AuthorsSee attached file.
Comments for author File: 
Comments.pdf
The quality of the English language is appropriate.
Author Response
Please see the response to reviewer #3 (page 7) of the attached pdf.
Author Response File: Author Response.pdf
Reviewer 2 Report
Comments and Suggestions for AuthorsThis paper conducted regional flood frequency analysis (RFFA) across a large spatial domain in North America using a Bayesian hierarchical model based on the GEV and GP distributions conditioned on 130 static catchment-specific covariates. The models’ return levels align well with government design peak discharge data.
This study is of interest to the journal readership. However, some key points need to be addressed before publication. I hope that my comments help in preparing an improved paper. I believe this manuscript is valuable and deserves publication after the comments have been satisfactorily addressed.
Major Comments:
1. The abstract highlights certain covariates as key (L15-16), but the paper lacks transparency on how strongly these covariates are associated/correlated with floods. This should be clarified in the abstract and/or conclusions.
2. The paper states that a modification was made to include daily gauge discharge (L07-09), but only a very few events (exceedances over the 98th quantile) were included. This phrasing should be improved to reflect what was actually done. Additionally, the Generalized Pareto distribution should be clearly disclosed in the abstract.
3. The discussion and conclusion sections focus heavily on the study's limitations rather than the results. I suggest creating a separate subsection for limitations/future research, letting the discussion and conclusion concentrate more on the work done.
4. The separation of independent exceedances over a high threshold for the GP distribution assumed all events with a minimum spacing of three days are independent. This criterion seems questionable. For example, independent events have been selected using a minimum time spacing of 10 days + log(A), where A is the drainage area in square miles (Lang et al., 1999).
Minor comments:
5. Regarding major comment #2: Recent papers leverage information on extremes within ordinary events (e.g., daily streamflow) and address EVT limitations, including the assumption of infinite independent observations per year (which is not the case for floods) through Metastatistical approaches (e.g., Miniussi et al., 2020). These models allow conditioning on covariates like time and temperature (e.g., Vidrio-Sahagún et al., 2022 and 2023), which can extend predictions to ungauged basins. This could interest the authors and may be worth highlighting as potential future research to provide a broader perspective.
6. L98-100: what other previous studies?
7. L01-03 (Abstract): Consider rephrasing to emphasize that the main problem is the scarcity of flood observations rather than issues related to hydrological model simulations, as the latter emerge as a potential solution to the former.
8. L04: Specify that the study did not cover all of North America, as no Mexican or Caribbean data was included (L137), and the results focused on a smaller portion than the entire USA and Canada (L285-289).
9. L176-177: Discuss the rationale behind transforming the covariates to follow Gaussian distributions and the limitations of assuming their normality in the limitations section.
References
Lang, M., Ouarda, T. B., & Bobée, B. (1999). Towards operational guidelines for over-threshold modeling. Journal of hydrology, 225(3-4), 103-117. https://doi.org/10.1016/S0022-1694(99)00167-5
Miniussi, A., Marani, M., & Villarini, G. (2020). Metastatistical Extreme Value Distribution applied to floods across the continental United States. Advances in Water Resources, 136, 103498. https://doi.org/10.1016/j.advwatres.2019.103498
Vidrio-Sahagún, C. T., He, J., and Pietroniro, A. (2023). Nonstationary hydrological frequency analysis using the Metastatistical extreme value distribution. Advances in Water Resources, 176 (June), 104460. https://doi.org/10.1016/j.advwatres.2023.104460
Comments on the Quality of English LanguageThe writing is overall in good shape, and no major revisions regarding the quality of English are needed.
Author Response
Please see the response to reviewer #1 (page 1) of the attached pdf.
Author Response File: Author Response.pdf
Reviewer 3 Report
Comments and Suggestions for AuthorsSee my comments in the file.
Comments for author File: Comments.pdf
Author Response
Please see the response to reviewer #2 (page 3) of the attached pdf.
Author Response File: Author Response.pdf
Round 2
Reviewer 1 Report
Comments and Suggestions for AuthorsThank you for the answers you provided. However, I must insist on some points that were not treated satisfactorily in the previous round of revision. In particular:
1. Third point of issue 3: The term "posterior predictive distribution" has a precise meaning and is not subjective. I suggest the authors refer to page 7 of the cited book "Bayesian Data Analysis"; similar definitions can be found in any book on Bayesian statistics. A posterior predictive distribution strictly refers to a prediction of an unobserved outcome (i.e., in your case, an unobserved river discharge), and not a parameter. Equation 17 represents a posterior distribution, not a posterior predictive distribution. If you read only the formulas, you might think that your equation 17 is connected to the equation on page 41 of BDA. However, there is a significant difference: the latter equation refers to a normal model and includes $\tilde{y}$, i.e., a future observation under the normal model! This allows the use of "predictive" on page 41. In your case, you have a GEV model, and equation 17 does not contain any prediction; it is purely posterior. I must emphasize this: your paper heavily relies on statistics, and such fundamental errors need to be avoided.
2. Fourth point of issue 3: These quantities cannot be interpreted as residuals in a linear regression setting, as they are not residuals! You are checking that a quantity that your model forces to be normal is estimated as normal by your model itself. The model cannot assign a distribution different from normal to such a quantity! Performing this check leads to a vicious circle that should be avoided.
3. The comment on Figure 6 remains unchanged, despite the checks you illustrated in response to question 5 showing that it is not correct.
4. I am puzzled by the answer to question 4. You have many data points to check, yet you only show 3 plots (not mentioning the exercise in the paper). I agree that checking each station is not feasible, but it is usually useful to summarize such checks, focusing on the output of interest. If this model-assessment exercise aligns with your literature, I will refrain from further comments. However, I caution you that, from a statistical viewpoint, it is not rigorous, and the conclusions you draw are ex
Author Response
Please refer to the word document.
Author Response File: Author Response.pdf
Reviewer 3 Report
Comments and Suggestions for AuthorsThank you to the authors for addressing all my comments satisfactorily.
Author Response
Thank you for reviewing our manuscripts.
