|Dear authors, |
I reviewed the revised manuscript „Influence of antecedent conditions on flood risk in sub-Saharan Africa“. The authors tried to extent their statistical analysis and compare the antecedent short- and seasonal scale conditions of reported floods by MunicRe with the conditions in years, where no flood occurred. Simple statistical tests are conducted in order to test the significance of precipitation anomalies at different temporal scales.
I appreciate the effort and I believe, that the manuscript has improved from a methodological point of view. However, I am unfortunately not convinced that the findings are yet sufficient for publication. The major conclusion, that floods are triggered by a combination of the catchment state (as represented by SPEI) and high precipitation during the build-up period is rather trivial and has been investigated in various studies, mostly by means of more complex methods. The differences reported in the manuscript are mostly not very clear and often not statistically significant. Further the shortcomings of the datasets and methods impede the interpretation of the results.
Thus, as much as I regret it, I cannot recommend publication at the current state. I would like to encourage the authors to further advance their methods (particularly to use more complex inferential methods, which are suitable for the data) and re-submit a new-version, if more robust results could be achieved.
In the following I will raise my three major concerns:
1) Introduction, climatic and hydrological setting:
The study covers a large target region with various different climatic settings. No information is provided on climate-variability, large scale climatic circulation modes and on typical flood generation processes in different parts of the region. The Köppen classificastion (Fig.4) is not sufficient to describe the climate of the region (especially the very basic version, which classifies the center of the continent as oceanic?). It might be a step into the right direction to analyse the spatial climatic variability (SPEI). Maybe one could then find different flood types in different regions, which are characterized by different SPEI-flood relationships.
One major problem of the study remains the quality of the MunicRe data set. The comparison of observed and non-observed floods assumes, that the data set is a) somehow complete and b) that the climate-flood link is stationary. However, as the authors admit, there is a strong trend in the data, which points to increasing settlements in flood-prone areas or an increase of flood reports. Those problems might blur important statistical relationship.
In their recent publication (“Should seasonal rainfall forecasts be used for flood preparedness?”) the authors show (based on modelling results), that the precipitation-flood link varies over the target region and is highly dependent on the climatic conditions. In the presented manuscript, all floods are pooled, which also might blur clear results. I fear, that the quality MunicRe data set alone might be too poor and the number of reported floods too low to draw statistically significant conclusions and to derive interesting/new results.
3) Statistical Methods:
There are quite a few methodological problems. The assumptions of statistical tests are often violated by data sets. E.g. event-precipitation (Pre7 and Max7) have been z-normalized, although short-term precipitation is certainly not normal-distributed. Likewise the significance of different mean SPI values for flood and non-flood are based on a z-test and results are questionable. The 75%-quantiles in Fig. 6 clearly overlap, doesn’t that actually indicate, that differences are not significant?
Also the comparison of “flood probabilities” (Fig. 7) is not very robust, since it includes classes with different numbers of cases. Thus, single floods with anomalous pre-conditions can change the entire plot (and thus the interpretation of results). E.g. the sudden drop of SPEI0 is certainly a statistical artefact. Again I would rather use boxplots or similar methods, which include the range of values and some measure of significance. Further I wonder, why flood-likelyhood-ratio is always >1, even if SPEI is clearly negative. This might indicate either a data problem or a problem of the method.