the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
Causal interactions between ENSO and the North Tropical Atlantic
Abstract. The global climate is impacted by several major climate modes including the North Tropical Atlantic mode (NTA) and the El Niño–Southern Oscillation (ENSO). Although NTA and ENSO are suggested to have connections, there is uncertainty regarding the causal relationship between these climate modes. While previous works focused on the correlation between NTA and ENSO, causal analyses accounting for the influence of other tropical climate modes are lacking. Here we investigate the causal links between ENSO and NTA using outputs from high-resolution climate model simulations and reanalysis data. Our results suggest robust causal effects of ENSO on NTA and provide insights on the unstable impacts of NTA on ENSO. We observe high consistency between reanalysis data and the models in mimicking the impacts of ENSO on North Tropical Atlantic region. Specifically, most models (14 over 20) and reanalysis data revealed that ENSO is very unlikely to have no causal impacts on NTA. However, there is diverse response of the tropical Pacific to NTA between reanalysis data and the models. While reanalysis data indicates possible impacts of NTA on ENSO and sea surface temperature over the equatorial Pacific, the majority of models (18 over 20) suggest that the NTA is likely to have no causal effects on ENSO. Hence, the models may underestimate the causal effects of NTA on ENSO, implying that better representation of NTA variability and NTA-ENSO causal connections in the models may improve the predictability of ENSO variations.
- Preprint
(3308 KB) - Metadata XML
-
Supplement
(630 KB) - BibTeX
- EndNote
Status: closed
-
RC1: 'Comment on esd-2023-1', Anonymous Referee #1, 11 Apr 2023
The manuscript "Causal interactions between ENSO and the North Tropical Atlantic" by Thanh Le and Deg-Hyo Bae presents an assessment of the causal link between ENSO and NTA using both high-resolution climate models and reanalysis data. The idea of the manuscript is surely interesting and worthy of investigation, however it is difficult to read and the assessment needs to be strengthened from a statistical point of view. Below are my comments.
Major comments
- The authors used monthly data but reconstructed indices at annual resolution (Figure 1). Then, they use Eq. S1 to assess the causality which considers yearly resolution time series. But then the authors discuss seasonal effects of ENSO and NTA. I would suggest the authors to better clarity how the analysis is carried out, unless it is not possible to replicate the results they obtained.
- Another missing information is the number of lagged observations used for each model (p in Eq. S1) that is also strictly linked with my previous comment on the resolution. This can also affect the results since it depends on the resolution used and on the present/past observations. Which is the statistical threshold for significance?
- The authors claim for a causal link between NTA and ENSO with directionality pointing from ENSO to NTA. The causality has been assessed, if I correctly understood, by using indices at annual resolution (although they used monthly data, see comment 1). Indices are firstly standardized and then the Granger causality has been evaluated, whose assessment is based on the p-value based IPCC-based recommendations. However, the Granger causality needs to be assessed with respect to a null-hypothesis that requires a statistical basis. Which is this statistical basis? Apart the p-value, did the authors performed some statistical tests based, as an example, on bootstraps procedures or random phases?
- If the authors used yearly resolution, how the results are affected by the reduced size of samples (finite size effects)? It is the same for monthly resolution, thus a discussion and further additional tests are required to assess the robustness.
Minor comments
- Line 23: remove the comma before "typically".
- Line 24: remove the comma after "spring" and change with "and it".
- Lines 66-67: should be the "first principal component"?
- The quality of all figures needs to be improved.
Citation: https://doi.org/10.5194/esd-2023-1-RC1 - AC1: 'Reply on RC1', Thanh Le, 26 Jun 2023
-
RC2: 'Comment on esd-2023-1', Anonymous Referee #2, 11 Jun 2023
Review of MS No.: esd-2023-1
Causal interactions between ENSO and the North Tropical Atlantic
Author(s): Thanh Le and Deg-Hyo BaeThe authors study causal interactions between the North Tropical Atlantic mode (NTA) and the El Niño–Southern Oscillation (ENSO), accounting for possible compounding effects of the Indian Ocean Dipole (IOD) and the Indian Ocean Basin mode
(IOB), using a form of Granger causality (GC) analysis.
This is an interesting topic potentially bringing important results.
Reading the paper, a number of technical questions emerged.
Sampling: The authors use monthly data, however they evaluate seasonal indices, leading
to yearly data, i.e. they obtained the results using 65 samples. Questions:
1. Model order- they mention two criteria, but do not specify which was finally used and which order
of models was applied. Considering 65 samples, any order higher than 1 or 2 can be problematic,
considering 4 variables in the models.
2. The results are asymmetric, which might be interpretable as ENSO being more important climate mode.
On the other hand, by data construction, the effect of ENSO on NTA was evaluated using a time delay
of 3 months, while the other direction has inherent time lag 9 months. Can this play any role?
Would not be analysis using monthly data and different time lags interesting?
3. The results in Fig. 3 and other maps - are they obtained in the same way using yearly data, just
ENSO effect was evaluated on any grid-point separately? Is p-value mapped?
The question 3 leads to:
4. If many tests are presented, is any correction to multiple testing considered? This can be very critical esp.
in Fig. 7, where a few significant spots could appear by chance. Not to speak about very weak criteria
taking as significant also values of p 0.1 -0.3. Good to remind that typical conservative approach relies
on p<0.05. I would conclude that no effect of NTA on ENSO was detected, and very few models were able
to reproduce the effect of ENSO on NTA, observed in the reanalysis data. Btw. in Fig. 2 the results
for the reanalysis data is not visible because of small p-value? This should be mentioned in the caption and
the value should be written.
5. Another remark to significance levels - if they are obtained from an analytic expression, there are always
some data requirements. I propose to add some computational statistics such as surrogate data or bootstrap,
to avoid false significance.
6. Multiplicity correction applies also for sliding window results, e.g. Fig. 9e.
7. What would be the results without accounting for IOD and IOB?
8. If p-values is mapped in Figs. 3 and similar, I would not talk about the causal effect.
p-value gives the reliability of inference (of causality in this case) which is not generally
equivalent to (physical) causal effect of the cause on the effect variable.Citation: https://doi.org/10.5194/esd-2023-1-RC2 - AC2: 'Reply on RC2', Thanh Le, 26 Jun 2023
Status: closed
-
RC1: 'Comment on esd-2023-1', Anonymous Referee #1, 11 Apr 2023
The manuscript "Causal interactions between ENSO and the North Tropical Atlantic" by Thanh Le and Deg-Hyo Bae presents an assessment of the causal link between ENSO and NTA using both high-resolution climate models and reanalysis data. The idea of the manuscript is surely interesting and worthy of investigation, however it is difficult to read and the assessment needs to be strengthened from a statistical point of view. Below are my comments.
Major comments
- The authors used monthly data but reconstructed indices at annual resolution (Figure 1). Then, they use Eq. S1 to assess the causality which considers yearly resolution time series. But then the authors discuss seasonal effects of ENSO and NTA. I would suggest the authors to better clarity how the analysis is carried out, unless it is not possible to replicate the results they obtained.
- Another missing information is the number of lagged observations used for each model (p in Eq. S1) that is also strictly linked with my previous comment on the resolution. This can also affect the results since it depends on the resolution used and on the present/past observations. Which is the statistical threshold for significance?
- The authors claim for a causal link between NTA and ENSO with directionality pointing from ENSO to NTA. The causality has been assessed, if I correctly understood, by using indices at annual resolution (although they used monthly data, see comment 1). Indices are firstly standardized and then the Granger causality has been evaluated, whose assessment is based on the p-value based IPCC-based recommendations. However, the Granger causality needs to be assessed with respect to a null-hypothesis that requires a statistical basis. Which is this statistical basis? Apart the p-value, did the authors performed some statistical tests based, as an example, on bootstraps procedures or random phases?
- If the authors used yearly resolution, how the results are affected by the reduced size of samples (finite size effects)? It is the same for monthly resolution, thus a discussion and further additional tests are required to assess the robustness.
Minor comments
- Line 23: remove the comma before "typically".
- Line 24: remove the comma after "spring" and change with "and it".
- Lines 66-67: should be the "first principal component"?
- The quality of all figures needs to be improved.
Citation: https://doi.org/10.5194/esd-2023-1-RC1 - AC1: 'Reply on RC1', Thanh Le, 26 Jun 2023
-
RC2: 'Comment on esd-2023-1', Anonymous Referee #2, 11 Jun 2023
Review of MS No.: esd-2023-1
Causal interactions between ENSO and the North Tropical Atlantic
Author(s): Thanh Le and Deg-Hyo BaeThe authors study causal interactions between the North Tropical Atlantic mode (NTA) and the El Niño–Southern Oscillation (ENSO), accounting for possible compounding effects of the Indian Ocean Dipole (IOD) and the Indian Ocean Basin mode
(IOB), using a form of Granger causality (GC) analysis.
This is an interesting topic potentially bringing important results.
Reading the paper, a number of technical questions emerged.
Sampling: The authors use monthly data, however they evaluate seasonal indices, leading
to yearly data, i.e. they obtained the results using 65 samples. Questions:
1. Model order- they mention two criteria, but do not specify which was finally used and which order
of models was applied. Considering 65 samples, any order higher than 1 or 2 can be problematic,
considering 4 variables in the models.
2. The results are asymmetric, which might be interpretable as ENSO being more important climate mode.
On the other hand, by data construction, the effect of ENSO on NTA was evaluated using a time delay
of 3 months, while the other direction has inherent time lag 9 months. Can this play any role?
Would not be analysis using monthly data and different time lags interesting?
3. The results in Fig. 3 and other maps - are they obtained in the same way using yearly data, just
ENSO effect was evaluated on any grid-point separately? Is p-value mapped?
The question 3 leads to:
4. If many tests are presented, is any correction to multiple testing considered? This can be very critical esp.
in Fig. 7, where a few significant spots could appear by chance. Not to speak about very weak criteria
taking as significant also values of p 0.1 -0.3. Good to remind that typical conservative approach relies
on p<0.05. I would conclude that no effect of NTA on ENSO was detected, and very few models were able
to reproduce the effect of ENSO on NTA, observed in the reanalysis data. Btw. in Fig. 2 the results
for the reanalysis data is not visible because of small p-value? This should be mentioned in the caption and
the value should be written.
5. Another remark to significance levels - if they are obtained from an analytic expression, there are always
some data requirements. I propose to add some computational statistics such as surrogate data or bootstrap,
to avoid false significance.
6. Multiplicity correction applies also for sliding window results, e.g. Fig. 9e.
7. What would be the results without accounting for IOD and IOB?
8. If p-values is mapped in Figs. 3 and similar, I would not talk about the causal effect.
p-value gives the reliability of inference (of causality in this case) which is not generally
equivalent to (physical) causal effect of the cause on the effect variable.Citation: https://doi.org/10.5194/esd-2023-1-RC2 - AC2: 'Reply on RC2', Thanh Le, 26 Jun 2023
Viewed
HTML | XML | Total | Supplement | BibTeX | EndNote | |
---|---|---|---|---|---|---|
774 | 192 | 46 | 1,012 | 80 | 43 | 37 |
- HTML: 774
- PDF: 192
- XML: 46
- Total: 1,012
- Supplement: 80
- BibTeX: 43
- EndNote: 37
Viewed (geographical distribution)
Country | # | Views | % |
---|
Total: | 0 |
HTML: | 0 |
PDF: | 0 |
XML: | 0 |
- 1