This work investigates the association of Atlantic sea surface temperature anomalies (SSTA) with the early winter North Atlantic Oscillation (NAO). The NAO is represented by two indices in November, December, and January. One (regional mean index) is defined as 500 hPa geopotential height difference between areal averages in low and high latitudes, while another (station index) is the sea level pressure difference between two stations. Although the two indices are highly correlated, they exhibit clearly different statistical associations with the antecedent Atlantic SSTA. The regional mean index in early winter is significantly correlated with a tripole pattern of SSTA in the Atlantic up to the preceding spring. Therefore its predictability may be claimed for up to 7–9 months. However there is little predictability for the station index beyond SSTA 2 months before the early winter. The interseasonal potential predictability shown in the regional mean index mainly results from the contribution of a few years that are not strong anomaly years according to the station index. That suggests that the potential predictability displayed in association with the regional mean index mainly results from the average of a few anomaly years. The SSTA contribution to the NAO predictability comes mostly from the North Atlantic and northern subtropical Atlantic. The association of equatorial Atlantic SSTA with the NAO for both indices is insignificant. There is no linear connection between the tropical Atlantic SSTA and the tripole SSTA pattern, and the tropical and North Pacific SSTA may be a factor affecting NAO predictability.
If you can't find a tool you're looking for, please click the link at the top of the page to "Go to old article view". Alternatively, view our Knowledge Base articles for additional help. Your feedback is important to us, so please let us know if you have comments or ideas for improvement.
 The variability of the North Atlantic Oscillation (NAO) is traditionally represented by several indices, which are usually defined in terms of the pressure gradient between two regional means or between two stations in low and high latitudes over the North Atlantic basin, or as the time series of a related empirical orthogonal function (EOF) pattern of sea level pressure (SLP) or geopotential height at 500 hPa (H500) [Hurrell, 1995; Wallace, 2000; Li and Wang, 2003]. Some investigations have suggested that the low-frequency variation of the NAO may be partially forced by sea surface temperature (SST) anomaly (SSTA) [e.g., Rodwell et al., 1999]. In recent years, the influence of North Atlantic SSTs in preceding months on the winter NAO have been studied extensively [e.g., Czaja and Frankignoul, 1999, 2002; Saunders and Qian, 2002; Rodwell and Folland, 2002; Cassou et al., 2004; Fletcher and Saunders, 2006]. By applying a maximum covariance analysis (MCA) to SST and H500 of the National Center for Environmental Prediction (NCEP) and National Center for Atmospheric Research (NCAR) reanalysis data, Czaja and Frankignoul  (hereinafter referred to as CF2002) indicated that a Pan-Atlantic SST pattern precedes the NAO in early winter by up to 6 months and the bulk of the NAO signal comes from the midlatitudes of the North Atlantic Ocean. A so-called “North Atlantic horseshoe pattern,” with warm SST southeast of Newfoundland and cold SST to the northeast and southeast, precedes a positive phase of the NAO and provides a NAO prediction of up to 15% of its monthly variance several months in advance. Tropical Atlantic forcing of the NAO is nearly uncorrelated with and weaker than that resulting from the midlatitudes. It is associated with shorter lead times and reduced predictive skill.
 At the moment, there is no consensus about the origins of the forcing for NAO variability. Some model investigations have suggested that tropical forcing plays a dominant role [Latif et al., 2000; Hoerling et al., 2001; Peng et al., 2005] while others indicated that the bulk of the oceanic influence on the NAO comes from the local air-sea interaction [e.g., Rodwell et al., 1999]. By computing the cross-validated hindcast skill for predicting a range of winter NAO indices from lagged SSTA, Fletcher and Saunders  found that it is the Northern Hemisphere subpolar regions rather than the tropics which provide the best winter NAO seasonal predictability in lead of two to three seasons. Robertson et al.  argued that tropical and subtropical South Atlantic SSTA may also contribute to the NAO variation. On the other hand, Robertson  indicated that both the spatial pattern and power spectra of the winter Arctic Oscillation are not influenced by air-sea coupling in his simulations. Rodwell et al. , Watanabe and Kimoto [2000a], Pan , and Hu and Huang  suggested that there exist positive feedbacks between tripole SSTA and NAO. The atmospheric anomalies associated with the NAO produce tripole-like SSTA. In return, the tripole-like SSTA force NAO-like atmospheric anomalies [Pan, 2005]. It is clear that there are conflicting results regarding the role of extratropical North Atlantic SSTA in forcing North Atlantic climate at present [see Kushnir et al., 2002]. The combined response of tropical and extratropical SSTA becomes even more complicated.
 Because of these conflicting results regarding NAO predictability and the origins of the forcing for the NAO variability, it is necessary to reexamine the significance of the relationship between Atlantic SST and NAO with different approaches. In this work, we investigate the association of Atlantic SSTA with the NAO in early winter (November, December, and January (NDJ)) using two indices. One is referred to as the regional mean index, which is defined as the H500 difference between two areal averages in low and high latitudes. The other one is referred to as the station index, which is the difference between two stations in low and high latitudes. There are some similarities in the associations of the two indices with the antecedent SSTA, but the differences are distinct and interesting. The paper is organized as follows. In section 2, the data and the analysis strategy are briefly described. Section 3 presents the results. Section 4 contains a summary and further discussion of the results.
2. Data and Analysis Strategy
 Monthly mean SST and H500 from the NCEP/NCAR reanalysis [Kalnay et al., 1996] are analyzed. The reanalysis data span the period of January 1948 to April 2004 with a horizontal resolution of 1.875° × 1.875° for SST and of 2.5° × 2.5° for H500. Our analyses are mainly confined to 20°S–70°N, 100°W–20°E. Most of the analyses are carried out using the reanalysis data. No time filter or spatial smoothing is applied in the following calculations.
 We use two indices averaged in NDJ to represent the NAO variability in early winter (Figure 1). The reason for using the NDJ mean instead of the conventional winter mean of December, January, and February (DJF) is that NAO in NDJ has a more significant correlation with SSTA in previous months than that of NAO in DJF (see Figure 1 of CF2002). As we have mentioned above, the regional mean index (Figure 1a) is defined on the basis of the investigation of CF2002 as the difference of H500 between the areal average over (30°N–50°N, 100°W–20°E) and that over (50°N–70°N, 100°W–20°E). The station index (Figure 1b) is the SLP difference between two stations: Ponta Delgada, Azores and Stykkisholmur/Reykjavik, Iceland. The station index was calculated using the definition of Hurrell , given in http://www.cgd.ucar.edu/∼jhurrell/nao.stat.other.html. The skewness and kurtosis of the two time series shown in Figure 1 indicate that they obey normal distributions at the 99% significance level, according to the Lillifors test for normality [Thiēbaux, 1994]. These two indices are highly correlated with each other with a correlation coefficient of 0.9 (Figure 1).
 We then calculate the regression of the SST for three consecutive monthly means from the leading 12 months to the lagging 12 months onto the two indices in NDJ. The focus of the analysis is on the association of the SSTA up to 12 months preceding the NDJ NAO. Although the regression is done locally, what we are looking for is large-scale SST patterns that are connected with the NAO variation. The significance of the regression patterns is measured by the number of the significant regression grid points over the Atlantic Ocean basin (20°S–70°N, 80°W–20°E). Global SSTA composites for anomalous NAO years are also discussed to explain the difference in the association of SSTA with the two indices.
3.1. SSTA Linked to the Regional Mean NAO Index
Figure 2 shows the regressions of SSTA onto the normalized regional mean NAO index in early winter (NDJ) of Figure 1a from the index leading the SSTA 2 months (L = 2, Figure 2a) to the SSTA leading the index 8 months (L = −8, Figure 2f). The SSTA are three consecutive monthly mean departures from the monthly mean surface temperature of the reanalysis data. The effective sample size [Wilks, 1997], which is 19 for the regional mean NAO index, is used when testing the significance of the regressions. The sea ice grids are excluded in this and following regression calculations. The regression results are consistent with the SSTA composites (figure not shown) of the mean from the selected major NAO anomaly years with the regional mean index larger than one standard deviation (STDV) minus that for the index smaller than one negative STDV. In general, the regressions are negative in the high latitudes and the subtropics and positive in the middle latitudes around 30°–40°N, similar to the tripole pattern discussed in CF2002. However, comparing Figure 2 of CF2002 and Figure 2 in this work, there are obvious differences in the tropics. In Figure 2 of CF2002, there are clearly positive SSTA in the tropics (within 15° latitude from the equator) when SSTA is leading H500. In this work, some of the regressions are positive and some of them are negative, but almost no regression along the equator is significant when SSTA leads the regional mean NAO index (Figures 2c–2f). This is also different from some of the model results, which suggest that tropical SSTA play a dominant role in the low-frequency variation of the NAO [e.g., Latif et al., 2000; Robertson et al., 2000; Hoerling et al., 2001; Rodwell et al., 2004; Peng et al., 2005]. That may be because models overestimate the influence of the local air-sea interaction on the NAO due to large ensemble averages [Bretherton and Battisti, 2000]. The contrast relating to the impact of tropical SSTA on the NAO in this work and in CF2002 will be discussed further below.
 The statistical significance of the regression pattern of SSTA onto the normalized regional mean NAO index from the SSTA leading the index 12 months to the index leading the SSTA 12 months is measured by the ratio of the significant regression grid points to the total grid points in the Atlantic Ocean (20°S–70°N, 80°W–20°E) at different lags, which is shown as a percentage in Figure 3a. The corresponding criteria for the significance test of the regression at the 95% significance level was calculated using 1,000 Monte Carlo simulations [Livezey and Chen, 1983]. There are two steps to obtain the significance criteria. First, the above regression calculations are repeated 1,000 times by replacing the normalized regional mean NAO index by 1,000 time series randomly sampled from a normal distribution. In the second step, we count the number of grid points in each regression pattern which pass the 95% significance level. We then form a distribution of the number of significant grid points within the 1,000 samples. From this distribution we get the criteria for the 95% significance level of the regression pattern in the Atlantic Ocean region, which is plotted as the dashed line with dots in Figure 3a. The Monte Carlo test indicates that the ratio of significant regression grid points to total grid points must exceed 13% for a given regression pattern to be significant at the 95% level.
 The shaded regions seen in Figure 3a, which exceed the criteria (dashed line), are significant at the 95% level in the field test. According to the criteria, precursors of the early winter NAO variation can be found up to 9 months before in SSTA and the NAO is associated with SSTA up to 5 months later. The most significant regression pattern exists when the regional mean NAO index leads the SSTA by 2 months, different from one month lead indicated by CF2002 on the basis of MCA. However, both results suggest that the ocean responds to the NAO forcing most strongly in boreal early winter. The significance of the regressions then decreases monotonically, disappearing by the time the SSTA leads the regional mean NAO index by 5 months. The second maximum of significance occurs when the SSTA leads the regional mean NAO index by 8 months. The significance change is also visible from variation of the shaded area in Figure 2. The second peak extends potential predictability of the regional mean NAO index to about 7–9 month lead, which is 2 months longer than that obtained by CF2002. Using 10-year high-pass filter data, the recalculation shows a similar result, except that the peak in about 8 months in Figure 3a disappears and a new peak occurs at about 3 months (figures not shown). The high-pass filter results seem more consistent with CF2002 (see their Figure 1) as compared with the unfiltered results (Figures 2 and 3a). This may imply that a different mechanism is involved in different timescale phenomenon. The long-lead predictability may suggest the existence of air-sea interaction or oceanic influence in boreal spring that affects the evolution of the regional mean NAO index in the coming early winter. Robinson  indicated the possible linkage of SSTA in spring and the winter tripole SSTA pattern (see his Figure 2). Therefore the spring SSTA could be an oceanic response to the NAO in the previous winter. It is unclear why such a long memory exists in the ocean and how it is preserved.
3.2. SSTA Associated With the Station NAO Index
 Similarly, the results of the regressions of SSTA onto the normalized station NAO index of Figure 1b are shown in Figures 4 and 3b. The effective sample size [Wilks, 1997], which is 31 for the station NAO index, is used when testing the significance of the regressions. The regression results are also consistent with the corresponding SSTA composites (figure not shown), which are the mean when the index is larger than one STDV minus that when the index is smaller than one negative STDV. The regression pattern in Figure 4 is qualitatively similar to that in Figure 2. The regressions in the tropics are primarily negative and also insignificant. However, the differences between the regressions in Figures 2 and 4 are clear. The major difference is in the extratropical North Atlantic (Figures 2 and 4), resulting in the different features in Figures 3a and 3b. Comparing with Figure 2, it is seen that the number of significant regression grid points (the shaded area) in Figure 4 decreases more quickly with increasing lead time of SSTA to the NAO index.
 The significance of the regressions of SSTA on the normalized station NAO index is displayed in Figure 3b. The most significant regression exists when the station NAO index leads the SSTA by one month, which is more consistent with CF2002. The significance of the regression decreases almost monotonically with increasing lead time of the SSTA to the NAO index during the 12 month period. In contrast to Figure 3a, the second maximum of the significance does not exist in Figure 3b. The regression becomes insignificant after the SSTA leads the station NAO index by 2 months (Figure 3b). It is clear that in this case, the earlier SSTA provides little predictability for the station NAO index when SSTA leads by more than 2 months. This is consistent with the strong temporal stochastic nature of the NAO demonstrated by many previous investigations [e.g., Hurrell, 1995; Wunsch, 1999].
 It is clear that the significance level will be higher if only the North Atlantic region is examined. However, enlarging and reducing the field significance test region will only change the amplitude, but not the main features of the curves in Figures 3a and 3b. The main features of the curves are confirmed by the evolution of significant regression areas shown in Figures 2 and 4.
3.3. Possible Reasons for the Differences
 It is interesting that despite the strong correlation between the two indices, the regional mean index seems to have much higher predictability than the station index. In order to understand the reasons for the difference, we further examined the major NAO events based on these two indices. The years for which the indices are larger than one STDV or less than one negative STDV are listed in Table 1. It is seen from Table 1 that most of the anomaly years are the same for the two indices with just a few years different. Figure 5 shows the SSTA composites for the common anomaly years of the two indices, which displays a pattern more similar to Figure 4 than to Figure 2. This probably indicates that for most of the years, the NAO predictability estimated on the basis of the station index is more accurate than that on the basis of the regional mean index. The different predictability likely comes from those years when the two indices disagree. One possibility is that the difference results from those years when both indices are small. However, we find that the composite results using different anomaly thresholds, for example, 0.5 STDV and 0.75 STDV, are consistent with that using one STDV. That implies a lack of sensitivity of the results to the criteria used to choose the anomalous years. The regression onto both indices (Figure 6), calculated similarly to Figures 2 and 4, but excluding the noncommon anomaly years (1953, 1960, 1965, 1969, 1972, 1978, 1981, 1983, 1986) (see Table 1), shows high consistency for the regression pattern of the two indices, which is similar to Figure 5. This suggests that the different predictability results from the anomaly years that are not common between the two indices.
Table 1. Years for the Regional Mean NAO Index and the Station NAO Indexa
Regional Mean Index
Years in bold are larger than 1 standard deviation. All other years are smaller than 1 negative standard deviation.
Figure 7 shows the composites for the years not in common in Table 1. The contrast between the composites in the left and right panels is clear in both the Pacific and Atlantic Oceans. For the composite of the mean of 1983 and 1986 minus that of 1969, which are anomaly years identified from the regional mean index (Figure 1a), the large SSTA with the tripole pattern in the Atlantic is intensified in the preceding summer and spring (Figures 7c and 7e). Meanwhile, the negative SSTA in the North Pacific is pronounced and the SSTA is small in the tropical Pacific (Figures 7c and 7e). In contrast, the SSTA composites of the mean of 1953, 1960, and 1972 minus the mean of 1965, 1978, and 1981 (Figures 7b, 7d, and 7f), which are anomaly years identified from the station index (Figure 1b), display almost an opposite pattern in the North Pacific. It should be pointed out that the contour interval in Figures 7a, 7c, and 7e is two times that in Figures 7b, 7d, and 7f, considering the different sample sizes. Clearly positive SSTA exist in the tropical and the North Pacific. It is also noted that the tripole mode in the North Atlantic is replaced by a persistent positive SSTA in the central part of the North Atlantic (Figures 7b, 7d, and 7f). The negative SSTA to the north and south of the positive SSTA region are almost gone (Figures 7b, 7d, and 7f). The above analyses suggest that the higher potential predictability of the NAO with 8 months lead time in the regional mean index probably results from only a few years. It should be pointed out that almost all the composites in Figure 7 are not significant at the 90% level because of small sample size.
 It is noted from Table 1 and Figure 7 that two of the three years (i. e., 1969 and 1983) used for the composite in Figures 7a, 7c, and 6e accompany the NAO of the same polarity in preceding years while none accompanies such a consecutive event for the composite in Figures 7b, 7d, and 7f. This suggests that the reemergence mechanism may play a role in the multiyear persistence of SST anomalies in the North Atlantic winter [Alexander et al., 1998; Bhatt et al., 1998; Watanabe and Kimoto, 2000b; Timlin et al., 2002; Deser et al., 2003]. The reemergence mechanism emphasizes the recurrence of SSTA in next winter. Late winter ocean temperature anomalies are sequestered beneath the shallow summer mixed layer and are reincorporated into the deepening fall mixed layer. The enhanced mixing causes the subsurface anomaly to reappear in SST in the next winter [Bhatt et al., 1998]. The oceanic anomalies generated in winter are “frozen” in permanent thermocline in summer when a new seasonal thermocline developed during this period. The subsurface anomalies persistent from winter are decoupled from the atmosphere until the next fall. However, it is necessary to further investigate the influence of a consecutive event on NAO predictability.
Figure 5 shows that apart from Atlantic SSTA, a strong NAO case is preceded by warm SST in the eastern equatorial Pacific from spring to summer and the North Pacific is generally cold. The composites in Figure 7 suggest that SSTA in the Pacific may affect the NAO predictability. It is unclear how significant the linkage of the tropical and North Pacific SSTA is with the NAO predictability because of the small composite sample size. It is still unknown if the Pacific SSTA are part of a global anomaly pattern, including the SSTA in the Atlantic that are connected with the NAO.
3.4. Role of Tropical SSTA
 One of the differences between this work and CF2002 is the role of tropical Atlantic SSTA. CF2002 found that positive SSTA in the tropical Atlantic favors the negative phase of the NAO, although the contribution of the tropical Atlantic forcing is weaker than that from the midlatitude Atlantic Ocean. In this work, both the regional mean and station NAO indices show insignificant regressions with the tropical Atlantic SSTA.
 Since both CF2002 and this work use the same reanalysis data for largely the same period, the difference between the two most likely results from the different methods of analyses used in the investigations. To verify the connection of SSTA among different regions, we calculate the correlations between some regional mean SSTA and SSTA in the Atlantic. We use correlation to demonstrate the connection due to the fact that it can be more directly and more clearly revealed by straightforward regression or correlation than by some other techniques, such as EOF (rotated or not), as indicated by Dommenget and Latif . Figure 8 shows the correlations between the SSTA in the Atlantic and regional mean SSTA using the reconstructed data [Reynolds and Smith, 1994]. The tropical mean SST (Figure 8a) is the average of SST in 15S°–5°N, 30°W–10°E. This region is chosen according to Penland and Matrosova . The other three regions adopted in Figures 8b, 8c, and 8d are based on Figures 2 and 4 of this work and Figure 2 of CF2002. The correlations between the SSTA and a few combinations of four regional mean SSTA are also conducted. Additional calculations show that results are not sensitive to details of the chosen regions.
Figure 8a shows that the tropical SSTA does not significantly correlate with extratropical SSTA in the North Atlantic Ocean. The absence of the connection between the tropical and extratropical SSTA is also confirmed by the correlations displayed in Figures 8b–8d. Figures 8b and 8d show that the SSTA in the subtropics and in the high latitudes are highly correlated. There is a marginally significant correlation between the SSTA in the middle latitudes (around 30°–40°N) and that in the high latitudes or in the subtropics (Figures 8b–8d). The tripole mode (Figure 8e), which is similar to that in Figures 2 and 4 and the results of CF2002, is evident in the correlations between the SSTA in the Atlantic and the combination of the three regional mean SSTA of Figures 8b–8d. However, if the combination is extended to include the uncorrelated tropical SSTA, the correlations may show a false connection between the tripole mode and the tropical SSTA (Figure 8f).
 It is also noted from Figure 8 that there are no significant correlations between SSTA in tropical North and South Atlantic. This is consistent with Houghton and Tourre , Enfield and Mayer , and Mehta , who found no interhemispheric coherence variation. However, some studies showed the interhemispheric SST correlation may change sign according to timescales. SST anomalies tend to be negatively (positively) correlated between the tropical North and South Atlantic on the decadal (interannual) timescale [Tanimoto and Xie, 1999; Enfield et al., 1999].
4. Summary and Discussion
 This work investigates the association of Atlantic SSTA with the early winter NAO. In this study, the NAO is represented by two indices in November, December, and January, that are based on previous studies. One is defined as 500 hPa geopotential height difference between areal averages in low and high latitudes, while another is the sea level pressure difference between two stations: Ponta Delgada, Azores and Stykkisholmur/Reykjavik, Iceland, that is traditionally used to characterize the NAO behavior. These two indices are referred to as the regional mean and station indices, respectively. Although the two indices are highly correlated, they exhibit clearly different statistical associations with the antecedent Atlantic SSTA. The regional mean NAO index in early winter is significantly correlated with a tripole pattern of SSTA in the Atlantic up to the preceding spring. Therefore its predictability may be claimed for up to 7–9 months. However, there is little predictability for the station NAO index beyond SSTA two months before the early winter. Further composite analysis based on the anomaly years of the two indices suggests that the strong potential predictability of the NAO shown in the regional mean index at about 8 months lead mainly results from the contribution of a few years that are not strong anomaly years according to the station index. Therefore little predictability in interseasonal timescales is suggested for most of the NAO anomaly years and the potential predictability displayed in association with the regional mean index mainly results from the average of a few anomaly years. We also find that the SSTA contribution to the NAO predictability comes primarily from the North Atlantic and northern subtropical Atlantic. The association of equatorial Atlantic SSTA with the NAO for both indices is insignificant. There is no linear connection between the tropical Atlantic SSTA and the tripole SSTA pattern. Our analysis also shows that the tropical and North Pacific SSTA may be a factor affecting NAO predictability. On the average, cooling (warming) in the North Pacific may enhance (reduce) the NAO predictability.
 Although it has been found that the tropical and North Pacific SSTA may play a role in the variability and predictability of the Atlantic climate [Enfield and Mayer, 1997; Huang et al., 2002], it is unclear how the Pacific influences the NAO. There are two possibilities. One is that the impact is realized first by influencing SST in the North Atlantic Ocean, which then affects the NAO. The other possibility is that the tropical and North Pacific SSTA affect the NAO directly through the atmospheric circulation. By reexaming Figure 7, we found that the composite SSTA in the North Atlantic displays a pattern with some similarities to the regression pattern in Figures 2 and 4. However, the difference in the North Atlantic is also clear between the composites and the regressions, which may imply that the two possibilities cannot be ruled out.
 It is still open how to explain the different features of the two indices in association with the preceding SST. One possible reason is that the station NAO index does not take into account the change in position of the center of action of the NAO. The change of position is evident in both observations and in some increased greenhouse gas concentration simulations [e.g., Glowienka-Hense, 1990; Ulbrich and Christoph, 1999; Hilmer and Jung, 2000; Hu and Wu, 2003]. Broad-scale features of the NAO in the Atlantic sector may be better represented by the regional mean index than by the station index. Frederiksen and Zheng  found that the NAO-related spatial patterns are varied with timescales, which are associated with different physical processes. In addition, there are many other definitions of the NAO index besides the regional mean index and the station index that were used in this manuscript, such as principal component (PC) of the first EOF of SLP in the Atlantic area; PC of a rotational EOF of the 500 mb height. Even with the station-based index there are different choices of the southern point.
 The NAO variability in early winter is represented by the mean of November, December, and January in this work. Since the centers of the NAO show a strong seasonality in both their locations and intensities in observations [e.g., Portis et al., 2001], the seasonality of impact of the Atlantic SSTA on the NAO is unclear and another interesting research topic in the future. In addition, further work is required to understand the role of local air-sea interaction in the evolution of the NAO and the different features of the two indices in their connection with SSTA in antecedent seasons. Unfortunately, like the models used by Rodwell et al. , our coupled GCM at present cannot well reproduce the observed connection between the NAO in early winter and SSTA in antecedent seasons.
 The authors thank the editor R. Lieberman and anonymous reviewers for their suggestions and comments to significantly improve the manuscript. We are grateful for the advice and comments of J. Kinter III, L. Bengtsson, and D. Straus. The manuscript was kindly edited by K. Pegion. The discussion with R. Wu was very helpful for doing the normality test. This work was supported by the NOAA CLIVAR Atlantic Program (NA04OAR4310115).