Abstract

In this study, the correlation between the Northern Hemisphere winter Pacific and Atlantic storm tracks is examined using the NCEP–NCAR reanalysis and the 40-yr ECMWF Re-Analysis (ERA-40), as well as unassimilated aircraft observations. By examining month-to-month variability in the 250-hPa meridional velocity variance, the correlation between the two storm track peaks is found to be as high as 0.5 during the winters between 1975/76 and 1998/99. Here, it is shown that the correlation between the two storm tracks can be clearly detected from the aircraft data. Further analyses of the reanalysis data show that the correlation can also be seen in other eddy variance and covariance statistics, including the poleward heat flux at the 700-hPa level.

The correlation between the two storm tracks, as seen in both reanalysis datasets, is shown to be much weaker during the period 1957/58–1971/72, suggesting a possible regime transition from largely uncorrelated storm tracks to highly correlated storm tracks during the 1970s. However, during this earlier period, the number of aircraft observations is insufficient to verify the low correlation seen in the reanalyses. Thus, low biases in the reanalyses during the earlier period cannot be ruled out.

An ensemble of four GCM simulations performed using the GFDL GCM forced by global observed SST variations between 1950 and 1995 has also been examined. The correlation between the two storm tracks in the GCM simulations is much lower (0.18) than that observed, even if the analysis is restricted to the GCM simulations from the period 1975/76–1994/95. A Monte Carlo test shows that the observed correlation and the GCM correlation are statistically distinct at the 1% level.

Correlations between the Southern Hemisphere summer Pacific and Atlantic storm tracks have also been examined based on the reanalyses datasets. The results suggest that the amplitude of the SH summer Pacific and Atlantic storm tracks are not significantly correlated, showing that seeding of the Atlantic storm track by the Pacific storm track does not necessarily lead to significant correlations between the two storm tracks.

1. Introduction

In the midlatitudes, during the cool season, changes in the weather are dominated by passages of cyclones and anticyclones. Apart from affecting weather on the earth's surface, these storms are usually the surface manifestation of deep baroclinic waves that frequently extend into the lower stratosphere and participate in the large-scale transport of energy, momentum, and moisture. In the Northern Hemisphere (NH), cyclone occurrences are maximal over the midlatitude ocean basins (e.g., Pettersen 1956; Whitaker and Horn 1984). Blackmon (1976) noticed that the band of maximal bandpass transient eddy variance was more or less collocated with the region of maximum cyclone activity, and was first to use the term “storm tracks” to describe the geographically localized maxima in bandpass transient variance. In this paper, we will follow this definition and use eddy variance/covariance statistics to indicate storm track activity.

With their strong links to weather and global transports, changes in storm tracks certainly form an important component of climate variability. Until recently, storm track variability has not been extensively studied, one possible contributing reason being the lack of consistent datasets of sufficient duration to make such analyses physically meaningful. However, during the past decade, with the release of the 50-yr National Centers for Environmental Prediction–National Center of Atmospheric Research (NCEP–NCAR) reanalysis data (Kalnay et al. 1996; Kistler et al. 2001) and, more recently, the 40-yr European Centre for Medium-Range Weather Forecasts (ECMWF) Re-Analysis (ERA-40), climate researchers now have easy access to high-quality gridded analyses that are produced by consistent data assimilation systems. Thus, a large number of studies on storm track variability have been carried out during the past few years (for a partial list, please refer to Chang et al. 2002), most of which are based on the NCEP– NCAR reanalysis data.

Apart from documenting and understanding the role of storm track variability in weather and climate changes, a further motivation for studying storm track variability is to generate a baseline for the validation of GCM simulations. Since the baroclinic waves in the storm tracks play such a crucial role in atmospheric transports, changes in eddy forcing due to changes in the location or amplitude of storm tracks can have significant impacts on the large-scale circulation (e.g., Held et al. 1989; Branstator 1992; Peng and Whitaker 1999). Hence, the ability of GCMs to correctly simulate both the climatology and variability of storm tracks should be an important requirement for model predictions of climate changes to be considered credible. During the past few years, a number of studies have been conducted to validate this aspect of GCM simulations (e.g., Christoph et al. 1997; Carillo et al. 2000).

Most past studies of storm track variability have treated the Pacific and Atlantic storm tracks as separate entities and analyzed their variability separately (e.g., Lau 1988; Carillo et al. 2000; Graham and Diaz 2001; Geng and Sugi 2001). Recently, Chang and Fu (2002) examined the variability of the NH winter storm tracks by performing an EOF analysis on the December–January– February (DJF) mean hemispheric 300-hPa bandpass-filtered meridional velocity variance based on NCEP– NCAR reanalysis data and found that the leading EOF has a hemispheric structure and represents the simultaneous strengthening/weakening of both the Atlantic and Pacific storm tracks. They also showed that the two storm tracks are significantly correlated, even when month-to-month variations are considered. However, Harnik and Chang (2003) compared variances computed based on unassimilated radiosonde observations to those computed based on the NCEP–NCAR reanalysis data over the storm track entrance and exit regions and found that the correlation between the Pacific and Atlantic storm track entrances and exits is weaker in the radiosonde observations than in the reanalysis, raising the possibility that the correlation between the two storm tracks might have been spuriously increased by the reanalysis process. Unfortunately, there are very few radiosonde observations near the storm track peaks where Chang and Fu (2002) found the strongest correlations.

While there are very few radiosonde observations over the oceans, since the mid- to late 1960s, there have been frequent aircraft observations over the NH midlatitude oceans close to the storm track peaks. In this study, we will examine aircraft observations to see whether the correlation between the two storm tracks can be observed from unassimilated aircraft observations. In section 2, we will show that the signal can indeed be found in aircraft observations. In section 3, we will examine how this correlation shows up in different eddy statistics, and in section 4, some issues raised by this observed correlation will be discussed. Section 5 contains the conclusions.

2. Comparison of VV2501df with aircraft data

The main data analyzed in this study are the reanalysis data produced by the 50-yr NCEP–NCAR reanalysis project (Kistler et al. 2001) and ERA-40. Harnik and Chang (2003) compared variances computed from rawinsonde observations over NH land areas to those computed from the NCEP–NCAR reanalysis data and found that there appears to be substantial biases in the NCEP–NCAR-based variance statistics prior to the mid-1970s; hence in this section we will focus on the results based on data starting from 1975.

To validate the results based on the reanalysis datasets, we computed variances from unassimilated aircraft observations over the ocean basins. As part of the NCEP–NCAR reanalysis project, all observations used in the reanalysis have been archived (Kalnay et al. 1996). Most of the aircraft observations near the storm track maxima are taken between the 200- and 300-hPa level, hence most of the comparisons shown in this section will be at the 250-hPa level. All aircraft report by pilot (AIREP) and aircraft-to-satellite data relay (ASDAR) reports have been processed. Chang (2003) analyzed this dataset and showed that the midwinter minimum in the Pacific storm track can be verified based on unassimilated aircraft observations. [Please refer to Chang (2003) for more details concerning this dataset.] In this study, we will use this dataset to see whether we can verify the correlation between the Pacific and Atlantic storm tracks found in the NCEP–NCAR reanalysis data by Chang and Fu (2002).

Let us first examine the variability of the storm track activity near the tropopause. In this paper, storm track activity is defined by 24-h difference-filtered (Wallace et al. 1988) 250-hPa meridional velocity variance:

 
formula

The overbar in (1) represents monthly mean. This filter is chosen because of its ease of comparison with data sources with frequent data gaps (Chang and Fu 2002; Harnik and Chang 2003). As shown in Wallace et al. (1988), the peak response for the filter is at a period of 2 days, with half-power points at 1.2 and 6 days. Chang and Fu (2002) examined different variance and covariance quantities computed using different filters and found that they show very similar variability. Note that the peak response of the filter is 2; hence, variance and covariance statistics computed using the 24-h difference filter should be divided by 4 before quantitative comparisons to results from other filters can be made.

The climatological winter mean (DJF of 1975/76– 1998/99) distribution of VV2501df is shown in Fig. 1a and its month-to-month variability (standard deviation) is shown in Fig. 1b. In Fig. 1, we can see the two storm tracks peak over the eastern Pacific and central Atlantic, with maximum month-to-month variability over the eastern Pacific and the eastern Atlantic. As discussed in Chang and Fu (2002), when an EOF analysis (with area-weighted covariance) is performed on the monthly anomalies, the leading EOF corresponds to the simultaneous strengthening/weakening of both storm tracks. The spatial structure of this leading EOF (which accounts for 18.6% of the month-to-month hemispheric variance in VV2501df) is shown in Fig. 2. In Fig. 2a, the correlation between the leading principal component (PC1) with the monthly anomaly is shown, and the regression with PC1 is shown in Fig. 2b. Clearly, this EOF has a structure that straddles both storm tracks.

Fig. 1.

(a) DJF mean of VV2501df, based on NCEP–NCAR reanalysis from the winters of 1975/76–1998/99. Contour interval is 50 m2 s−2. (b) Std dev in monthly mean VV2501df for the same period. Contour interval is 20 m2 s−2

Fig. 1.

(a) DJF mean of VV2501df, based on NCEP–NCAR reanalysis from the winters of 1975/76–1998/99. Contour interval is 50 m2 s−2. (b) Std dev in monthly mean VV2501df for the same period. Contour interval is 20 m2 s−2

Fig. 2.

(a) Correlation between PC1 and VV2501df. Contour interval is 0.1. (b) Regression of VV2501df based on PC1. Contour interval is 20 m2 s−2. In both (a) and (b), regions where the correlation between PC1 and VV2501df are over 0.25 (95% significant for 72 months) have been shaded

Fig. 2.

(a) Correlation between PC1 and VV2501df. Contour interval is 0.1. (b) Regression of VV2501df based on PC1. Contour interval is 20 m2 s−2. In both (a) and (b), regions where the correlation between PC1 and VV2501df are over 0.25 (95% significant for 72 months) have been shaded

It is well known that since EOF analysis is designed to maximize the amount of explained variance, the appearance of a hemispheric pattern does not necessarily imply that the entire structure is physically coherent (e.g., Dommenget and Latif 2002). Examples can be found in which the different maxima/minima in the pattern do not correlate significantly with each other. To check that the two storm tracks are in fact correlated, two storm track indices are defined based on the position of the two peaks in the pattern shown in Fig. 2. A Pacific (PAC) index is defined by the area average of VV2501df over the area 40°–50°N, 180°–140°W, and an Atlantic (ATL) index is defined by the average over 42.5°– 52.5°N, 50°–10°W. The resultant time series have a correlation of 0.50, which is highly significant.1 We have also defined indices based on larger areas (35°–55°N, 170°E–130°W and 37.5°–57.5°N, 60°W–0°), and the resultant correlation is only slightly smaller (0.46). The correlation and regression of these two indices with the monthly VV2501df anomalies are shown in Fig. 3. In Fig. 3, regions where the correlation is larger than 0.25 are shaded. Clearly, both indices are well correlated with the storm track variations in the other ocean basin. A Monte Carlo test of field significance has been performed (Zwiers 1987; see also Livezey 1995), and the results show that the interbasin correlation patterns shown in Fig. 3 are statistically significant. More details of the field significance test are presented in the appendix.

Fig. 3.

(a) Correlation between the PAC index and VV2501df and (b) between the ATL index and VV2501df. (c) Regression of VV2501df based on the PAC index and (d) based on the ATL index. Contour intervals are 0.1 in (a) and (b) and 20 m2 s−2 in (c) and (d). Regions in (a)–(d) where the correlation between the corresponding index and VV2501df are over 0.25 have been shaded. In (c) and (d) (and all regression maps shown later), the amount of variances in VV2501df explained by the regression over the Pacific sector (30°–65°N, 170°E– 130°W) is printed in the lower-left corner, and that over the Atlantic sector (30°–65°N, 60°W–0°) is printed in the lower-right corner. Only the sector between 140° and 10°E is shown

Fig. 3.

(a) Correlation between the PAC index and VV2501df and (b) between the ATL index and VV2501df. (c) Regression of VV2501df based on the PAC index and (d) based on the ATL index. Contour intervals are 0.1 in (a) and (b) and 20 m2 s−2 in (c) and (d). Regions in (a)–(d) where the correlation between the corresponding index and VV2501df are over 0.25 have been shaded. In (c) and (d) (and all regression maps shown later), the amount of variances in VV2501df explained by the regression over the Pacific sector (30°–65°N, 170°E– 130°W) is printed in the lower-left corner, and that over the Atlantic sector (30°–65°N, 60°W–0°) is printed in the lower-right corner. Only the sector between 140° and 10°E is shown

In Figs. 3c,d, the amount of explained variance accounted for by the regressions, in the two sectors 170°E– 130°W and 60°W–0°, between 30° and 65°N, are displayed. The regression based on the PAC index accounts for 40.6% of the variance in the Pacific sector and 8.2% in the Atlantic sector, while the regression based on the ATL index accounts for 30.1% of the variance in the Atlantic sector and 10.4% in the Pacific sector. So while the correlation between the two storm tracks is quite significant, the amount of storm track variations “predictable” from the other storm track is not that large. Statistics similar to those shown in Figs. 1–3 have been computed based on ERA-40 data—the results are similar and will not be shown here.

While the analyses based on the two reanalysis datasets suggest that the two storm tracks are significantly correlated, we would like to verify this correlation using observational data. Harnik and Chang (2003) examined rawinsonde data over the storm track entrance and exit regions. They showed that the correlation between the two storm tracks over those regions, as computed based on unassimilated rawinsonde data, is weaker than that seen in the NCEP–NCAR reanalysis data. This raises the question of whether the correlation between the two storm tracks has been spuriously increased by the reanalysis process. However, there are no rawinsonde stations near the storm track peaks, thus necessitating the use of other data sources.

In this paper, we will use unassimilated aircraft observations to compute variance to compare with those computed based on the reanalysis datasets. As in Chang (2003), 6-hourly 5° × 5° aircraft observation grids are generated by averaging all2 observations between 200 and 300 hPa, within a 5° × 5° grid box, and within 3 h of each 6-hourly synoptic hour. The spatial distribution of the observations has been shown in Chang (2003; can also be inferred from Fig. 5; see discussions below). The average monthly mean number of observations in each 5° × 5° grid box in the central Pacific (40°–50°N, 170°E–150°W) is shown in Fig. 4a by the solid line, and that in the central Atlantic (40°–50°N, 50°–10°W) is shown by the dashed line. We can see that there were few observations prior to 1965, with the number of observations staying relatively constant (especially over the Pacific) between the 1970s and early 1980s, apart from the peak around 1979 [the First Global Atmospheric Research Programme (GARP) Global Experiment (FGGE) year], and then taking off after the mid-1980s. In Fig. 4b, the average number of 6-hourly periods per 5° × 5° grid box per month with one or more observations is shown (the maximum possible value is 120). We can see that over the central Pacific, the average number hovers around 40 since 1967/68, and over the Atlantic, the value has stayed around 70 since 1970/71. Figure 4 shows that there should be a sufficient amount of data to be used to verify the correlation between the two storm tracks.

Fig. 5.

(a) Regression of VV2501df (in m2 s−2) computed using unassimilated aircraft observations based on the PAC index. (b) Same as in (a), but based on the ATL index. All grid boxes over which the correlation between VV2501df and the corresponding index is above 0.31 are shaded. The regression is only defined for grid boxes over which there are more than 15 Dec, 15 Jan, and 15 Feb with over 20 data pairs within the grid box

Fig. 5.

(a) Regression of VV2501df (in m2 s−2) computed using unassimilated aircraft observations based on the PAC index. (b) Same as in (a), but based on the ATL index. All grid boxes over which the correlation between VV2501df and the corresponding index is above 0.31 are shaded. The regression is only defined for grid boxes over which there are more than 15 Dec, 15 Jan, and 15 Feb with over 20 data pairs within the grid box

Fig. 4.

(a) Average number of aircraft observations per DJF month between 200 and 300 hPa per 5° × 5° grid box over the central Pacific (40°–50°N, 170°E–150°W; shown by solid line) and the central Atlantic (40°–50°N, 50°–10°W; shown by dashed line). (b) Same as in (a), except for the average number of 6-hourly periods per grid box per month that has observations

Fig. 4.

(a) Average number of aircraft observations per DJF month between 200 and 300 hPa per 5° × 5° grid box over the central Pacific (40°–50°N, 170°E–150°W; shown by solid line) and the central Atlantic (40°–50°N, 50°–10°W; shown by dashed line). (b) Same as in (a), except for the average number of 6-hourly periods per grid box per month that has observations

For each grid box, monthly variances are computed using aircraft observations based on (1), averaging only over those times when observation pairs are available. To limit the effect of possible biases introduced by data gaps (e.g., Kidson and Trenberth 1988; Chang 2003) and changes in the spatial distribution of the observations in time, variances are only defined over a grid box when the number of data pairs in a month exceeds 20, and the grid box is only used if the number of DJF months with greater than 20 data pairs exceeds 15 for each calendar month (i.e., all data points shown with numerical value in Fig. 5 have at least 45 months of defined variance). This dataset is referred to as AIROBS below. To estimate the error introduced by the uneven distribution of aircraft observations, a “simulated aircraft observations” dataset is generated based on the reanalysis data. This dataset consists of sampling the reanalysis data only when aircraft observations exist within the grid box (referred to as RSAMP below). Statistics computed based on the AIROBS and RSAMP datasets are compared to those computed based on the full reanalysis dataset (REAN).

To compare with the reanalysis data, the correlation and regression of AIROBS with the PAC and ATL indices defined above are computed, and the results are shown in Fig. 5. The regressed values are shown by numbers and grid boxes where the correlation is larger than 0.31 (95% significant limit for 42 DOFs) are shaded. It is clear that the PAC index correlates significantly with much of the Atlantic storm track as defined by AIROBS, and the ATL index correlates well with the Pacific storm track over the central Pacific. The regressed values compare favorably with that shown in Figs. 3c,d computed based on REAN. In Table 1, the regressed values averaged over the central Pacific (40°– 50°N, 170°E–150°W) and central Atlantic (40°–50°N, 50°–10°W) for all three datasets are displayed. We see that the strength of the signal computed based on AIROBS is very close to that computed based on RSAMP and REAN, suggesting that there are no significant biases between the reanalysis and aircraft observation datasets.

Table 1.

Area average of regression based on the PAC and ATL indices in m2 s−2 for the Pacific means (40°–50°N, 180°–140°W) and Atlantic means (40°–50°N, 50°–10°W)

Area average of regression based on the PAC and ATL indices in m2 s−2 for the Pacific means (40°–50°N, 180°–140°W) and Atlantic means (40°–50°N, 50°–10°W)
Area average of regression based on the PAC and ATL indices in m2 s−2 for the Pacific means (40°–50°N, 180°–140°W) and Atlantic means (40°–50°N, 50°–10°W)

The month-to-month variations in area-averaged VV2501df, computed based on AIROBS and REAN, are shown in Fig. 6. We can see that the time series are well correlated with each other—the correlation between AIROBS and REAN for the Pacific time series is 0.93, and the correlation for the Atlantic time series is 0.97 (correlations between AIROBS and RSAMP are even higher). The two time series computed based on AIROBS have a correlation of 0.39—this value, while still significant at the 99% level, is quite a bit smaller than that based on REAN (0.48). However, the same correlation based on RSAMP is 0.41, suggesting that much of the reduction in the correlation between the two storm tracks seen in AIROBS is probably due to the effects of data gaps. Thus, we can conclude that the significant correlation between the Pacific and Atlantic storm tracks can be detected in unassimilated aircraft observations.

Fig. 6.

(a) VV2501df (in m2 s−2) monthly anomalies, averaged over the area 40°–50°N, 170°E–150°W, based on NCEP–NCAR reanalysis data (solid line) and aircraft observations (dots). (b) Same as in (a), but for the area 40°–50°N, 50°–10°W

Fig. 6.

(a) VV2501df (in m2 s−2) monthly anomalies, averaged over the area 40°–50°N, 170°E–150°W, based on NCEP–NCAR reanalysis data (solid line) and aircraft observations (dots). (b) Same as in (a), but for the area 40°–50°N, 50°–10°W

3. Correlation/regression for quantities other than VV2501df

In section 2, we focused on meridional wind variance at 250 hPa because there are observations at that level with which to validate the reanalysis. However, storm track activity can be defined based on other quantities, such as low-level meridional heat flux (e.g., Nakamura et al. 2002) or surface pressure/geopotential height anomalies (e.g., Wallace et al. 1988). In this section we will examine other storm track statistics to see how the correlation between Pacific and Atlantic storm tracks manifest themselves in these other quantities.

Both NCEP–NCAR reanalysis and ERA-40 data for the period 1975/76–1998/99 have been analyzed, and only the results based on the NCEP–NCAR reanalysis will be shown here since the results are very similar. In Fig. 7, regression maps of 24-h filtered poleward flux of zonal momentum (uυ′) at the 250-hPa level, poleward flux of temperature (υT′) at the 700-hPa level, and 1000-hPa geopotential height variance, based on the PAC (left) and ATL (right) indices defined in section 2 above, are shown. Again, the amount of variance accounted for by the regression in the Pacific and Atlantic sectors is printed on the bottom left and right corners of each panel. In all the panels, areas over which the correlation is larger than 0.25 are shaded.

Fig. 7.

(a) Regression of 24-h filtered eddy poleward flux of zonal momentum (uυ′; in m2 s−2) based on the PAC index. (b) Same as in (a), but based on the ATL index. (c), (d) Same as in (a), (b), but for the poleward flux of temperature (υT′; in K m s−1). (e), (f) Same as in (a), (b), but for variance of 1000-hPa geopotential height (in m2). In (a)–(f), shaded areas represent correlations between the regressed quantity and the reference index is over 0.25. As in Fig. 3, the amount of variance explained by the regression over the Pacific and Atlantic is printed in each panel

Fig. 7.

(a) Regression of 24-h filtered eddy poleward flux of zonal momentum (uυ′; in m2 s−2) based on the PAC index. (b) Same as in (a), but based on the ATL index. (c), (d) Same as in (a), (b), but for the poleward flux of temperature (υT′; in K m s−1). (e), (f) Same as in (a), (b), but for variance of 1000-hPa geopotential height (in m2). In (a)–(f), shaded areas represent correlations between the regressed quantity and the reference index is over 0.25. As in Fig. 3, the amount of variance explained by the regression over the Pacific and Atlantic is printed in each panel

First, let us examine the results for eddy momentum flux at the 250-hPa level (Figs. 7a,b). The regression in the same sector clearly shows a strengthening of the momentum flux dipole [for climatological structure of uυ, see, e.g., Blackmon et al. (1977)] toward the main body and downstream end of both storm tracks. The cross correlation between the two storm tracks is quite weak in this statistic, with only a weak negative center north of each storm track being statistically significant. This is not altogether surprising since eddy momentum flux depends quite strongly on details in the eddy and flow structure; hence, its correlation with storm track amplitude alone (which is basically what the PAC and ATL indices represent) may not be that strong.

The regression for the 700-hPa poleward heat flux (VT7001df) is shown in Figs. 7c,d. Corresponding to an increase in activity of either storm track (in terms of VV2501df), we can see a strong increase in the eddy heat flux, as well as a statistically significant increase in the poleward heat flux in the other ocean basin. Out of the three quantities shown in Fig. 7, the meridional heat flux is the one most strongly coupled to VV2501df. Further analysis of this quantity will be presented below.

The regression with the 1000-hPa geopotential height variance (ZZ10001df) is shown in Figs. 7e,f. Corresponding to the increase in upper-level eddy variance, surface geopotential variance also increases in both storm tracks. However, the relation is weaker than that for the 700-hPa heat flux, probably because surface pressure/height variations depend more on variations in environmental parameters such as stability, SST, and ocean frontal structures than mid- and upper-level variance and covariance statistics. Some cross–storm track correlation can be seen in Figs. 7e,f, but the correlation between the Atlantic storm track with the PAC index is especially weak.

To further examine the cross correlation between the two storm tracks in terms of lower-tropospheric quantities, we define alternative storm track indices based on the 700-hPa poleward heat flux. PACvT is defined as the average of VT7001df over the area 40°–50°N, 150°E–170°W, and ATLvT is the average over the area 45°–55°N, 50°–10°W. The two time series have a correlation of 0.39—not quite as high as that between PAC and ATL but still statistically significant at the 99% level. The correlation and regression maps based on these two indices are shown in Fig. 8. In Figs. 8a,b, the results for VV2501df are shown. These should be compared to Figs. 3c,d. The results for VT7001df and ZZ10001df are shown in Figs. 8c–f (these should be compared to Figs. 7c–f). In all of these figures, a statistically significant cross correlation between the two storm tracks can be clearly seen. In particular, the correlation between PACvT and ZZ10001df over the Atlantic is stronger than that for PAC shown in Fig. 7e. Thus, Fig. 8 shows that the correlation between the two storm tracks not only show up in VV2501df but can also be clearly seen in the 700-hPa poleward heat flux.

Fig. 8.

(a) Regression of VV2501df (m2 s−2) based on the PACvT index. (b) Same as (a), but based on the ATLvT index. (c)–(f) Same as Fig. 7c–f, but based on the PACvT and ATLvT indices

Fig. 8.

(a) Regression of VV2501df (m2 s−2) based on the PACvT index. (b) Same as (a), but based on the ATLvT index. (c)–(f) Same as Fig. 7c–f, but based on the PACvT and ATLvT indices

4. Discussion

a. Correlation between the storm tracks prior to the mid-1970s

In sections 2 and 3, we focused on the correlation between the two storm tracks after the mid-1970s because of the possibility of biases in the NCEP–NCAR-analyzed variance/covariance statistics prior to that time (Harnik and Chang 2003; Iskenderian and Rosen 2000). Nevertheless, with the availability of ERA-40 data, it is of interest to examine this issue and see whether the two reanalyses agree with each other. Chang and Fu (2002) found a large jump in storm track intensity in the NCEP–NCAR reanalysis data between 1972 and 1975, and Harnik and Chang (2003) suggested that while part of this jump could be seen in radiosonde data, part of the signal could have been spurious. Hence we will focus on data prior to 1972. ERA-40 starts from late 1957, and hence we will examine the correlation from 1957/58 to 1971/72 (15 winters, 45 DJF months3).

The correlation based on the PAC and ATL indices for this earlier period is shown in Fig. 9 for both the NCEP–NCAR reanalysis and ERA-40. The PAC58 and ATL58 indices are not significantly correlated in either case; the correlation equals 0.24 in the NCEP–NCAR reanalysis and 0.18 in ERA-40. Each individual index, though, is highly correlated between the two reanalyses, with the correlation between the PAC58 indices being 0.96 and that between the ATL58 indices being 0.97. In any case, Fig. 9 suggests that the correlation between the two storm tracks is not significant during 1957/58– 1971/72.

Fig. 9.

(a) Correlation between VV2501df and the PAC58 index, based on NCEP–NCAR reanalysis data, for the period DJF 1957/58– 1971/72. (b) Same as in (a), but for the ATL58 index. (c), (d) Same as in (a), (b), but based on ERA-40 data. Shaded areas represent correlations over 0.31

Fig. 9.

(a) Correlation between VV2501df and the PAC58 index, based on NCEP–NCAR reanalysis data, for the period DJF 1957/58– 1971/72. (b) Same as in (a), but for the ATL58 index. (c), (d) Same as in (a), (b), but based on ERA-40 data. Shaded areas represent correlations over 0.31

What is the reason behind the low correlation prior to the mid-1970s? One possibility is that there is a regime change during the mid-1970s, such that the atmosphere favors covarying storm tracks after that time and independent storm tracks prior to that time. Other studies have suggested that there may have been a climate regime transition during the 1970s (e.g., Hurrell 1995; Seager et al. 2000; Trenberth 1990; Mantua et al. 1997).

However, since there are few observations over the storm track peaks during this period, one cannot immediately rule out the alternative possibility that the low correlation could in fact be due to analysis errors introduced by model biases toward low correlations (see next section). The fact that the PAC and ATL indices computed based on the two reanalyses agree quite well would suggest that this is probably not likely, but it is not difficult to construct two slightly different time series based on the analyzed PAC and ATL indices by adding in a correlated part to enhance/diminish the correlation between the two time series. The case in point is that as stated above; the correlation between the ATL and PAC indices in the NCEP–NCAR reanalysis is 0.24, while that in the ERA-40 is 0.18, even though the two analyses are very similar. Hence, it is easy to imagine that it does not take too many changes to the NCEP– NCAR PAC and ATL indices to boost the correlation from 0.24 to higher values,4 which could make the correlation more statistically consistent with that observed during 1975/76–1998/99.

b. Correlation in a GCM simulation

To investigate this issue further, we will examine the correlation between the Pacific and Atlantic storm track in a GCM simulation. The simulation is an ensemble of four experiments using the Geophysical Fluid Dynamics Laboratory (GFDL) climate AGCM (Gordon and Stern 1982; Broccoli and Manabe 1992; Alexander and Scott 1996), forced using observed global SST [referred to as the Global Ocean–Global Atmosphere (GOGA) experiments below] from 1950 to 1995. The experiment is run with a spectral model at a resolution of R30 with 14 unevenly spaced sigma levels. For this set of experiments, only daily 200-hPa meridional velocity is available to us; hence, comparisons will be made with 24-h filtered 200-hPa meridional velocity variance computed based on the NCEP–NCAR reanalysis. We define PAC200 and ATL200 indices by averaging VV2001df over the areas 35°–55°N, 170°E–130°W and 37.5°– 57.5°N, 60°W–0°, respectively. These areas are chosen to be larger than those used to define the PAC and ATL indices in order to allow for the possibility that the storm track peaks in the GCM may be located at a slightly different position from those in the reanalysis.

Even when the PAC and ATL are defined using the same areas, the correlation at 200 hPa is slightly lower than that at 250 hPa, and with a larger area, the correlation drops a bit further, with the correlation between the PAC200 and ATL200 indices being 0.43 based on the NCEP–NCAR reanalysis (for the years 1975/76– 1998/99). The correlation maps based on the two indices are shown in Figs. 10a,b. Its clear that at the 200-hPa level, the two storm tracks are also significantly correlated.

Fig. 10.

(a) Correlation between VV2001df and the PAC200 index, based on NCEP–NCAR reanalysis data for the period DJF 1975/76– 1998/99. (b) Same as in (a), but for the ATL200 index. (c), (d) Same as in (a), (b), but based on 80 winters of GCM simulations forced by observed SST from the years 1975/76–1994/95. Shaded areas represent correlations over 0.25 in (a) and (b) and 0.13 in (c) and (d)

Fig. 10.

(a) Correlation between VV2001df and the PAC200 index, based on NCEP–NCAR reanalysis data for the period DJF 1975/76– 1998/99. (b) Same as in (a), but for the ATL200 index. (c), (d) Same as in (a), (b), but based on 80 winters of GCM simulations forced by observed SST from the years 1975/76–1994/95. Shaded areas represent correlations over 0.25 in (a) and (b) and 0.13 in (c) and (d)

The same correlation, based on the GCM ensemble simulations, is shown in Figs. 10c,d. In this figure, the correlations are computed based only on the data from the years 1975/76–1994/95 (total of 240 months from the four experiments). However, results based on the entire 540 months of GCM data are very similar, suggesting that at least in the GCM simulations, a regime transition like that seen in the reanalysis is not present. The correlation between the PAC200 and ATL200 indices from the GCM experiments is only 0.18. While this is statistically significant at the 95% level for 240 months of data, this correlation is much smaller than that computed based on the reanalysis. The patterns in Fig. 10 also show that the correlation between the two storm tracks, while statistically significant, is weak at best. For all four simulations, during the last 24 yr (1972/73–1994/95), the correlations between PAC200 and ATL200 are 0.28, −0.07, 0.33, and 0.16, respectively.

Since the 24-yr correlation between the two storm tracks is seen to vary so much, a Monte Carlo test is performed to see how different the correlations seen in the reanalysis and the GCM simulations are in a statistical sense. From the 80 yr of GCM simulations (this has also been done for the full 180 yr of GOGA simulations), 24 yr are selected at random and the correlation between the Pacific and Atlantic storm tracks is computed. This preserves the month-to-month correlation in the data but assumes that the interannual correlation is insignificant, which is confirmed to be the case. Ten thousand random samples are formed, and the results show that given the GCM samples, a correlation of over 0.43 is obtained in less than 1% of the samples, suggesting that the GCM simulation and the reanalysis have come from statistically distinct populations. We have also analyzed a 100-yr GCM simulation using the same GCM but forced by a climatological seasonally varying SST distribution. For that experiment, both upper-level meridional velocity variance as well as the 700-hPa meridional heat flux are available, and the results for both quantities are very similar to that based on the GOGA experiments, with the mean correlation computed based on the full experiment (99 winters) much less than that seen in the reanalysis. Monte Carlo tests have also been done with this experiment, and the results again suggest that the GCM and reanalysis data are statistically distinct at the 1% level.

The results in this section show that for the GCM simulations examined, the correlation between the Atlantic and Pacific storm tracks is much lower than that seen in the reanalysis, and no regime transition in terms of change in correlation between the two storm tracks is seen during the 1970s in the set of GCM experiments forced by observed SST variations. However, the 24-yr correlation between the Pacific and Atlantic storm tracks is seen to vary substantially among the four GCM simulations. This raises the question as to why the GCM experiments fail to simulate the high correlation between the two storm tracks and whether the observed significant correlation is just accidental. Though we cannot completely rule out the possibility that the observed relatively high correlation between the two storm tracks between 1975/76 and 1998/99 could just be accidental, the probability for such a chance occurrence appears to be statistically low.

While we do not have information regarding how well the forecast models can simulate the strong correlation between the two storm tracks, if it turns out that the models have low biases (when run as a freely evolving climate model) similar to the GFDL GCM, the results shown here also open up the possibility that the low correlation seen in the reanalyses prior to the mid-1970s could partly be due to biases in the reanalysis model background forecast. Clearly, more work needs to be done in order to understand the physics behind the correlation, and more GCM simulations should be examined to see whether the strong correlation between the two storm tracks is successfully simulated in other GCM experiments.

c. Correlation between the Pacific and Atlantic storm tracks in the Southern Hemisphere

Previous works (e.g., Orlanski and Sheldon 1993, 1995; Chang and Yu 1999) have shown that baroclinic waves in the Atlantic storm track are frequently triggered by upstream waves over the eastern Pacific. One possibility behind the significant correlation between the two storm tracks could simply be that the Atlantic storm track is just the downstream extension of the Pacific storm track; hence, the Atlantic storm track amplitude could conceivably be affected by changes in Pacific storm track intensity.

To test this hypothesis, we examine the variability of the Southern Hemisphere (SH) summer storm track. Lee and Held (1993) and Chang (1999) have shown that baroclinic wave propagation is very coherent along the entire SH summer storm track. Hence, the Pacific and Atlantic storm tracks in the SH summer can clearly be considered as part of a single storm track. Two storm track indices, PACSH and ATLSH, are defined based on the average of VV2501df over the areas 40°–60°S, 170°E–130°W and 40°–60°S, 60°W–0°, respectively, and correlations between the two storm tracks are analyzed as above. Again, we have computed the correlations based on the period DJF 1975/76–1998/99, when there are satellite observations to constrain the reanalysis. Both NCEP–NCAR reanalysis and ERA-40 data have again been analyzed.

The correlations based on these two indices are shown in Fig. 11. The PACSH indices (Figs. 11a,c) clearly do not correlate significantly with the SH Atlantic storm track. Meanwhile, the ATLSH indices do show some significant correlation upstream, with positive correlation around 60°S and negative correlation around 40°S over the Pacific. This suggests that the amplitude of the Atlantic storm track is correlated with the meridional position of the Pacific storm track rather than the intensity of the Pacific storm track. The two indices do not show significant correlation—the correlation between PACSH and ATLSH is 0.12 in the NCEP–NCAR reanalysis data and 0.21 in the ERA-40 data.

Fig. 11.

(a) Correlation between VV2501df and the PACSH index, based on NCEP–NCAR reanalysis data for the period DJF 1975/76– 1998/99 for the SH sector from 140° to 10°E. (b) Same as in (a), but for the ATLSH index. (c), (d) Same as in (a), (b), but based on ERA-40 data. Shaded areas represent correlations over 0.25

Fig. 11.

(a) Correlation between VV2501df and the PACSH index, based on NCEP–NCAR reanalysis data for the period DJF 1975/76– 1998/99 for the SH sector from 140° to 10°E. (b) Same as in (a), but for the ATLSH index. (c), (d) Same as in (a), (b), but based on ERA-40 data. Shaded areas represent correlations over 0.25

As expected, the agreement between the two reanalyses is not as good as that in the Northern Hemisphere. PACSH computed based on the two reanalyses has a correlation of 0.81, while the correlation between the two ATLSH indices is 0.72. Hence, once again we cannot rule out the possibility that the low correlation computed based on the reanalyses may contain biases introduced by biases in the forecast models. However, taken at their face value, the results shown in this section suggest that the mere fact that the Atlantic storm track is a downstream extension of the Pacific storm track does not necessarily imply that the Atlantic storm track amplitude should be significantly correlated with the Pacific storm track amplitude.

5. Conclusions

In this study, the correlation between the Northern Hemisphere winter Pacific and Atlantic storm tracks is examined using the NCEP–NCAR reanalysis and ERA-40, as well as unassimilated aircraft observations. Examining month-to-month variability in 250-hPa meridional velocity variance, the correlation between the two storm track peaks is found to be as high as 0.5 during the winters between 1975/76 and 1998/99. Here, it is shown that the correlation between the two storm tracks can be clearly detected from the aircraft data. Further analyses of the reanalysis data show that the correlation can also be seen in other eddy variance and covariance statistics, including poleward heat flux at the 700-hPa level.

Our analyses also show that the correlation between the two storm tracks, as seen in both reanalysis datasets, is much weaker during the period 1957/58–1971/72, suggesting a possible regime transition from largely uncorrelated storm tracks to highly correlated storm tracks during the 1970s. However, during the earlier period, the number of aircraft observations is insufficient to verify the low correlation seen in the reanalyses. Thus, low biases in the reanalyses during the earlier period cannot be ruled out.

We have also examined an ensemble of four GCM simulations performed using the GFDL GCM forced by observed SST variations between 1950 and 1995. The correlation between the two storm tracks in the GCM simulations is much lower (0.18) than observed, even if we restrict the analysis to the GCM simulations from the period 1975/76 to 1994/95. A Monte Carlo test shows that the observed correlation and the GCM correlation are statistically different at the 1% level.

Correlations between the Southern Hemisphere summer Pacific and Atlantic storm tracks have also been examined based on the reanalyses datasets. The results suggest that the amplitude of the SH summer Pacific and Atlantic storm tracks are not significantly correlated, showing that seeding of the Atlantic storm track by the Pacific storm track does not necessarily lead to significant correlations between the two storm tracks.

Several questions are raised by the results discussed in this study. What is the physical mechanism behind the high correlation between the two storm tracks? Did a transition from a state of largely uncorrelated storm tracks to highly correlated storm tracks occur during the 1970s? Why is the correlation between the two storm tracks seen in the GFDL GCM simulations so much weaker than that observed? Is this a general characteristic of most GCMs, or is it just a characteristic of the GFDL GCM? To answer these questions, more GCM simulations will need to be examined, and idealized modeling and theoretical studies should be conducted to reveal the physics behind this correlation. Without a concrete understanding of this phenomenon, we will not be able to assess whether the inability of GCMs to simulate the correlation represents a serious flaw in the physics/dynamics simulated by these models.

Acknowledgments

The NCEP–NCAR reanalysis and aircraft data were obtained from NCAR, and assistance from the NCAR SCD is much appreciated. The ERA-40 data were obtained from the ECMWF Web site. The author would like to thank Peter Philips and Mary Jo Nath for providing him with the GFDL model data. This work is supported by NSF Grant ATM0296076 and NOAA Grant NA16GP2540.

REFERENCES

REFERENCES
Alexander
,
M. A.
, and
J. D.
Scott
,
1996
:
Atlas of Climatology and Variability in the GFDL R30S14 GCM.
Doc. 1996-774-842, U.S. Government Printing Office, 121 pp
.
Blackmon
,
M. L.
,
1976
:
A climatological spectral study of the 500 mb geopotential height of the Northern Hemisphere.
J. Atmos. Sci
,
33
,
1607
1623
.
Blackmon
,
M. L.
,
J. M.
Wallace
,
N-C.
Lau
, and
S. L.
Mullen
,
1977
:
An observational study of the Northern Hemisphere wintertime circulation.
J. Atmos. Sci
,
34
,
1040
1053
.
Branstator
,
G.
,
1992
:
The maintenance of low-frequency atmospheric anomalies.
J. Atmos. Sci
,
49
,
1924
1945
.
Broccoli
,
A. J.
, and
S.
Manabe
,
1992
:
The effects of orography on midlatitude Northern Hemisphere dry climates.
J. Climate
,
5
,
1181
1201
.
Carillo
,
A.
,
P. M.
Ruti
, and
A.
Navarra
,
2000
:
Storm tracks and zonal mean flow variability: A comparison between observed and simulated data.
Climate Dyn
,
16
,
219
228
.
Chang
,
E. K. M.
,
1999
:
Characteristics of wave packets in the upper troposphere. Part II: Hemispheric and seasonal differences.
J. Atmos. Sci
,
56
,
1729
1747
.
Chang
,
E. K. M.
,
2003
:
Midwinter suppression of the Pacific storm track activity as seen in aircraft observations.
J. Atmos. Sci
,
60
,
1345
1358
.
Chang
,
E. K. M.
, and
D. B.
Yu
,
1999
:
Characteristics of wave packets in the upper troposphere. Part I: Northern Hemisphere winter.
J. Atmos. Sci
,
56
,
1708
1728
.
Chang
,
E. K. M.
, and
Y.
Fu
,
2002
:
Interdecadal variations in Northern Hemisphere winter storm track intensity.
J. Climate
,
15
,
642
658
.
Chang
,
E. K. M.
,
S.
Lee
, and
K. L.
Swanson
,
2002
:
Storm track dynamics.
J. Climate
,
15
,
2163
2183
.
Christoph
,
M.
,
U.
Ulbrich
, and
P.
Speth
,
1997
:
Midwinter suppression of Northern Hemisphere storm track activity in the real atmosphere and in GCM experiments.
J. Atmos. Sci
,
54
,
1589
1599
.
Dommenget
,
D.
, and
M.
Latif
,
2002
:
A cautionary note on the interpretation of EOFs.
J. Climate
,
15
,
216
225
.
Geng
,
Q.
, and
M.
Sugi
,
2001
:
Variability of the North Atlantic cyclone activity in winter analyzed from NCEP–NCAR reanalysis data.
J. Climate
,
14
,
3863
3873
.
Gordon
,
C. T.
, and
W.
Stern
,
1982
:
A description of the GFDL global spectral model.
Mon. Wea. Rev
,
110
,
625
644
.
Graham
,
N. E.
, and
H. F.
Diaz
,
2001
:
Evidence for intensification of North Pacific winter cyclones since 1948.
Bull. Amer. Meteor. Soc
,
82
,
1869
1893
.
Harnik
,
N.
, and
E. K. M.
Chang
,
2003
:
Storm track variations as seen in radiosonde observations and reanalysis data.
J. Climate
,
16
,
480
495
.
Held
,
I. M.
,
S. W.
Lyons
, and
S.
Nigam
,
1989
:
Transients and the extratropical response to El Niño.
J. Atmos. Sci
,
46
,
163
174
.
Hurrell
,
J. W.
,
1995
:
Decadal trends in the North Atlantic Oscillation: Regional temperatures and precipitation.
Science
,
269
,
676
679
.
Iskenderian
,
H.
, and
R. D.
Rosen
,
2000
:
Low-frequency signals in midtropospheric submonthly temperature variance.
J. Climate
,
13
,
2323
2333
.
Kalnay
,
E.
, and
Coauthors
,
1996
:
The NCEP/NCAR 40-Year Reanalysis Project.
Bull. Amer. Meteor. Soc
,
77
,
437
471
.
Kidson
,
J. W.
, and
K. E.
Trenberth
,
1988
:
Effects of missing data on estimates of monthly mean general circulation statistics.
J. Climate
,
1
,
1261
1275
.
Kistler
,
R.
, and
Coauthors
,
2001
:
The NCEP–NCAR 50-year reanalysis: Monthly means CD-ROM and documentation.
Bull. Amer. Meteor. Soc
,
82
,
247
267
.
Lau
,
N-C.
,
1988
:
Variability of the observed midlatitude storm tracks in relation to low-frequency changes in the circulation pattern.
J. Atmos. Sci
,
45
,
2718
2743
.
Lee
,
S.
, and
I. M.
Held
,
1993
:
Baroclinic wave packets in models and observations.
J. Atmos. Sci
,
50
,
1413
1428
.
Livezey
,
R. E.
,
1995
:
Field intercomparison.
Analysis of Climate Variability, H. von Storch and A. Navarra, Eds., Springer-Verlag, 159–176
.
Livezey
,
R. E.
, and
W. Y.
Chen
,
1983
:
Statistical field significance and its determination by Monte Carlo techniques.
Mon. Wea. Rev
,
111
,
46
59
.
Mantua
,
N. J.
,
S. R.
Hare
,
Y.
Zhang
,
J. M.
Wallace
, and
R. C.
Francis
,
1997
:
A Pacific interdecadal climate oscillation with impacts on salmon production.
Bull. Amer. Meteor. Soc
,
78
,
1069
1079
.
Nakamura
,
H.
,
T.
Izumi
, and
T.
Sampe
,
2002
:
Interannual and decadal modulations recently observed in the Pacific storm track activity and East Asian winter monsoon.
J. Climate
,
15
,
1855
1874
.
Orlanski
,
I.
, and
J.
Sheldon
,
1993
:
A case of downstream baroclinic development over western North America.
Mon. Wea. Rev
,
121
,
2929
2950
.
Orlanski
,
I.
, and
J.
Sheldon
,
1995
:
Stages in the energetics of baroclinic systems.
Tellus
,
47A
,
605
628
.
Peng
,
S.
, and
J. S.
Whitaker
,
1999
:
Mechanisms determining the atmospheric response to midlatitude SST anomalies.
J. Climate
,
12
,
1393
1408
.
Pettersen
,
S.
,
1956
:
Weather Analysis and Forecasting. 2d ed.
Vol. 1, McGraw-Hill, 422 pp
.
Seager
,
R.
,
Y.
Kushnir
,
M.
Visbeck
,
N.
Naik
,
J.
Miller
,
G.
Krahmann
, and
H.
Cullen
,
2000
:
Causes of Atlantic Ocean climate variability between 1958 and 1998.
J. Climate
,
13
,
2845
2862
.
Trenberth
,
K. E.
,
1990
:
Recent observed interdecadal climate changes in the Northern Hemisphere.
Bull. Amer. Meteor. Soc
,
71
,
988
993
.
Wallace
,
J. M.
,
G. H.
Lim
, and
M. L.
Blackmon
,
1988
:
Relationship between cyclone tracks, anticyclone tracks, and baroclinic waveguides.
J. Atmos. Sci
,
45
,
439
462
.
Whitaker
,
L. M.
, and
L. H.
Horn
,
1984
:
Northern Hemisphere extratropical cyclone activity for four midseason months.
J. Climatol
,
4
,
297
310
.
Zwiers
,
F. W.
,
1987
:
Statistical considerations for climate experiments. Part II: Multivariate tests.
J. Climate Appl. Meteor
,
26
,
477
487
.

APPENDIX

Test of Field Significance

The issue of field significance is addressed in Livezey and Chen (1983) and reviewed in Livezey (1995). Basically, when correlating a climate index with either an analyzed or a model-generated field (e.g., geopotential height or, in our case, eddy variance and covariance), we would expect that even if there is no real correlation between the index and the field, since we are effectively computing the correlation many times over, by chance some grid points will exhibit significant correlations. Hence, one needs to test the field significance of the correlation patterns obtained.

In this study, field significance is assessed using the permutation test (Zwiers 1987). Technically, since the lag-1 autocorrelation for both the Pacific and Atlantic storm track indices is not statistically significant, the storm track index time series are randomly reordered, and the resultant time series are correlated with the desired field. In each case, some grid points will exhibit significant correlation. The test is repeated 10 000 times, and the number of grid points exhibiting significant correlations in each trial is tabulated and then sorted in ascending order. The number of grid points showing significant correlations exceeded by exactly 5% (i.e., 500) of the trials is obtained from this ordered tabulation. This level is then compared to the number of grid points found to be statistically significant when correlated with the original time series to assess the field significance of the correlation pattern.

This test has been performed for the patterns shown in Figs. 3a,b, 5, 7, and 8. In each case, only the interbasin correlation is tested; that is, when the Pacific storm track index is used, only grid points within 60°W and 0° are considered, while if the Atlantic storm track index is used, only grid points within 160°E to 140°W are considered. The reason for this is that we certainly expect each storm track to be significantly correlated with a storm track index based on itself, and our goal is to see whether the storm track index is significantly correlated with the other storm track.

The results of the tests show that the interbasin correlation patterns shown in Figs. 3a,b, 5a,b, 7a,c,d,f, and 8a,b,c,d,f are field significant at the 5% level (many of these patterns are significant at the 1% level or better). The patterns shown in Figs. 7b,e are not significant even at the 10% level, while the pattern shown in Fig. 8e is only significant at the 10% level but not at the 5% level. As discussed in the main text, the interbasin correlations between the storm track indices and the poleward flux of zonal momentum (Figs. 7a,b), as well as the 1000-hPa geopotential height (Figs. 7e,f and 8e,f), appear to be weaker than those of the other parameters. Nevertheless, several of these panels do manage to pass the field significance test.

Footnotes

Corresponding author address: Dr. Edmund K. M. Chang, Institute for Terrestrial and Planetary Atmospheres, Marine Sciences Research Center, State University of New York at Stony Brook, Stony Brook, NY 11794-5000. Email: kmchang@notes.cc.sunysb.edu

1

For 69 degrees of freedom (DOF) a correlation of 0.24 is 95% significant, while a correlation of 0.28 is 99% significant. Even if we assume only 50 DOF, the 95% and 99% significance levels are 0.28 and 0.36, respectively. Note that the lag-1 correlation for both the PAC and ATL indices is not statistically significant, meaning that we can treat the number of months (minus 3 since anomalies are considered here) as the number of DOF.

2

As in Chang (2003), to eliminate erroneous out-of-range observations, the 4D reanalysis grid is linearly interpolated to the spatial and temporal location of each observation, and observations that differ from the interpolated reanalysis value by more than 30 m s−1 are dropped. We have experimented with different cutoff values and found that the results are not sensitive to the cutoff value.

3

For 42 DOFs, a 95% significant level requires a correlation of about 0.31.

4

For example, simply by adding the differences between the NCEP–NCAR and ERA-40 indices to the NCEP–NCAR indices (i.e., doubling the differences between the two reanalyses) immediately raises the correlation between the two indices to 0.265.