• Andrade, E. R., Jr., and W. D. Sellers, 1988: El Niño and its effect on precipitation in Arizona and western New Mexico. J. Climatol.,8, 403–410.

  • Barnett, T. P., and R. Preisendorfer, 1987: Origins and levels of monthly and seasonal forecast skill for United States surface air temperatures determined by canonical correlation analysis. Mon. Wea. Rev.,115, 1825–1850.

  • Barnston, A. G., 1994: Linear statistical short-term climate predictive skill in the Northern Hemisphere. J. Climate,7, 1513–1564.

  • ——, and R. E. Livezey, 1987: Classification, seasonality and persistence of low-frequency atmospheric circulation patterns. Mon. Wea. Rev.,115, 1083–1126.

  • ——, ——, and M. S. Halpert, 1991: Modulation of Southern Oscillation–Northern Hemisphere mid-winter climate relationships by the QBO. J. Climate,4, 203–217.

  • Bell, G. D., and J. E. Janowiak, 1995: Atmospheric circulation associated with the Midwest floods of 1993. Bull. Amer. Meteor. Soc.,76, 681–695.

  • Fu, C., H. F. Diaz, and J. O. Fletcher, 1986: Characteristics of the response of sea surface temperature in the central Pacific associated with warm episodes of the Southern Oscillation. Mon. Wea. Rev.,114, 1716–1738.

  • Halpert, M. S., and C. F. Ropelewski, 1992: Surface temperature patterns associated with the Southern Oscillation. J. Climate,5, 577–593.

  • Hoerling, M. P., A. Kumar, and M. Zhong, 1997: El Niño, La Niña and the nonlinearity of their teleconnections. J. Climate,10, 1769–1786.

  • Ji, M., A. Kumar, and A. Leetmaa, 1994: A multiseason climate forecast system at the National Meteorological Center. Bull. Amer. Meteor. Soc.,75, 569–577.

  • Kousky, V. E., 1993: The global climate of December 1991–February 1992: Mature-phase warm (ENSO) episode conditions develop. J. Climate,6, 1639–1655.

  • Kumar, A., and M. P. Hoerling, 1995: Prospects and limitations of seasonal atmospheric GCM predictions. Bull. Amer. Meteor. Soc.,76, 335–345.

  • ——, and ——, 1997: Interpretation and implications of the observed inter–El Niño variability. J. Climate,10, 83–91.

  • ——, A. Leetmaa, and M. Ji, 1994: Simulations of atmospheric variability induced by sea surface temperatures and implications for global warming. Science,266, 632–634.

  • Livezey, R. E., 1995: The evaluation of forecasts. Analysis of Climate Variability: Applications of Statistical Techniques, H. von Storch and A. Navarra, Eds., Springer-Verlag, 177–196.

  • ——, and W. Chen, 1983: Statistical field significance and its determination by Monte Carlo methods. Mon. Wea. Rev.,111, 46–59.

  • ——, and K. C. Mo, 1987: Tropical–extratropical teleconnections during the Northern Hemisphere winter. Part II: Relationships between monthly mean Northern Hemisphere circulation patterns and proxies for tropical convection. Mon. Wea. Rev.,115, 3115–3132.

  • ——, J. D. Hoopingarner, and J. Huang, 1995: Verification of official monthly mean 700-hPa height forecasts: An update. Wea. Forecasting,10, 512–527.

  • ——, M. Masutani, and M. Ji, 1996: SST-forced seasonal simulation and prediction skill for versions of the NCEP/MRF model. Bull. Amer. Meteor. Soc.,77, 507–517.

  • Murphy, A. H., and E. S. Epstein, 1989: Skill scores and correlation coefficients in model verification. Mon. Wea. Rev.,117, 572–581.

  • Parker, D. E., P. D. Jones, C. K. Folland, and A. Bevan, 1994: Interdecadal changes of surface temperature since the late nineteenth century. J. Geophys. Res.,99, 14 373–14 399.

  • Philander, G. S., 1990: El Niño, La Niña, and the Southern Oscillation. Academic Press, 289 pp.

  • Reynolds, R. W., 1988: A real-time global sea surface temperature analysis. J. Climate,1, 75–86.

  • Ropelewski, C. F., and M. S. Halpert, 1986: North American precipitation and temperature patterns associated with the El Niño/Southern Oscillation (ENSO). Mon. Wea. Rev.,114, 2352–2362.

  • ——, and ——, 1989: Precipitation patterns associated with the high index phase of the Southern Oscillation. J. Climate,2, 268–284.

  • ——, and ——, 1996: Quantifying Southern Oscillation–precipitation relationships. J. Climate,9, 1043–1059.

  • Rowell, D. P., C. K. Folland, K. Maskell, and M. N. Ward, 1995: Variability of summer rainfall over tropical North Africa (1906–92): Observations and modeling. Quart. J. Roy. Meteor. Soc.,121, 669–704.

  • Sittel, M. C., 1994: Differences in the means of ENSO extremes for maximum temperature and precipitation in the United States. Center for Ocean–Atmospheric Prediction Studies Tech. Rep. 94-2, 50 pp. [Available from Dr. James J. O’Brien, The Florida State University, Tallahassee, FL 32306.].

  • Smith, T. M., R. W. Reynolds, R. E. Livezey, and D. C. Stokes, 1996: Reconstruction of historical sea surface temperatures using empirical orthogonal functions. J. Climate,9, 1403–1420.

  • Wallace, J. M., and D. Gutzler, 1981: Teleconnections in geopotential height field during the Northern Hemisphere winter. Mon. Wea. Rev.,109, 784–812.

  • View in gallery

    Departures from 1950–79 normals (Reynolds 1988) of global SSTs for January 1992 used to force the J92 experiment. Contour interval is 0.5°C and negative anomalies are shaded.

  • View in gallery

    Monthly mean 700-hPa height anomaly composites for the J92 experiment. Contour interval is 20 m, the zero contour is dotted, and negative contours are dashed.

  • View in gallery

    Monthly mean U.S. surface temperature anomaly composites for the J92 experiment. Contour interval is 0.5°C, the zero contour is omitted, and negative contours are dashed.

  • View in gallery

    Monthly mean U.S. precipitation composites for the J92 experiment. Contour interval is 10 mm, the zero contour is omitted, and negative contours are dashed.

  • View in gallery

    Observed monthly mean 700-hPa height anomaly composites for cases with average positive SST anomalies in the central equatorial Pacific (area described in the text) that exceed the criterion (in degrees Celsius times 10) in the lower-right corner of each panel. Contour interval is 10 m, the zero contour is dotted, and negative contours are dashed. Light and dark shading denote 10% and 5% local statistical significance, respectively.

  • View in gallery

    Same as Fig. 5 except for the DJ-window composite.

  • View in gallery

    Same as Fig. 5 except for negative SST anomalies.

  • View in gallery

    Same as Fig. 5 except for ON- through MJ-window composites and negative SST anomalies.

  • View in gallery

    Same as Fig. 5 except for U.S. surface temperature composites, contour interval of 0.5°C, and no-zero contour.

  • View in gallery

    Same as Fig. 9 except for (a)–(c) January–March warm ENSO episode composites and (d) the JFM-window composite.

  • View in gallery

    Same as Fig. 9 except for (a) MAM-, (b) JA-, (c) JAS-, and (d) DJ-window composites.

  • View in gallery

    Same as Fig. 9 except for negative-SST anomalies.

  • View in gallery

    Same as Fig. 9 except for the JF-window composite and negative SST anomalies.

  • View in gallery

    Same as Fig. 5 except for U.S. precipitation, contour interval of 10 mm, and no zero contour.

  • View in gallery

    Same as Fig. 14 except for (a) the January warm ENSO episode and (b) JFM-, (c) MJ-, (d) ON-, and (e) ND-window composites.

  • View in gallery

    Same as Fig. 14 except for negative SST anomalies.

  • View in gallery

    Same as Fig. 14 except for (a) JF- and (b) OND-window composites and negative SST anomalies.

  • View in gallery

    Same as Fig. 5 except for 45Y experiment composites.

  • View in gallery

    Monthly dependence of measures of quality and skill of the match between corresponding observational and 45Y experiment 700-hPa height composites for both (a) positive- and (b) negative-SST anomalies. The solid, dashed, and dotted curves are for anomaly correlation ρFX, unconditional bias C, and skill score β, respectively.

  • View in gallery

    Same as Fig. 7 except for 45Y experiment composites.

  • View in gallery

    Same as Fig. 9 except for 45Y experiment composites.

  • View in gallery

    Same as Fig. 19 except for U.S. surface temperature.

  • View in gallery

    Same as Fig. 12 except for 45Y experiment composites.

  • View in gallery

    Same as Fig. 14 except for 45Y experiment composites and contour interval of 5 mm.

  • View in gallery

    Same as Fig. 19 except for U.S. precipitation.

  • View in gallery

    Same as Fig. 16 except for 45Y experiment composites and contour interval of 5 mm.

All Time Past Year Past 30 Days
Abstract Views 0 0 0
Full Text Views 30 30 12
PDF Downloads 20 20 8

Teleconnective Response of the Pacific–North American Region Atmosphere to Large Central Equatorial Pacific SST Anomalies

View More View Less
  • 1 Climate Prediction Center, NCEP/NWS/NOAA, Washington, D.C.
  • 2 Research and Data Systems Corporation, Greenbelt, Maryland
  • 3 Environmental Modeling Center, NCEP/NWS/NOAA, Washington, D.C.
  • 4 Research and Data Systems Corporation, Greenbelt, Maryland
  • 5 Environmental Modeling Center, NCEP/NWS/NOAA, Washington, D.C.
© Get Permissions
Full access

Abstract

A prominent year-round ensemble response to a global sea surface temperature (SST) anomaly field fixed to that for January 1992 (near the peak of a major warm El Niño–Southern Oscillation episode) was observed in a 20-yr integration of the general circulation model used for operational seasonal prediction by the U.S. National Weather Service. This motivated a detailed observational reassessment of the teleconnections between strong SST anomalies in the central equatorial Pacific Ocean and Pacific–North America region 700-hPa heights and U.S. surface temperatures and precipitation. The approach used consisted of formation of monthly mean composites formed separately from cases in which the SST anomaly in a key area of the central equatorial Pacific Ocean was either large and positive or large and negative. Extensive permutation tests were conducted to test null hypotheses of no signal in these composites. The results provided a substantial case for the presence of teleconnections to either the positive- or negative-SST anomalies in every month of the year. These signals were seasonally varying (sometimes with substantial month to month changes) and, when present for both SST-anomaly signs in a particular month, usually were not similarly phased patterns of opposite polarity (i.e., the SST–teleconnected variable relationships were most often nonlinear).

A suite of 13 45-yr integrations of the same model described above was run with global SST analyses reconstructed from the observational record. Corresponding composites from the model were formed and compared visually and quantitatively with the high-confidence observational signals. The quantitative comparisons included skill analyses utilizing a decomposition that relates the squared differences between two maps to phase correspondence and amplitude and bias error terms and analyses of the variance about composite means. For the latter, in the case of the model runs it was possible to estimate the portions of this variance attributable to case to case variation in SSTs and to internal variability. Comparisons to monthly mean maps and analyses of variance for the 20-yr run with SSTs fixed to January 1992 values were also made.

The visual and quantitative comparisons all revealed different aspects of prominent model systematic errors that have important implications for the optimum exploitation of the model for use in prediction. One of these implications was that the current model’s ensemble responses to SST forcing will not be optimally useful until after nonlinear correction of SST-field-dependent systematic errors.

Corresponding author address: Dr. Robert E. Livezey, Climate Prediction Center, NCEP/NWS/NOAA, Washington, DC 20233.

Email: livezey@sgi84.wwb.noaa.gov

Abstract

A prominent year-round ensemble response to a global sea surface temperature (SST) anomaly field fixed to that for January 1992 (near the peak of a major warm El Niño–Southern Oscillation episode) was observed in a 20-yr integration of the general circulation model used for operational seasonal prediction by the U.S. National Weather Service. This motivated a detailed observational reassessment of the teleconnections between strong SST anomalies in the central equatorial Pacific Ocean and Pacific–North America region 700-hPa heights and U.S. surface temperatures and precipitation. The approach used consisted of formation of monthly mean composites formed separately from cases in which the SST anomaly in a key area of the central equatorial Pacific Ocean was either large and positive or large and negative. Extensive permutation tests were conducted to test null hypotheses of no signal in these composites. The results provided a substantial case for the presence of teleconnections to either the positive- or negative-SST anomalies in every month of the year. These signals were seasonally varying (sometimes with substantial month to month changes) and, when present for both SST-anomaly signs in a particular month, usually were not similarly phased patterns of opposite polarity (i.e., the SST–teleconnected variable relationships were most often nonlinear).

A suite of 13 45-yr integrations of the same model described above was run with global SST analyses reconstructed from the observational record. Corresponding composites from the model were formed and compared visually and quantitatively with the high-confidence observational signals. The quantitative comparisons included skill analyses utilizing a decomposition that relates the squared differences between two maps to phase correspondence and amplitude and bias error terms and analyses of the variance about composite means. For the latter, in the case of the model runs it was possible to estimate the portions of this variance attributable to case to case variation in SSTs and to internal variability. Comparisons to monthly mean maps and analyses of variance for the 20-yr run with SSTs fixed to January 1992 values were also made.

The visual and quantitative comparisons all revealed different aspects of prominent model systematic errors that have important implications for the optimum exploitation of the model for use in prediction. One of these implications was that the current model’s ensemble responses to SST forcing will not be optimally useful until after nonlinear correction of SST-field-dependent systematic errors.

Corresponding author address: Dr. Robert E. Livezey, Climate Prediction Center, NCEP/NWS/NOAA, Washington, DC 20233.

Email: livezey@sgi84.wwb.noaa.gov

1. Introduction

Arguably the most important source of atmospheric predictability on intraseasonal to interannual timescales is slow-changing sea surface temperature (SST) anomalies in the tropical oceans and the forcing associated with them. The most well known and perhaps most dramatic of these are the multiyear swings in SSTs in the central equatorial Pacific Ocean collectively referred to as the El Niño–Southern Oscillation (ENSO) system (Philander 1990). Although many of the consequences of ENSO and other tropical SST variability are well documented [e.g., Ropelewski and Halpert (1986, 1989, 1996) and Halpert and Ropelewski (1992) for U.S. surface temperatures and precipitation], in some respects our understanding is somewhat crude and limited. This is especially true in the case of extratropical teleconnections, which are emphasized in this study.

In this setting, we have readdressed the problem using a somewhat different strategy than previous studies. Because this is our first attempt using this approach (described below), our focus has been narrowed to examination of Pacific–North American (PNA) region responses to substantial SST anomalies in a particular area of the central equatorial Pacific Ocean where such anomalies have often occurred during recent strong ENSO episodes. The reasons for this specific emphasis will be made clear in the following discussion.

The most important features of our analysis are 1) the symbiotic use of an atmospheric general circulation model (GCM) and observations and 2) a compositing technique in which selection of constituent cases is based on the underlying tropical forcing regardless of whether or not the situation is classified as an ENSO episode.

The advantages of use of the observational record to build confidence in the dynamic models, identify systematic errors, and suggest improvements to the GCMs are obvious, but advantages of the models to guide exploration of the observations are less obvious. The latter is best understood in the context of the limits of the observational record. Currently, analyses of global SST and northern extratropical circulation data of sufficient accuracy and detail span a period of less than 50 yr, although efforts are under way to extend the record (e.g., Smith et al. 1996). Thus, from a signal to noise perspective only the grossest low-frequency teleconnections will be detectable without a priori guidance of what to look for. Discrimination of different responses between cold/warm or weak/strong ENSO episodes even in the northern winter when responses are apparently strongest has been difficult and not altogether conclusive (e.g., Ropelewski and Halpert 1986; Livezey and Mo 1987; Barnston et al. 1991). Consistent identification of definite tropical–extratropical connections for other seasons with smaller signal to noise ratios has been even more difficult, although some progress has been made recently for the PNA region by Bell and Janowiak (1995) and Barnston (1994) for the spring and summer, respectively.

These signal detection problems do not exist in principle for dynamic models; any number of realizations for a prescribed SST anomaly pattern can be generated to isolate the model response regardless of how unimportant it may be (Kumar and Hoerling 1995). The usefulness of this approach has traditionally been limited by available computer resources and lack of confidence in how representative the model responses are. Technological advances have reduced the first obstacle, while a performance analysis of the atmospheric GCM used here suggests that it should be a useful tool at least for the subtropics and PNA region extratropics (Livezey et al. 1996).

The other important aspect of the work reported below is our choice of methodology for analysis of the observational data. Two approaches have been taken most commonly in the past: linear statistical analysis or composite analysis (or some other stratification of the data) based on classification by ENSO episodes. Examples of the former are the sophisticated linear analyses by Barnett and Preisendorfer (1987) and Barnston (1994) and examples of the latter are the studies by Barnston et al. (1991), Halpert and Ropelewski (1992, hereafter HR92), Sittel (1994), and Ropelewski and Halpert (1996, hereafter RH96), with the study by Livezey and Mo (1987) reflecting some aspects of both. Each of the traditional approaches has its drawbacks. We believe that sought-after signals are at best piecewise linear (empirical evidence for this can be found in all five of the composite studies cited in the previous sentence and in the results here), so any linear analysis will only reflect a compromise of different regimes of behavior. On the other hand, methods that rely on stratification by ENSO episode suffer partially by the arbitrary classification process and more so by the fact that ENSO episodes often differ substantially from one another in the temporal evolution, strength, or spatial pattern of SST anomalies.

Given these problems, we have adopted a simple approach that exploits advantages of both of the most frequently used approaches, namely composite analysis with constituent cases based on forcing (i.e., SST) criterion rather than classification by ENSO state. The criterion used here (suggested by model experiments) is admittedly crude, but was sufficient to address our focus: Teleconnected responses to large SST anomalies in the central equatorial Pacific Ocean characteristic of (but not necessarily linked to) recent strong ENSO episodes. More sophisticated approaches to case selection are being explored in preparation for future studies, which will examine effects of structural and amplitude differences in the SST anomaly field. An example of the same philosophical approach to the data analysis here, with a narrower focus but more sophisticated case selection, is the work of Montroy et al. (1997, manuscript submitted to J. Climate, hereafter MRL). Finally, composites developed from an ENSO-episode classification also were examined to supplement those based on SSTs.

The analyses emphasized monthly rather than seasonal averages. Although this placed a heavier burden on the observational data because of signal/noise problems, the model experiments suggested that month to month changes in sought-after signals are often large. In those instances where this was not the case, we have enhanced the representation of the common monthly mean signal by pooling cases over two- and three-month windows.

In the next section the observational data, model experiments, and analysis techniques will be described. This will be followed by results of prescribed idealized-SST model runs, analyses of the observational data, and results of prescribed observed-SST experiments in sections 3–5, respectively. The narrative will conclude in section 6. Although the model results motivate, complement, and support the observational results, a reader who is principally interested in the latter can skip sections 2a, 2c(3), 3, and 5, and the description of analysis of variance calculations in section 2c(2) without incurring much difficulty with the rest of the narrative.

2. Methods

As implied above, two sets of model runs of somewhat different character were used in this study. The first of these strongly influenced the subsequent search for signals in the observational data, while the second was used to confirm some of the observational analysis results and better place some of the first set of model results in context. The work proceeded in this order, but because both model experiments utilized the same model these will be described first together.

a. Model experiments

Both sets of GCM runs utilized a modified T40 version of the National Centers for Environmental Prediction’s (NCEP) Medium Range Forecast (MRF) model. Modifications included changes to vertical mixing, convection, and cloud–radiation interaction parameterizations to improve the spatial distribution and other properties of tropical convection. The model has been shown to be useful under certain conditions for production of seasonal forecast guidance [cf. Ji et al. (1994) for description of the model and Livezey et al. (1996) for a performance analysis of it]. Similarly, both suites of integrations were forced with prescribed SSTs. The first of these (hereafter referred to as the J92 experiment) consisted of a single 20-yr run initialized with the analyzed atmospheric conditions on 1 February 1992, and subsequently forced month by month with a global SST field consisting of the annually varying operational climatology (Reynolds 1988) but with anomalies fixed at those observed for January 1992 (Fig. 1). Because the fixed SST anomalies were taken from the height of a major ENSO warm episode, the J92 run can be thought of as a “perpetual” strong warm episode simulation. Monthly mean atmospheric anomalies from this run are defined as departures from the monthly mean climatology of a corresponding 20-yr run with zero SST anomalies. Both the J92 and zero-anomaly runs are briefly referred to by Kumar et al. (1994).

Two aspects of J92 experimental design impact subsequent work and interpretations, one probably less seriously than the other. First, because January 1992 SST anomalies rather than standardized anomalies are persisted, the relative (with respect to the historical record) strength of the prescribed tropical anomalies for part of the year will be greater than for January. This is a contributing factor for our inability in the observational part of our study to find as many good analogs to the J92 tropical forcing in the warm half of the Northern Hemisphere year as in the cold half.

Second, extratropical SST anomalies are persisted as well as the tropical anomalies. Obviously the timescales of the J92 anomalies are artificial and unrealistic, but this is a less serious problem for the Tropics where observed timescales, both of SST anomalies and air–sea interaction, are longer. In contrast, extratropical SST anomalies have relatively short timescales and are highly responsive to low-frequency atmospheric variability. Thus, without additional experimentation, J92 run signals cannot be unambiguously ascribed to tropical forcing alone because of the presence of nontrivial extratropical SST anomalies (i.e., those in the North Pacific Ocean in Fig. 1).

The second suite of model runs (hereafter referred to as the 45Y experiment) consists of 13 45-yr (1950–1994) integrations where monthly mean SSTs were prescribed by the reconstructed global SST analyses of Smith et al. (1996). These are simulation runs and represent extensions of the modern-era-only integrations evaluated in Livezey et al. (1996). Here only composite responses corresponding to those formed from the observations were examined. A more detailed performance analysis will be subsequently published. Simulated monthly mean atmospheric anomalies for each run through 1993 were formed by month by month subtraction of the 45-yr model climatology calculated from the first 9 of the 13 runs (addition of the last 4 had inconsequential effect). For a given variable field, all 13 anomaly map series were then averaged to form the ensemble mean simulation series. A very slow mass leak in the 45Y runs necessitated adjustments to account for a spurious trend in the upper-air height fields. These adjustments will be reported on in a later paper.

b. Data

The atmospheric datasets used in the observational portions of this study, monthly mean Northern Hemisphere extratropical 700-hPa geopotential heights, and monthly mean temperatures and precipitation for U.S. climate divisions, were the same described by Livezey et al. (1996). In all cases anomalies are defined as departures from 1950–1994 normals, consistent with the 45Y experiment normal period.

In addition to the global reconstructed SST analyses used to drive the 45Y runs, the Global Sea-Ice SST (GISST) analysis (Parker et al. 1994) was used to select composite members as described next.

c. Composite analysis

1) Formation

For the J92 run fields all 20 Januarys, Februarys, etc. were averaged to define the SST-forced signal. Charts derived separately from the first and second 10 yr of the run (not shown) with few exceptions were remarkably similar in all important respects. Thus, we are confident the J92 responses presented in section 3 almost universally represent statistically significant signals rather than noise residuals. Where there was instability in a particular case will be clearly noted in the narrative.

Cases for final composites of both the observations and corresponding 45Y run maps were the same, but their selection was based only on the observational data. Specifically, the index used for case selection was the average SST anomaly in an area bounded by the date line and 150°W and 5°N and 5°S. It was chosen because our interest was the teleconnected responses to strong central equatorial Pacific forcing, like that imposed in the J92 runs, and in making meaningful comparisons between the three sets of composites. The key area encompasses the largest negative outgoing longwave radiation (OLR) anomalies observed for January 1992 (Kousky 1993) and the largest SSTs (in excess of 28.5°C, see Fig. 1) along the Pacific equator. Comparably large OLR anomalies generally have not been observed in areas much farther east, even in cases where they exhibit comparable SST anomalies. This is because of considerably lower normal SSTs in these areas.

Thus, the average SST anomaly in the area we have selected should reasonably reflect the anomalous forcing. Obviously, this is a crude index and there will be exceptions to the circumstances described above. More sophisticated criteria for composite membership are possible and their use will be motivated in section 6. Nevertheless, the index used here has proved to be a useful first criterion. Incidentally, it was formed from the GISST analyses rather than the reconstructed analyses because the critical area was well sampled through the record and the GISST analyses have little inherent smoothing in sampled area. We did not, however, believe the maps from this version of GISST were suitable for forcing the GCM.

Actual formation of a sequence of monthly composites of a particular variable proceeded in two steps. First, two different paired sets of both warm and cold SST anomaly composites were formed from two different membership selection criteria, one where the cutoff for selection for either the warm or cold anomaly composites was a mean absolute GISST anomaly for the composite month in the key area in excess of 0.8°C, and the other in excess of 1.0°C. These two sets of paired composite sequences were subjected to statistical significance tests (described in the next section) to determine to what extent either the warm or cold SST anomaly composites would be given further consideration.

Final composite membership was a priori constrained to be from 4 to 10 cases. The lower bound was selected to reduce sampling errors and the upper bound to limit sample size disparities for different times of the year. These membership limits occasionally mandated, for a particular composite, use of the 0.8°C and 1.0°C charts, respectively, because the former always contained at least 4 members and the latter at most 10. In those instances where both charts fell within the preset limits for membership, the one with the strongest relative signal (with respect to statistical significance) was selected for further analysis.

Cases that met the two criteria are listed for monthly composites in Table 1. Note that there is no particular bias toward selection of either the earlier or later parts of the record, so we were not particularly concerned about trends complicating interpretation of the composites. Composites of monthly mean anomaly maps over two- and three-month windows were formed from cases that satisfied the 1.0°C criterion and were not limited to a maximum number of members. These will be referred to below by two and three capital letters, respectively, to denote the months within a window, that is, JF for January and February and JFM for January–March.

Composites were also formed from the observational data based on a classification of 12-month periods from August through July as either strong warm or cold ENSO episodes for purposes of comparison. This classification was not quantitatively based but was rather the result of a combination of the previous work of others (e.g., Fu et al. 1986) and experience. The seven warm and six cold episodes selected are listed in Table 2. Note the absence of the very unusual warm episode of 1976–1977 in which very large SST anomalies were confined to the far eastern equatorial Pacific basin. This exclusion is supported by the appearance of only three entries for the period in Table 1, one of which signals the beginning of the 1977–1978 warm episode, which is included in Table 2.

2) Statistical and practical significance

Statistical significance of J92 composites was assessed by comparison of composites constructed separately from the first and second halves of the run. However, this approach is not practical for the observed and 45Y monthly mean composites because composite membership never exceeds 10 and is barely adequate in many instances for resolving any signal.

Instead, a permutation procedure (cf. Livezey 1995) was used to determine the field significance (Livezey and Chen 1983) of all composite maps other than those for the J92 run. For each composite, a local two-tailed t test was conducted at the 5% level at every data point of a prescribed domain (30°–180°W, 25°–90°N for the upper-air grid and the United States for the surface data). The percent of total area over a prescribed domain where the null hypothesis of no signal was rejected was then determined and tested against an estimate of its empirical distribution function for the particular variable–domain–month or window.

A particular empirical distribution function was generated by computation of the test statistic for 1000 random selections of the same number of large warm anomaly and cold anomaly cases that make up the composites being tested. For monthly and ENSO-episode composites, this amounts to randomly relabeling the years on the appropriate 45 anomaly charts followed by recalculation of the test statistic, with the whole process repeated 1000 times. However, in the case of two- and three-month window monthly mean composites, all cases in a window were first pooled before random relabeling. These approaches are equivalent to resampling without replacement.

Once field significances were available, final sequences of charts were formed following the rules outlined in the previous section. In order to determine whether a particular variable’s 12-month composite sequence should be given further serious consideration outside of the January–March period (where there is already ample prior evidence of a signal), several steps were taken to screen the results. First, to partially account for the multiplicity of field significance tests conducted, the binomial distribution was utilized for each calendar sequence to test the null hypothesis that no signal was present for the April–December period. This was done separately for each variable–domain’s cold and warm SST anomaly composites. For this step it was assumed (under the null hypothesis) that the nine months were all practically statistically independent of each other. This was reasonable for the variables for which composites were formed here because their autocorrelations imply times between independent samples that are not very different from the sampling interval of one month.

If the null hypothesis could be rejected at a reasonable level for a sequence, we proceeded with its further examination. On the other hand, if it was not possible to reasonably reject the null hypothesis (the situation for three out of the six monthly composite sequences), the smoother two- and three-month window monthly average sequences (from those beginning in April through those ending in December) were subjected to a similar binomial test, but with the assumption that only one out of two 2-month window and one out of three 3-month window maps were statistically independent. This step provided the necessary justification for presentation of the remaining sequences. Of course the three warm-SST anomaly sequences were also examined for features that reflected behavior in the J92 experiment as additional support for their representativeness.

Informally, consistency of monthly composite features with those for an adjacent month or same-month ENSO-episodes whose composite was formally statistically significant was also used from time to time to enhance the credibility of a signal. Additionally, as already mentioned, in those instances where signals did not seem to change much from month to month, we sharpened their presentation by concatenation of 1.0°C cutoff results from two and three adjacent months. In this regard, selected corroborative ENSO-episodes and two- and three-month window composites will also be displayed. Caution must be exercised in the use of a strong cold SST anomaly composite that resembles an inverted (with respect to anomaly sign) weak warm SST anomaly composite, or vice versa, to support at least the presence of a linear signal in the latter. This is because the anomaly data is centered on the full-record mean, so there will always be a tendency, even in the absence of signal, for cold and warm SST anomaly charts to exhibit inverted features. In this context, statistical significance of individual composite maps becomes crucial, and examples will be noted where appropriate.

Composite sequences for the 45Y runs were formed, tested, and displayed exactly as the final observational monthly sequences to facilitate comparisons.

To some extent the permutation tests indicated the practical and potential importance of the observational and model composites, respectively. In order to more explicitly address these issues we have also computed a signal to noise ratio for each composite in the form of a percent of explained variance. Denote an anomaly at a location by ( )′, its composite average by an overbar, and its area-weighted sum over its domain (same as for the significance tests) by angle brackets. Then the percent of variance that a composite explains for the N maps (Table 3) that make up the composite, EV, can be written as
i1520-0442-10-8-1787-e1
where the summations are over the composite membership.
To further contrast the observed and two different types of model composites and to investigate the relative importance of SST-forced and internally generated variability in the 45Y experiment, analyses of variance were performed. Let M denote the number of different global SST fields used in a composite, and Q the number of realizations for each of these fields. For the J92 run M = 1 and N = Q, for the observational composites Q = 1 and N = M, and for the 45Y runs Q = 13 and N = MQ (Table 3). It can be easily shown that the mean squared amplitude of members of a composite (the denominator in the second term of [1] divided by N) can be written
i1520-0442-10-8-1787-e2
where VSC is the mean squared amplitude of the composite map, VE is the mean squared difference between the M case ensemble averages and the composite, and VR is the mean squared difference between the MQ case realizations and their respective M case ensemble averages. With these definitions, VE = 0 for the J92 experiment composites and VR = 0 for the observational composites. Note of course that EV = VCVT. This quantity will be tabulated for all monthly mean composites displayed.
For the 45Y experiment, unbiased estimates of total variance about the composite mean V and variances separately attributable to SST VSC and internal variability VI can be expressed (Rowell et al. 1995) in terms of VE and VR as
i1520-0442-10-8-1787-e3
Unbiased variances for the J92 experiment and observational composites are given by VI and VS without the second term, respectively. The ratios of VC, VI, and VS to V were formed by pooling the numerators and the denominator separately over all months. The pooling consisted of a weighted average (by M) for the observational and 45Y run composites. These results are presented in section 5d for all positive-SST anomaly composites.

3) Comparison

Visual comparisons were made month by month between corresponding pairs of composites from the two suites of model runs and the observations. Those between one or another of the model sets and the observations should reveal something about model skill and systematic error. Both of these topics are being pursued more thoroughly in work to be reported later, but they were also partially examined quantitatively here as described in the next paragraph. Comparisons between the positive-SST anomaly composites of the two model runs along with those for observational–model pairs may reveal something about the extent either model experiment’s composites are representative of the other’s. For instance, if the model composites are quite different from each other in some cases, it is possible that the key area selection criteria in the 45Y run is inadequate for representation of the J92 SST forcing. The fault, however, may lie in the design flaws in the J92 experiment if they compromise the ability under any circumstances to select a sufficient number of cases from the 45Y experiment for meaningful comparisons.

Let X and F denote the composite averages of corresponding anomalies from two different composites at a location, that is, shorthand for the prime and overbar operations defined above. Then, analogous to (1) the percent of squared anomaly of the map of X explained by F, β, is written
βXF2X2
Let s2F, s2X, and sFX denote the individual map variances and covariance of the two maps, respectively, and define the anomaly correlation as
ρFXsFXsF sX
The score β can then be expanded through the Murphy and Epstein (1989) decomposition as
i1520-0442-10-8-1787-e6a
The term D2 in (6a) acts as a normalization factor for β such that for perfect matches (A2 =1, B2 = C2 = 0) β = 1, and for F ≡ 0 (A2 = B2 = 0, C2 = D2) β = 0. With reasonable climatologies, C2 and D2 are usually small and substantially offset each other, especially for the larger domains examined here. Thus,
βA2B2
The practical implication of approximation (7) is that ρFX usually must exceed 1/2(sF /sX) for β > 0, that is, for the two maps to have smaller squared difference than for when F ≡ 0. For more details of this type of analysis, see either Livezey et al. (1995) or Livezey (1995).

In section 5, terms of (6) of interest between observed and 45Y experiment monthly composites of different variables will be presented. The comparisons will be made over the same domains used for the other quantitative analyses: the PNA-region domain for upper-air heights and the United States for surface variables.

3. J92 experiment composites

Monthly mean composites for the J92 experiment are displayed for 700-hPa height and U.S. surface temperature and precipitation anomalies in Figs. 2–4, respectively, while EV for these 36 maps are listed in Table 4. Each set of maps is now discussed in turn.

a. 700-hPa heights

As already noted, a large majority of the major features in the panels of Fig. 2 were reproduced in separate composites made up of the first and second 10 yr of the J92 run. From January to June the only exception was the center of action of the area of positive height anomalies at high latitudes of Eurasia and the Arctic Ocean in January. For July and August the details of the anomalous flow over particularly the western Arctic Ocean and adjacent Alaska and Canada were unstable, but other features, including the contrast between the Bering Sea and Gulf of Alaska and the ocean farther south, were insensitive to the sample. Finally the location of the centers of action of the band of negative height anomalies across the north Atlantic Ocean into Europe on the November chart shifted substantially between the two samples. All in all the composites are remarkably stable.

The implication of the sensitivity analysis is that the model atmosphere displays SST-forced signatures throughout the year. Moreover, these signatures are prominent enough to explain from 20% to over 40% of the total monthly variance of 20 cases over the entire PNA region (Table 4).

From December–April the signal changes little and is clearly characteristic of Barnston and Livezey’s (1987) tropical–Northern Hemisphere (TNH) pattern. Livezey and Mo (1987) and others have shown that this is the dominant strong ENSO episode response over the PNA region for January through March. However, unlike in the model, little evidence exists for its occurrence in nature in December and April. The form of the stable part of the signal changes abruptly in May with the expansion of the positive part of the PNA-region signal to the northwest. This is followed by successive repositioning of the positive anomaly farther to the west, culminating in its location in the Bering Sea in August and September and its successive readvancement to its TNH pattern location in December. During the May–June part of the transition period, the SST-response shares many of the characteristics of Barnston and Livezey’s (1987) North Pacific (NP) pattern, which Bell and Janowiak (1995) have suggested can be related to ENSO in March through May.

Thus, the 700-hPa J92 composites reflect observed winter and spring equatorial SST-driven signals but with a wider seasonal window of TNH-pattern dominance and an accompanying delay of the appearance of the NP pattern in the spring. The high-latitude signal the rest of the year and its gradual transition to an early winter TNH pattern has not been previously indicated by observational analyses. One of the tasks of section 4 will be to reexamine this model result in the light of the observational composites.

b. U.S. surface temperature

Like the upper-air composites, the J92 monthly mean surface temperature composites in Fig. 3 exhibited extraordinary insensitivity to division of the sample into two parts. For example, the December through April anomaly patterns are all clearly related (as one would expect from Fig. 2) but have important differences. In particular, the position of the axis of the temperature anomaly ridge is systematically positioned farther west through this part of the sequence. These details, as well as all others shown on the December through July maps, are reproduced in both samples. On the other hand, the negative anomaly in the northwestern United States on the August composite is missing for one sample, as is the southeast to northwest gradient in November, while the focus of the maximum negative anomaly in October shifts from the far west to Lake Superior between the two samples.

The range of EV for the temperature composites is substantially larger than those for the upper-air heights (Table 4), varying from as little as 11% in November (perhaps the most unstable composite) to as high as 60% in April. Apart from some instability described above, the low values from July through November also reflect the location of the main center of action of the signal far to the northwest of the contiguous states. When the fact that the EVs are computed over a sample of 20 is taken into account, it is clear that all of the composites represent reproducible signal to some extent and that the model response is impressively consistent in over one-half of the months.

As for the patterns themselves, for one-half of the year (December–May) they are dominantly warmer than normal with a south to north gradient and straightforwardly emulate the pattern expected from a dominant TNH-pattern in the 700-hPA flow. Moreover, they are consistent with positive temperature anomalies in the Northwest found by HR92 for warm ENSO episodes. From July through October the maps exhibit mostly cold anomalies, but the distribution of these colder than normal conditions varies from month to month and in the two cases mentioned above is somewhat uncertain. These warm-season features previously have not been clearly noted in observational studies, nor should this have been expected because of the ENSO-episode emphasis of these studies.

c. U.S. precipitation

All of the important features of the precipitation anomaly composites in the 12 panels of Fig. 4 were reproduced in the two samples, including most of the smaller-scale centers. The explained variances (Table 4) are more uniform than for the other two composite sets already displayed and are mostly big enough to lend additional confidence that significant responses to large central equatorial Pacific SST anomalies have been identified in the model. Nevertheless, of the three variables examined, precipitation is the one that would least be expected to reflect the observational composites because of its heightened sensitivity to imperfections in either the model or the experimental design.

Thus, the presence of a year-round signal in Fig. 4 should be taken seriously, if not the details of its form. It should be noted though that the model clearly does reproduce the well-known cold-season, El Niño–related wetness in the southeast United States along the gulf and Atlantic coasts (RH96). Of some interest also is the presence of a very large scale positive anomaly of precipitation in the western United States in June. Bell and Janowiak (1995) have argued that the excessive rainfall in the spring of 1993 in the north-central and northwest United States that preceded the June–July floods was linked to anomalously warm central equatorial Pacific SSTs.

4. Observational composites

The discussion of the composites formed from the observations will be somewhat more involved than those for the two model-based composites for three reasons. First, results of significance tests and the selection procedures that were the basis of both the observed and 45Y composites must be described. This step was additionally complicated by the occasional use of ENSO-episodes based or 2- and 3-month window composites. Second, unlike in section 3, cold SST anomaly results must be examined in addition to the warm anomaly ones. Finally, more careful attention will have to be paid to details of the composites and their relationship to the results of others, in particular for precipitation and the recent work of MRL.

a. 700-hPa heights

The results of the composite formation and selection process described in section 2c(1) for monthly mean 700-hPa heights are summarized in Table 5, along with explained variances and the results for each map of the permutation tests described in section 2c(2). The latter are expressed as numbers of randomly formed composites of the same size that had greater percent of area in the PNA region locally significant at the 5% level. By binomial arguments the positive-SST anomaly April–December sequence of monthly mean composites was statistically significant at about the 0.1% level, so the full sequence was displayed in Fig. 5. This was not the case for the negative-SST anomaly monthly sequence, but note that spring and fall two- and three-month window composites, independent of the highly significant composites for the months of February and March, were highly significant in their own right. Consequently, the cold anomaly sequence is displayed as well in Fig. 7 after charts supplementing Fig. 5 in Fig. 6.

Before the features of the composites are discussed, their map significances (Table 5) should be compared to those for the ENSO-episode composites. While corresponding map significance results (and implied signal strengths) for SST-anomaly-based composites were not universally superior to the ENSO-episode composites, the emphasis on the former was clearly justified from the tables. This was especially obvious in almost all respects for the “cold” sequences and in the month to month consistency in the “warm” sequences. Hereafter, we will cite an ENSO-episode map significance only if it was particularly notable or if its corresponding composite was used to support an SST-anomaly-based composite. In general, the term “composite” will refer to one based on an SST-anomaly criterion.

Turning now to the composite maps themselves, the warm-SST anomaly sequence (Fig. 5) began (as expected from previous work) with a clear TNH signature that was notably consistent from January to March everywhere except for January to February over the Arctic Ocean, Europe, and the adjacent northeast Atlantic Ocean. Recall in section 3a that the January J92 composite was sensitive to its sample in the same area where the January–February inconsistency appears here.

From April–October (Fig. 5), the composites showed a consistent repositioning of the main anticyclonic anomaly over the extratropical PNA region, first to the west and then northwest into July and then southeast into October. This scenario was predicted by the J92 experiment but with a delayed start. The April composite reflected both the breakdown of the TNH pattern and a transition to the NP pattern emphasized for the spring by Bell and Janowiak (1995). The gross features of this composite were supported by the MA 2-month window and April ENSO-episode composites (not shown), both of which were field significant at the 5% level. The JA 2-month window and ENSO-episode composites likewise underpinned the key features of this northwest PNA-region scenario in August and September, respectively, when the monthly mean composites were otherwise in the noise level over the entire PNA region. Not only were the corroborative maps (not shown) field significant at 3% and 6% levels, respectively, the key features were locally significant as well.

The weakest part of the sequence occurs for the last two months of the year. The November composite appeared to have no temporal continuity with adjacent charts and no support from other composites, so no real case existed for a consistent November signal. The situation was not as bleak for the December composite because the DJ 2-month window monthly mean composite was considerably more field significant (2% level). This map reflected the beginning of the establishment of a TNH pattern after the apparent chaos of November (Fig. 6).

Overall, with both the theoretical support of the J92 experiment and the empirical evidence presented here, there is a substantial case for the existence of a high-latitude, Pacific basin teleconnection to central equatorial Pacific positive-SST anomalies in 11 months of the year. The practical implications of this finding can be assessed to some extent with the EVs (Table 5). On the average these were smaller than their counterparts (Table 4) for the J92 experiment, as expected, and varied widely month to month because of the small number of members in each composite and the change in this number from month to month. Nevertheless, they are large enough (around 15%–20% on the average) to impact long-range prediction (given accurate enough SST forecasts). Because the signal is highly localized (and the local explained variance considerably higher than the regional average) during much of the warm half of the year, this impact may potentially be substantial and important. Incidentally, explained variances for negative-SST anomaly composites have the same characteristics as those just discussed, and all conclusions based on these characteristics apply equally well to both sets.

In contrast to the positive-SST anomaly sequence, the window of the annual cycle with clear negative-SST anomaly teleconnections at 700 hPa was considerably narrower, specifically from October through March. The 2- and 3-month window composites suggested that echoes of part of the March pattern may extend into June. These teleconnections (Fig. 7) were particularly interesting from two more standpoints. First, the negative-SST anomaly teleconnections mirrored their positive-SST anomaly counterparts in only a very few respects; the SST-anomaly signals were not linear. Second, there appeared to be three distinct patterns for October–December, January–February, and March, respectively, with only part of the patterns sharing common characteristics in the December–January and February–March transitions. On the other hand, the extent of the common features shared by the October and December composites, which were not field significant in the PNA region, with the November composite, which is highly significant, lent considerable support to the signal content in all three maps. The same was true for January (not field significant) and February (highly significant).

The salient and common characteristics of the fall part of the sequence (Fig. 7) were the anomaly couplet in the northwestern part of the Pacific basin and a prominent positive anomaly in the North Atlantic between Greenland and the British Isles. This latter feature was replaced by a strong negative anomaly in January, which was sequentially repositioned to the southwest in February and March, with an accompanying growth of the band of positive anomalies to the south. In February this Atlantic couplet was phased very much like the North Atlantic Oscillation (NAO; Barnston and Livezey 1987). Over North America and the northwestern Pacific in January and February, the anomaly pattern most resembled a negative PNA pattern (Wallace and Gutzler 1981), which was practically in quadrature with the TNH pattern. This was not present at all in the March composite, leaving a north–south West Coast couplet mimicking that in the western Atlantic Ocean. Finally, note that in April there appeared to be a remnant of the Atlantic couplet.

For a smoothed perspective of the negative-SST anomaly cold season sequence, a partial sequence of two-month window monthly mean 700-hPa height anomaly composites is presented in Fig. 8. The contrasts between both the very similar ON and ND composites on one hand and the JF on the other, everywhere outside northwest North America and between the JF and MA composites outside the North Atlantic Ocean, are dramatic. Also note the similarity between AM and MJ composites in the vicinity of western North America and the reflection of the March signal in this area.

b. U.S. surface temperature

Table 6 contains analogous information to that listed in Table 5 but for monthly mean U.S. surface temperature composites. Unlike the 700-hPa composites, the collective statistical significance of the April through December negative-SST anomaly composites was more readily established than that for the positive anomaly composites. The former were at least significant at the 3% level by binomial arguments, but the latter (the positive-SST anomaly composites) required invocation of the three-month window monthly mean composites (significant at least at the 1% level), the map to map continuity, and the reproduction of behavior predicted by the J92 experiment to justify their display here. Despite the collectively stronger relative signal in the negative anomaly sequence, there was little difference between the two in their respective average EVs, which were both about 20%. As expected these averages were considerably smaller than their J92 counterparts but still substantial.

The first thing that was noted in the positive-SST anomaly sequence (Fig. 9) was the consistency between the temperature anomaly patterns for January–March; all three maps exhibited a mostly north–south contrast in temperature and southernmost extent of positive anomalies over the Great Basin and the north central Great Plains. These features were even more consistently represented in the three ENSO-episode composites (Figs. 10a–c). The form of the signal was sharpened further in the highly significant JFM-window composite (Fig. 10d). The wintertime contrast between the northwest and southeast (Figs. 9 and 10), as well as opposite-signed temperature anomalies in the Northwest for negative-SST anomaly composites (Fig. 12, later in this section), were all consistent with the findings of HR92.

From April through July there was month to month continuity in the shift of focus of warmer than normal temperatures to the northwestern United States and their eventual replacement by colder than normal temperatures. The April and July patterns were respectively reflected in the MAM-window (significant at the 6% level) and JA-window (significant at the 5% level) monthly mean composites (Figs. 11a,b). More importantly, the same spring to summer scenario was evident in the J92 sequence (Fig. 3) but lagged a month or so from the observational composites (Fig. 9).

After July the field significances for the positive-SST monthly mean anomaly composites (Table 6) were quite low (with virtually no signal in August), but through October their corresponding maps (Fig. 9) provided striking confirmation of the hypothetical scenario predicted by the J92 experiment (Fig. 3). For example, note the similar positioning of negative temperature anomalies in the western United States in the July and September maps and their relocation to the center of the country in October in both Figs. 3 and 9. The highly significant (4% level) JAS-window monthly mean composite (Fig. 11c) reflected the former feature well.

Interestingly, Barnston (1994) also identified a summertime SST-related signal that had its principal focus over the western United States. Barnston’s scenario leading to, for example, colder than normal conditions over the key area involved existence of tropical-basin-wide negative SST anomalies most of the previous year and the waning phase of a major ENSO cold episode. This does not dovetail well with the teleconnection argued here, unless there has been a systematic historic tendency for Barnston’s scenario to lead to rapid increases in SST in our key region in the central equatorial Pacific.

Lastly, note for the positive-SST anomaly composites that, like their 700-hPa counterparts, the November composite was extremely weak and the December composite lacks any coherent connection to adjacent months or the J92 run, although the DJ-window composite (Fig. 11d) did exhibit a statistically significant (5% level) compromise between transitional December and the heart of the cold season.

Because of a lack of model support, less could be inferred about negative-SST anomaly monthly mean temperature signals (Fig. 12). For example, there was little obvious continuity from map to map with mostly weak field significances (Table 6) for one-half of the year from May through October. The August pattern (the most significant of the six) was weakly reflected in September, and the July pattern (the next most field significant) was supported in the southwest by the JA-window monthly mean pattern (11% level; not shown).

Much more could be said about the negative-SST anomaly composites in the colder half of the year. Like their corresponding 700-hPa height composites (but to a lesser extent), the November and December maps had some features in common as did the January and February maps, but the two pairs were somewhat different from each other. The former pair also shared some features with the March and April composites; this was especially true of the December and April patterns.

Overall, there was little evidence of linearity in the U.S. temperature teleconnections to opposite signed central equatorial Pacific SST anomalies. In January, and more so in February, the temperature anomaly contrasts were more east–west for negative-SST anomalies (Fig. 12) than the north–south contrasts that dominate the same months for positive-SST anomalies (Fig. 9). The former were reminiscent of those associated with a negative PNA upper-air pattern, and the latter a TNH 700-hPa anomaly pattern. These differences were emphasized by comparison of the highly significant positive-SST anomaly JFM-window composite (Fig. 10d) and the likewise highly significant (3% level) negative-SST anomaly JF-window composite (Fig. 13). Only the March chart (Fig. 12) resembled the signature of a negative TNH pattern.

c. U.S. precipitation

As before, final selection criteria, field significances, and explained variances are reported for SST-anomaly and ENSO-episode-based composites of U.S. mean precipitation in Table 7. Likewise, the final sequences for positive- and negative-anomaly monthly mean composites are displayed in Figs. 14 and 16, respectively, while several composites to supplement these are displayed in Figs. 15 and 17.

Like the field significance results for U.S. monthly mean temperature, those for precipitation composites outside of the January–March period were so weak for positive-SST anomaly signals that 3-month window monthly mean precipitation composites (collectively significant at better than the 10% level) had to be cited to argue their existence. In dramatic contrast, the negative-SST anomaly composites had more overall field significance from April–December than any of the other monthly mean observed sequences presented here, including that for positive anomaly composites of 700-hPa heights (Table 5). Because there was such a strong contrast in collective significance between the positive- and negative-SST anomaly composites, there was a noticeable difference in their average EVs as well, with the former averaging about 18% and the latter 23%. Both of these averages were comparable to the average for the J92 experiment but, as expected, were somewhat more variable from month to month.

The description of the positive-SST anomaly signals (Fig. 14) for January–March was augmented by the January warm ENSO-episode (Fig. 15a; 2% field significant) and JFM-window positive-SST anomaly (Fig. 15b; 6% field significant) composites. The five charts collectively suggested that the contrast between coastal rainfall excesses and inland deficits in the southeastern United States, a tendency for above-normal precipitation in at least the southern part of the Great Plains, and the excess–deficit contrast between California and the Pacific Northwest, respectively, were all common responses to positive-SST anomalies in our key area of the equatorial Pacific during winter months. By April (Fig. 14) only a vestige of the California excess remained.

Over the remainder of the year a number of features of the positive-SST anomaly sequence were noted either because of recurrence from month to month or local significance. For example, a contrast between positive precipitation anomalies in Texas and negative ones in the central Mississippi and lower Missouri River valleys appeared in both the May and June composites and had considerable support in terms of both local and field significances (around 5%) from both the MJ-window (Fig. 15c) and AMJ-window (not shown) composites. There was likewise a tendency on all three composites from July to September for heavier than normal precipitation to the southwest of Lake Michigan and lighter than normal in the southeast, but these features had considerably less statistical underpinning than those already mentioned. It was tempting to associate the July composite pattern with the 1993 Midwest floods because it looked very much like the pattern of that July. However, July 1993 was one of the seven cases in the composite and the large positive rainfall anomaly was not locally significant. Finally, several features recurred at least twice in the October to December composites, appeared on both of the ON- and ND-window composites, and were locally significant on at least one of the latter pair (5% level; Figs. 15d,e): negative rainfall anomalies in the Pacific Northwest and positive anomalies in the southwest and lower Mississippi River valley.

It is important to place the above precipitation composite results in the perspective of previous and current work of others. For example, the results shown in Figs. 14 and 15 were entirely consistent with those of Andrade and Sellers (1988) for Arizona in the spring and fall and Ropelewski and Halpert (1986) and RH96 for the south-central, southern coastal, and northwestern areas of the United States in the cold half of the year.

However, our results failed to consistently reproduce a positive rainfall anomaly found by RH96 over the southern and central Great Basin in the warm half of the year, perhaps because monthly rainfall amounts are generally quite small for this region and season. It was only at the end of this period and into the fall that such an anomaly appeared in our composites first in the central Great Basin and later (as already noted for Arizona) for the southern part of the region. Additionally, the composites of RH96 (or those of Sittel 1994) did not reflect the wintertime inland negative precipitation anomaly in the southeast.

In contrast MRL, whose focus was limited to the United States east of the Rocky Mountains, reproduced virtually all of the features emphasized in Figs. 14 and 15 for the common domain. In their study the selection of cases for composite membership was based variously on the sign and amplitude of several different principal components of SST over the tropical Pacific basin. The results associated with three of these principal components are important for the discussion here and later. Two of them, the first unrotated and first varimax rotated modes, represent ENSO’s principal pattern of variability, while the third, the second rotated mode, is associated with modification of this main pattern.

For their domain of interest, MRL’s maps reflect all of the signals we have emphasized for January–March, the positive precipitation anomaly in the south-central Great Plains in May and June, the contrast in anomaly sign between the southeast and the area southwest of Lake Michigan in July–September, and the rainfall excesses in the south and central Mississippi River valleys in November and December.

A comparison of the four 2- and 3-month window composites in Fig. 15 with their monthly mean counterparts from the J92 experiment (Fig. 4) was also useful and revealed some important differences. For ON and ND all three locally significant observed features were predicted by the idealized model run but too far to the south, while the January–March J92 composites missed the prominent observed inland features and reversed the Pacific coast anomalies. Lastly, none of the important features of the observed MJ-window positive-SST anomaly composite were anticipated by the J92 experiment. These differences could not be attributed to systematic error of the model until the 45Y composites had also been compared to the observed composites because some of the differences may have been a consequence of meaningful differences (at least for the precipitation signal) in the composite global SST fields that corresponded, respectively, to the observational and J92 composites.

The negative-SST anomaly composites (Fig. 16) contain many prominent and recurrent or locally significant features. For example, from November–February a pattern characterized by drier than normal conditions along coastal areas in the southeastern United States with wetter than normal conditions inland was only partially broken up in December. In the center of the country and to some extent in the southwest both prior to (in September and October) and subsequent to (in March and April and to a lesser extent in May) this winter period, negative precipitation anomalies dominated the composites. This feature, as well as the strong negative rainfall anomalies in the upper Midwest in June and July, lent substantial support to a contributing role for a developing cold-ENSO episode in the great Midwestern drought of spring and early summer 1988. From May to July the maps showed mainly positive rainfall anomalies in the southeast.

In the western United States three distinct rainfall regimes were apparent in the negative-SST anomaly composites. Along the West Coast in January and February departures from normal rainfall were substantially more negative in California than to the north in eastern Oregon and Washington. Before (October–December) and after (March) this, all but the southernmost end of Pacific coastal regions were covered by positive precipitation anomalies. The rest of the year the western United States was dominated by composite rainfall deficits, except for a break in July, the only month in the entire negative-SST anomaly sequence not field significant at least at the 20% level.

The results just described for negative-SST anomaly composites are generally consistent with Andrade and Sellers (1988) and RH96 for the times of the year these studies identified signals. This is also true for all times of the year for the composites of MRL, but the correspondences are more complicated than for positive-SST anomaly composites. Specifically, from November to March our negative-SST anomaly results match MRL’s composites based on their first unrotated tropical Pacific SST mode, but they mainly match those based on their second rotated mode the remainder of the year.

In fact, all of the important features for April–October in Fig. 16, including the prominent and recurrent negative precipitation anomalies in the north-central or central United States, are reproduced in a combination of MRL’s first and second rotated mode composites. Additionally, their August composite for the second rotated mode provides a consistent bridge between our July and September maps, which our August composite failed to do. The results suggest that the ENSO structure that modulates SST anomalies in our critical central equatorial Pacific area is somewhat different for warm-season cold episodes than otherwise.

Incidentally, MRL’s cold episode results also enhance those here for the winter by extension of the January and February dipole anomaly pattern in the southeast (Fig. 16) to March, while clearly reflecting the onset of drier than normal conditions for the central and south-central United States in the spring.

Unlike analyses of 700-hPa height and U.S. temperature anomalies described earlier, corresponding composite maps for positive- and negative-SST anomalies (Figs. 14 and 16 respectively) exhibited some linearity, especially from September through February. Keep in mind, however, that the nature of our compositing process implies a tendency for opposite-appearing patterns, so that it is important for there also to be some local statistical significance on corresponding features of both negative- and positive-SST anomaly composites to support a linear response. In this context the strongest cases for linearity could be made for November and February. This linearity is highlighted by comparisons of the negative-SST anomaly JF-window (Fig. 17a; 3% level) and OND-window (Fig. 17b; 0.1% level) composites to those for the positive-SST anomaly JFM-window (Fig. 15b) and ON- and ND-windows (Figs. 15d,e), respectively.

The observed composites just described will now be compared to 45Y experiment composites formed from exactly corresponding cases.

5. 45Y experiment composites

The availability of the 45Y experiment permitted the examination of several additional aspects of the SST anomaly signal. First, for at least the case of positive-SST anomalies, it was possible to check the relevance of the J92 experiment to the observational composites, that is, the extent to which differences between the J92 global SSTs and those for composited cases determined the salient characteristics of the J92 experiment signals. Next, the characteristic errors of the 45Y experiment composites in reproducing the observational composites can be determined through the use of the error decomposition (6) described in section 2c(3). Finally, because the 45Y experiment composites were made up of multiple realizations of each case, it was possible to assess the relative importance of case to case differences in global SSTs and internal variability in the degradation of EVs through the partitioning of variances summarized in (3). This discussion is deferred to section 5d. Otherwise complete sets of field significances and EVs for the 45Y experiment are listed in Table 8.

a. 700-hPa heights

The sequence of positive-SST anomaly composites for the 45Y experiment (Fig. 18) had much in common with both the J92 experiment (Fig. 2) and corresponding observed (Fig. 5) composites.

The differences in placement of previously highlighted features between the two sets of model experiment composites were actually quite minor. April composites were somewhat different; the earlier breakdown of the TNH pattern in the 45Y experiment was more in accordance with the observations (Fig. 5). The only other notable difference was that the 45Y experiment anomaly centers over the northern Pacific Ocean and northwestern North America led those of the J92 experiment in June and September. Thus, except for April, the gross pattern (if not the amplitudes) of the J92 SST anomalies was representative overall of the average forcing for composited cases, at least in terms of modeled 700-hPa height responses. However, the facts that anomalies were mostly stronger and EVs for all variables (Table 8) were uniformly much larger in the J92 experiment than their counterparts in the 45Y experiment provided evidence that at least differences in average strength of forcing were important.

The visual impression of good correspondences between 45Y experiment (Fig. 18) and observed (Fig. 5) positive-SST anomaly composites was confirmed by the results of the skill decomposition (Fig. 19a). Average anomaly correlations ρFX for the PNA region were quite high for all but two months (May and December), notable in view of the fact that hypothesized signals covered only a small part of the domain in many months. Phase errors reflected in the anomaly correlations generally consisted of upstream displacement of features by the model from January through May, with the displacement somewhat small from February to April, a bit larger in January, and quite large in May. By June the observational composite features caught up to those for the model, and up to November these features were phased well on the corresponding maps. Premature establishment of the TNH pattern by the model and downstream phase errors characterized the last two composites in the sequence.

Most of the higher anomaly correlations translated into positive skill scores β. Exceptions were for August to October because of large positive unconditional map biases C, which also seriously degraded the score for July. Smaller but substantial positive unconditional biases were present in seven out of the eight remaining model composites.

Here 10 out of 12 of the 45Y experiment positive-SST anomaly composites were field significant at least at the 1% level, with the remaining two (May and June) no worse than 20% (Table 8). In contrast, corresponding observed composites (Table 5) were generally not field significant at any acceptable level from August to the end of the year. Thus, confirmation of features of the observed composites, particularly for August–October, by highly significant 45Y experiment counterparts was especially important.

Anomaly correlations (Fig. 19b) between 45Y experiment and observational composites for negative-SST anomalies (Figs. 20 and 7, respectively) were almost all lower than the corresponding positive-SST anomaly maps from March to October. In fact, there was really little to note in either sequence of maps from June to September. On the other hand, the much higher combined overall ρFX in November and December for the negative-SST anomaly composites reflected an important result: the 45Y experiment confirmed the existence of the late fall–early winter high-latitude observed negative-SST signal. Three of the four observed composite anomaly centers were clearly reproduced by the model with a modest downstream phase error, and the fourth was hinted at in the November 45Y experiment map and weakly present with a larger downstream position error in the December chart. Unlike the observations, the model failed to show the initial development of the pattern in October.

In winter and spring months the 45Y experiment continued to exhibit downstream position errors as well as a failure to reproduce well the observed features over the North Atlantic. Specifically, the model composites contained a TNH rather than PNA/NAO signature in January and February and missed completely the NAO/western Atlantic centers in March and April. Moreover, the model was much slower than the observations in diminishing features over North America, retaining them in the composites into May in contrast to their disappearance from the observational composites by March. This led to a huge amplitude error B2 (not shown) and very negative β (Fig. 19b) in April.

The quality of the model reproduction of the observed composites in November and December were not reflected in the skill scores, mainly because of large negative C for both matches. These negative unconditional biases were present in nine out of 10 of the remaining negative-SST anomaly model composites (March is the exception; Fig. 19b) and were larger in six of them (dominantly summer to fall). Thus, PNA-region 700-hPa biases, with the same sign as the SST anomalies (whether positive or negative) on which the composite was based, exist in the model for every month except March (when they were oppositely signed with modest magnitude).

To once more underline the importance of the 45Y experiment negative-SST anomaly 700-hPa height composites for November and December, it should be noted (Table 8) that both were field significant at least at the 0.001 level. Overall field significances crudely exhibited the same seasonality as those for observational composites.

b. U.S. surface temperature

As might have been anticipated from 700-hPa results, the 45Y experiment positive-SST anomaly U.S. surface temperature anomaly composite sequence (Fig. 21) shared many features with J92 experiment (Fig. 3) and observed (Fig. 9) sequences. The 45Y experiment generally was in closer conformity with observations than the J92 experiment. Differences in the two model-based sequences included more westward positioning of negative temperature anomalies from July to September, the opposite in October, and a considerably weaker signal in November in the 45Y run. In four out of these five months the J92 experimental composites showed some degree of sensitivity to the sample. Anomaly amplitudes in the 45Y run were moderately smaller than in the J92 experiment composites, in conformity with the results for 700-hPa height anomalies.

Anomaly correlations between the positive-SST anomaly 45Y run and observational composites (Fig. 22a) were exceptionally high from January to April and for September but by themselves masked notable correspondence in the scenarios for the warm half of the year. On the other hand very large unconditional warm biases degraded the β score substantially in March and April. The principal cause of this bias in the model, the extension of positive temperature anomalies too far south in the predominantly zonally oriented pattern, was also present to a lesser extent in the January and February composites. Half of the remaining eight composites also had positive unconditional biases but mostly of a different character.

As suggested above, the model scenario from May to October (Fig. 21) mutually reinforced the observational sequence (Fig. 9) far more than either the ρFXs and βs (Fig. 22a) would suggest. The observations overall reflected a decline of the warmer than normal temperatures accompanied by replacement with colder than normal temperatures—first in the southwest, then over most of the west, with a final shift of focus to the east at the end of the period. This was precisely the model’s scenario. In the 45Y run positive-SST anomaly composites, the negative temperature anomalies were established too early in the southwest in May and June. The June and July scores were further degraded because of failure of cold anomalies in the east to match up between model and observations (the June observational feature was not locally significant). For August the enormous errors were a consequence of a model composite entirely consistent with those preceding and following it being compared to an observational composite with almost no character. The poor October score was mainly the result of a poor match in the southwest [the large C was more than compensated by D (not shown)]. Interestingly, the absence of any real signal in the November model composite strongly corroborated the same observational result.

As usual field significances for the 45Y run surface temperature composites (Table 8) were generally strong, both for positive- and negative-SST anomalies, and in both cases reflected semiannual variability crudely present in corresponding observational field significances (Table 6).

In view of the results for 700-hPa height anomaly composites, it came as no surprise that matchups between negative-SST anomaly U.S. temperature composites for the 45Y experiment (Fig. 23) and the observations (Fig. 12) were less successful than those just described. The various systematic phase errors and difficulties in reproducing features over the North Atlantic Ocean in the model height anomaly composites inevitably must have led to systematic errors in the temperature anomaly composites. The anomaly correlations ρFX (Fig. 22b) reflected this; only the one for the March negative-SST anomaly matchup was better than mediocre.

For all practical purposes, the model negative-SST anomaly composite response all the way from December through April was the opposite of the model positive-SST anomaly composite signal (matching up well with observations only in March). The observational contrast between the January and February patterns on the one hand and the family of four patterns flanking them in the sequence (Fig. 12) on the other, as well as the observed nonlinearity between positive- and negative-SST anomaly composites, were completely missing.

In view of this, it is somewhat surprising that Hoerling et al. (1997) simulated this nonlinearity with integrations of the same model used here. However, the SST fields used to drive their integrations represented composite ENSO-episode scenarios, which, besides being fixed from run to run, differed from the collective scenarios used here in two other ways. First, in Hoerling et al. (1997) SST anomalies were only prescribed for the tropical Pacific and were set to zero elsewhere. Additionally, cold-episode anomalies were set as exactly the inverse of warm-episode anomalies. Further tests have so far ruled out only the difference in domains for prescribing SST anomalies as the source of the difference in outcomes.

While there was substantial reason to presume the gross differences in cold-season, negative-SST anomaly, temperature composites were mainly the result of model errors, those for the warm season could not be so easily ascribed to deficiencies in either the model or the observations. Statistically significant negative temperature anomalies in the southwest and west on observed composites for July–September were emphatically contradicted by opposite-signed, highly significant model features. The latter again represented linear model responses to SST anomalies not noted in the observations. This suggests caution in the use of either the observed or modeled summertime patterns.

Incidentally, the linearity exhibited by the model was also reflected in unconditional temperature biases (Fig. 22b). Like the positive-SST anomaly composites, 8 of 12 of the negative-SST had unconditional biases of the same sign as the SST anomaly and most of them were for months in the cold half of the year. This repeated (in a physically consistent way) the pattern of unconditional biases noted for model 700-hPa height anomaly composites. This prominent systematic error is currently under investigation because it has been noted as well in operational model forecasts.

c. U.S. precipitation

Apart from expected amplitude differences (see previous discussions), the two model sequences of positive-SST anomaly U.S. precipitation composites (Figs. 4 and 24) were more dissimilar than alike. However, the two model signals were similarly phased for October through December, and the 45Y experiment results were closer to observational composites (Fig. 14) than those for the J92 run for January–April. This was fortunate because there was an a priori expectation for a SST-forced response during the cold part of the year. During this seven-month period, anomaly correlations between 45Y run and observational composites were uniformly positive, but mostly small except for moderate values for November, January, and February (Fig. 25a). The three 45Y experiment maps for these months (which were all field significant at better than the 1% level; Table 8) corroborated the locations of almost all of the observational signals highlighted for the cold half of the year for positive-SST anomalies. In April, another highly significant 45Y run composite matched the positions of observational anomalies well in the southeastern third of the United States. Generally, 45Y experiment composite anomalies were much weaker than those in observational composites; the contour interval in Figs. 14 and 16 is twice that in Figs. 24 and 26.

For May and June the corresponding model composites were largely alike but did not reflect the locally and field-significant signal in the central United States in the observations (Fig. 15c). Thus, the failure of the J92 experiment to capture these features could not be attributed to the SST field being nonrepresentative. However, the similarity of the model results constituted a strong case for the presence of an SST-forced signal in the observations. Given this signal’s strength (Fig. 15c and Table 6), we believe it should be taken seriously and that the models were in systematic error.

In contrast, for all three months from July to September there was little correspondence between the two sets of model composites or between either one and the observational set. This reinforced our skepticism about the presence of a positive-SST anomaly forced signal in the U.S. precipitation field in these months.

Finally, the 45Y experiment reproduced (Fig. 26) many of the features noted on the negative-SST anomaly observed composites (Fig. 14), although none of them particularly well. The model sequence actually had poorer overall field significance (Table 8) than the observational sequence (Table 7), mainly from composites for the first half of the year. This was not the case for any other model–observation composite sequence pairs.

There were three months with moderate ρFX (Fig. 25b): November, December, and June. For the first two of these the model actually reproduced the observed composites well except for failure to spread negative precipitation anomalies along the Gulf Coast in November and incorrectly extending positive ones over the entire Great Basin and northern Rocky Mountain region. The biggest error in the model’s June composite was the large westward displacement of the center of the important negative precipitation anomaly observed over the north-central United States. This upstream relocation of the principal central United States rainfall deficit was a consistent feature of model composites from at least March through July. The model again exhibited a very consistent, but quite different, systematic error pattern for the August–October period. Composite negative-SST anomaly charts for the 45Y experiment for these three months displayed a strong pattern that was dominantly wet and bore little resemblance to corresponding observed composites. The month to month consistency of these model errors for both periods making up the warm two-thirds of the year further reinforced our confidence in the presence of negative-SST anomaly signals throughout the period.

In the first two months of the year, the model only reproduced some aspects of the observed composite in the east in January and in the west in February. After December, little indication of observed excess rainfall in the northwest was present in the model negative-SST anomaly composites. Because this feature was prominently present in five out of six observed composites from October through March, we were confident the difference in observed and modeled composites in the latter half of the period was the result of a model deficiency rather than sampling problems.

d. Analysis of variance

The results of application of the analyses of variance (2) and (3) for positive-SST anomaly composites were largely representative of those for negative SST, so only the former will be discussed. The huge differences between the pooled composite signal to noise ratios VC /V for the two model experiments (Table 9) were largely the result of a single global SST field in the J92 experiment and the partitioning of SST-forced “signal” in the 45Y experiment between VC and VS. We are confident that this is the case because the SST anomaly for our key area in January 1992 was equalled or exceeded in only 14 of the 95 months listed in Table 1 for positive-SST anomalies. Livezey and Mo (1987) found (see their Fig. 9b) a strong linear relationship for Januarys and Februarys between positive-SST anomalies in an area to the east of the area used here and the amplitude of the TNH pattern.

These observations can be coupled to the fact that, for the 45Y experiment, the root-mean standard differences between case ensemble averages and composite averages (the square root of VS) were about the same (for 700-hPa heights and temperatures) or greater than (for precipitation) the composite amplitudes (Table 9). Thus, J92 experiment composite signals one or more standard differences greater than those for the 45Y experiment would have signal to noise ratios at least four times greater with comparable variances V.

These arguments reinforce our view that amplitude differences were the most important components of case to case SST-field differences to case-average differences. Kumar and Hoerling (1997) reached the same conclusion with anomaly correlation analyses performed separately on the observations and the 45Y experiment. One other factor that may have played a smaller role in the deflation of the 45Y run signal with respect to the J92 signal was the transiency of the forcing in the former and the stationarity in the latter. Finally, for the two model experiments, note the inconsistent differences from variable to variable between variances V, even after discounting VSs in each case.

Two of the most interesting features of observational analyses of variance (middle column of Table 9) were the consistency of signal to noise ratios over all three sets of composites and the relative success of the three sets in representing their members compared to those for the 45Y experiment. These advantages were the result of a much smaller noise level V and comparable signal VC for 700-hPa heights, somewhat smaller V and somewhat larger VC for temperatures, and comparable V and much larger VC for precipitation. The latter for the 45Y run composites was only one-fifth of precipitation VC for observational composites and one-tenth of that for J92 run composites. Except for precipitation, the model had unrealistically large variability about the composite averages. Fortunately, this can be largely removed by averaging of case realizations.

On a final note, the results of partitioning of variance of composite memberships for the 45Y experiment clearly indicated that fractional portions attributable to case to case differences in SSTs were comparable to those for the squared composites themselves and of the same order of magnitude as the stronger observational composites. An optimistic view is that these differences can be accounted for by a combination of simple observational relationships, for example, separate linear fits of positive- and negative-SST anomalies in the key area to teleconnected pattern amplitude and ensemble GCM integrations corrected for systematic errors.

6. Concluding remarks

Teleconnections between moderate to large SST anomalies in the central equatorial Pacific Ocean and monthly mean Northern Hemisphere 700-hPa heights and U.S. surface temperature and precipitation have been examined through use of separate sets of composite maps for positive and negative SST anomalies. A considerable mass of observational and model evidence has been presented that confirmed previously described relationships, refined their description, and established new teleconnections. Highlights of our findings are as follows.

  • Almost year-round teleconnections were found between all three fields and positive-SST anomalies. Confident signals could not be identified only in prewinter for 700-hPa heights and U.S. surface temperatures and in late summer for U.S. precipitation.
  • Distinct teleconnections were found in prewinter, midwinter, and prespring between negative-SST anomalies and both 700-hPa heights and U.S. surface temperatures. Year-round teleconnections were identified between U.S. precipitation and negative-SST anomalies.
  • Teleconnections from prewinter through prespring for 700-hPa heights and U.S. surface temperatures exhibited substantial nonlinearity between positive- and negative-SST anomalies, while those for U.S. precipitation were substantially linear.

Implications of these findings for monthly and seasonal prediction are twofold: 1) new opportunities exist for successful forecasts of the three fields examined here given high quality tropical Pacific SST predictions, and 2) this success will be compromised unless the nonlinearity of the teleconnections is taken into account.

The composite analyses were initially motivated by results from the J92 GCM experiment and were substantially strengthened by the opportunity to examine composites from this fixed-SST experiment and the suite of variable-SST 45Y GCM runs. Three-way comparisons between model-based and observational composites revealed a number of things of interest. For example, because phase similarities between the J92 and 45Y experiment composites far outweighed their dissimilarities, aside from amplitude, J92 results were quite representative of model teleconnections to positive-SST anomalies. A satisfying result was that in most instances that corresponding model composites differed, those for the 45Y experiment matched the observational composites better. This suggested that the model had some degree of realistic sensitivity to differences in the SST field, at least in those instances when the Pacific SST anomaly was moderately large and positive. This issue will be revisited below.

The visual and quantitative comparisons between the 45Y experiment and observational composites uncovered major systematic errors in the GCM. Besides a number of differences in timing of transitions and positioning of centers, other very important model errors included the following:

  • physically consistent overall biases of the same sign as the SST anomalies in both 700-hPa heights and U.S. surface temperatures;
  • overly linear response of both 700-hPa heights and U.S. surface temperatures in the winter to opposite-signed SST anomalies;
  • considerably less success in reproduction of negative-SST anomaly observed teleconnections than of those for positive anomalies. This was the case despite the fact that there was more overall confidence attached to the former for the cold one-half of the year in the observational composites.

For the GCM to have optimum impact on monthly and seasonal prediction, these errors must be corrected, perhaps through a combination of model changes and application of nonlinear statistical postprocessing. Many of the errors noted earlier were gross enough that even linear statistical corrections would have considerable beneficial effect. Recently, completed work, which will be reported on later, has confirmed this and work is under way to explore nonlinear model postprocessing. The strength of the signals noted here lends confidence that this effort will be successful.

A final 45Y experiment result important to this discussion was the demonstration through analysis of variance that responses to differences in the SST field were comparable in amplitude to that for the average response (i.e., the composite). The result was corroborated by the differences in composite amplitudes between the J92 and 45Y experiments and other evidence. From both of these exercises the following is clear:

  • Case to case differences in forcing within a composite lead to substantive differences in response, both in amplitude and form, but mostly in the former.

Once again, the implication of this finding for prediction is that these more subtle (than the sign of the SST anomaly in one area) differences in forcing need to be accounted for as well. Empirical approaches to this problem will be hampered by the limited available sample with which to work. Nevertheless, we intend to explore more discriminating case selection techniques where simple field to field relationships will be explored within subsamples of the complete dataset. Appropriately designed GCM experiments may suggest directions that should be pursued for this effort. The difficulties involved here make the use of the models, because of their sensitivity to case to case differences in forcing, even more attractive for prediction if suitable postprocessed corrections for systematic errors can be made. The analyses of variance suggest that success will yield large relative gains in prediction skill.

As a last comment here, the collective results presented in sections 3a, 4a, and 5a on Northern Hemisphere extratropical 700-hPa heights imply that nonlinear teleconnections to equatorial Pacific SSTs must exist for surface temperature and precipitation over Alaska, Canada, and some parts of high-latitude Eurasia for some months. Obviously, failure to treat this nonlinearity has hampered previous efforts to identify such relationships. We intend to pursue analyses for these regions following our exploration of more sophisticated case selection.

Acknowledgments

The final version of the manuscript benefited substantially from reviews provided by Gerry Bell, Chet Ropelewski, Ed Schneider, Tom Smith, and an anonymous reviewer. The multiple-month window analyses that proved enormously useful in bolstering the credibility of many new results were done at the urging of Mike Wallace. Likewise, conversations with Dave Montroy and his associates have mutually strengthened our work. Finally, M. M. and H. R. have been partially supported by NOAA’s Office of Global Programs.

REFERENCES

  • Andrade, E. R., Jr., and W. D. Sellers, 1988: El Niño and its effect on precipitation in Arizona and western New Mexico. J. Climatol.,8, 403–410.

  • Barnett, T. P., and R. Preisendorfer, 1987: Origins and levels of monthly and seasonal forecast skill for United States surface air temperatures determined by canonical correlation analysis. Mon. Wea. Rev.,115, 1825–1850.

  • Barnston, A. G., 1994: Linear statistical short-term climate predictive skill in the Northern Hemisphere. J. Climate,7, 1513–1564.

  • ——, and R. E. Livezey, 1987: Classification, seasonality and persistence of low-frequency atmospheric circulation patterns. Mon. Wea. Rev.,115, 1083–1126.

  • ——, ——, and M. S. Halpert, 1991: Modulation of Southern Oscillation–Northern Hemisphere mid-winter climate relationships by the QBO. J. Climate,4, 203–217.

  • Bell, G. D., and J. E. Janowiak, 1995: Atmospheric circulation associated with the Midwest floods of 1993. Bull. Amer. Meteor. Soc.,76, 681–695.

  • Fu, C., H. F. Diaz, and J. O. Fletcher, 1986: Characteristics of the response of sea surface temperature in the central Pacific associated with warm episodes of the Southern Oscillation. Mon. Wea. Rev.,114, 1716–1738.

  • Halpert, M. S., and C. F. Ropelewski, 1992: Surface temperature patterns associated with the Southern Oscillation. J. Climate,5, 577–593.

  • Hoerling, M. P., A. Kumar, and M. Zhong, 1997: El Niño, La Niña and the nonlinearity of their teleconnections. J. Climate,10, 1769–1786.

  • Ji, M., A. Kumar, and A. Leetmaa, 1994: A multiseason climate forecast system at the National Meteorological Center. Bull. Amer. Meteor. Soc.,75, 569–577.

  • Kousky, V. E., 1993: The global climate of December 1991–February 1992: Mature-phase warm (ENSO) episode conditions develop. J. Climate,6, 1639–1655.

  • Kumar, A., and M. P. Hoerling, 1995: Prospects and limitations of seasonal atmospheric GCM predictions. Bull. Amer. Meteor. Soc.,76, 335–345.

  • ——, and ——, 1997: Interpretation and implications of the observed inter–El Niño variability. J. Climate,10, 83–91.

  • ——, A. Leetmaa, and M. Ji, 1994: Simulations of atmospheric variability induced by sea surface temperatures and implications for global warming. Science,266, 632–634.

  • Livezey, R. E., 1995: The evaluation of forecasts. Analysis of Climate Variability: Applications of Statistical Techniques, H. von Storch and A. Navarra, Eds., Springer-Verlag, 177–196.

  • ——, and W. Chen, 1983: Statistical field significance and its determination by Monte Carlo methods. Mon. Wea. Rev.,111, 46–59.

  • ——, and K. C. Mo, 1987: Tropical–extratropical teleconnections during the Northern Hemisphere winter. Part II: Relationships between monthly mean Northern Hemisphere circulation patterns and proxies for tropical convection. Mon. Wea. Rev.,115, 3115–3132.

  • ——, J. D. Hoopingarner, and J. Huang, 1995: Verification of official monthly mean 700-hPa height forecasts: An update. Wea. Forecasting,10, 512–527.

  • ——, M. Masutani, and M. Ji, 1996: SST-forced seasonal simulation and prediction skill for versions of the NCEP/MRF model. Bull. Amer. Meteor. Soc.,77, 507–517.

  • Murphy, A. H., and E. S. Epstein, 1989: Skill scores and correlation coefficients in model verification. Mon. Wea. Rev.,117, 572–581.

  • Parker, D. E., P. D. Jones, C. K. Folland, and A. Bevan, 1994: Interdecadal changes of surface temperature since the late nineteenth century. J. Geophys. Res.,99, 14 373–14 399.

  • Philander, G. S., 1990: El Niño, La Niña, and the Southern Oscillation. Academic Press, 289 pp.

  • Reynolds, R. W., 1988: A real-time global sea surface temperature analysis. J. Climate,1, 75–86.

  • Ropelewski, C. F., and M. S. Halpert, 1986: North American precipitation and temperature patterns associated with the El Niño/Southern Oscillation (ENSO). Mon. Wea. Rev.,114, 2352–2362.

  • ——, and ——, 1989: Precipitation patterns associated with the high index phase of the Southern Oscillation. J. Climate,2, 268–284.

  • ——, and ——, 1996: Quantifying Southern Oscillation–precipitation relationships. J. Climate,9, 1043–1059.

  • Rowell, D. P., C. K. Folland, K. Maskell, and M. N. Ward, 1995: Variability of summer rainfall over tropical North Africa (1906–92): Observations and modeling. Quart. J. Roy. Meteor. Soc.,121, 669–704.

  • Sittel, M. C., 1994: Differences in the means of ENSO extremes for maximum temperature and precipitation in the United States. Center for Ocean–Atmospheric Prediction Studies Tech. Rep. 94-2, 50 pp. [Available from Dr. James J. O’Brien, The Florida State University, Tallahassee, FL 32306.].

  • Smith, T. M., R. W. Reynolds, R. E. Livezey, and D. C. Stokes, 1996: Reconstruction of historical sea surface temperatures using empirical orthogonal functions. J. Climate,9, 1403–1420.

  • Wallace, J. M., and D. Gutzler, 1981: Teleconnections in geopotential height field during the Northern Hemisphere winter. Mon. Wea. Rev.,109, 784–812.

Fig. 1.
Fig. 1.

Departures from 1950–79 normals (Reynolds 1988) of global SSTs for January 1992 used to force the J92 experiment. Contour interval is 0.5°C and negative anomalies are shaded.

Citation: Journal of Climate 10, 8; 10.1175/1520-0442(1997)010<1787:TROTPN>2.0.CO;2

Fig. 2.
Fig. 2.

Monthly mean 700-hPa height anomaly composites for the J92 experiment. Contour interval is 20 m, the zero contour is dotted, and negative contours are dashed.

Citation: Journal of Climate 10, 8; 10.1175/1520-0442(1997)010<1787:TROTPN>2.0.CO;2

Fig. 3.
Fig. 3.

Monthly mean U.S. surface temperature anomaly composites for the J92 experiment. Contour interval is 0.5°C, the zero contour is omitted, and negative contours are dashed.

Citation: Journal of Climate 10, 8; 10.1175/1520-0442(1997)010<1787:TROTPN>2.0.CO;2

Fig. 4.
Fig. 4.

Monthly mean U.S. precipitation composites for the J92 experiment. Contour interval is 10 mm, the zero contour is omitted, and negative contours are dashed.

Citation: Journal of Climate 10, 8; 10.1175/1520-0442(1997)010<1787:TROTPN>2.0.CO;2

Fig. 5.
Fig. 5.

Observed monthly mean 700-hPa height anomaly composites for cases with average positive SST anomalies in the central equatorial Pacific (area described in the text) that exceed the criterion (in degrees Celsius times 10) in the lower-right corner of each panel. Contour interval is 10 m, the zero contour is dotted, and negative contours are dashed. Light and dark shading denote 10% and 5% local statistical significance, respectively.

Citation: Journal of Climate 10, 8; 10.1175/1520-0442(1997)010<1787:TROTPN>2.0.CO;2

Fig. 6.
Fig. 6.

Same as Fig. 5 except for the DJ-window composite.

Citation: Journal of Climate 10, 8; 10.1175/1520-0442(1997)010<1787:TROTPN>2.0.CO;2

Fig. 7.
Fig. 7.

Same as Fig. 5 except for negative SST anomalies.

Citation: Journal of Climate 10, 8; 10.1175/1520-0442(1997)010<1787:TROTPN>2.0.CO;2

Fig. 8.
Fig. 8.

Same as Fig. 5 except for ON- through MJ-window composites and negative SST anomalies.

Citation: Journal of Climate 10, 8; 10.1175/1520-0442(1997)010<1787:TROTPN>2.0.CO;2

Fig. 9.
Fig. 9.

Same as Fig. 5 except for U.S. surface temperature composites, contour interval of 0.5°C, and no-zero contour.

Citation: Journal of Climate 10, 8; 10.1175/1520-0442(1997)010<1787:TROTPN>2.0.CO;2

Fig. 10.
Fig. 10.

Same as Fig. 9 except for (a)–(c) January–March warm ENSO episode composites and (d) the JFM-window composite.

Citation: Journal of Climate 10, 8; 10.1175/1520-0442(1997)010<1787:TROTPN>2.0.CO;2

Fig. 11.
Fig. 11.

Same as Fig. 9 except for (a) MAM-, (b) JA-, (c) JAS-, and (d) DJ-window composites.

Citation: Journal of Climate 10, 8; 10.1175/1520-0442(1997)010<1787:TROTPN>2.0.CO;2

Fig. 12.
Fig. 12.

Same as Fig. 9 except for negative-SST anomalies.

Citation: Journal of Climate 10, 8; 10.1175/1520-0442(1997)010<1787:TROTPN>2.0.CO;2

Fig. 13.
Fig. 13.

Same as Fig. 9 except for the JF-window composite and negative SST anomalies.

Citation: Journal of Climate 10, 8; 10.1175/1520-0442(1997)010<1787:TROTPN>2.0.CO;2

Fig. 14.
Fig. 14.

Same as Fig. 5 except for U.S. precipitation, contour interval of 10 mm, and no zero contour.

Citation: Journal of Climate 10, 8; 10.1175/1520-0442(1997)010<1787:TROTPN>2.0.CO;2

Fig. 15.
Fig. 15.

Same as Fig. 14 except for (a) the January warm ENSO episode and (b) JFM-, (c) MJ-, (d) ON-, and (e) ND-window composites.

Citation: Journal of Climate 10, 8; 10.1175/1520-0442(1997)010<1787:TROTPN>2.0.CO;2

Fig. 16.
Fig. 16.

Same as Fig. 14 except for negative SST anomalies.

Citation: Journal of Climate 10, 8; 10.1175/1520-0442(1997)010<1787:TROTPN>2.0.CO;2

Fig. 17.
Fig. 17.

Same as Fig. 14 except for (a) JF- and (b) OND-window composites and negative SST anomalies.

Citation: Journal of Climate 10, 8; 10.1175/1520-0442(1997)010<1787:TROTPN>2.0.CO;2

Fig. 18.
Fig. 18.

Same as Fig. 5 except for 45Y experiment composites.

Citation: Journal of Climate 10, 8; 10.1175/1520-0442(1997)010<1787:TROTPN>2.0.CO;2

Fig. 19.
Fig. 19.

Monthly dependence of measures of quality and skill of the match between corresponding observational and 45Y experiment 700-hPa height composites for both (a) positive- and (b) negative-SST anomalies. The solid, dashed, and dotted curves are for anomaly correlation ρFX, unconditional bias C, and skill score β, respectively.

Citation: Journal of Climate 10, 8; 10.1175/1520-0442(1997)010<1787:TROTPN>2.0.CO;2

Fig. 20.
Fig. 20.

Same as Fig. 7 except for 45Y experiment composites.

Citation: Journal of Climate 10, 8; 10.1175/1520-0442(1997)010<1787:TROTPN>2.0.CO;2

Fig. 21.
Fig. 21.

Same as Fig. 9 except for 45Y experiment composites.

Citation: Journal of Climate 10, 8; 10.1175/1520-0442(1997)010<1787:TROTPN>2.0.CO;2

Fig. 22.
Fig. 22.

Same as Fig. 19 except for U.S. surface temperature.

Citation: Journal of Climate 10, 8; 10.1175/1520-0442(1997)010<1787:TROTPN>2.0.CO;2

Fig. 23.
Fig. 23.

Same as Fig. 12 except for 45Y experiment composites.

Citation: Journal of Climate 10, 8; 10.1175/1520-0442(1997)010<1787:TROTPN>2.0.CO;2

Fig. 24.
Fig. 24.

Same as Fig. 14 except for 45Y experiment composites and contour interval of 5 mm.

Citation: Journal of Climate 10, 8; 10.1175/1520-0442(1997)010<1787:TROTPN>2.0.CO;2

Fig. 25.
Fig. 25.

Same as Fig. 19 except for U.S. precipitation.

Citation: Journal of Climate 10, 8; 10.1175/1520-0442(1997)010<1787:TROTPN>2.0.CO;2

Fig. 26.
Fig. 26.

Same as Fig. 16 except for 45Y experiment composites and contour interval of 5 mm.

Citation: Journal of Climate 10, 8; 10.1175/1520-0442(1997)010<1787:TROTPN>2.0.CO;2

Table 1.

Years (add “19” prefix) in which monthly mean SST anomaly averaged over 150°–180°W and 5°N–5°S exceeded |0.8°C|. Those that also exceeded |1.0°C| are indicated in boldface.

Table 1.
Table 2.

ENSO episodes examined in this study.

Table 2.
Table 3.

Parameters for three sets of composites for analyses of mean squared amplitude of members of composite.

Table 3.
Table 4.

Explained variances (EV) times 100 for J92 experiment monthly mean composites.

Table 4.
Table 5.

Monthly mean PNA region 700-hPA height composite characteristics. Listed for each type of composite and for each month (first or middle month for 2- or 3-month windows, respectively) are number of members in the composite (N) and number of cases in a 1000 trial randomization test that had larger percents of area locally significant at the 5% level in boldface, that is, field significance (FS). Also listed for each one-month window monthly composite is the membership criterion (C) for the preferred composite in absolute degrees Celsius departure from normal with an asterisk (*) if this selection criterion was mandatory because of preset membership limits, and explained variances (EV) times 100 in italics.

Table 5.
Table 6.

Same as Table 5 except for U.S. surface temperature. Here N is not shown for 2- and 3-month windows and ENSO episode composites because they are the same as those in Table 5.

Table 6.
Table 7.

Same as Table 5 except for U.S. precipitation. Here N is not shown for 2- and 3-month windows and ENSO episode composites because they are the same as those in Table 5.

Table 7.
Table 8.

Characteristics of monthly mean PNA region 700-hPA height, U.S. surface temperature, and U.S. precipitation composites from 45Y run experiment. All parameters are defined as in Table 5 and are for 1-month window composites. The parameters C and N are not tabulated because they are identical to their observed composite counterparts in Tables 5–7, with the exception of negative SST anomaly N for months 1 and 2 (January and February), which are one less because 45Y run data were not available for January and February 1950.

Table 8.
Table 9.

Results of analyses of variance for all three sets of positive-SST anomaly composites. Numerators and denominators of ratios and other quantities are weighted (by M) averages over all months. All ratios are times 100.

Table 9.
Save