1. Introduction

A large fraction of current operational seasonal predictability across North America is derived from knowledge of the state of the El Niño–Southern Oscillation (ENSO) cycle (Trenberth 1997). Across the southwestern United States, winter and spring precipitation anomalies are positively correlated with central and eastern equatorial Pacific Ocean surface temperatures. The Southwest is anomalously wet during the winter and spring seasons following El Niño, when the Tahiti–Darwin Southern Oscillation index (SOI) is negative, and anomalously dry following La Niña, when the SOI is positive (Redmond and Koch 1991). This relationship has been established observationally for more than a decade (Ropelewski and Halpert 1986, 1987). The correlation is particularly pronounced using spring streamflow data that effectively integrate high-elevation snowpack over the entire cold season (Cayan and Webb 1992). The ENSO–Southwest precipitation correlation is now routinely considered each year by the National Oceanic and Atmospheric Administration (NOAA) Climate Prediction Center and other climate forecast centers when long-lead winter outlooks are prepared.

The ENSO cycle is closely, though not perfectly, phase-locked to the seasonal cycle in the tropical Pacific, such that by late boreal summer each year it is usually apparent whether the equatorial Pacific is moving toward an ENSO extremum (either cold or warm) the following boreal winter (Rasmusson and Carpenter 1982; Wallace et al. 1998). Midlatitude teleconnections forced by atmospheric heating anomalies associated with tropical rainfall anomalies are also most pronounced in the boreal winter season (Trenberth et al. 1998). Thus the seasonality of both the ENSO cycle and of tropical–extratropical teleconnection dynamics acts to create a winter seasonal maximum in ENSO-based long-lead predictability of North American climate anomalies.

Longer-term fluctuations of Pacific Ocean temperatures and North American precipitation have been scrutinized more recently (Graham 1994; Latif and Barnett 1994, 1996). Mantua et al. (1997) described a low-frequency mode of Pacific Ocean variability centered off the equator in the Northern Hemisphere the spectrum of which contains more power at decadal periods than the ENSO signal. This Pacific decadal oscillation (PDO, as it shall be denoted in this paper) apparently is correlated with oceanic conditions near the North American coast (Bernal et al. 2001) and with precipitation across the southwestern United States (Gershunov and Barnett 1998; Higgins et al. 2000; Liles 2000) and northwestern Mexico (Brito-Castillo et al. 2002).

Higgins et al. (2000) present maps showing that PDO-like variability (which they called “tropical interdecadal” fluctuations) accounts for at least as much interannual precipitation variance across the southwestern United States as a higher-frequency interannual ENSO-related index. Barlow et al. (2001) reached similar conclusions for summer-season hydroclimatic variability. McCabe and Dettinger (1999) showed pronounced decadal variations in interannual correlations between ENSO-related indices, especially the Tahiti–Darwin Southern Oscillation index, and winter precipitation across western North America. They related these decadal variations to the PDO.

A goal of this study is to quantify the extent to which prediction schemes incorporating both ENSO and PDO indices might enhance the statistical relationship between Pacific SST and southwestern U.S. precipitation. Covariate relationships with a season of lead time [SST leading precipitation, as carried out in the analysis by Redmond and Koch (1991)] that may have practical predictive value are established to determine whether additional information available in late summer pertaining to the state of the PDO significantly modifies or enhances the ENSO-based predictability of Southwest winter precipitation anomalies.

Gershunov and Barnett (1998) calculated composites of the seasonal frequency of heavy precipitation across the United States with respect to ENSO extrema separately for high and low North Pacific SST years [their North Pacific oscillation (NPO) index is similar to the Mantua et al. (1997) PDO index]. They concluded that the ENSO signal in wintertime North American climate anomalies is significantly enhanced when the PDO and ENSO constructively interfere (El Niño + high PDO, La Niña + low PDO) and is diminished when PDO and ENSO destructively interfere.

This study extends the Gershunov and Barnett (1998) and McCabe and Dettinger (1999) analyses by considering in detail a region of the United States with a well-known high-amplitude ENSO winter climate signal. Results are expressed using a statistical framework similar to operational long-lead outlooks. The first issue addressed is the relative importance of ENSO and PDO indices as individual predictors of Southwest precipitation anomalies. This is followed by an examination of multivariate predictability incorporating information from both ENSO- and PDO-related conditions in the Pacific. The ultimate goal is to inquire whether some practical enhancement of seasonal predictability of Southwest precipitation is possible.

2. Data and analysis procedure

3. ENSO, PDO, and Southwest winter precipitation

Scatterplots were constructed for all climate divisions in New Mexico and Arizona. The plot for NM4 is shown in Fig. 2 as an example. The headwaters of the Gila River, whose spring streamflow anomalies are highly correlated with ENSO fluctuations (Molles and Dahm 1990), are located within NM4. The x coordinate of each point is its JAS Niño-3 value for a particular year, and the y coordinate is the corresponding September–August PDO index. The subsequent DJFM precipitation anomaly is represented by the symbol plotted for each year, divided into terciles. An × represents a precipitation anomaly in the medial tercile (i.e., near-climatological precipitation). An open circle denotes a negative (dry) precipitation anomaly, in the lowest tercile of the 48 yr, with the larger circles denoting dry anomalies in the lower half of the dry tercile (i.e., an extreme negative anomaly). Solid dots denote positive (wet) precipitation anomalies in the same way, with the larger dots representing the wettest 1/6 of the DJFM values.

The vertical lines in Fig. 2 split the 48 yr of data into Niño-3 terciles corresponding to the extrema marked by the solid and open squares in Fig. 1a. The trio of numbers 5/1/10 to the right of the x-axis label indicates the distribution of dry/medial/wet DJFM seasons in NM4 following upper-tercile (warm) values of JAS Niño-3 SST. Using the definitions introduced in section 2c, x/y/z = 5/1/10, N = 16, and S₊ = 0.69.

The fraction of wet winters following warm Niño-3 is z/N = 0.62, which exceeds the fraction expected by chance (0.33) by 0.29 or 29% (of N). Following the convention employed by the National Centers for Environmental Prediction (NCEP) in its operational long-lead outlook maps, +29 would be the numerical value assigned to this region for a winter long-lead outlook issued in October when Niño-3 is warm. Enhanced probabilities exceeding 20 are unusual on NCEP's operational long-lead outlooks, demonstrating the historically strong El Niño effect on winter precipitation in the NM4 climate division.

The corresponding distribution of winter precipitation anomalies following cold JAS Niño-3 (also shown on the x-axis label) is 6/7/3, yielding a skill score of S₋ = 0.81. The observed chance of a dry winter following cold Niño-3 is 6/16 (0.37), exceeding the null hypothesis of 0.33 by only 4%. On an NCEP long-lead outlook map, this value, less than +5, would be insufficient to plot, and (unless other long-lead forecast tools indicate otherwise) this region would be labeled CL, meaning that there is little or no basis for a long-lead outlook different from climatology, based just on Niño-3.

The horizontal lines in Fig. 2 split the data into terciles based on the preceding September–August PDO index. The trios of numbers just above and below the y-axis label show the distributions of NM4 winter precipitation anomalies following negative or positive August–September PDO years. Negative PDO yields no predictive skill at all (5/6/5 yields S₋ = 1.0 with no enhanced probability of a dry winter).

The corresponding precipitation outlook for positive PDO yields an observed distribution of 3/5/7, so S₊ = 0.73 and the enhanced probability of a wet winter is +13 (7/15 = 46.7%, about 13% higher than the 33% climatological expectation). The S₊ metric indicates that positive PDO generates slightly less seasonal predictive skill of NM4 winter precipitation than does warm Niño-3.

The trios of numbers just within the upper-right and lower-left corners of the scatterplot illustrate multivariate predictions, based on both SST predictors exhibiting lower- or upper-tercile anomalies. The seasonal predictability implied using both ENSO and PDO predictors is obtained by examining the data points in the upper-right or lower-left corners of the plot. The set of four warm Niño-3 + PDO years (upper right) yields a precipitation anomaly distribution of 1/0/3, so S₊ = 0.5 and the fraction of wet winters is enhanced by 42% over the climatological probability of 33%, but that result is based on a very small sample. The multivariate prediction yielded by lower-tercile values of both Niño-3 and PDO is 2/3/1 (lower right), so S₋ = 0.83, which in this case is little different from random chance.

Scatterplots like Fig. 2 were calculated for each climate division in Arizona and New Mexico. Comparison of the NCEP-style enhanced probabilities with plots with S₋ and S₊ values gives us a sense of how to interpret the skill scores. By construction, S scores of 1 or more demonstrate no skill at all relative to climatology. Values greater than about 0.9 are so slightly skewed that no operationally useful predictability is indicated. The S values between about 0.75 and 0.9, derived from sample sizes larger than the multivariate predictions described above, indicate modestly meaningful predictive skill. Values between 0.6 and 0.75, or smaller, come from distributions that are markedly skewed toward the predicted tercile.

Skill scores for all Arizona and New Mexico climate divisions are illustrated separately for cold or warm Niño-3 predictors and negative or positive PDO predictors. On the Niño-3-based maps (Fig. 3), all divisions except Arizona division 4 exhibit some useful predictability for both cold and warm years. Seasonal predictability is generally somewhat higher (lower S scores) for warm events leading wet winters (Fig. 3b) than for cold events leading dry winters (Fig. 3a). Skill scores based on cold Niño-3 years (Fig. 3a) are mostly in the range of very modest predictability. Somewhat lower S scores are associated with wet winters following warm Niño-3 seasons (Fig. 3b), with enhanced precipitation probabilities of greater than 20% in five divisions scattered around the two states.

The corresponding S scores derived from PDO terciles are mapped in Fig. 4. Like the NM4 example, S scores across Arizona and New Mexico are broadly comparable to, or somewhat higher than, the Niño-3-based scores. Five of the 15 divisions in Arizona and New Mexico exhibit S₋ scores greater than 0.9, indicative of no predictive skill during cold/negative years. Like the Niño-3-based results, positive PDO anomalies provide more predictive skill than do negative PDO anomalies. With these results, it can be concluded that Niño-3 is a slightly better overall seasonal predictor than PDO for wintertime southwestern U.S. precipitation anomalies. This conclusion is different from results presented by Gershunov and Barnett (1998), who show distinctly more anomalous frequencies of heavy precipitation associated with La Niña (cold Niño-3) than with El Niño (warm Niño-3).

Multivariate S scores and enhanced probabilities, such as those discussed for NM4 in Fig. 2, were examined for each division but are not shown. As was the case for NM4, in the divisions with the lowest warm Niño-3-based S scores, addition of the PDO generally increases statistical predictability slightly. There is no corresponding enhancement of forecast skill derived from multivariate prediction incorporating cold Niño-3 and negative PDO.

However, as the NM4 scatterplot exemplifies, the sample size of the multivariate prediction is small, and so the robustness of this result is uncertain. Equally important from an operational perspective, the very small sample size indicates that such a multivariate prediction cannot be made in most years (i.e., years of coincident ENSO and PDO extrema are uncommon). Thus the multivariate predictions of winter precipitation based on simultaneous PDO and Niño-3 anomalies, derived from historical records, are probably not useful operationally.

4. Modulation of seasonal predictability by the 1970s regime shift

The results described in the previous section suggest that, considered separately, the Niño-3 and PDO indices have similar effects on seasonal predictability of winter precipitation in the Southwest, with Niño-3 providing considerably more predictive skill. In this section, a different approach to constructing predictability scatterplots is defined that clarifies the effect of the PDO. The data are explicitly separated into subperiods before and after the PDO regime shift in 1977. This year represents the only major shift described by Gershunov and Barnett (1998) in their twentieth-century NPO time series manifested in the 1950–97 data record; another midcentury shift in their record occurred in 1947, a few years prior to the start of our record.

Predictability scatterplots for NM4 for the 1950–76 and 1977–97 predictor periods are shown in Fig. 5. These plots collectively depict the same data points shown in Fig. 2 but now separated by the 1977 PDO regime shift. The upper- and lower-tercile limits of Niño-3 and PDO derived from the full period of record are maintained, with the result that there is only one upper-tercile PDO year in the 1950–76 data (Figs. 1a and 5a) and one lower-tercile PDO year in the 1977–97 data (Figs. 1b and 5b).

There is a pronounced difference in ENSO-based predictability before and after the regime shift. In the preshift data subset (Fig. 5a) cold Niño-3 values systematically lead to a highly skewed distribution of NM4 winter precipitation: the cold Niño-3-based precipitation tercile values are 5/5/1, yielding a skill score S₋ = 0.64 and an enhanced predictability of a dry winter of 12%. Warm Niño-3 values, however, lead to a precipitation tercile distribution of 4/1/4, generating no significant predictability of NM4 precipitation (dry and wet winters are equally likely; S₊ = 1).

This asymmetry in NM4 predictability is reversed in the later, post-1977 data subset (Fig. 5b). After 1977, when the PDO is generally positive (Fig. 1a), the set of five cold Niño-3 values leads to an insignificantly skewed precipitation distribution (1/2/2; S₋ = 1.2) whereas warm Niño-3 values systematically lead to positive NM4 winter precipitation anomalies (1/0/6; S₊ = 0.29, with a 52% enhanced predictability of a wet winter).

Figures 6 and 7 illustrate the Niño-3-based S scores for all divisions in Arizona and New Mexico, with the data divided into pre-PDO-shift and post-PDO-shift years. The smaller numbers of years contained in each subsample lead to greater sampling uncertainty in each score relative to the statistics derived from the entire dataset, but the pronounced asymmetry in skewness in the two groups of years is readily apparent. In the pre-PDO-shift data (before 1977, when PDO was primarily negative) S₋ scores associated with cold Niño-3 years (Fig. 6a) are smaller (more skill) than the corresponding values derived from the entire period of record (Fig. 3a) in every division except AZ7 (southeastern Arizona) and NM3 (New Mexico northeastern plains). On the other hand, seasonal predictability associated with warm Niño-3 is markedly worse in most divisions (Fig. 6b). Ten of 15 divisions in these two states exhibit S₊ scores above 0.9 in this subperiod, indicating that predictability during El Niño winters was practically nonexistent during these decades.

In the post-PDO-shift subset of years (1977–97) the asymmetry in ENSO-based predictability reverses across the Southwest. The small set of five cold Niño-3 years yields S₋ scores greater than or equal to unity in 14 of 15 divisions (Fig. 7a). However the set of seven warm Niño-3 years exhibits a dramatic shift toward wet winters (Fig. 7b), with S₊ scores that are less than 0.75 in every division and less than 0.6 everywhere except the two westernmost divisions in Arizona.

Skill scores associated with PDO variability within the preshift and postshift data subsets (Fig. 8) are generally slightly lower than the analogous overall PDO-based scores (Fig. 4) but are much less impressive than the preshift and postshift Niño-3-based scores. In the 27 preshift years prior to 1977 there is only one upper-tercile PDO value, so only statistics associated with lower-tercile extrema are meaningful (Fig. 8a). Skill scores are less than 0.75 in only 2 of the 15 divisions in Arizona and New Mexico. When comparing Fig. 8a with Fig. 6a, it is seen that Niño-3 again provides more information than the PDO for prediction of dry winters.

The same comparisons hold for positive PDO as a predictor in the postshift decades. The postshift PDO-based S₊ scores in Fig. 8b are mostly a small amount lower than the S₊ scores based on PDO for the entire dataset (Fig. 4b); six divisions in Fig. 8b have scores less than 0.75. The PDO-based scores throughout the Southwest are not nearly as low as the Niño-3-based scores in Fig. 7b, however.

In summary, pronounced decadal variability in the predictive skill associated with ENSO SST is associated with the regime shift in 1977. Before the shift, cold Niño-3 temperature in late summer was an excellent predictor of a dry winter across the Southwest, but warm Niño-3 anomalies did not provide much predictive skill for subsequent winter precipitation. After the shift, warm Niño-3 values systematically led to wet winters but cold Niño-3 values were not of much use. Both before and after the shift, Niño-3 is a better interannual predictor of Southwest precipitation anomalies than are the interannual PDO fluctuations.

5. Discussion and conclusions

The goal of this study has been to clarify the seasonal statistical predictability of winter seasonal anomalies of precipitation in Arizona and New Mexico derived from antecedent knowledge of Pacific SST anomalies along the equator and in the North Pacific. Autumn-season (JAS) Niño-3 SST and the annually averaged Mantua et al. (1997) PDO index were used to characterize Pacific SST prior to each winter season. Predictability was assessed using a modified Brier score [Eq. (1)] that conveniently encapsulates the skewed three-bin precipitation distributions used operationally to describe probabilistic seasonal anomaly forecasts. The 48-yr period of record was initially considered as a single dataset. The ENSO cycle provided more predictability than the PDO index, and the ENSO-based seasonal predictability was modified only modestly by multivariate consideration of the PDO.

Using the PDO regime shift in 1977, the empirical predictability statistics were recalculated after splitting the data into preshift and postshift subsets. The statistics calculated on the partitioned data illustrated dramatic differences in ENSO-based predictive skill in the two subsets. During the pre-1977 period, when the PDO was largely negative, predictive skill was higher for cold Niño-3 temperatures leading dry winters in the Southwest than for warm Niño-3 leading wet winters; that is, La Niña provided a better basis for a winter seasonal forecast than El Niño did. After the regime shift in 1977, the asymmetry in predictability reversed, and El Niño was a better basis for a seasonal forecast than was La Niña. This single regime shift probably is responsible for most of the large effect of the PDO on seasonal climate anomalies that was documented by Gershunov and Barnett (1998) and Higgins et al. (2000). Echoing a principal conclusion in those two papers, and the McCabe and Dettinger (1999) study, the results presented here emphasize the importance for seasonal climate prediction of incorporating information about the phase of the PDO when formulating ENSO-based winter forecasts.

The PDO exhibits interannual variability in addition to the 1977 regime shift, however. Within both the pre-1977 and post-1977 subsets, fluctuations of the PDO only modestly affect the anomaly distribution of subsequent Southwest wintertime precipitation as compared with the predictability associated with Niño-3.

These results have several implications for operational seasonal forecasting. First, the practical utility of the PDO for seasonal forecasting is limited to some extent, because in real time it is difficult to distinguish a long-term regime shift (which apparently occurred just once in our 48-yr data record) from a shorter-timescale fluctuation. Thus, even if the PDO regime shift had been identified in 1977, the importance of that shift relative to other fluctuations (e.g., PDO variations in 1958 or 1991 in Fig. 1a) could not have been identified for the purposes of a seasonal forecast.

In this regard, the possibility that the PDO shifted regimes back into a negative phase in the late 1990s adds uncertainty to ENSO-based seasonal forecasts for the next few years. If the PDO has now returned to a long-term negative phase, then the very high ENSO-based predictability of wet winters in recent decades (Fig. 7b) may weaken, and instead the cold phase of ENSO will generate better predictability, as occurred before 1977 (Fig. 6a).

A second implication is somewhat more speculative, that is, whether North Pacific Ocean temperatures are truly responsible for the shift in ENSO-based predictability that unquestionably seems to have occurred in the late 1970s. On physical grounds it is unclear why short-term, year-to-year fluctuations of North Pacific SST should have any less influence on winter-season precipitation in the Southwest than multidecadal SST fluctuations do. However most of the PDO influence seems to be related to just one pronounced regime shift, because PDO-based interannual predictability within each regime is modest (Fig. 8).

Furthermore, studies of North Pacific SST variability identify other “regime shifts” that do not have effects on Southwest precipitation as pronounced as those of the 1977 shift. For example, Chao et al. (2000) point to 1957–58 as the year of another interdecadal Pacific SST regime shift in the latter half of the twentieth century. A shift in that year is difficult to see in Fig. 1a, and there is no obvious ENSO predictability change associated with that year. Perhaps some related component of the climate system that shifted in the late 1970s is the physical agent responsible for the different effect of ENSO fluctuations on southwest North American climate before and after 1977. Support for this possibility is found in a recent study by Pierce (2002), who found that specifying North Pacific SSTs in a coupled model did not force the circulation changes associated with observed decadal modulations of ENSO teleconnections.

The foregoing speculation cannot be addressed in our empirical study and probably cannot be resolved through purely empirical analyses of any kind. Better fundamental understanding of the dynamics of storm tracks and tropical–extratropical interactions will be required to make progress on this problem. Improvements in characterization and understanding of the PDO, including extending a record of its variability backward in time (Biondi et al. 2001), are necessary to define the PDO's far-field effects on climate. Meanwhile, even in the absence of dynamical understanding, it is appropriate to consider the asymmetry in ENSO-based seasonal predictability between the low-PDO and high-PDO climate regimes we have documented here.