A Precipitation Climatology for Stations in the Tropical Basin;
Effects of ENSO
The influence of ENSO on Pacific Basin precipitation is described in a general way by spatial maps of the linear correlation between an index of ENSO and the rainfall at the 66 stations shown in Fig. 1 . The ENSO indices chosen here are (1) the Southern Oscillation Index (SOI) derived from Tahiti minus Darwin sea level pressure, (2) the SST anomaly in the Niño 3.4 region (5.5N-5.5S, 120-170.5W), and an equally weighted average of both indicators. To simplify interpretation, the sign of the SOI is reversed so that positive SOI is associated with El Niño as is a positive SST anomaly. In forming the average of SOI and SST, standardized anomalies of 3-month means of each index are first computed, and then the sign-reversed standardized SOI anomaly is added to the standardized SST anomaly, and the sum is divided by 2. Regardless of whether the SST alone or the SOI alone shows stronger correlations with the basin station rainfalls, using an average of the two indices is done in order to reduce noise and sampling variability to obtain a more complete, balanced indicator of the ENSO state. It is of interest to examine ENSO-rainfall correlations both contemporaneously and at lag. When a lag is used, the ENSO index occurs before the rainfall and is thus viewed as a predictor.
The resulting correlation maps are shown in Figure 4 for four target (predictand) seasons: Jan-Feb-Mar, Apr-May-Jun, Jul-Aug-Sep and Oct-Nov-Dec. Panels a, b, c and d show simultaneous correlations, panels e through h show correlations in which the ENSO index leads the rainfall by 3 months, panels I through L for ENSO index leading rainfall by 6 months, and m through p ENSO leading by 9 months. This sequence of correlation maps is shown first where the SOI is used as the ENSO index, then where the (SOI+SST)/2 is used as the ENSO index, and finally where just the Niño 3.4 SST is used. Light shading denotes statistically significant negative correlations at the 0.05 significance level, and dark shading shows significant positive correlations. (Recall that the SOI is sign-reversed; therefore, positive correlations indicate wetness in association with warm ENSO, as do positive correlations with the SST.) With 42 cases (41 cases for lagged relationships), the correlation magnitude required for statistical significance is 0.32. A thick contour is drawn in Fig. 4 for zero correlation, and light solid (dashed) contours are drawn for positive (negative) correlations of 0.50 and 0.75. Because the contouring algorithm works on the basis of gradients as well as actual values, a contour for a given correlation level may be drawn even when no actual individual station correlation (shown by numerical labels at the station locations) attains that level. Moreover, smoothing is used, such that contradictions between the contouring and the individual values may occur, especially in regions having great local variation in correlation. The user should therefore examine the numerical labels as well as the contours in making independent judgements about the overall correlation magnitudes in given vicinities, taking highly local factors into consideration in interpreting the individual station correlations.
Results using any of the three ENSO indices are fairly similar. This is expected in view of the high correlation between the SOI and the Niño 3.4 SST--near or just above 0.8 for much of the year, and more than 0.85 from late boreal summer through fall (see Table 2 in Barnston et al. 1997). Comparing results for the ENSO indices in more detail, it is found that the SOI and SST alternate fairly evenly with respect to which produces stronger results, depending on the region within the basin. For the simultaneous correlations for Jan-Feb-Mar, for example, stronger correlations are found using the SOI alone at the Hawaiian stations, Johnston, and the southeastern near-equatorial stations (Atuona and Takaroa) that are longitudinally close to the Tahiti part of the SOI. The Australian stations are also more closely related to the SOI than the SST, being close to Darwin. Stations along the immediate equator near the Niño 3.4 region, such as Fanning and Christmas islands as well as the other five Kiribati stations and Nauru just west of the date line, are more closely related to the Niño 3.4 SST than the SOI. The SST also produces stronger results for the northwestern stations of Wake, Kwajalein, Chuuk, Yap and Koror. Many of the south-of-equator stations that are dry with El Niño are more strongly related to the SOI than the SST. A greater relevance of the SST to precipitation in the northwestern stations might be expected because neither Tahiti nor Darwin is close to that portion of the basin, both being south of the equator. The (SOI+SST)/2 ENSO index, while surpassing both SOI alone and SST alone only in a few cases for the simultaneous correlations for Jan-Feb-Mar (e.g. Guam WSMO, Pohnpei and Majuro; and tying at several other stations), the combined index yields results about equally strong as those of the SOI or SST alone when considered over all four seasons and all three lead times. This occurs largely because of stations located in between the regions where the SOI is more relevant and the region where SST is more relevant; these stations benefitting from a more complete ENSO index. The combined index tends to produce correlations that are slightly higher than the average of the individual SOI and SST correlations, but this "advantage" may vanish when squared coefficients (reflecting the percent of variance explained) are considered. It is basically a more conservative index that is less vulnerable to sampling problems than the individual SST or SOI indices. For all three ENSO indices, the simultaneous relationships appear strongest during the Oct-Nov-Dec and Jan-Feb-Mar periods at the off-equator stations, and relatively weaker in Apr-May-Jun and Jul-Aug-Sep. This result is somewhat different for the near equatorial stations of Kiribati and Nauru where highest correlations occur in Jul-Aug-Sep and Oct-Nov-Dec, and lowest occur during the other two periods. Factors possibly relevant to these findings are (1) the ENSO signal in the SST tends to be highest compared with random variations during the boreal late winter and spring (Kumar and Hoerling 1998), and (2) the normal seasonal cycle of SST, in which SST is highest (and closest to the rain-producing convection threshold, which would be easier to exceed with a given positive anomaly) during boreal spring. During Jan-Feb-Mar, the off-equator stations show rainfall deficits (surpluses) during warm (cold) phase ENSO periods, while the stations along the equator from about 160.5E eastward respond in the opposite manner. During the other seasons, the same geographical distribution of direction of response is generally observed, but with the equatorial region receiving enhanced rainfall with warm ENSO extending slightly farther westward. In Jan-Feb-Mar, the latitudinal range with which simultaneous El Niño-related rainfall surpluses are found is about 5.5N-10.5S near the date line and about 15.5N-20.5S near 150.5W longitude. Stations in this study that tend to respond with significant positive anomalies to El Niño and negative anomalies to La Niña throughout a sizeable portion of the annual cycle include Fanning, Christmas, Arorae, Beru, Tarawa, Banaba, Nauru, Nui, Atuona, and Takaroa. In Oct-Nov-Dec and Jan-Feb-Mar, Funafuti, Atafu, Pukapuka, Rakahanga, and Penrhyn join this set as well. While Henderson (near 160.5E) is drier than normal with El Niño (and wet with La Niña) in boreal fall and winter, it tends to be wet with El Niño (dry with La Niña) in Jul-Aug-Sep. While many of the stations not mentioned have the opposite response to ENSO (dry with El Niño), some of them are in nodal locations and have unpredictable impacts, or tend respond variably or oppositely as a function of the time of the year (like Henderson). Examples of nodal locations are Rabaul, Bora Bora and Papeete, and a location whose response depends on the time of year is Koror (dry with El Niño only in boreal fall and winter).
Figure 4 indicates that when the ENSO index leads precipitation by 3, 6 or 9 months, the sense of the relationships tend to remain as they are for simultaneous relationships but weaken for most locations and target seasons. This makes sense in view of the increased uncertainty of the ENSO state that will be realized at the time of the rainfall. However, for some stations/target seasons the weakening is not substantial, such as at the four Hawaiian stations, Nukualofa, and some of the Fijian stations for Jan-Feb-Mar. The Hawaiian result at 1 and 2 season lag was also obtained by Chu (1989), and implies that useful precipitation forecasts can be made for Jan-Feb-Mar as early as the beginning of October, providing three months for strategic planning on the parts of water management interests on the affected islands. Significant relationships at these lag times at locations other than Hawaii, parts of Fiji and southern Tonga are also noteworthy, and form the familiar horseshoe-shaped pattern of off-equator El Niño-related dryness surrounding the equatorial wet zone, with northern and southern dry regions nearly meeting in the western equatorial Pacific. For example, moderately strong correlations for Jan-Feb-Mar at 6 months lag appear close to the equator both east and slightly west of the date line, at most of the Kiribati stations. The general geographical extent of the predictive potential shown here is qualitatively similar to that described by Ropelewski and Halpert (1987, 1996) and in the CCA studies of Barnston and He (1996) for Hawaii, and He and Barnston (1996) and Yu et al. (1997) for the tropical Pacific islands in general. Specifically, when the ENSO phase has been established by boreal mid-summer (i.e., it appears to be warm or cold rather than neutral), that phase tends to persist through the remainder of the calendar year (Barnston and Ropelewski 1992). This causes the lagged correlation relationships with Jan-Feb-Mar rainfall to be somewhat more similar to the Jan-Feb-Mar simultaneous relationships than is the case for other target seasons. On the other hand, the "spring barrier" in the continuity of the ENSO state causes the ENSO-rainfall relationships to weaken more quickly for boreal summer and fall target periods when lag time is introduced. This is illustrated, for example, by the deterioration in the strength of the Jul-Aug-Sep simultaneous relationships at most stations (Fig. 4c ) over the three lag times (Figs. 4g, 4k and 4o).
The influence of ENSO episodes on Pacific Basin precipitation is described in greater detail using composite analysis, in which responses to warm ENSO episodes are considered separately from responses to cold episodes. The degree of realism of the assumption of linearity in the ENSO-rainfall relationship, which is needed in the overall interpretation of the correlations shown in Fig.5, is evaluated in composite analysis. If rainfall anomaly composites are fairly equal-but-opposite for warm versus cold ENSO conditions, approximate linearity is confirmed. The composite analysis is conducted as follows: Three-month running total precipitation is composited (averaged) over the cases assigned to a particular ENSO status (warm or cold episode), for 3-year periods centered on the boreal winter of the mature phase of the episode, and compared with the climatological average rainfall. The set of winters defined as warm or cold episodes, shown in Table 4, is based on an examination of the Tahiti-minus Darwin sea level pressure-based Southern Oscillation Index (SOI) and the tropical Pacific SST leading up to and including the boreal winter season in question. While the resulting set of winters is in general agreement with the sets as defined earlier by Loon and Madden (1981), Rasmusson and Carpenter (1983) and Ropelewski and Jones (1987), a few differences exist.
The current categorization is based on the standardized anomalies of running 3-month mean SOI and running 3-month mean SST in the Niño 3.4 region (5.5N-5.5S, 120-170.5W). The Niño 3.4 region is selected because it has been suggested to be more highly correlated to the core ENSO phenomenon than other traditional regions such as Niño 3 or Niño 4 (Barnston et al. 1997). The 1950-1996 period is used as the basis for the climatology upon which standardized anomalies of the SOI and the Niño 3.4 SST are computed. While the SST in the Niño 3.4 region is highly correlated with the SOI (on the order of -0.80 to -0.85, varying somewhat with season), use of both variables creates a slightly more complete and balanced indication of the ENSO state than either alone. Because the SST in the ENSO-related tropical Pacific is now considered to be the hallmark of the ENSO phenomenon, it is weighted more heavily in the categorization process than the SOI. Hence, even with a strong SOI anomaly, the SST anomaly must reach a critical level in the Niño 3.4 region in order for a "warm" or "cold" boreal winter to be established. However, a strong SST signal is not sufficient; the SOI must also assume a more lenient corroborating anomaly value in order for the winter to qualify. The details of the procedure used to classify the winters, and some consequences of this procedure, are provided in the appendix.
Figure 5 shows ENSO composite rainfall results for the 66 Pacific basin stations. The dotted line shows the climatological mean rainfall for all years, the solid line the mean for the composited 12 warm ENSO episode years, and the dashed line that for the 8 cold episode years. Vertical lines denote calendar year changes. Each plot shows composites for the period beginning about 5 seasons before the boreal winter of a mature episode and ending about 5 seasons after that winter. The notation for relative year identity (0 and +1) in Fig. 5 is as defined in Ropelewski and Halpert (1987). Results for small portions of year -1 (late boreal autumn) and year +2 (early boreal spring) are also shown. Thus, the mature episode winter (year 0 to +1) is located near the midpoint of the abscissa range. The differences between the composited rainfall totals for the samples representing the warm or cold phases of ENSO versus the totals of the remaining years (neutral-plus-oppositely phased years) were statistically tested with the Student's t-test. It should be noted that the statistical assumptions underlying the t-test (e.g. Gaussian distributions) may not be sufficiently satisfied. Nonetheless, we use it as a rough guide for indicating significant mean differences in rainfalls as a function of ENSO category. For a given station, those seasons whose warm or cold episode years' mean passed a 2-sided test with a significance level of 0.05 or better are indicated with a hollow or solid square, respectively, along the solid or dashed curve in Fig. 5 . For example, near the time of the mature episode boreal winter at Kahului, warm episodes are associated with deficient precipitation with 0.05 or stronger statistical significance in Dec-Jan-Feb, Jan-Feb-Mar, Feb-Mar-Apr and Mar-Apr-May. Cold episodes associate with enhanced rainfall at the 0.05 significance level only during Feb-Mar-Apr and Mar-Apr-May. In looking at the time of mature episode boreal winter over all 66 stations, one notes the general relationship between ENSO and rainfall described in Ropelewski and Halpert (1987): At the off-equator stations (such as many of the U.S.-affiliated stations) warm ENSO is associated with suppressed rainfall, while at the near-equatorial stations from the date line eastward to the South American coast, where the SST anomaly is positive, rainfall is enhanced. Enhanced rainfall with El Niño is particularly dramatic at the seven Kiribati stations (Banaba, Butaritari, Tarawa, Beru, Arorae, Fanning and Christmas), at Nauru, and at two of the four Tuvalu stations (Nui and Atafu). All of these effects tend to occur in reverse for cold ENSO episodes.
While ENSO phase-specific rainfall influences can be distinguished in Fig. 5 that might not show up in the overall correlation results of Fig.4 , certain responses of even greater specificity remain unresolved in both analysis schemes. For example, if a station's rainfall response only occurs reliably in the very strongest two or three warm events (e.g. 1982-83 and its runners-up), this response might not show up in the composite that includes the more moderate events as well. Examples might be Apia and Pago Pago, where ENSO effects as shown by Figs. 4 and 5 are weak but where strong rainfall deficits have occurred during the strongest warm events (e.g. 1972-73, 1982-83) as noted in Fig. 3a,b and Fig. 7 (discussed below).
In general, ENSO effects during periods other than the boreal late summer through fall, winter and spring of the mature phase, are not very noteworthy. However, in some cases an apparent ENSO effect can be noted in the boreal winters a year before or a year after the mature episode boreal winter. At Kahului, for example, there is a significant tendency toward a wet winter the year following the mature warm episode, which itself tends to be dry. A similar feature is found at Willis Island, Wallace Island, Nacocolevu, Niuatoputapu, Alofi and Rarotonga Island. Adjacent winter responses may be associated with the episode that peaks a year beforehand or a year afterward (e.g. positive temperature anomalies in Hawaii are seen to occur the boreal winter one year after a mature El Niño as much as during the El Niño boreal winter itself; Barnston and He 1996). However, they may also be explained in part by adjacent year mature episodes in their own right, which are usually of opposite phase but may occasionally be like-phased (perhaps seen as one extended episode). Figure 6a (top) shows the temporal distribution of the warm and cold episodes, where each episode is assigned to a boreal winter straddling two years. Figure 6b (bottom) shows the distribution of frequencies of types of adjacent year ENSO status sequences. It is noted that there is a greater tendency for boreal winter mature warm episodes to be followed by a cold episode a year later than by another warm episode. Cold episodes, on the other hand, do not appear to be followed by warm episodes. Figure 6b shows that the frequency of occurrence of a warm episode the year before a cold episode is greater than the frequency of a neutral or another cold episode the year before (frequencies of 5, 2, and 1, respectively). This helps explain why the cold episode composites sometimes indicate the opposite sign of rainfall anomaly for the preceding winter (e.g. see Johnston, Pohnpei, Chuuk, Butaritari, Beru, Arorae, Nui, Atafu, Henderson, Koumac, and Rakahanga). While these preferred sequences may suggest a potential for predictability, the sample size of ENSO episodes as presented here is too small to draw such a conclusion.