There are several well-documented studies showing the shifts in seasonal mean rainfall and temperature conditional on El Niño–Southern Oscillation (ENSO) phase. Here the shifts in the seasonal histograms of daily rainfall over South America conditional on ENSO phase are examined. The authors are motivated to analyze daily rainfall statistics over seasons by the demands for information on the shorter temporal scales voiced by users of climate data. In the first stage of the analysis the Kolmogorov–Smirnov (K-S) test is used, comparing El Niño to La Niña histograms of daily station data, to identify regions where there are significant shifts in the histograms. The K-S statistic analyses of daily station data are then compared to the same analyses performed on existing publicly available gridded station datasets. The degree to which the station and gridded data agree in showing geographical regions of significance provides evidence that the gridded fields might provide guidance on the nature of the ENSO signal where station data are not available. Further, the analysis of the gridded datasets can be used to motivate and guide efforts to obtain more complete daily data where the gridded datasets suggest an ENSO signal. As an example a detailed comparison of one station in southern Brazil and its nearest neighbors in the gridded data are presented, suggesting that, despite biases, the gridded fields are generally consistent with the station data where both are available.
For many regions of the world neither daily station data nor daily gridded datasets are available for analysis. Thus despite documented and well-known regional biases in the precipitation fields available in the NCEP–NCAR reanalysis the extent to which shifts in the daily statistics of the NCEP–NCAR reanalysis precipitation are consistent with station and gridded station analyses is also examined. The preliminary work described here suggests that while the reanalysis does not ideally replicate the gridded station results the reanalysis may be useful as a tool for indicating candidate regions for further analysis with station or gridded data.
Over the past two decades climate scientists have made significant gains in understanding how seasonal rainfall amounts are modulated by El Niño–Southern Oscillation (ENSO) phase. This knowledge has been extensively used in empirical seasonal climate forecasts and as a way to evaluate the performance of numerical climate models. The ability to understand and predict (in a probabilistic way) seasonal rainfall amounts is a great achievement. However, this capability often falls short of the information demands made by potential users of climate forecasts. Users and decision makers often ask for information relating to the beginning and ending of rainy seasons, the duration of dry periods within a season, and the amounts of rainfall that will fall in critical periods of the growing season. Thus far, our ability to provide this kind of detailed predictive information has not met with wide success, particularly on the spatial scales of interest to many users. Our current understanding of the climate system indicates that these specific deterministic forecasts of these quantities are not likely to ever be successful.
On the other hand work in the agricultural community, for example, has suggested that some gains in the utility of forecast information can be achieved through characterizing daily weather within a month or season, for example, Hansen and Ines (2005). In their application, stochastic weather generators based on disaggregation of monthly data are used to produce input to crop models. The exploratory work outlined in this paper suggests that examination of the historical record can be used to determine how the statistics of the seasonal weather are modified, or shifted, in relation to ENSO phase. This information might be used directly in some applications or could inform the parameters that go into the stochastic weather generators. Likewise, examination of the historical record can also guide other statistical techniques for generating daily weather sequences, such as the hidden Markov model (Robertson et al. 2004).
In this study we present an example of a statistical analysis that can be used to identify stations that show shifts in the observed histograms of daily precipitation. We illustrate, using data from one station in southern Brazil, how the histograms of observed daily weather at a station are modulated by ENSO phase at this station in two different seasons. In addition, we compare the analysis at a single station to analysis of gridded data nearest the station since these gridded data often provide a longer time history than can be obtained from any single station and are often interpolated into regions with very sparse station data. However, even gridded observational data are not available for large portions of the world. Thus, fully aware of the limitations of the precipitation data derived from model-based data assimilation systems, we also examine the possibility of using the daily historical precipitation record available from the National Centers for Environmental Prediction–National Center for Atmospheric Research (NCEP–NCAR) reanalysis. We emphasize here that the purpose of this paper is to motivate interest and research into the statistics of weather within seasonal climate. Further, our purpose is to motivate efforts to collect and make available daily station rainfall and temperature data, and efforts to examine the possibility of using gridded station data and model-based analyses where the station data do not exist or cannot be made available.
There have been a number of studies demonstrating statistically significant shifts in average, or median, seasonal rainfall amounts associated with ENSO for large areas of the globe (e.g., Ropelewski and Halpert 1987; Kiladis and Diaz 1989). Subsequent studies for South America and Brazil (e.g., Aceituno 1988; Grimm et al. 1998, 2000; Pisciottano et al. 1994; Uvo et al. 1998; Liebmann and Marengo 2001) have documented in more detail the regional ENSO-related shifts in seasonal mean rainfall. These studies, using more detailed data than available in global studies, have shown that the changes in seasonal precipitation amounts associated with the warm (El Niño) and cold phases (La Niña) of ENSO occur in different seasons, in some regions.
Almost all previous studies of ENSO relationships with daily rainfall have focused on extremes (e.g., Cayan et al. 1999; Gershunov and Barnett 1998; Liebmann et al. 2001). Increasingly, however, the potential users of seasonal climate information are asking for more information about shifts in the overall character of the seasons as they are modulated by ENSO. In this paper we motivate expansion of previous work by examination of the statistical properties of the entire observed daily rainfall histograms over 3-month seasons as they might be modified in conjunction with ENSO phase. The purpose of this paper is to illustrate the shifts in the seasonal histograms of daily precipitation, as they are manifest in analysis of station data, gridded station data, and outputs from NCEP–NCAR reanalysis. The comprehensive analysis and interpretation of these ENSO-related changes in the character of daily rainfall within a season for several stations and over specific regions are not addressed here.
This study is possible only through the efforts of several investigators and institutions who developed historical daily rainfall and temperature datasets from station observations and made them available to the research community, for example, Peterson and Vose (1997) and Agência Nacional de Energia Elétrica (2000). The station data used in this study were screened for missing values, and only those stations with data records that are at least 90% complete for the 1961 to 1990 period were included in our analysis. We augment the station data analyses through examination of daily gridded datasets compiled by Liebmann and Allured (2005, hereafter LA05) for all of South America, referred to as LA in the following text, and for Brazil only (Silva et al. 2007) referred to as CPC in the following text. We also examine the daily rainfall statistics obtained from the NCEP–NCAR reanalysis (Kalnay et al. 1996). The El Niño and La Niña seasons were identified using the Niño-3.4 index computed from the extended sea surface temperature data of Kaplan et al. (1998). Thresholds of +0.5°C for El Niño and −0.5°C for La Niña averaged over each 3-month period were used to define ENSO seasons.
a. Station example
We choose as an example to illustrate the modulation of the daily precipitation over a season a station in southern Brazil. This example is chosen because it lies in a region with a complex but well-documented ENSO signal in seasonal means (see, e.g., Grimm et al. 1998). Nova Palmira, Brazil, is located in an area of southeast South America associated with ENSO-related enhanced seasonal rainfall totals for the October of the El Niño year through the following February (Ropelewski and Halpert 1987). The Nova Palmira station is also located within a region, identified by Grimm et al. (1998) as having enhanced rainfall in the August through November season of El Niño years and less rainfall during October to December of La Niña years. Given the differences in these two analyses we look at the statistics of daily rainfall for the August–October (ASO) and the October–December (OND) seasons. Both seasons show enhanced seasonal total rainfall associated with El Niño compared to La Niña in the station data for Nova Palmira (Table 1). However, there are only small differences in total rainfall that occur during neutral years compared to La Niña years during ASO and neutral conditions and El Niño years during OND.
Histograms of daily rainfall of at least 1 mm day−1 for ASO and OND show broader distributions of daily rainfall amounts associated with El Niño compared to La Niña (Fig. 1). The Kolmogorov–Smirnov (K-S) tests (Smirnov 1948; Keeping 1995, 259–260) on the histograms of daily rainfall for Nova Palmira show that the El Niño and La Niña distributions are significantly different from each other at greater than the 98% level in both seasons (Table 2). The K-S test uses the maximum absolute difference between the cumulative distribution curves from two samples to test if the samples come from different populations. The null hypothesis of the test is that the two samples come from populations with the same distribution function.
There are several differences between histograms as illustrated in Figs. 1e and 1f. For example, the seasons associated with El Niño years clearly have many fewer (more) occurrences of days with rainfall less (greater) than 5 mm (20 mm) than the seasons occurring during La Niña years. Time series at Nova Palmira for the ASO and OND seasons confirm the histogram results by, for example, showing more days with rainfall greater than 20 mm during El Niño years than in either neutral or La Niña years (Fig. 2). During ASO all but one of the El Niño seasons in the 1961 to 1990 period experienced more than the average of 8 days with rainfall greater than 20 mm, and all but one of the La Niña seasons experienced fewer days (Fig. 2a). If these two years are excluded from the sample the average number of days with rainfall greater than 20 mm during El Niño years during ASO is more that twice the average for La Niña (11.9 versus 4.8 days). For this example neutral years experience an average of about 8 days with rainfall amounts greater than 20 mm day−1, the same as the overall mean, suggesting that the opposing ENSO phases are symmetric with respect to the number of days with relatively heavy rainfall in the ASO season. However, we note in Table 1 that there are only small differences between seasonal total rainfall in La Niña and neutral years (450 mm versus 440 mm). This suggests that during the La Niña years the daily rainfall distribution shifts toward more days of lighter rainfall than in neutral conditions. This is clearly the case for the difference in the El Niño compared La Niña histograms but it is less clear for the differences between the comparisons of the La Niña versus neutral histograms (not shown). However, the time series of the number of days with ASO rainfall less than 5 mm (Fig. 2c) shows that while half of the neutral years (8 out of 16) experience more than the average number of days with less than 5 mm of rainfall, all but one of the La Niña years have more than the average number of days with light or no rainfall (5 out of 6).
In the OND season, the mean total rainfall amount shows only a small difference between El Niño and neutral years (482 mm versus 452 mm, respectively) but large differences between La Niña totals and either El Niño or neutral years, consistent with Grimm et al. (1998). As noted above, however, the histograms of daily rainfall for the apposing ENSO phases are significantly different from each other at the 98% level by the K-S test. The time series of the number of days with greater than 20 mm of rain during OND (Fig. 2b) shows that half of the warm ENSO episode years experience more than the average number of days with over 20 mm for all years, and all except 1 of the 8 cold episode years have fewer than the average number of days with over 20 mm of rainfall. On average, cold episode years have about 4 days with rainfall greater than 20 mm in OND, while in neutral years the average is about 6 days, and for warm-episode years about 8 days. Despite the seeming symmetry (about neutral years) in the number of days with rainfall greater than 20 mm in this example, the K-S statistic shows no significant differences between La Niña and neutral conditions during ASO or El Niño and neutral conditions in OND. We also do not see a comparable “symmetry” about the number of days with less than 5 mm of rainfall (Fig. 2d). There are clear differences between El Niño years, with 8 of 10 yr showing fewer light rain days, and La Niña, with 7 out of 8 yr showing more days with rainfall less than 5 mm.
In the example given above we chose the thresholds for the time series based on the histograms. We leave the choice of significant thresholds for other studies involving users who need to manage climate-related risk. We also leave for further research the interesting question of whether at some locations there is some precipitation threshold value, or range, critical to some users, for which the ENSO-related differences in the seasonal totals are not significant but where the precipitation values of interest to users do show significant differences.
b. Gridded station datasets
Our long-term goal is to identify datasets and analyses that will allow us to study areas over the globe that may experience ENSO-related shifts in the character of the daily rainfall over a season, that is, significant shifts in daily rainfall histograms. Here we examine the gridded datasets over South America. High-quality, continuous daily rainfall records that extend for at least a 30-yr period are very sparsely distributed; see, for example, Fig. 3 for Brazil. Researchers have attempted to lessen the impact of the inadequate station records by averaging observed station rain data over regular latitude–longitude grids or interpolating data from several stations to a set of grid points. These techniques allow the integration of data from stations that may be less than complete, not continuous in time, or comprised of a number of short-record stations into grid-averaged or point estimates on a regular grid. For brevity we will refer to both sets of analyzed data as gridded datasets. The techniques and rules for forming gridded datasets introduce biases and uncertainties but have the advantage of greatly enlarging the geographical coverage and length of record of available daily data. By its very nature the averaging of daily rainfall over the grid boxes will produce smoother time series of daily rainfall than time series from individual stations; the larger the grid “size” the greater the smoothing (e.g., Robeson and Ensor 2006). As mentioned above, in this study we examine the daily rainfall statistics from two such datasets: LA05 and the dataset produced by the Climate Prediction Center/National Oceanic and Atmospheric Administration (CPC/NOAA), both on a 1° longitude by 1° latitude grid. The CPC uses the same methods to grid the data in North and South America. There is no published comparison of stations to the gridded CPC analysis for South America; however, a comparison of station and the gridded CPC/NOAA analysis for North America can be found in Silva et al. (2007).
A full comparison of station and gridded datasets is beyond the scope of this paper. However, as an example to examine compatibility of these types of data, at least at this location, we compare the daily rainfall histograms for grid boxes that contain Nova Palmira and note that the statistics of the gridded data show the same general behavior as the station data with respect to the opposite phases of ENSO (Fig. 4). In general, the daily precipitation histograms for warm-episode conditions are “broader” than those for the cold ENSO phase. In particular, the warm ENSO phase tends to show relatively more occurrences of daily precipitation greater than 20 mm day−1 than the cold phase. However, neither of the gridded analyses shows as strong a dichotomy between the daily rainfall in the warm and cold ENSO phases as does the station data. While the K-S test for the histograms for Nova Palmira station data show significant differences between the warm and cold ENSO phases at the 99.5% level in ASO, the significance level falls to 93% in the LA gridded analyses, and in the CPC analysis the differences are not statistically significant (Table 2). A similar analysis for OND, also in Table 2, shows some improvement in the gridded data significance levels, but the levels of significance are still less than those for the station data. Given the general properties of gridded datasets based on station observations pointed out in Robeson and Ensor (2006), the gridded analyses tend to underestimate the extreme daily precipitation amounts and conversely to overestimate the number of days with very light precipitation. It is not surprising that in the example presented here the gridded data underestimate the strength of the ENSO-related differences in the character of daily rainfall compared to the analysis of a representative station within the grid, and illustrates the need for station data that have long records of reliable daily observations.
A comparison of the time series of precipitation greater than 20 mm day−1 for the LA grid box containing the Nova Palmira station (Fig. 5) shows generally good agreement between the single station and the grid values for both ASO and OND. Somewhat surprisingly there is no bias between the mean number of days with rainfall greater than 20 mm between the station and the LA grid for ASO. Both the LA grid and stations show an average of 8 days with precipitation greater than 20 mm. The time series based on the CPC grid shows a smaller number of days with rainfall equal to or greater than 20 mm (6.5 days per season) than the Nova Palmira station, but the overall behavior of the time series are very similar. For the OND season the LA analysis shows slightly fewer days with greater than 20 mm (6.5 days for the station date versus 6 days for the LA grid). The CPC gridded analysis shows a larger bias in the mean number of days with precipitation greater than 20 mm (5.0 days in the CPC grid). Despite larger biases the time series of rainfall equal to or greater than 20 mm for the LA and CPC gridded analyses for OND are also generally consistent with the Nova Palmira station.
Maps of areas with statistically different daily precipitation histograms based on the K-S statistics for the gridded daily rainfall data in South America are compared to available station data for ASO (Fig. 6) and for OND (Fig. 7). In the limited areas where the station and gridded analyses overlap, primarily in southern Brazil, there is general agreement between the station and gridded analyses for both seasons. In ASO the LA analysis shows an area, defined by several grid points, where the ENSO influence on daily rainfall distributions is in the same direction as discussed for Nova Palmira and the coincident grid box, that is, relatively wetter during the warm phase (Fig. 6a). The analysis based on the CPC grids (Fig. 6b) is in general agreement, but with even greater statistical significance in the areas south of Nova Palmira. Both gridded analyses also indicate areas with a tendency for drier conditions during warm ENSO episodes for the states of Maranhao, Para, and Amapa in Brazil that extend into French Guiana and Surinam in the LA analyses. These are consistent with the areas of drier bimonthly (Aceituno 1988) and seasonal rainfall totals (e.g., Ropelewski and Halpert 1996; Liebmann and Marengo 2001; Uvo et al. 1998). The shifts toward drier conditions have greater spatial extent and stronger statistical significance in the CPC gridded analysis (Fig. 6b) than in the LA maps (Fig. 6a).
Both gridded datasets for ASO also suggest that there are areas with significant differences in the histograms of daily rainfall in central Brazil. However, there is little spatial coherence in central-west and northwest Brazil in either gridded set. Thus, while the gridded analyses might be useful in identifying potential areas with significant ENSO-related shifts in daily rainfall, the suggestion of significant shifts in the histograms of daily precipitation in central Brazil during the ASO season requires closer examination with daily station data.
The CPC analysis for the OND (Fig. 7b) season shows a continuation of dry conditions in the Guiana Highlands of Amazonas and a tendency for wetter conditions in west-central Brazil that does not appear either in the station analysis or the LA gridded analysis (Figs. 7a and 7c). This discrepancy will require further investigation. In general, however, the LA and CPC gridded analyses agree in identifying regions with shifts toward wetter conditions in southern Brazil, spilling into Uruguay and Argentina in the LA analysis. Both gridded analyses suggest wetter conditions during OND of warm ENSO episodes extending from eastern Mato Grosso eastward to the Atlantic coasts (Figs. 7a and 7b). However, these wetter conditions in OND are not supported by station data or the NCEP–NCAR reanalysis. Analyses by Grimm (2003, 2004) suggest that the tendency for wetter conditions in this part of Brazil are more evident in southern summer, December–February (DJF), with a transition between dry and wet conditions associated with ENSO occurring between October and December. In general, the CPC analyses for Brazil appear to show overall higher levels of statistical significance in the differences between the daily rainfall histograms for warm and cold ENSO episodes.
An examination of the K-S statistics, in both gridded datasets, for each of the 12 overlapping 3-month seasons through the annual cycle, shows good agreement in the areas identified with statistically significant differences in the daily rainfall histograms and the previous analyses based on seasonal totals (e.g., Ropelewski and Halpert 1987, 1989, 1996; Aceituno 1988; Pisciottano et al. 1994; Grimm et al. 1998, 2000; Uvo et al. 1998). In addition to these well-documented areas with ENSO-influenced precipitation in northeastern Brazil and southeastern South America the gridded daily rainfall is consistent with the more recent analysis of Grimm (2003, 2004) showing ENSO-related shifts of the daily rainfall in regions of central Brazil, discussed above. The gridded analyses show a continuous evolution through the annual cycle, with maximal spatial extent and statistical significance occurring in December–February (Figs. 8c and 9c). The analysis in Grimm (2004) suggests that statistically significant shifts in total seasonal rainfall amounts are primarily associated with La Niña episodes in the important agricultural areas of Mato Grosso eastward through Bahia.
The analyses using the LA gridded data show the evolution of dry conditions associated with ENSO warm episodes in northeastern South America (Fig. 8). A relatively small area of dry conditions appears in Surinam and French Guiana in the November–January season (Fig. 8b) and slowly expands, then moves into northeastern Brazil by February–April (Fig. 8e) and continues through the April–June season (Fig. 8g). Indications of drier, warm ENSO-related conditions then weaken and migrate northward. A similar evolution is seen in the CPC analysis, which is limited to Brazil only, but there are also some interesting differences (Fig. 9). In the CPC gridded analysis drier conditions related to warm-episode years are evident during OND (Fig. 9a) grow in strength during the period from November through February (Figs. 9b and 9c), expand, and migrate southeastward to the state of Ceara (Figs. 9d and 9e). A maximum in the areal extent of highly significant differences between the histograms of daily rainfall related to warm and cold ENSO episodes, that is, with warm-episode years showing shifts to drier conditions, occurs in MAM and AMJ (Figs. 9f and 9g). This evolution continues with a northwestward migration of the areas of statistical significance with both a decrease in the areal extent and decrease in statistical significance (Figs. 9h–l). The areal extent, duration, and statistical significance of these dry conditions associated with ENSO are all larger in the CPC analysis than in the LA analysis. The relative differences between the LA and CPC analyses can also be seen in the evolution of the warm-episode-related enhanced precipitation in southern and southeastern Brazil from August through December for the LA gridded data (Figs. 8j–b), and for the CPC grids (Figs. 9j–b). Differences include, for instance, the larger area of K-S statistic significance in the CPC in each of the 3-month seasons in this sequence. Further analysis of grids with respect to station data, where such datasets exist, is required to fully understand the differences between the gridded datasets. However, we note that the evolution in the area where the daily histograms are statistically significant in the CPC gridded analyses is more consistent with analysis of seasonal totals in Grimm et al. (1998) than the evolution depicted in the LA analysis.
c. NCEP–NCAR reanalysis
It is well known that the precipitation in the NCEP–NCAR reanalysis is biased compared to observations and has particular errors for northeastern Brazil (e.g., Kistler et al. 2001). Smith and Ropelewski (1997) document spatial biases in the ability to replicate ENSO-related precipitation patterns in a version of the NCEP model and data assimilation system similar to the version used in the NCEP–NCAR reanalysis That study suggested that even though absolute values of rainfall are poorly replicated in the reanalysis in several parts of the globe, the ENSO-related shifts in rainfall totals do a credible job of replicating signs and relative shifts seen in observations of seasonal rainfall amounts. Since there are large parts of the world without adequate daily precipitation records, and because the few records that do exist are too sparse to develop gridded rainfall datasets such as those discussed here, it seems worthwhile to examine the ability of the NCEP–NCAR reanalysis to identify places in the world with significant ENSO-related shifts in daily rainfall histograms, even if the exact nature of these shifts is not well represented by the reanalysis. To this end we repeat the analyses presented in Figs. 8 and 9 using the NCEP–NCAR rainfall estimates for the 1961–90 period. The K-S statistics in the NCEP–NCAR reanalysis appear to have generally higher levels of statistical significance through the annual cycle (Fig. 10) than either of the gridded analyses based on station data (Figs. 8 and 9). Since the reanalysis grid size of ∼1.875° latitude by 1.875° longitude for daily precipitation (identified as PRATE in the reanalysis dataset) is much larger than the 1° latitude by 1° longitude station-based gridded analyses, we expected that the reanalysis would show overall lower values of statistical significance because of the greater spatial smoothing of daily rainfall. However, in the trade-off between the tendency to reduce the magnitude of the daily statistics because of the spatial averaging and the tendency of models to be very sensitive to tropical forcing, the model sensitivity appears to dominate. Nonetheless, to a first approximation the reanalysis seems to do a credible job of replicating the areas with statistically significant shifts in daily rainfall histograms, with some notable exceptions. In particular, the reanalysis does not capture the timing and areal extent of the ENSO-related shifts in precipitation for southeastern South America that are well documented in previous studies (e.g., Grimm et al. 2000) and apparent in our analysis of station data for Brazil as well as in the gridded station analyses. We note further that a comparison of the time series of the number of days with rainfall greater than 20 mm at Nova Palmira and the closest reanalysis grid point for OND (not shown) indicate large differences.
There is not close agreement in comparisons of the reanalysis to the gridded data in central Brazil. However, the reanalysis does show a tendency for wet conditions from Mato Grosso eastward during ENSO warm episodes that peaks in the DJF season shown in Grimm (2003, 2004). However, the reanalysis appears to overemphasize the warm-episode-related dry conditions in Ceará compared to the gridded analyses, for example, Figs. 10d–h, compared to the CPC (Figs. 9d–h) and the LA analyses (Figs. 8d–h). This may be related to the tendency of the reanalysis to set up a fictitious, local, geographically induced, land–sea circulation and corresponding unrealistic rainfall pattern (V. Kousky 2005, personal communication).
Despite these shortcomings the reanalysis appears able to capture the general character of the ENSO influences on daily rainfall consistent with the gridded station analysis. As one example, in addition to the features discussed above, the reanalysis appears to have captured the tendency for wetter conditions in Venezuela also evident in the gridded station analysis in a comparison of Figs. 8e–h and 10e–h for the February to July period.
4. Summary and discussion
In this study we first present an example of how the histograms of observed daily weather at a station, Nova Palmira, are modulated by ENSO phase and how that modulation evolves in two different seasons, ASO and OND. We illustrate, using the same station example, that the character of the rainy season (i.e., the statistics of daily weather) may have significant differences for El Niño versus La Niña years even though seasonal rainfall totals compared to amounts typical of neutral conditions may differ by relatively small amounts. Not all of these differences are statistically significant but suggest further analyses with more stations with complete daily station records may be warranted.
We also examine gridded station observations, formed from all available observations within the grid area, and compare them to station data. Even though the gridded analyses tend to bias the daily rainfall records by overestimating the number of days with rainfall and underestimating the number of days with larger rainfall amounts compared to stations, comparison of two of the gridded analyses available for South America, CPC and LA, with station data suggests that these analyses can be used to guide the identification of regions with significant shifts in the character of daily rainfall conditional on ENSO. There is general agreement between each of the gridded analyses and the few stations with nearly complete daily records available for analysis. The existence of some differences in the comparisons of these analyses should motivate the development of more complete daily station data. On the other hand, the gridded analyses based on station observations are likely to be more useful for comparisons to models with similar grid scales than comparisons between model-generated precipitation and station data.
We also perform a preliminary comparison between the NCEP–NCAR reanalysis daily precipitation statistics and the gridded datasets. The reanalysis shows many areas of agreement with the gridded stations analysis but fails to identify some regions of significant ENSO influence, for instance, southeastern South America. The reanalysis also has some regions of disagreement, particularly in Nordeste, Brazil. This suggests that the unadjusted NCEP–NCAR reanalysis data cannot be relied on as a faithful substitute for observations in the analysis of the statistics of daily rainfall in South America. Nonetheless, recent research (Ines and Hansen 2006) also suggests that model-based precipitation such as the NCEP–NCAR reanalysis can be adjusted through application of fairly simple techniques, such as removal of the constant bias and shifts of the daily distribution of the reanalysis rainfall to conform more closely to the observations. The extent to which these bias corrections and adjustments are a function of location and season is a research topic that requires thorough investigation.
In summary, while the climate system will not allow the prediction of the exact sequence of weather events within a season, this preliminary study suggests that it may be possible to provide some useful probabilistic information on the shifts in the general character of the rainfall season, at least those associated with ENSO. It is hoped that the examples shown here will motivate further efforts to gather and archive daily rainfall data and make them available for analysis.
We wish to thank our colleagues at the IRI, especially Jim Hansen for his comments on an earlier version of this work and Benno Blumenthal and the Data Library staff for technical support with the analysis. We also wish to thank the reviewers, particularly Reviewer B, for extremely helpful comments and suggestions. This work was funded by a grant/cooperative agreement from the National Oceanic and Atmospheric Administration, NA050AR4311004. The views expressed herein are those of the authors and do not necessarily reflect the views of NOAA or any of its subagencies.
* Current affiliation: UCAR Visiting Scientist, NOAA/Climate Program Office, Silver Spring, Maryland
Corresponding author address: M. A. Bell, IRI, The Earth Institute at Columbia University, 61 Rte. 9W, Monell Building, Palisades, NY 10964-8000. Email: firstname.lastname@example.org