An Exploratory Analysis of a Record of El Nino Events, 1800-1987
Solow, Andrew R., Journal of the American Statistical Association
The El Nino-Southern Oscillation (ENSO) refers to a quasi-periodic redistribution of heat and momentum in the Pacific Ocean. A comprehensive review of this phenomenon, which is the largest source of interannual climatic variability on a global scale, is contained in the monograph by Philander (1990). The El Nino phase of ENSO, which occurs on average every four years, is characterized by a dramatic warming of the eastern equatorial Pacific. This warming is associated with large-scale meteorological and oceanographic anomalies. Some of these anomalies, such as reduced tropical cyclone activity in the western Pacific, weaker upwelling along the west coast of South America, and increased winter rainfall in the southern United States, are consistent from event to event. More dramatic but less consistent anomalies - such as the recent flooding of the Mississippi River system - are also associated with El Nino events. The cumulative economic effects of El Nino are substantial, and there is considerable interest in understanding and ultimately predicting the occurrence and magnitude of El Nino events.
Current efforts aimed at predicting El Nino events fall into two broad categories: methods based on physical models (see, for example, Cane and Zebiak 1987 and Inoue and O'Brien 1984) and empirical methods (see, for example, Barnett 1984 and Barnston and Ropelewski 1992). On the empirical side, the general approach has been to apply multivariate techniques to historical records of spatial variables like sea surface temperature in the equatorial Pacific to identify patterns (called precursors) that tend to precede El Nino events. This approach is limited by the availability of comprehensive historical oceanographic and meteorological data.
A second line of research, for the most part unconnected to prediction, has focused on the construction of a long-term historical record of the occurrence of El Nino events. For example, using a variety of instrumental and documentary records, Quinn, Neal, and Antunez de Mayolo (1987) constructed a record of major El Nino events covering the past 450 years. Moreover, they were able to classify each event since 1800 as either moderate or strong. This part of the record is reproduced in Table 1.
In fact, Quinn et al. (1987) provided a somewhat finer classification of event strengths. I use the coarser classification in this article for two reasons. First, the classification scheme is subjective to some extent, and it was felt that the coarser classification would contain fewer errors. Second, because the record contains only 50 events, the benefits in terms of model economy achieved by lumping the finer classification were felt to outweigh the costs in terms of description.
Beyond their historical interest, the data in Table 1 can, in principle, be used in El Nino prediction. One approach is to exploit regularities in these data in making probabilistic forecasts about the timing and magnitude of future El Nino events. A more useful approach is to use these data in conjunction with model experiments or observations of El Nino precursors. In either case, it is necessary to develop a statistical model of the data in Table 1. The purpose of this article is to take a step in that direction.
In statistical terminology, the record in Table 1 is called a marked point process (see, for example, Cox and Isham 1980, p. 131), with each event in the point process having a binary mark. The simplest model of such a process is one in which the events follow a homogeneous Poisson process and their marks are both independent of the point process and of each other. With this model as a starting point, the approach taken herein is to identify the most parsimonious model consistent with the data. This approach is implemented by testing the current model against certain alternatives. Because there is little scientific guidance about plausible models for these data, the alternatives considered are those that seem natural from a statistical viewpoint. …