Climate teleconnections, such as El Niño's influence on precipitation in distant regions, have long been sought using correlations between time series. However, correlation does not distinguish cause from effect, and its significance strongly depends on series length and autocorrelations. This work stands alongside studies that use network methods and statistical tests to identify climate links, but most of them rely on correlation. The authors propose an alternative that explicitly tests the predictive power of one variable for another.
We used MERRA-2 reanalysis data on surface temperature and precipitation from 1980 to 2018, averaged to daily values on a 1°×1.25° grid. Anomalies were detrended and deseasonalized, then smoothed with weekly means. For each land cell, two measures of association with temperature in the Niño 3.4 region were calculated: maximum lagged Pearson correlation and bivariate Granger causality with lags up to 90 or 180 days. Correlation significance was assessed via surrogate data (AAFT method), while causality used an F-test with chi-square approximation. A persistence metric was introduced: the number of temporal resolutions (from 7 to 30 days) at which the association is significant.
Both methods reproduce known El Niño teleconnection regions well: the equatorial Pacific, Australia, Amazonia, southern Africa, and the Horn of Africa. However, Granger causality revealed a significant link with precipitation in central India, where correlation is weak and unstable. For western North America, the Granger method detected a signal in southern California that correlation missed. Granger causality is sensitive to the maximum lag, whereas correlation barely changes. Global metrics (AUC, F1) show that Granger causality has higher precision and recall at strict significance thresholds.
This work shows that Granger causality can complement correlation methods, especially in cases of weak signal-to-noise ratio. Unlike neighboring studies that often restrict themselves to correlation networks, here the direction of influence is explicitly tested and information from all lags is considered. This allows not only finding links but also estimating their time scale. New regions identified only by causality require physical explanation, which may stimulate the search for mechanisms.
The method can be extended to multivariate Granger causality to distinguish direct and indirect influences and build causal graphs. This would help trace chains of signal propagation through the atmosphere and ocean. It is also worth investigating the influence of maximum lag and filtering methods on the robustness of results.
This work could affect climate modeling, seasonal forecasting, and attribution of extreme events.
Verify the newly identified teleconnections on independent data and with climate models. Develop a multivariate version of the method to analyze indirect links.
An open question in climatology is how to distinguish true teleconnections from spurious correlations caused by common trends or autocorrelation. This method directly addresses that problem by testing predictive power rather than just synchronous association.
The correlation between precipitation in India and El Niño has weakened in recent decades, but Granger causality still finds a significant link.