Detection of forced changes in the precipitation distribution using ridge regression

Detection and attribution (D&A) of anthropogenically forced changes to precipitation is challenging due to the high internal variability of precipitation and the limited spatial and temporal coverage of the observational records. These factors result in a low signal-to-noise ratio of potential regional and even global trends.

Here, we use a statistical method – regularised linear regression, or ridge regression – to create physically interpretable detection metrics (fingerprints) for D&A of changes in the precipitation distribution with a high signal-to-noise ratio. The regression coefficients that make up the fingerprints of forced change are based on the CMIP6 multi-model archive data masked to match observational coverage, and are then applied to gridded precipitation observations to assess the degree of forced change detectable in the real-world climate. 

We show that the signature of forced change is detected and attributed to external forcing in two different observational datasets in global metrics of mean and extreme precipitation (PRCPTOT, and Rx1d, respectively). If the global mean trend is removed from the data, forced changes are still detected, indicating that climate change affects the spatial patterns of precipitation, and increasing confidence in the results of this method for D&A of precipitation, as well as in climate models capturing the relevant processes that contribute to the regional patterns of change. Furthermore, we show the sensitivity of our D&A results to several ‘design choices’, including target metric of forced change, regularisation parameter, season of interest, (spatial coverage of) observational dataset used, the forced trend length and the region of interest (tropics vs. subtropics). 

The method is largely insensitive to target metric and regularisation parameter, increasing confidence in the robustness of the results. However, we find that June-July-August generally has low forced trend signal-to-noise ratio in both mean and extreme precipitation. Furthermore, the observational dataset choice affects detectability not only through coverage differences but also dataset disagreement, and the chosen trend length can result in different forced trends when comparing observations to model projections. These sensitivities may explain apparent contradictions in recent studies on whether models under- or overestimate the observed increase in extreme precipitation. Lastly, the detection models are found to rely primarily on the signal in extratropical northern hemisphere data, which is at least partly due to observational coverage, but potentially also due to presence of a more robust signal in the northern hemisphere in general. When these regions are excluded, detection of significant forced changes is no longer possible, which may also have implications for the ability to assess risks and inform adaptation policies in the tropics and the global south.

Ridge regression is powerful for D&A of the precipitation distribution, opening up possibilities for extension of the method to learn more about mechanisms driving forced changes in precipitation. However, internal variability, limited coverage in time and space, and dataset disagreement in precipitation data continue to play a large role.