Bias in the variance of gridded data sets leads to misleading conclusions about changes in climate variability

Many studies addressing climate change and climate variability over large regions rely on gridded data. Grids are preferred to station-based data sets because they help avoiding bias arising from the irregular spatial distribution of the observations. However, while spatial interpolation techniques used for constructing gridded data are good at preserving the mean of the data, they do not offer an adequate representation of their variance. In fact, the grid's variance depends largely on the spatial density of observations used for constructing it. Most global and regional climate data sets are characterized by large temporal changes in the number of observations available for interpolation, with a strong reduction in the last 30 years. These changes in the sample size result in changes in the variance of gridded data that are merely an effect of the interpolation process, and ignoring this fact may lead to erroneous conclusions about changes in climate variability and extremes. We discuss this problem and we demonstrate its importance with a widely used global dataset of temperature and precipitation. We propose to move from interpolation techniques towards statistical simulation approaches that provide a better representation of climate variability when constructing climatic grids.