Sensitivity of trends to estimation methods and quantification of subsampling effects in global radiosounding temperature and humidity time series

Abstract

Climate trends estimated using historical radiosounding time series may be significantly affected by the choice of the regression method to use, as well as by a subsampling of the dataset often adopted in specific applications. These are contributions to the uncertainty of trend estimations, which have been quantified in literature, although on specific pairs of regression methods, and in the not very recent past characterized by smaller trends in temperature than those observed over the last two decades. This paper investigates the sensitivity of trend estimations to four linear regression methods (parametric and nonparametric) and to the artificial subsampling of the same dataset using historical radiosounding time series from 1978 onwards, available in the version 2 of the Integrated Global Radiosonde Archive (IGRA). Results show that long-term decadal trends may have not negligible uncertainties related to the choice of the regression method, the percentage of data available, the amount of missing data and the number of stations selected in the dataset. The choice of the regression methods increases uncertainties in the decadal trends ranging from −0.10 to −0.01 KÁda-1 for temperature in the lower stratosphere at 100 hPa and from 0.2 to 0.8% da-1 for relative humidity (RH) in the middle troposphere at 300 hPa. Differences can also increase up to 0.4 KÁda-1 at 300 hPa when the amount of missing data exceeds 50% of the original dataset for temperature, while for RH, significant differences are observed in the lower troposphere at 925 hPa for almost all datasets. Finally, subsampling effects on trend estimation are quantified by artificially reducing the size of the IGRA dataset: Results show that subsampling effects on trend estimations when at least 60 stations, up to 76% of data available, are considered for temperature and at least 40 stations for RH.

Potential Contribution of Climate Conditions on COVID-19 Pandemic Transmission over West and North African Countries

“…Note that the main reason for using the non-parametric Kendall and Spearman statistical tests is that they tend to be more powerful and better suited for non-normally distributed variables compared with parametric statistical tests such as the Pearson test [48]. As the distribution of COVID-19 cases data and climatic variables is not necessarily Gaussian, the non-parametric Kendall and Spearman statistical tests might be more appropriate for testing the null hypothesis of the association/correlation between the two given variables [49,50].…”

Potential contribution of climate conditions on COVID-19 pandemic transmission over West and North African countries

“…The correlation between the basic meteorological variables and virus transmission over 275 days was investigated from March 1, 2020 to November 30, 2020 at each of the considered country and at regional level. To assess the correlation, differently from other studies [29–31], here, we considered as a reliable variable the residuals of the daily new positive cases with respect to the robust and non-parametric Median-Based Linear Model (MBLM) defined in [49]. Further, when working on the residuals, makes our analysis independent on the analyzed time period.…”

“…The main reason for using the non-parametric Kendall and Spearman statistical test is that it tends to be more powerful and better suited for non-normally distributed variables compared with parametric statistical tests such as the Pearson test [48]. As the distribution of COVID-19 cases data and climate variables are not necessarily Gaussian, the non-parametric Kendall and Spearman statistical tests might be more appropriate for testing the null hypothesis of the association/correlation between the two given variables [49,50].…”

“…In addition, the correlation coefficients are often computed between two time series (e.g., two time series representing, respectively, the number of daily COVID-19 cases and daily average temperatures or humidity) [29–31], without taking into account for the possible presence of cycles or temporal trends in the data, which can strongly affect the estimated correlation values and yield artefactual associations. In the present paper, the daily cases of COVID-19 are fitted with the robust and non-parametric Median-Based Linear Model (MBLM) defined in [49]. As it can be easily observed (see section next section) that the COVID-19 cases data does not follow a Gaussian distribution, however as discussed above, the non-parametric MBLM model might be more powerful and better suited compared with parametric linear models [49].…”

“…In the present paper, the daily cases of COVID-19 are fitted with the robust and non-parametric Median-Based Linear Model (MBLM) defined in [49]. As it can be easily observed (see section next section) that the COVID-19 cases data does not follow a Gaussian distribution, however as discussed above, the non-parametric MBLM model might be more powerful and better suited compared with parametric linear models [49]. Considering the residuals, makes the analysis independent on the period of lockdown or restriction, being the correlation analysis strongly dependent on the considered time period.…”