key: cord-0442040-00y2cf5d authors: Raimondo, Sebastian; Benigni, Barbara; Domenico, Manlio De title: Environmental conditions and human activity nexus. The case of Northern Italy during COVID-19 lockdown date: 2020-10-15 journal: nan DOI: nan sha: e3ac3b5c9e63bf14e47ec66f1428aa692ebad93b doc_id: 442040 cord_uid: 00y2cf5d During COVID-19, draconian countermeasures forbidding non-essential human activities have been adopted worldwide, providing an unprecedented setup for testing sustainability policies. We unravel causal relationships among 16 environmental conditions and human activity variables and argue that, despite a measurable decrease in NO2 concentration due to human activities, locking down a region is insufficient to significantly reduce emissions. Policy strategies more effective than lockdowns must be considered for pollution control and climate change mitigation. conditions -i.e air pollution and meteorological conditions data -and human activity -i.e. mobility and energy load data -over a survey period of four months -February, March, April and May -both in 2019 and in 2020, for a total of 16 variables (Fig. 1 ). By taking into account all these variables we reduce the possible disturbances due to latent confounders. Afterwards, we evaluated the differences in the dynamics of such variables between the period related to a situation with extremely reduced activity and baseline periods. Since the lockdown imposed social distancing and, consequently, drastically reduced human mobility, we referred to an indicator of air quality which is sensitive to mobility changes, i.e. the nitrogen dioxide (NO 2 ) concentrations, which proves to highly depend on emissions from transportation means and industrial activity 18 24 and Granger causality (GC) 25 . Finally, we assessed the robustness of the results found with statistical tests, by taking inspiration from a relatively recent Bayesian technique, based on a state-space model, used to infer the causal impact of advertising campaign on the market sales 26 . By specifying which period in the data should be used for training the state-space model (pre-intervention period) and which period for computing a counterfactual prediction, this technique assesses the impact of the attributable intervention 26 . trends due to lack of publicly available data for the year 2019. The results of the causal analysis are reported in Fig. 2 . The partial correlation network in Fig. 2 a) reveals the stronger relations between all human activity variables, which vary synchronously, driven by the lockdown interventions. In particular the variable "change in movement" is the most connected. The negative relation between "residential" and "work-places" is well captured by the method, likewise for the "total precipitation", "surface radiation" and "temperature". The NO 2 variable is related just to the solar radiation, which may be interpreted as the influence of the seasonal effects mentioned in the introduction. Fig. 2 c) ) obtaining a cumulative concentration of 132.37 · 10 3 mol/m 2 which would have been equal to 178.31 · 10 3 mol/m 2 in absence of the lockdown. Even though the results seems to reveal a clear causal impact of the lockdown on the NO 2 concentration, the computed probability of obtaining this effect by chance is p = 0.114. This suggests that the effect might be due to chance, or to other uncontrollable issues related to the data; for example, because the lockdown period includes data where the effect of the lockdown has already worn off, or because the regressors are not well correlated with the NO 2 . In this work, we investigated the relationship between human activity and environmental conditions by leveraging the extraordinary event of the Italian lockdown due to COVID-19. To this aim we have fused heterogeneous data sources -including human mobility, total energy load, meteorological conditions and NO 2 concentrations -in four different periods over 2019 and 2020. We have found evidence that, concomitantly with the reduction in both the mobility and the energy demand, the NO2 average concentration significantly decreases in 2020 lockdown w.r.t the same period in the previous year. The lower variance of NO 2 during the 2020 lockdown is even more visible than the one in average concentration. This is attributable to the complex nature of the human-environmental system where internal factors may act as filter on the rapid variations of the NO 2 . Shedding light on this factor would unveil new possible features of the system under consideration that could be useful for the development of more reliable, data-informed, local climate-change and pollution control policies. Given the complexity of the system, where drivers of different nature contribute in affecting the air quality, we determined and quantified the causal roles of these drivers on the system itself. Overall, the analysis detected the sign of a causal relation between the relaxation of a broad spectrum of human activities and air pollution abatement during the lockdown in Northern Italy. However, the statistical significance of the results is questionable, and can be improved with additional data about human mobility and considering additional variables to reduce the noise of latent confounders. Nevertheless, the causal impact may still remain nebulous because of the inherent behaviour of the system. In the light of these findings, we consider our approach to be indicative, but not definitive, for investigating the nexus between environmental conditions and human activity during COVID-19 lockdown in Italy. Further developments could complement our analysis with mobility data of 2019 and could include more detailed data such as stratified transport (i.e. heavy and light transport). In fact, it is known that motorway traffic has decreased considerably during the lockdown, but the reduction in the circulation of heavy transport has not been so similarly comparable, due to the delivery of essential commodities. For example, in March even though the highway total vehicles dropped by 63%, the heavy vehicles decreased by just 27%: the latter pollutes four times more than light transport, according to the Italian National Autonomous Roads Corporation (ANAS). Here, we proved the power of coupling data analysis with causal inference, to study the nexus between environmental conditions human activity from a systemic perspective. Moreover, we shed some light on the causal relations between human activities and NO 2 concentrations, highlighting that a lockdown may be not enough in changing pollutants concentrations and consequently in being regarded as a desirable or even a potential strategy for climate change mitigation and sustainability. While the continuity in the essentials supply chain proves the efficiency of the region in providing indispensable services during emergencies such as the COVID-19 pandemic, it also testifies that the backbone of human activities has never really stopped. Consequently, it is plausible to hypothesize that such a backbone might play a more significant role than non-essential one, providing serious challenges to policy and decision making to build a sustainable society in response to climate change. We firmly believe that a paradigm shift towards a complex systems view is necessary for the optimal management of resources. Such an approach would be of tremendous help for decision making processes allowing for more informed and integrated choices, especially with a view to the development of mitigation policies in accordance with climatic and environmental goals (e.g. Sustainable Development Goals of Agenda 2030). In this work we relied on data of nitrogen dioxide (NO 2 concentrations) from Copernicus Sentinel-5P satellite from the 1 st of January 2019 to the 1 st of June 2020 (TROPOMI Level 2 Nitrogen Dioxide total column products. Version 01. European Space Agency 27 ). In particular, we referred to high-resolution daily concentrations of the tropospheric NO 2 over Lombardia region. Data of meteorological conditions are retrieved from the Copernicus Climate Change Service 28 (ERA5-Land reanalysis dataset) and consist of hourly data of 6 variables -i.e surface pressure, surface net solar radiation, temperature, total precipitation, wind direction and wind speed -over the Lombardia region. These data were averaged daily. Google mobility data 19 are provided in terms of daily length of stay at different places -e.g. residence, grocery, parks, etc. -aggregated at regional level. Apple data 20 are provided in terms of variation in the volume of driving directions requests while Facebook data 21 in terms of positive or negative change in movement relative to baseline (February 2020). As for NO 2 , we considered the case of the Lombardia region covering collectively the time period from 13 th of January to the 27 th of July 2020. All these mobility data are available only for the year 2020. For what concern energy load, we considered data of Northen Italy for the period from 1 st of January 2019 to the 27 th of July 2020 from the Italian transmission system operator Terna 29 . "Northen Italy" is the smallest available space aggregation which includes Lombardia; since Lombardia has the highest energy demand in this area, Northen Italy data are deemed appropriate for our purposes. To determine whether data reflect the regime shift of the lockdown we implemented a shifting point detection technique based on an information-theoretic measure of similarity. Through this data-driven procedure we found the date which splits the survey period in pre-lockdown and lockdown periods. This is used both for statistical tests and for the Bayesian state-space model. In particular, we applied the Jensen-Shannon divergence to shifting subsets of the NO 2 time series (of length N ) for the 2020. The main result of the process is the date of the tipping point in which the dynamics meets a regime shift. The date of the "information-theoretic" lockdown turns out to be the 14 th of March (p-value = 6.12 · To rigorously assess the differences in the time series of the considered variables, we tested the differences in mean between each period with t-tests and surrogate data tests. In addition, we evaluated the differences in the variances between the above periods with F-test, Bartlett's test and Fligner test. The magnitude of the differences observed in the time series are evaluated through two effect-size measures: the Cliff-δ and the C.L.E.S. (Common Language Effect Size). The Cliff-δ is computed by enumerating the number of occurrences of an observation from one group having a higher response value than an observation from the second group, and the number of occurrences of the reverse: where the two time series are of size n 1 and n 2 . The C.L.E.S. instead is defined as the probability that a randomly selected individual from one group have a higher score on a variable than a randomly selected individual from another group. We computed this measure numerically with a brute-force approach. Afterwards, to obtain a statistical indication of the possible causal relations between the 16 variables, we build the partial correlation matrix. The partial correlation measures the strength and the direction of the (rank) dependence of two variables from a set of random variables when the influence of the remaining variables is removed. We further investigated the causal dependence through the "predictive causality" measure by C. Granger, and we compared the results with a step-wise linear regression with the AIC selection criteria. It is to be noticed that not all the required constraints for Granger causality are strictly respected (i.e. linearity and stationarity of the observations). For this reason and to assess the robustness of the results found with the statistical tests, we employed a Bayesian modeling technique. This technique uses a state-space model to predict what would have been the system evolution after an "intervention", if the intervention had never occurred. For this method to work correctly, the covariates themselves must not be affected by the intervention and the relationship between covariates and treated time series, as established during the pre-intervention, must remains stable throughout the post-intervention period. This procedure does not require linearity nor stationarity in the data but, in turn, it cannot find a direct causal nexus between NO 2 concentration and the variables affected by the lockdown measures. In our case, the idea is to detect a possible deviation of the NO 2 time series from the counterfactual prediction, when the lockdown is the "intervention" and the meteorological conditions are the covariates. If the meteorological conditions were the only variables that influence the NO 2 concentration, we would expect the data to follow the counterfactual prediction for the post-treatment period. If, instead, other variables (the ones related to human activity restrictions) were responsible for the variation in the NO 2 concentrations the deviation from the counterfactual would be clear. Thanks to this approach we stated the impact of the reduced human mobility and energy consumption due to the lockdown, considered as the attributable intervention on NO 2 concentrations. Changes in contact patterns shape the dynamics of the covid-19 outbreak in china Nowcasting and forecasting the potential domestic and international spread of the 2019-ncov outbreak originating in wuhan, china: a modelling study Estimating the burden of sars-cov-2 in france Introductions and early spread of sars-cov-2 in the new york city area Reducing transmission of sars-cov-2 Assessing the impact of coordinated covid-19 exit strategies across europe Have deaths from covid-19 in europe plateaued due to herd immunity? Mathematical models to guide pandemic response The effect of travel restrictions on the spread of the 2019 novel coronavirus (covid-19) outbreak Inferring change points in the spread of covid-19 reveals the effectiveness of interventions Effects of non-pharmaceutical interventions on covid-19 cases, deaths, and demand for hospital services in the uk: a modelling study Which interventions work best in a pandemic? Changes in air quality during the lockdown in barcelona (spain) one month into the sars-cov-2 epidemic Assessing air quality changes in large cities during covid-19 lockdowns: The impacts of traffic-free urban conditions in almaty, kazakhstan Does lockdown reduce air pollution? evidence from 44 cities in northern china Covid-19 lockdowns cause global air pollution declines Severe air pollution events not avoided by reduced anthropogenic activities during covid-19 outbreak. Resources, Conservation and Recycling Rapporto sulla qualità dell'aria di milano e provincia Google COVID-19 Community Mobility Reports Retrieved on July 2020 from Facebook A common language effect size statistic It's the effect size, stupid: What effect size is and why it is important Partial correlation coefficients. Encyclopedia of statistics in behavioral science Investigating causal relations by econometric models and cross-spectral methods Inferring causal impact using bayesian structural time-series models