Latest: Using machine learning and synthetic control methods to study the impact of Wuhan's closure on air pollution and health!

Anyone who engages in econometrics pays attention to this account

Manuscript: [email protected]

All the code programs, macro and micro databases and various software of the econometric circle methodology are placed in the community. Welcome to the econometric circle community for exchanges and visits.

For machine learning methods, you can refer to the following articles: ①"Machine learning methods appeared in top journals such as AER, JPE, QJE, etc.!", ②Frontier: Summary of machine learning application classification in the fields of finance and energy economy, ③Text analysis How to do the steps, tools, pathways and visualizations? ④The application of text big data analysis in economics and finance, the most comprehensive literature review, ⑤the most comprehensive: the summary of the application status and frontier outlook of deep learning in the field of economic and financial management, young and middle-aged scholars can't help but pay attention! ⑥Top Frontier: Machine learning in agriculture and applied economics, its comparison with econometrics, you will be out if you don't read it or not! ⑦Machine learning and big data econometrics, you must read this, ⑧Machine learning and Econometrics book recommendations, worthy classics, ⑨The latest trends in the application of machine learning in micrometrics: big data and causal inference, ⑩machines Study the first book, data mining, reasoning and prediction, ⑩Top, machine learning is an applied econometric method, and I don’t understand the danger of being eliminated in the future!

Regarding the synthesis control method, you can refer to the following articles: ①The 33 selected Articles on the synthesis control method SCM! The group alarmed Alibaba! ②Let’s talk about the synthetic control method SCM, to help you find the control group of causal inference, ③What’s the ghost of synthetic control method? Guide to the cutting-edge method of causal inference, ④Generalized synthetic control method gsynth, causal inference based on interactive fixed effects, ⑤Generalized synthetic control method gsynth, Stata run program release, ⑥synthesis control method and HCW method, who can go further? , ⑦DID, synthesis control, matching, RDD comparison of four methods, applicable scope and characteristics, ⑧ aid and Dutch disease, discovery of the Wenchuan earthquake, causal inference literature

For COVID-19, lockdowns, etc., you can refer to the following articles: ①About COVID-19, the articles in the latest Chinese and English periodicals are here, ②The legacy of SARS virus in the Chinese media, economy, society and other fields! ③The statement of 621 economists in the United States on supporting employment and enterprises to respond to the new coronavirus pandemic. ④For the first time, Academician of Metrology used the DID method to analyze the impact of China's lockdown on the spread of the new coronavirus! ⑤How do American Ivy League universities arrange teaching, research and doctoral dissertations under COVID-19? ⑥It's on fire! Top5 articles on Ctrip.com Liang Jianzhang's Top 5 hit the European and American academic circles during COVID-19! ⑦50 selected papers on incident research methods! Help analyze the impact of COVID-19! ⑧Compilation of COVID-19 related working papers , NBER sorts out for you! ⑨Latest: Top 5 journals are unlikely to publish quantitative papers related to COVID-19! Don't go after hot spots in vain! ⑩The two Nobel Prize winners actually quarreled on "Nature"! Is "New Crown" really an excellent natural experiment?

Text br /> on the bottom of the text content, author: Wei Zi Yi, Dongbei University of Finance and Economics College, communication mail: [email protected]
before the author of the article: Top, Machine Learning is an application of econometric methods, do not know the future facing Eliminate the danger!
Using machine learning and enhanced synthetic control methods to study the impact
of Wuhan's lockdown on air pollution and health during the epidemic Matthew A Cole, Robert JR Elliott, and Bowen Liu, The Impact of the Wuhan Covid-19 Lockdown on Air Pollution and Health: A Machine Learning and Augmented Synthetic Control Approach, 2020 April.
The author of this article uses a two-step method to quantify the impact of Wuhan’s closure on four air pollutant concentration indicators. First, the author uses machine learning to eliminate the confounding effects of weather conditions on pollution levels. Second, the author used a new enhanced synthetic control method, using unclosed cities as the control group to assess the impact of closed cities on climate pollution. The author found that the NO2 concentration dropped as much as 24μg/m3 during the lockdown period and 63% for the unclosed period; the PM10 concentration dropped to the same level in a shorter period of time; the lockdown had no obvious effect on the SO2 or CO concentration. Finally, the author estimated that the reduction of NO2 concentration in Wuhan may have avoided the deaths of 496 people in Wuhan, 3368 in Hubei Province, and 10,822 in the country.
One. The line of thinking is
that people are confined to their homes due to the epidemic, which is also known as "closing the city." The author found that the implementation of "closing the city" brought about an improvement in air quality. With the closure of many shops and businesses, the city’s industrial activities and vehicle use have dropped sharply, so many reports have shown that pollution levels during the lockdown period were significantly lower than those experienced under normal conditions.
At the same time, this discovery also brought the author more thinking. Which pollutants have actually fallen? How to measure the reduction of pollutants? (Considering that pollution levels have changed in time and space, comparing them with data from previous years or with data from other cities may be misleading.) How can pollution reduction improve people's health?
First, considering the correlation between pollution, the number of hospital admissions, and mortality, since China’s air pollution often exceeds WHO standards, the number of hospital admissions will definitely be higher without reducing pollution. Secondly, because the exposure of pollutants may increase the death rate of the new crown virus, if the closure of the city does not improve the air quality, the death rate of the new crown virus in cities such as Wuhan is more likely to rise. All in all, the reduction in air pollution caused by the lockdown may increase the capacity of hospitals to accommodate COVID-19 patients and directly reduce the number of COVID-19 deaths.
However, the concentration of pollutants is not only affected by the policy of closing the city, but also affected by climatic conditions and emission levels. And the influence of wind speed, wind direction and temperature on pollution concentration is often greater than the effect of any policy intervention. Traditional econometric methods that try to solve the impact of weather on pollution trends ignore that the trends of pollution trends are non-linear and affected by interactions. Therefore, in order to capture the impact of meteorological variables on pollution more effectively, this paper adopts predictive machine learning methods.
The author applied a combination of the most advanced machine learning and synthetic control methods to quantify the impact of Wuhan’s lockdown on air pollution levels. First, the author applies the latest weather standardization technology to eliminate the influence of local weather conditions on the pollution concentration. Second, using normalized climate concentration data, the Ridge ASCM was applied to investigate the causal impact of the Wuhan blockade on local air pollution levels. Third, select mortality estimates from the existing literature and calculate the number of potential deaths due to the improvement of air quality in Wuhan City, Hubei Province and the whole country.
two. Method introduction
2.1. Normalization
This article uses two steps to identify the causal impact of the Wuhan blockade on local air pollution levels. The first step is to perform random forest-based climate normalization on each of the four pollutants in 30 Chinese cities to obtain normal pollution concentrations observed every hour. The second step is to normalize the hourly weather.

There are two main reasons for weather normalization:

• Since changes in the concentration of pollutants are affected by both meteorological conditions and emission levels, it is difficult to evaluate the impact of policies on the concentration of atmospheric pollutants in environmental economics policy evaluation. This means that it is difficult to clearly determine whether the improvement in air quality is due to a real drop in emissions, or is simply the result of weather conditions that have reduced the measured concentration. Therefore, the best way to account for local meteorological conditions is to eliminate their effects from the observations of pollution concentrations and eliminate local weather effects, so that policy makers and social planners can make more informed decisions on the effects of previous air pollution interventions. , Which in turn will help guide future policy decisions.

• A key principle behind the synthetic control method is comparative case studies, in which the impact of policy interventions can be estimated by comparing variable changes between a single processing unit and multiple control units. Ideally, the control unit should be as similar to a single processing unit as possible, but not exposed to policy intervention. In our case, the daily air pollution concentration level in Wuhan is very unstable, which will lead to the potential problem of overfitting. Therefore, a machine learning algorithm is used to remove weather noise from the observed air pollution concentrations in all thirty cities. The result is a more reliable estimate of the impact of the Wuhan city closure on the local air pollution level, that is, the removal of pollution and natural changes caused by pure human activities under weather conditions.

2.2. Machine learning

• Decision tree and random forest

The decision tree model has strong interpretability for the training set. However, the decision tree may be prone to overfitting. Therefore, the prediction obtained only from the decision tree is optimal for a given training set, but may result in low prediction accuracy for the new data set.
In order to overcome the inherent shortcomings of decision trees, Breiman introduced the random forest algorithm. The principle is that considering the low prediction accuracy of the decision tree model on the new data set, if the algorithm can be used on a larger number of data sets, the performance of the algorithm will be improved. If there is only one set of data, then the randomness of the data can be increased by using the bootstrapping method and the bagging method. The random forest algorithm is essentially composed of a large number of individual decision trees, and is obtained by averaging the estimates from the entire forest. Compared with a single decision tree, the random forest method can greatly improve the prediction performance. Therefore, the meteorological normalization technique used in this article is based on the random forest algorithm. That is, the regression tree is obtained by recursively dividing a single predictor on a threshold until the purity of the node is reached.

• Enhanced Synthesis Control Method (ASCM)

The design of the enhanced synthesis control method is similar to the traditional DID, and its goal is to find a control unit similar to the processing unit. In this article, since we are interested in testing the impact of Wuhan’s lockdown on local air pollution levels, the ideal solution is to find a city in China that has not experienced the lockdown, but with different characteristics (for example, economic development level, It is very similar to Wuhan in terms of industrial structure and population. Wait. However, in fact, no city can match as closely as Wuhan. By adopting the SCM method, we use a data-driven program to simulate or construct artificial or "synthetic" Wuhan using a weighted average set of control cities. The goal of Synthetic Wuhan is to reproduce the air pollution level trajectory of real Wuhan before the closure of the city. In this way, after the closure of the city, the difference in the trajectory between synthetic and real Wuhan can be summarized as the causal effect of the closure.
ASCM extends SCM to situations where good pre-intervention matching cannot be achieved between processing and synthesis units. ASCM uses the resulting model to estimate the deviation due to poor pre-intervention matching, and then corrects the deviation in the original SCM estimate.
three. Empirical results
3.1 Machine learning results

The figure above shows the average and normalized SO2, NO2, CO, and PM10 concentrations observed in Wuhan from January 2013 to February 2020. (The gray is the observed value and the red is the normalized result) It can be seen that the observed concentration level generally decreases with time, especially for SO2. More importantly, the above figure illustrates the influence of our weather normalization process. There is a clear difference between the observed concentration and the normal weather concentration. The latter is a smoother data series.

To further narrow the time interval, the above figure shows the daily curves of SO2, NO2, CO and PM10 observed and normalized in Wuhan from December 21, 2019 to February 3, 2020. It can also be seen that the trend of climate normalized pollutants fluctuates less than the observed value, indicating the degree to which weather conditions affect the pollution level of the recording station. Therefore, it can be seen that it is confusing if only the observed pollution value is used to identify the impact of the closure of the city (the dotted line in the figure) on the pollution concentration.
3.2 The impact of Wuhan's lockdown on local air pollution
In order to present the results, we consider our four pollutants in turn. The left side of the figure below plots the difference in NO2 levels between "synthetic" Wuhan and Wuhan. The trend of “synthetic” Wuhan and Wuhan NO2 levels is plotted on the right. The vertical line also refers to the lock date in Wuhan. We found that these two trends remained between 45～52μg/m3 from December 21st to December 27th (especially at the WHO safety limit of 40μg/m3), and began to drop to 35μg/m3 in January 2020. ~40μg/m3. This coincided with the sharp decline in economic activity during the Chinese New Year. After January 21, 2020, there is a huge and significant difference between the NO2 emissions of Wuhan and Synthetic Wuhan, with a peak difference of about 24μg/m3, a decrease of 63%. In the last 12 days, the gap between the two was reduced, but it was still greater than 15μg/m3. It is worth noting that NO2 is already lower than the WHO standard of 40μg/m3. The synthetic weather NO2 level in Wuhan is always between 33 and 40μg/m, while the actual level in Wuhan basically dropped to about 20μg/m3 3 days after the closure of the city, and remained below 20μg/m3 before the end of the study. The results show that the blockade has led to a significant drop in NO2 levels in Wuhan’s climate normalization.

Next, let’s look at SO2. As shown in the figure, the difference between the synthesis of SO2 and the actual Wuhan is negligible. This shows that the other 29 cities are doing a good job of simulating the trajectory of Wuhan's pollution concentration. After the lockdown, the level of SO2 in Wuhan was about 3μg/m3 lower than when Wuhan was not locked down. However, it decreased three to four days after the city was closed, and returned to the same trend as the other 29 cities.

The figure below plots the results of CO levels. It can be seen from the figure that Synthetic Wuhan is not one that has high consistency with the real Wuhan before the city was closed. This means that we cannot rashly conclude that the Wuhan blockade has a significant impact on CO levels.

Finally, the author analyzes the differences in PM10 levels. Within four or five days of closing the city, the difference in PM10 levels continued to increase, with a maximum difference of 22μg/m3. After seven or eight days, the difference in trends gradually narrowed. As a result, the lockdown in Wuhan resulted in a significant but short-term reduction in PM10 levels.

3.3 Placebo test

• Timely placebo trial

For the timely placebo trial, it is assumed that the Wuhan shutdown occurred on the same date, but in 2018 or 2019. The figure below shows the results of a timely placebo trial of NO2 and PM10. On the left side of the figure, we focus on the data cycle between December 21, 2018 and February 3, 2019, and then set the simulated lockdown to January 21, 2019. The right side of the figure shows a period of time between December 21, 2017 and February 3, 2018. The lockdown device was set on January 21, 2018. For the placebo trials at these two times, we did not find a significant difference.

• In-situ placebo trial

We randomly give the lockdown order to one of the 29 other controlled cities. The figure below depicts the changes in NO2 levels before and after the closure of the city to show the difference between the synthetic trends of 29 different controlled cities. It can be seen that Wuhan is very obvious in the 29 lines, and no other city shows a similar decline in value.

In the results of PM10, we found that the four synthetic Wuhan lines have similar results (Shijiazhuang, Jinan, Hangzhou, and Hohhot, respectively). However, the results for these four cities are not significant. The picture on the right shows the experiment again after we gave up the four lines, and we can find that the red line was very obvious at the beginning of the lockdown. The results are consistent with our baseline results of normal PM10 levels in Wuhan weather, and we only noticed a significant drop within two to seven days after lock-in.

• Alternative control group

The final sensitivity check is to use a series of different control groups to run the Ridge ASCM model to check whether the results are sensitive to our initial selection of 29 large cities. In addition to the complete 29 city control group, we also used four alternative control groups to re-evaluate the results, which we call synthetic control groups 1, 2, 3, and 4 (Syn_CG1, Syn_CG2, Syn_3, and Syn_CG4 in the figure). The figure below shows the results of the synthesis of NO2 and PM10 from the four alternative control groups in Wuhan. All five control groups are consistent with the pre-closing trends of NO2 and PM10. We found that the effect of Wuhan’s lockdown on NO2 and PM10 levels was not sensitive to the selection of the control group.

four. The impact of pollution decline on health After the impact
of Wuhan on the concentration of pollution, the author began to consider the potential number of lives saved due to the improvement of air quality. The following table reports about 1228 to 3368 lives saved by reducing NO2 concentration by 20μg/m in the entire region of Hubei Province. The number of lives saved by reducing NO2 concentration by 10μg/m is between 614 and 1,684. For completeness, the author extends the analysis to all areas that are blocked within China. As of early February 2020, a total population of more than 233 million (including 59 million in Hubei) has been formally blocked. The results provided in the following table show that a 20μg/m3 reduction in NO2 can save lives between 3940 and 10822, while a reduction of 10g/μm3 saves lives between 1970 and 5411. Our results show that the improvement in air quality brought about by China's blockade has led to a significant reduction in mortality.

Robert Elliott, the former dean of the Department of Economics at the University of Birmingham, took the pleasure of studying Chinese economy and collaborated with Professor Sun Puyang of Nankai University on some articles. His representative works are as follows:

Elliott, R.J.R., Liu, Y., Strobl, E. and Tong, M. (forthcoming), Estimating the Direct and Indirect Impact of Typhoons on Plant Performance: Evidence from Chinese Manufacturers, Journal of Environmental Economics and Management.
Elliott, R.J.R, Jabbour, L. and Vanino, E. (forthcoming) Innovation and the Intensive and Extensive Margins of Trade: Evidence from French Firms. Oxford Bulletin of Economics and Statistics.
Beltran, A., Maddison, D. and Elliott, R.J.R. (forthcoming) The impact of flooding on property prices: A repeat-sales approach. Journal of Environmental Economics and Management.
Cole, M.A., Elliott, R.J.R., Okubo, T. and Strobl, E. (2019), Natural Disasters and the Birth, Life and Death of Plants: The Case of the Kobe Earthquake. Journal of Economic Geography, Vol. 19, 2, pp. 373-408. ABS 4.
Elliott, R.J.R. Sun, P. and Zhu, T. (forthcoming), Electricity Prices and Industry Switching: Evidence from Chinese manufacturing firms. Energy Economics.
Elliott, R.J.R., Horsewood, N.H. and Zhang, L. (2018), Importing exporters and exporting importers: A study of the decision of Chinese firms to engage in international trade, Review of International Economics.
Beltran, A., Maddison, D. and Elliott, R.J.R. (2018), Assessing the economic benefits of flood defences: A repeat-sales approach, Risk Analysis, Vol. 38, 11, pp. 2340-2367.
Cole, M.A., Elliott, R.J.R., Occhiali, G. and Strobl, E. (2018), Power Outages and Firm Performance in Sub-Saharan Africa, Journal of Development Economics, Vol. 134, pp. 150-159.
Beltran, A., Maddison, D. and Elliott, R.J.R. (2017), Is Flood Risk Capitalised into Property Values? Ecological Economics.
Cole, M.A., Elliott, R.J.R., Okubo, T. and Strobl, E. (2017), Natural Disasters and the Birth, Life and Death of Plants: The Case of the Kobe Earthquake. Journal of Economic Geography.
Elliott, R.J.R. and Lindley, J. (2017), Environmental Jobs and Growth in the United States? Ecological Economics, Vol. 132, pp. 232–244.
Elliott, R.J.R., Jabbor, L. and Zhang, L. (2016), Firm Productivity and Importing: Evidence from Chinese Manufacturing Firms. Canadian Journal of Economics, Vol. 49, 3, 1086-1124.
Elliott, R.J.R., Strobl, E. and Sun, P. (2015), The Local Impact of Typhoons on Economic Activity in China: A View from Outer Space, Journal of Urban Economics. Vol. 88, pp. 50–66.
Elliott, R.J.R., Sun, P. and Qiqin X. (2015), Energy Distribution and Economic Growth: An empirical test for China, Energy Economics, Vol. 48, pp. 24-31.
Elliott, R.J.R. and Zhou, Y. (2015), Globalization, Firm Co-location and Wage Inequality in China: A Spatial Analysis, World Development, Vol. 66, pp. 629-644.
Albornoz, F., Cole, M.A., Elliott, R.J.R. and Ercolani, M.G. (2014), The Environmental Management of Firms: Motives, Obstacles and the Role of Foreign Ownership, Journal of Environmental Management, Vol. 146, pp. 150-163.
Cole, M.A., Elliott, R.J.R. and Strobl, E. (2014), Climate Change, Hydro-dependency and the African Dam Boom, World Development, Vol. 60, pp. 84-98.
Elliott, R.J.R., Sun, P. and Chen, S. (2013), Energy Efficiency and the Impact of Foreign Direct Investment in China: A City-Level Study, Energy Economics, Vol. 40, C, pp 484-494.
Cole, M.A., Elliott, R.J.R. and Kemerat, K. (2013), Local Exposure to Toxic Releases: Examining the Role of Ethnic Fractionalization and Polarization, Ecological Economics, Vol. 93, pp. 249-259.
Cole, M.A., Elliott, R.J.R., Okubo, T. and Zhou, Y. (2013), The Carbon Dioxide Emissions of Firms: A Spatial Analysis, Journal of Environmental Economics and Management, Vol. 65, 2, pp. 290-309.

Long press the above QR code to read the original PDF

The following short-linked articles belong to a collection, you can collect them and read them, or you won't find them in the future.
In 2.5 years, nearly 1,000 non-weighted measurement articles in the econometric circle,

You can search for any measurement related issues directly in the official account menu bar,

Econometrics Circle