Lockdown, Infection, and Expressed Happiness in China

Our study covers municipalities, deputy-provincial-level cities, prefecture-level cities, and prefectures. There are different types of cities in mainland China, characterized by official ranks in the administrative hierarchy. There are 4 major cities as municipalities, specifically Beijing, Shanghai, Tianjin, and Chongqing. They have equal standing to a province, under the governance of central government directly. The next level is called deputy-provincial-level cities, which are major cities in their province. These cities enjoy administrative dependence though they are still governed by their province. Among 15 of them, 10 are provincial capitals. For example, Wuhan as a deputy-provincial-level city, is the provincial capital of Hubei province. The next level is prefecture-level cities. These cities are governed directly by their provincial government. Some of them are provincial capitals. There are also prefectures which have the same administrative level as prefecture-level cities, mostly located in less developed areas. Below prefecture-level city and prefecture, there are county-level cities or counties. Our study excludes them due to their relatively smaller sizes. For simplicity, we will just use “city” to refer to these three types of cities and prefectures in the rest of the paper.

Other cities in Wuhan province are excluded from our analysis to mitigate the potential spillover effect.

They excluded China from their analysis on impacts of lockdowns on expressed happiness.

https://​www.​reuters.​com/​article/​us-china-health-who-idUSKBN1ZM1G9. Assessed on August 24, 2021.

March 18 was the first date with no new infections, but later there were some new infections (one for about 2 days).

Since we controlled for date fixed effects in our models, the common information (both on the infection risks and the lockdown) shock should have already captured. Unless we believe the information effects were quite different on the treatment and control groups over time, the lockdown effect we estimated should be mostly the actual lockdown effect not the information effect.

The falsification test follows the specification as below:

The event study specification is as follows, with interaction terms between the Wuhan indicator and date indicators:

Data on infections during the initial outbreak are under-reported but can indeed represent perceived infection risks.

Our original data only consist of posts by active Weibo users. Active users were defined and collected by Hu et al. (2020) by the following four rules: (1) follows > 50; (2) fans > 50; (3) tweets > 50; and (4) recent post < 30 days. Based on this definition, active users account for 8% of all users. The reasons for including only the active users are mainly to bypass the limitation of Weibo search interface and cost of traversing all Weibo users during the COVID-19 outbreaks and to facilitate the efficiency and timeliness of data collection. Given the rapid spread of COVID-19 and abrupt lockdown measures during the sample period of this study, we believe this dataset provides a reliable collection of Weibo posts and happiness measurement, which is more real-time and less subject to censorship. We observe that about 70% of active Weibo users are self-identified as female. However, there is no user id to link the dataset of Weibo posts to the user dataset, and therefore, we are not able to distinguish individual users from institutional users. We also dropped re-tweeted posts, which mostly consisted of news and posts representing other users’ sentiment. We further dropped abnormal Weibo posts with fewer than 3 Chinese characters or more than 170 characters, to improve the efficiency of the sentiment analysis.

DXY started to publish COVID 19 data only after the Wuhan lockdown on January 23, 2020.

The information was retrieved from http://​news.​ifeng.​com/​c/​7u8Wvj66XuS on March 1, 2021.

Alternatively, we used the daily mean of AQI calculated by the Chinese standard using six atmospheric pollutants—AQI, CO, NO₂, O₃, PM10, PM2.5, and SO₂—and the results appear to be consistent.

All meteorological variables are inverse-distance-weighted mean.

Summary statistics of the two variables are reported in Panel A of Appendix Table 6.

The complete list of keywords is available upon request.

Using pretreatment outcome variables as the predictors are crucial to control for unobserved covariates (Abadie, 2021; Abadie & Vives-i-Bastida, 2022). We include pre-treatment outcome variables on January 19, 18, and 17 in the model, however, our results are robust to including more pre-treatment outcome variables.

We also conducted the analysis with propensity score weighting. The results are similar to our baseline ones, and they will be available upon requests.

Results are essentially the same if we use all cities, as in column 5.

Summary statistics of the set of lockdown and unlocked variables are reported in Panels B (for column 1 in Table 5) and C (for column 1 in Table 5) of Appendix Table 6 respectively.

From March 4th, our control cities mostly had no government-implemented restrictions. However, we acknowledge the possibility of self-protection. If the self-protection of the control group cities has a net positive influence on their expressed happiness, our estimated lockdown effect was a lower bound.