2.1 Lockdowns and socio-economic data

The COVID-19 pandemic caused around 7 million deaths worldwide (Our World in Data, 2020). Some countries enacted minimum restrictions, while others imposed strict measures. Hale et al. (Reference Hale, Webster, Petherick, Phillips and Kira2020) tracks each country's policies in categories such as containment and closure, economic response, and health systems. The main specification uses the date when workplace closure orders started in each country. Specifically, we construct a dummy “Lockdown=1” when Hale et al. (Reference Hale, Webster, Petherick, Phillips and Kira2020) indicates the government “requires closing (or working from home) for some sectors or categories of workers” or “requires closing (or working from home) for all-but-essential workplaces (e.g., grocery stores, doctors)”; and the dummy “Lockdown=0” when no measures or closing is recommended but not required. As a robustness check, we also use the date when stay-at-home orders were in place or restrictions on internal movement began. We also determine inhabitants’ level of compliance with the lockdown orders using data from Google's mobility reports (Google LLC, 2020).

Table 1, panel A presents summary statistics at the country level for the 70 countries included in the analysis. On average, the countries introduced restrictions in bi-week 6.7 of the year 2020, which is around March 31. Most countries had not ended the restrictions at the end of the study period (July 12), so there are only 21 observations for the bi-week when the lockdown ended. See figure A1(A) for the evolution of the percentage of countries under lockdown by bi-week.

Table 1. Summary statistics

Notes: Panel (A) presents data on the countries in the sample used for the regression analysis. Column (1) is the mean, and column (2) the standard deviation. Columns (3) and (4) are the minimum and maximum values of the variable for a country, respectively. There are fewer observations for the lockdown start/end bi-week because some countries did not impose workplace closure restrictions or had not ended restrictions by July 12, 2020. Panel (B) presents the deforestation data at the municipality bi-week level, the level used in the regression analysis. Alerts as forest share were calculated as the number of pixels with deforestation alerts per 10,000 forest pixels. Panel (C) is restricted to Colombian municipalities.

We study the heterogeneous effects of the lockdown orders by the percentage of a country's gross domestic product (GDP) vulnerable to lockdown restrictions. For lockdown-vulnerable GDP, we take as a proxy the share of wholesale, retail trade, restaurants, and hotels because these sectors depend on lockdown-restricted human interactions. We use data from United Nations Department of Economic and Social Affairs (2020).Footnote ² Table 1 shows that the share of vulnerable GDP varies from 0 to 70 per cent with a mean of 46 per cent. This country proxy is not perfect because it is influenced by cities not included in the analysis. In cities, services jobs are prone to remote working, while in forested municipalities, they represent wholesale and retail trade (including hotels and restaurants) that we want to capture. We explore the economic channel further using municipality-level data for Colombia from the National Planning Department (DNP, 2018). We obtain the share of each municipality's GDP attributed to restaurants, hotels, and sales. These sectors were largely affected by the lockdowns, so this information allows us to study the economic channel of the lockdowns’ effect on deforestation.

We use an index of government effectiveness provided by the World Bank (2020) to measure the likelihood of compliance with lockdown orders. This index captures perceptions of the quality of the civil service and the quality of policy formulation and implementation. It is standardized to have a global mean of 0 and a standard deviation of 1. The global index varies from approximately $-2.5$ to 2.5 (higher values indicate better governance). Colombia, the country we use for the within-country analysis, has a government effectiveness of $-0.09$. For Colombian municipalities, we use the effectiveness index constructed by DNP (2018). We standardized the index to make the Colombian estimate comparable to the global one.

Around a third of the countries in the sample are in the Americas, 39 per cent in Africa, and 27 per cent in Asia-Oceania. We obtained the map of the administrative divisions of each country from University of California Davis (2018). For most countries, we use the second administrative division, equivalent to counties and municipalities. But some small countries have only one level of division, the equivalent of states. See table B1 in the online appendix for details of each country. We measure the municipality's remoteness as the distance to the country's capital. We standardized the distance within each country to have a mean of 0 and a standard deviation of 1.

To study compliance with lockdown restrictions, we use data from Google's mobility report (Google LLC, 2020). This data uses anonymized phone locations and time patterns to identify aggregate mobility compared to a pre-lockdown day.Footnote ³ We use the percentage change in mobility to workplaces, measured by the change in total visitors.

2.2 Deforestation

Data for vegetation cover change alerts was obtained from Hansen et al. (Reference Hansen, Krylov, Tyukavina, Potapov, Turubanova, Zutta, Ifo, Margono, Stolle and Moore2016) for every bi-week from January 1 to July 12 for 2019 and 2020.Footnote ⁴ This data is constructed by analyzing satellite imagery and searching for disturbances in vegetation cover. The alerts are reported in squared pixels of $0.00025^{\circ }$, approximately 28 meters at the equator. We aggregate the data at the municipality bi-week level and calculate the number of deforestation alerts per 10,000 forest pixels. Although the share of forest pixels that are deforested is our main dependent variable, we also use the number of alerts as a robustness check, and the results remain the same.

There are three main issues with the alerts data.Footnote ⁵ First, it does not differentiate between the deforestation of natural forests and tree logging in plantations (Fergusson et al., Reference Fergusson, Saavedra and Vargas2020). Second, the date an alert is reported is not necessarily the date the deforestation happened because clouds might obstruct immediate detection. Finally, false positives are removed if not confirmed in a follow-up image in the next six months. To address the first issue, we restrict the alerts to areas classified as primary forests by Turubanova et al. (Reference Turubanova, Potapov, Tyukavina and Hansen2018), and call them deforestation alerts. To address the other two issues, we include region bi-week fixed effects to absorb under/over-reporting of alerts in a given area. An alternative way to deal with the third issue is to study whether the results vary based on the data's processing date. For example, instead of using the 2019 alert data as of 2020, where most alerts have been purged, we can use the 2019 alerts as of July 2019. This criteria will leave both years with a similar lag of processing noise.

Figure 1(A) presents the forest areas of the world in green and the gray “pathrow” squares for the areas included in the alerts. The Hansen et al. (Reference Hansen, Krylov, Tyukavina, Potapov, Turubanova, Zutta, Ifo, Margono, Stolle and Moore2016) data is for humid tropical forests, so the forests of Canada, Russia, and Europe are not included in this study. As the alerts’ data comes in squares of neighboring regions, we will use these squares in one of the estimation strategies to compare lockdown areas within neighboring regions.

Figure 1. Deforestation alerts data. (A) Forest and deforestation data, (B) Countries, (C) Number of alerts by bi-week. Notes: Panel (A) shows forests in green; and squares areas with deforestation alerts in gray. Panel (B) shows in dark green countries included in the regression because they have data on deforestation alerts and lockdown information. Panel (C) presents the total number of deforestation alerts in 2019 (blue circles) and 2020 (red diamonds). The x-axis is the bi-week of the respective year, and the y-axis the number of deforestation alerts. Map source: http://glad-forest-alert.appspot.com/.

Figure 1(B) presents the map of countries with deforestation alert data and lockdown information, which are consequently included in the analysis. The largest countries in the sample are Brazil, the Democratic Republic of the Congo, and Indonesia.Footnote ⁶

Figure 1(C) presents the total number of tropical deforestation alerts in all countries in the sample by bi-week. Globally there are around 800,000 deforestation alerts each bi-week. There is no change around bi-week 7 (end of March), when most countries introduced lockdown orders. Panel B of table 1 presents summary statistics at the municipality bi-week level, the dependent variable of the main regressions. On average, around 77 per cent of the municipality bi-week observations include a deforestation alert, equivalent to around one alert per 10,000 pixels (0.01 per cent) of the forest area. Note that only municipalities with primary forests are included.

2.3 Estimating equations

We estimate whether deforestation was reduced during COVID-19 lockdowns using difference-in-differences strategies. That is, we compare deforestation before and after the lockdowns were imposed in countries under lockdown, to countries that did not have a lockdown that bi-week. We have data at the municipality bi-week level for 8965 municipalities with forests in 70 tropical countries. The primary sample comprises 14 bi-weeks from January-July 2020, the first year of the pandemic. We use data for January-July 2019 for a triple difference-in-differences strategy. We estimate equation (1):

(1)\begin{equation} Y_{mcsnw} = \beta_1 Lockdown_{cw} + \gamma_m + \gamma_{nw} + \varepsilon_{m,sw}, \end{equation}

where $Y_{mcsnw}$ is a measure of deforestation in municipality $m$, country $c$, deforestation data square $s$, continent $n$, in bi-week $w$. In the main specifications, we use the number of pixels with deforestation alerts per 10,000 forest pixels. $Lockdown_{cw}$ indicates whether country $c$ was under lockdown in bi-week $w$. $\gamma _m$ are municipality fixed effects to capture time-invariant characteristics that affect deforestation, like distance to the country's capital, the presence of roads and rivers, the terrain's slope, and whether the forest is in a protected natural area (Deng et al., Reference Deng, Huang, Uchida, Rozelle and Gibson2011; Busch and Ferretti-Gallon, Reference Busch and Ferretti-Gallon2017). $\gamma _{nw}$ are continent-bi-week fixed effects to control for shocks like COVID-19 cases and the weather. We weigh the observations according to the area of primary forest in the municipality. Consequently, we estimate the effect on the average forest pixel. Finally, $\varepsilon _{m,sw}$ is an error term that we cluster two ways: at the municipality level and the data square bi-week level. As mentioned in the previous sub-section, the deforestation data is processed by squares of satellite images; therefore, observations for municipalities in the same data square are correlated due to errors in each satellite image. The identification for $\beta _1$ comes from differential deforestation rates for municipalities under lockdown compared to those in the same continent that were not under lockdown.

Equation (1) assumes a common shock for all countries in the same continent bi-week. However, there is considerable variation in COVID-19 cases and the weather within the same continent. Consequently, we exploit the satellite image square $s$ to which each municipality belongs. Figure 2 presents an example of two different deforestation data squares that cover Colombian municipalities. The square on the left covers the southern border of Colombia with Ecuador, Peru, and Brazil. The square on the right includes the eastern border of Colombia with Venezuela, Brazil, and Guyana. The square-bi-week fixed effects ($\gamma _{sw}$) in equation (2) absorb any common shock to the forest in each square during a given bi-week. If the square covers many countries, we can compare deforestation in municipalities of countries with different lockdown timings. For example, as Colombia started its lockdown on bi-week seven and Brazil on bi-week nine, our empirical strategy compares differential deforestation rates in those two bi-weeks in the border Amazon municipalities in the same square. The identification for $\beta _2$ in equation (2) comes from differential deforestation rates in municipalities under lockdown compared to those in nearby countries that were not locked down. Note that if a square covers only one country, the square bi-week fixed effect will absorb the lockdown effect in that country.

(2)\begin{equation} Y_{mcsw} = \beta_2 Lockdown_{cw} + \gamma_m + \gamma_{sw} + \varepsilon_{m,sw} \end{equation}

One possible concern with the two specifications described above is that the estimated coefficients ($\beta _1$, $\beta _2$) might confound the lockdown effect with differences in seasonal variation across countries. We use two approaches to address this concern. First, we conduct a placebo estimation of equation (2) using 2019 data and assuming lockdowns happened in the same bi-weeks as in 2020. Second, we simultaneously use 2020 and 2019 deforestation data in a triple difference-in-differences specification. This specification allows us to use country bi-week of the year fixed effects to capture specific country seasonality. Specifically, we estimate

(3)\begin{equation} Y_{mcswy} = \beta_3 Lockdown_{cwy} + \gamma_{cy} + \gamma_{m} + \gamma_{cw} + \gamma_{swy} + \varepsilon_{m,sw}. \end{equation}

$Y_{mcswy}$ is a measure of deforestation in municipality $m$, country $c$, square $s$, in bi-week $w$ of year $y$. $Lockdown_{cwy}$ is an indicator for whether country $c$ was under lockdown in bi-week $w$ of year $y$, that is, only for observations in 2020 ($y=2020$). $\gamma _{cy}$ are country-year fixed effects, $\gamma _{cw}$ are country bi-week of the year fixed effects, and $\gamma _{swy}$ square bi-week year fixed effects. The identification for $\beta _3$ comes from differential deforestation rates in municipalities under lockdown compared to municipalities in the same square that were not under lockdown, net of deforestation differences the year before.

Figure 2. Example of deforestation data squares covering multiple countries. (A) Southern Colombia, (B) Eastern Colombia. Notes: This figure presents two squares in which deforestation alerts are reported. Using data square bi-week fixed effects absorbs any common shock to the area, such as the weather or COVID-19 cases. Thus, we can estimate the effect of differential lockdown timings across neighboring countries on equation (2).