Subnational Income, Growth, and the COVID-19 Pandemic

ABSTRACT

Using real-time data and machine-learning methods, we produce monthly aggregates on gross national income (GNI) for 147 Pakistani districts between 2012 and 2021. We use them to understand whether and how the COVID-19 pandemic affected the growth and subnational distribution of income in Pakistan. Three findings emerge from our analysis. First, districts experienced a sizable decline in income during the pandemic, as their monthly growth rate dropped on average by 0.133 percentage points. Second, a larger income drop took place in districts with a higher COVID-19 incidence, which correspond to urban areas characterized by a higher population density. Third, COVID-19 caused a decline in income inequality across districts, with richer districts experiencing more negative income growth during the pandemic.

1. Introduction

The COVID-19 pandemic has produced dramatic changes around the globe, as the effects of the virus and government containment policies have disrupted our societies and economies starting from February 2020. While there is ample knowledge on how high-income countries and their subnational units reacted (Chetty et al. 2020; Woloszko 2020; Chen et al. 2020; Delle Monache, Emiliozzi, and Nobili 2021; Giannone, Paixão, and Pang 2022), the same level of analysis and evidence is lacking in low-income countries because of the scarce availability of recent data.

This paper examines the effect of the COVID-19 pandemic on subnational income in a low-income country, Pakistan, and its districts. However, one of the main challenges was the lack of real-time data on gross national income (GNI). To compensate for this, we gathered real-time data from a variety of sources that could, a priori, have a potential relationship with economic activity. We therefore developed a machine-learning algorithm to now-cast the GNI covering 147 districts in Pakistan from 2012 to 2021 by combining traditional administrative data with night-lights and other satellite data. Our approach builds on the frontier in this literature by relying on multiple satellite products (Asher et al. 2021; Ch, Martin, and Vargas 2021; Beyer, Hu, and Yao 2022), integrating our prediction exercise with a robust empirical inference (Athey 2017), and connecting micro and macro data (Vavra 2021).

Three novel findings contribute to the current debate on the effects of the COVID-19 pandemic and containment policies on the economy. First, we observe that Pakistani income slowed down during the pandemic, when the growth rate dropped on average by 0.133 percentage points. Second, there exists a robust and negative correlation between the incidence of COVID-19 (cases, deaths, and recoveries), which hit urban and densely populated areas more aggressively, and income growth. Third, we observe that the previous two effects led to lower income inequality across districts due to higher-income districts experiencing higher negative income growth during the pandemic. This creates a sort of “convergence to the bottom”—albeit only temporary. Our findings bring a perspective to the latest literature on convergence (Patel, Sandefur, and Subramanian 2021; Kremer, Willis, and You 2022; Pande and Enevoldsen 2021; Acemoglu and Molina 2021), and in particular to the economic effects of COVID-19 on growth and inequality in low- and middle-income countries. Our monthly data set suggests that conditional convergence has been taking place in Pakistan since before the pandemic. In fact, districts with a lower level of income in 2012 have been growing faster than districts with higher incomes before the pandemic. During the COVID-19 period, the gap between the growth trajectories of districts seems to have further diminished. However, a key distinction exists between pre-pandemic and pandemic-induced periods. While the hypothesis of convergence before the pandemic is supported by higher growth across low-income districts and may represent a permanent move toward a new steady state, during the pandemic, growth dynamics are mainly governed by high-income and urban districts slowing down the most. This suggests it may only be a temporary shock rather than an effective convergence to the bottom. These results parallel the findings of Gupta, Malani, and Woda (2021) in India, who use representative panel data on household finance and consumption instead of satellite now-casting. Our work is also connected to the literature on growth in regions and regional convergence (Gennaioli et al. 2013, 2014; Ganong and Shoag 2017; Lessmann and Seidel 2017; Giannone, Paixão, and Pang 2022; Giannone unpublished manuscript; Hsieh and Moretti unpublished manuscript) showing two key determinants of the recent pandemic-induced recession: urbanization and COVID-19 incidence. We highlight the differential effect of the pandemic across urban and rural districts, and we find that this heterogeneous incidence is mainly due to the high density of population in cities. Our results are also in line with Moeen et al. (2021), who show that the service sector was the most hit by COVID-19, followed by industry, while agriculture was only lightly affected. Moreover, they find that richer households lost more than poorer ones and that urban districts were more affected than rural ones. These findings corroborate the ones in this paper and, in particular, highlight that shocks to manufacturing and highly productive districts can create long-term effects on investment and productivity, in line with the work of Choudhary and Limodio (2022). Second, our results are in line with the work of Saez and Zucman (2016) showing that inequality declines during recessions, though this specific case may be due to a decline in contact-intensive activities in urban centers (Koren and Pető 2020) rather than financial returns. In this respect, our results are aligned with the findings of Deaton (2021) on the lower pandemic-induced within-country inequality and the recent World Bank report suggesting that years of poverty eradication vanished in a few months.¹ Finally, this paper contributes to an emerging literature in macro-development assessing the effects and costs of COVID-19 on low- and middle-income countries (Alfaro, Becerra, and Eslava 2020; Alon et al. 2020; Gottlieb et al. 2021b,a).

The remainder of the paper is organized as follows: The next section introduces some key papers in this literature; in the Section “Data and Methodology”, we illustrate the data gathering procedures and methodology; in the Section “Results”, we present the main results; and a technical guide on the employed algorithms can be found in the supplementary online appendix. Finally, the Section “Conclusions” offers some concluding remarks.

2. Related Literature

As stressed by the Bank for International Settlements (Tissot et al. 2020), the current crisis has called into question the traditional statistical aggregates. The constant mutations of the virus result in a rapidly escalating framework where the economic impact varies heterogeneously among sectors and geographic areas. Standard statistical aggregates are often available at the national level, with several months of delays. For this reason, the literature exploring novel sources of data is rapidly expanding.

Chetty et al. (2020) is an important contribution in this field. Exploiting real-time and granular data on American companies, it tracks the crisis’ impact on consumption and the labor market. Through a different approach, Woloszko (2020) proxies them from Google Trend and now-casts the national GDP for 46 OECD and G20 countries. A wider approach is proposed by Chen et al. (2020), integrating search queries with electricity and unemployment data. Following a similar approach, Delle Monache, Emiliozzi, and Nobili (2021) build a weekly economics index for Italy through granular administrative data. Using a social accounting matrix multiplier, Moeen et al. (2021) assess the impact of COVID-19 on macroeconomic variables in Pakistan. Their study reveals a 26.4 percent decline in GDP from mid-March to the end of June 2020, with services experiencing the most significant losses (17.6 percent).

Similar studies are not reproducible in emerging markets with a structural deficiency of administrative data and low Internet penetration. To overcome this obstacle, a growing number of researchers are referring to satellite data (see Donaldson and Storeygard (2016) and Nagaraj and Stern (2020)). This novel source of information is available at a very granular level for the entire globe and almost in real time. Following this literature, Beyer, Franco-Bedoya, and Galdo (2021) combine VIIRS night-lights and electricity consumption to monitor the pandemic impact in India. This study shows that the drop in habitual activities persists after the restrictions’ lifting. It also suggests that the pandemic particularly affects the manufacturing and in-migration areas, while the out-migration areas seem to experience a reduced decline. Also, the work of Roberts (2021) obtains similar results, using night-lights to study COVID-19’s impact on Morocco.

In this literature, the work of Henderson, Storeygard, and Weil (2012) has popularized the use of night-lights as a popular proxy for economic development in emerging markets. Nevertheless, the recent findings of Asher et al. (2021) cast some shadow on their effectiveness in time-series analysis, since their elasticity with the local output varies according to the level of aggregation and the context. In other words, night-lights can be correlated to several development indicators, and discerning what they are proxying in different regions is difficult. Some papers overcome this issue by adopting different and more detailed proxies for local economic output. Engstrom, Hersh, and Newhouse (2017) proves that the extraction of daytime features from satellite data explains 60 percent of average log consumption in emerging markets; Jain (2020) shows that all the satellite data hide implicit biases (for example, clouds, saturation, non-random misclassification, meteorological variables) in the realization process, whereas Burke et al. (2021) specify that the errors attributed to these models tend to be overestimated and related to the low-quality administrative data used as reference.

Our work includes insights from this literature: we use satellite lights in line with Henderson, Storeygard, and Weil (2012), but consider different moments of these series and include other real-time data sets as discussed by Asher et al. (2021). In addition to these data sets, we also partnered with local electricity providers to build a district-level electricity data set, in line with Beyer, Franco-Bedoya, and Galdo (2021), and add a finer split between electricity for domestic, commercial, and industrial use.

3. Data and Methodology

This section provides an overview of the data and illustrate the methods used in the analysis.

3.1. Data

For our analysis, many different databases have been used. In particular, we collected data from various sources to account for as many variables as possible that might be relevant predictors of economic activity.

Pakistan is among the low- and middle-income countries offering the most detailed and extensive administrative data. Most of these resources are available in traditional wide economic macro-aggregates, while micro-aggregates are often available only upon request. The principal statistical publications are released by the State Bank of Pakistan, the Ministry of Finance, and the Pakistani Bureau of Statistics. From the latter, we used three sources: (a) the annual Pakistan Economic Survey (PES), which contains an extensive set of variables such as wages, doctor fees, and import/export of cargo; (b) the Monthly Bulletins of Statistics, which report price indexes for over 400 items at the city-month level; and (c) the Survey on COVID-19, which provides information on migratory movements during the pandemic. Additionally, the National Electric Power Regulatory Authority produces granular data on electricity consumption at the tehsil-month level, disaggregated by various destination uses (commercial, domestic, industrial, and others).

In accordance with the growing literature on satellite imagery, we included VIIRS night-lights as a potential predictor of economic activities. In particular, we opted for the high-frequency and less pre-processed VNP46A1-VIIRS/NPP Daily Gridded Day Night Band 500m Linear product. Furthermore, to capture economic activities beyond those directly associated with artificial lights, such as agriculture, we incorporate a far wider range of potential economic activity detectors. Among these, there are weather-related variables such as sun hours, temperature, and humidity, which are significant inputs in agricultural production. Simultaneously, we included a wide range of satellite data from NASA that encompass measurements related to vegetation health, fires, and cloud characteristics, including presence, thickness, and water content, as well as the detection of pollutants such as ozone, CO₂, and NO₂. All these variables could predict the quality of the harvest, influence agricultural yields, or be a good proxy for industrial activity. Considering the results of Asher et al. (2021), we also measure the primary destination use of land using the yearly landcover map produced by the European Space Agency (ESA) under the Climate Change Initiative (CCI).

The provincial-level aggregates for gross national income in Pakistan are taken from the United Nations’ Sustainable Development Goals data set. This data set combines country-level information with periodic household surveys. The compilation of these data is carried out by the Global Data Lab, hosted by the Nijmegen Center for Economics (NiCE) at Radboud University in the Netherlands. Figures are available in 2011$ PPP up until 2018. We decompose the provincial aggregate at the tehsil level using the share of population obtained from the United Nations’ WorldPop platform. To further validate this decomposition, we use an indicator for living standards provided by the United Nations Development Program (UNDP) in 2017. We also collect the sectoral composition of GDP for the main four provinces of Pakistan (Punjab, Sindh, Balochistan, and Khyber Pakhtunkhwa), provided by the Institute for Policy Reforms.

Finally, statistics on COVID cases, deaths, and recoveries were produced at the district level by provincial authorities and released upon our request. Due to institutional factors, it was not possible to obtain COVID data at the tehsil level; hence, we decided to conduct the analysis at the district-month level. Lastly, the number of available health facilities is retained from the Humanitarian Data Exchange of the United Nations Office for the Coordination of Humanitarian Affairs (OCHA).

Supplementary online appendix S1 offers a more schematic and complete list of all the specific data sets used and their respective sources.

3.2. GNI Prediction

In many low- and middle-income countries, such as Pakistan, data are usually scarce and slow to become available. The main challenge we had to face for this analysis was the lack of real-time disaggregated data on GNI. To remedy these shortcomings, we assigned to each tehsil a fraction of the province’s GNI proportional to its population and we gathered real-time data from a variety of sources that could a priori have a potential relationship with economic activity. We therefore obtained a data set at the tehsil-year level that we can use to produce monthly estimates for GNI after 2018. These are then aggregated at the district-year level to produce a statistical analysis of income growth in relation to the local incidence of the COVID-19 pandemic.

We do not assume a specific ex ante relationship between income and any of the potential predictors; instead, we employ a set of machine-learning (ML) algorithms that operate through supervised learning. Supervised learning implies that algorithms need to be trained on a set of already-existing data before being able to make predictions. GNI data in our setting are available yearly up until 2018; therefore, for the training we use a data set with observations at the tehsil year level from 2012 to 2018. The potential real-time GNI predictors in our data set exhibit significant heterogeneity in their ranges. To mitigate the potential dominance of certain variables solely based on their scale, we employ a standard technique called min-max normalization. This normalization allows us to rescale all the variables between 0 and 1, ensuring a fair comparison and preventing undue influences based on the original ranges. We opt for this normalization over standardization, as we do not have any prior knowledge of the underlying distributions of the variables.

We focused on a total of five classes of algorithms: elastic net, random forest, bagging, boosting, and support vector machines. In order to assess which one performs better, it is necessary to split the data set into a train sample (comprising 75 percent of observations) and a test sample (representing the remaining 25 percent of observations). This division allows the algorithm to learn from the training sample, identifying which variables should be considered and their respective predictive powers. Subsequently, we can use the trained algorithm to make predictions using the test sample, enabling a comparison between the predicted values and the actual data. The best result is achieved by the bagging algorithm, with an overall mean square error of 0.01506. Supplementary online appendix S2 provides supplementary details on the methodology for more in-depth reference, including the mean square error values for all the algorithms.

Once we had identified the best-performing algorithm, we employed it to generate predictions for the years following 2018. Furthermore, as we wanted to carry out the analysis at the month-year level, to produce the estimates, we employed a data set in which predictors were disaggregated at the tehsil-month-year level. With this approach, we were able to obtain GNI predictions for each tehsil-month, accomplishing two objectives simultaneously: on the one hand, we were able to obtain real-time figures for GNI in Pakistan; on the other hand, we could develop a data set that is both geographically and temporally disaggregated.

Once estimates at the tehsil level were produced, we collapsed the data set at the district level to carry out the analysis. To verify that our data are correctly attributed to each district, despite initially obtaining the geographical disaggregation using the proportion of population, we use the Human Development Index provided by the United Nations Development Program (UNDP). This indicator is available for 2017 for each Pakistani district, and one of its components is a measure of living standards, calculated through a survey on the living conditions of households. Therefore, we decomposed provincial GNI in 2017, taking into account the different living conditions in each district. The correlation between this newly disaggregated measure of GNI and our estimated GNI is very high, 0.98, and statistically indistinguishable from unit, as graphically represented by fig. S2.2 in Supplementary online appendix S2.

3.3. Empirical Analysis

The primary aim of our analysis is to investigate the role of COVID-19 on GNI growth in Pakistan. To explore this, we employ the following empirical model:

where growth_dmy represents the GNI growth rate of district d in month m and year y and is regressed on four different COVID-19 indicators. The four measures of COVID incidence are (a) Covid19_my, a dummy variable which takes value 1 from May 2020 onward for all districts in Pakistan; (b) Cases_dmy, the natural logarithm of the number of COVID-19 cases in district d during month m of year y; (c) Deaths_dmy, the natural logarithm of the number of COVID-19 deaths in district d during month m of year y; (d) Recoveries_dmy, the natural logarithm of the number of COVID-19 recoveries in district d during month m of year y.

In addition, we augment equation (1) by including an interaction with a dummy variable for urban districts, Urban_d, which takes unit value for districts containing at least one of the 20 largest Pakistani cities. The augmented model is therefore as follows:

The purpose of this specification is to extrapolate and analyze the differential impact of COVID-19 between urban and rural districts. Even in this case, we use the four different measures of COVID incidence and standard errors are clustered at the district level.

Finally, to further explore the dynamics of income growth in Pakistani districts, we examine the differential effects of COVID-19 based on the starting economic condition. We explore the following empirical model:

The income growth of district d in month m and year y is regressed on the dummy for COVID that takes value 1 from May 2020 onwards, Income_d2012, which represents the level of income of district d in 2012, and the interaction between these two variables. As before, standard errors are clustered at the district level. Across all of these specifications, the standard errors are clustered at the district level.

4. Results

Figure 1 presents the variables selected as those with the biggest predictive power for district GNI by the bagging algorithm. Night-lights emerge as the most significant predictor in our analysis, further reinforcing the strong evidence that they serve as a good proxy for economic activity. The second most influential indicator is the presence of urban areas, indicating that cities in Pakistan generally exhibit higher levels of wealth and concentrate a significant portion of the country’s production. Another noteworthy predictor is the female share of the population, which ranks third in terms of relevance. This suggests that the gender composition of the population has an impact on the local income dynamics.

The results of our estimation are reported in fig. 2. The top panel displays GNI in billions of US$ aggregated at country level from 2018 until 2021. It shows that there is a steady growth path from early 2018 to August 2020. The COVID-19 outbreak halts this trend in September 2020, leading to a gradual decline. Figure S3.3 in the supplementary online appendix shows a similar picture, with the time range extended from 2012 to 2021. To understand the drivers of this decline, we define a district as being “urban” if it contains one of the top 20 cities by size, as defined by the Pakistani Bureau of Statistics in its 2017 census.² As a result, 20 districts are classified as “urban” and the remaining 127 as “rural.” The bottom panel of fig. 2 depicts the GNI trend divided by urban and rural areas: the blue line with dots represents the average GNI in urban districts, while the dashed red line with squares represents the average GNI in rural districts. It highlights that the average GNI of urban districts (on the left y-axis) is four times bigger than the average GNI of rural districts (on the right y-axis). It is important to note that while urban districts exhibit a steep decline in income as the pandemic begins, this decline is much milder for rural districts. This smaller loss might be due to the suspension of high-interaction activities mostly concentrated in cities, or to the strategy of smart lockdown, promptly imposed by the Pakistani government only on certain hot spots across the country. Figure S3.4 in the supplementary online appendix offers a version of fig. 2 in which urban districts are classified differently: in the top panel, only districts containing one of the top 10 cities are classified as urban, while in the bottom panel, urban districts are those with one of the top 50 cities. The results are qualitatively similar, with urban districts exhibiting a steeper decline than rural districts.

In order to explore the spatial distribution of income, fig. 1 reports three pictures. The top-left panel presents a map with the average income per district between 2018 and 2019. In this map, high-income districts are indicated with dark green colors, and it is notable how the concentration of economic activities takes place mainly along the Indus River and the metropolitan areas (Islamabad, Karachi, Lahore, Peshawar, and Quetta). The arid and sparsely populated lands of Balochistan appear to be the poorest, followed by the mountain regions of Gilgit-Baltistan and Khyber Pakhtunkhwa. The top-right panel shows the average income growth between 2020 and 2021. In this case, darker colors indicate a stronger decline (or a smaller increase) in growth. The darkest areas are once again in the densely populated Punjab districts and in major urban areas. By comparing these two maps, it is already clear that districts with high incomes before the pandemic were the most severely hit after the outbreak. The bottom panel of fig. 3 shows exactly this negative correlation between GNI growth during 2020–2021 (on the y-axis) and the log level of GNI in the previous years (on the x-axis). The correlation is −0.67 and statistically different from zero below the 1 percent significance threshold. Figure S3.5 in the supplementary online appendix reports the same descriptive evidence in terms of income per capita: results are similar, including the negative and significant correlation between the pre-pandemic level of income and the growth of per capita income during the pandemic. We prefer to present the analysis with income levels rather than in per capita terms, given that the numbers for the population may not be adjusted based on the incremental COVID-19 mortality. Figure S3.6 in the online appendix shows the overall income growth during the pre-pandemic period from 2012 to 2019. Table 1 presents the baseline summary statistics for most variables included in our empirical analysis.

After these descriptive figures, we want to explore the relationship between income growth and COVID-19 incidence. Panel A of table 2 presents the empirical results of equation (1). Column (1) shows that COVID-19 had a negative and statistically significant effect on income growth between 2018 and 2021. After the COVID outbreak, districts’ growth was on average 0.133 percentage points lower. Similarly, the remaining three columns of panel A show that districts exhibiting a higher incidence of COVID-19 cases, deaths, or recoveries experience lower income growth throughout the period. Columns (2), (3), and (4) show that a 100 percent increase in COVID-19 cases, deaths, and recoveries implies a 0.0216, 0.0485, and 0.0297 percentage-point decline in income growth, respectively.

Panel B of table 2 further investigates whether and how the results of panel A differ across urban and rural districts using the specification of equation (2). Column (1) shows a key result of our analysis. The COVID-19 dummy has a negative effect on growth on average, but this effect is much bigger in urban districts than in rural ones. Before COVID, the difference in income growth between rural and urban districts appeared not to be statistically different from 0. During the COVID period, instead, the growth rate of rural districts declined on average by 0.104 percentage points, while the growth rate of urban districts declined by an additional 0.216 percentage points. Columns (2), (3), and (4) show that when controlling for COVID-19 cases, deaths, and recoveries, there is no difference between urban and rural areas. In other words, for a given number of cases, deaths, or recoveries, the effects do not differ based on the degree of urbanization.

The most straightforward explanation for this is given by the first three columns of table 3, which show the regressions of the logarithm of COVID cases, deaths, and recoveries on the Urban dummy. Looking at the results, it becomes apparent that COVID incidence is much higher in urban districts. This implies that the underlying explanation for the bigger decline in GNI growth that urban districts experience is that they have been hit the hardest by the pandemic. These findings are also in line with the policy of smart lockdown adopted by the Pakistani government, which imposed a partial lockdown only on selected hot spots nationwide. To have a better understanding of the factors that led to cities being impacted the most, we look at the potential drivers of this heterogeneous incidence. In the last three columns of table 3, the COVID variables are regressed on the standardized population density in each district. The findings demonstrate how population concentration has a significant role in the spread of COVID.

In supplementary online appendix S3 it is possible to find a battery of additional tables that verifies the robustness of these results. Tables S3.2, S3.3, S3.4, S3.5, and S3.6 replicate table 2 with some modifications. Table S3.2 includes district fixed effects: while the results employ a different source of variation, namely within-districts only, the magnitudes, signs, and significance of the coefficients are very similar. Table S3.3 investigates the determinants of the economic recession; its main message is that districts with a high income share in the service sector appear to be the most hit by the incidence of the COVID-19 pandemic. Table S3.4 includes in the regressions a control for the number of health facilities in each district. This serves to check that results are not driven by the difference in administrative capacity between urban and rural districts. This may be an issue if the allocation of doctors, or health inputs, or bureaucratic skills may be strategically allocated more to poorer-performing districts, as highlighted by Limodio (2021). However, despite this potential threat to identification, we do not observe a change in the key results. Table S3.5 uses two alternative definitions for the urban dummy: in panel A, districts classified as urban are those including one of the top 10 biggest cities, while in panel B, urban districts are those with one of the biggest 50 cities in the country. Even with this different definition of urban, the interpretation of the results does not considerably differ. Finally, table S3.6 uses as a dependent variable the growth rate of income per capita, leading once again to similar results.

Table 4 presents five versions of equation (3): Column (1) provides an indication of conditional convergence in the last decade in Pakistan, by showing that districts with a 1-standard-deviation-lower income in 2012 are growing by 0.03 percentage points more between 2012 and 2021. Column (2) introduces in the regression the COVID-19 dummy and its interaction with the standardized level of income in 2012. With this specification, Income_2012d has a lower but nonetheless negative and significant coefficient. The COVID-19 dummy is negative and statistically different from zero. What is most relevant is, however, the coefficient of the interaction term, as it expresses the effects of the pandemic on the growth trajectories of districts. Districts with a 1-standard-deviation-higher income in 2012 were already growing 0.016 percentage points less. During the pandemic period, they slowed down by an additional 0.14 percentage points. Columns (3), (4), and (5) include district and/or time fixed effects, showing how the result remains quite unchanged regardless of which fixed effects are included in the specification.

To further validate these findings, supplementary online appendix S3 offers two additional tables. First, table S3.7 uses the natural logarithm of the mean income per district in 2012 rather than the standard deviation. The findings remain the same; in particular, the interaction term’s coefficient is still negative and significant. Second, table S3.8 adds dummies that, in 2012, assigned each district to a particular income tercile. According to the findings, richer districts (third tercile) are those experiencing a steeper decline in growth, especially during COVID. This result underlines how the already existing differences in growth trajectories across Pakistani districts have been further reinforced by the pandemic, which P hit hardly richer districts. It is crucial to emphasize that this evidence does not necessarily signify a decrease in overall inequalities during the pandemic period. While horizontal inequalities, which refer to disparities across districts, may have decreased, vertical inequalities may have persisted. In essence, the correlation between income growth and inequality is not straightforward, as the adverse effects of the pandemic may have disproportionately affected the most vulnerable individuals, particularly those residing in wealthier districts.

5. Conclusions

In this paper, we apply a method at the frontier of the machine-learning literature to calculate monthly aggregates on gross national income (GNI) for 147 Pakistani districts between 2012 and 2021 using machine learning and real-time satellite data. Our work shows that Pakistani districts experienced a decline in income growth during the COVID-19 pandemic, as the average monthly growth rate dropped by 0.133 percentage points. We verify that the incidence of COVID-19, measured through cases, deaths, and recoveries, was higher in cities and appears to have a negative and sizable effect on income. Finally, we show that COVID-19 induced a sizable within-country difference in growth patterns, as districts with high pre-pandemic income experienced negative and strong growth during the pandemic. While, on the one hand, this may reduce district inequality and the prominence of urban centers, on the other hand, this process may lower the long-term prospects of the most dynamic Pakistani districts and harm long-term growth.

Data Availability Statement

Data can be accessed at https://www.dropbox.com/scl/fo/t2f4rvxwvm9hbbm3pquth/h?rlkey=tk6bs3ml505jf1dnha25v7or0&dl=0.

Author Biography

M. Ali Choudhary is a professor of Economics and Public Policy at Loughborough Business School and the State Bank of Pakistan, I.I. Chundrigar Road, Karachi, Pakistan, and the Centre for Economic Performance, 32 Lincoln’s Inn Fields, WC2A 3PH, London, UK, and the Loughborough Business School, Sir Richard Morris Building, Loughborough University, Epinal Way, Loughborough, Leicestershire, LE11 3TU; his email addresses are [email protected] and [email protected]. Ilaria Dal Barco is Predoctoral Associate at Bocconi University, Via Roentgen 1, 20136 Milan, Italy; her email address is [email protected]. Ijlal A. Haqqani is a economist at the State Bank of Pakistan, I.I. Chundrigar Road, Karachi, Pakistan; her email address is [email protected]. Federico Lenzi is a PhD student from the Northwestern University, Kellogg School of Management, 2211 Campus Drive, Evanston, IL 60208; his email address is [email protected]. Nicola Limodio (corresponding author) is a associate Professor of Finance at Bocconi University, Department of Finance, BAFFI CAREFIN and IGIER, Via Roentgen 1, 20136 Milan, Italy; his email address is [email protected]. The research for this article was financed by the International Growth Center and the State Bank of Pakistan. The authors would like to thank the members of the Monetary Policy Committee of the State Bank of Pakistan, namely, Governor Reza Baqir, Murtaza Syed, Jameel Ahmed, Asad Zaman, Naved Hamid, Azam Faruqee, Hanid Mukhtar, and Tariq Hassan, for the encouragement to find innovative data sources to measure economic growth during the COVID-19 pandemic. The authors also express their gratitude to Minister Asad Umar, Ijaz Nabi, Brigadier Saeed, Lieutenant Colonel Adnan, Major Sami and his team, Imtiaz Ahmed, Kashif Sahazad, and Nadeem Hanif. The authors would also like to acknowledge the useful suggestions of Giorgia Barboni, Johannes Boehm, Nicola Gennaioli, Thiemo Fetzer, Federico Rossi, Chris Roth, Nicolas Serrano Velarde, Tom Schmitz, and participants at various conferences, seminars, and workshops. A supplementary online appendix is available with this article at The World Bank Economic Review website.

Footnotes

References