Testing the differentiated impact of the COVID-19 pandemic on air travel demand considering social inclusion.

The economic downturn and the air travel crisis triggered by the recent coronavirus pandemic pose a substantial threat to the new consumer class of many emerging economies. In Brazil, considerable improvements in social inclusion have fostered the emergence of hundreds of thousands of first-time fliers over the past decades. We apply a two-step regression methodology in which the first step consists of identifying air transport markets characterized by greater social inclusion, using indicators of the local economies' income distribution, credit availability, and access to the Internet. In the second step, we inspect the drivers of the plunge in air travel demand since the pandemic began, differentiating markets by their predicted social inclusion intensity. After controlling for potential endogeneity stemming from the spread of COVID-19 through air travel, our results suggest that short and low-density routes are among the most impacted airline markets and that business-oriented routes are more impacted than leisure ones. Finally, we estimate that a market with 1% higher social inclusion is associated with a 0.153%-0.166% more pronounced decline in demand during the pandemic. Therefore, markets that have benefited from greater social inclusion in the country may be the most vulnerable to the current crisis.

Introduction

The COVID-19 outbreak triggered a dramatic worldwide shock to air transport demand. The significant health challenges, intensified by the fear of contracting and spreading the Coronavirus,1 provoked the closure of several international borders and forced an unprecedented wave of flight cancelations. This has strongly impacted commercial aviation and related tourism sectors and brought the risk of a severe economic crisis that may last for many years. In many emerging economies, the short-, medium-, and long-term consequences of the pandemic pose a substantial threat to the middle-income consumer segment, which has expanded considerably over the past few decades.2

In Brazil, considerable advances in social inclusion since the 1990s have allowed the emergence of hundreds of thousands of middle class-travelers (Neri, 2015; Klein et al., 2018). Macroeconomic stabilization policies, economic growth, and improvements in living conditions, combined with the industry's economic deregulation, decisively contributed to the greater participation of a rising middle class. Additionally, labor income inequality indicators began to improve in 2019, after having worsened for four years since the start of the recession.3 However, with the advent of the coronavirus crisis, income inequality has been intensifying across the country, only partially mitigated by the temporary emergency aid grant from the government to more than 60 million people.4 As a result of the slump in economic activity and the coronavirus-induced health crisis, air travel dropped by 81%—from 45.7 million revenue passengers in the second and third quarters of 2019 to only 8.7 million for the same period in 2020.

This study investigates the impact of the coronavirus crisis on air transportation, focusing on the future dynamics of market positioning in the airline industry. In Brazil, the low-cost carrier (LCC) business model has been losing strength over the years, with only Gol Airlines operating in this segment but without large differentials in ticket prices and mean yield than its primary rival, Latam.5 Since its merger with the regional trip airlines, Azul Airlines, another incumbent, has been positioning itself more strongly toward high-yield passengers and the monopolistic regional markets. The coronavirus crisis has brought relevant challenges to each of these firms as the traditional passenger segment has changed its purchasing behavior, with business travelers swapping trips for videoconferencing. In this context, in which market segments has the current crisis most impacted non-business passengers’ demand? The markets in which non-business passengers strongly participate are usually those related to tourism and the regions marked by greater social inclusion over the past decades.

Our econometric methodology for identifying markets characterized by greater social inclusion in the Brazilian air transportation industry allows us to quantify the drivers of the pandemic's short-term demand shrinkage throughout 2020, to estimate the demand evolution for the coming years, and to figure out the airlines' attendant business model's needs for adjustment. This methodological framework employs indicators that capture income distribution, credit availability, and internet access at the local level as proxies for social, financial, and digital inclusion factors in the industry, allowing us thereby to estimate their impacts on air travel. We contribute to the literature by carrying out an empirical analysis of the determinants of the sharp drop in demand for air travel since the pandemic began by differentiating the markets according to their estimated intensity of social inclusion. Furthermore, this approach addresses endogeneity issues that arise due to the possible spread of COVID-19 through air travel, a factor that may cause inconsistent estimation and bias in air transport demand coefficients when considering 2020 data. Our results suggest that, after controlling for the intensity of the coronavirus contagion, social inclusion factors are relevant in explaining the fall in traffic on domestic routes over the period.

This paper is organized as follows. Section 2 reviews the relevant literature on the impact of the COVID-19 pandemic on air travel markets; Section 3 describes the Brazilian air transport industry, discusses the improvement in social inclusion indicators and the emergence of first-time air transport consumers, and presents the data sample, the empirical model, and the estimation strategy; Section 4 shows and interprets the estimation results; and Section 5 concludes.

The impacts of the COVID-19 pandemic on air travel markets

The literature regarding the COVID-19 pandemic's impacts on air transport markets worldwide is thriving. To date, these studies have focused on airlines, air transport-related sectors, and the local economy and seek to discuss government actions to mitigate the associated effects.

Taking the passenger perspective, Monmousseau et al. (2020) investigate the impact of the travel restriction measures implemented during the COVID-19 pandemic on the U.S. airline industry. They compute, close to real time, indicators extracted from social media to measure how travel restrictions impacted the relationship between passengers and airlines. Iacus et al. (2020) develop a forecasting model to project air passenger traffic during the pandemic. They document that in the first quarter of 2020, the impact of aviation losses may have translated to a global GDP decline of 0.02%–0.12%. Maneenop and Kotcharin (2020) estimate the short-term impact of COVID-19 on the stock returns of 52 listed airline companies worldwide using an event study methodology. Focusing on the effects of historical uncertainty shocks on airline employment, Sobieralski (2020) finds that job loss is nearly 7% in the airline workforce, with an upper bound of over 13%, and that recovery following the shocks takes between four and six years. Brown and Kline (2020) study the managerial preparedness of U.S. airlines in the COVID-19 pandemic and conclude that airline management teams failed to learn from previous outbreaks. Czerny et al. (2020) focus on the post-pandemic recovery of the Chinese aviation market.

Macilree and Duval (2020) discuss International Civil Aviation Organization ICAO's role in coordinating safety provisions during the pandemic and nationwide state aid in airline bailouts, recapitalization, and ownership. Abate et al. (2020) study government support for airlines in the aftermath of the COVID-19 pandemic, indicating that governments prioritize sustaining air transport connectivity to protect economic activity and jobs in both the aviation and tourism sectors.

Naboush and Alnimer (2020) investigate the circumstances under which an air carrier is liable for the transmission of COVID-19 and the scope of the safety measures required by the ICAO to prevent its spread. Prince and Simon (2020) focus on the geographic heterogeneity of the impact of international travel on the spread of COVID-19 in the United States. Similarly, Nakamura and Managi (2020) calculate the overall relative risk of COVID-19 importation and exportation for every airport in local municipalities worldwide.

By contrast, Lamb et al. (2020) consider travelers' perspectives and investigate the factors associated with their willingness to fly during the COVID-19 pandemic. They survey 632 participants in the United States, extracting characteristics of demographics, personality, emotional state, and travel purposes, and they develop regression models for both business and pleasure travel. They find that the perceived threat from COVID-19, agreeableness, affect, and fear are statistically significant regressors for both the business and pleasure types of fliers. The absolute value of their estimated standardized coefficients associated with “the perceived threat from COVID-19″ and “fear” variables are higher for pleasure travelers than for business travelers: −0.172 against −0.159 for “perceived threat from COVID-19” and −0.063 against −0.050 for “fear.” Their approach allows us to better understanding the passenger behavior in the aftermath of the crisis. With the authors’ breakdown of passenger types—business and pleasure segments—it is possible to project adjustments to market positioning and even restructure the airline business models in the aftermath of the crisis.

In line with Lamb et al. (2020), we perform a market segmentation for analysis. However, unlike them, we focus on an emerging market country's air transport industry, wherein social inclusion is a critical factor in determining the consumer segments' relative participation: the higher the social inclusion, the lower the participation of the mainstream segment of business-related travelers. In contrast to their study, our econometric approach uses panel data from city pairs across the country to identify the routes that rely most on middle-class consumers and to investigate the vulnerability of these market segments to the impacts of the economic crisis brought about by the pandemic. Based on Prince and Simon (2020), who suggest that geographic heterogeneity of the spread of COVID-19 in the U.S. may be related to population density, cultural differences, public response, and potential virus carriers' inflow, we use a geographic heterogeneity lens to understand the effects of the pandemic on air transport demand.

The surveyed literature guides our empirical approach design. Based on the findings of Monmousseau et al. (2020), insights can be obtained regarding the demand induction factors resulting from greater digital inclusion that were possibly relevant to mitigating the fall in demand observed during the pandemic period. The digital inclusion factor is one of the key aspects considered in our work. The demand forecast of Iacus et al. (2020) motivated us to extend the literature to decompose the observed drop in air passenger traffic into its possible determinants. From the study by Maneenop and Kotcharin (2020) and Brown and Kline (2020), we can see the market pressures on airlines from the capital markets and management perspectives, respectively. Under such circumstances, airlines certainly need to adjust their business models to suit the new post-pandemic market reality—implications that we consider when interpreting our estimation results. With the analyses of Macilree and Duval (2020) and Abate et al. (2020), it is possible to see the impacts of public institutions' interventions on air transport during the crisis. We also consider variables related to public and corporate policies, such as the impact of the Brazilian government's emergency financial aid grant and the “Essential Air Network,” conceived by the airlines and the government. Finally, inspired by Naboush and Alnimer (2020), Nakamura and Managi (2020), and Lamb et al. (2020), we study the role of the number of COVID-19 cases and its possible endogeneity with air transport demand.

Social inclusion, demand for air travel, and the impacts of the COVID-19 pandemic

Traditionally, air travel demand is driven by factors related to income and price elasticities of demand (Gillen et al., 2007). In general, when economic activity accelerates, passenger travel is induced for business reasons and, as the disposable income of families increases, for their higher spending on leisure travel. With increased competition among airlines, prices fall, and thus the number of trips rises. However, in many emerging countries, poor national income distribution can hinder the air transport sector's expansion. As sustainable economic development is promoted with less concentrated income, a potentially significant emerging middle class may gain access to the air travel market, promoting traffic growth. This phenomenon is certainly in the context of the advancement of social inclusion in developing countries. According to the World Bank, “Social inclusion is the process of improving the terms on which individuals and groups take part in society—improving the ability, opportunity, and dignity of those disadvantaged on the basis of their identity."6

The United Nations emphasizes the role of some forms of social progress and social inclusion, highlighting issues of financial inclusion (access to useful and affordable financial products and services)7 and digital inclusion (access and use of information and communications technologies).8 According to the United Nations Department of Economic and Social Affairs, these issues have become even more relevant with “successive lockdowns and measures confinement put in place by governments to curb the spread of COVID-19."9

Before the pandemic, with the reality of continuous social progress, the growth of the middle class played a crucial role in driving structural changes in the economies’ consumption patterns as people came out of poverty. According to a consultancy report, from 1990 to 2005, the size of the middle class impressively grew from 1.4 to 2.6 billion worldwide.10 The rise of the middle class in Asia, Latin America, and Africa has boosted the propensity to fly in these regions. For this reason, the International Air Transport Association estimated in 2015 deliveries of more than 6000 new jets with a capacity of between 70 and 130 seats by 2035 in the region.11

With the advent of health and economic crises triggered by the COVID-19 pandemic, much uncertainty has emerged regarding the continuation of the spillover effects of social inclusion on air transport demand that had marked many emerging markets for years. The abrupt drop in air travel in 2020, although global, has had varying intensities across countries and has been dictated by the unpredictable advancement of new coronavirus variants. Additionally, government restrictions as policy responses, such as business closures and lockdowns, have been heterogeneous and, in many cases, far from efficient. Ferraresi et al. (2020) find that countries characterized by low development and digitalization levels, among other factors, have adopted less stringent measures. As a result, the higher levels of unemployment and impoverishment brought about by the continuation of the pandemic status quo may produce dramatic and enduring socioeconomic consequences in many countries. In this context, a plausible hypothesis is that the pandemic may have so far been more harmful for the air travel demand in markets previously characterized by recent advances in social inclusion. Our study seeks to examine this hypothesis empirically in the context of domestic air transport in Brazil.

Research design

Our methodology for identifying air transport markets related to greater social inclusion relies on data for the period when social inclusion significantly advanced in Brazil. Since the macroeconomic stabilization in the 1990s, Brazil has experienced spurts of economic growth and welfare improvement for the less favored classes. At the same time, income transfer programs have helped combat extreme poverty in the country, while a declining income inequality gave rise to the emergence of a “new middle class” (Neri, 2015), who started consuming new goods as well as basic health, education, financial, and technology services. Combined with the airline industry's deregulation in the early 2000s, this social development hugely contributed to the emergence of hundreds of thousands of new air transport consumers in the 1990s and the 2000s. Air transport has become less associated with consumer elites, becoming a more frequent item in the consumption basket of many households. Nonetheless, the economic downturn in the mid-2010s, along with the coronavirus outbreak in 2020, has interrupted such trend.

The coronavirus outbreak in Brazil

On March 11, 2020, the World Health Organization (WHO) characterized the spread of the disease caused by the new Coronavirus, COVID-19, as a pandemic.12 Officially, the first cases of the disease appeared in China in December 2019, but soon after, new cases were diagnosed in Europe and other parts of Asia, subsequently spreading to the rest of the world. In Brazil, the first official record of the disease was on February 26, 2020.13 On March 17, 2021, COVID-19 had already caused more than 284,000 deaths in the country.14

Commercial aviation's total revenue passenger kilometers (RPK) grew in Brazil by 0.6% in 2019.15 The record passenger movement in January 2020 (more than 9.8 billion RPKs), as compared with January in the previous years,16 raised optimistic performance expectations for the sector. However, the alarming diffusion of the new Coronavirus in Brazil in early 2020 led the Ministry of Health to launch precautionary measures to deter its spread over the country, which soon proved insufficient. This sequence of events forced the government to announce mandatory quarantine in mid-March.17

The severity and high rate of contagion of the disease and the government's attendant sanitary measures to contain it dramatically affected most sectors of the economy and cast uncertainty in the near future. The tourism-related sectors have been among the most affected by the pandemic, notably air travel, due to the perception that they could be a conduit to the virus's spread. As in other countries, Brazilian airlines had to deal with an unprecedented adverse demand shock. Passenger traffic fell sharply with the nationwide spread of the disease. According to the National Civil Aviation Agency (ANAC),18 Brazilian airlines suffered a record 94.5% plunge in passenger movement in April 2020 as compared with the same period in 2019. The aircraft occupancy rate also registered a sharp drop in the same period, from 81.9% to 65.4%.19 Moreover, losses of leading airlines exceeded R$ 15 billion in the first six months of 2020.20

In March and April 2020, the government and airlines tried to tackle the crisis by relaxing slots rules at controlled airports (i.e., slots waiver), postponing airport concessionaires' contractual payments for the concession, and implementing an “essential airline network,” among other measures.21 In search of solutions to accelerate recovery and avoid further losses, Azul and Latam signed a codeshare agreement in August.22 Combining Latam's presence in the country's leading hubs and Azul's capillarity on flights to destinations with lower demand and infrastructure support, the agreement expanded their network, and the frequency of flights offered to passengers.

Sources of data

We rely on two datasets. The first dataset is a panel of directional city pairs of the domestic airline industry in Brazil, with monthly observations between July 2010 and December 2018. We include only passenger flights. To compute a city pair, we group multiple airports in the same catchment area. We only consider routes with six or more observations and a hundred passengers in each period. The full sample size is 65,452, but it reduces to 48,957 in some specifications owing to the social inclusion regressor's data availability. The second dataset reports only data for 2020 and comprises 763 original city pairs that still had scheduled flights in 2019. To examine thoroughly the drivers of the notable decline in air travel demand in 2020, we consider only the first full quarters since the pandemic outbreak. We then develop a cross-section of routes considering the second (Q2) and third (Q3) quarters of 2020 and compare the figures of 2020Q2–Q3 with 2019Q2–Q3 to analyze the pandemics' short-run effects on the air travel markets.

Most data for both datasets are available from the ANAC in online public databases.23 The agency provides information on all domestic and international scheduled flights on its air transportation market statistics database and its active scheduled flight report (VRA). For the social inclusion proxies, we use the following sources. First, regarding the standards of living, health, and education, we use data from Firjan - the Industry Federation of the State of Rio de Janeiro -, specifically, the Firjan Municipal Development Index.24 For the financial inclusion proxies, we use the Central Bank of Brazil's ESTBAN database, a publicly available dataset on the number of branches, and a detailed balance sheet for each local bank at the city-month level. We use data from the National Telecommunication Agency (ANATEL) for the digital inclusion proxies at the city-month level. Other data sources include the Brazilian Institute of Geography and Statistics (IBGE), the National Agency for Petroleum, Natural Gas and Biofuels (ANP), the National Land Transport Agency (ANTT), the Ministry of Health (regarding coronavirus cases), and the Brazilian government's Transparency Portal25 on emergency aid benefits.

Empirical model

Equations (1), (2) present the specifications of our proposed empirical model. First, Equation (1) presents our model of air travel demand in Brazil (first data set):and second, Equation (2) presents a framework to pinpoint the effects of the COVID-19 pandemic on air travel demand (second data set):where k denotes the city-pair, t the periods (t = 1, …, 102 months). u k,t and ε k are the error terms, and the β′s, γ′s and δ′s are the parameters to be estimated. In Table 1 , we provide a presentation of the variables in Equations (1), (2). Detailed information on each variable is available in the Appendix.

Table 1

Description of model variables.a

Variable	Description
Equation (1)
PAX	total airline tickets sold (ln)
INC	per capita gross domestic product (ln)
P	mean airline price (ln)
PBUS	mean bus price (ln)
TOUR	a proxy for the size of the tourism market (ln)
NET	number of cities served (ln)
LEIS	high season period (dummy)
TREND	time trend
REC	mid-2010s recession (dummy)
HDI(DSL)	Human Development Index (HDI) 's dimension for “decent standard of living”, as a proxy for social inclusion (ln)
CELL	cell phones, as a proxy for digital inclusion (ln)
LOAN	households' access to credit, as a proxy for financial inclusion (ln)
DEBT	household indebtedness, as a proxy for financial inclusion (ln)
Equation (2)
PAXVAR	variation in total air tickets sold between the two first quarters after the pandemic, 2020, and the same period in the previous year, 2019 (fraction)
DIST(j)	j-th route distance interval: 0–500 km (base case), 500–1000 km, 1000–1500 km, 1500–2000 km, 2000 km or higher (dummies)
PVAR	variation in mean air travel price between the two first quarters since the pandemic outbreak, 2020, and the same period in the previous year, 2019 (fraction)
INCPRE	per capita gross domestic product previous to the pandemic
DENSPRE	route density previous to the pandemic
LEISPRE	proportion of leisure passengers previous to the pandemic
ESSENTNET	“Essential Air Network”, an emergency play carried out by the government and the airlines
AMAZ	Amazon state countryside (dummy)
CODESHARE	Latam-Azul airlines codeshare agreement operations (dummy)
COVID	confirmed COVID-19 infection cases (ln)
DIST	route distance in kilometers (ln)
EMERGAID	emergency financial aid grants (ln)
SOCINCL	a proxy for social inclusion, predicted from Equation (1) 's coefficients of HDI(DSL), CELL, LOAN, and DEBT.

Note that we omit subscripts k and t. See details of each variable in the Appendix.

Estimation strategy

The empirical approach of the present study involves estimating the two demand equations. Equation (1), which has PAX as the regressand, estimates the effects of social, digital, and financial inclusion factors that were possibly very relevant in generating air travel demand in the early 2010s. For this purpose, we employ the regressors HDI (DSL), CELL, LOAN, and DEBT. In the first specification, we use HDI—the overall Human Development Index of the endpoint cities—to capture the effects of socioeconomic factors in the most general way. Equation (2), which has PAXVAR as the regressand, has the form of demand variation, to capture the main drivers of the remarkable fall in demand for air travel in the first two quarters of the pandemic (2020Q2–Q3) as compared with the same period of the previous year. For this equation, we focus on the results of the regressor COVID and its interactions (COVID × DIST, COVID × EMERGAID, and COVID × SOCINCL). Through interaction variables, it is possible to test the existence of moderation effects (intensifiers or attenuators) on the relationship between COVID and PAXVAR in the sample period. We focus on the hypothesis test of the estimate of COVID × SOCINCL because SOCINCL is calculated from a prediction using the estimated coefficients of variables HDI (DSL), CELL, LOAN, and DEBT in Equation (1). The idea of this procedure is to test whether the routes assigned with the most prominent social, digital, and financial inclusion indicators in the early 2010s would have some differentiated impact caused by the COVID-19 pandemic.

Our two-step procedure can be summarized as follows. In the first step, we estimate the social inclusion parameters of demand in Equation (1). In the second step, we calculate the estimated social inclusion metric SOCINCL and plug it into Equation (2) as an interaction with COVID. The estimation problem here is that the standard errors of the second-step parameters must be adjusted to reflect that the social inclusion variables are previously estimated. To accomplish this, we compute bootstrap-adjusted standard errors computed using a stratified bootstrapping procedure with 2000 replications.

Our empirical approach addresses two additional specification challenges. First, we need to account for unobserved factors that drive air travel demand across a large country like Brazil. Undoubtedly, many regional idiosyncrasies either stimulate or hinder demand that our framework of a few covariates in Equations (1), (2) is unable to control for. For example, in Equation (2), we suspect that psychological effects related to the fear of the coronavirus contagion may drive the assessment of risk and the purchase behavior of consumers (Lamb et al., 2020; Kim et al., 2020), and, for a theoretical framework, we refer to Kahneman and Tversky (1979). Additionally, although we control for the government's emergency financial aid granted since the first months of the pandemic, the actual evolution of local economies is a critical unobserved factor of our empirical model. Therefore, we consider an amplified set of control variables in our robustness checks to account for such unobservables in our analysis. In Equation (1), we employ region-specific seasonal dummies, local temperature, and panel time dummies to account for local seasonality and global time effects. In Equation (2), we use regional and region-interacted dummies to account for specificities that are particular to the endpoint airports' regions of the routes. We provide details of these settings in the robustness check section.

The second specification challenge is endogeneity. In Equations (1), (2), we have a typical problem of demand-side endogeneity of the right-hand side price variables: P, P × LEIS, and PVAR. Therefore, we need proper instrumental variables (IV) to proceed with a consistent estimation. For these variables, we use cost shifters, Hausman-type, Stern-type, and BLP-type IVs (See Mumbower, Garrow, and Higgins (2014) for a detailed discussion of these classes of IVs). All IVs are computed in natural logarithms.

A more subtle endogeneity issue is related to the pandemic's effect on the decline in air travel intensity. In Equation (2), COVID is potentially correlated with the unobserved factors that drive PAXVAR. The endogeneity of COVID is due to the possible spread of the SARS-CoV-2 virus through air travel (Adiga et al., 2020, Barnett and Fleming, 2020; Nakamura and Managi, 2020). Schuchat et al. (2020) describe the continued travel-associated importation of the Coronavirus from outside the U.S. as an early contributor to the initiation and acceleration of domestic COVID-19 cases. Prince and Simon (2020) investigate the effect of international air travel on the spread of COVID-19 in the United States. They consider environmental and exposure factors that explain the substantial geographic heterogeneity of the COVID-19 spread. They employ population density, cultural differences, and public response proxies as environmental factors, and the inflow of international airline travelers as exposure factors. They find statistically significant impacts of the inflow of international airline travelers on the COVID-19 spread in the U.S. territory when the air traffic originated from pandemic hotspots.

As the coronavirus contagion is possibly positively correlated with the unobserved dynamics of the local economies, we consider that COVID may be correlated with the error term of Equation (2), thus raising the risk of inconsistent estimation—more specifically, a positive bias that may underestimate the downward effect of the COVID variable. To address this issue, we utilize other factors that may be correlated with the spread of COVID-19, such as the endpoint cities' population densities, areas, and public health and education conditions.26 We also use these variables to interact with the LEIS as additional instruments. Finally, we employ the origin and destination's minimum, mean, median, and maximum local temperatures as instruments.27 We build upon our IV strategy on recent studies suggesting some of the factors that might contribute to the spread of Coronavirus, either directly or indirectly. Schuchat et al. (2020) point to high population density as a probable contributor to the spread, among other factors. Pramanik et al. (2020) and Tosepu et al. (2020) find evidence of the role of climatic predictors. Kejela (2020) suggests factors such as inadequate knowledge and public hygiene conditions, the society's socioeconomic nature, and the knowledge, attitude, and practice of health care workers. Robertson (2021) employs proxies for pre-existing health conditions in the population and local economic and demographic conditions, among others.

Estimation results

Table 2, Table 3 present the results of our empirical models of air travel demand (PAX) and air travel demand decline due to the COVID-19 pandemic (PAXVAR) in the Brazilian airline industry. To simplify the exposition of the results, we omit indexes k and t of the model variables.

Table 2

Estimation results: air travel demand (PAX).

	(1)	(2)	(3)	(4)	(5)	(6)	(7)	(8)	(9)
	PAX	PAX	PAX	PAX	PAX	PAX	PAX	PAX	PAX
INC	1.8229***	1.8614***	1.4219***	1.4621***	1.2579***	1.3145***	1.2267***	0.9260***	0.4603***
P	−1.3307***	−1.3069***	−1.3200***	−1.3381***	−1.3020***	−1.3118***	−1.3616***	−1.4593***	−1.5521***
PBUS	0.4355***	0.2625***	0.5161***	0.5267***	0.6626***	0.4388***	0.4984***	0.1606**	0.1518*
TOUR	0.0059***	0.0064***	0.0070***	0.0071***	0.0086***	0.0083***	0.0077***	0.0057***	0.0055***
NET	0.1751***	0.1263***	0.1010***	0.1066***	0.1311***	0.1412***	0.1207***	0.1072***	0.1295***
P × LEIS					−0.0117***	−0.0234***	−0.0236**	−0.0515***	−0.0387***
TREND	−0.0044***	−0.0042***	−0.0042***	−0.0030***	−0.0022***	−0.0022***	−0.0026***	−0.0029***
TREND × REC				−0.0012***	−0.0017***	−0.0017***	−0.0016***	−0.0012***
HDI		1.3169***
HDI (DSL)			0.5366***	0.3833***	0.4091***	0.3708***	0.4935***	0.6736***	0.2996***
CELL			0.6071***	0.5076***	0.5313***	0.4159***	0.5463***	0.6595***	0.7381***
LOAN			−0.0001	−0.0007	−0.0010*	0.0027***	−0.0012**	−0.0010*	−0.0005
DEBT			−0.0460***	−0.0455***	−0.0296***	−0.0285***	−0.0261**	−0.0196*	−0.0749***
Estimator	FE/IV/LASSO	FE/IV/LASSO	FE/IV/LASSO	FE/IV/LASSO	FE/IV/LASSO	FE/IV/LASSO	FE/IV/LASSO	FE/IV/LASSO	FE/IV/LASSO
Alternative Version	No	No	No	No	No	SUBSET	No	No	No
City-Pair Clusters	1041	1013	1013	1013	1013	1013	1013	1013	1013
Local Temp Controls	No	No	No	No	No	No	15/48	11/48	11/48
Reg/Seas Controls	No	No	No	No	No	No	No	250/300	242/300
Panel Time Controls	No	No	No	No	No	No	No	No	72/102
IV Count	20/438	16/438	16/438	16/438	43/876	26/876	43/876	42/876	41/876
AIC Statistic	19,572	6611	5530	5847	4326	5449	4596	3586	3649
BIC Statistic	19,626	6673	5618	5944	4431	5555	4834	5987	6586
Adj R2 Statistic	0.4800	0.4693	0.4809	0.4776	0.4936	0.4818	0.4893	0.5022	0.5022
RMSE Statistic	0.2832	0.2615	0.2586	0.2595	0.2555	0.2584	0.2562	0.2522	0.2521
RMSE CV Statistic	1.1862	1.2378	1.2703	1.2574	1.2376	1.289	1.2533	1.3082	1.3218
Nr Observations	65,452	48,957	48,957	48,957	48,957	48,957	48,748	48,748	48,748

Notes: Estimation results produced by the instrumental variables, post-double-selection LASSO-based methodology of Belloni et al. (2012, 2014a,b), with fixed effects (IV-LASSO). Post-LASSO estimation is performed with a Two-Stage Least Squares, fixed-effects, procedure with standard errors robust to heteroskedasticity and autocorrelation. LASSO penalty loadings account for the clustering of city-pairs. Control variables estimates omitted. Blank cells indicate that the variable was not used. INC, P, PBUS, TOUR, and NET not penalized by LASSO. Endogenous variables: P and P × LEIS. Column (6)'s model employs alternative versions of the following subset of variables: PBUS, TOUR, and NET. Cross-validation performed with a 4-fold procedure. P-value representations: ***p < 0.01, **p < 0.05, *p < 0.10.

Table 3

Estimation results: air travel demand change due to the COVID-19 pandemic (PAXVAR).

	(1)	(2)	(3)	(4)	(5)	(6)	(7)	(8)	(9)
	PAXVAR	PAXVAR	PAXVAR	PAXVAR	PAXVAR	PAXVAR	PAXVAR	PAXVAR	PAXVAR
DIST 500–1000	0.0606***	−0.1932***	−0.0872***	−0.0938***	−0.0942***	−0.1573***	−0.1490***	−0.1648***	−0.1686***
DIST 1000–1500	0.1189***	−0.3091***	−0.1285**	−0.1407***	−0.1349**	−0.2371***	−0.2229**	−0.2477***	−0.2513***
DIST 1500–2000	0.1600***	−0.3796***	−0.1548**	−0.1731**	−0.1566**	−0.3000***	−0.2697**	−0.3216***	−0.3165***
DIST 2000 HIGH	0.1971***	−0.4834***	−0.2046**	−0.2295***	−0.2169**	−0.3757***	−0.3581***	−0.4078***	−0.4018***
PVAR	−0.5759***	−0.2621***	−0.4827***	−0.4796***	−0.4746***	−0.6835***	−0.6014**	−0.6956***	−0.7851***
INCPRE	−0.1133	0.0290	0.0051	−0.0142	0.0034	0.0009	0.0659	0.0309	−0.0057
DENSPRE	0.0359***	0.0272***	0.0309***	0.0245***	0.0278***	0.0275***	0.0326***	0.0260***	0.0370***
LEISPRE	0.1523***	0.1776***	0.1205***	0.1226***	0.1211***	0.0947*	0.0954*	0.0822	0.0658
ESSENTNET	−0.0224	−0.0040	−0.0186	−0.0202	0.0416	0.0600	0.0701	0.0612	0.0517
ESSENTNET × AMAZ	0.1651**	0.2160***	0.1946***	0.2065***	0.1397*	0.1829	0.1752	0.1856	0.2156*
CODESHARE					−0.0932	−0.0851	−0.1099	−0.0952	−0.0868
COVID	−0.0201	−0.2845***	−0.2350***	−0.2588***	−0.2568***	−0.3385***	−0.4770***	−0.3612***	−0.5490***
COVID × DIST		0.0328***	0.0173***	0.0199***	0.0190***	0.0287***	0.0370***	0.0301***	0.0432***
COVID × EMERGAID			0.0268***	0.0320***	0.0324***	0.0368***	0.0600***	0.0382***	0.0692***
COVID × SOCINCL				−0.0133**	−0.0119**	−0.0164**	−0.0224**	−0.0197***	−0.0395***
Estimator	IV-LASSO	IV-LASSO	IV-LASSO	IV-LASSO	IV-LASSO	IV-LASSO	IV-LASSO	IV-LASSO	IV-LASSO
Alternative Version	No	No	No	No	No	No	COVID	SOCINCL	SUBSET
City Clusters	88	88	88	88	88	88	88	88	88
Region Controls	5/5	5/5	5/5	5/5	5/5	4/5	4/5	4/5	3/5
Region/Interact Controls	No	No	No	No	No	27/70	20/70	26/70	21/70
IV Count	8/163	11/115	17/163	17/163	17/163	9/163	7/163	9/163	7/163
AIC Statistic	−551	−633	−666	−667	−664	−629	−611	−612	−519
BIC Statistic	−473	−550	−578	−574	−566	−411	−425	−398	−334
Adj R2 Statistic	0.3842	0.4474	0.4713	0.4729	0.4712	0.4645	0.4464	0.4513	0.3759
RMSE Statistic	0.1649	0.1561	0.1526	0.1522	0.1524	0.1506	0.1539	0.1526	0.1634
RMSE CV Statistic	0.1594	0.1588	0.1599	0.1605	0.1595	0.1525	0.1560	0.1530	0.1529
Nr Observations	763	763	763	763	763	763	763	763	763

Notes: Estimation results produced by the instrumental variables, post-double-selection LASSO-based methodology of Belloni et al. (2012, 2014a,b), with fixed effects (IV-LASSO). Post-LASSO estimation is performed with a Two-Stage Least Squares, fixed-effects, procedure with standard errors robust to heteroskedasticity and autocorrelation. LASSO penalty loadings account for the clustering of city-pairs. Control variables estimates omitted. Blank cells indicate that the variable was not used. DIST, PVAR, INCPRE, DENSPRE, LEISPRE, ESSENTNET, ESSENTNET × AMAZ, CODESHARE, and COVID not penalized by LASSO. Endogenous variables: PVAR, COVID, and CODESHARE. Column (6)'s model employs alternative versions of the following subset of variables: COVID, SOCINCL, INCPRE, and DENSPRE. Cross-validation performed with a 4-fold procedure. P-value representations: ***p < 0.01, **p < 0.05, *p < 0.10.

Table 2 shows the estimation results for the demand for air travel using data from 2010 to 2018. The nine columns of the table show the coefficient estimation results from using the FE-IV-LASSO procedure. Column (1) presents the baseline model, which contains no social inclusion proxies. This column shows that the estimated income (INC) and price (P) elasticities of demand are 1.8229 and −1.3307, respectively, indicating a relatively high sensitivity of demand to these variables. The PBUS, TOUR, and NET variables have positive, statistically significant estimated coefficients of 0.4355, 0.0059, and 0.1751, respectively, indicating that increases in these factors are associated with demand increases. The trend variable TREND suggests a persistent decrease in demand over time, with an estimated coefficient of −0.0044. These results remain relatively invariant throughout the remaining columns in Table 2.

In Column (2) of Table 2, we insert our first social inclusion metric—the Human Development Index regressor (HDI). The estimated coefficient of this variable is positive and statistically significant, suggesting that enhancements in quality of life and socioeconomic conditions are associated with a higher market size for air transport. In Column (3), we present a breakdown of social inclusion indicators to capture phenomena related to improvements over the geographic space and time in metrics of life conditions: HDI (DSL), digital inclusion (CELL), and financial inclusion (LOAN and DEBT variables). The results in Column (3) and the remaining columns suggest that some of these factors are statistically related to the air travel demand in the sample. The variables HDI (DSL), CELL, and DEBT have statistically significant estimated coefficients; the first two are positive and the last negative. These results indicate that demand is apparently positively driven by socioeconomic factors associated with social inclusion, such as income distribution and access to the Internet: HDI (DSL) and CELL. The result related to household indebtedness (DEBT) shows that financial inclusion has the downside of increasing household indebtedness and discouraging air travel. It is important to note that the LOAN variable, another proxy for financial inclusion through access to bank credit, is not statistically significant in most specifications; this provides some evidence that families did not count on bank credit for airline ticket purchases over the sample period.

In Columns (4) and (5) of Table 2, we insert interaction variables to capture possibly more complex phenomena associated with the demand for air travel: a break in the trend due to the 2014–2016 recession (TREND × REC) and a possible moderating effect of price elasticity of demand on routes with a higher proportion of leisure travelers (variable P × LEIS). Both estimated coefficients are negative and statistically significant, suggesting an intensification role for the respective variables. We consider the model in Column (5) as our preferred specification.

Table 3 presents the estimation results of Equation (2), which is representative of the remarkable drop in demand for air travel due to the COVID-19 outbreak (PAXVAR). Again, we present the results in nine columns, with the first specification being the baseline model. In all versions, we employ the IV-LASSO estimator. As discussed in the empirical strategy section, we apply stratified bootstrapping to adjust the standard error of estimates in the specifications in which we include the social inclusion variables (SOCINCL).28

The estimation results in Column (1) of Table 3 indicate the positive effects of distance with respect to PAXVAR. In fact, the estimated coefficients of variables DIST 500–1000, DIST 1000–1500, DIST 1500–2000, and DIST 2000 HIGH are all positive and statistically significant, allowing us to infer that longer routes suffered less demand fall than shorter routes. Additionally, as expected, PVAR has a negative estimated coefficient (−0.5759), suggesting that airlines could partially manage the intensity of the demand drop with price adjustments. Regarding the variables of the routes’ characteristics before the pandemic, namely, INCPRE, DENSPRE, and LEISPRE, we observe that only the last two are statistically significant and positively related to PAXVAR. Concerning the essential air network outlined by the major carriers and the government, we have some evidence that only the most inaccessible region of the Brazilian Amazon has benefited from the commitment to the maintenance of operations, as indicated by the results of the variables ESSENTNET (nonsignificant) and ESSENTNET × AMAZ (positive and statistically significant). Finally, the results in Column (1) point to an apparent absence of statistical significance of the COVID variable: the estimated coefficient of this variable is negative but statistically nonsignificant in Column (1).

In Column (2) of Table 3 we insert the interaction variable COVID × DIST. Comparing the results of Columns (1) and (2) reveals some interesting changes. First, the estimated coefficient of COVID maintains its negative sign, but becomes statistically significant at the 1% level. This result is clearly associated with the introduction of COVID × DIST, which has a statistically significant estimated coefficient with a positive sign (0.0328). These results indicate that the downward effect of the spread of COVID-19 on the air transport demand is attenuated by route distance: the longer the route, the lower the decline in demand caused by the pandemic. This finding is consistent with consumers' difficulty in substituting air travel with road trips in medium-to-long-haul markets.29 It is thus in line with Bauer et al. (2020), who provide evidence that airline business models based on long-haul operations may gain competitive advantage in the COVID context and in the adjustments to the new-normal.

Also notable is the complete flip in the coefficients of the DIST dummy variables' signs (DIST 500–1000, DIST 1000–1500, …): unlike Column (1), they are all negative in Column (2). This apparent contradiction does not alter the finding in Column (1) that PAXVAR is positively correlated with distance. With the interaction term COVID × DIST, the full effect of distance on PAXVAR in (2) is equal to the coefficient of the respective distance dummy plus the coefficient of the interaction term multiplied by COVID. This second term outweighs the first for the majority of the observations, such that the estimated full influence of distance on PAXVAR is still positive, consistent with the results in Column (1).30

In Columns (3), (4), and (5), we introduce regressors COVID × EMERGAID, COVID × SOCINCL, and CODESHARE, respectively. These variables have estimated coefficients that were positive, negative, and nonsignificant, respectively. From these specifications, we can infer that: 1) the emergency aid granted by the government in response to the pandemic had an attenuation effect on the relationship between COVID and PAXVAR; 2) the routes marked by greater social inclusion in the early 2010s had a more significant impact on demand during the first quarter of the COVID-19 pandemic crisis; and 3) the codeshare agreement between Latam and Azul airlines does not seem to have produced any immediate soothing effect on demand drop.

We consider the specification (5) as our preferred PVAR model. With this model, we estimate the full elasticity of PAXVAR with respect to COVID to be between −0.0082 and −0.0880, considering the 25 and 75 percentiles of DIST, EMERGAID, and SOCINCL. Furthermore, as we are particularly interested in the interpretation of the estimates of the COVID × SOCINCL variable, we also report its associated full effect. Based on our estimation results, we have evidence that ceteris paribus, a 1% increase in SOCINCL is associated with a 0.153%–0.166% decrease in air travel demand in the sample period.31 This result has important implications for airlines’ corporate policy that may help their market repositioning and business model adjustments in the challenging years to come. With fewer passengers from the middle class, demand tends to shrink and become less price elastic. However, with the recent phenomenon of greater substitutability of air travel for virtual meetings by corporate passengers, the demand from this mainstream market segment of business travel may also diminish. Consequently, it becomes a matter of market survival for carriers to find ways to foster non-business passenger air travel demand. Our results may provide evidence in favor of the necessity of expanding the low-cost business model to recover at least part of the profitability lost in 2020 from the exploitation of leisure travel segments.

Robustness checks

We consider some robustness checks to inspect the sensitivity of our main estimation results displayed in the fifth columns of Table 2, Table 3 We present the results of these experiments in Columns (6), (7), (8), and (9) of each table.

First, we check the robustness of the PAX model. In Column (6) of Table 2, we employ different metrics for some of the key variables of our PAX model. Unlike the remaining columns of Table 2, the model in (6) employs alternative versions of PBUS, TOUR, NET, LEIS, CELL, LOAN, and DEBT, as described in the Appendix. Note that it represents a relevant perturbation, with changes in more than half of the regressors. As a result of this experiment, the vast majority of variables kept their estimated coefficients relatively stable. The only change in the results concerns the LOAN variable, whose coefficient becomes positive and statistically significant. In the other robustness checks, this variable does not follow such a result, becoming nonsignificant in most specifications. We consider this particular estimate of LOAN an isolated finding, whose case should be further investigated in future studies.

In Columns (7)–(9), we introduce a broad set of controls to better approach the unobserved factors that may influence air travel demand during the pandemic. In Column (7), we introduce 48 local temperature controls32 ; in Column (8), we add 300 region-specific seasonal dummy variables33 ; and finally, in Column (9), we introduce 102 panel time controls. Note that in the latter estimation, due to perfect collinearity with the time dummies, the TREND and TREND × REC variables are not included in the regressor set. All the proposed controls were included in the LASSO penalization procedure to avoid overfitting. The final step of the estimation reported in Column (9) comprises 325 controls (11 local temperatures + 242 region/seasons + 72 times) selected by the LASSO estimation, from the 450 (48 + 300 + 102) originally assigned to the model. In all these robustness checks, the key results in Column (5) remain roughly unchanged.

Second, we check the robustness of the PAXVAR model. In Column (6) of Table 3, we insert a set of interaction terms as controls for the economic conditions of the local economies during the pandemic. To construct these terms, we multiply each regional dummy by the geometric mean and the simple mean of the gross domestic product and the population of the endpoints of each route. We then multiply each of these new variables by dummies of trunk routes (i.e., routes between state capitals) and dummies of regional routes (i.e., routes between country towns or between a capital and a country town). In total, we create 70 regions-interacted controls to capture the socioeconomic idiosyncrasies of each route's local economies in the sample period. Again, the proposed controls are included in LASSO's penalization for model selection. The procedure selected 27 controls from 70 subjects originally assigned. The results of this robustness check are roughly the same as those in Column (5).

Finally, in Columns (7)–(9) we employ alternative versions of COVID, SOCINCL, INCPRE, and DENSPRE. The description of the alternative versions of the regressors is provided in the Appendix. See Table 4, Table 5, Table 6. Note that in these experiments, we also insert the region-interacted controls developed for Column (6). In all of the proposed experiments of Columns (7), (8), and (9), most of the estimated coefficients remain unchanged, thus providing further evidence of the robustness of our key findings.

Table 4

Alternative version of regressors - PAX model (Table 1).

Regressor	Main model	Alternative Version
PBUS	This variable is centered on a reference of regulated bus yield - equal to 0.1524 BRL per kilometer in July 2015.	Instead of the reference of bus yield, we employ an inflation-adjusted bus price index (base 100) computed from Brazil's consumer price index data base. It is equal to the geometric mean yield centered at 100 on a route-specific basis, and without multiplying by distance.
TOUR	The number of total charter flights on the route divided by the maximum population between the origin and destination cities.	The number of total charter flights on the route divided by the maximum number of flights between the origin and destination cities.
NET	The geometric mean (between O and D) of the number of network points (cities) served with flights.	The maximum (between O and D) of the number of network points (cities) served with flights.
LEIS	A dummy of the high-season period - the Summer in the Southern hemisphere, from December to March.	A dummy of the high-season period - the Summer in the Southern hemisphere, from December to March -, interacted with the proportion of leisure passengers traveling on the city-pair (Source: EPL - National Logistics and Planning Company’s 2014 passenger survey).
CELL	The geometric mean (between O and D) of the number of cell phones per 100 inhabitants.	The maximum (between O and D) of the number of cell phones per 100 inhabitants.
LOAN	The geometric mean (between O and D) of the difference between the amount of new loans and those paid off in the last twelve months.	The maximum (between O and D) of the difference between the amount of new loans and those paid off in the last twelve months.
DEBT	The geometric mean (between O and D) of the household bank debt per capita.	The maximum (between O and D) of the household bank debt per capita.

Table 5

Alternative version of regressors - PAXVAR model (Table 2).

Regressor	Main model	Alternative Version
COVID	The maximum number of confirmed Coronavirus infection cases between the endpoint cities.	The maximum number of confirmed Coronavirus deaths between the endpoint cities.
SOCINCL	To avoid contaminating this variable's effect with the recession period of 2014q2-2016q4, we consider only predicted values until 2013.	The recession began in the mid-2014, then it is debatable if we can use either 2013 or 2014 as a cutoff year. Instead of considering 2013 as a cutoff, we consider 2014.
INCPRE	The geometric mean (between O and D) of the per capita gross domestic product.	The maximum (between O and D) of the per capita gross domestic product.
DENSPRE	It is equal to the number of total sold airline tickets on the city-pair in the second and third quarters of 2019.	Instead of using only 2019 figures, we alternatively consider the city-pair specific mean of 2017, 2018, and 2019 (second and third quarters).

Table 6

Descriptive Statistics of model variables.

Eq.	Variable	Description	Level of measurement	Metric	N. Obs.	Original scale
						Mean	Std. Dev.	Minimum	Maximum
(1)	PAX	airline tickets sold	city-pair	Count	65,452	5436.37	12,425.45	100	2,03,831
	INC	GDP per capita	geom. mean (O, D cities)	BRL constant values	65,452	2945.54	938.56	1004.52	7460.04
	P	airline price	city-pair	BRL deflated	65,452	473.40	214.35	45.96	2047.70
	PBUS	bus price	geom. mean (O, D cities)	BRL deflated	65,452	189.70	127.98	18.30	641.13
	PBUS (alt)	bus index price	geom. mean (O, D cities)	index (mean = 100)	65,452	101.36	4.47	78.20	119.77
	TOUR	charter flights per population	city-pair; max (O, D cities)	ratio (K per milion)	65,452	51.63	213.77	0.00	10,689.45
	TOUR (alt)	charter flights per city flights	city-pair; max (O, D cities)	ratio (per milion)	65,452	90.57	453.32	0.00	23,728.81
	NET	served cities	geom. mean (O, D cities)	Count	65,452	19.25	9.68	2.00	63.94
	NET (alt)	served cities	max (O, D cities)	Count	65,452	38.87	17.07	2.00	74.00
	LEIS	high season	systemwide	Dummy	65,452	0.33	0.47	0	1
	LEIS (alt)	leisure passengers during high season	city-pair; systemwide	fraction × dummy	65,452	0.13	0.21	0.00	0.94
	TREND	time trend	systemwide	discrete sequence	65,452	49.35	29.06	1	102
	REC	recession (Apr 2014–Dec 2016)	systemwide	Dummy	65,452	0.32	0.47	0	1
	HDI	human development index	geom. mean (O, D cities)	index × 100	48,965	73.22	6.48	48.99	85.06
	HDI (DSL)	human development index (DSL)	geom. mean (O, D cities)	index × 100	48,965	67.61	7.45	40.33	80.80
	CELL	cell phones per population	geom. mean (O, D cities)	ratio × 100	65,452	96.29	39.80	9.41	266.46
	CELL (alt)	cell phones per population	maximum (O, D cities)	ratio × 100	65,452	134.68	50.89	22.01	481.59
	LOAN	new loans per population	geom. mean (O, D cities)	ratio × 100	65,452	1497.20	3491.69	0.00	53,203.53
	LOAN (alt)	new loans per population	maximum (O, D cities)	ratio × 100	65,452	6066.53	11,464.21	0.00	60,066.26
	DEBT	household bank debt per population	geom. mean (O, D cities)	Ratio	65,452	527.11	463.37	14.34	4141.37
	DEBT (alt)	household bank debt per population	maximum (O, D cities)	Ratio	65,452	1472.61	1616.83	16.69	6336.12
(2)	PAXVAR	change in tickets from 2019 to 2020	city-pair	change rate	765	−0.67	0.21	−1.00	−0.01
	DIST 500 - 1000	distance between 500 and 1000 km	city-pair	Dummy	765	0.30	0.46	0	1
	DIST 1000 - 1500	distance between 1000 and 1500 km	city-pair	Dummy	765	0.19	0.39	0	1
	DIST 1500 - 2000	distance between 1500 and 2000 km	city-pair	Dummy	765	0.14	0.35	0	1
	DIST 2000 HIGH	distance higher than 2000 km	city-pair	Dummy	765	0.15	0.36	0	1
	PVAR	change in price from 2019 to 2020	city-pair	change rate	765	−0.27	0.25	−0.80	2.09
	INCPRE	GDP per capita previous to the pandemic	geom. mean (O, D cities)	BRL deflated	765	36,205.37	11,060.59	15,152.16	72,434.62
	INCPRE (alt)	GDP per capita previous to the pandemic	maximum (O, D cities)	BRL deflated	765	47,477.85	18,987.11	15,775.61	91,453.09
	DENSPRE	tickets previous to the pandemic	city-pair	Count	765	22,330.78	51,096.69	117.00	5,58,086.00
	DENSPRE (alt)	tickets previous to the pandemic (2017–2019)	city-pair	mean over 3 years	765	23,051.77	53,444.26	85.00	6,37,286.31
	LEISPRE	leisure passengers previous to the pandemic	city-pair	Fraction	765	0.42	0.16	0.05	0.94
	ESSENTNET	essential air network routes	city-pair	Dummy	765	0.08	0.27	0	1
	AMAZ	amazon state routes	city-pair	Dummy	765	0.02	0.13	0	1
	CODESHARE	codesharing between Latam and Azul	city-pair	Dummy	765	0.13	0.33	0	1
	COVID	confirmed covid infections	maximum (O, D cities)	Count	765	2,68,393.97	2,90,974.48	25,505.00	9,31,341.00
	COVID (alt)	confirmed covid deaths	maximum (O, D cities)	Count	765	10,012.13	11,530.76	576.00	34,408.00
	DIST	distance	city-pair	Km	765	1157.43	739.27	103.87	3290.29
	EMERGAID	emergency financial aid grants per population	maximum (O, D cities)	ratio × 100	765	25.97	4.74	17.00	35.48
	SOCINCL	predicted social inclusion proxy (ref. 2013)	city-pair	Fraction	765	0.48	0.16	0.14	0.91
	SOCINCL (alt)	predicted social inclusion proxy (ref. 2014)	city-pair	Fraction	765	0.47	0.16	0.14	0.92

Conclusion

Summary and results

This study developed an empirical framework to pinpoint some of the drivers of the sharp decline in domestic air transport demand in Brazil since the coronavirus outbreak began. More specifically, we focused on the pandemic's impact on the routes linking cities marked by higher social inclusion. First-time consumers from the emergent classes have entered the air travel market with more intensity since the mid-nineties. The proposed methodology aimed to identify air transport markets possibly characterized by greater social inclusion and inspect their behavior during the first quarter of the pandemic period. Our contribution lies in assessing the possible impacts on the air travel demand of social inclusion indicators at the local economy level, such as income distribution, credit availability, and access to the Internet. Another contribution is the approach to address endogeneity regarding the possible COVID-19 contagion through air travel, thus avoiding inconsistent estimation of demand during the virus outbreak.

We produce evidence of the air travel market's differentiated reactions to the pandemic shock according to their social inclusion intensity. We find strong evidence that the demand plunge on routes marked by short haul and low density was the most dramatic. Moreover, we find suggestive evidence that business-oriented routes are impacted more strongly than leisure ones. We also estimate that a market with 1% higher social inclusion is associated with a roughly 0.153%–0.166% more pronounced decline in demand during the pandemic. Our results show that the markets most benefited by greater social inclusion in Brazil are the most vulnerable to COVID-19 persistence. Finally, we find a positive correlation between COVID-19 cases and the error term of demand variation during the first quarters of the pandemic. We tackle this endogeneity by employing a LASSO-based IV approach.

Policy implications

Our findings provide some groundwork for taking stock of the adjustments in business models that airlines need to hasten to return to the pre-pandemic traffic levels. In particular, our evidence underpins the view that markets that are more reliant on social inclusion should be targeted with deeper discounts to bring back price-sensitive consumers. Carriers that react more quickly to these new market conditions are likely to get a head start in the post-pandemic environment. In addition, we conclude that the public policies implemented to relieve the problems caused by the pandemic to the air transportation sector had conflicting results. On the one hand, the emergency financial aid grant was successful to maintain some activity in the local economies in order to moderate the effects of COVID-19 on the demand for air transport; on the other hand, the launch of an “Essential Air Network” by the government and the major airlines apparently was not effective in mitigating the intensity of the fall in demand for air travel. Greater articulation of public and private policies could have enhanced the effectiveness of those measures.

Limitations and possible extensions

In this study, due to data availability, we worked with separate data samples to address the impact of the COVID-19 pandemic on air transportation in Brazil. The ideal approach would be to work with a single, combined dataset to exploit how the pandemic may have induced exogenous variation in the air travel demand change. We believe that future studies could address this limitation with a more complete dataset. Additionally, we emphasize that our analysis is restricted to the first two quarters of the pandemic period. We believe that the inclusion of data from other periods after those used in this sample could reveal other phenomena and serve as an empirical test of the evidence raised in our study. Finally, we consider that applying to realities of other emerging countries that have also recently experienced greater social inclusion could contribute to the findings of this work being assessed in a more general international context.

Authors’ contribution

All authors contributed to the study conception and design. Material preparation, data collection and compilation were performed by Alessandro V. M. Oliveira. All authors contributed to the writing of the first draft of the manuscript and commented on its subsequent versions. All authors read and approved the final manuscript.

Declaration of competing interest

On behalf of all authors, the corresponding author states that there is no conflict of interest.