Content uploaded by Rob Vos
Author content
All content in this area was uploaded by Rob Vos on Jun 10, 2020
Content may be subject to copyright.
Estimating the Poverty Impact of COVID-19
The MIRAGRODEP and POVANA frameworks1
David Laborde, Will Martin, and Rob Vos (IFPRI)
1 Introduction
Cases of COVID-19 worldwide have grown exponentially since our previous analyses of the pan-
demic’s impacts on global staple food markets and poverty and hunger. On March 10, the number in-
fections had just passed 110,000 with about 4,000 deaths. In the three months since, the number of
people with COVID-19 has increased near 70-fold (to over 7 million) and the number of deaths more
than 100-fold (to over 400,000). The epicenter of the pandemic shifted from China to Europe and then
to the United States and Latin America. The coronavirus is now also spreading rapidly in low- and mid-
dle-income countries, many of which lack robust health systems or strong social safety nets that can
soften the pandemic’s public health and economic impacts.
More than half of the world population has been under some form of social distancing regime to contain
the health crisis. As a result, millions of businesses have had to close shop. The ILO anticipates 200
million workers could be thrown into unemployment. In the U.S. alone, virtually overnight, 40 million
people lost their jobs in April and May. Governments in Europe and the U.S. have promised unprece-
dented fiscal and monetary stimulus measures to compensate for the income losses of businesses and
workers and contain an inevitable economic crisis. But the relief responses of low- and middle-income
countries have thus far been more limited.
With COVID-19 and its economic fallout now spreading in the poorest parts of the world, many more
people will become poor and food-insecure. In a new scenario analysis, we estimate that globally, ab-
sent interventions, 148 million people could fall into extreme poverty (measured against the PPP$1.90
poverty line) in 2020—an increase of 20% from present levels. This in turn would drive up food insecu-
rity. A global health crisis could thus cause a major food crisis—unless steps are taken to provide un-
precedented economic emergency relief.
Assessing the poverty impact of COVID-19 is not trivial, however. This is so not only because the crisis
is still unfolding and available information of the precise socio-economic consequences is incomplete,
but also because the channels of influence are multiple and interconnected globally. Furthermore, while
1 The MIRAGRODEP model and the development of the POVANA database are supported by various donors, in particular the Policies, Institu-
tions and Markets (PIM) CGIAR research program.
TECHNICAL NOTE
06/09/2020
2
a number of other analyses of the poverty impacts have assumed uniform shifts in the distribution of
income per country, we are concerned that this assumption fails to take into account the complexity of
the channels of effect.
For these reasons, we use a global modelling framework to assess the potential impacts of the COVID-
19 crisis on global poverty and food security. Specifically, we combine two economic modeling frame-
works: the MIRAGRODEP global computable general equilibrium (CGE) model and the POVANA
household dataset and model. The two frameworks are linked in top-down fashion; that is, results of the
CGE model-based scenario analysis are introduced as inputs in the household survey data and model
to assess poverty outcomes. This process is summarized in Figure 1.
Figure 1 Implementation of the Covid-19 scenario
By using data on the full income distribution for over 300,000 representative households, this approach
requires no ex-ante or ad-hoc assumptions about how the economic shocks caused by COVID-19 are
changing the distribution of income in any given country. In this approach, real incomes of households
change endogenously with the simulated changes in the full vector of prices of goods, services, and
factors (including wages), and other income determinants (productivity). Changes in poverty levels are
calculated based on the thus adjusted income distribution. This approach has great advantages over
other approaches which either assume uniform proportional changes in income, as in a standard appli-
cation of the tools provided by PovcalNet,2 or through exogenous assumptions about shifts in income
inequality.
2 http://iresearch.worldbank.org/PovcalNet/ . For an application of this approach for a World bank assessment of COVID-19 impacts on global
poverty, see Mahler, Daniel Gerszon, Christoph Lakner, R.Andres Castaneda, and Haoyu Wu. 2020. “The impact of COVID-19 (Coronavirus)
on global poverty: Why Sub-Saharan Africa might be the region hardest hit.” World Bank Blog. April 20.
https://blogs.worldbank.org/opendata/impact-covid-19-coronavirus-global-poverty-why-sub-saharan-africa-might-be-region-hardest. A similar
approach is used in Sumner, A., Hoy, C. & Ortiz-Juarez, E. (2020) Estimates of the impact of COVID-19 on global poverty. WIDER Working
Paper 2020/43. UN WIDER, Helsinki.
3
This framework has been used previously to study the impact of a macroeconomic slowdown on global
poverty in Laborde and Martin (2018).3 The main differences between the current work and the previ-
ous study are twofold. First, the Laborde-Martin study looks at a change of economic growth projections
from 2015 to 2030 and was compared poverty outcomes in 2030, using the dynamic version of the
CGE and projecting household surveys until 2030. In the current exercise, we focus on 2020 scenario
results under a range of assumptions of short-term impacts of COVID-19, as explained further below.
Second, in Laborde and Martin (2018) alternative IMF projections for global growth are regenerated by
imposing commensurate changes in total factor productivity on the corresponding MIRAGRODEP pa-
rameter values. In the current exercise, no IMF growth projections are used. Instead, the factors under-
lying the socio-economic impacts of COVID-19, such as health impacts, social distancing measures,
restrictions on (labor) mobility, international transport, etc. are translated into MIRAGRODEP’s model
terms to simulate the impacts on economic growth, incomes, employment, consumption, prices, trade,
and ultimately poverty.
In this note, we spell out the core features and assumptions of the MIRAGRODEP global CGE model;
the POVANA framework for poverty analysis, and the assumptions underlying the construction of the
Covid-19 scenario.
2 The MIRAGRODEP model
MIRAGRODEP is a global Computable General Equilibrium (CGE) model based on MIRAGE (Model-
ling International Relations under Applied General Equilibrium).4 The model was developed and im-
proved with the support of The African Growth and Development Policy Modeling Consortium
(AGRODEP). It is a multi-region, multi-sector, dynamically recursive CGE model. The core
MIRAGRODEP model5 is an open-source resource distributed by the AGRODEP network,6 with train-
ing sessions held periodically to teach researchers how to use the model.
This model allows for a detailed and consistent representation of the economic and trade relations be-
tween countries. International economic linkages are captured through international trade in goods and
services, as well as through capital flows. A dynamic, recursive solution is obtained by solving the
model sequentially and moving the equilibrium from one year to another. In our study, we assume per-
fect competition in all sectors, which allows us to have a detailed geographic and sector decomposition,
but also avoids tackling the challenging issue of endogenous mark-up behavior during crisis.
In each country, a representative consumer maximizes a CES-LES (Constant Elasticity of Substitution
– Linear Expenditure System) utility function subject to an endogenous budget constraint to generate
the allocation of expenditures across goods. It is equivalent to replacing the Cobb-Douglas structure of
the Stone-Geary function (that is, LES) by a CES structure. The LES system allows for different income
elasticities of demand, with those for food typically lower than those for manufactured goods and ser-
vices. The demand system is calibrated on the income and price elasticities estimated by Muhammad
3 Laborde Debucquet, D. and Martin, W. (2018), Implications of the global growth slowdown for rural poverty. Agricultural Economics, 49: 325-
338. doi:10.1111/agec.12419
4 Decreux, Yvan and Valin, Hugo, (2007), MIRAGE, Updated Version of the Model for Trade Policy Analysis: Focus on Agriculture and Dy-
namics, Working Papers, CEPII research center.
5 Laborde, D., Robichaud, V., Tokgoz, S., 2013. MIRAGRODEP 1.0: Documentation. AGRODEP Technical Note, Washington DC, IFPRI.
6 http://www.agrodep.org/model/miragrodep-model
4
et al. (2017).7 Once total consumption of each good has been determined in the top level, the origin of
the goods consumed is determined by another CES nested structure, following the Armington assump-
tion of imperfect substitutability of imported and domestic products.
On the production side, demand for intermediate goods are determined through a Leontief production
function that assumes intermediate inputs are used in fixed proportions to output. Total value added is
determined through a CES function of unskilled labor and a composite factor of skilled labor and capi-
tal. This specification assumes a lower degree of substitutability between the last two production fac-
tors. In agriculture and mining, production also depends on land and natural resources. In the present
application of the model, we assume that new capital investment is perfectly mobile across sectors,
while installed capital is immobile. Furthermore, skilled labor is assumed to be perfectly mobile across
sectors, while unskilled labor is imperfectly mobile between agricultural sector and non-agricultural sec-
tors. Due to the, presumed, short-term nature of the COVID-19 shock, we divide the original substitu-
tion elasticity for factors of production in the production tree by a factor two, as substitution effects tend
to be smaller in the short run. Indeed, we consider very limited capacity of producers to change the
capital-labor utilization ratio within a single year.
For the present scenario analysis, we assume further that investments are savings-driven, and that the
real exchange rate is flexible, that is, it adjusts endogenously such that the current account of the bal-
ance of payments remains constant as a share of each individual country’s GDP. To guarantee the
supply of external finance matches demand in the global capital market, capital inflows towards coun-
tries with a current account deficit are “scaled” up or down by an homogenous factor to capture the
scarcity or abundance of international capital. Hence, we assume portfolio preferences and capacity to
borrow on international markets remain constant for all countries.
The treatment of the government sector in MIRAGRODEP differs from that in MIRAGE and many other
global general equilibrium models. In MIRAGRODEP, the government sector is explicitly modeled as
distinct from private agents, while this is not the case in many other global models. The income of the
government consists of taxes levied on production, factors of production, exports, imports, consump-
tion, and household income. For the present analysis, we do not consider endogenous tax policy re-
sponses and we consider that, except for those countries where we model a budgetary policy response
(such as the economic stimulus measures taken by many of the richer nations; see under scenario as-
sumptions), a reduction in tax income is actually associated with a reduction in public expenditure per
capita while maintaining the public deficit/surplus to GDP constant. The relevance of this default as-
sumption is not to confound stimulus measures with “automatic” macroeconomic stabilizers induced by
increased public deficits and which would be associated with a welfare gain by supporting current con-
sumption through an increase in public debt without considering future welfare costs of the debt.
As in Laborde and Martin (2018), we use the GTAP 9.1 database as MIRAGRODEP’s main source to
determine values of variables and parameters. This database allows us to readily use up to 140 re-
gions/countries and 65 products and production sectors. In addition, the database is enhanced by and
harmonizes additional datasets on land use, agricultural production, food balance sheets, agricultural
domestic support measures (obtained through, e.g., the Ag-Incentives Consortium, and WTO notifica-
tions) and trade policies (HS6 tariff and trade data), as well as updated Social Accounting Matrices for
all individually-specified countries. A realistic baseline is constructed aligned with the United Nations’
7 The Influence of Income and Prices on Global Dietary Patterns by Country, Age, and Gender, by Andrew Muhammad, Anna D'Souza, Birgit
Meade, Renata Micha, and Dariush Mozaffarian, ERS, March 2017
5
demographic projections and updated IMF economic growth estimates to bring the base year values
(2011) to those of the actual year of simulation (2020)
For this specific study, we use a condensed version of the model with 29 sectors, of which 18 are re-
lated to agri-food activities (primary production and downstream activity) and 36 regions/countries.8 For
a given year (2020 in the present case), the model consists of 310,345 equations and (non-zero) varia-
bles.
3 The POVANA framework
To translate the CGE model simulation results into poverty impacts, we rely on the POVANA dataset
and follow an approach similar to Ivanic and Martin (2018).9 The coverage of the POVANA dataset, and
the source are always evolving and the most recent documentation is available online.10 For the sake of
comparison using the most recent peer-review publications on the topic, we use the same version of
the POVANA dataset as in Laborde and Martin (2018), and detailed in Appendix A.1.
While the household coverage, largely composed of LSMS survey data, for 31 countries, includes more
than 300,000 representative households, we retain and use directly the information available from
household surveys on the income sources and expenditure patterns in each of just over 285,000 sam-
ple households. Our approach requires consistency between the expenditure and income information
for each household, and we make a number of adjustments to reconcile data across sources. However,
household records requiring very large adjustments for this reason are excluded.
The impacts of macroeconomic changes on the incomes of the poor are perhaps simplest when a de-
cline in productivity lowers the efficiency of production within an unincorporated or household-run busi-
nesses (such as family farms). Another relatively simple case arises from a shock that results in a size-
able change in the price of an agricultural commodity for which a poor household is a net seller or net
buyer. The short-run effect of such a shock on real household incomes can be estimated with the avail-
able information regarding the degree to which households are net sellers or net buyers of the product.
Changes in productivity and prices also affect wages, while household incomes are further influenced
by changes in remittances (and, indirectly, real exchange-rate adjustments). In short, the variables that
determine household incomes as observed in the household datasets are adjusted in line with the
changes in their matching variables in the global CGE model. This not only applies to consumer prices,
wages, and employment, but also for non-labor factor endowments (capital, land, natural resources).
Input/output ratios for household business activities are kept constant, while labor productivity for each
activity/crop is adjusted in accordance with the change generated in the CGE model scenario analysis.
Input and output prices for household businesses adjust in accordance with the domestic price changes
resulting from the CGE simulations.
To understand the poverty measures reported in this study, it is helpful to represent the income of
household i as
=(,,,)+
×+++
8 Details available at https://public.tableau.com/profile/laborde6680#!/vizhome/IFPRI_Blog_Coronavirus_LMV_032020/MainStory , select
“model nomenclature”.
9 Ivanic, M., Martin, W., 2018. Sectoral productivity growth and poverty reduction: National and global impacts. World Development. 109 (Sep-
tember):429-439. Available at https://doi.org/10.1016/j.worlddev.2017.07.004
10 https://public.tableau.com/profile/laborde6680#!/vizhome/POVANA_Surveys/POVANA
6
where represents the profit function for each business activity of an unincorporated business, de-
fined over a vector of output prices , input prices (goods and wages) , output quantities and input
requirements (hired labor and other inputs in terms of good and services) . Actually, is defined by
the production function such as =(). Other elements in the income function include the vector
of factor prices, simplified to include only two elements: unskilled wages and other factor income,
,
the quantity of these factors of production sold by the household , the net public transfers re-
ceived/paid by the household , the net international remittances received/paid by the household and
the remaining net domestic private transfers.
Similarly, expenditures is defined by the expenditure function (,)= with the vector of
consumer prices, adjusted for auto-consumption cases), and the vector of consumed quantities.
Initially, we check that =+, where represents the positive savings of household i. In this sim-
ulation, we consider that at the household level is exogenous. Indeed, we want to compute the com-
pensation measures of welfare changes at fixed current utility () and net savings, to not associate
with decreasing savings – a normal coping strategy – with positive utility outcome (increased consump-
tion). In addition, when using the extreme poverty line, households at this income level or around it,
have extremely scarce available savings and this assumption is not critical in our assessment. For
many households, (transfers) will need to adjust to balance income and expenditures. Please note
that, as explained above, first round impacts on savings are considered in the CGE model.
Because we assume in the present scenario analysis that the COVID-19 shock is short lived (the shock
and the response take place within the same year), we follow Deaton (1989)11 and consider only the
first-order impacts on and e are interested only in the direct impact of real income changes on con-
sumption, we neglect the second order effects on ().
Therefore, we measure the changes in welfare or real income. such as: =++. We
define =
as the relative change of a given variable and its value in the initial dataset, we will get,
for a given and by dropping the household i index but introducing k as an index on goods and ser-
vices:
(,,,)+
×
+++
=
+
+.
.
+.
.
+ +
.
.
+
.
.
++ + +
+
+
where represents the labor endowment of household i sold on the labor market at wage rate ,
and is a composite of other factors of production (capital, land, etc.…) owned by the household and
rented out on markets at rental rate . Importantly, for many poor households, the initial values of
many of these variables are zero in the base run, such as property taxes. We show the general specifi-
cation here.
11 Deaton, A. 1989. “Rice Prices and Income Distribution in Thailand: A Non-Parametric Analysis.” Economic Journal 99 (395): 1–37.
7
Due to the presumed short-term nature of the shock and limited coping capacities of household, we ne-
glect a number of possible second-order effects, such as sales of non-labor assets ( = 0), produc-
tion inputs (= 0) and consumption pattern (= 0). The latter is to avoid having to make specific as-
sumptions about which coping strategy household choose to mitigate reductions in their consumption
bundle. As indicated above, we further assume households do not reduce savings as a coping strategy
( = 0). In addition, we further assume governments do not adjust income tax rates as a COVID-19 re-
sponse (=0).
Adjustment of other variables in the above specifications are endogenous as determined by the
MIRAGRODEP-structural model equations. Household, firm and government behavior all vary by
country (or region). At the national level (i.e. the geographical entity represented in the model to which
the individual household belongs):
• is the relative producer price change as defined in the CGE for the good or service k, or
group of goods or services in which it is included.
•
is the relative consumer price change as defined in the CGE for the commodity k, or group of
commodities in which it is included. This price includes the import price index based on the Arm-
ington assumption (true price index of the associated good or services at the consumer level).
• For any goods or services y, for which a specific household i, is considered to produce a signifi-
cant amount for self-consumption, we consider that
is assumed to be equal to the producer
price change, instead of the consumer price change.
• Due to our focus on the left tail of the income distribution, we consider that all labor is unskilled
labor and is the relative change in the wage of unskilled workers. Since MIRAGRODEP con-
siders two labor markets, rural and urban, is specified separately for rural and urban work-
ers, depending on the household location. The location of workers remains constant in the simu-
lation.
• is the relative change of payments to non-labor endowments in the CGE (country level
weights), including land, capital, and natural resources.
• For each household, we implement a reduction in labor supplied to the market,
, identical to
the reduction of unskilled labor supply introduced in the CGE (exogenous scenario shifter) as a
consequence of lockdown and/or disease. We assume that “unsold” labor by the household,
captured by , is not recycled in the incorporated business activity and leads to additional pro-
duction (put it differently:
= 0). Similarly, the initial amount of labor used internally by the
household is not assumed to change due to confinement measures.
• is assumed to be equal to . Various assumptions have been experimented in the past and
there is no perfect solution. captures many elements, including statistical adjustment. Without
indexing would lead to a significant amount of “dark matter” in the system, stabilizing without
any justification the system. While it expected by various transfers and “rents” could be stabiliz-
ing, or not indexed on labor payments, assuming = 0 was an excessive assumption, and an
inconsistent one since from a CGE point of view, no “values” should remain fixed. Any
value/price should be indexed on at least one price in the system to avoid a violation of the
Walras law underpinning our analysis. Since is capturing other income by definition, we link
and ;
• We consider that is only impacted by the changes in labor productivity driven by the CGE
model. Indeed, one of our strong assumption is that we do not consider disruption, including for
hired labor, in the availability of inputs used by the household, so we do not consider changes in
8
. Logically, no change in does not lead to changes in . Actually, this assumption has very
limited implication for the assessment of the impacts of COVID-19 on global poverty, since
households living in extreme poverty which are self-employed and own a microenterprise or
small farm typically rely on few assets, intermediate inputs, or hired labor, and mostly rely on
the labor, administrative and management skills of their families. Still, we want to capture a
productivity effect by indexing to the relative change in output per unit of labor and by sector,
to guarantee the consistency of the framework regarding relative wage changes, productivity
changes and prices changes. Since, we use wage changes from the CGE that captures the
evolution of the marginal labor productivity in value, due to price and productivity effect, we
need to have the various elements in the framework to avoid a systematic bias about self-em-
ployment.
With these assumptions, the model estimates a new income level for each household and hence a new
income distribution for each new scenario. The per capita incomes for each household in each house-
hold survey are subsequently compared with the international (extreme) poverty line of $1.90 per per-
son, per day at 2011 purchasing power parity (PPP) (using PPP conversion factors as available at Pov-
cal.net to convert the poverty line into domestic prices). The poverty rate is calculated as the share of
the population with an income below the indicated poverty line. This calibration process allows us to de-
fine the nominal per capita income, , at base prices associated with our poverty definition. We
can also define this value as a real income poverty line at base prices. For any level of income, the
number of poor people is defined by =×()
where i the household index, N the overall
household set in the household survey for the country r , the demographic weight of this household
and ( ) a dummy variable defined on the income level indicating if the household is above or below
the poverty line such as: ()=0 >
1 . The poverty rate of the total population is equal to
=
. We compute poverty headcount and poverty rate for sub-groups of population, e.g. ur-
ban/rural, farmer/non farmer for instance = + , by changing the composition of N.
But the poverty line is not specific to any group. When doing simulations, we look at the real income
change of households and see how many households are actually crossing this poverty line in one di-
rection or another, considering a new vector (+), with as defined previously and the ini-
tial income.
Our sample of countries has wide geographic coverage and includes 65% of the world’s extreme poor.
In order to obtain estimates of poverty changes at the global level, we need to associate each country,
not included in the POVANA sample, to a weighted vector of in-sample countries. These weights (see
details in Annex A.2) are done by minimizing the quadratic distance between a vector, for the real coun-
try, defined by a set of ex-ante variables (initial level of poverty, share of rural population from the World
Development Indicators) and ex-post variables (impact on GDP and farm value added), and the same
vector for the weighted constructs. Combining ex-ante and ex-post elements is important since two
countries with similar structural features at the macroeconomic level (poverty rate, GDP per capita,
share of rural population) could be impacted differently due to various sectoral specialization, or idio-
syncratic shocks (infection rates) or policy responses (confinement).
To summarize, for a country r included in the POVANA dataset, we define the changes in the number
of poor people as =×(+)
×()
. Since, the current size of the popula-
tion is not significantly modified by the nature of the shock (low morbidity), the changes in poverty rate
is driven by the changes in the numerator, i.e. the number of poor people.
9
For a country r not included in the POVANA sample, we compute the such as:
=
×,
+
×,
with , is the weight of country j in the
linear combination used for country r, and
the initial number of poor people in the base data for
each relevant country group/countries. This approach allows to capture variations in impacts between
urban and rural poor in a consistent manner.
Finally, the POVANA data base also provides basic information about household consumption patterns.
This also allows to identify the impacts of economic shocks (like the consequences of COVID-19) on
food consumption. Income losses and food price shocks will disproportionately hurt poor people’s food
security, since they spend most of their income on food: as much as 70 percent. Rich people spend
only a small share—perhaps around 15 percent—of their incomes on food (Figure 2). The most imme-
diate threat of COVID-19 to food security arises from reductions in the incomes of poor and vulnerable
people. Some of these losses arise from income losses in agriculture, but a much larger share of these
income losses arises from disruption to non-agricultural income sources.
Figure 2 Engel’s Law: Declining food expenditure shares with rising incomes
Source: POVANA database. Authors’ computation.
Note:The blue line represents estimated share of food consumption in total expenditures estimated through a pol-
ynome of degree 3 on the log of individual income household, normalized by their own country poverty line.
4 Modelling the COVID-19 scenario
The COVID-19 represents a unique shock, potentially a single occurrence by century event, by its mag-
nitude, global scale, and diversity of impacts. Beyond the direct effects of the disease, income losses
arise from people’s very real desire to avoid catching the disease and their altruistic concerns to avoid
infecting other people, and from policy responses designed to reduce the adverse externalities associ-
ated with an unmitigated pandemic.
10
No global, economy-wide model incorporating these features is available to fully assess these potential
impacts. Many of the changes in behavior and in the functioning of economies are, as yet, poorly un-
derstood and their impacts on economic activity were still largely not fully known when we prepared this
scenario analysis. It is also difficult to rely on experience from past events, since no events like COVID-
19 have occurred on this scale in today’s globalized and technologically advanced world. Therefore, we
have had to make a number of assumptions and some ad-hoc choices about the responses of eco-
nomic agents to this unprecedented situation.
For the same reason, we do not want to use other economic models estimates as inputs since we do
not believe they are necessarily more robust than our own model to assess this unique situation. Using
our own model further allows us to keep control of all assumptions made.
In crafting the scenarios used here, we have based our choices on earlier work, such as the analysis of
Laborde, Martin and Vos in March 2020,12 where differentiated impacts on productivity and trade costs
were analyzed for a 1% global economic slowdown during 2020.
Before looking at individual assumptions, it is important to keep in mind that the model operates on an
annual timestep and the impacts of any shock are calculated as the average impact for the year. There-
fore, a disruption lasting 10 days will be associated with a 10/365 impact and a price shock, e.g. such
as the decline in oil prices, must be calibrated on the shift in annual average prices and not on the
“peak” value.
Our key assumptions for the base COVID-19 scenario can be organized around four dimensions: do-
mestic supply disruptions, international trade distortions, household behavioral responses and policy
responses.
Domestic supply disruptions
Disruptions in labor markets
• The first impact is linked to the direct impact of the disease. We use the estimates provided by
the Imperial College for each country in the world on March 26, 2020,13 in the case of the “So-
cial distancing of the whole population” scenario for all countries. Since the online excel ap-
pendix does not provide results by age cohorts, we re-estimate them by considering that for a
given country r =, and =, with ,
the population size in the age cohort c in country r, the default probability of infection by age
cohort, and the mortality rate by age cohort. The last two parameters are taken from ob-
served values in existing studies.14 and are calibrated for each country. This allows us to
recompute consistent distributions of infections and deaths by age cohorts. We then consider
the number of cases in the active population (defined over the 15 to 65 years old population) to
impact the share of lost days of work. We consider that a death occurring in March, leads to a
loss of 9/12 of annual labor supply for a given individual while sickness is associated with 15
days of lost labor supply (15/365). In this scenario, we do not differentiate cases by degree of
12 https://www.ifpri.org/spotlight/ifpri-resources-and-analyses-covid-19-also-known-coronavirus
13 See Walker and al. https://www.imperial.ac.uk/media/imperial-college/medicine/mrc-gida/2020-03-26-COVID19-Report-12.pdf
14 https://www.worldometers.info/coronavirus/coronavirus-age-sex-demographics/
11
severity with differentiated coefficients for infected/hospitalized/Intensive Care treatments cases.
The direct relative reduction in labor supply due to the disease directly is estimated as
=,×
[;], ×
,
[;]
Note that this direct effect is generally quite minor due compared to the next type of disruption.15
• Due to the confinement measures used in attempting to internalize the externalities associated
with the COVID-19 pandemic, we also allow for the fact that some willing workers become una-
ble to sell their labor. We use as a base value the “social-distancing” parameter included in the
Imperial College estimates, and assume that 12 weeks of confinement is applied in each coun-
try, except in Africa when we limit it to 8 weeks, due to the more limited ability of poor popula-
tions to manage long periods of economic disruption; the younger average age of people in the
region and the consequent more relaxed implementation of confinement policies. These as-
sumptions result in reductions in the labor supply of 23 percent in most countries or 15 percent
in Africa. We consider that 1/3 of skilled workers impacted by social distancing are able to con-
tinue working through telecommuting. This crude estimates is based on our review of the ILO
early review of Covid-19 impact on jobs (3rd edition)16 and Dingel and Neiman (2020).17 So, con-
finement measures lead to an additional reduction of
for country r and level of skill h,
{,} such as
=Social_distance×××
with = 2/3 if r in Africa,
south of Sahara, and 1 otherwise, =2/3 if = "" and 1 otherwise.
Disruptions in specific value chains
• While agriculture and food sectors have been identified as essential in most countries, we also
assume some supply disruption caused by reduced labor mobility (e.g., for seasonal migrant la-
bor) and further, that perishable farm products suffer greater post-harvest losses due to logistics
problems and demand fallout. An increase in post-harvest losses of perishable products (fruits,
vegetables, meat and dairy) by 5 points. While this estimate is conjectural, it is clear that there
are substantial losses in some cases and very small losses in others making an average loss of
5 percent seem a reasonable guesstimate in this preliminary analysis
• Total factor productivity in transportation is assumed to decline by 5 percent to capture loss of
logistical efficiency. This number is extrapolated based on anecdotal evidence ranging from
monitoring of GPS tracking devices on truck fleet in the United States (see the work of ATRI)18
to surveys in West Africa (see, e.g., the work of CILLS).19 While crude, this estimate provides at
least a reasoned estimate of the extent of disruption to transportation sectors, especially in de-
veloping countries.
• Limitations on face-to-face services result in losses of output in distribution systems and recrea-
tional activities (including restaurants) and tourism (domestic and international).. While these
15 When providing numbers per capita in our results and models, we did not correct the total population used in the counter-factual by the
number of deaths. This omission has limited implications due to limited number of deaths in total population.
16 https://www.ilo.org/wcmsp5/groups/public/@dgreports/@dcomm/documents/briefingnote/wcms_745963.pdf
17 https://bfi.uchicago.edu/wp-content/uploads/BFI_White-Paper_Dingel_Neiman_3.2020.pdf
18 American Transportation Research Institute, for instance https://truckingresearch.org/2020/05/05/joint-research-confirms-covid-19-impact-
on-trucking/
19 https://www.reuters.com/article/us-health-coronavirus-food-africa/west-african-food-trade-under-strain-as-covid-19-shuts-borders-
idUSKBN2330RV
12
outcomes are the result of both the lockdown policies discussed above and household re-
sponses (fear of contamination and altruistic desire to avoid contaminating others) discussed
further below, we introduce a shadow tax20 of about 25% for both final and intermediate con-
sumption of these services primarily to capture the first explanation. This reduces the demand
for these services, ceteris paribus, by about one-third on average. This we consider to be con-
sistent with both our social distancing assumptions.
International distortions
• To capture the effects of the “oil war” between Saudi Arabia and Russia that pre-dated COVID-
19, we calibrate an exogenous expansion of the supply of oil. The combined effects of this pro-
duction and expansion and the decline of demand linked to the COVID-19 crisis leads a net de-
cline in real energy prices (oil and natural gas) of 25% for raw products and 17% for refined
products at the global level. For comparison, the June 2019 to June 2020 price change of the
WTI Crude contract is 33% (US$53 per barrel to US$35 per barrel), but experts are expecting a
recovery especially following the OPEC response.21 Since the model combines natural gas and
crude oil into one variable, the energy price response will be more muted in our model because
of the lesser impact on the price of natural gas.
• The containment measures cause bottlenecks and delays in international freight and transport.
In terms of the model parameters, this assumption has been translated into an increase in the
average cost of international freight by 3%, not considering any feedback on energy prices. We
calibrate these numbers to capture the increased time needed to doing the business of trade
owing to logistical delays in harbors and at airports caused by new regulations, lack of inspec-
tors, and so on. The ad-valorem equivalent of lost additional days is based on Hummels and
Schaur (2013).22
Household Responses
• We assume that households and firms use part of their accumulated savings as a coping mech-
anism to compensate for reduce their fall in current incomes.23 In the CGE model, the savings
reduction is defined for each country/region iteratively by two imposed constraints: (a) to the ex-
tent they can, households try to limit their welfare loss to 5 base points relative to actual con-
sumption, but (b) households cannot cut their savings rates by more than 6 basis points and
cannot let their savings become negative. The choice of these boundaries was based on the
evolution of gross saving rates observed in the previous crisis.24 For instance, in the United
States, between 2006 and 2009, the gross saving rate fell from 19.3% to 14.0%, while the world
20 We use a shadow tax instead of a preference shifter in the model to avoid changing the utility function and therefore compromising the wel-
fare analysis.
21 We do not model the actual reduction in production taken by OPEC decided in May https://www.bloomberg.com/news/articles/2020-06-
06/opec-agrees-to-extend-output-cuts-as-quota-cheats-offer-penance. However, for our analysis a first order impact for many developing
countries, including OPEC members like Nigeria, is the overall impact on their export revenue, so the combined effects of price and quantity
and we consider that evern if prices will rebound more in the second half of the year, it will be driven by quantity reduction, and the net effects
on current account will be the same.
22 Hummels, David L., and Georg Schaur. 2013. "Time as a Trade Barrier." American Economic Review, 103 (7): 2935-59. DOI:
10.1257/aer.103.7.2935
23 In the model, gross positive firms’ savings are redistributed to the households in each period.
24 https://data.worldbank.org/indicator/NY.GNS.ICTR.ZS. The definition of gross savings in the WDI and in the model differs mainly regarding
the public savings inclusion in the former. Still, we consider that the order
13
average moved first from 23.0% to 26.7% and then declined to 23.4%. In the model, the imple-
mentation of the described rule, leads to a global saving rate reduction of 2.8 points with large
variation including countries where the adjustment reaches the 6 percent limit (e.g. Guatemala
with a 6 point reduction), countries where limited adjustment is needed (e.g. China with a 2 point
reduction) and countries with no adjustment, mainly due to lack of initial savings (e.g. most
LDCs in West Africa). Due the macroeconomic closure of the model with a balanced capital
markets leading to the value of investments to be equal, in each country, to the available net
savings (household savings + government savings + foreign savings), the drop in savings is as-
sociated with a commensurate fall in investments, which is further reflected in lower demand of
capital goods (a demand effect that is stronger than the drop in food demand, for instance).
• Households reduce demand for fresh products (such as fruits, vegetables, and fish) as they per-
ceive (erroneously or not) that such products are less safe with COVID-19. Such changes in
perceptions and related food demand shifts have been observed in a number of locations (see
e.g., Tamru et al. 2020).25 However, we do not account for such shifts in perceptions for lack of
sufficient information. However, the scenario results do endogenously account for food demand
shifts from price and income impacts of COVID-19 (see next section).
Policy Responses
• Due to their limited actual role and our close monitoring of these policy instruments, we did not
include specific export restrictions measures regarding food products (see Laborde et al.,
2020,26 for more information).
• The present scenario accounts for the economic stimulus packages being implemented by
countries in North America and in Europe, including significant income transfers to households.
For the OECD countries, except Mexico, Chile. Israel and Turkey, we implement a stimulus
package of 3.2% of each country GDP in average. It is done by an increased transfer/reduction
in income tax from the government to the representative household in each region.
• While remittance flows are impacted through falls in wage income in the remitting countries, we
do not consider any “relocation” of foreign workers in the present COVID-19 scenario analysis.
The existing stock of foreign workers in each country has been kept unchanged.
5 Some basic scenario results
Under the assumptions as spelled out in the preceding sections, we project a downturn in global eco-
nomic growth of 5% in 2020 (see Table 1). This projection is broadly similar to the recent IMF forecast,
which shows a downturn of the world economy from the 2%-3% growth anticipated pre-pandemic to an
actual decline of 3%. The scenario further indicates that the poorest nations face significantly greater
adversity. The recession under way in Europe and the U.S. is projected to depress economic activity
across developed countries by 6% on average in 2020, despite an expected rebound later in the year
as social distancing measures are relaxed and stimulus measures take effect. This recession will spill
over to the rest of the world through lower demand for trade and lower commodity prices. Developing
25 Tamru, Seneshaw, Hirvonen, Kalle, and Minten, Bart. 2020.Impacts of the COVID-19 crisis on vegetable value chains in Ethiopia, IFPRI
blog, available at:https://www.ifpri.org/blog/impacts-covid-19-crisis-vegetable-value-chains-ethiopia
26 Laborde Debucquet, David; Mamun, Abdullah; and Parent, Marie. 2020. Documentation for the COVID-19 food trade policy tracker: Track-
ing government responses affecting global food markets during the COVID-19 crisis. COVID-19 Food Trade Policy Tracker Working Paper 1.
Washington, DC: International Food Policy Research Institute (IFPRI). https://doi.org/10.2499/p15738coll2.133711
14
economies will be hurt by the economic fallout caused by their own social distancing measures and by
increased morbidity affecting the labor supply for farming and other business activity.
For developing countries as a group, the economic fallout would lead to a decline of aggregate GDP of
3.6%, but economies in Africa south of the Sahara, Southeast Asia and Latin America would be hit
much harder due to their relatively high dependence on trade and primary commodity exports. The re-
cession is expected to be less severe in China and the rest of East Asia, where we expect the eco-
nomic recovery to start sooner with the earlier lifting of containment measures.
We expect economies in Africa to be hit hardest (almost a 9% decline relative to our baseline projec-
tion). But agri-food sectors may be spared and expand, as the collapse in export earnings and reduced
ability to import food push up domestic production. Lower labor demand in urban service sectors may
push workers to return to agriculture, also contributing to greater domestic food production. With more
workers in the sector, however, individual incomes would remain low.
As shown in Table 1, our adjustment rule regarding consumption smoothing limits the impact of the re-
duction in GDP on household consumption, since savings and investments absorbs part of the shock.
Also, we have more limited impact on consumption since GDP is being measured in real terms through
a Fisher index, while consumption is tracked through a welfare metric (equivalent variation) capturing
the mitigating effects of the household response to fall in prices. These price considerations should al-
ready be kept in mind when looking at the column with the results for exports (measured as changes in
values at constant international dollars). Therefore, both prices and volume effects are responsible for
the collapse of world trade of goods and services by about 18%.
Without social and economic mitigation measures such as fiscal stimulus and expansion of social
safety nets in the global South (scenario assumption), the impact on poverty would be devastating as
shown in Figure 3. In addition to the 20% global increase in extreme poverty noted above, the scenario
indicates urban and rural populations in Africa south of the Sahara would suffer most, as 80 million
more people would join the ranks of the poor, a 23% increase. While most poor people is located in ru-
ral areas, we estimate that, in relative terms, the increase in poverty will be bigger i in urban areas. For
instance, we estimate that, in Sub-Saharan Africa, the number of poor people will increase by 15% in
rural areas, but by 44% in urban areas. Detailed poverty numbers at the country and regional level are
available online.27
27 https://public.tableau.com/profile/laborde6680#!/vizhome/EconomicImpactEconomicslowdownduetoCovid-19IFPRIb-
logApril16th2020/Blog04162020_IFPRI
15
Table 1 Macroeconomic impacts of MIRAGRODEP-COVID 19 scenario by regions and
countries
(Percentage change from baseline values)
Real Consumption
Real GDP
Exports
World
-1.0
-5.1
-17.9
Developed Countries
-0.2
-6.2
-19.3
Developing Countries
-2.5
-3.6
-16.0
Africa
-3.5
-7.3
-28.5
Africa South of Sahara
-3.2
-8.9
-31.2
North Africa
-4.0
-6.4
-28.6
Asia (ex. Central Asia)
-3.9
-4.6
-23.3
Central Asia (excl. Russia)
-4.1
-9.9
-21.6
Latin America
-4.4
-5.9
-27.5
Central America
-6.2
-8.7
-20.2
South Asia
-3.7
-5.0
-22.9
South East Asia
-4.2
-7.0
-23.9
Bangladesh
-3.0
-7.2
-26.4
Cameroon
-4.1
-7.4
-17.8
China
-4.2
-4.5
-21.8
Ecuador
-3.4
-7.6
-40.5
Ethiopia
-1.8
-2.9
-18.2
Guatemala
-3.8
-5.9
-26.7
Indonesia
-3.5
-6.5
-33.1
India
-3.9
-5.9
-21.8
Kenya
-3.0
-4.6
-25.4
Cambodia
-10.2
-8.5
-21.8
Sri Lanka
-2.0
-3.8
-29.4
Malawi
-4.7
-5.7
-26.2
Nigeria
-2.7
-4.8
-37.7
Nicaragua
-3.5
-6.4
-20.8
Nepal
-3.7
-6.6
-36.3
Pakistan
-3.2
-5.2
-30.8
Panama
-13.2
-9.7
-19.7
Peru
-3.5
-8.0
-20.7
Rwanda
-2.9
-4.2
-29.1
Tanzania
-2.5
-2.8
-14.6
Uganda
-3.3
-2.1
-25.6
Viet Nam
-7.0
-9.8
-30.9
Yemen
-6.5
-8.5
-32.9
Zambia
-2.0
1.2
-12.5
South Africa
-2.2
-2.4
-19.2
Rest of West Africa
-7.4
-5.0
-22.7
Rest of Africa, South of Sahara
-2.9
-4.8
-31.7
Rest of LAC
-4.4
-5.7
-27.5
Source: MIRAGRODEP Simulation
Note: Regions in bold aggregated results computed post simulations, weighted by the relevant country level variable. Detailed
for rich countries are omitted. Real consumption is limited to household private consumption and defined as the equivalent
variation (welfare) metrics. Real GDP is computed following national accounting principles. Fisher price indices between base
prices and simulation prices are used. Exports of goods and services are measured FOB, in value, at constant international
dollars but final export prices.
16
Figure 3 Global and Regional Poverty Impacts of MIRAGRODEP-COVID 19 scenario by
slected regions
(Absolute and percentage change from baseline values)
Source: MIRAGRODEP and POVANA Simulations
The number of poor people in South Asia would increase by 15% or 42 million people. As these esti-
mates refer to the extreme poor, i.e., those who typically lack sufficient means to buy enough food, we
expect a commensurate rise in the number of food-insecure people.
These income and price impacts leads to significant changes in diets as shown in Figure 3, capturing
both the macroeconomic impacts (income, real exchange rate) and the value chain specific drivers
(perishability, labor intensity) mechanisms. We see a shift from high value products, including fruits and
vegetables, to staples to maintain calories intake. Still, these global results should not mask some level
of heterogeneity across countries, as shown in Table 1. Indeed, depending on the country, and the
magnitude of the recession (income loss) and the origin of the staples (e.g. imported or locally pro-
duced), the changes in their affordability vary substantially. See, for instance, the differentiated re-
sponse for wheat between developed and developing countries, where this product is mainly imported
by the latter group.
20%
15%
23%
15% 15% 14%
0%
5%
10%
15%
20%
25%
0
20
40
60
80
100
120
140
160
Total
Population Rural
Population Total
Population Rural
Population Total
Population Rural
Population
World Africa South of Sahara South Asia
Relative Increase in the number of poor
people
Additional Poor people ($1.90 poverty ine)
Increase in number of poor people ($1.90 poverty line)
Relative increase in the number of poor
17
Figure 4 COVID-19 impacts on diets (average effect for world)
(Percentage change in average global household consumption by product)
Source: MIRAGRODEP Simulation
Note: Global average based on weighted changes at the estimated at the country or regional levels. Weights are based on
base value of consumption, while changes are computed on the evolution of the volume of consumption for each national rep-
resentative household.
Figure 5 COVID-19 impacts on diets (average effect for world)
(Percentage change in average global household consumption by product)
Source: MIRAGRODEP Simulation
Note: Income group average based on weighted changes at the estimated at the country or regional levels. Weights are based
on base value of consumption, while changes are computed on the evolution of the volume of consumption for each national
representative household.
-8.54
2.81
-5.89
-0.08 -1.14
5.1
-1.34
5.16
-4.91 -4.87 -6.22
-10
-8
-6
-4
-2
0
2
4
6
u
Food commodities (excl. processed food)
-50
-40
-30
-20
-10
0
10
Changes in consumption, volume , Percentage
Food commodities
Bangladesh Nigeria Developed Countries Developing Countries
18
Annexes
A.1 Household data inventory
Table A.1 Household surveys used in this study
Country name
Year
Survey name
Albania
2005
Living Standards Measurement Survey
Armenia
2004
Integrated Survey of Living Standards
Bangladesh
2005
Household Income-Expenditure Survey
Belize
2009
Household Income and Expenditure Survey
Cambodia
2003
Household Socio-economic Survey
China
2002
Chinese Household Income Project
Côte d'Ivoire
2002
Enquete Niveau de Vie des Menages
Ecuador
2006
Encuesta Condiciones de vida
Guatemala
2006
Encuesta Nacional de Condiciones de Vida
India
2005
India Human Development Survey (IHDS)
Indonesia
2007
Indonesia Family Life Survey
Malawi
2004
Second Integrated Household Survey
Moldova
2009
Cercetarea Bugetelor de Familie
Mongolia
2002
Household Income and Expenditure Survey
Nepal
2002
Nepal Living Standards Survey II
Nicaragua
2005
Encuesta Nacional de Hogares sore Medicion de Nivel de
Vida
Niger
2007
Enquete National sur Le Budget et la Consommation des
Menages
Nigeria
2003
Nigeria Living Standards Survey
Pakistan
2005
Pakistan Social and Living Standards Measurement Survey
Panama
2003
Encuesta de Niveles de Vida
Peru
2007
Encuesta Nacional de Hogares
Rwanda
2005
Integrated Household Living Conditions Survey
Sierra Leone
2011
Sierra Leone Integrated Household Survey
Sri Lanka
2007
Household Income and Expenditure Survey
Tajikistan
2007
Living Standards Measurement Survey
Tanzania
2008
National Panel Survey
Timor-Leste
2007
Poverty Assessment Project
Uganda
2005
Socio-Economic Survey
Viet Nam
2010
Household Living Standard Survey
Yemen
2006
Household Budget Survey
Zambia
2010
Living Conditions Monitoring Survey
Source: MIRAGRODEP model database.
19
A.2 Weighting procedure for countries not included in the household database
The distance that minimizes a weighted squared difference between country’s i specific variables and
those of a group of reference countries is defined as:
,=.,
,,
,
(
,)/()
+.,,,,
,
()
,
s.t.
,,
,
,,
,
,,
=
and
,
where I is the set of every country in the world in our global household dataset; K is the same set as I; J
is the set of 31 countries included in our global household dataset28 and j an element of J; ,, is the
weight of country j in the linear combination used for country i; ,
is the relative change in
the total value of agricultural production of country i in a given scenario s. S being the scenario space.
This change is computed by combining FAOSTAT data on individual crop production and prices for
each country and the quantity and price changes obtained in the CGE, either for the country, if singled
out in the model, or the region to which the country belongs. This top-down approach allows us to cap-
ture, at the country level, for each country of the world, one of the key drivers, i.e. farm income, of the
results for a given scenario. Indeed, we need to rely on this approach, capturing ex-post elements,
since the price changes driven by the productivity changes are region, production, and scenario specific
and focusing on ex-ante clustering analysis would miss this point.
, stands for a set of country-level variables that are used to bring together similar coun-
tries. The set F includes the following variables for 2013, extracted from the World Development Indica-
tors database: GDP per capita in PPP (2011), poverty rate (measured against the PPP$1.90 pp/pd pov-
erty line), prevalence of undernourishment, share of agriculture in total GDP, and share of rural popula-
tion.
28 Card(K) is the cardinal of the set K, and therefore card(K)=211. For missing countries in FAOSTAT, like the Democratic Republic of Congo,
we use a proxy country, in this case, the Central Africa Republic.
