Initial Impacts of the Pandemic on Consumer Behavior: Evidence

NBER WORKING PAPER SERIES

INITIAL IMPACTS OF THE PANDEMIC ON CONSUMER BEHAVIOR:

EVIDENCE FROM LINKED INCOME, SPENDING, AND SAVINGS DATA

Natalie Bachas

Peter Ganong

Pascal J. Noel

Joseph S. Vavra

Arlene Wong

Diana Farrell

Fiona E. Greig

Working Paper 27617

http://www.nber.org/papers/w27617

NATIONAL BUREAU OF ECONOMIC RESEARCH

1050 Massachusetts Avenue

Cambridge, MA 02138

July 2020

This paper was prepared for the Brookings Papers on Economic Activity Conference on June 25,

2020. Authors were paid an honorarium by Brookings for the preparation of this research with an

amount under $5,000 per author. We thank Therese Bonomo, Peter Robertson, and Tanya

Sonthalia for their outstanding analytical contributions to the report. We thank Jonathan Parker,

Jan Eberly, and Erik Hurst for helpful discussions. We are additionally grateful to Samantha

Anderson, Maxwell Liebeskind, Robert McDowall, Shantanu Banerjee, Melissa Obrien, Erica

Deadman, Sruthi Rao, Anna Garnitz, Jesse Edgerton, Michael Feroli, Daniel Silver, Joseph

Lupton, Chris Knouss, Preeti Vaidya, and other members of the JP Morgan Chase Institute for

their support, contributions, and insights. The views expressed herein are those of the authors and

do not necessarily reflect the views of the National Bureau of Economic Research.

At least one co-author has disclosed a financial relationship of potential relevance for this

research. Further information is available online at http://www.nber.org/papers/w27617.ack

NBER working papers are circulated for discussion and comment purposes. They have not been

peer-reviewed or been subject to the review by the NBER Board of Directors that accompanies

official NBER publications.

paragraphs, may be quoted without explicit permission provided that full credit, including ©

notice, is given to the source.

Initial Impacts of the Pandemic on Consumer Behavior: Evidence from Linked Income, Spending,

and Savings Data

Natalie Bachas, Peter Ganong, Pascal J. Noel, Joseph S. Vavra, Arlene Wong, Diana Farrell, and

Fiona E. Greig

NBER Working Paper No. 27617

July 2020

JEL No. E21,E6,E62,H31

ABSTRACT

We use U.S. household-level bank account data to investigate the heterogeneous effects of the

pandemic on spending and savings. Households across the income distribution all cut spending

from March to early April. Since mid April, spending has rebounded most rapidly for low-income

households. We find large increases in liquid asset balances for households throughout the income

distribution. However, lower-income households contribute disproportionately to the aggregate

increase in balances, relative to their pre-pandemic shares. Taken together, our results suggest that

spending declines in the initial months of the recession were primarily caused by direct effects of

the pandemic, rather than resulting from labor market disruptions. The sizable growth in liquid

assets we observe for low-income households suggests that stimulus and insurance programs

during this period likely played an important role in limiting the effects of labor market disruptions

on spending.

Natalie Bachas

Princeton University

[email protected]

Peter Ganong

Harris School of Public Policy

University of Chicago

1307 East 60th Street Chicago,

IL 60637

and NBER

[email protected]

Pascal J. Noel

University of Chicago

Booth School of Business 5807

South Woodlawn Avenue

Chicago, IL 60637

[email protected]

Joseph S. Vavra

Booth School of Business

University of Chicago

5807 South Woodlawn Avenue

Chicago, IL 60637

and NBER

[email protected]

Arlene Wong

Department of Economics

Princeton University

192A Julis Romo Rabinowitz Building

Princeton, NJ 08544

and NBER

[email protected]

Diana Farrell

JPMorgan Chase Institute

601 Pennsylvania Avenue, NW

Washington, DC 20004

[email protected]

Fiona E. Greig

JP Morgan Chase Institute

601 Pennsylvania Avenue

Suite 07

Washington, DC 20004

[email protected]

Introduction

The Covid-19 pandemic led to a large and immediate decline in U.S. aggregate spending and an

increase in aggregate private savings. In this paper, we use anonymized bank account information

on millions of Chase customers to measure the microeconomic dynamics underlying these aggre-

gate patterns. Speciﬁcally, we use our household level account data to explore how spending

and savings over the initial months of the pandemic vary with household-speciﬁc demographic

characteristics, such as pre-pandemic income and industry of employment.

Figure 1: Aggregate Consumption and Savings

−20

−10

Percent

2019m4 2019m7 2019m10 2020m1 2020m4

Personal Consumption Expenditures Y−o−Y Growth

Personal Savings Rate Y−o−Y Change

Figure shows the year-over-year growth of Personal Consumption Expenditures and change in the Personal Savings

rate, calculated from monthly Bureau of Economic Analysis data.

Measuring and understanding the link between income, spending, and savings is useful for

understanding the causes and dynamics of this recession. For instance, the relationship between

individual income, spending and savings can shed light on the role of supply factors (such as shut-

downs and reducing activities with high infection risk) versus demand factors (such as Keynesian

spill-overs across sectors as unemployed workers reduce spending). Understanding these factors

can be informative about the effectiveness of different stimulus policies for targeting different

households and businesses. Many data sets have already been used to study the dynamics of

geographic level spending during the pandemic, but aggregated relationships may or may not

be identical to those at the individual household level at which economic behavior is ultimately

determined.

Our paper provides an initial step in analyzing these household-level dynamics.

Focusing ﬁrst on aggregate results, we ﬁnd that overall spending fell over 35% in the second

See e.g. Chetty et al. (2020).

To be clear, our current analysis does not run regressions at the household level, but it does crucially rely on

individual household data to deﬁne groups and outcomes of interest. We also focus for now on sorting households

by pre-pandemic characteristics like income level, rather than by changes during the pandemic.

half of March. In April, spending began to increase from its nadir, but it remained substantially de-

pressed through the end of our sample on May 30. Declines in non-essential spending accounted

for most of the declines in spending.

Amongst non-essential categories, declines were particu-

larly large for restaurants, hotel accommodations, and clothing and department stores. Amongst

essential categories, declines were most dramatic for healthcare, ground transportation and fuel.

Reassuringly, these patterns are similar to those found using other aggregate sources of spending

data.

, EarnestResearch (2020), Baker et al. (2020), Chetty et al. (accessed on 06/15/2020), Facteus

(accessed on 06/15/2020) and Karger and Rajan (2020). However, this also implies that these re-

sults do not rely on the unique features of our micro data, so they are not the main contribution

of our paper.

We next turn to results which do rely on our micro data linking household-level observables

on income, spending and savings. First, we ﬁnd that during the initial stages of the pandemic in

March, there are extremely large declines in spending for all quartiles of the pre-pandemic income

distribution.

Spending by the top quartile of the income distribution falls by modestly more than

any other quartile (in percentage terms). However, this difference is small relative to the broad

decline in spending by all income groups. Beginning in mid April when aggregate spending be-

gins to recover, substantial differences by income emerge: spending recovers much more rapidly

for low-income households than for high-income households so that large differences arise by the

end of May. We show that these relationships between income and spending over the pandemic

hold both in general, as well as within narrow geographic areas like zip codes.

Second, we explore differences in spending by individual’s industry of employment. This

variation is interesting because industries vary substantially in both their exposure to labor market

disruptions and in average income levels. Exploiting joint variation in industry of employment

and household income is thus helpful for better understanding the source of heterogeneity in

spending patterns. We ﬁnd that spending cuts are pervasive, with declines for workers in all

industries of employment. Consistent with the patterns we ﬁnd by income, workers in industries

with low average pay initially cut spending slightly less and then have spending which recover

more rapidly. For example, grocery store workers have the smallest declines in spending and

the most rapid rebound, while white collar professional workers’ spending is recovering more

We deﬁne these categories precisely later, but loosely speaking non-essential stores are those which are subject to

government restrictions as a result of the pandemic.

See for example

As we discuss more in Section I, since our data arises from bank account information, we under sample the very

lowest income households, but the sample is otherwise broadly representative.

High and low income people live in different locations, which might have different exposure to the pandemic.

Using within zip code variation shows that income-spend relationships are not driven by confounding effects of

physical location.

slowly. We then further split workers within given industries of employment by their individual

pre-pandemic income levels. We ﬁnd that income appears to matter more for spending than

industry of employment. For example, low-income workers in all industries have rapid increases

in spending in mid April, while these increases are muted for high-income workers.

Finally, we turn to evidence on the distribution of household savings over the pandemic to

provide further insight into the effects of changing income and spending on household liquid-

ity. To the best of our knowledge, we are the ﬁrst paper to explore these distributional effects.

Aggregate savings have increased substantially over the last two months. Information on the un-

derlying distribution of increases is useful for understanding the sources and consequences of this

increase. There are several forces during the pandemic that likely affected aggregate savings rates:

1) As discussed above, spending has fallen. This decline is most dramatic at the top of the income

distribution, which will tend to boost savings for these households. 2) Massive increases in unem-

ployment have reduced labor income, and these effects are especially concentrated on low-income

workers. This will tend to reduce savings for low-income households. 3) Stimulus and social in-

surance programs like Economic Impact Payments (EIP) and expanded unemployment insurance

(UI) provide transfers which represent a larger share of income for low than high-income house-

holds. This will tend to increase liquidity and savings by low-income households. 4) Delayed tax

ﬁling dates may increase short-term savings if those who owe money delay ﬁling more often than

those who are owed refunds.

Consistent with aggregate savings data, we ﬁnd a large initial increase in savings during the

pandemic. By the end of May 2020, average liquid balances are 36% higher than at the same point

in 2019. While increases in liquid balances are pervasive throughout the income distribution,

we ﬁnd that lower-income households contribute disproportionately to the aggregate increase

in balances, relative to their initial pre-pandemic shares. That is, liquid balances at the end of

May are slightly more equally distributed over the income distribution than liquid balances in

February. However, in dollar terms, high-income households contribute most to the aggregate

increase in savings.

Taken together, our results suggest several conclusions. First, labor market disruptions were

unlikely to be a primary factor driving initial spending declines during the recession. Overall

declines in spending were much larger than what could be explained by the rise in unemployment

in this recession, given historical relationships. Furthermore, spending actually declines by less

for households with greater exposure to labor market disruptions. This does not mean that labor

market disruptions have no effects on spending or that demand spillovers are unimportant, but it

does suggest that at least in these initial months of the recession, the direct effects of the pandemic

are the primary factor driving spending.

Second, the composition of typical spending is important for understanding spending de-

clines. Aggregate spending declines by more in non-essential sectors which are more exposed to

shutdowns and health risk. Furthermore, spending declines more for high-income households,

who tend to consume more of these non-essential goods in normal times.

Third, various stimulus and social insurance programs like EIP and expanded UI likely played

a sizable role in helping to stabilize spending and liquid balances, especially for low-income

households. Since ﬁscal stimulus was ramped up at the same time that many states began to

reopen, it is difﬁcult to disentangle general "re-opening" effects from effects of this ﬁscal stimulus

by looking just at aggregate spending. However, stimulus checks and expanded UI beneﬁts rep-

resent a larger share of monthly income for low-income workers than for high-income workers,

and would thus naturally explain the more rapid recovery in spending we observe for low-income

workers. Finally, expanded transfers could also explain the disproportionate increase in savings

that we observe for lower-income households. It is important to note that many of these transfer

programs are likely temporary – the EIP payments are a one-off stimulus, while the expanded

component of UI beneﬁts is slated to end in late July 2020. Households may be less likely to

immediately consume, and more likely to save, these payments because they are non-permanent.

It is important to emphasize that our evidence for now focuses on time-series patterns for

relatively aggregated household groups, and so we do not provide any causal evidence on the

strength of any particular channels driving spending decisions. Thus, our evidence is suggestive

rather than conclusive on this front. The early patterns we ﬁnd in this paper may also change as

the pandemic progresses and new policy decisions are made. Future work exploring even more

detailed household level results as this recession progresses will hopefully shed further light on

the economic consequences of this pandemic and associated policy responses.

I Data Description

Our analysis of spending and checking account balances is based on the universe of transactions

from Chase checking accounts, debit cards, and credit cards through May 30, 2020. Our main mea-

sure of total spending includes all debit and credit card purchases as well as cash withdrawals.

In robustness checks in the Appendix we show that our conclusions are similar if we add paper

checks to our measure of total spending.

While we observe credit, debit, cash, and check trans-

actions, we are still working to process electronic checking account transactions such as ACH

payments and so this type of spending is not included in our analysis. For all checking accounts,

We do not include paper checks in our main analysis for two reasons. First, we do not know whether the checks

reﬂect spending, debt payments, or transfers. Second, due to delays in depositing and processing checks, there is a

lag between when the check was used and when it appears as a withdrawal in the bank account. Hence, it is hard to

interpret the patterns of paper check outﬂows at the high frequency we use in this analysis.

Table 1: Income Distribution and Credit Cards Spending

Quartile cut-offs Mean income

% Sample with Credit

Card

Avg. Weekly Credit Card

Spend

Quartile 1 $12,000-$27,707 $20,948 30% $205

Quartile 2 $27,707-$41,255 $34,185 36% $228

Quartile 3 $41.255-$63,462 $50,927 46% $329

Quartile 4 $63,462 + $108,914 57% $639

N 5,014,672

Initial Balances

Income Quartiles

∗

The last two columns provide the percentage and average spending of any individuals with a Chase credit card.

we also observe checking account balances.

We impose income and activity screens in order to focus on a sample of individuals who

primarily use their Chase account to manage their ﬁnances. Speciﬁcally, we ﬁlter on those who

have a non-business account, had at least ﬁve checking account transactions and at least three

card transactions in every month between January 2018 and March 2020, and had at least $12,000

in labor income in both 2018 and 2019.

This leaves us with a sample of just over ﬁve million

individuals.

We measure labor income using information on payroll direct deposits. We further measure

industry of employment based on the payer associated with direct deposits in February 2020.

However, there is an important caveat that we can match the payer associated with payroll income

to an identiﬁed payer for only 24% of households, and most of these payers tend to be large

employers. Finally, it is important to note that while we observe labor income through February,

2020, we are still working to process and interpret data on labor income and government transfers

during the pandemic. As a result, data on income changes during the pandemic are not available

for our current analysis. For this reason, we report various results based on pre-pandemic income,

but do not yet have results on how spending has responded to individual income changes.

Given that our sample is drawn from account holders at a single ﬁnancial institution, we use

income data from the CPS to measure how representative it is of the US population. Table 1

reports quartiles of the labor income distribution for our sample. Figure A.12 plots the average

labor income by quartile for the Chase sample compared to average labor income for the CPS

population (adjusted for income and payroll taxes since the Chase measure is post-tax). This

ﬁgure suggests that our sample is broadly representative, although it somewhat overstates income

at the lowest end of the distribution and slightly understates income at the highest end of the

distribution. The overstatement at the lowest end of the distribution is due to two factors. First,

We have explored different thresholds on transactions and results are similar.

reliable measurement requires us to impose a minimum threshold of $12,000 in labor income.

In the CPS, 7.7% of households have labor income below this cutoff. They would be excluded

from our analysis.

Second, every household in our data set has a bank account. Therefore,

we do not include unbanked households, who are disproportionately low-income. The FDIC

reports that 6.5% of U.S. households did not have a bank account in 2017.

A ﬁnal caveat for our

analysis is that we report average outcomes in terms of spending and liquid balances by income

quartile. However, there may be heterogeneity within quartiles. For example, not all households

were eligible for EIP payments and some unemployed households faced long delays in receiving

UI payments. For all these reasons, our ﬁndings that average spending and average balances

are relatively higher for low-income households should not be interpreted to mean that all low-

income households are doing relatively well during the pandemic. There is compelling evidence

that this is not the case (Bitler, Hoynes, and Schanzenbach, (2020).

Our data are unique in their size, sample coverage, and in their individual-level view of in-

come, spending, and balances. Other data sources used to research the consumer response to

COVID-19 tend to be aggregated over region, store, or time (e.g. Earnest, Womply, or Afﬁnity),

which limits the analysis of household balance sheet dynamics. By observing covariates at the

individual level, like geography and industry of employment, we can also directly control for

confounding factors that might be correlated with income and changes in spending. For exam-

ple, our data can be used to look at how spending varies with income within narrow geographic

areas like zip codes, and thus help control for the fact that high-income locations differ from low-

income locations along a number of dimensions. Our data also allows us to look at how spending

changes by income groups within industry of employment.

Our sample, which captures households across the income distribution, complements the

work done using Facteus data (e.g. Karger and Rajan (2020), Alexander and Karger (2020)) and

proprietary Fintech data (Baker et al. (2020)) which is primarily focused on low-income house-

holds. Finally, the size of the Chase customer base allows for additional precision when calculat-

ing statistics of interest as well as for substantially more disaggregated data cuts, relative to data

sets with smaller sample sizes. Our data is closest in structure to that in Andersen et al. (2020),

which uses similar bank account data from a Scandinavian bank. The most important distinction

is that our data covers U.S. households, and thus a dramatically different institutional environ-

ment with different social safety nets and government responses to the pandemic.

We require at least $12,000 in labor income since it is difﬁcult to distinguish truly low income households from

mis-measured higher income households without reliably captured direct deposits.

After conditioning on households with labor income above $12,000 in the CPS, mean and median income in the

bottom quartile of our sample is very similar to mean and median income in the bottom quartile of the CPS.

See 2017 FDIC National Survey of Unbanked and Underbanked Households, Executive Summary.

II Household Spending

II.A. Overall Change In Spending

We begin by measuring the change in total spending. Appendix A.1 provides changes in spending

for each of the components of total spending (credit card, debit card, and cash), as well as paper

checks. Panel (a) of Figure 2 plots the 2020 to 2019 year-over-year percentage change in weekly

spending. Panel (b) shows the average dollar amount of spending in 2020 and 2019. Changes in

spending follow a distinctive pattern: spending is stable through the beginning of March, then

declines precipitously by over 35 percent relative to 2019 from the second through fourth week of

March. The size of the spending drop is largely consistent with other estimates from similar data

sources during the same time frame.

These declines are somewhat larger than the aggregate

spending declines in Figure 1, but this is not particularly surprising. Personal consumption ex-

penditures includes substantial spending on components like housing services, which likely had

little to no decline. Spending showed signs of recovery in May, but remains roughly 15 percent

below pre-pandemic levels as at the end of May.

Figure 2: Average Spending Changes

(a) Percent Change (b) Levels

The left panel plots the year-over-year percentage change in weekly spending. The right panel plots the dollar change

in weekly spending. Vertical lines show the date when the national emergency was declared and the date on which

a majority of EIP payments arrive.

The timing of the initial spending drop mirrors the spread of the virus and staggered national

See for example, Baker et al. (2020), Chetty et al. (accessed on 06/15/2020), and Facteus (accessed on 06/15/2020).

implementation of government social distancing orders. A national emergency was declared on

March 13, 2020. Over the following three weeks, the number of states with stay-at-home orders

increased from zero to forty-ﬁve. The prevalence of COVID-19 also increased dramatically over

the course of March.

At the same time, the drop in spending also closely tracks the pattern of initial job losses.

Unemployment Insurance (UI) claims began spiking in the third week of March, with more than

20 million UI claims ﬁled by April 11. Conversely, spending begins to recover in the weeks after

April 15 when a majority of EIP payments arrive and as many of the unemployed workers who

ﬁle claims in March and early April begin to receive beneﬁts. This raises a question of how much

of the drop in spending is due to the pandemic itself, the social distancing policies, or income

losses.

It is useful to calibrate the size of the spending drop relative to what we have observed among

those who lose a job involuntarily during normal times. Ganong and Noel (2019) measure the

spending drop around job loss among UI recipients, and observe an initial spending drop of

roughly 6%. In other words, the spending drop in March 2020 is roughly six times larger than

the average household spending drop in the ﬁrst month of unemployment for UI recipients in

normal times. This puts into perspective how dramatic the spending drop is and suggests that

the pandemic and policies aimed at preventing its spread are contributing substantially to the

drop in spending.

II.B. Change In Household Spending By Categories

While Figure 2 shows a sharp drop in aggregate spending over March and April, there is reason

to think that speciﬁc spending categories would be differentially impacted. Many non-essential

businesses, like bars and salons, were closed by state and local governments. Similarly, stay-at-

home orders limited the ability of individuals to travel. Beyond the mechanical effect of social

distancing regulations, individuals may also have independently curtailed spend in certain cate-

gories to avoid risk of infection or as a response to income loss.

While we do not have information on debit card or cash spending by categories, we do have

detailed category splits for credit card spending. We begin by disaggregating total credit card

spending into essential and non-essential categories, as commonly deﬁned in state "stay-at-home"

orders. Figure 3, panels a and b, show a dramatic difference in the path of essential and non-

essential spending. Essential spending spiked in early March as households stockpiled goods

like groceries. It then fell substantially before eventually stabilizing at a year-over-year decline

of around 15 percent.

In contrast, spending on non-essential categories fell sharply throughout

The downward spike in year-over-year essential spending in the week ending April 18, 2020 likely arises because

March, bottoming at a decline of just over 50 percent, and then began to slowly recover through

late April and May.

Figure 3: Credit Card Spending on "Essential" and "Non-Essential" Categories

(a) Percent Change (b) Levels

We use state social distancing orders that restricted non-essential goods and services to categorize spend. "Essential"

categories include fuel, transit, cash, drug stores, discount stores, auto repair, groceries, telecom, utilities, insurance,

and healthcare. "Non-essential" includes department stores, other retail, restaurants, entertainment, retail durables,

home improvement, professional and personal services, and miscellaneous. Although ﬂights, hotels, and rental cars

are sometimes categorized as "essential" and not technically closed, we include them in the "non-essential" group

because they are affected by stay-at-home restrictions on non-essential travel.

Given the fact that households were ordered to stay at home except to make essential trips

in most states, one might ask why households were still spending roughly $50 a week on non-

essential categories in April. First, there was variation in both the degree of closures and in

what was deemed non-essential across locations. Second, our spending categories do not map

perfectly to each speciﬁc non-essential category. Third, households may be able to switch some

non-essential services from in-person to remote; for example from movie theater entertainment to

online streaming or from in-restaurant dining to take-out.

We also quantify how much each category contributed to the aggregate drop in credit card

spend. Table 2 shows what share of aggregate spending went towards essential and non-essential

categories before and during the pandemic. Multiplying the pre-pandemic shares by their relative

percentage drops, we ﬁnd that non-essential spending accounted for 84% of the aggregate decline,

and essential spending accounted for 16%.

of the timing of Easter, which occurred during this week in 2020 but during the previous week in 2019. Many grocery

stores are closed on Easter, which may explain a dip during this week in 2020 relative to the same week in 2019, which

Table 2: Credit Card Spending Changes for "Essential" and "Non-Essential" Categories

Share of

spending

Year-over-year

percent change

Share of

spending

Year-over-year

percent change

Apr-19 35% 65%

Apr-20 46% -18% 54% -49%

Contribution to

Aggregate Drop in

Spend

16% 84%

Essential

Non-Essential

* Percent contribution to aggregate drop in spend is calculate as: (% Drop in Category A)*(Baseline Share of Category

A)/(% Drop in Aggregate).

To further illustrate the divergence in spending patterns across categories, we split essential

and non-essential spending into more disaggregated categories in Figure 4. Total essential spend-

ing spiked by roughly 20% in early March before dropping by 20% by end-March. However,

there are a wide range of spending responses among goods and services deemed essential. In the

ﬁrst few weeks of March there was a temporary surge in spending on groceries, discount stores,

and pharmacies. Spending at grocery stores, which contributes the largest share of total essential

spending, remained elevated through the end of our sample, aside from a brief decline in the

week including Easter, when many grocery stores are closed. In contrast, spending fell in several

other essential categories like "hospital", "other healthcare", "transit and ground transportation",

and "fuel". Total dollar declines in these categories exceed the dollar increases in grocery spend-

ing, so that overall essential spending declines. Focusing on non-essential spending, declines are

strongest in restaurants, hotel accommodations, and clothing and department stores. Overall,

these results largely mirror those computed in other aggregate data sets and provide reassurance

that our data is consistent with external evidence.

II.C. Heterogeneity In Spending Changes By Income

SPENDING CHANGES OVER THE INCOME DISTRIBUTION

We next explore whether spending reductions (both in aggregate and by category) vary with pre-

pandemic income. We stratify our sample into income quartiles based on total labor inﬂows in

2019.

For context, those in the bottom quartile make less than $28,000 in take-home labor income

per year, while those in the top quartile earn more than $63,000. As discussed in Section I, our

bottom quartile misses unbanked households and the 8% of U.S. households with labor income

did not include the Easter closures.

In future work, we plan to explore also the relationship to income changes during the pandemic.

Figure 4: Credit card spending growth across spending categories

below $12,000, since we cannot reliably measure their income.

Figure 5 plots the year-over-year change in spending for each quartile, both in percentage and

dollar terms. The top income quartile reduces spending by about 39 percent, or $400, by the fourth

week of March, while the bottom quartile reduces spend by 32 percent, or $100. The difference

in the spending drop between income quartiles is starker in dollar terms than percentages, since

high-income households have a higher baseline level of spending. However, divergence in spend-

ing over the income distribution starting in the second half of April is more striking. By the end

of April, the decline in spending partially recovers, with the recovery most pronounced for the

lowest income quartiles. The recovery in spending for the lowest income quartiles occurs in the

same week when many stimulus payments are made in mid April. The timing of the divergence

in spending by income suggests that stimulus payments may have played an important role in

restoring the ability of low-income households to maintain spending during the pandemic.

Appendix A.2 provides the spending changes over the income distribution by form of pay-

ment (debit, credit, cash and check). We observe similar patterns of spend changes across all

forms of payments.

Table A.1 reports the cumulative change in spending by income quartile in 2020 relative to

2019 for the 11 "pandemic" weeks in our dataset between March 15 and May 30. The highest

income quartile contributes disproportionately to the change in spending, accounting for 37% of

initial spending and 50% of the spending decline. As a result, the share of spending for the highest

income quartile declined.

While the results so far show that households with higher income cut spending by more and

Figure 5: Spending by income quartiles

(a) Percent Change (b) Levels

have slower recoveries in spending than low income households, it is important to note that in-

come is correlated with many other factors which might also affect spending responses, so these

are not necessarily causal relationships. One particular concern in the context of the pandemic

is that income is correlated with physical location, and locations vary in the strength of the pan-

demic. In particular, high income individuals tend to live in cities, which have greater disease

burden and more restrictive shutdowns. This means that the relationship between income and

spending dynamics could reﬂect features of where high income households live, rather than ef-

fects of income itself.

To differentiate the role of income from the role of physical location, we look at the relationship

between income and spending over the pandemic within narrow geographic areas. In particular,

we compute the following regression

2020,z,q

− c

2019,z,q

2019,q

= Q uartile

+ ZIP

+ ε

z,q

where c

t,z,q

is average spending per customer with t as the year (for the time period April 15-

May 28), z is zip code and q is the income quartile. We take two steps to minimize the inﬂuence

Note that measuring spending at the geography rather than household-level introduces additional concerns on

this front: high income households are more likely to leave cities than low income households in response to the

pandemic, which might induce spurious declines in spending in locations with many high income households prior

to the pandemic.

of outliers. First, note that the denominator is

2019,q

which uses everyone in the income quartile.

This prevents having one very large or very small ratio from skewing our results. Second, c

t,z,q

is winsorized at the 1st and 99th percentile. We focus primarily on speciﬁcations with geography

z equal to 5-digit zip codes but also explore more aggregated 3-digit zip codes to again limit the

inﬂuence of measurement error.

Comparing odd columns without geographic ﬁxed effects to even columns with ﬁxed effects

shows that relationships between income and spending over the pandemic within zip codes are

very similar to unconditional relationships. That is, high income households cut spending more

during the pandemic relative to low income households living in the same zip code.

The sim-

ilarity of results with and without ﬁxed effects shows that these relationships are not driven by

any observed or unobserved differences across locations where high and low income household

live.

Table 3: Income-Spend Relationships within Geography

Dependent variable:

Spending Gr

owth

(1) (2) (3) (4) (5) (6) (7) (8)

Income Q2 −0.032

∗

∗∗

−0.028

∗∗∗

−0.048

∗∗∗

−0.039

∗∗∗

−0.035

∗∗∗

−0.034

∗∗∗

−0.032

∗∗∗

−0.033

∗∗∗

(0.004) (0.004) (0.002) (0.001) (0.003) (0.002) (0.009) (0.008)

Income Q3 −0.080

∗∗∗

−0.079

∗∗∗

−0.118

∗∗∗

−0.092

∗∗∗

−0.094

∗∗∗

−0.085

∗∗∗

−0.082

∗∗∗

−0.081

∗∗∗

(0.004) (0.004) (0.002) (0.001) (0.003) (0.002) (0.009) (0.008)

Income Q4 −0.154

∗∗∗

−0.142

∗∗∗

−0.217

∗∗∗

−0.161

∗∗∗

−0.179

∗∗∗

−0.159

∗∗∗

−0.149

∗∗∗

−0.147

∗∗∗

(0.004) (0.004) (0.002) (0.001) (0.003) (0.002) (0.009) (0.008)

Constant −0.043

∗∗∗

−0.047

∗∗∗

−0.049

∗∗∗

−0.061

∗∗∗

(0.003) (0.001) (0.002) (0.006)

Geography FE

NO YES NO YES NO YES NO YES

Observations 70,189

70,189 70,189 70,189 25,432 25,432 3,608 3,608

Adjusted R

0.024 0.235 0.238 0.557 0.137 0.506 0.082 0.272

Table regresses year over year spending growth for the period April 15-May 28, 2020 for individual geography × income

quartiles on income quartile dummies with and without geography ﬁxed effects. Columns (1) and (2) deﬁne geography as

5-digit zip codes and equally weight. Columns (3) and (4) use 5-digit zip codes and weight by number of customers in each

zip code × income quartile. Columns (5) and (6) use equal weights and 5-digit zip codes but restrict to zip code × income

quartiles with at least 20 customers. Columns (7) and (8) use equal weights and deﬁne geographies as 3-digit zip codes.

Signiﬁcance: * p < 0.1, ** p < 0.05, *** p < 0.01.

Finally, we further decompose the decline in credit card spending by income quartiles into

Note that our information on address is as-of early 2020.

essential and non-essential categories.

Figures 6 and 7 show that the spending declines for

essential categories are indistinguishable across income groups, while non-essential credit card

spending diverges more across income groups.

Figure 6: Share of credit card spending decline accounted for by essential and non-essential credit

card spending by income quartiles

(a) Essential (b) Non-essential

Figure 7: Reduction in essential vs. non-essential spending by income quartiles

Although all households cut spending dramatically, the fact that high-income households cut

spending by somewhat more may be surprising. Recent research suggests that lower-income

Unfortunately, as mentioned above, we do not have the spending split by categories for the other forms of pay-

ment (debit, cash and check).

households work in jobs that are harder to perform at home, require higher physical proximity,

and therefore may be more impacted by distancing restrictions (Mongey, Pilossoph, and Weinberg

(2020)). Perhaps as a result, recent evidence from administrative ADP data shows that job losses

were four times larger for workers in the bottom income quintile than those in the top income

quintile, with a staggering 35 percent employment decline for the lowest-income workers (Cajner

et al. (2020)). In response to greater income losses, we might have expected lower-income workers

to have cut their spending by more. In fact, we ﬁnd the reverse: higher-income households cut

their spending by slightly more and their spending recovers more slowly.

Differences between high and low-income households in the composition of spending may

be one reason why spending falls by more for high-income households. Non-essential categories

represent a larger share of spending for high-income households – 67 percent of spending in

April 2019 for households in the top income quartile compared to 59 percent for those in the bot-

tom income quartile. In addition, higher-income households have slightly larger drops in their

essential spending. Together, these facts imply that reductions in non-essential spending account

for a somewhat larger share of total spending declines for high- versus low-income households

(85 percent compared to 79 percent, Figure 7). Since these non-essential categories are most af-

fected by the pandemic shut-downs, overall spending of higher-income households may be more

affected by supply-side restrictions. In other words, the effective price of consumption rises more

for higher-income households relative to lower-income households. Thus, the composition of

spending of higher-income households likely contributed to the larger decline in their spending.

As discussed above, the widening of these initial spending declines during the recovery phase

may reﬂect an important role for economic stimulus and transfer programs. The tax rebate checks

that began to arrive in April amount to a larger share of total income for a low-income household

than for a high-income household. Ganong, Noel, and Vavra (2020) also show that the $600 ex-

pansion in UI beneﬁts enacted through Federal Pandemic Unemployment Compensation (FPUC)

boosted wage replacement rates to well over 100 percent for many low-income unemployed work-

ers, providing a substantial income boost once they began receiving beneﬁts.

Finally, higher-income households may be more exposed to negative wealth effects. Higher-

income households hold more ﬁnancial assets, and therefore are exposed to declines in asset prices

during the initial stages of the pandemic. However, wealth effects are unlikely to be a key driver

of the heterogeneous spending responses by income, given previous estimates on the strength of

wealth effects together with the fact that the stock market had recovered most of its pandemic

related losses by the end of May.

CHANGE IN SPENDING BY INDUSTRY OF EMPLOYMENT

We next examine whether workers in sectors most affected by employment disruptions adjust

spending in ways that differ from workers in less affected sectors.

Figure 8 plots spending changes by industry of employment, for each industry where we have

signiﬁcant sample size. We aggregate to industries at the two-digit NAICS code. The one excep-

tion is retail, which we break out into grocery stores, drug stores, and discount stores–generally

considered essential businesses and kept open under social distancing policies–and clothing and

department stores, which were generally deemed non-essential businesses and where layoffs have

been greater (Cajner et al. (2020)).

Overall, it is hard to discern systemic patterns between spending declines and the distribution

of employment losses by industries. It is true that essential workers like those in grocery stores

exhibit smaller spending declines. At the same time, professionals exhibit the largest spending

declines, even though many jobs in this category can more easily be performed remotely.

While industry of employment is closely related to job losses, it is important to note that it is

also highly correlated with income levels and that this may be explaining some of these differ-

ences.

For example, grocery store workers are typically low income, while professional workers

are typically high income. In this sense, patterns when splitting by industry of employment in

many ways mirror those when splitting by income: the largest drops and spending and slowest

recoveries occur in higher-income industries of employment.

To provide a further sense of the separate role of income and industry, in Figure A.11, we

compute spending by industry of employment separately for workers in the highest and lowest

quartile of pre-pandemic income. Comparing variation across industries within income quartile

in Figure A.11 to variation across industries without conditioning on income in Figure 8 shows

that controlling for income substantially reduces the role of industry of employment. Similarly,

comparing the same colored line between panels (a) and (b) in Figure A.11 vs. comparing different

colored lines in Figure 8 also shows that income generally has a greater correlation with spending

dynamics than does industry of employment.

One potential interpretation is that the income channel accounts for only a small share of

spending changes through the end of May. This may not be surprising given the magnitude of

the spending decline. As mentioned previously, we document that average household spending

fell over 35 percent, while the typical unemployed worker receiving UI only cuts spending around

6 percent in normal times (Ganong and Noel (2019)).

However, there are several reasons for caution in concluding that income losses play a small

role in spending effects. First, industry of employment may not fully proxy for job loss in our

See Appendix Table A.2.

Figure 8: Spending Changes Split by Industry of Employment

sample. To the extent that we can ascertain industry of employment primarily for employees of

large ﬁrms, we may not be capturing the income losses for employees of small businesses.

Second, current conditions of the pandemic make comparing the magnitude of the spending

response in April 2020 to that of UI recipients during normal times highly uncertain. On the one

hand, the economic situation is highly uncertain, and labor markets weakened at an unprece-

dented pace. This might cause the unemployed to cut spending by more than during normal

unemployment spells. On the other hand, as a result of the CARES Act, UI beneﬁts are much

more generous in level and duration, and available to many more workers. Furthermore, sizable

stimulus checks were also sent out in April. These income supports might buffer against labor

income-related spending declines if this stimulus continues. The more rapid recovery of spend-

ing for low-income households suggest this channel is at work. The rest of the paper looks at the

behavior of household savings to provide additional evidence on these channels.

III Household Liquid Balances

Given the unprecedented reduction in spending across income and industries documented above,

we next explore whether there were changes in the distribution of household liquid balances. Fig-

ure 1 shows that aggregate private savings increased substantially over the pandemic, reﬂecting

the combination of large declines in spending and large increases in government transfers from

stimulus programs. However, there are reasons to think that the pandemic could have hetero-

geneous impacts on household savings and substantial resulting effects on the distribution of

wealth: households experiencing job loss may draw down on savings (or further draw on sources

of borrowing), while those with job security may be essentially forced to save more, as consump-

tion of many non-essential goods and services is more restricted. In addition, stimulus payments

and other income support programs represent a larger share of pre-pandemic income for high-

than for low-income households.

To explore these effects, we calculate how the distribution of end-of-week balances in house-

hold checking accounts evolved during the pandemic. Speciﬁcally, we explore how various un-

conditional moments of checking account balances evolved as well as how balances changed

across the income distribution and by industry of employment. While checking account balances

are only a subset of total savings and wealth, they represent some of the most liquid and easily

accessible "cash-on-hand" available for households to smooth consumption and self-insure. A

large literature has shown that liquid assets of this form play a crucial role in consumption. Fur-

thermore, checking account balances have the practical advantage of being precisely and easily

measured since checking accounts are one of our primary data sources.

We begin by plotting the average level of liquid balances, and the percentage year-over-year

change from January through the end of May 2020. Figure 9 shows that by the last week of May,

average balances increased by 33 percent year-over-year, or about $1500 dollars relative to earlier

in the year. This increase is consistent with the large increase in the personal savings rate shown

in Figure 1 and with the growth in the stock of commercial bank deposits shown in Appendix

Figure A.13.

Much of the year-over-year growth in checking balances occurred during and

after the week when most EIP stimulus checks were deposited, which suggests that the increase

was driven by these income inﬂows, in addition to the reduced spending we documented in the

previous section.

Figure 10 plots additional moments of the distribution of liquid balances over time. Panel

(a) shows that increases in liquid balances are pervasive, with increases observed at various per-

centiles of the distribution. The dollar increase in balances is greater for households with larger

initial pre-pandemic balances. However, it is important to note that scale effects would be ex-

pected to drive that type of pattern: for example, if all balances double, the accounts with the

largest initial balances will have the largest absolute increases. Panel (b) shows that the lower end

of the distribution is growing more than the top end of the distribution. Interestingly, the year-

over-year growth for lower percentiles shoots up around the time of stimulus payments and then

Note that Figure 9 should not be compared directly to the personal savings rate in Figure 1, since aggregate

personal savings is a ﬂow variable while checking account balances are a stock variable. Appendix Figure A.13

provides evidence on the growth in a more comparable aggregate stock variable to Figure 9.

We further decompose this trend into checking account inﬂows and outﬂows in the Appendix.

Figure 9: Level and Year-on-Year Change in Average Checking Account Balances.

(a) Levels (b) Percent Change

trends down. This suggests that households with low initial liquidity received a large increase in

liquidity from stimulus payments, but they may be fairly rapidly using up this additional cash.

Figure 10: Change in Distribution of Checking Account Balances.

(a) Levels (b) Percent Change

While the results in Figure 10 show that increases in liquid balances are pervasive, it is inter-

esting to explore the relationship with pre-pandemic income. In particular, it is useful to know

whether the increase in aggregate liquid balances was primarily driven by gains at the top of

the income distribution (e.g. by individuals who cut spending most dramatically while generally

maintaining labor income), or by gains at the bottom of the income distribution (e.g. individuals

who cut spending somewhat less and faced larger declines in labor market income – but also had

larger government transfers). Figure 11 plots checking account balances (in levels and growth

rates) by income quartiles. Similar to the unconditional distribution of balances, we see pervasive

increases in balances with increases observed for all groups. Also similar to the unconditional dis-

tribution, there are clear scale effects: the highest income quartile posted the largest dollar gains

of around $2,000. The lowest income quartile increased balances by more than $1,000, which was

the largest increase in year-on-year percentage terms.

Figure 11: Change in Average Checking Account Balances by Income Quartile.

(a) Levels (b) Percent Change

This ﬁgure plots both average dollar balances and year-on-year percentage change in checking account balances by

income quartile. Balance increases are larger in dollar terms for high-income households (who have higher pre-

COVID balances), and in percent terms for low-income households (who have lower pre-COVID balances).

Given these scale effects, what should we conclude about the relative role of high- vs. low-

income households in driving the increase in liquid wealth? One way to answer this question is to

compare each group’s contribution to the aggregate increase, relative to that group’s initial share

of savings. If all groups savings grow by the same amount, then each group’s contribution to the

aggregate increase is equal to its initial share and the wealth distribution is unchanged. If low-

income households have higher savings growth, then they will contribute more to the aggregate

increase than their initial share and wealth inequality will decline.

Table 4: Decomposition of Total Liquid Balances Changes by Income Quartile

Initial

Balances

Share of

Initial

Balances

Increase in

Balances

Share of

Increase in

Balances

Final

Balances

Share of

Final

Balances

Quartile 1

$2.67B 11.4% $1.28B 19.0% $3.95B 13.1%

Quartile 2

$3.44B 14.7% $1.39B 20.7% $4.83B 16.1%

Quartile 3

$5.22B 22.3% $1.60B 23.8% $6.82B 22.7%

Quartile 4

$12.02B 51.5% $2.45B 36.4% $14.47B 48.1%

Total

$23.35B 100.0% $6.72B 100.0% $30.07B 100.0%

Top Decile

$7.19B 30.8% $1305M 19.4% $8.49B 28.2%

Top One Percent

$1.84B 7.9% $294M 4.4% $2.13B 7.1%

Initial balances are computed in February 2020 and Final Balances are calculated in May 2020.

To explore this more formally, Table 4 reports the initial balances in February 2020, the increase

in balances from February to May, and the ﬁnal balances in May for each income quartile. Unsur-

prisingly, higher income quartiles contribute more to the level and change in total liquid balances,

since these households have much more liquid wealth. For example, 51.5% (12.02/23.35) of total

liquid balances come from the top income quartile.

It is also true that the top income quartiles drive the majority of the increase in liquid balances

over the pandemic (36%), but importantly, this increase is less than proportional to the initial

share of liquid wealth held by the top quartile. Table 4 shows that lower-income quartiles are ac-

tually driving more of the aggregate increase in balances than would be expected from their initial

balance shares, so liquid wealth inequality decreases between February and May. This provides

a concrete sense in which the poor are disproportionately increasing savings relative to the rich

during this pandemic. While this shift in the wealth distribution towards low-income households

many not seem huge, it implies a more than three percentage point decline in the share of liquid

wealth held by the richest quartile occurring over a matter of weeks. One important caveat is that

we only measure checking account balances. If higher-income households transferred more assets

out of the checking account, it is possible that we understate the increase in their total assets.

This increase in savings for the poor very likely reﬂects the fact that stimulus checks and

expanded UI beneﬁts provide a disproportionate increase in income for these households. This

also means that this shift may reverse in the near future if stimulus is reduced. For example, the

expanded federal supplement to UI insurance which has led to replacement rates above 100% for

many families, is set to expire at the end of July 2020. The magnitude of the additional spending

drop induced by initial disease avoidance and social distancing restrictions may also dominate

the consumption response caused solely by income loss. This could lead to an increase in savings,

even for those experiencing job loss, but it might not continue as social distancing is relaxed.

On the other hand, if delayed tax payments contributed to the growth in cash balances among high-income

families, liquid asset growth could be short-lived.

Figure 12: Growth of Balances by Industry of Employment

Finally, Figure 12 shows liquid balance growth by industry of employment. While increases

are again pervasive, we ﬁnd that grocery store and department store workers have the largest

growth in checking account balances. This is directly in line with checking account growth by

income, since these are also the lowest income industries in our split.

IV Conclusion

We ﬁnd that all individuals across the income distribution cut spending at the start of the pan-

demic. These declines are massive relative to typical responses to spending responses to unem-

ployment. While high-income households cut spending more than low-income households, these

differences are small relative to the huge common declines in spending. However, beginning

in mid-April, substantial differences by income emerge: while spending begins to recover for

all groups, it does so much more rapidly for the lowest income quartile. Similar patterns emerge

when cutting by industry of employment, with workers in all industries initially cutting spending

dramatically and then workers in low wage industries seeing spending recover more quickly.

One limitation of this paper is that Chase micro data on income during the pandemic period

is still being processed at the time of writing and not yet available for analysis. We therefore turn

to public use data to explore how the income distribution has changed in recent months. Speciﬁ-

cally, we simulate how income has likely changed in the ﬁrst few months of the pandemic using

statutory provisions of the CARES Act, information from the Current Population Survey, and the

unemployment insurance calculator in Ganong, Noel, and Vavra (2020). Although labor income

fell the most for lower-income households, we estimate that total income including transfers actu-

ally increased the most for those at the bottom of the income distribution for two reasons. First,

the EIPs were a ﬂat payment and therefore constituted a larger share of income for low-income

households. Second, because the temporary $600 supplement to UI beneﬁts under the CARES Act

is the same for all unemployed workers, it drives up the replacement rate and resulting income

disproportionately for low income workers. In fact, UI beneﬁts now replace more than 100 per-

cent of lost earnings for low-income households (Ganong, Noel, and Vavra (2020)). The details of

this simulation are described in the Appendix.

Figure 13: Estimated changes in income and spending by income quartiles

This ﬁgure shows the change in income and spending by income quartile. The change in income compares March, April, and May 2020 to

average quarterly income in the prior year. The change in income reﬂects the decline in labor income, the EIPs, and unemployment beneﬁts. The

change in spending compares April 15-May 30 to spending at the same time in the prior year. Our estimate focuses on this narrower time horizon

after the most immediate impacts of the lockdown—which depressed spending across the income distribution—had subsided.

Figure 13 juxtaposes simulation-based estimates of the change in income alongside the change

in spending from Figure 2. There is a suggestive correlation between the pattern of income

changes and the relative pattern of spending changes. Spending falls the least for the group

receiving the most income support, and decreases the most for the group with the least income

support. In future work, when Chase micro data on income during the pandemic becomes avail-

able, we plan to explore the joint dynamics of income and spending at the household-level to

better understand these patterns.

Two other pieces of evidence in our paper suggest that government income support could be

driving spending during this period. First, the timing of the more rapid rebound in spending

for low-income households coincides closely with the timing of EIP stimulus and expanded UI

beneﬁts, suggesting an important role for government support in stabilizing spending during

the pandemic, especially for low-income workers. Second, while increases in liquid balances are

widespread during the pandemic and driven in large part by general declines in spending, we

see that households at the bottom end of the income distribution –who see the largest stimulus

relative to pre-pandemic income – have the largest growth in liquid savings during this period.

As a result, liquid wealth inequality falls between February and May.

Taken together, our results suggest that labor market disruptions were unlikely to be a pri-

mary factor driving spending declines in these initial months of the recession. Many of the effects

of labor market disruptions on spending were likely offset by sizable ﬁscal stimulus and insur-

ance programs. Instead, direct effects of the pandemic were likely the primary factor driving

overall declines in spending. Our analysis does not claim to disentangle the effect of pandemic-

related channels — i.e regulatory shut-downs vs. disease prevalence and fear of infection — on

the spending decline. It instead focuses on the impact of income changes brought about by job

loss and government transfers.

There are some important cautionary implications for future policy. First, it is important to

note that even though aggregate spending has recovered substantially from its nadir, it remains

well below normal. Spending on May 31st, when our sample currently ends, remains very low

in absolute terms, even when compared to spending declines in other severe episodes like the

Great Recession. Spending has partially recovered, but still remains severely depressed relative

to pre-pandemic levels. Policy makers should thus not be too quick to conclude that the economy

has rapidly recovered to normal. Even more importantly, our results suggest that an important

share of this spending recovery has in fact been driven by aggressive ﬁscal stimulus and insurance

payments. While we see a large spike in savings for low-income households immediately after

EIP, these increases may erode as the EIP gets used and if UI beneﬁts get scaled back. This suggests

that new support may be needed to maintain spending for low-income, vulnerable households in

the near future. Phasing out broad stimulus too quickly could potentially transform a supply-side

recession driven by direct effects of the pandemic into a broader and more persistent recession

caused by declines in income and aggregate demand.

References

Alexander, Diane and Ezra Karger. 2020. “Do Stay-at-Home Orders Cause People to Stay at

Home? Effects of Stay-at-Home Orders on Consumer Behavior.” Tech. Rep. WP 2020-12, Fed-

eral Reserve Bank of Chicago.

Andersen, Asger, Emil Hansen, Niels Johannesen, and Adam Sheridan. 2020. “Pandemic, Shut-

down and Consumer Spending: Lessons from Scandinavian Policy Responses to COVID-19.”

Baker, Scott R, R.A. Farrokhnia, Steffen Meyer, Michaela Pagel, and Constantine Yannelis. 2020.

“How Does Household Spending Respond to an Epidemic? Consumption During the 2020

COVID-19 Pandemic.” Working Paper 26949, National Bureau of Economic Research. URL

http://www.nber.org/papers/w26949.

Bitler, Marianne, Hilary Hoynes, and Diane Schanzenbach. (2020). “The social safety net in the

wake of COVID-19. BPEA Conference Drafts.” Tech. rep.

Cajner, Tomaz, Leland Dod Crane, Ryan Decker, John Grigsby, Adrian Hamins-Puertolas, Erik

Hurst, Christopher Johann Kurz, and Ahu Yildirmaz. 2020. “The U.S. Labor Market During the

Beginning of the Pandemic Recession.” Tech. Rep. WP 2020-58, University of Chicago, Becker

Friedman Institute for Economics.

Chetty, Raj, John N. Friedman, Nathaniel Hendren, Michael Stepner, and the Opportunity In-

sights Team. 2020. “How Did COVID-19 and Stabilization Policies Affect Spending and Em-

ployment? A New Real-Time Economic Tracker Based on Private Sector Data.”

———. accessed on 06/15/2020. Economic Tracker. https://tracktherecovery.org/.

EarnestResearch. 2020. Where States Stand: Measuring the Reopenings One

Step at a Time. https://www.earnestresearch.com/insights/

where-states-stand-measuring-the-reopening-one-step-at-a-time/.

Facteus. accessed on 06/15/2020. Insights on Changing Consumer Spending. Facteus Insight

Report on Consumer Spending and Transactions. https://www.facteus.com/reports/

first-report-6-10-2020/.

Ganong, Peter and Pascal Noel. 2019. “Consumer Spending during Unemployment: Positive and

Normative Implications.” American Economic Review 109 (7):2383–2424. URL https://www.

aeaweb.org/articles?id=10.1257/aer.20170537.

Ganong, Peter, Pascal Noel, and Joseph Vavra. 2020. “US Unemployment Insurance Replacement

Rates During the Pandemic.” Tech. Rep. WP 2020-62, University of Chicago, Becker Friedman

Institute for Economics.

Karger, Ezra and Aastha Rajan. 2020. “Heterogeneity in the Marginal Propensity to Consume:

Evidence from Covid-19 Stimulus Payments.” Tech. Rep. WP 2020-15, Federal Reserve Bank of

Chicago.

Mongey, Simon, Laura Pilossoph, and Alex Weinberg. 2020. “Which Workers Bear the Burden of

Social Distancing Policies?” Tech. Rep. WP 2020-51, University of Chicago, Becker Friedman

Institute for Economics.

APPENDIX

Natalie Cox Peter Ganong Pascal Noel Joseph Vavra Arlene Wong

Diana Farrell Fiona Greig

A.1 Change in Spend by Payment Type

Figure A.1: Average Spending Changes on Credit Cards

(a) Percent Change (b) Levels

The left panel plots the year-over-year percentage change in weekly credit card spending. The right panel plots the

dollar change in weekly credit card spending.

A.1

Figure A.2: Average Spending Changes on Debit Cards

(a) Percent Change (b) Levels

The left panel plots the year-over-year percentage change in weekly debit card spending. The right panel plots

the dollar change in weekly debit card spending. Here we demarcate the date of the ﬁrst EIP payments from

Treasury. Over 50% total EIP payments were distributed by April 17 (see https://home.treasury.gov/news/press-

releases/sm1025). Therefore, while the delivery of EIP payments was somewhat "staggered" over time, the majority

were received close to this point in time.

Figure A.3: Average Changes in Cash Spending

(a) Percent Change (b) Levels

The left panel plots the year-over-year percentage change in weekly cash withdrawals. The right panel plots the

dollar change in weekly cash withdrawals. Cash withdrawals at ATMs using debit cards are reﬂected in the debit

card series, so this series primarily reﬂects cash withdrawals with bank tellers.

A.2

Figure A.4: Average Spending Changes in Checks

(a) Percent Change (b) Levels

The left panel plots the year-over-year percentage change in weekly spending via checks. The right panel plots the

dollar change in weekly spending via checks.

Figure A.5: Average Total Spending Changes (Including Check Spending)

(a) Percent Change (b) Levels

The left panel plots the year-over-year percentage change in weekly total spending (inclusive of spending via checks).

The right panel plots the dollar change in weekly total spending (including spending via checks).

A.3

A.2 Spending by Payment Type, Split by Income Quartiles

Figure A.6: Credit card spending by income quartiles

(a) Percent Change (b) Levels

Figure A.7: Debit card spending by income quartiles

(a) Percent Change (b) Levels

A.4

Figure A.8: Cash spending by income quartiles

(a) Percent Change (b) Levels

Cash withdrawals at ATMs using debit cards are reﬂected in the debit card series, so this series primarily reﬂects cash

withdrawals with bank tellers.

Figure A.9: Check spending by income quartiles

(a) Percent Change (b) Levels

A.5

Figure A.10: Total spending by income quartiles (Inclusive of Checks)

(a) Percent Change (b) Levels

Table A.1: Decomposition of Total Spending Changes by Income Quartile

Initial

Spending

Share of

Initial

Spending

Decrease in

Spending

Share of

Decrease in

Spending

Final

Spending

Share of

Final

Spending

Quartile 1

$7.66B 17.8% -$0.89B 10.2% $6.78B 19.7%

Quartile 2

$8.93B 20.8% -$1.30B 15.0% $7.63B 22.2%

Quartile 3

$10.71B 24.9% -$2.14B 24.6% $8.57B 25.0%

Quartile 4

$15.73B 36.5% -$4.37B 50.2% $11.36B 33.1%

Total

$43.04B 100.0% -$8.70B 100.0% $34.34B 100.0%

Top Decile

$7.58B 17.6% -$2394M 27.5% $5.19B 15.1%

Top One Percent

$1.04B 2.4% -$409M 4.7% $0.63B 1.8%

Initial spending is computed as total spending in 2019 for the 11 weeks between March 15 and May 30. Final spending

is computed as total spending over the same 11 "pandemic" weeks in 2020.

A.6

A.3 Spending by Industry of Employment

Figure A.11: Spending Changes Split by Industry of Employment

(a) Spend by Income Quartile 1 (b) Spend by Income Quartile 4

A.4 CPS and Chase Sample Income Comparison

A.7

Figure A.12: Comparison of income quartiles for Chase sample and CPS population

This ﬁgure compares average labor income by quartile in our analysis sample to those in the CPS. Since we only

observe post-tax income in the Chase sample, we adjust the CPS income measure downwards to account for income

and payroll taxes.

Table A.2: Average Income by Industry of Employment

Industry

Total Monthly Pay Average Monthly Pay

Clothing/Dept Store $76,279,559 $2,266

Grocery/Drug/Discount Store $309,150,823 $3,932

Education $169,907,642 $3,918

Government $1,215,885,445 $3,762

Health Care $114,163,903 $5,704

Manufacturing $651,859,889 $5,772

Finance $344,412,048 $5,423

Professional $251,134,078 $2,366

This table lists the average monthly income for individuals in the debit card sample (column 2) and credit card sample

(column 3) that are employed in each industry.

A.8

A.5 Additional Results

Figure A.13: Year-over-year growth in aggregate bank deposits

Percent

2020m1 2020m2 2020m3 2020m4 2020m5

This ﬁgure plots aggregate year over year growth in all commercial bank deposits, using data from the Fed Board of

Governors Release H.8

A.6 Simulation of distribution of income changes

We simulate the changes in income for households using the March 2019 Current Population Sur-

vey Annual Social and Economic Supplement (CPS ASEC). We require average hourly earnings

of at least the federal minimum wage and to avoid issues with eligibility for transfers we also

require that earners are citizens. We cut the sample into weekly earnings quintiles and allow the

unemployment rates to vary between quintiles and month. We calculate a separate unemploy-

ment rate for each quintile and year using the CPS merged to the Earner Study in the CPS. We

merge to the earner study taken at least 8 month prior to the observation period where we mea-

sure unemployment. For 2019, we take the unemployment rates pooling across April, August

and December. For 2019, we match to the April CPS only. To produce variation by month, we

re-scale these unemployment rates so as to match the national unemployment rate in aggregate

in each month, including among the unemployed those who were misclassiﬁed as absent from

work for other reasons. We randomly assign individuals in the ASEC to unemployment at the

rates described above.

If an unemployed worker receives unemployment beneﬁts, we calculate their regular Unem-

ployment Compensation (UC) according to the calculator used in Ganong, Noel, and Vavra (2020).

In our simulation, a worker cannot receive regular UC if they have insufﬁcient earnings history

A.9

according to the rules of their state. Additionally we calculate a recipiency rate among mone-

tarily eligibles, using the ratio of actual beneﬁts paid out in DOL ETA Form 5159 to the implied

value of beneﬁts if every eligible person in our simulation was receiving beneﬁts. For example

in April 2020, we calculate that 55% of eligible beneﬁts were paid out and assume that 55% of

eligibles receive beneﬁts. In 2020, we include the possibility of additional unemployment beneﬁts

from the CARES Act: both the $600 weekly supplement (Federal Pandemic Unemployment Com-

pensation) and the insurance for those with insufﬁcient wage earnings to qualify for regular UC

(Pandemic Unemployment Assistance). For both of these policies we use data on the rollout of

these policies from the Hamilton Project (Nunn, Parsons, and Shambaugh 2020). An unemployed

worker receives an additional $600 in their UC at random, with the probability corresponding to

the share of workers who live in a state that had rolled out FPUC at that time. An unemployed

worker who is monetarily ineligible for beneﬁts receives a total compensation of $600 at random,

with the probability corresponding to the share of workers who live in a state that had rolled out

FPUC at that time and adjusting for the recipiency rate described above.

We additionally model Economic Impact Payments (EIPs) assuming that Adjusted Gross In-

come (which is used to determine EIP eligibility) is equal to household income in the CPS. We

allow each CPS household at most one EIP, which is $1200 plus $500 per child in the household.

We allow a household an economic impact payment if they have an income which is sufﬁciently

low to receive the full payment: we do not model the phase out, instead we give a household no

EIP if they would have received a reduced payment.

To calculate changes, we ﬁnd the percentage change in expected income for each household

with and without the transfers. We then average the expected changes over households within the

income quartiles. The income quartile cutoffs are taken from the Chase sample (so exclude anyone

with income below 12,000), and are calculated on household labor income, which we adjust for

tax.

A.10