Tracking the Pandemic in Real Time: Administrative Micro Data in Business Cycles Enters the Spotlight — Joseph Vavra, Journal of Economic Perspectives, Volume 35, Number 3, Summer 2021
Summary
Joseph Vavra, professor of economics at the Booth School of Business, University of Chicago, writes in the Journal of Economic Perspectives (Volume 35, Number 3, Summer 2021) about how economists used administrative micro data to follow the COVID-19 recession as it unfolded. Vavra notes the US unemployment rate rose from 3.2 percent in February 2020 to 4.1 percent in March and 13.1 percent in April, but the April rate was not reported by the Bureau of Labor Statistics until early May. Researchers turned to data from the payroll processing firm ADP, the scheduling firm Homebase, the JPMorgan Chase Institute and cell phone location firms like PlaceIQ and SafeGraph. The essay discusses Economic Impact Payments of $1,200 for most adults and the $600 weekly unemployment supplement, which it says disproportionately benefited low-income households.
Summary drafted by a model from the document's text below and checked by script against that text before publication. It is a navigation aid, not a reading of what the document proves. Where AI is used
Full text
Journal of Economic Perspectives—Volume 35, Number 3—Summer 2021—Pages 47–66
Tracking the Pandemic in Real Time:
Administrative Micro Data in Business
Cycles Enters the Spotlight
Joseph Vavra
L
ooking back, we now know that the US unemployment rate at the start of the
COVID-19 pandemic rose from 3.2 percent in February 2020 to 4.1 percent
in March and 13.1 percent in April. However, the April unemployment
rate was not reported by the Bureau of Labor Statistics until early May. Preliminary
data on April retail sales was not released by the Census Bureau until mid-May, and
the first release of gross domestic product data by the Department of Commerce
covering April did not occur until the end of July. Thus, a number of economists
turned to private-sector micro data to try to understand the recession while it
was still unfolding: for example, data on employment patterns from the payroll
processing firm ADP and the scheduling firm Homebase, data on bank accounts
and credit card payments from sources like the JPMorgan Chase Institute and firms
that provide financial planning services like mint.com and SaverLife, and even data
on locations of cell phone users from firms like PlaceIQ and SafeGraph. The use of
administrative micro data from these and other sources allowed pandemic-related
research to be produced in nearly real-time and the scope for analysis of individual
behavior, which would be impossible using traditional aggregate data.
In this essay, I loosely define administrative data as that arising as a by-product
of some non-research activity, which contrasts with traditional data sources that
are primarily collected for research purposes, like the Panel Survey of Income
Dynamics, the American Community Survey, or the Consumer Expenditure
Survey. The applications I discuss in this paper use administrative data collected by
■ Joseph Vavra is Professor of Economics, Booth School of Business, University of Chicago,
Chicago, Illinois. His email address is joseph.vavra@chicagobooth.edu.
For supplementary materials such as appendices, datasets, and author disclosure statements, see the
article page at https://doi.org/10.1257/jep.35.3.47.
48 Journal of Economic Perspectives
mortgage servicers, cell phone apps, credit bureaus, financial services firms, and
payroll processing firms. These companies collect vast amounts of data in the course
of their regular business, but this data is not collected with academic research as
an end-goal. While this definition of administrative data could also include data
produced by government administrations as a product of non-research activity,
such as micro data on taxes and households from the Internal Revenue Service or
data based on information from state-level unemployment insurance agencies like
the Longitudinal Household-Employment Dynamics data, I focus my discussion on
privately collected data.
Over the last decade, there has been an explosion in the availability and use
of administrative micro data for economic research—this trend has cut across most
empirical subfields of economics. But in this paper, I discuss ways in which this data
has shaped macroeconomic research on recessions and stimulus policy. The Great
Recession was the first business cycle to occur in this new age of administrative data
availability, and although the research was using this data retrospectively, I begin the
paper with a brief discussion of some macro lessons from this research. However,
I focus mostly on the pandemic, because administrative data has played a crucial,
early role in shaping our understanding of this period. For example, the massive
and incredibly rapid increases in unemployment at the start of the pandemic were
particularly concentrated in low income, service-sector jobs, while high income
workers were largely insulated from job loss and saw only modest cuts in nominal
wages. Expanded unemployment benefits had a substantial and immediate effect
on spending, but did little to discourage job search.
In some cases, the patterns uncovered with studies of administrative data can
also be seen with traditional data sources, but they were apparent weeks or months
sooner because administrative data is often available in nearly real-time. This opens
up the possibility for faster policy reactions. In other cases, administrative data leads
to insights which cannot be obtained with traditional data sources. Traditional
survey-based data typically either have small sample sizes or a limited panel element,
they often have non-trivial measurement error, and they are often released with
substantial lags. Administrative data can provide novel insights by measuring vari-
ables with more precision than traditional data methods or by measuring variables
that are not captured by any traditional data sources. In addition, vast sample sizes
can enable very detailed data cuts and statistical precision and can sometimes allow
for new sources of variation and identification strategies.
Administrative data also raises challenges and concerns. The raw data itself was
collected for other purposes, and for researchers, it can often be messy and difficult
to interpret. The nature of administrative data means it often has a narrow lens of
focus, with great depth but more limited breadth. Representativeness and external
validity of administrative data are often big concerns. The greater statistical preci-
sion does not necessarily mean that estimates are unbiased, because large sample
sizes do not themselves solve identification challenges and resolve issues of causality.
I do not attempt to provide a comprehensive review of all research that uses
administrative micro data to understand recessions; indeed, given the explosion of
Joseph Vavra 49
work in this area and the many administrative datasets now in use, that task would be
enormous indeed. Instead, I intentionally choose a small number of applications,
and only a few papers within each application to illustrate some of the breadth
of administrative micro data available as well as some common challenges. I focus
on applications to the US macroeconomy, but there is clearly also a vast amount
of administrative micro data, policy variation, and applications of interest in other
countries.
Administrative Data in the Great Recession
The Great Recession from 2007 to 2009 offered the first widespread appli-
cation of administrative micro data to understanding recessions. The key role of
house price declines and mortgage market disruptions in the Great Recession is
widely established, most prominently with the work of Mian and Sufi (for an over-
view, see their 2018 essay in this journal). The fact that the recession itself originated
in mortgage markets made data on mortgages (and credit more broadly) vital for
understanding this period, and the simultaneous rise in “big data” information
technology infrastructure meant that credit providers had such data. Many lenders
were willing to make this data available to researchers, and macroeconomists took
advantage. Here, I focus on some applications from my own work to illustrate some
of the insights and the limits of such data.
In Beraja et al. (2019), we use administrative micro data to show that in addi-
tion to this direct effect of house prices on the economy, the house price boom–bust
also substantially constrained the monetary policy response to the Great Recession.
Collateralized borrowing in the housing market is an important part of the mone-
tary transmission mechanism, because interest rate cuts encourage households
to refinance their mortgage and extract home equity to fund current consump-
tion. However, using the Credit Risk Insight Servicing McDash (CRISM) dataset
produced by merging mortgage servicing records from Black Knight Financial
Services (BKFS) with credit bureau data from Equifax, we show that this transmis-
sion mechanism was substantially dampened during the Great Recession.
This monthly panel data includes detailed loan-level characteristics including
loan balances, interest rates, and origination characteristics for tens of millions of
loans which are serviced through BKFS. These loans are, in turn, linked to Equifax
borrower credit records, allowing us to measure consolidated borrowing posi-
tions, creditworthiness, and, most importantly, letting us link successive loans for
a specific borrower across time. This monthly data ultimately allows us to measure
refinancing and home equity extraction after the large declines in mortgage rates
induced by the first round of the Federal Reserve’s quantitative easing program in
November 2008.
During the Great Recession, house prices fell substantially on average, but
declines varied greatly across space. Using the CRISM micro data, we show that
interest rate declines during the Great Recession had the smallest effects on
50 Journal of Economic Perspectives
Figure 1
Mortgage Refinancing and Cash-Outs during the Great Recession
6,000
1.5
Monthly refinance
Highest equity quartile
Total amount
Highest equity quartile 5,000
1 4,000 Lowest
equity
cashed out (mn $)
3,000
propensity, in percent
quartile
0.5
2,000
Lowest equity quartile
0 1,000
09
20 9:9
0
20 9:7:1 1 09
20 9:9
0
20 :7 :1 1
0
20 9:5
0
20 9:3
0
20 9:1 09
20 9:5
0
20 9:3
0
20 :1
0 1
20 8:1
0
20 8:9
0
20 8:7 09 1
20 8:1
0
20 :9
08
20 8:7
0
20 8:5
0
20 8:3
0
20 8:1 0
20 :5
08
20 8:3
0
20 8:1
20 0 20 0
Source: This figure is reproduced from Beraja et al. (2019, Figures 3a and 4a).
Note: Left panel shows refinancing propensities in the lowest equity (largest house price decline)
compared to highest equity metropolitan statistical areas, and the right panel shows cash-out volumes.
refinancing in the locations with the largest house price declines and increases in
unemployment. For example, both refinancing in general and cash-out activity,
in particular, rose much less after quantitative easing (marked by the red lines in
Figure 1) in these locations with little housing equity. An obvious explanation is that
many households were underwater in these locations, making it difficult or impos-
sible to refinance, and for households who could refinance, they had more limited
equity to extract.
Translating this cross-region evidence through the lens of a macro model, we
conclude that the house-price bust substantially constrained the stimulative power
of monetary policy during the Great Recession. In closely related work, Berger et al.
(forthcoming) uses this same CRISM data over a much longer period of time to
show that the strength of this refinancing channel of monetary policy is influenced
not just by house price movements but also by the past behavior of monetary policy
itself. Keeping interest rates low for an extended period of time, as has been done
since the Great Recession, means that many households lock in low interest rates
and become less sensitive to future rate stimulus—even after rates return to more
normal levels.
This research using administrative data provides a new basis for thinking about
the aggregate strength of this refinancing channel. First, while refinancing has been
studied empirically for almost 40 years, older research tended to focus on data with
relatively small samples, which may or may not be representative when trying to
draw inference for the economy as a whole. In contrast, the CRISM data covers
around 60 percent of the entire mortgage market, with a broad cross-section of
loan types and characteristics, so it is much more likely to be representative. Second,
prior research typically studies loan-level datasets without links across time, which
means earlier studies cannot separately distinguish loan prepayment arising from
rate refinancing, cash-out refinancing, moving, and default.
Tracking the Pandemic in Real Time 51
However, this data is not a panacea, and it faces some limitations and pitfalls
in answering some important questions. The data provides a very detailed lens
into mortgage characteristics and household liabilities, but it has essentially no
useful information on other important aspects of the household balance sheet like
spending, assets, or income. Because this data has greater depth than breadth in
its lens of coverage, it is difficult to answer important questions like how spending
responds to mortgage refinancing or how refinancing responds to income shocks.
While some papers (like Bartlett et al. 2019) have linked this data to other admin-
istrative datasets to expand the set of demographic information available, to my
knowledge, no links to many key covariates like income exist.
On this particular point, it is useful to highlight a potential pitfall that arises
frequently in the use of administrative micro data: imputed data. Casual users of
Equifax CRISM data may misleadingly think that Equifax does collect information on
income because their data reports such information. However, these variables are
entirely imputed rather than coming from any actual data on income, so they are
of little practical use. This particular data is well-documented and the imputation
points are clear from inspection of data codebooks. However, the point illustrates a
broader practical concern with the recent rise of various “big data” insight providers
that mark micro data products. These products often draw from many different
sources with little transparency and market the breadth of data that they provide—
even though some of this data is imputed or predicted using machine learning.
It is important, but not always possible, to distinguish actual from imputed data
in these sources when commercial motives mean there is little transparency about
underlying sources or methods. While this concern is clearly data-set specific, I
recommend that researchers invest time to really understand the collection and
construction of the data that they use in order to limit these potential problems.
While a lot of research is focused on mortgage markets, many papers also used
administrative micro data to explore more typical macroeconomic questions during
the Great Recession. For example, Stroebel and Vavra (2019) and Grigsby, Hurst,
and Yildirmaz (2021) study the cyclicality of prices and nominal wages, respectively.
Stroebel and Vavra (2019) use weekly store-UPC pricing data from the marketing
firm IRI to construct local price indices and use these indices in a cross-zip-code-
based identification strategy to argue for procyclical price markups. This type of
analysis would be impossible with traditional price indices, which are only avail-
able at much higher levels of spatial aggregation. Grigsby, Hurst, and Yildirmaz
(2021) use ADP data (discussed in more detail in the following section) to show
that nominal wage cuts were much less common during the Great Recession than
implied by traditional datasets. In traditional datasets, it is very difficult to measure
the frequency and size of nominal wage adjustment because even tiny measurement
errors can contaminate results. ADP data measures the actual wages paid using
administrative data from those paying the wage and can thus eliminate wage “cuts”
arising from measurement error.
All of the Great Recession research highlighted in this section exploits admin-
istrative micro data for research, which could not be performed with traditional
52 Journal of Economic Perspectives
datasets. In general, this type of unique analysis is where administrative data has the
highest value added. However, even in these research applications which would be
impossible with traditional sources, it is important to highlight that administrative
data does not supplant traditional data. This research still relies on traditional data
sources for crucial benchmarking steps and validation of representativeness, which
are pervasive concerns with administrative data.
Great Recession research demonstrated the fundamental value of administra-
tive micro data for macroeconomics. Furthermore, the relationships established
between academic researchers and data providers through this research, in turn,
played a crucial role in speeding analysis of the pandemic when it arrived.
Administrative Data in the Pandemic
Economists have responded to the worldwide health crisis with an unusually
rapid and focused outpouring of research on its economic effects. This analysis
has been produced much more rapidly than in the Great Recession, often being
released weeks or even just days after relevant events. This pace of research opens
new opportunities for influencing policy as it unfolds rather than later analyzing
the consequences of policy, but it also introduces a number of novel challenges.
There are obvious trade-offs between producing deep and careful research that will
stand the test of time and producing research quickly. Indeed, the fast pace of this
research means that more findings will likely eventually need to be revised or clari-
fied relative to research conducted at a more typical academic pace.
In my discussion here, I choose applications with two goals in mind: 1) I want to
highlight several broadly different types of administrative data used to understand
the pandemic recession; and 2) I want to highlight results and conclusions that have
received some amount of support in multiple administrative data sources or with
traditional data. Most of the research I mention here focuses on the period from
March to September 2020. We know that underlying health and economic condi-
tions have changed rapidly across time, so it is important to note that conclusions
from research looking at this early stage of the pandemic may differ from research
examining the current stage of the pandemic or the eventual recovery over the
coming months or years.
Labor Market Data
Some of the first empirical research on the pandemic focused on measuring
labor market disruptions using administrative micro data. Cajner et al. (2020) and
Grigsby et al. (2021) use data from the payroll processing firm ADP, and Bartik et al.
(2020) use data from the scheduling firm Homebase to document a number of
labor market facts in the early stages of the pandemic.
ADP is a large human resources firm providing payroll processing for around
26 million US workers each month. This data is broadly representative of private
sector employment using a variety of external benchmarks, although it modestly
Joseph Vavra 53
overweights medium-large firms (Cajner et al. 2020; Grigsby, Hurst, and Yildirmaz
2021). Homebase is a scheduling firm providing services to tens of thousands of
small businesses that employ hundreds of thousands of workers. This dataset is
much less representative because it is skewed towards small firms in sectors like
restaurants and retail that disproportionately employ hourly workers. However,
these firms were among those most disrupted by the pandemic and they are other-
wise somewhat underrepresented in the ADP data. Furthermore, Bartik et al. (2020)
complement the raw data with an additional survey of Homebase users, allowing for
some additional insights using this data.
Both papers show the striking distributional effects of the pandemic: lower-
wage workers were much more likely to lose their jobs than higher-wage workers
during this time period. Figure 2 illustrates these findings. In part, this pattern
arises because low-wage workers tend to be concentrated in sectors like food service,
which were particularly hard hit by the pandemic. Furthermore, this specific low-
wage segment of the population is particularly vulnerable because these workers are
also less likely to have substantial savings.
These broad distributional patterns are masked when focusing on the aggregate
unemployment rate, and they were first established in these administrative datasets.
Several papers have now documented this same pattern of greater unemployment
for low-wage workers using traditional public datasets (for example, Ganong, Noel,
and Vavra 2020; Cortes and Forsyth 2020). Thus, a main advantage of administrative
data in this context was its speed, rather than a unique lens. Using administrative
data to understand labor market trends 4-6 weeks earlier is of great use for policy-
making decisions, but is arguably less crucial at the typical horizons of academic
research.1
However, these administrative data studies also offered some more unique
insights. Bartik et al. (2020) decompose the total reduction of worker hours and
find it was primarily driven by extensive margin effects with firms shutting down
entirely or reducing the size of their workforce, rather than intensive margin effects
where hours were reduced but workers remained employed.
As discussed in the previous section, ADP data has a unique ability to measure
nominal wage adjustment because it measures actual payments made to workers
and thus does not suffer from measurement error, which contaminates traditional
survey-based data. Following Grigsby, Hurst, and Yildirmaz (2021), Grigsby et al.
(2021) find that 6 percent of workers (mostly at the top of the income distribution)
received nominal wage cuts early in the pandemic, but that 30 percent of these
wage cuts were reversed by November. The pace of this nominal wage adjustment
1
Indeed, publicly available labor market data is itself already available quite rapidly. Thus, speed will
generally be a greater comparative advantage for administrative data on spending, rather than for labor
market data, since public data on spending is produced with moderately longer lags. See Chetty et al.
(2020) for an effort to produce and publicly distribute daily statistics on consumer spending, business
revenues, employment rates, and other key indicators disaggregated by ZIP code, industry, income
group, and business size, based on anonymized data from a group of companies. For details, see https://
tracktherecovery.org/.
54 Journal of Economic Perspectives
Figure 2
Employment Change Relative to February 1, 2020, by Income Quintile
1
0.95
Quintile 5
Employment change relative to February 1
0.9
0.85
Quintile 4
0.8
0.75 Quintile 2
Quintile 3
0.7
0.65
Quintile 1
0.6
3/7/20 4/4/20 5/2/20 6/6/20
Note: Computed with ADP data in Cajner et al. (2020, their Figure 4a). The lines show pre-pandemic
wage quintiles.
is substantially greater than during the Great Recession, but it still implies modest
effects on overall earnings relative to the layoffs at the bottom end of the distribu-
tion. Overall, this data shows that the main labor market effect of the pandemic has
been a large increase in unemployment at the bottom of the distribution, and that
there is a more modest decline in wages but with continuing employment at the top
of the distribution.
Moving forward, there will undoubtedly be much more research using this
administrative data to understand labor markets. Two of the biggest advantages of
these administrative data relative to traditional data sources are the ability to link
individual workers together with firms so that workers can be tracked over time,
and the fact that pay and hours can be measured exactly without the measurement
error from self-reported data. On the other side, a potentially significant concern
is that this data captures small employers (Homebase) or broader private-sector
employment (ADP), but it has essentially no information on public-sector employ-
ment. If state and local budget cuts (early in the pandemic) or surpluses (later in
the pandemic) lead to public sector employment changes, this data will largely miss
these trends.
Tracking the Pandemic in Real Time 55
Financial Accounts Data
Data from individual bank accounts can be used to study various household-
level outcomes. The first and most direct source of account-level data are banks
themselves. The primary source of such data in the United States is the JPMorgan
Chase Institute (JPMCI), a think tank within JPMorgan Chase & Co., which has
developed a strictly controlled process to use anonymized account-level data on
the universe of Chase customers directly for academic and policy research. The
second common source of bank account data comes from financial service compa-
nies, which often require the user to first provide bank account log-in information
to obtain some service; once the company obtains this data, they then make anony-
mized versions available for research purposes. For example, users of mint.com and
SaverLife users enter all of their various account information, and these companies
then provide financial planning services and budgeting information to their users.
Users of Earnin can sign up to receive free payday loans, but they must first link to
a bank account in order to do so.2
Bank account information provides a detailed and high-frequency lens into
individual economic behavior. It typically contains transaction-level information
on both account inflows (like direct deposits) and account outflows (like debit
card transactions), allowing researchers to measure the connection among high
frequency income, spending, and savings. Other datasets have detailed information
on individual components (for example, the ADP data described in the previous
section for income, or data produced by Visa or credit card processing companies
for spending), but cannot link these components at a household level. Such links
turn out to be crucial for some of the insights using this data to study the pandemic.
The JPMCI data has the further advantage of large sample sizes: as of 2015, it
includes 27 million checking accounts (Farrell and Greig 2015). In addition, Cox
et al. (2020) shows that the distribution of income in this data is generally similar
to that of the population as a whole (although by construction it does not include
any “unbanked” households). While the JPMCI data has essentially a complete,
transaction-level accounting of everything that occurs within Chase accounts, a
corresponding disadvantage of this data is that it has a limited lens for anything
that occurs outside of Chase accounts, such as activities in second bank accounts or
on non-Chase credit cards. For this reason, most research using JPMCI introduces
various screens so that inactive or barely active accounts are not included in the
analysis, but it is nevertheless possible that some important non-Chase activity is
missed.
An advantage of bank account data from financial aggregators like SaverLife
is that users have strong incentives to include all of their active accounts in order
to obtain reliable planning information. These datasets do tend to have much
smaller sample sizes, and selection is more of a concern because users choosing to
2
Facteus also provides some similar information from card processing, combining individuals using
debit cards, payroll cards, and load cards at an account level, although I have not worked with this data
and am less clear on the underlying nature of the sample.
56 Journal of Economic Perspectives
use financial planning apps may not be representative of the broader population.
However, Baker (2018) provides substantial benchmarking and validation exercises
to argue that these selection concerns are of more limited import in his applica-
tions and that data from these types of platforms are indeed informative about
broader behavior. While few studies have used Earnin data, these selection issues
are likely to be an even greater concern for that data because this sample is built
on those seeking payday loans. However, that data may be useful for understanding
the behavior of particular vulnerable populations of interest that might be under-
represented in other datasets.
What was learned from administrative bank account data during the pandemic?
Cox et al. (2020) complement the analysis of labor income losses discussed earlier
by using JPMCI data to show how spending and savings have changed across the
income distribution over the same period. This analysis requires linking individual
income, spending, and savings, which is possible only as a result of the unique lens
offered by administrative financial account data.3 Using a sample of around five
million active account-holders for which they can measure pre-pandemic income
using direct deposit information, Cox et al. (2020) find dramatic declines in
spending across the income distribution at the end of March 2020. These spending
declines are strongest in certain categories like entertainment and hotel accommo-
dations, which require in-person activity.4 However, starting in mid-April, spending
recovers much more rapidly for low-income households. At the same time, these
low-income households also see the largest growth in checking account balances.
Figure 3 illustrates these patterns. This finding seems surprising in light of the
evidence from Cajner et al. (2020) and Bartik et al. (2020) that these households
had the largest declines in labor income over this same period in time.
How can the households experiencing the most job loss during the pandemic
fare best in terms of spending and savings growth? The timing suggests an impor-
tant role for government support programs. The divergence in spending patterns
occurred shortly after the passage of the Coronavirus Aid, Relief, and Economic
Security (CARES) Act, which created large transfers that disproportionately bene-
fited low-income households. In particular, the CARES Act provided one-time
broad-based Economic Impact Payments of $1,200 for most adults and created a
Federal Pandemic Unemployment Compensation program that added $600 per
week on top of regular state unemployment insurance benefits from April through
July. The Economic Impact Payments were the same absolute size for all but the
3
Some of these spending patterns could potentially be explored in the publicly available consumer
expenditure data once it becomes available covering this period. In addition, although Chetty et al.
(2020) do not have individual income data, they find similar zip code–level spending patterns across zip
codes with high and low income, suggesting that these spending patterns over the income distribution
are not unique to JPMCI data. Finally, it is useful to note that the large run-up in savings observed in
JPMCI is consistent with the spike in savings observed in aggregate data once it was released.
4
These category spending patterns do not require account-level links and have been observed in a variety
of credit card data sources including Affinity (Chetty et al. 2020) and Womply (Alexander and Karger
2020). These papers include a much more extensive discussion of these category patterns.
Joseph Vavra 57
Figure 3
Changed in Total Spending and Average Daily Balances during the Pandemic
Year-over-year percent change in total Year-over-year percent change in average
spending by income quartile daily balance per household
40
Percent change in daily balances
Percent change in total spending
National emergency 50
National emergency
declared, March 13 declared, March 13
20
40
Income quartile
0
30 1 (lowest)
Income quartile 2
−20
(average over the week)
1 (lowest) 3
20
−40 2 4 (highest)
EIP payments
3 from EIP payments
−60 10
4 (highest) Treasury, from Treasury,
April 15 April 15
−80 0
Feb-01 Mar-07 Apr-11 May-16 Feb-01 Mar-07 Apr-11 May-16
End of week End of week
Source: Cox et al. (2020, Figure 5a and 11b).
highest-income individuals, so they resulted in larger income growth for low-income
households. The $600 unemployment insurance supplements even more dispropor-
tionately benefited low-income households, in part because low-income households
were more likely to be unemployed, and in part because a flat weekly $600 supple-
ment represents a larger share of pre-job loss income for low income individuals.
Figure 3 uses vertical lines to show the date when the national pandemic emergency
was declared and when Economic Impact Payments were first distributed by the US
Treasury.
These correlations with policy timing are merely suggestive, but several papers
have used administrative data to make an argument for causal effects of these
expanded transfers. Baker et al. (2020) use the SaverLife data described above
to analyze high frequency spending responses to the one-time Economic Impact
Payments. Using daily data on 38,000 active account users from December 2019
through May 2020 they identify just over 23,000 Economic Impact Payments by
looking for direct and check deposits in certain categories with sizes corresponding
to common amounts of these payments, like $1,200 and $2,400. They then estimate
a marginal propensity to consume from 7 days before to 23 days after the receipt of
the payments, using a distributed lag regression controlling for individual and time
fixed effects to estimate responses. Overall, they find a cumulative marginal propen-
sity to consume of around $0.37 in their raw sample, or $0.27 when they reweight
their relatively low-income sample to instead match the distribution of income and
several other observables in Current Population Survey data. Users with earnings
under $1,000 per month had a marginal propensity to consume roughly twice as
large as users earning $5,000 a month or more. There is an even steeper gradient
with liquidity, as users with the highest account balances have marginal propensi-
ties to consume around 0.1 while those with balances under $100 have marginal
58 Journal of Economic Perspectives
propensities to consume of 0.4 or more. Overall, these results are consistent with
evidence from prior regressions in Johnson, Parker, and Souleles (2006) and Parker
et al. (2013) that stimulus checks significantly boosted spending.
However, even though the SaverLife data is vastly superior in frequency, detail,
and accuracy to the Consumer Expenditure Survey data used in studies of earlier
recessions, the policy variation itself during the pandemic is harder to interpret
from a causal identification perspective. Unlike in the 2001 and 2008 recessions,
the timing of stimulus checks during the pandemic was not random and was
concentrated over a shorter period. Paper checks during the pandemic were priori-
tized by income, and direct deposit timing depended on past tax filing status and
differed for those on Social Security and other government support programs, all
of which are correlated with income. The non-random timing of payments along
with the differential trends in spending by income shown in Cox et al. (2020) intro-
duce some concerns about interpreting their estimates of marginal propensity to
consume, especially at longer horizons, as pure causal effects. I note this not as a
critique of their general conclusion that stimulus checks increased spending. Given
the very high frequency of sharp breaks in spending observed in Baker et al. (2020),
there is little doubt that the stimulus checks did boost spending. But I do want to
highlight the broader point that using administrative data, even when it might have
enormous sample sizes and more precise measurement, does not in-and-of-itself
solve identification challenges.
Overall, Economic Impact Payments were a large part of the initial US govern-
ment response to the pandemic, totaling $270 billion by the end of May 2020, and
the evidence in Baker et al. (2020) implies they had an important role in increasing
spending, especially for lower-income households as shown in Figure 4. In addi-
tion to these broad-based payments, the more targeted $600 weekly supplements
to unemployment insurance also played a particularly important role in increasing
spending for low-income households. Overall, the aggregate scale of these expanded
unemployment benefits was similar to the Economic Impact Payments, with roughly
$260 billion in expanded benefits paid out from April through the end of July 2020.
However, unlike Economic Impact Payments, these $600 payments were targeted at
households with declines in labor market income. These $600 supplements nearly
tripled typical benefit levels and resulted in benefits that replaced about 145 percent
of lost income for the median worker.
Ganong et al. (2021) use JPMCI data to show that expanded unemployment
benefits substantially boosted the spending of unemployed households, while
having a comparatively muted effect on job search at the time. They show that
while the $600 supplements were available, the net income and spending of unem-
ployed households actually rose rather than declined after job loss, both in absolute
terms and relative to that of employed households. Using a variety of identification
strategies, they estimate causal spending responses to the start of the $600 unem-
ployment insurance supplement in April 2020, to its expiration in August, as well as
to additional short-term supplements of $300 that were paid out in September. For
example, they estimate causal responses to the start of expanded unemployment
Tracking the Pandemic in Real Time 59
Figure 4
The Timing of Unemployment Insurance and Increased Spending
$2,000
Weekly unemployment
$1,500
●
● ●
insurance benefit
$1,000 ● ●
● ●
●
$500
$0 ● ● ● ● ● ● ● ● ● ● ● ● ● ●
Jan Feb Mar Apr May Week of first benefit
● April 5
April 19
Spending (card and cash)
$700 ●
● May 10
●
● ● May 31 (control)
●
● ●
●
$600 ● ● ●
● ● ●
●
● ● ●
$500
●
●
$400 ●
Jan Feb Mar Apr May
Source: Ganong et al. (2021)
benefits by comparing the spending and income of households who all become
unemployed at the end of March 2020 but begin receiving unemployment benefits
at different dates.
The spending of these different cohorts is nearly identical up to the start
of unemployment benefits and then jumps immediately when benefits begin.
Exploiting these differences across cohorts, they estimate a marginal propensity to
consume out of benefits of 0.43. Strikingly, they also find high marginal propensity
to consume at the time of benefit expiration in August as well as in response to
an additional short-lived $300 benefit increase paid out in September 2020, even
though unemployed households had built up substantial liquidity by then through
prior receipt of expanded benefits.
In contrast to these large spending responses, Ganong et al. (2021) find small
effects of the $600 on job finding. Simple job finding models calibrated to pre-
pandemic evidence predict a very large and sustained increase in job finding after
the expiration of unemployment benefits—the job finding rate was quite stable
from May through October. Moreover, they find an important role for recall to
previous employers, rather than transitions to new employers, in explaining what
fluctuations in job finding rates do exist over this period. Fitting a job search model
with various elements to match the patterns in job finding and recalls, they estimate
that employment distortions induced by the $600 unemployment insurance supple-
ments were much lower than implied by pre-pandemic distortion estimates. They
60 Journal of Economic Perspectives
also use this data to document a number of novel labor market facts that traditional
data sources do not measure: for example, traditional unemployment data does not
track individuals, but they show that repeat unemployment is particularly important
during the pandemic.
Overall, this research leverages administrative account-level data in crucial ways,
which cannot be done using other datasets currently available, by linking account-
level income measures to account-level measures of spending and saving. However,
this account-level data is ill-suited for answering certain other questions of great
interest. For example, what fraction of households who lost jobs received unemploy-
ment insurance and how long do they have to wait to receive benefits? It might seem
that bank account data could be used to answer this question, but in fact a large
fraction of individuals now receive their unemployment benefits via pre-paid debit
cards. These cards are unobserved in JPMCI data, and given their transitory nature,
they are also unlikely to be linked in financial account aggregators like SaverLife.
As a result, this financial account data cannot distinguish a worker who is waiting-
for/denied/never-filed-for unemployment benefits from one is currently receiving
unemployment benefits via a prepaid card.
Cell Phone Data
Cell phones produce near-constant streams of data that allow for detailed infor-
mation on geographic location at very high temporal frequencies in near real-time.
During the pandemic, movement and social interaction was of even more direct
interest than usual. This data can also be used to proxy for shopping activity in
narrow geographic areas, which can be used to identify the effects of government
shutdown and reopening policies.
Couture et al. (2021) use data on roughly 75 million unique cell phones
from PlaceIQ to construct a “daily location exposure index” that captures county-
to-county movements together with a “device exposure index” that measures the
exposure of devices to each other within venues. These exposure indices, which
they post publicly every weekday, have been used in a variety of papers. In addition
to providing this public good, Couture et al. (2021) provide an extensive discus-
sion of representativeness and advantages and disadvantages of this cell phone data
in addition to documenting several interesting patterns of movement during the
pandemic.
I will not repeat that discussion here but will highlight a few key observations.
First, the data is broadly representative of general population distributions and
flows across space when benchmarked against various external data sources, but
it is more representative when studying broader geographic areas like counties or
states than when studying vary narrow geographies. Second, the PlaceIQ and most
other US-based cell phone data is collected through smartphone apps with loca-
tion-tracking services rather than directly from cell-service providers. Because older
adults are less likely to own smartphones, older households are less represented
in this data. This can, in turn, be important in studies of the pandemic, given that
COVID-19 exhibits a sharp age-gradient in disease outcomes. They also caution
Joseph Vavra 61
that cross-location level comparisons are likely to be less reliable than time-series
variation within locations across time due to differences in coverage and represen-
tativeness across space. Furthermore, device IDs turn over frequently, which means
that the panel element at the level of individual devices is typically limited to around
six months.
Couture et al. (2021) is primarily focused on the development and validation
of their exposure indices rather than on particular applications, but they do demon-
strate a number of interesting results. During the pandemic, for example, the
indices show that a sharp decline in movement in and out of Manhattan is detect-
able in near real-time in the early stages of the pandemic. More generally, they also
explore the relationship between cell phone visit data and credit card spending
data, which track each other very closely in some categories like arts and entertain-
ment but sharply diverge in other categories like grocery spending. This divergence
in grocery spending likely reflects a substitution towards online purchases together
with the consolidation of multiple trips with smaller expenditure into single trips
with larger expenditure per trip. Thus, while visits and spending generally track
each other, this is not uniformly true. As a result, questions focused on physical pres-
ence, like in-person shopping time, are likely to be more reliably answered with cell
phone data than questions about ultimate expenditures.
Goolsbee and Syverson (2021) use similar cell phone data from 45 million cell
phone users produced by SafeGraph to try to understand the factors driving declines
in consumer traffic from March 1 to May 16.5 In particular, they seek to differentiate
the role of government-imposed restrictions from households voluntarily staying
home in driving changes in consumer behavior. To do this, they combine local store
visit data from SafeGraph with county- and city-level shutdown policies and imple-
ment a cross-border identification strategy, which compares weekly shopping visits
across counties with different restrictions within commuting zones. In particular,
commuting-zone fixed effects should help to control for unobserved factors, like
health fear, that are common to the commuting zone within that week. Thus, the
effects of government restrictions will be identified only from variation in consumer
behavior across counties with different policies all within the same commuting zone.
This identification strategy reduces the concern that correlations between govern-
ment restrictions and declines in consumer activity reflect a common response to
rising health risk rather than a causal effect of the restrictions themselves.
Overall, they find that while consumer traffic fell by 60 percentage points,
legal restrictions explain only 7 percentage points of this decline, which means that
declining economic activity was predominantly driven by direct consumer responses
to the virus rather than by government shutdown policies. Of course, this result
needs to be interpreted with the usual caveat that cross-sectional causal effects may
differ from aggregate effects; for example, any restrictions in one county that have
5
Updates after publication of the original paper, available at (https://bfi.uchicago.edu/insight/
research-update-drivers-of-economic-decline/), extend the analysis to the period of re-opening in the
summer and to the second round of closures in the fall and find nearly identical results.
62 Journal of Economic Perspectives
spillover effects on the commuting zone as a whole will be missed by this empirical
strategy. Nevertheless, such spillovers would need to be implausibly large to undo
the main message. Also, it is important to reiterate the point raised above that cell
phone data captures store visits, not expenditures. However, Alexander and Karger
(2020) use a similar identification strategy with store-level credit card spending data
from Womply and arrive at very similar conclusions.
While the applications discussed in this section used a variety of different data
sources, several lessons emerge. First, the pandemic caused substantial declines in
spending which were very concentrated in certain service sectors. Second, these
sectors employ many low-income households, so the pandemic led to much larger
unemployment for low-income households. Third, these declines in spending and
increases in unemployment were largely unavoidable in the sense that they were
caused directly by health fears rather than by mandated government shutdowns.
Fourth, US government transfers, in the form of broad-based stimulus checks and
expanded unemployment benefit checks, led to substantial increases in spending
for many of those otherwise hardest hit by the recession.
Additional Discussion and Conclusions
Some of the research results from using administrative micro data rely crucially
on the particular lenses offered by this data. In other cases, similar results can even-
tually be obtained using traditional datasets, but the use of administrative data can
allow research to proceed more quickly in a way that can offer more timely input
for policy decisions. I close with a few additional reflections on the challenges facing
researchers who are contemplating the use of administrative micro data.
First, administrative data can be difficult to interpret because samples are
often unrepresentative and because they may contain a limited set of covariates of
interest. For this reason, I stress that administrative micro data should be viewed as
a complement rather than substitute for traditional data sources. Without represen-
tative data sources for benchmarking and validation, it is very difficult to interpret
results from particular administrative data sources. Indeed, there has also been a
dramatic expansion of traditional data surveys themselves during the pandemic.
For example, the Census now conducts rapid household and business pulse surveys.
Some of this data has, in turn, been used for similar near real-time analysis (as in
Dube 2021).
Second, administrative datasets are often very large and thus will require
substantial computational resources and time to analyze. Their size can also exacer-
bate the first concern about interpretation because it can be harder to notice data
anomalies and other issues when the data itself is cumbersome to analyze.
Third, administrative datasets often have significant barriers to access: they
may be expensive to purchase or might depend on personal connections for access.
Further, continued access and data availability is often not assured. For example,
data access can disappear because the provider goes out of business, changes
Tracking the Pandemic in Real Time 63
business models, becomes subject to new legal restrictions, changes licensing terms,
or for many other unanticipated reasons. Even public-spirited firms quail at the
prospect of making a commitment to maintaining these data-sharing arrangements
over a period of years for each research project they approve.
These issues with access and the institutional risk of private-sector administra-
tive data raise some concerns for the economics profession. Research directions may
be overly influenced by the interests of a small number of individuals privileged with
access. Other researchers may have limited or no ability to test the reproducibility
and robustness of findings and to carry out extensions of the analysis. Journals are
increasingly requiring detailed replication code and detailed data access instruc-
tions in online data repositories as conditions of publication. These repositories and
associated access information are often a useful starting point for those interested
in using the same data for follow-up work, but they do not themselves eliminate
access barriers.
However, other kinds of administrative data are becoming more publicly
available. For example, some academic institutions have been taking on a role as
data intermediaries. The Kilts Center at the University of Chicago Booth School
of Business acquires data from AC Nielsen and other private providers and then
administers widely available academic licenses for this data. During the pandemic,
Opportunity Insights began publicly publishing data that they obtain from a very
large variety of private data providers (Chetty et al. 2020), although confidentiality
agreements mean that this is not micro data and is instead aggregated to zip code
or higher levels of aggregation. Finally, in response to the pandemic, many data
providers have reduced the barriers to entry for acquiring micro data. For example,
the SafeGraph data discussed above is now widely available for academic use.
Some government institutions have also embraced the use of administrative
micro data and might play a role in this process. The Federal Reserve, the Consumer
Financial Protection Bureau, the Office of the Comptroller of the Currency, and
other regulators have access to a wide variety of administrative micro data, and
they increasingly allow the use of this data to enrich their internal research: for
example, Aladangady et al. (2019) discusses the potential of high-frequency admin-
istrative data for informing public statistics. These institutions have also increasingly
provided paths for external researchers to access this data. By acting as an interme-
diary between private data providers and the research community as a whole, this
model can potentially break down some barriers and democratize access to this
data.
For individual researchers, working with administrative data often requires a
significant investment of time and a degree of risk. Nevertheless, this comes with
the opportunity for transformative research insights, which could not be made with
other sources.
64 Journal of Economic Perspectives
References
Aladangady, Aditya, Shifrah Aron-Dine, Wendy Dunn, Laura Feiveson, Paul Lengermann, and
Claudia Sahm. 2019. “From Transactions Data to Economic Statistics: Constructing Real-Time,
High-Frequency, Geographic Measures of Consumer Spending.” https://www.nber.org/
books-and-chapters/big-data-21st-century-economic-statistics/transactions-data-economic-
statistics-constructing-real-time-high-frequency-geographic-measures.
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.” https://doi.org/10.21033/wp-2020-12.
Baker, Scott R. 2018. “Debt and the Response to Household Income Shocks: Validation and Application
of Linked Financial Account Data.” Journal of Political Economy 126 (4): 1504–57.
Baker, Scott R., R.A. Farrokhnia, Steffen Meyer, Michaela Pagel, and Constantine Yannelis. 2020.
“Income, Liquidity, and the Consumption Response to the 2020 Economic Stimulus Payments.”
NBER Working Paper 27097.
Bartik, Alexander W., Marianne Bertrand, Feng Lin, Jesse Rothstein, and Matthew Unrath. 2020.
“Measuring the Labor Market at the Onset of the Covid-19 Crisis.” Brookings Papers on Economic
Activity 1–72.
Bartlett, Robert P., Adair Morse, Richard Stanton, and Nancy Wallace. 2019. “Consumer Lending
Discrimination in the FinTech Era.” https://doi.org/10.2139/ssrn.3063448.
Beraja, Martin, Andreas Fuster, Erik Hurst, and Joseph Vavra. 2019. “Regional Heterogeneity and the
Refinancing Channel of Monetary Policy.” Quarterly Journal of Economics 134 (1): 109–83.
Berger, David, Konstantin Milbradt, Fabrice Tourre, and Joseph Vavra. Forthcoming. “Mortgage Prepay-
ment and Path-Dependent Effects of Monetary Policy.” American Economic Review.
Cajner, Tomaz, Leland D. Crane, Ryan A. Decker, John Grigsby, Adrian Hamins-Puertola, Erik Hurst,
Christopher Kurz, and Ahu Yildirmaz. 2020. “The U.S. Labor Market during the Beginning of the
Pandemic Recession.” Brookings Papers on Economic Activity Summer (2020).
Chetty, Raj, John N. Friedman, Nathaniel Hendren, Michael Stepner, and The Opportunity Insights
Team. 2020. “How Did COVID-19 and Stabilization Policies Affect Spending and Employment?
A New Real-Time Economic Tracker Based on Private Sector Data.” NBER Working Paper 27431.
Cortes, Guido Matias, and Eliza C. Forsythe. 2020. “The Heterogeneous Labor Market Impacts of the
Covid-19 Pandemic.” Upjohn Institute Working Papers 20-327.
Couture, Victor, Jonathan I. Dingel, Allison Green, Jessie Handbury, and Kevin R. Williams. 2021. “JUE
Insight: Measuring Movement and Social Contact with Smartphone Data: A Real-Time Application
to COVID-19.” Journal of Urban Economics.
Cox, Natalie, Peter Ganong, Pascal Noel, Joseph Vavra, Arlene Wong, Diana Farrell, Fiona Greig, and
Erica Deadman. 2020. “Initial Impacts of the Pandemic on Consumer Behavior: Evidence from
Linked Income, Spending, and Savings Data.” Brookings Papers on Economic Activity: 35–82.
Farrell, Diana, and Fiona Greig. 2015. Weathering Volatility: Big Data on the Financial Ups and Downs of U.S.
Individuals. JP Morgan Chase Institute. https://www.jpmorganchase.com/content/dam/jpmc/
jpmorgan-chase-and-co/institute/pdf/54918-jpmc-institute-report-2015-aw5.pdf.
Dube, Arindrajit. 2021. “Aggregate Employment Effects of Unemployment Benefits during Deep Down-
turns: Evidence from the Expiration of the Federal Pandemic Unemployment Compensation.”
NBER Working Paper 28470.
Ganong, Peter, Fiona Greig, Max Liebeskind, Pascal Noel, Daniel M. Sullivan, and Joseph Vavra. 2021.
“Spending and Job Search Impacts of Expanded Unemployment Benefits: Evidence from Adminis-
trative Micro Data.” Becker Friedman Institute for Economics Working Paper 2021-19.
Ganong, Peter, Pascal Noel, and Joseph Vavra. 2020. “US Unemployment Insurance Replacement Rates
during the Pandemic.” Journal of Public Economics 191.
Goolsbee, Austan, and Chad Syverson. 2021. “Fear, Lockdown, and Diversion: Comparing Drivers of
Pandemic Economic Decline 2020.” Journal of Public Economics 193.
Grigsby, John, Erik Hurst, and Ahu Yildirmaz. 2021. “Aggregate Nominal Wage Adjustments: New
Evidence from Administrative Payroll Data.” American Economic Review 111 (2): 428–71.
Grigsby, John, Erik Hurst, Ahu Yildirmaz and Yulia Zhestkova. 2021. “Nominal Wage Adjustments During
the Pandemic Recession.” American Economic Review: Papers and Proceedings 11: 258–62.
Johnson, David S., Jonathan A. Parker, and Nicholas S. Souleles. 2006. “Household Expenditure and the
Income Tax Rebates of 2001.” American Economic Review 96 (5): 1589–1610.
Joseph Vavra 65
Mian, Atif, and Amir Sufi. 2018. “Finance and Business Cycles: The Credit-Driven Household Demand
Channel.” Journal of Economic Perspectives 32 (3): 31–58.
Parker, Jonathan A., Nicholas S. Souleles, David S. Johnson, and Robert McClelland. 2013. “Consumer
Spending and the Economic Stimulus Payments of 2008.” American Economic Review 103 (6):
2530–53.
Stroebel, Johannes, and Joseph Vavra. 2019. “House Prices, Local Demand and Retail Prices.” Journal of
Political Economy 127 (3).
66 Journal of Economic Perspectives
File and source
- File
- vavra-2021-jep-35-3-47-tracking-the-pandemic-in-real-time.pdf
- Size
- 726,137 bytes
- SHA-256
- 9483d8321a3d25377bcdcc7752ab4437e5dc3d0d565faa60f5fe6a722554471f
- Original
- No public link identified.