The Short-Term Consequences of COVID-19 on the German Labor Market

(1)

University of Groningen

Master Thesis

MSc Economic Development & Globalization

The Short-Term Consequences of COVID-19

on the German Labor Market

Natalia Wiersma

Student number: S3155420

E-mail address: n.wiersma.6@student.rug.nl

Supervisor

Konstantin. M. Wacker, Ph.D. Co-assessor

prof. dr. J. de Haan

(2)

Abstract

In this paper, the short-term consequences of COVID-19 on the German labor market are documented. This relationship is examined by testing whether COVID-19 incidence has an impact on the underemployment in 401 German counties and districts. The main finding reveals that COVID-19 increased the overall underemployment in Germany. The low magnitude of the effect can be explained by Germany’s policies to mitigate the economic impact of COVID-19 on the labor market. One of these policies includes the

short-time work programs, known as Kurzarbeit. Even though there seems to be a

cor-relation between COVID-19 incidence and Kurzarbeit applications, the empirical results of this study reveal there is no evidence to assume a causal relationship when considering month fixed effects. This paper also documents whether the effect of COVID-19 cases on underemployment is stronger for regions specialized in specific industry sectors (pri-mary, secondary, and tertiary). The outcome suggests that the negative impact of COVID-19 on labor market outcomes is larger for regions with higher employment in the tertiary sector. This paper concludes that, while the COVID-19 pandemic will result in unprecedented consequences for the labor market, the severity of these consequences will be asymmetric within Germany.

(3)

1 |

Introduction

The world has been in pandemic mode since the COVID-19 outbreak that marked the beginning of 2020. The World Health Organization (WHO) first declared COVID-19 a world health crisis of international concern in January 2020. Since then, the virus has spread across every continent, except Antarctica. This novel coronavirus disease, known as COVID-19 can cause mild to severe respiratory infections in humans. As of December 2020, the global death toll from the virus has passed 1.5 million (Johns Hopkins Univer-sity, 2020).

Besides the tragic human consequences, the virus has also created enormous uncer-tainty for the global economy, triggering a socio-economic crisis. The crisis has the po-tential to leave longstanding and deep scars for the economy if not countered with ade-quate measures. Governments around the world have taken draconian measures to con-tain the spread of the virus. In particular, social distancing, isolation, quarantine, and lock-down policies play a pivotal role in this. These measures required many businesses to shut down temporarily, leaving many people confined to their homes. As a result, business activity slowed down, strongly disturbing economic life and the labor market. There is no doubt that the COVID-19 pandemic will result in unprecedented consequences for the labor market, but the severity will be highly asymmetric across and within countries. Countries or regions specialized in the services sector (e.g., tourism and restaurants) will be hit particularly hard, due to business closures and lockdown measures caused by the pandemic.

In this paper, the economic consequences of COVID-19 on the labor market in Germany will be explored. This will be done by creating an econometric model that investigates whether the number of COVID-19 cases has an effect on underemployment. The results of this research rely on two main data sources. The Robert Koch Institut, a German research institute and Federal Government Agency, provides a large sample size on the number of confirmed COVID-19 cases for 401 German counties and districts. This data is combined with the monthly underemployment data from the Federal

Em-ployment Agency (Bunderagentur für Arbeit).

(5)

Figure 1— Daily Confirmed COVID-19 Cases per Million People

Notes: Data from John Hopkins University CSSE COVID-19 Data. The graph spans from the 1st_{of February till}

the 26th_{of December 2020. Due to limited testing, actual COVID-19 cases are higher than the confirmed cases}

displayed in the graph.

One might say, by looking at the graph, that Germany learned to contain the spread of the virus when the second wave hit. During the first wave, all countries were hit with approximately the same numbers. The first wave seems to coincide with the underemployment rate fluctuations shown in Figure 2. It shows a sharp increase in the underemployment rate in March, which coincides with the first wave (March-May) of COVID-19 that can be observed again in Figure 1.

The results of this study confirm that there is a significant relationship between COVID-19 and labor market outcomes. Namely, an increase in COVID-19 cases led to an increase of 0.0024 percent in underemployment. The low estimate is likely to be ex-plained by certain policies implemented by Germany to fight the pandemic’s economic effects on the labor market. The pandemic has particularly triggered renewed interest

in short-time work schemes, known as Kurzarbeit which is explained in further detail

in section 3.2. 0 200000 400000 600000 800000

01feb2020 01apr2020 01jun2020 01aug2020 01oct2020 01dec2020

Date

(6)

Figure 2 — Monthly underemployment rates in Germany since December 2019

Notes: Data from Bundesagentur für Arbeit.

Upon closer examination it is found that an increase in COVID-19 cases led to an

in-crease of 0.638 percent in Kurzarbeit applications. However, in this case the model does

not consider time fixed effects.

Afterwards, it is documented whether the effect of COVID-19 cases on underem-ployment is stronger for specific industry sectors (primary, secondary and tertiary). The outcome is that the effect of COVID-19 cases on underemployment becomes more

posi-tive in states (Bundesländer) with higher tertiary employment.

This paper contributes to a relevant and recent study conducted by Eichenbaum et al. (2020) in which the interaction between economic decisions and epidemics is ex-amined. According to the authors, epidemics give rise to negative shifts in both aggregate demand and supply. They argue that policies to reduce the effects of these shifts such as containment measures imply a trade-off between short-term economic consequences and health outcomes of an epidemic. This paper adds to this literature by addressing the short-term economic consequences on the German labor market and reporting the heterogeneous impacts of COVID-19 by sector, using data on COVID-19 cases and un-deremployment. Moreover, this paper builds on the growing literature on the effect of pandemics on labor market outcomes and consumer behavior such as McKibben and Sidorenko (2006); Keogh-Brown et al. (2010); Santos et al. (2013); Baker et al. (2020); Barro et al (2020). 6.5 7 7.5 8 U nde re m pl oym ent ra te (%)

01dec2019 01feb2020 01apr2020 01jun2020 01aug2020 01oct2020 01dec2020

(7)

The structure of the paper is as follows: Section 2 examines the possible channels through which a pandemic can affect labor market outcomes and the heterogeneous effects of pandemics by industry sector. Sector 3 introduces the data and how it was collected. The empirical strategy is discussed in section 4. Finally, the results and con-clusions are presented in section 5 and 6, respectively.

2 |

Literature Review

The literature on the aggregate effects of pandemics is increasing at a fast pace, in large part due to the need to estimate the economic implications of the COVID-19 pandemic. The following part will give some insight on how a pandemic translates into outcomes for the labor market.

2.1 Macroeconomics of pandemics

Since most pandemics throughout history have had negative consequences for the econ-omy, the COVID-19 pandemic will be no different. Eichenbaum et al. (2020) study the interaction between epidemics and economic decisions. According to their model, an epidemic has aggregate supply as well as aggregate demand effects and it will affect the labor market through both channels. This paper considers the number of COVID-19 cases to be the key force in driving supply and demand. National and regional policies (lockdowns) will be considered shortly because they are expected to be correlated with the number of COVID-19 cases.

Supply channel

(8)

Particularly because screening and health-care procedures are more advanced now, com-pared to a century ago. For this reason, the change in labor supply is probably limited.

Another factor that McKibben and Sidorenko (2006) consider, is that affected family members need to be cared for. According to their viewpoint, this attributes to a reduction in labor supply.

Besides deaths and illness, there are several other reasons why pandemics can cause a change in labor supply. Keogh-Brown et al. (2010) noted the impact of day-care and school closures on the supply of labor. In the absence of a caregiver, working parents will have to stay home or quit jobs because of childcare obligations.

Behavioral effects also play an important role in the reduction of labor supply. For instance, people who have to travel to their workplace have an increased risk of being infected and therefore try to minimize this risk by lowering their labor supply.

This fear of contagionleads to lower labor force participation as a result of avoidance of

close contact with other people (Eichenbaum et al., 2020). In the supply-and-demand model of labor markets, this would mean the labor supply curve shifts to the left in case labor is affected significantly.

One could think about the diverse supply effects for the variable of interest, un-deremployment. Considering the labor curve shifts to the left, people drop out of the labor force. Note that this does not have an effect on the underemployment (rate).

Demand channel

A large part of the literature on epidemics is focused on the influence of epidemics on consumer and firm behavior. A more recent study by Baker et al. (2020) finds that the early stages of consumer spending are characterized by stockpiling, which drives an in-crease in direct demand for goods in the retail sector. While the subsequent weeks are characterized by a decrease in demand as a result of a sharp decline in consumption. An explanation for this is that people will try to reduce exposure to a virus, by decreasing the purchasing and consumption of products and services. As a result, aggregate demand will fall causing a potential self-reinforcing downward spiral in employment, income, and output.

(9)

These effects included a reduction in domestic employment and income. Especially the people employed in service sectors, such as restaurants and hotels, were affected (World Bank, 2014).

Furthermore, in order to contain the virus, lock-down measures can be imple-mented by governments instructing the majority of businesses to close. As a result, some firms will have to lay off existing workers to decrease their costs. The anticipation that a pandemic will result in a recession also influences firm decisions to hire a lower quantity of labor resulting in lower demand on the labor market. Looking again at the supply-and-demand model of the labor market, this would mean the labor demand curve shifts to the left.

The combination of these effects reduces the severity of the epidemic; however, it exacerbates the magnitude of the recession (Eichenbaum et al., 2020).

For convenience, it is assumed that supply and demand shocks are determined by the number of COVID-19 cases. In order to investigate whether in the short term the COVID-19 virus will lead to increased underemployment in Germany, the following hy-pothesis is analyzed:

2.2 The heterogeneous effect of COVID-19

Depending on the economic activities people are engaged in, the probabilities of getting infected with the virus will differ. The shock of COVID-19 will be a negative externality no matter what, but the extent to which sectors are impacted will differ. A recent study by Del Rio-Chanona et al. (2020) studied the supply and demand shocks faced by dif-ferent industries as a result of pandemics. They find that at the industry level, (tertiary) sectors relating to tourism, entertainment, and restaurants experience large supply and demand shocks due to pandemics. While sectors such as mining, and manufacturing (primary and secondary) particularly experience larger supply shocks compared to de-mand shocks. (Del Rio-Chanona et al. 2020).

It is no surprise that regions specialized in the tertiary sector will be hit particularly hard. These sectors often require customers and providers to be together (e.g., restau-rants) and this will be disrupted by measures to contain the spread of the virus. Not

Hypothesis 1:

(10)

only will the impact be larger for contact-intensive services, but it is also very likely that the effect in these regions will be more persistent because of several reasons. First, as it is necessary to keep a safe distance, the size of restaurants and shops is a confining capacity constraint when opening up again. Complying with the safety requirements for these contact-intensive services brings along high investments for sectors that have been impacted the hardest. Second, the labor in non-routine jobs is more difficult to substitute for capital, i.e., robotization. This is because the tasks associated with these kinds of jobs include interpersonal interaction, flexibility, adaptability, and problem solving, which require human capabilities (Autor, 2015).

Furthermore, some sectors will be affected, but to a lesser extent. Particularly communication and information activities, considering it will be easier for these workers to move to telework. However, teleworking or working from home is not possible for all occupations. Sanchez et al. (2020) introduces new estimates of the share of jobs that can be carried out at home. They claim that on a global scale, one of every five jobs can be performed from home. Teleworking and income are correlated which explains why one of every 26 jobs can be done from home in low-income countries and in high-income countries one of every three jobs. For that reason, poor countries will have to carry a larger burden of COVID-19 because a smaller share of workers is able to work from home and because social safety nets are less advanced or lacking altogether. Sanchez et al. (2020) conclude that especially young, poorly educated workers will be more exposed to the labor market shocks of COVID-19 as the likelihood for them to work from home is the lowest.

Not all sectors are losing as a consequence of the virus. In fact, some sectors are benefitting from the situation, such as the digital sector. Companies like Google, Ama-zon, Netflix, and Apple have seen substantial increases in their profits as a result of shifts from face-to-face services to online services (Anderson et al. 2020).

To sum up, the sectoral composition of a country will determine the asymmetric effect the shock can have on the labor market. In order to empirically verify whether the tertiary sector is indeed impacted the most, the following hypothesis is to be investi-gated:

Hypothesis 2:

(11)

3 |

Data Collection

In this section, I will elaborate on how the data on COVID-19 incidence and labor market outcomes were collected and how they differ geographically and over time. The analysis of this study will cover an 8-month time frame, starting with the month of February 2020, when the first corona cases were confirmed. The following paragraph will familiar-ize the reader with some general employment statistics. Thereafter, the sources that were used to collect data are presented.

3.1 Employment

According to the Federal Statistical Office (Destatis), approximately 44.8 million Ger-man residents were employed in October 2020 – somewhat more than half of its popula-tion (83 million in 2020). Approximately ¾ of the working populapopula-tion is employed in the

services sector, with expectations of continuing growth.1_{On the contrary, less and less}

people are employed in the manufacturing sector. Jobs in agriculture only represent little over 1 percent (Destatis, 2020).

Furthermore, the number of persons in employment is considerably below the pre-pandemic level (Figure 3). In comparison with October 2019, persons in employment decreased by 1.3 percent. The upward trend in number of persons in employment ended with the COVID-19 pandemic. The consequences of the pandemic on employment can be seen in Figure 3 starting from March 2020 onwards. Since May 2020 the year-on-year decrease (green bar) in number of persons in employment remained at a similar negative level. Note that these results count short-time workers as employed instead of unem-ployed.

In October 2020, 1.93 million people were unemployed according to the

calcula-tions based on the German Labor Force Survey.2_{Moreover, the unemployment rate}

reached 4.4 percent in October 2020 (Destatis, 2020)

1_{Berlin is the state with the highest share of jobs in the services sector. The state with the lowest share} in services is Baden-Württemberg

(12)

Figure 3 — Persons in Employment January-October 2020

Notes: Data from Statistisches Bundesamt (Destatis), 2020. The horizontal axis shows the date. The vertical axis shows the percentage change in employed persons. Numbers in the figure are based on persons with place of residence in Germany.

3.2 COVID-19 cases

The reported numbers on confirmed cases are collected from the Robert Koch Institut (RKI), a German research institute and Federal Government Agency. The Robert Koch Institut is responsible for monitoring and evaluating the situation around diseases, such as COVID-19, and provides recommendations to health professionals. Disclosed under the German Protection against Infection Act (IfSG), the institute provides a database of notifiable diseases and corresponding cases. The IfSG states that the Robert Koch Institut obtains data on notifiable diseases through state and local health departments.

The Robert Koch Institut offers an application, SurvStat@RKI 2.0, that makes it possible to retrieve aggregated data from the system database and to produce graphic

displays.3_{The data on the absolute number of cases was collected by creating a query}

according to the desired needs of this study.4_{Important to keep in mind is that the case}

3_{Appendix A illustrates the maps that are produced as a result of the created query. It can be observed} that the north-eastern parts of Germany reported lower incidence rates compared to the south-western parts.

4_{See https://survstat.rki.de/Content/Query/Create.aspx for the application offered by the Robert Koch} Institute -1500 -1000 -500 0 500 1000 1500 2000

Jan-18 Jul-18 Jan-19 Jul-19 Jan-20 Jul-20

Month

Change on the same month a year earlier

(13)

numbers do not say anything about the place of infection. Instead, it reveals the place of residence of the infected person. The query displays a selection of diseases, time peri-ods, and places. To be more specific, the COVID-19 disease was selected, starting at the

6th_{calendar week (2}nd_{of February - 8}th_{of February) of the year 2020.}5_{Moreover, this}

study considers COVID-19 disease based on three different levels of NUTS. The NUTS (Nomenclature of territorial units for statistics) categorization is used as a hierarchical system to divide the economic territory. Germany is divided into three different levels.

NUTS 1, 2, and 3, respectively moving from states (Länder), to government regions

(Regierungsbezirke), and counties or districts (Kreise and kreisfreie Städte).

The data from SurvStat@RKI 2.0 is updated weekly or monthly depending on disease or pathogen (e.g., Syphilis, HIV) (Robert Koch Institute, 2020)

3.3 Underemployment, Kurzarbeit and Tertiary Sector

The data on the labor market conditions are retrieved from the Federal Employment

Agency (Bunderagentur für Arbeit). More specifically, data on underemployment is used,

which is unemployment plus persons in specific employability programs (excluding

short-time work).6_{Thus, this concept of underemployment shows people who are unemployed,}

looking for work, and also people who already found their way into subsidized employ-ment. At the district level, short-time work cannot be included in underemployment because short-time work cannot be assigned to the place of residence but only to the place of work. The data presents the underemployment numbers for the NUTS 3 level (401 counties and urban districts). Moreover, the underemployment data is given for each month, but this study will only consider the data from February to September 2020.

Furthermore, as mentioned at the beginning of this paper, the pandemic has

triggered renewed interest in short-time work schemes, known as Kurzarbeit. This is a

state-funded social insurance program that allows companies to reduce their employees’

5_{The query uses the time format “season year and week 27”, i.e., the first season week is the 27}th_week of the calendar year. The fact that they start counting in the middle of the year could have to do with their accounting cycle however, this is unclear from the information provided. For convenience, the text refers to the first season week as the 1st_{week of the calendar year.}

(14)

working hours to discourage layoffs.7_{During the turbulent times of the global financial}

crisis of 2007-2008, these short-time work programs were instrumental in keeping the country’s labor market stable (IMF, 2020). This paper will take Kurzarbeit into account and data on this is also retrieved from the Federal Employment Agency. The Federal Employment Agency discloses the data on Kurzarbeit at NUTS 1 and NUTS 3 level, and both will be investigated. At NUTS 3 level, the Kurzarbeit has only been reported for the months from February till May 2020, whereas at NUTS 1 level this data is

available for the months from February till August 2020.8

Besides, data on employment by state (Bundesländer) in primary, secondary, and

tertiary sector is also retrieved from the Federal Employment Agency. These

employ-ment statistics are reported separately, for all 16 states in absolute values.9_According

to the economic view of economic sectors, sectors can be subdivided into primary (agri-culture), secondary (manufacturing), and tertiary sector (services). For the empirical part of this paper, the share of employment in the tertiary sector will be considered. The reason for focusing on the tertiary sector is because this study is interested in whether the effects of COVID-19 on underemployment are stronger in states that have higher employment in the tertiary sector (hypothesis 2).

The Federal Employment Agency provide labor market statistics which are up-dated on a monthly basis (Bundesagentur für Arbeit, 2020)

3.4 Data visualization

In this part, some graphs will be reported that visualize key information and data col-lected from the main data sources. Figure 4 depicts the development of underemploy-ment and 19 cases across all the counties and districts. On average, the COVID-19 cases reach a maximum in the month of April, followed by a steep decline and followed again by an increase. The underemployment trend is lagging a little bit behind the incidence trend, which makes sense since getting sick precedes any labor consequences. Although both trends show a steep increase around February-March, this does not

7_{The government covers 60 percent (67 percent for parents) of employees’ lost income. In other words,} employees will receive 60 percent of their income for hours not worked and they receive the usual wage for hours worked. It is considered a highly sufficient tool to protect workers’ income and to support aggregate demand (IMF, 2020)

8_{The data at the district level has a waiting period of 4 months which explains the limited data} availa-bility.

(15)

necessarily mean that there is a causal relationship. Whether or not this is the case, will be researched with the econometric model in chapter 4 of this paper.

Figure 5 shows the average for confirmed cases and applications for Kur-zarbeit by state over a period of 8 months. Interestingly, it seems that both bars follow the same trend. Nordrhein-Westfalen reported the highest average for COVID-19 cases from February to September. In addition to that, this state also reports the highest average of Kurzarbeit applications. The same goes for Bayern, showing the second-high-est average for COVID-19 cases and Kurzarbeit applications. Figure 6 reinforces the observation that both variables seem to follow the same trend.

Figure 4 — Development of Underemployment and Cases at NUTS 3 level from February-September 2020

Notes: Data from the Robert Koch Institute and the Bundesagentur fur Arbeit. The averages on the y-axis are in absolute values. The numbers on the horizontal axis illustrate the months in which the cases and underemployment were measured. The first vertical axis shows the average underemployment in Germany per month. The second vertical axis shows the average number of confirmed cases in Germany per month. Important to note is that these averages are computed based on a county level.

0 50 100 150 200 250 Ca se s (A ve ra ge ) 8200 8400 8600 8800 9000 9200 U nde re m pl oym ent (A ve ra ge ) 2 4 6 8 10 Month

(16)

Figure 5 — Comparison of Kurzarbeit Applications and Covid-19 Incidence by State

Notes: Data from the Robert Koch Institute and the Bundesagentur fur Arbeit. The horizontal axis illustrates the German States. The vertical axis shows the average numbers of Kurzarbeit applications and COVID-19 cases over a period of 8 months (February-September 2020).

Figure 6 — Development of Kurzarbeit Applications and Cases at NUTS 1 level from

February-September 2020

Notes: Data from the Robert Koch Institute and the Bundesagentur fur Arbeit. The numbers on the horizontal axis illustrate months in which cases and Kurzarbeit applications were measured. The first vertical axis shows average numbers of Kurzarbeit applications per month. The second vertical axis shows the average number of confirmed cases in Germany per month.

0 5,000 10,000 15,000 20,000 25,000 01 S chles wig-H olstei n 02 H ambur g 03 N iede rsachs en 04 Bre men 05 N ordrhe in-W estfa len 06 H esse n 07 Rhe inland-P falz 08 Ba den-W ürttem berg 09 Ba yern 10 S aarla nd 11 Be rlin 12 Bra ndenbur g 13 M eckle nbur g-Vorpom mern 14 S achs en 15 S achs en-A nhalt 16 Thüri ngen

Kurzarbeit applications COVID-19 incidence

0 2000 4000 6000 Ca se s (A ve ra ge ) 0 10000 20000 30000 40000 K urz arbe it a ppl ic at ions (A ve ra ge ) 0 2 4 6 8 Month

(17)

4 |

Methodology

The main aim of this research is to investigate the short-term economic consequences of COVID-19 on the German labor market. To test this relationship, this study uses panel data from the Robert Koch Institute and the German Federal Employment Agency.

In order to combine the data from both sources, the data from the Robert Koch Institute on confirmed cases had to be converted from weeks to months. This way it matches the monthly underemployment numbers from the German Federal Employment

Agency. The data from the query starts with the 1st_{calendar week and ends with the}

53rd_{week of the following calendar year.}10_{Since this paper takes February as a starting}

point for investigating the effect of COVID-19 cases on underemployment, week 6 is the

first week considered. Week 6 (2nd_{of February to 9}th_{of February 2020) and its}

corre-sponding number of cases are assigned to February in this example. A potential concern in merging the data this way is that some weeks begin in one month and end in another.

An example of this is week 14 (29th_{of March to 5}th_{of April 2020). Here, the case numbers}

from March the 30th_{and 31}st_{are assigned to April since week 14 falls in April for the}

most part. Only if the exact numbers for March 30th_{and 31}st_{were known, they could}

have been assigned to March.

Merging COVID-19 incidence and underemployment from 401 counties and urban

districts produces rich panel data with 2,813 observations. Panel data allows us to con-trol for variables that change over time but not across regions (e.g., federal regulations). It can also control for hidden or unobserved variables in case they do not change over time (e.g., cultural differences). In other words, it accounts for this heterogeneity.

Since we are interested in analyzing the impact of variables that change over time, fixed effects are used. The relationship between the independent variable (COVID-19 cases) and the outcome variable (underemployment and Kurzarbeit) is being explored within an entity (counties and districts). Every county or district has its own character-istics which could have an effect on the independent variable. For instance, counties or districts with a bigger share of elderly population could influence the number of con-firmed cases, as older people are more susceptible to becoming sick.

(18)

The following regression equation will be used to test the first hypothesis and there-fore investigates the relationship between the overall underemployment and COVID-19 cases:

ln(𝑢𝑛𝑑𝑒𝑟𝑒𝑚𝑝𝑙𝑜𝑦𝑚𝑒𝑛𝑡)_𝑟,𝑡= 𝛽₁ln(𝑐𝑎𝑠𝑒𝑠)_𝑟,𝑡+ 𝜃_𝑟 + ∆_𝑡+ 𝜀_𝑟,𝑡 (1)

ln(𝑢𝑛𝑑𝑒𝑟𝑒𝑚𝑝𝑙𝑜𝑦𝑚𝑒𝑛𝑡)_𝑟,𝑡 is an economic outcome (log of underemployment) for region r

(counties and districts) and month t. ln(𝑐𝑎𝑠𝑒𝑠)𝑟,𝑡 is the log of the number of confirmed

cases in region r in time t. 𝜃𝑟 and ∆𝑡 are included to represent region and time (month)

fixed effects. Fixed effects are considered because there are individual effects that are unique to each region. These are what we call, unobserved heterogeneity and by not accounting for these we may have a bias in our estimates. Different levels of underem-ployment are not a problem, but it can be if this heterogeneity is somehow correlated to the COVID cases variable. This would result in an omitted variable bias. An omitted variable bias refers to the problem that some variables are useful in the regression but are not included. There are a variety of reasons one could think of, why this is likely to be the case. For example, underemployment in the east of Germany can be higher, people can be more obedient to the state for historical reasons and follow the state rules more strictly, which would lower COVID-19 cases. Thus, if you don’t control for regional fixed effects it could be that you get a biased estimate for the variables included. Therefore, regional and time fixed effects are included in the model. Using regional fixed effects will absorb the initial level of underemployment because regional fixed effects capture every aspect that does not change over time. Finally, the error term (𝜀) is included.

The relationship between Kurzarbeit applications and COVID-19 cases can be

empirically tested with the same model (1) however, the outcome variable log of

under-employment should be replaced by log of Kurzarbeit applications.

(19)

whether the effect of COVID-19 cases on underemployment is stronger for states that have higher employment in tertiary sectors:

ln(𝑢𝑛𝑑𝑒𝑟𝑒𝑚𝑝𝑙𝑜𝑦𝑚𝑒𝑛𝑡)_𝑟 = 𝛽₁ln(𝑐𝑎𝑠𝑒𝑠)_𝑟∗ 𝐸𝑚𝑝𝑙𝑜𝑦𝑚𝑒𝑛𝑡_𝑆𝑒𝑐𝑡𝑜𝑟_𝑆ℎ𝑎𝑟𝑒_𝑟 + 𝜃_𝑟 + 𝜀_𝑟 (2)

The difference between regression equation (2) and equation (1), is that equation (2) drops the t-subscript since the variable does not change over time. In addition, the

independent variable ln(cases) is being interacted with 𝐸𝑚𝑝𝑙𝑜𝑦𝑚𝑒𝑛𝑡_𝑆𝑒𝑐𝑡𝑜𝑟_𝑆ℎ𝑎𝑟𝑒𝑟, an

employment sector share variable. More specifically, it will be interacted with the

em-ployment share in the tertiary sector. Again, 𝜃𝑟 is included to represent region fixed

effects and the error term (𝜀) is included.

As shown above, both models include a logarithm of the variables

underemploy-ment and cases. The logarithmwill be used because each variable seems to have a

right-skewed distribution (also known as positively right-skewed). This implies that the mean is larger than the median and therefore most data fall on the right side of the graph’s peak. Since the data does not follow a normal distribution, taking the log of both variables will transform these variables into more symmetric ones. It is important to use logarith-mic transformations because non-normality would increase the chance of producing skewed errors. When making a prediction, it is desirable to have an error term that is as small as possible. Figure 7 illustrates the transformation that occurs when taking the

logarithm of the independent variable cases.

Figure 7 — Normality of Distribution Transformation

0

.002

.004

.006

0 500 1000 1500 2000 2500

Density kdensity cases

0 .1 .2 .3 .4 0 2 4 6 8

Density kdensity lncases

(20)

In Figure 8 a scatter plot is used to illustrate the correlation between the long differences of underemployment and cases between February and September 2020. The two observations that seem to divert the most from the linear fitted curve, are the cities Berlin and München. A log specification could decrease the influence of those observa-tions. The outliers can cause problems as the presence of them might bias the regression estimates. It is undesirable to have results of the regression analysis depend to a great extent on the values of a few extreme or unusual data points.

Figure 8 — Correlation of the Long Differences of Underemployment and Cases, February-September 2020

Notes: figure illustrates the correlation of the long differences of underemployment and COVID-19 cases between February and September 2020 (i.e., DUnderemployment = Underemployment09 – Underemployment05). A linear fitted curve is added that considers all observations, including the outliers. Figure created based on data from the Robert Koch Institute and the Bundesagentur fur Arbeit.

4 |

Empirical Results

To test for heteroskedasticity in the fixed-effect model, I compute the modified Wald statistic for groupwise heteroskedasticity. The null hypothesis is rejected so it can be concluded that there is a presence of heteroskedasticity (non-constant variance). There-after, the Lagrange-Multiplier test is used to investigate for serial correlation. The null

0 10000 20000 30000 40000 0 1000 2000 3000 DCases

DUnderemployment Fitted values

Berlin

(21)

hypothesis tells us there is no serial correlation. With a p-value of 0.000 the null hypoth-esis is rejected, and it can be concluded that the data has first-order autocorrelation. Now that we are aware of this, the model can be specified accordingly. Since we are dealing with a large number of observations and small t (time) dataset, clustering stand-ard errors can manage both autocorrelation and heteroskedasticity. In other words, the analysis will continue by using robust standard errors.

To confirm my reasoning for using a fixed-effects model we can compare it with the random-effects model. In this specific case, we cannot perform the Hausman test to decide whether to proceed using the fixed effects model or the random-effects model,

since we are using robust standard errors. Therefore, we use the command xtoverid. This

command (test for identifying restrictions) can be used because the random effects esti-mator uses the additional orthogonality conditions that the regressors are uncorrelated

with the group-specific error 𝑢𝑖 (the “random effect”). These additional orthogonality

conditions are overidentifying restrictions and can thus be tested. The null hypothesis is rejected; therefore, the fixed-effect model should be preferred. The result of these tests can be found in Appendix C.

Besides, it needs to be investigated whether time fixed effects should be included.

For this, a joint test (testparm) can be conducted to see if the monthly dummies are

equal to 0. If so, no time fixed effects are needed. The p-value is equal to 0.000 so the null that the coefficients for all years are jointly equal to zero is rejected. As expected, the analysis will continue using time fixed effects.

A possible concern that should be noted, is the presence of a systematic measure-ment error. Some states could have had more progress in testing at earlier stages or they could generally be more proactive in testing compared to other states. Vice versa, less testing could have been present in the earlier stages of the pandemic. There are several reasons why I doubt this is an issue. First, Germany began nationwide COVID-19 testing in contrast to the regional testing done in Italy. Here, they only started testing in the northern provinces. Still, Germany’s 16 federal states are responsible for their own healthcare system, and for that reason they make their own decisions regarding testing. Thus, there could be a different testing pattern going on in Germany. This may be controlled for by adding the monthly fixed effects. Omitting, let’s say “testing progress”,

biases the coefficient of the variable ln(cases). Controlling for monthly fixed effects

(22)

5.1 Main results

Table 1 reports regression equation estimations for Pooled OLS (1), RE model (2), FE model with robust standard errors (3), FE model with month fixed effects (4), FE model

with a 1-month lag of ln(cases), and finally a FE model with an interaction term (6).11

Column 1 reports the effect of ln(cases) on the dependent variable ln(underemployment).

The results from model 3 reveal that ln(cases) has a significant coefficient (at the 1%

significance level) with a negative sign. When adding time fixed effects, the coefficient of ln(cases) increases from -0.0097 to 0.0024. In model 5 a lagged variable is added to the fixed effects model because past cases can also have a potential impact on

underem-ployment. By adding a 1-month lag to the ln(cases) variable the first month (February)

will be lost. The results show that by adding the lagged variable, the effect of COVID-19 cases on underemployment becomes less strong but remains significant at the 5% significance level.

Since the model can be specified as a log-log model, the coefficients represent the

elasticity of ln(underemployment) with respect to ln(cases). In other words, the

coeffi-cient represents the estimated percent change in the dependent variable for a percent change in the independent variable. Another way to put this is that when COVID-19 cases double, the number of underemployed people goes up with the corresponding per-centage times a hundred:

Result model 3: A 100% increase in ln(cases), decreases (ln)underemployment in Germany with 0.91%. This effect is significant at the 1% significance level

Result model 4: A 100% increase in ln(cases), increases (ln)underemployment in Germany with 0.24%. This effect is significant at the 5% significance level.

Result model 5: A 100% increase in ln(cases), increases ln(underemployment) in Germany with 0.17%. This effect is significant at the 5% significance level. The coefficient for each month represents, around the constant, the variation in underemployment by month on the whole German level. It seems that for the last months of the sample, especially June, July, August, and September, the coefficients appear to be higher compared to the first months.

(23)

Table 1 — Effect of log cases on log underemployment at NUTS 3 level (1) (2) (3) (4) (5) (6) Dependent variable: log underemploy-ment

Pooled OLS RE FE FE with

month fixed effect FE with a 1-month lag FE with in-teraction ln(cases) 0.191*** -0.00877*** -0.00907*** 0.00239** 0.00173** -0.00877*** (0.00910) (0.000910) (0.000909) (0.00106) (0.000710) (0.000874) March 0.00635 (0.00747) April 0.0606*** 0.0795*** (0.00821) (0.00815) May 0.0994*** 0.116*** (0.00718) (0.00845) June 0.107*** 0.127*** (0.00685) (0.00793) July 0.118*** 0.141*** (0.00723) (0.00778) August 0.126*** 0.148*** (0.00804) (0.00822) September 0.0936*** 0.114*** ln(cases)_l1 Tertiary × (0.00834) (0.00855) 0.00220*** (0.000840) 0.0505** ln(cases) (0.0236) Observa-tions 2,813 2,813 2,813 2,813 2,375 2,813 R-squared 0.135 0.052 0.616 0.531 0.054 Number of Region_n 401 401 401 401 401 Regional FE

YES YES YES

Month FE YES YES

Cluster- robust standard errors in parentheses

*** p<0.01, ** p<0.05, * p<0.1

(24)

Furthermore, model 6 in Table 1 includes the interaction between the COVID-19 cases and tertiary employment share. The sign of the regression coefficient tells us whether there is a positive or negative relationship between the interaction term and the dependent variable. In this case, the regression coefficient of the interaction term is 0.0505, and the effect is significant at the 5 percent significance level. The interaction variable will describe how much the effect of COVID-19 incidence depends on the initial employment rate in the tertiary sector. The higher the employment in the tertiary sector, the less negative the relationship between underemployment and cases. This result con-firms the expectation that a higher share of employment in the tertiary sector will make the impact of COVID-19 stronger.

Result model 6 (Table 1): The higher a state’s employment in the tertiary sector,

the less negative the effect of ln(cases) on ln(underemployment). The effect

be-comes positive at tertiary rate levels of 0.170 and above.

Model 6 shows the main effect of 19 cases in the first row. The effect of COVID-19 cases is -0.00877 when tertiary rate = 0. The model predicts that at a non-existent level of tertiary employment, COVID-19 decreases underemployment instead of in-creases. As the tertiary employment rate increases, the effect of COVID-19 cases also increases (which means, it becomes less negative). Eventually, at a certain level tertiary

rate, the effect will turn positive. This will happen at rates of 0.170 or above. Since the

lowest tertiary level in this sample is .6376613 for Baden-Wurttemberg, the effect will be positive for all states. In other words, an increase in COVID-19 cases will increase underemployment in all states. The effect will be stronger in states with higher employ-ment rates in the tertiary sector. For example, Berlin is the state with the highest share in the tertiary sector. With a tertiary rate of .8661993. the effect will be 0.03597. This effect is lower but still positive in Baden-Württemberg with a magnitude of 0.02343. Even though the main effect is negative, increasing levels of tertiary rate will shift the effect towards a more positive sign, indicating more underemployment in states with a higher share of tertiary employment.

(25)

Kurzarbeit

The renewed interest in the short-time work programs (Kurzarbeit), as introduced at the beginning of this paper, could be a possible explanation for the low estimates that are presented in Table 1. It is important to note that short-time work is not accounted for in the underemployment data. Consequently, these short-time work programs offer an interesting follow-up in the analysis of this paper. For that reason, the following

results consider ln(Kurzarbeit) instead of ln(underemployment) as the dependent

varia-ble.

Table 2 and 3 report regression equation estimations for the FE model with robust standard errors (1) and FE with month fixed effects (2) at NUTS level 3 and NUTS

level 1 respectively. Column 1 reports the effect of the predictor variable ln(cases) on

the dependent variable (ln)Kurzarbeit. Table 2 reports that ln(cases) has a significant

coefficient at the 1% significance level. The coefficient decreases from 0.280 to 0.007 and

becomes insignificant when adding the monthly fixed effects. Furthermore, results from

Table 3 model 1 reveal that ln(cases) has a significant coefficient with a positive sign at

the 1% significance level. When adding time fixed effects, the coefficient of cases de-creases from 0.638 to 0.068, becoming statistically insignificant. In sum:

Result model 1 (Table 2): A 100% increase in (ln)cases, increases (ln)Kurzarbeit

in Germany with 28%. This effect is significant at the 1% significance level.12

in Germany with 0.74%.

in Germany with 63.8%. This effect is significant at the 1% significance level.

in Germany with 6.81%.

(26)

Table 2 — Effect of COVID-19 cases on Applications for Kurzarbeit at NUTS 3 level

(1) (2)

Dependent variable: log of applications for short-time

work (Kurzarbeit)

FE FE with month fixed effect

ln(cases) 0.280*** 0.00737 (0.0336) (0.0159) March 3.249*** (0.173) April 4.101*** (0.176) June 4.062*** (0.168) Observations 1,244 1,244 R-squared 0.200 0.915 Number of Region_n 401 401 Regional FE YES Month FE YES

Cluster robust standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.1

Notes: Data from the Robert Koch Institute and Bundesagentur fur Arbeit. Robust standard errors are in parenthe-ses, adjusted for clustering by county. The dependent variable in the top panel is a log specification of the variable underemployment. In column 1, 2, 3, and 4 the econometric models are pooled OLS, random effects model, fixed effects model and fixed effects model with month fixed effects, respectively. Column 3 and 4 include regional fixed effects. Column 4 includes both regional and time fixed effects. Ln(cases) is a variable that takes the logarithm of the absolute number of confirmed COVID-19 cases. The time period is February-September 2020.

(27)

Table 3 — Effect of log Cases on log Kurzarbeit at NUTS 1 level

(3) (4)

Dependent variable: log of applications for short-time

work (Kurzarbeit)

FE FE with month fixed effect

ln(cases) 0.638*** 0.0681 (0.0428) (0.0695) February -0.517* (0.245) March 3.758*** (0.615) April 5.078*** (0.696) May 2.944*** (0.610) June 1.908*** (0.549) July 1.221* (0.581) August 0.921 (0.616) Observations 112 112 R-squared 0.602 0.973 Number of States 16 16 Regional FE YES Month FE YES

Cluster robust standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.1

(28)

6 |

Conclusion, Discussion, and Limitations

6.1 Conclusion and discussion

As the COVID-19 virus has spread across the globe, it has become indisputable that the pandemic has the potential to derail the world economy. As of today, the size and per-sistence of the economic impact is still not known for certain. Instead, it is safe to say that the impact of the pandemic will be highly asymmetric across and within countries. In this paper the question of whether COVID-19 incidence has an impact on underemployment has been addressed empirically for 401 German counties and districts between February and September 2020. The data on COVID-19 cases has been retrieved from the Robert Koch Institut. This data has been combined with data on

underem-ployment and emunderem-ployment sectors from the Federal Emunderem-ployment Agency (

Bunderagen-tur für Arbeit). The main finding reveals there is empirical support for the assumption that an increase in COVID-19 cases leads to an increase in underemployment. Results vary when using different estimators (OLS, RE, FE). It can also be observed that the main model (FE with fixed effects), is sensitive to the inclusion of the 1-month lag in the independent variable. Including the past cases, decreases the positive effect.

This analysis also reports the heterogeneous effects of cases across different

in-dustry sectors. The results confirm findings by Del Rio-Chanona et al. (2020) that the

economic shocks from pandemics have a larger impact on activities in the service sector. The findings suggest that the effect of COVID-19 cases on underemployment becomes more positive at increasing levels of tertiary rates. In other words, states with higher rates of employment in the tertiary sector will experience higher underemployment at increasing levels of COVID-19 incidence. A possible explanation for this finding is the relationship between containment measures being implemented to contain the virus and

the closure of businesses. Particularly businesses that require intensive contact will be

impacted. These businesses are often instructed to close or to make significant changes to their day-to-day operations in order to maintain distance.

Whether the low estimates for the fixed-effect model can be explained by the

German’s Kurzarbeit cannot be ruled out. Even though there seems to be a correlation

(29)

A debate has emerged about whether the lock-down measures being implemented have caused more harm than good. Policymakers should recognize the trade-offs between disease prevention and employment. Employment losses are mostly concentrated in sec-tors directly affected by restricting measures such as lockdowns. The findings in this paper highlight those sectors that are most in need of assistance. Governments should try to help those workers most affected.

6.2 Limitations and further research

Reflecting back on the analysis done, there are several limitations. The first limitation relates to the time span of this study. The study uses data from February to September 2020. It could be argued that the effect of COVID-19 cases might not reveal itself very

rapidly in underemployment numbers. This could also be a possible explanation for the

low estimates of the models. In addition, the data of this study was collected in Septem-ber 2020. Since then, the impact of COVID-19 on the economy has become way more severe, and for that reason, a stronger effect could be expected when analyzing a more extensive period.

Moreover, one could argue that the effects in the model only happen in the region that is being investigated. The model assumes that each county or district’s error term is independent of each other. However, this is deemed unlikely since labor markets are generally related between districts. A way to correct for this is to cluster the standard error at the NUTS 2 (state) level, instead of the NUTS 3 (county and district) level. By constructing a new set of standard errors, the error term of each district within a state could be correlated.

The third limitation concerns the dependent variable underemployment. The

var-iable likely follows a dynamic process. The demand for labor may depend on past reali-zations of labor input (i.e., present demand depends on past demand). The current value of labor demand depends on its past value in addition to exogenous explanatory varia-bles. Therefore, it can be said that the employment structure follows a dynamic process. This could be a concern for the empirical finding and should be investigated in future research.

(30)

county-specific testing patterns vary over time, and therefore a county-specific and random pattern remains. This could be dealt with by adding state or county-specific time trends since these specific time trends would take out all the increase in underemployment over time. For example, if underemployment in Bavaria increases with 5 percent per month, then around that trend the fluctuations you are left with are to be explained with the COVID-19 incidence variable.

(31)

Appendix A

Map 1: SurvStat@RKI 2.0 - Incidence rate COVID-19 States, Germany, 2020

(32)

Appendix B

Table 4 — Summary Statistics

Variable Obs Mean S.D. Min Max

(33)

Appendix C

. xttest3

Modified Wald test for groupwise heteroskedasticity in fixed effect regression model

H0: sigma(i)^2 = sigma^2 for all i chi2 (401) = 51461.66

Prob>chi2 = 0.0000 . xtoverid

Test of overidentifying restrictions: fixed vs random effects

Cross-section time-series model: xtreg re robust cluster(Region_n) Sargan-Hansen statistic 177.921 Chi-sq(1) P-value = 0.0000

(34)

Bibliography

Andersen, A. L., Hansen, E. T., Johannesen, N., & Sheridan, A. (2020). Consumer responses to the COVID-19 crisis: Evidence from bank account transaction

data. Covid Economics, 88-115.

Autor, D. H. (2015). Why Are There Still So Many Jobs? The History and Future of

Workplace Automation. Journal of Economic Perspectives, 3-30.

Bundesagentur für Arbeit. (2020). Arbeitslosigkeit und Unterbeschäftigung. Retrieved

from Statistik Arbeitsagentur:

https://statistik.arbeitsagentur.de/DE/Navigation/Grundlagen/Definitionen/Ar

beitslosigkeit-Unterbeschaeftigung/Arbeitslosigkeit-Unterbeschaeftigung-Nav.html

Destatis. (2020). Labour Market Employment Statistisches Bundesamt. Retrieved from

destatis.de: https://www.destatis.de/EN/Themes/Labour/Labour-Market/Employment/_node.html

Baker, Farrokhina, R., S., Meyer, S., Pagel, M., & Yannelis, C. (2020). How Does Household Spending Respond to an Epidemic? Consumption During the 2020

COVID-19 Pandemic. NBER Working Paper, 26949.

IMF. (2020). Kurzarbeit: Germany’s Short-Time Work Benefit. Retrieved from

imf.org: https://www.imf.org/en/News/Articles/2020/06/11/na061120-kurzarbeit-germanys-short-time-work-benefit

John Hopkins University. (2020, February 19). Mapping COVID-19. Retrieved from

ARCgis: https://coronavirus.jhu.edu/map.html

Keogh‐Brown, M. R., Wren‐Lewis, S., Edmunds, J. W., Beutels, P., & Smith, D. R. (2010). The macroeconomic impact of pandemic influenza: estimates from

models of the United Kingdom, France, Belgium and The Netherlands. The

European Journal of Health Economics , 11(6):543-54.

McKibben, W., & Sidorenko, A. (2006). Global Macroeconomic Consequences of

Pandemic Influenza. Sydney, Australia: Lowy Institute for International Policy. del Rio-Chanona, M. R., Pichler, A., Lafond, F., Farmer, D. J., & Mealy, P. (2020).

Supply and demand shocks in the COVID-19 pandemic: An industry and

occupation perspective. Covid Economics, 65-103.

Robert Koch Institut. (2020). COVID-19 (Coronavirus SARS-CoV-2). Retrieved from

(35)

https://www.rki.de/DE/Content/InfAZ/N/Neuartiges_Coronavirus/nCoV.htm l

SurvStat@RKI 2.0. (2020). SurvStat@RKI 2.0. Retrieved from survstat.rki.de:

https://survstat.rki.de/Content/Query/Create.aspx

World Bank. (2014). The economic impact of the 2014 Ebola epidemic: short and