B 108 - 2021-22 - Endeligt svar på spørgsmål 1: Spm. om beregninger af effekten af digitale samtaler fra første dag på baggrund af de opdaterede regneprincipper, til beskæftigelsesministeren

Beskæftigelsesudvalget 2021-22
B 108
Offentligt

IFN Working Paper No. 1307, 2019

Does Job Search Assistance Reduce

Unemployment? Experimental Evidence on

Displacement Effects and Mechanisms

Maria Cheung, Johan Egebark, Anders Forslund,

Lisa Laun, Magnus Rödin and Johan Vikström

Research Institute of Industrial Economics

P.O. Box 55665

SE-102 15 Stockholm, Sweden

[email protected]

www.ifn.se

Electronic copy available at: https://ssrn.com/abstract=3515935

B 108 - 2021-22 - Endeligt svar på spørgsmål 1: Spm. om beregninger af effekten af digitale samtaler fra første dag på baggrund af de opdaterede regneprincipper, til beskæftigelsesministeren

Does job search assistance reduce unemployment?

Experimental evidence on displacement eﬀects and

mechanisms

∗

Maria Cheung

Johan Egebark

Anders Forslund

Lisa Laun

Magnus R¨din

Johan Vikstr¨m

November 20, 2019

Abstract

This paper uses a large-scale two-level randomized experiment to study direct

and displacement eﬀects of job search assistance. Our ﬁndings show that the as-

sistance reduces unemployment among the treated, but also creates substantial

displacement leading to higher unemployment for the non-treated. By using de-

tailed information on caseworker and job seeker behavior we show that vacancy

referrals passed on from caseworkers to job seekers is the driving mechanism be-

hind the positive direct eﬀect. We also examine explanations for the displacement

eﬀect and show that displacement is not due to constrained resources, but arises

in the labor market. A comparison between diﬀerent meeting formats suggests

that face-to-face meetings and distance meetings are more eﬀective than group

meetings. Despite the existence of displacement eﬀects, when we incorporate our

results into an equilibrium search model we ﬁnd that a complete roll-out of the

program would lead to lower unemployment and reduced government spending.

Keywords: vacancy referrals, counseling, job search, randomized experiment.

JEL codes: J68, J64, C93

We are grateful for helpful suggestions from Pierre Cahuc, Bruno Cr´pon, Marie Gartell, Lena

Hensvik, Paul Muller, Knut Røed, Michael Rosholm, Arne Uhlendorﬀ, Gerard van den Berg, Aico

van Vuuren, Olof ˚slund, and seminar participants at CREST, IFAU, LISER, Stanford University,

University of Utah, CAFE workshop, CREST/IZA/OECD conference, IIPF conference, IZA World

Labor Conference, Nordic conference on register data and economic modelling, SOLE conference, and

the Swedish PES conference on Labour Market Policies. The experiment was conducted using support

from the European Commission. Vikstr¨m, Laun and Forslund acknowledge support from FORTE

(2016-00886). Egebark acknowledges support from Handelsbanken’s research foundations.

Swedish Public Employment Service.

Swedish Public Employment Service and Research Institute of Industrial Economics (IFN), Stock-

holm.

IFAU and UCLS, Uppsala University.

IFAU.

IFAU and UCLS, Uppsala University. Corresponding Author: IFAU, Box 513, S-751 20 Uppsala,

Sweden. E-mail: [email protected].

∗

Electronic copy available at: https://ssrn.com/abstract=3515935

B 108 - 2021-22 - Endeligt svar på spørgsmål 1: Spm. om beregninger af effekten af digitale samtaler fra første dag på baggrund af de opdaterede regneprincipper, til beskæftigelsesministeren

Introduction

Job search assistance (JSA), aimed at helping job seekers search for jobs more eﬃciently,

constitutes an important component of active labor market programs (ALMPs) in many

OECD countries. This widespread use of JSA as a policy instrument raises the impor-

tant question to what extent such programs eﬀectively reduce unemployment. Previous

evidence suggests that job search assistance is one of the most powerful tools in the

ALMP toolbox. As summarized in Card et al. (2010, 2017), both observational studies

and randomized experiments from a wide range of countries document generally posi-

tive impacts of JSA.

Despite the numerous previous studies, there are still important

questions where evidence is lacking. Few studies have been able to explain why job

search assistance is comparably eﬀective. In particular, little is known about how JSA

aﬀects caseworker and job seeker behavior and how this operates to the positive em-

ployment eﬀects. This is unfortunate since a deeper understanding of the mechanisms

may help to ﬁne-tune programs and improve their eﬃciency. In addition, one concern

is that the positive results for those who participate in JSA policies reﬂect negative

impacts on non-treated job seekers (Cr´pon et al., 2013; Ferracci et al., 2014; Gautier

et al., 2018). Since such displacement eﬀects, which mainly represent a re-ordering of

job queues, have important implications for the overall eﬀectiveness of JSA programs,

more evidence is needed on their size and origin.

This paper contributes to the literature on JSA in several ways. We present evidence

from a large-scale two-level randomized experiment designed to detect both direct and

displacement eﬀects. We exploit rich information on caseworker actions and job seeker

search behavior, which allow us to study mechanisms in a number of dimensions that,

to the best of our knowledge, have not been studied before. We also analyze the

The evidence from experiments includes for instance Gorter and Kalb (1996), Dolton and O’Neill

(1996), Dolton and O’Neill (2002), van den Berg and van der Klauuw (2006), H¨gglund (2011),

Graversen and van Ours (2008a), Graversen and van Ours (2008b), Cr´pon et al. (2013), Arni (2015)

and Maibom et al. (2017). Two recent US studies include McConnell et al. (2016) and Manoli et al.

(2018).

Electronic copy available at: https://ssrn.com/abstract=3515935

B 108 - 2021-22 - Endeligt svar på spørgsmål 1: Spm. om beregninger af effekten af digitale samtaler fra første dag på baggrund af de opdaterede regneprincipper, til beskæftigelsesministeren

origin of any displacement by contrasting displacement due to resource constraints and

displacement of jobs. Comparing three diﬀerent meeting formats further adds to the

understanding of the mechanisms. Finally, we investigate the potential impact of a

full-scale roll-out of the program, by developing and estimating an equilibrium search

model using the results from the experiment.

The experiment, conducted in 2015, consisted of more frequent meetings with a

caseworker during the ﬁrst quarter of unemployment and included randomization of

treatment both across and within local employment oﬃces. It targeted newly unem-

ployed job seekers at 72 PES oﬃces in Sweden (one quarter of all oﬃces), where each

oﬃce typically served one entire local labor market. We randomly selected 36 of the 72

oﬃces to provide the JSA program and within these oﬃces job seekers were randomly

assigned to the JSA program. The randomization at two levels allows us to credibly

estimate the overall eﬀect of JSA. Whereas the direct eﬀect is captured by compar-

ing the treated and the non-treated at the active oﬃces, displacement is captured by

comparing the non-treated at the active and the non-active oﬃces.

In line with the previous literature, we ﬁnd that JSA reduces unemployment among

the treated. This raises the question what explains the eﬀectiveness of job search

assistance. To study the mechanisms in detail, we use administrative data to follow

the actions caseworkers take during the program, and pair this with information on

search behavior obtained from monthly activity reports submitted by the job seekers.

By exploiting data on vacancy referrals for both caseworkers and job seekers, we show

that the key mechanism behind the positive eﬀects of the JSA program is an increased

number of vacancy referrals passed on from the caseworkers to the job seekers and

an accompanying increase in the number of referrals the job seekers apply to. Simply

put, the caseworkers use their expertise to ﬁnd and point job seekers to suitable job

openings, and the job seekers take advantage of this information by applying to the

jobs they are referred to. We also show that the increased number of vacancy referrals

Electronic copy available at: https://ssrn.com/abstract=3515935

B 108 - 2021-22 - Endeligt svar på spørgsmål 1: Spm. om beregninger af effekten af digitale samtaler fra første dag på baggrund af de opdaterede regneprincipper, til beskæftigelsesministeren

does not lead to broader search in terms of occupations or geographical area, but rather

streamlines the search process by helping the treated workers apply to the most relevant

jobs early in the unemployment spell.

We have also examined competing explanations. First, using data on all regis-

tered violations of the job search rules along with information on job applications, we

ﬁnd no evidence that caseworkers increase the monitoring of the job seekers and no

corresponding impact on job seekers’ search eﬀort. Second, using information from in-

dividual action plans and program participation data along with information on search

activities, we do not ﬁnd that the treated job seekers receive more job search training

and support, and we see no impact on job seekers’ search strategies.

Besides facilitating eﬃcient policy, this new evidence on mechanisms ﬁlls an impor-

tant gap in the literature. Previous evaluations of JSA polices have contributed to the

understanding of mechanisms by providing results for diﬀerent types of interventions

and target populations. For instance, Meyer (1995) and Ashenfelter et al. (2005) com-

pared diﬀerent policies, and found that assistance combined with monitoring produced

desired results, whereas monitoring alone did not. However, there is less evidence show-

ing in which way a speciﬁc intervention alters caseworker and job seeker behavior, and

how such changes translate into positive employment eﬀects.

In addition to the positive direct eﬀects, we ﬁnd that the JSA program creates

substantial displacement leading to higher unemployment for the non-treated. While

the exit rate from unemployment for treated job seekers increases by 4.6 percent it

falls by 3.8 percent for the non-treated, implying a smaller, although still positive,

overall employment eﬀect. We see no impact on wages, but indications of an increase

in the number of posted vacancies in the local labor market. All this suggests that

This adds to other studies on the role of vacancy referrals in the job search process. Examples

include Van den Berg et al. (2019), Fougere et al. (2009), Engstr¨m et al. (2012) and Bollens and

Cockx (2017).

A recent exception is Arni (2015), who combines a randomized experiment with survey data to

study behavioral mechanisms of an intensive counseling program targeted at older job seekers at two

PES agencies in north-western Switzerland.

Electronic copy available at: https://ssrn.com/abstract=3515935

B 108 - 2021-22 - Endeligt svar på spørgsmål 1: Spm. om beregninger af effekten af digitale samtaler fra første dag på baggrund af de opdaterede regneprincipper, til beskæftigelsesministeren

JSA is associated with equilibrium eﬀects, which imply that the eﬀectiveness of JSA

documented in many previous studies is exaggerated. These results are consistent with

the results from a small but growing literature on displacement eﬀects.

Most notably,

Cr´pon et al. (2013) provide novel experimental evidence of JSA provided to young,

long-term unemployed college graduates in France. They ﬁnd that the positive eﬀects

for the treated were smaller than the negative displacement eﬀects for the non-treated,

suggesting that more jobs were lost than found. We ﬁnd more positive results in our

experiment that targeted a more general group of all newly unemployed job seekers

and oﬀered JSA earlier in the unemployment spell. This suggests that the setting

is important for the overall assessment of JSA policies. Other recent studies ﬁnding

evidence of displacement eﬀects include Ferracci et al. (2014), who develop methods to

study displacement eﬀects with non-experimental data, and Gautier et al. (2018), who

use an equilibrium search model to study JSA in Denmark.

We complete the analysis on displacement eﬀects by presenting evidence that dis-

criminates between displacement due to resource constraints and displacement in the

labor market. Separating between these two sources is central since the policy impli-

cations are diﬀerent. While displacement due to resource constraints can be avoided

by carefully considering the funding arrangements, displacement in the labor market

is more challenging to address with policy arrangements. Still, this division has not

been analyzed before. In our experiment, the intention was to give more assistance to

the treated with unchanged support to the non-treated. However, since it is diﬃcult

in practice to control every feature of a policy program we cannot automatically rule

out displacement of resources. By using information on resource allocation at the local

oﬃce level, we show empirically that there is no crowding out of resources. Instead,

we document substantially larger displacement eﬀects in weak labor markets compared

Earlier evidence on displacement include Blundell et al. (2004), Pallais (2014), Lalive et al. (2015)

as well as Dahlberg and Forslund (2005) and Albrecht et al. (2009) for Sweden. Previous studies with

similar two-level randomization designs as we use include Miguel and Kremer (2004), Banerjee et al.

(2010) and Cr´pon et al. (2013).

Electronic copy available at: https://ssrn.com/abstract=3515935

B 108 - 2021-22 - Endeligt svar på spørgsmål 1: Spm. om beregninger af effekten af digitale samtaler fra første dag på baggrund af de opdaterede regneprincipper, til beskæftigelsesministeren

to labor markets with many job openings. Taken together this suggests that the size-

able displacement eﬀects are due to displacement of jobs. It also adds to the existing

evidence that displacement can be limited under good labor market conditions.

To further contribute to the understanding of the mechanisms, we analyze whether

the impact of JSA depends on the meeting format. Here, we exploit that the 36 active

oﬃces were randomly assigned to provide face-to-face meetings with a caseworker,

distance meetings using the internet or telephone, or group meetings. Ultimately, policy

makers are looking to allocate resources to interventions with the largest potential. We

ﬁnd positive employment eﬀects for face-to-face and distance meetings, but not for

group meetings.

This ﬁnding further supports the vacancy referral mechanism, since

we see an increase in referrals for face-to-face and distance meetings, but not for group

meetings. Our interpretation is that, since group meetings involve support to several

job seekers at the same time, caseworkers are unable to discuss vacancies with each

participant.

The evidence on the displacement of jobs implies that JSA programs create search

externalities. Since the size of the externalities depends on the share of program partic-

ipants, the reduced form estimates alone are insuﬃcient to study the implications of a

full-scale roll-out of the program. Such an assessment can be done by incorporating the

estimated responses into a structural model. To do this, we build upon the Diamond-

Mortensen-Pissarides (DMP) model in Gautier et al. (2018), which was designed to

study the equilibrium eﬀects of a Danish JSA program.

A key feature of the model

is the endogenous matching function, which speciﬁes that the success of an application

Maibom et al. (2017) also ﬁnd that face-to-face meetings outperform group meetings. However,

in their case each treatment was given in only one region, whereas, in this paper, we have a design

that explicitly allows for inter-treatment comparisons. Another recent study is Cr´pon et al. (2015),

which ﬁnds positive employment eﬀects of frequent group meetings in the form of search clubs with

meetings several times a week.

Gautier et al. (2018) also study displacement using a randomized trial, but in this case there is no

randomization over local oﬃces, only over unemployed individuals in two non-random Danish regions.

With their equilibrium search model, they conclude that increasing the share of treated will raise

equilibrium unemployment and decrease welfare.

Electronic copy available at: https://ssrn.com/abstract=3515935

B 108 - 2021-22 - Endeligt svar på spørgsmål 1: Spm. om beregninger af effekten af digitale samtaler fra første dag på baggrund af de opdaterede regneprincipper, til beskæftigelsesministeren

depends on the number of applications sent by other workers, hence creating search

congestion. One contribution we make is that we adjust the Gautier et al. (2018) model

to ﬁt the JSA policy evaluated in this paper.

Simulations using the estimated model show that a larger share of participants leads

to a lower unemployment rate. Increasing the treatment share from 0 to 100 percent

lowers the unemployment rate by around 0.2 percentage points, suggesting that the net

eﬀect of a full-scale roll-out is positive—despite the substantial displacement of jobs.

The program has a small negative eﬀect on government spending, since the reduction

in beneﬁt payments due to the lower unemployment rate is only slightly larger than the

program costs. Welfare is decreasing in the share of participants, because the program

implies lost non-market time, and because vacancy costs go up as the vacancy rate

increases. In sensitivity analyses, we also consider a model with a delayed vacancy

response to explore the fact that it may take time for ﬁrms to observe and react to

the new market conditions. This delayed vacancy model predicts larger reductions of

unemployment and government expenditures, and reverses the welfare eﬀects to positive

numbers.

Section 2 details the experiment, and Section 3 presents the data sources and the

empirical strategy. Section 4 gives the main results for the program and displacement

eﬀects. In Section 5, we shed light on the mechanisms behind the direct eﬀect of the JSA

program. Section 6 compares the three types of meetings, and Section 7 investigates

the origin of the displacement eﬀects. Section 8 presents our equilibrium search model

and reports the simulation results. Section 9 concludes.

Electronic copy available at: https://ssrn.com/abstract=3515935

B 108 - 2021-22 - Endeligt svar på spørgsmål 1: Spm. om beregninger af effekten af digitale samtaler fra første dag på baggrund af de opdaterede regneprincipper, til beskæftigelsesministeren

2.1

The experiment

Randomization

The experiment took place during six months in 2015 (March–May and August–Novem-

ber). It involved 72 Public Employment Service (PES) oﬃces, which corresponded

to roughly one quarter of all oﬃces in Sweden. The oﬃces were selected in order

to be a representative sample with respect to geographical dispersion and size. The

target population consisted of all newly unemployed job seekers at the 72 oﬃces, only

exempting job seekers who had been unemployed in the last three months and newly

arrived immigrants. Most local labor markets have one employment oﬃce. Thus,

randomization over oﬃces implies randomization over local labor markets, facilitating

estimation of displacement eﬀects. In the metropolitan areas with more than one oﬃce

we selected oﬃces where job seekers were less likely to compete for the same jobs, e.g.,

one oﬃce in the northern part and one oﬃce in the southern part of the city.

For the 72 local oﬃces, we applied a two-level randomization strategy over both

oﬃces and job seekers. Randomization over job seekers within active oﬃces identiﬁes

the direct eﬀect of the JSA program by comparing treated and non-treated job seekers in

the same labor market. Randomization over oﬃces identiﬁes the displacement eﬀects

by comparing non-treated job seekers at active and non-active oﬃces. To achieve

a balanced sample of active and non-active oﬃces we used stratiﬁed randomization.

Based on a model developed by the PES, we divided the oﬃces into sixfolds with

similar economic and demographic conditions. Within each strata, we then randomized

each oﬃce into diﬀerent categories: one oﬃce was assigned face-to-face meetings, one

distance meetings, one group meetings, and the remaining three constituted non-active

oﬃces. In total, this gave 12 oﬃces per meeting format and 36 non-active oﬃces that

continued with the baseline services oﬀered to all job seekers.

We excluded the smallest local labor markets (monthly inﬂow less than 20 job seekers), since it

would have been diﬃcult to assist job seekers using group meetings in these oﬃces.

Electronic copy available at: https://ssrn.com/abstract=3515935

B 108 - 2021-22 - Endeligt svar på spørgsmål 1: Spm. om beregninger af effekten af digitale samtaler fra første dag på baggrund af de opdaterede regneprincipper, til beskæftigelsesministeren

The active oﬃces randomly assigned 50% of the target population to the treatment

group.

Since the randomization within the active oﬃces was based on the job seekers’

date of birth, we know the treatment assignment according to the treatment protocol.

We use this theoretical treatment assignment to perform balancing tests in Table 1.

Columns 1 and 2 present group averages for treated and non-treated job seekers at the

active oﬃces, and Column 3 gives the sample statistics for job seekers at the non-active

oﬃces. Columns 4 and 5 show p-values for comparisons across the groups. The results

are reassuring: there are no signiﬁcant diﬀerences between the treated and the non-

treated job seekers at the active oﬃces, or between job seekers at active and non-active

oﬃces.

2.2

Treatment

Perhaps the most straightforward way to intensify job search assistance is to increase

the number of meetings between job seekers and professional caseworkers (see, e.g.,

Graversen and van Ours, 2008a,b). The program used in this study more than doubled

the meeting frequency during the ﬁrst quarter of the unemployment spell. All extra

meetings were mandatory

, although only job seekers with unemployment beneﬁts

could be subject to sanctions if they did not show up when summoned. The active

oﬃces were compensated by central project-speciﬁc funding, with the intention to fully

ﬁnance the increased costs from the program.

To compare diﬀerent types of JSA, the experiment included three meeting formats:

During the spring wave, the treatment intensity was randomly set to either 50 or 80%. Since

take-up varied across oﬃces with diﬀerent treatment intensities we had insuﬃcient power to use this

diﬀerence in the analysis.

All the estimates presented in the paper are weighted by the intention to treat share, i.e., the

observed share of job seekers at the local oﬃce who would be randomized to treatment based on the

treatment protocol. This corrects for the diﬀerent shares in the spring (50 and 80%) and for random

diﬀerences between the oﬃces (e.g., one oﬃce having 48% and another having 52% treated for a

treatment share of 50%).

The exception was the distance meetings during the spring wave when we could not use telephone

meetings due to security issues. This meant that job seekers who did not have access to a computer

with internet connection and an e-ID could not participate.

Electronic copy available at: https://ssrn.com/abstract=3515935

B 108 - 2021-22 - Endeligt svar på spørgsmål 1: Spm. om beregninger af effekten af digitale samtaler fra første dag på baggrund af de opdaterede regneprincipper, til beskæftigelsesministeren

Table 1: Balancing test of diﬀerences between treated and non-treated at active oﬃces

and between active and non-active oﬃces

Treated

Variables

(1)

Age

Male

Unemployment beneﬁts

Health disability

Matchable

Education level

Less than high school

High school

College

Place of birth

Sweden

Nordic countries

West Europe

Outside west Europe

Unemployment days

Year t–1

Year t–2

Year t–3

Year t–4

Unemployment spells

Year t–1

Year t–2

Year t–3

Year t–4

No. programs, last 4 yrs

Labor market education

Preparatory education

Labor market training

Subsidized employment

Observations

33.33

0.542

0.642

0.052

0.868

0.224

0.491

0.285

0.678

0.013

0.036

0.273

30.66

67.42

69.57

63.82

0.431

0.789

0.806

0.706

0.024

0.048

0.027

0.106

14,075

Non-

treated

(2)

33.39

0.539

0.638

0.052

0.862

0.223

0.493

0.285

0.667

0.015

0.034

0.284

30.43

68.49

71.83

63.68

0.440

0.800

0.815

0.721

0.020

0.047

0.030

0.108

12,463

Non-

active

oﬃces

(3)

33.54

0.555

0.634

0.054

0.864

0.225

0.486

0.289

0.641

0.015

0.036

0.308

30.98

68.25

71.22

66.14

0.442

0.793

0.819

0.716

0.022

0.047

0.031

0.106

31,240

p-val

diﬀ

T–C

(4)

0.707

0.571

0.546

0.931

0.117

0.873

0.833

0.932

0.059

0.197

0.256

0.035

0.767

0.449

0.133

0.921

0.427

0.518

0.630

0.397

0.172

0.869

0.389

0.711

26,538

p-val

diﬀ

TC–NA

(5)

0.637

0.186

0.747

0.602

0.936

0.878

0.690

0.844

0.399

0.547

0.857

0.385

0.777

0.914

0.868

0.438

0.793

0.972

0.840

0.960

0.965

0.983

0.370

0.920

57,778

Notes: Summary statistics by treatment status, weighted by the observed intention to treat share. Standard errors

in column 5 are clustered at the PES oﬃce level.

Unemployment beneﬁts

is an indicator variable for collecting

unemployment beneﬁts,

Health disability

is an indicator variable for having a functional disability and

Matchable

an indicator variable from an initial assessment about the job seeker’s potential to take a job on short notice.

Electronic copy available at: https://ssrn.com/abstract=3515935

B 108 - 2021-22 - Endeligt svar på spørgsmål 1: Spm. om beregninger af effekten af digitale samtaler fra første dag på baggrund af de opdaterede regneprincipper, til beskæftigelsesministeren

individual face-to-face meetings, individual distance meetings and group meetings. The

face-to-face meetings and the distance meetings were similar in many respects. They

added three extra meetings to the baseline services, which included a short meeting

when the job seeker registered, followed by a longer planning meeting, i.e., about two

meetings within the ﬁrst quarter. The extra meetings focused on personalized job

search assistance, but we did not provide any detailed instructions about the content

of the meetings. Instead, it was up to the caseworkers to oﬀer assistance according to

the needs of each job seeker. The basic idea with oﬀering distance meetings—online or

via telephone—was that this could create greater ﬂexibility, both for the caseworkers

and for the job seekers.

In contrast, the group meetings used a more detailed protocol. Since the meetings

gathered around 10–15 job seekers at once they were given in the form of seminars,

where each occasion concerned a speciﬁc topic, such as CV writing, interview training,

and advice for creating professional networks. The instruction was that job seekers

and caseworkers should meet frequently in an initial stage, with ﬁve seminars over

the ﬁrst two weeks of unemployment. Thereafter, the participants were divided into

smaller groups, which were supposed to meet on their own on a weekly basis during

two months (not visible in our data).

Table 2 looks at the diﬀerent treatments given within the experiment. It shows

the intention-to-treat comparison of treated and non-treated job seekers (according to

the treatment protocol) at the active oﬃces, for the full sample and for each of the

three meeting formats. We use information about all registered contacts between the

job seeker and the caseworker in the PES administrative registers, which covers all job

seekers. We distinguish between

physical meetings,

which include face-to-face meetings

and group meetings, and

distance meetings,

which include contacts over telephone and

online using an e-ID.

All meetings

are the sum of these two.

Panel A describes program participation. During the ﬁrst registration meeting,

Electronic copy available at: https://ssrn.com/abstract=3515935

B 108 - 2021-22 - Endeligt svar på spørgsmål 1: Spm. om beregninger af effekten af digitale samtaler fra første dag på baggrund af de opdaterede regneprincipper, til beskæftigelsesministeren

Table 2: Program participation and meetings in active oﬃces

All meeting

types

(1)

Panel A: Program participation

Informed about the program

At least one program meeting

Panel B: Number of meetings

All meetings quarter 1

0.500

∗∗∗

(0.026)

[3.159]

−0.017

(0.021)

[1.176]

0.414

∗∗∗

(0.039)

[3.150]

0.002

(0.033)

[1.152]

0.418

∗∗∗

(0.046)

[3.302]

−0.034

(0.039)

[1.247]

0.707

∗∗∗

(0.052)

[3.019]

−0.024

(0.039)

[1.131]

0.619

∗∗∗

(0.004)

0.230

∗∗∗

(0.004)

0.601

∗∗∗

(0.007)

0.265

∗∗∗

(0.006)

0.621

∗∗∗

(0.008)

0.247

∗∗∗

(0.007)

0.643

∗∗∗

(0.008)

0.164

∗∗∗

(0.006)

Face-to-face

meetings

(2)

Distance

meetings

(3)

Group

meetings

(4)

All meetings quarter 2

Panel C: Type of meeting

Physical meetings quarter 1

0.318

∗∗∗

(0.021)

[2.262]

0.182

∗∗∗

(0.015)

[0.897]

0.376

∗∗∗

(0.032)

[2.287]

0.038

∗

(0.022)

[0.863]

-0.015

(0.035)

[2.331]

0.433

∗∗∗

(0.030)

[0.971]

0.596

∗∗∗

(0.043)

[2.155]

0.112

∗∗∗

(0.027)

[0.864]

Distance meetings quarter 1

Panel D: Time pattern of meetings

All meetings month 1

0.193

∗∗∗

(0.014)

[2.115]

0.204

∗∗∗

(0.013)

[0.584]

0.104

∗∗∗

(0.011)

[0.460]

26,538

0.120

∗∗∗

(0.021)

[2.097]

0.162

∗∗∗

(0.019)

[0.585]

0.132

∗∗∗

(0.017)

[0.468]

10,567

0.155

∗∗∗

(0.024)

[2.176]

0.144

∗∗∗

(0.021)

[0.641]

0.119

∗∗∗

(0.020)

[0.486]

8,259

0.335

∗∗∗

(0.030)

[2.076]

0.324

∗∗∗

(0.026)

[0.523]

0.048

∗∗∗

(0.019)

[0.420]

7,712

All meetings month 2

All meetings month 3

Observations

Notes: The results are from a linear regression of each variable on a treatment indicator, weighted by the observed

intention to treat share. The sample includes the active PES oﬃces during the experiment period. Standard errors

in parentheses, control means in square brackets.

∗∗∗

p<0.01,

∗∗

p<0.05,

∗

p<0.1. Face-to-face meetings include

visits, Distance meetings include contacts over telephone or online using an e-ID, and All meetings are the sum of

the two, all based on administrative records from the PES.

Electronic copy available at: https://ssrn.com/abstract=3515935

B 108 - 2021-22 - Endeligt svar på spørgsmål 1: Spm. om beregninger af effekten af digitale samtaler fra første dag på baggrund af de opdaterede regneprincipper, til beskæftigelsesministeren

the caseworkers were instructed to use an experiment-speciﬁc tool that revealed the

treatment status of the worker (based on date-of-birth), and then inform the treated

job seekers about the program.

The ﬁrst row of Panel A shows that this tool was

used for 62% of the job seekers in the treatment group. The main explanations to the

take-up rate was that individual caseworkers failed to use the experiment-speciﬁc tool

and that some oﬃces failed to correctly capture the target population. However, since

we observe the entire target population and for whom the experiment-speciﬁc tool was

used, we have full control over how the experiment was implemented. The second row

of Panel A shows that 23% of the target population participated in at least one program

speciﬁc meeting. Information about the JSA program speciﬁc meetings were collected

from the written acts at the PES. The fraction who initiated treatment was highest

for face-to-face meetings (27%), and lowest for group meetings (16%). Reasons for not

initiating treatment was, for example, caseworkers failing to oﬀer the extra meetings,

job seekers ﬁnding a job before the ﬁrst meeting, and job seekers simply not showing

up for the extra meetings.

Panels B–D of Table 2 show that the impact of the program on the frequency, type

and time pattern of meetings was consistent with the treatment protocol. Panel B

shows that the treated job seekers received on average 0.5 more meetings, which, given

the take-up rate of 23%, corresponds to about two more meetings for those who actu-

ally showed up (not adjusted for early exits).

We also see that the meeting frequency

increased during the experiment period (quarter 1) but not after the experiment period

(quarter 2), and that the increase was similar for face-to-face and distance meetings

and slightly larger for group meetings. Panel C shows that job seekers assigned to face-

to-face meetings and group meetings received signiﬁcantly more physical meetings,

Participants in the group meetings were informed about all ﬁve meetings immediately, while

participants in the face-to-face and distance meetings typically were summoned by caseworkers to one

meeting at the time.

This ﬁgure is lower than the intended number of meetings, which was three for the face-to-face

and the distance meetings and ﬁve for group meetings. However, given that a substantial fraction of

the job seekers found a job within three months, this is consistent with the treatment protocol.

Electronic copy available at: https://ssrn.com/abstract=3515935

B 108 - 2021-22 - Endeligt svar på spørgsmål 1: Spm. om beregninger af effekten af digitale samtaler fra første dag på baggrund af de opdaterede regneprincipper, til beskæftigelsesministeren

whereas job seekers assigned to distance meetings received more distance meetings.

Finally, Panel D shows that face-to-face and distance meetings were fairly evenly dis-

tributed over the ﬁrst three months of the unemployment spell, while group meetings

were more concentrated to the ﬁrst two months.

3.1

Data and empirical strategy

Data

Our analysis beneﬁts from access to rich data. Here, we provide an overview of the

data records and then describe the data in more detail in the analyses. From the

Swedish PES we have information on individual characteristics as well as daily records

of unemployment status for all job seekers. The PES data also contains information on

all meetings between the caseworkers and the job seekers. We have access to wage data

on the full-time equivalent monthly wage rates from Statistics Sweden and detailed

vacancy data from the PES on the number of posted vacancies per municipality and

month.

Detailed information about caseworkers’ actions and job seekers’ search behavior

allow us to study the mechanisms behind the eﬀects of the JSA program. For caseworker

actions, we use several administrative data sources collected by the PES. First, to study

monitoring of job seekers, we use data on all registered violations of the job search

rules. Second, to study the provision of support and training, we use information

from individual action plans specifying all types of support given to each job seeker,

along with daily records of participation in ALMPs. Third, to analyze how caseworkers

communicate information about vacancies, we use data on vacancy referrals from the

caseworkers to the job seekers. Finally, we use meeting records for each caseworker to

All treatment assignments were associated with more distance meetings, but the association was

by far the strongest for the distance meetings group. Among other things, distance meetings include

all telephone contacts between job seekers and caseworkers.

Electronic copy available at: https://ssrn.com/abstract=3515935

B 108 - 2021-22 - Endeligt svar på spørgsmål 1: Spm. om beregninger af effekten af digitale samtaler fra første dag på baggrund af de opdaterede regneprincipper, til beskæftigelsesministeren

study any reallocation of resources from treated to non-treated job seekers.

We are also among the ﬁrst to use data on job seeker search behavior from monthly

activity reports. Since September 2013, job seekers in Sweden are required to submit a

monthly report on the number and type of all job search activities to the PES (e.g., job

applications and search channels). Failure to provide a report, or submitting a report

indicating too low level of search activity, may lead to a beneﬁt sanction. Moreover,

for each job application, the job seekers have to report the name and telephone number

of the ﬁrm, and the occupation to which they apply, so that caseworkers can validate

the information in the reports. Therefore, the activity report data should be a reliable

data source for measuring job search behavior.

3.2

Empirical strategy

We ﬁrst analyze the data the way it is usually done in evaluations of experiments: we

utilize the randomization over job seekers within oﬃces and compare the outcomes of

treated and non-treated job seekers at the active oﬃces. To do this, we assign treatment

status according to the treatment protocol of the experiment, and estimate:

1(Assigned to program

) +

(1)

where our main parameter of interest is the intention-to-treat (ITT) eﬀect of the JSA

program,

. We then estimate displacement eﬀects by exploiting the randomization

over oﬃces. Here, the model for individual

in oﬃce

is:

1(Assigned to program

) +

1(In a program area

) +

(2)

where 1(In a program area

) indicates an active oﬃce and 1(Assigned to program

)

indicates being in the treatment group. The displacement eﬀect is given by

, which

is identiﬁed through the comparison of non-treated job seekers at active and non-active

Electronic copy available at: https://ssrn.com/abstract=3515935

B 108 - 2021-22 - Endeligt svar på spørgsmål 1: Spm. om beregninger af effekten af digitale samtaler fra første dag på baggrund af de opdaterede regneprincipper, til beskæftigelsesministeren

oﬃces.

captures the direct comparison of the treated and the non-treated at the

active oﬃces, and the net eﬀect for the treated job seekers, compared to non-treated job

seekers at the non-active oﬃces, is given by ∆

≡

. The overall eﬀect takes the

share of treated into account and captures the average eﬀect across both treated and

non-treated job seekers at the active oﬃces compared to job seekers at the non-active

oﬃces.

The analysis focuses on the ITT eﬀect of the JSA program, but we will also report

instrumental variables (IV) estimates of the eﬀect of program participation. One reason

is that the structural model is set up for program participants. Since job seekers may

react already to information about the program and such responses can be considered

a program eﬀect, we take a cautious approach and deﬁne program participants as those

who were supposed to be informed about the program (ﬁrst row of Table 2). However,

since only a fraction of the program participants showed up for an extra meeting (second

row of Table 2) our IV estimates do not reﬂect the eﬀect of actual take-up of meetings.

In that sense, they are conservative.

To obtain the direct program eﬀect, we instrument program participation with the

program assignment indicator, 1(Assigned to program

), and estimate:

1(Program participation

) +

1(In a program area

) +

(3)

where

reﬂects the direct program eﬀect at the active oﬃces. The interpretation of

this LATE eﬀect depends on how representative program participants are compared to

the target population as a whole. Table A-1 in the appendix shows that the program

participants are similar to the target population in terms of observable characteristics,

which suggests that the LATE may be similar to the average program eﬀect. This is

expected since the dropout mainly occurred at the caseworker and the oﬃce level, and

was not due to individual job seekers selecting in or out of the program, as described

Electronic copy available at: https://ssrn.com/abstract=3515935

B 108 - 2021-22 - Endeligt svar på spørgsmål 1: Spm. om beregninger af effekten af digitale samtaler fra første dag på baggrund af de opdaterede regneprincipper, til beskæftigelsesministeren

in Section 2.2.

To estimate the net program eﬀect for the participants, we follow the

IV approach in Cr´pon et al. (2013).

4.1

Eﬀects of the job search assistance program

Unemployment

We begin with a visual inspection of the direct ITT eﬀect of the program. As shown in

Figure 1 there is a striking diﬀerence between treated (assigned to the JSA program)

and non-treated (assigned to the control group) at the active oﬃces: the unemployment

rate is lower for the treated than for the non-treated. The eﬀect of the program reaches

its maximum of about three percentage points around the time when treatment stops,

after 3–4 months.

The immediate eﬀect of the program is expected since the intensive

caseworker assistance aims at helping job seekers leave employment as fast as possible.

Hence, in contrast to, for instance, training programs there are no lock-in eﬀects. The

ﬁgure also shows that even though the eﬀect decreases after treatment ends, it is still

signiﬁcant until the tenth month since inﬂow to unemployment.

We next estimate treatment eﬀects in a regression framework using three outcomes:

(1) the probability of leaving unemployment during the ﬁrst quarter of unemployment,

and the number of days registered as unemployed during (2) the ﬁrst quarter and (3)

Not surprisingly, the picture is diﬀerent for the job seekers who attended at least one extra meeting.

Compared to the full target population, they receive unemployment beneﬁts more often, are disabled

to a lesser extent, are perceived as matchable to a higher degree, have a higher education level and

are natives to a larger extent. On the other hand, they have more extensive unemployment history,

with somewhat more days in unemployment over the last four years.

See Cr´pon et al. (2013) for details. Here, the IV scaling takes into account that the compliers ex-

perience positive program eﬀects, whereas the non-compliers experience negative displacement eﬀects.

The main identifying assumptions is that the displacement eﬀects are uncorrelated with treatment

status. In practice, this means that compliers and non-compliers on average should have similar po-

tential outcomes under no treatment. In our case, the compliers and the non-compliers have similar

observed characteristics, lending support to this assumption.

Since the eﬀect appears early in the unemployment spell, it may reﬂect pre-program responses.

The median number of days until the ﬁrst program meeting is 29 days and the direct eﬀect during the

ﬁrst month is relatively small, so that pre-program responses are unlikely to be the main explanation

to the observed pattern.

Electronic copy available at: https://ssrn.com/abstract=3515935

B 108 - 2021-22 - Endeligt svar på spørgsmål 1: Spm. om beregninger af effekten af digitale samtaler fra første dag på baggrund af de opdaterede regneprincipper, til beskæftigelsesministeren

-.06

Difference Treated-Non-treated

-.04

-.02

.02

Months since inflow to unemployment

Figure 1: Diﬀerence in the share of unemployed between treated and non-treated job

seekers at the active oﬃces, by months since inﬂow to unemployment

the ﬁrst year, where the last two measures include re-unemployment after a period of

employment. Unemployment includes full-time unemployment and participation in an

active labor market program. Panels A–C in Table 3 show the results for each of the

three outcomes. The ﬁrst two columns of the table, where we report the direct ITT

eﬀect,

, based on equation (1), conﬁrm the graphical evidence from Figure 1. The

JSA program increases the exit rate from unemployment by 3.5 percentage points, or

about 10% (column 2 with individual control variables). We also see that the treated

have 1.5 fewer days of unemployment during the ﬁrst quarter (Panel B), and 5.9 fewer

days of unemployment during the ﬁrst year (Panel C).

Since other interventions for unemployed job seekers previously studied in the lit-

erature diﬀer in many respects, such as content, intensity, target population, labor

market situation and evaluation horizon, it is diﬃcult to compare eﬀect sizes across

studies. We note that the direct eﬀect in our study is somewhat smaller than eﬀects

from experiments evaluating JSA in the form of early meetings in Denmark (Graversen

and van Ours, 2008b; Gautier et al., 2018). This is expected, however, since the Danish

programs were more intensive.

Electronic copy available at: https://ssrn.com/abstract=3515935

B 108 - 2021-22 - Endeligt svar på spørgsmål 1: Spm. om beregninger af effekten af digitale samtaler fra første dag på baggrund af de opdaterede regneprincipper, til beskæftigelsesministeren

The ﬁndings in columns 1 and 2 suggest that the JSA program successfully decreased

time in unemployment. However, the overall eﬀectiveness of the program ultimately

depends on the degree of displacement. If the positive eﬀect of job search assistance

comes at the expense of non-treated job seekers there is reason to question the beneﬁts

of the policy. We therefore turn to models that estimate both direct and displacement

eﬀects, and report estimates from equation (2) using data from both active and non-

active oﬃces (Column 3 of Table 3). As expected, the direct eﬀect is similar to the

estimates above. Thus, estimating the direct eﬀect using

from equation (1) or

from equation (2) lead to similar results.

Turning to the displacement eﬀect in column 3, the estimate of being in a program

area indicates sizeable displacement, but the precision is low. To increase precision, we

therefore add data from time periods prior to the experiment, back until year 2012.

This strategy is illustrated in Figure 2. Each line in the ﬁgure compares the number

of unemployment days during the ﬁrst year between two groups of job seekers. The

darker grey line compares the treated and the non-treated at the active oﬃces, and the

lighter grey line compares non-treated job seekers at active and non-active oﬃces (by

calendar month of inﬂow to unemployment). The dashed vertical lines, ﬁnally, indicate

the two experiment periods. Before the experiment there is no systematic pattern in

the data—the time series are noisy but stay around zero. In contrast, during the two

experiment periods the series diverge. Unemployment decreases for the treated relative

to the non-treated at the active oﬃces, and, at the same time, unemployment increases

for the non-treated at the active oﬃces relative to the non-treated at the non-active

oﬃces. This pattern gives a clear indication that the JSA program caused displacement

eﬀects.

These displacement eﬀect patterns are supported by the regression estimates in

We include unemployed job seekers during the entire period 2012–2015 using the same sample

restrictions for the target population as during the experiment in 2015. We use data beginning in

2012 since several reforms were introduced between 2011 and 2012, including public investments to

increase the number of caseworkers, and a new mandate for the PES to provide services to immigrants.

Electronic copy available at: https://ssrn.com/abstract=3515935

B 108 - 2021-22 - Endeligt svar på spørgsmål 1: Spm. om beregninger af effekten af digitale samtaler fra første dag på baggrund af de opdaterede regneprincipper, til beskæftigelsesministeren

Table 3: Direct and displacement eﬀects of the JSA program on unemployment

Experiment period

ITT

(1)

Panel A: Exit unemp. 1st quarter

Assigned to program

Program participant

In a program area

Net eﬀect treated

Control mean

Assigned to program

Program participant

In a program area

Net eﬀect treated

Control mean

Assigned to program

Program participant

In a program area

Net eﬀect treated

Control mean

Year dummies

Month dummies

PES oﬃce dummies

Controls

Clusters

Observations

196.1

26,538

196.1

Yes

26,538

7.627

∗

(4.504)

1.683

(4.478)

190.7

Yes

57,778

4.318

∗∗

(1.912)

−2.284

(1.612)

187.0

Yes

552,816

4.160

∗∗

(1.727)

−1.831

(1.484)

187.0

Yes

552,816

74.37

−6.692

∗∗∗

(1.527)

74.37

−5.893

∗∗∗

(1.414)

1.158

(0.731)

−0.355

(0.755)

73.51

−5.944

∗∗∗

(1.296)

0.725

∗∗

(0.344)

−0.780

∗∗

(0.334)

73.78

−6.602

∗∗∗

(1.778)

0.695

∗∗

(0.318)

−0.761

∗∗

(0.326)

73.78

−5.991

∗∗∗

(1.359)

−9.679

∗∗∗

(2.195)

4.160

∗∗

(1.727)

−5.519

(2.936)

187.0

Yes

552,816

0.354

−1.574

∗∗∗

(0.318)

0.354

−1.502

∗∗∗

(0.310)

−0.020

(0.014)

0.014

(0.013)

0.368

−1.512

∗∗∗

(0.328)

−0.016

∗∗

(0.007)

0.019

∗∗∗

(0.006)

0.390

−1.505

∗∗∗

(0.399)

−0.015

∗∗

(0.006)

0.018

∗∗∗

(0.005)

0.390

−1.456

∗∗∗

(0.353)

−2.352

∗∗∗

(0.570)

0.695

∗∗

(0.318)

−1.658

∗∗∗

(0.552)

73.78

0.036

∗∗∗

(0.006)

0.035

∗∗∗

(0.006)

0.035

∗∗∗

(0.005)

0.035

∗∗∗

(0.006)

0.034

∗∗∗

(0.006)

0.055

∗∗∗

(0.010)

−0.015

∗∗

(0.006)

0.039

∗∗∗

(0.010)

0.390

ITT

(2)

ITT

(3)

ITT

(4)

Pre-data, 2012–2015

ITT

(5)

(6)

Panel B: Unemp. days 1st quarter

Panel C: Unemp. days 1st year

Notes: Regression of each outcome variable on an indicator for active PES oﬃce (“In a program area”) and an

indicator for active PES oﬃce

intention to treat status is treated (“Assigned to program”). The controls include

the variables in Table 1. Standard errors in parentheses are clustered at the PES oﬃce level.

∗∗∗

p<0.01,

∗∗

p<0.05,

∗

p<0.1.

Net eﬀect in column (6) is for program participants.

Electronic copy available at: https://ssrn.com/abstract=3515935

B 108 - 2021-22 - Endeligt svar på spørgsmål 1: Spm. om beregninger af effekten af digitale samtaler fra første dag på baggrund af de opdaterede regneprincipper, til beskæftigelsesministeren

Diff unemp. days 1st year

-20

-10

2012m1

2013m1

2014m1

Month

2015m1

2016m1

Treated Active–Non-treated Active

Non-treated Active–Non-treated Non-active

Figure 2: Diﬀerences in days in unemployment during the ﬁrst year since registration

between treated and non-treated job seekers in active oﬃces, and between non-treated

job seekers at active and non-active oﬃces

columns 4–6 of Table 3. Here, we interact the variables

Assigned to program

and

In a

program area

in equation (2) with an indicator for becoming unemployed during the

Experiment period,

and adjust for oﬃce ﬁxed eﬀects and calendar time (year and month

dummies). This strategy of using pre-experiment data to capture oﬃce level hetero-

geneity gives a similar direct eﬀect and a slightly lower displacement eﬀect compared

to using only data from the experiment period in column 3, but due to the increased

precision the displacement eﬀects are now signiﬁcant.

Our preferred estimates in column 5 (with individual characteristics) suggest that

the program reduces the exit rate for the non-treated job seekers at the active oﬃces by

1.5 percentage points (3.8%) during the ﬁrst quarter of unemployment, a ﬁnding that is

consistent with substantial displacement eﬀects of the JSA program. The net eﬀect for

treated job seekers,

, is 1.8 percentage points, or 4.6%. The overall eﬀect, taking

both the impact on the treated and the non-treated job seekers at the active oﬃces into

account, depends on the share of treated job seekers which was roughly 50%. Thus,

Electronic copy available at: https://ssrn.com/abstract=3515935

B 108 - 2021-22 - Endeligt svar på spørgsmål 1: Spm. om beregninger af effekten af digitale samtaler fra første dag på baggrund af de opdaterede regneprincipper, til beskæftigelsesministeren

since the positive net eﬀect for the treated is larger than the negative displacement

eﬀect for the non-treated, the overall eﬀect of the program is positive. On average,

across treated and non-treated at the active oﬃces, the JSA program increases the job

ﬁnding rate by 0.25 percentage points or 0.7%.

Finally, by construction, the IV estimates in column 6 are larger than their ITT

counterparts in column 5, with a net program eﬀect corresponding to a 3.9 percentage

point (10%) increase in the exit rate during the ﬁrst quarter. The results for days in

unemployment during the ﬁrst quarter and the ﬁrst year in Panels B and C of Table 3

display similar patterns.

Summing up, the JSA program reduces unemployment among the treated job seek-

ers, but also leads to large displacement eﬀects for the non-treated job seekers. This is

similar to the results from the two-level randomized experiment reported in Cr´pon et

al. (2013). In their case, the net eﬀect for the treated is close to zero, insigniﬁcant, and

smaller than the displacement eﬀects for the non-treated, which implies that more jobs

were lost than found. Hence, we ﬁnd more positive overall eﬀects of JSA. This may

be due to the more general target population (all newly unemployed job seekers with

no unemployment during the last 3 months) compared to Cr´pon et al. (2013) (young

college graduates unemployed for at least six months), or that JSA was oﬀered earlier

during the unemployment spell. This strengthens the importance of studying displace-

ment of JSA policies in diﬀerent settings, to get a deeper understanding of when and

why displacement arises.

4.2

Direct and displacement eﬀect heterogeneity

The two-step randomization setup allows us to credibly identify displacement eﬀects

and, hence, the overall eﬀect of the program for the target population. However, this

ignores displacement for non-treated job seekers

outside

the target population, such as

The share of treated is 53%, so that the overall eﬀect is 0.53*0.18 (eﬀect for the treated)+0.47*-1.5

(eﬀect for the non-treated)=0.25 percentage points.

Electronic copy available at: https://ssrn.com/abstract=3515935

B 108 - 2021-22 - Endeligt svar på spørgsmål 1: Spm. om beregninger af effekten af digitale samtaler fra første dag på baggrund af de opdaterede regneprincipper, til beskæftigelsesministeren

the stock of job seekers that entered unemployment before the experiment, the inﬂow

of unemployed during the summer months when no one was assigned to the program,

and the inﬂow of unemployed during the experiment period that did not belong to

the target population. It can be questioned whether these groups compete for the

same jobs as the target population. Both the stock of unemployed at the beginning

of the experiment period and the inﬂow of unemployed outside the target population

during the experiment period (newly arrived immigrants and job seekers with a recent

unemployment spell) are likely to be further from the labor market to begin with. The

inﬂow during the summer months, on the other hand, mainly consists of short-term

unemployed, such as students looking for a job during the summer break.

Table A-2 in the appendix, nevertheless, presents estimates of the displacement

eﬀects for these groups outside the target population. The ﬁrst three columns study

the impact on individuals in the target population who entered unemployment before

the experiment, during the summer months, or after the experiment in 2015. The

results indicate that these groups may be somewhat aﬀected, but substantially less

than the non-treated job seekers in the target population during the experiment. The

last three columns study the impact on individuals who became unemployed during

the experiment period but did not belong to the target population, who appear to be

unaﬀected by the program. Overall, these ﬁndings speak against any large displacement

eﬀects of the program for groups outside the target population.

The existence of displacement eﬀects in our experiment also raises the important

question of

who

gains and

who

loses. From a welfare perspective we may accept dis-

placement if job search assistance beneﬁts job seekers who are less attached to the

labor market at the expense of those who are closer to ﬁnding a job. Table A-3 in the

appendix presents estimates by education, country of origin, unemployment history,

and gender. Since the analysis reduces the sample sizes dramatically, most diﬀerences

across sub-groups in the table are statistically insigniﬁcant. Still, overall, the results

Electronic copy available at: https://ssrn.com/abstract=3515935

B 108 - 2021-22 - Endeligt svar på spørgsmål 1: Spm. om beregninger af effekten af digitale samtaler fra første dag på baggrund af de opdaterede regneprincipper, til beskæftigelsesministeren

in columns 1–6 indicate that the least attached groups of job seekers beneﬁt the most

from the JSA program, in particular when considering the impact over the entire ﬁrst

year. Estimates of the direct eﬀect are larger for low educated, foreign born and job

seekers with long unemployment history (above median number of unemployment days

in the last four years). The displacement eﬀects, on the other hand, are more evenly

distributed across groups. We ﬁnally note some interesting gender diﬀerences; men ben-

eﬁt more than women in terms of direct eﬀects, but also suﬀer more from displacement

(columns 7 and 8).

4.3

Wages and vacancies

The positive eﬀect on job ﬁnding in our experiment could cause demand side responses

with eﬀects on wages and the number of vacancies created by ﬁrms in the market. Any

such eﬀects are important determinants of the market level unemployment rate in the

new equilibrium. To ﬁnd out how job search assistance aﬀects wages we estimate the

eﬀect on the log wage rate for the ﬁrst job after unemployment, using data on the full-

time equivalent monthly wage rate from Statistics Sweden.

The results in Table 4

reveal no signiﬁcant eﬀects on wages neither for the treated nor for the non-treated

at the active oﬃces. These results are also informative for the impact of JSA on job

quality. On the one hand, we may expect that professional support from caseworkers,

with expert knowledge on the local labor market, would help job seekers to ﬁnd higher-

quality jobs or improve the matching between job seekers and employers. On the other

hand, if job seekers feel pressured to exit unemployment early they may accept jobs of

lower quality or worse matches with employers.

Notably, neither of these hypotheses

are supported by our wage estimates.

To measure the vacancy eﬀect, we exploit disaggregated data on all vacancies posted

The data cover roughly 50% of all jobs, including the entire public sector, large private ﬁrms and

a sub-sample of small private ﬁrms. There is no diﬀerence across treated and non-treated job seekers

with respect to the sampling scheme (conditional on ﬁnding a job). Since treated job seekers ﬁnd jobs

Electronic copy available at: https://ssrn.com/abstract=3515935

B 108 - 2021-22 - Endeligt svar på spørgsmål 1: Spm. om beregninger af effekten af digitale samtaler fra første dag på baggrund af de opdaterede regneprincipper, til beskæftigelsesministeren

Table 4: Eﬀects of the JSA program on log wages

ITT

(1)

Assigned to program

In a program area

Net eﬀect treated

Control mean

Observations

−0.002

(0.004)

0.000

(0.005)

−0.001

(0.003)

10.09

237,598

(2)

−0.003

(0.006)

0.000

(0.005)

−0.002

(0.008)

10.09

237,598

Notes: Outcome is the log monthly wage on an indicator for active PES oﬃce (“In a program area”) and an indicator

for active PES oﬃce

intention to treat status is treated (“Assigned to program”). The regressions include year

dummies, month dummies, PES oﬃce dummies and the control variables in Table 1. The net eﬀect in column 2 is

for the program participants. Standard errors in parentheses are clustered at the PES oﬃce level.

∗∗∗

p<0.01,

∗∗

p<0.05,

∗

p<0.1.

at the PES per municipality and month over the period 2012–2015.

For each local

labor market (oﬃce) and month, we calculate the average number of vacancies facing

the target population, based on their municipality of residence. We use the log of

vacancies as outcome, and estimate a similar model as for the other outcomes above.

The main diﬀerence is that we now have one outcome measure per oﬃce and month.

This means that we focus on the program area dummy, which takes the value one

for the active oﬃces during the experiment period. Since the size of the local labor

markets varies, we use both weighted (by size of target population in each oﬃce) and

unweighted regressions.

The vacancy estimates in Table 5 show some interesting patterns. Column 1 indi-

cates that the JSA program increased the number of posted vacancies by 3.3% during

the experiment, but the eﬀect is insigniﬁcant. Column 2 allows for diﬀerent responses

during the spring and the fall wave of the experiment. The rationale for separating

between the immediate and somewhat longer run is that it may take time for ﬁrms to

faster we observe wages for a slightly larger share of treated compared to non-treated job seekers.

Cottier et al. (2018) provide recent evidence from Switzerland suggesting that JSA may push job

seekers into jobs of lower quality.

Not all Swedish ﬁrms post vacancies at the PES. Firms hiring high-skilled workers, e.g., engineers,

are less likely to use this channel. However, the vacancies posted at the PES should capture those

that are relevant for our target population.

Electronic copy available at: https://ssrn.com/abstract=3515935

B 108 - 2021-22 - Endeligt svar på spørgsmål 1: Spm. om beregninger af effekten af digitale samtaler fra første dag på baggrund af de opdaterede regneprincipper, til beskæftigelsesministeren

observe and react to the new market in the wake of the JSA program. This is supported

by the data. Whereas vacancies are unaﬀected during the spring (the point estimate

is virtually zero), we ﬁnd a rather sizeable eﬀect during the fall, which is signiﬁcant at

the ten percent level with the unweighted data (column 4). We return to this delayed

vacancy pattern when we estimate the equilibrium search model in Section 8.

Table 5: Eﬀects of the JSA program on vacancies

Weighted

(1)

In a program area

In a program area×Spring period

In a program area×Fall period

Control mean

Observations

5.896

16,269

0.033

(0.033)

−0.002

(0.033)

0.066

(0.041)

5.896

16,269

5.896

16,269

(2)

(3)

0.048

(0.035)

0.018

(0.043)

0.076

∗

(0.039)

5.896

16,269

Unweighted

(4)

Notes: Regression of the log number of vacancies facing the target population in each oﬃce and month on an

indicator for active PES oﬃce (“In a program area”), year dummies, month dummies and PES oﬃce dummies.

Weights are determined by the size of target population per oﬃce and month. Standard errors in parentheses are

clustered at the PES oﬃce level.

∗∗∗

p<0.01,

∗∗

p<0.05,

∗

p<0.1.

Why does the JSA help the unemployed?

We have shown that JSA, in the form of more meetings with a caseworker at the local

employment oﬃce, has a positive direct eﬀect on exit from unemployment. An impor-

tant question is how the direct eﬀect of the JSA program arises. While many previous

studies have documented positive eﬀects of JSA and monitoring policies (see, e.g., the

overview by Card et al., 2017), the evidence on the driving mechanisms behind the

eﬀects is still scarce. In particular, we know little about how JSA alters caseworkers’

and job seekers’ behavior, and how this operates to lower unemployment. This is unfor-

tunate since a better understanding of the underlying mechanisms facilitates eﬃcient

policy.

In this section, we analyze why the positive eﬀect of the JSA program in our ex-

Electronic copy available at: https://ssrn.com/abstract=3515935

B 108 - 2021-22 - Endeligt svar på spørgsmål 1: Spm. om beregninger af effekten af digitale samtaler fra første dag på baggrund af de opdaterede regneprincipper, til beskæftigelsesministeren

periment arises. To give a complete picture of what drives the results we study both

the actions taken by the caseworkers and the accompanying reactions by the job seek-

ers. The richness and the quality of the data allow us to study mechanisms in detail.

Information about caseworker actions is obtained from several diﬀerent administrative

records gathered by the PES. The information on search behavior comes from activity

reports ﬁled by the job seekers every month. The reports provide information about all

search activities during the past month. As mentioned in Section 3.1, the job seekers

risk beneﬁt sanctions if they fail to provide a report and caseworkers can validate the

job applications in the report.

The direct eﬀect of job search assistance may arise through a number of possible

channels. We broadly divide these mechanisms into three strains. First, caseworkers

may use the more frequent meetings to increase the monitoring, so that they detect

more violations of the rules that the job seekers should follow (Graversen and van Ours,

2008a,b). From a job seeker perspective, tighter monitoring should lead to increased job

search eﬀort (more job applications).

Second, the JSA program may give caseworkers

more time to provide job search training and related support. Caseworkers may, for

instance, use the meetings to help job seekers write better CVs and job applications,

prepare them for interviews, or provide valuable information about job search strategies.

Any intensiﬁed job search training may cause the job seekers to alter their search

strategies, perhaps by promoting other channels besides formal applications, which

in turn may lead to faster job ﬁnding.

Both monitoring and support have been in

focus in some previous studies (Meyer, 1995; Ashenfelter et al., 2005; Van den Berg

and Van der Klaauw, 2006). We contribute to this literature by presenting evidence

that builds on a large-scale randomized trial combined with detailed register data on

caseworkers’ actions and job seeker search behaviour.

Tighter monitoring may also aﬀect reservation wages. However, the results in Section 4.3 reveal

no wage eﬀects for our experiment.

Previous studies showing that labor market policies can aﬀect search strategies include Van den

Berg and Van der Klaauw (2006) and Bonoli et al. (2014).

Electronic copy available at: https://ssrn.com/abstract=3515935

B 108 - 2021-22 - Endeligt svar på spørgsmål 1: Spm. om beregninger af effekten af digitale samtaler fra første dag på baggrund af de opdaterede regneprincipper, til beskæftigelsesministeren

A third possible channel that has been discussed to a lesser extent is that casework-

ers can provide information about relevant vacancies. Vacancy referrals may lead to

broader job search, both in terms of geography and in terms of occupation. They may

also help the job seekers pinpoint the most relevant job openings, without changing

how broad the search is. Both explanations suggest that vacancy referrals can speed

up the exit from unemployment. To study this mechanism, we exploit data from the

administrative registers on the number of vacancy referrals provided by the casework-

ers. We pair this with data on job seekers’ search behaviour to study to what extent

the job seekers take advantage of the referrals.

Concerning data details, we focus on the ﬁrst quarter of unemployment, which is the

period during which the JSA program took place.

For each outcome, we estimate the

model in equation (1), so that the reported coeﬃcients represent the diﬀerence between

the treated and the non-treated job seekers at the active oﬃces during the experiment

in 2015. One concern is that the extra meetings aﬀect the likelihood to ﬁle an activity

report, for instance, if the caseworkers push the treated job seekers to submit reports.

However, as shown in Panel A of Table 6 there is no diﬀerence between the treated and

the non-treated in the propensity to report.

Panel A of Table 6 shows diﬀerences between treated and non-treated job seekers

with respect to monitoring and search eﬀort. In Sweden, caseworkers are responsible

for all monitoring of job search activities. If caseworkers observe a violation of the

job search rules, they should notify the unemployment insurance funds, which then

decide about beneﬁt sanctions.

Here, we use data on these registered violations of

We focus on activity reports of individuals who are still unemployed when the report is supposed

to be submitted. We exclude job seekers who, according to the PES registers, are not supposed to

report their activity. Data on caseworker actions include information for the entire target population.

Beneﬁt sanctions are monetary ﬁnes (suspension of unemployment insurance beneﬁts). Refusal of

job oﬀers, insuﬃcient job search, not showing up for meetings with PES, not applying for assigned jobs,

and job quits may lead to a sanction. The size of the sanction depends on the type of violation, but

the general rules are that the ﬁrst violation leads to a warning; the second to fourth violations imply

suspension for one day, ﬁve days, ten days, respectively; and the ﬁfth violation leads to a permanent

suspension.

Electronic copy available at: https://ssrn.com/abstract=3515935

B 108 - 2021-22 - Endeligt svar på spørgsmål 1: Spm. om beregninger af effekten af digitale samtaler fra første dag på baggrund af de opdaterede regneprincipper, til beskæftigelsesministeren

the rules. Initially, column 1 shows that the total number of registered violations is

higher for the treated job seekers. However, when we separate between diﬀerent types

of violations, it is only violations due to failure to show up for meetings that are aﬀected

(column 2). This diﬀerence between treated and non-treated job seekers is expected:

since the intensiﬁed JSA implies more meetings, we expect a mechanical eﬀect here. In

contrast, violations regarding insuﬃcient job search eﬀort (column 3) are unaﬀected,

suggesting that the caseworkers’ monitoring of search eﬀort did not change. Other

types of sanctions are also unaﬀected (not reported).

The ﬁrst three columns of Panel A focused on caseworkers’ actions. Next, we turn

to the corresponding changes to job seekers’ search behavior. Since the treated job

seekers do not face more intensive monitoring, we do not expect to see any diﬀerence in

terms of job search eﬀort. This is also supported by the data. There are no diﬀerences

in terms of the total number of activities, or the total number of job applications

(columns 6–7). We conclude that increased search eﬀort—induced by a higher degree

of monitoring—hardly explains the observed positive direct eﬀect of the JSA program.

Next, Panel B looks at diﬀerences with respect to job search support and training.

One of the caseworkers’ initial tasks after an individual has registered at the PES is

to set up an action plan. The plan constitutes a mutual agreement between the PES

and the job seeker, and speciﬁes both the type of support and training that the PES

commits to oﬀer, and what actions the job seeker should undertake to ﬁnd a job.

The action plan consists of up to eight diﬀerent support categories and can be revised

as the job search continues.

Columns 1–4 of Panel B show no diﬀerence between

the treated and non-treated with respect to support and training. The probability of

having an action plan is the same across groups. Both the total number of support

categories in the action plan and the number of categories that speciﬁcally capture job

The categories are: Search for jobs, Improve your search, Guidance to work, Education to work,

Start new business, Clarify your qualiﬁcations for work, Adapt your work situation, and Work prepara-

tory measures. We deﬁne Improve your search and Guidance to work to be categories that speciﬁcally

capture job search training.

Electronic copy available at: https://ssrn.com/abstract=3515935

B 108 - 2021-22 - Endeligt svar på spørgsmål 1: Spm. om beregninger af effekten af digitale samtaler fra første dag på baggrund af de opdaterede regneprincipper, til beskæftigelsesministeren

Table 6: Eﬀects of the JSA program on caseworker actions and job seeker search

behavior

Panel A:

Total

viola-

tions

(1)

Assigned to

program

Obs.

Panel B:

0.037

∗∗∗

(0.011)

[0.337]

19,674

Caseworker action:

Monitoring

Viola-

tions

absence

meeting

(2)

0.035

∗∗∗

(0.005)

[0.059]

19,674

Viola-

tions job

eﬀort

(3)

0.000

(0.001)

[0.010]

19,674

Search behavior:

Search eﬀort

Prob.

reporting

Total

activi-

ties

(6)

0.152

(0.337)

[14.46]

11,959

Total

job appli-

cations

(7)

0.032

(0.249)

[7.811]

11,959

(5)

−0.003

(0.006)

[0.520]

19,674

Caseworker action:

Job search support and training

Has

action

plan

(1)

Support

cat.

in plan

(2)

0.002

(0.015)

[1.520]

14,658

Job search

supp. cat.

in plan

(3)

0.011

(0.009)

[0.285]

14,658

ALMP

parti-

cipant

(4)

0.005

(0.004)

[0.082]

19,674

Search behavior:

Search channels

Unsolicited

job

applications

(5)

−0.140

(0.111)

[2.546]

11,959

Other job-

enhancing

activities

(6)

0.266

∗

(0.140)

[3.417]

11,959

Assigned to

program

Obs.

Panel C:

−0.010

(0.006)

[0.752]

19,674

Caseworker action:

Vacancy referrals

Total

(1)

Sugge-

sted

(2)

0.079

∗

(0.044)

[1.579]

19,674

Manda-

tory

(3)

0.034

∗∗

(0.016)

[0.139]

19,674

Total

(5)

Search behavior:

Applications to

vacancy referrals

Sugge-

sted

(6)

0.071

∗∗∗

(0.024)

[0.427]

11,959

Manda-

tory

(7)

0.029

∗∗∗

(0.007)

[0.033]

11,959

Assigned to

program

Obs.

0.112

∗∗

(0.048)

[1.718]

19,674

0.101

∗∗∗

(0.026)

[0.460]

11,959

Notes: Results from a linear regression on a treatment indicator, weighted by the observed ITT-share. All

outcomes measured during the 1st quarter of unemployment. Standard errors in parentheses, control means in

square brackets.

∗∗∗

p<0.01,

∗∗

p<0.05,

∗

p<0.1.

Electronic copy available at: https://ssrn.com/abstract=3515935

B 108 - 2021-22 - Endeligt svar på spørgsmål 1: Spm. om beregninger af effekten af digitale samtaler fra første dag på baggrund af de opdaterede regneprincipper, til beskæftigelsesministeren

search training are unaﬀected. Finally, we see no diﬀerence regarding the probability

of participating in an active labor market program.

Turning to the job seekers’ search behavior (columns 5 and 6) there is no trace of any

impact on job seekers’ search channels: column 5 shows no eﬀect on the probability

of ﬁling unsolicited job applications, and column 6 shows only a small eﬀect on the

likelihood of reporting other job-enhancing activities. Summing up, the evidence in

Panel B speaks against an increased amount of job search support and training as an

important channel.

Panel C looks at whether the job seekers who participated in the JSA program

became better informed in terms of which vacancies to apply to. We distinguish between

two types of vacancy referrals, suggested job openings and mandatory referrals. The

mandatory referrals are mainly used when caseworkers believe that there is a good

match between the vacancy requirements and the job seeker’s skills. They are only

used for individuals collecting unemployment beneﬁts, who risk beneﬁt sanctions in

case of insuﬃcient job search. For suggested vacancies, the job seeker is free to choose

whether to apply or not. They include both vacancies posted at the online PES job

board and job openings communicated from employers to the PES without being posted

as formal vacancies. Columns 1–3 show that the JSA program raised the total number

of vacancy referrals passed on from caseworkers to job seekers by 0.11 or 6.5%. This is

due to both more suggested job openings (up 5%) and more mandatory referrals (up

25%). Below we show that this increase is not due to displacement of resources with

fewer referrals given to the non-treated.

We next study if job seekers reacted to the information provided by the casework-

ers. Here, columns 5–7 show that the job seekers take full advantage of the vacancy

referrals. The number of applications to referred vacancies increases by 0.10 during

the ﬁrst quarter, which is close to the increase in the number of referrals provided by

the caseworkers. The fact that both sources of information—administrative records of

Electronic copy available at: https://ssrn.com/abstract=3515935

B 108 - 2021-22 - Endeligt svar på spørgsmål 1: Spm. om beregninger af effekten af digitale samtaler fra første dag på baggrund af de opdaterede regneprincipper, til beskæftigelsesministeren

caseworker actions and the monthly reports on job seekers’ search behavior—paint the

same picture gives strong support for the importance of the vacancy referral channel.

We take this as evidence that caseworkers use their expertise to prepare suitable va-

cancies, point the job seekers to these vacancies during the extra meetings, and that

the job seekers follow the advice and apply to the jobs.

The next question is why vacancy referrals speed up the exit from unemployment.

One explanation is that vacancy referrals lead to broader job search, for instance in

terms of occupation or geographical search area.

Job seekers who search broader

already from the start of the unemployment spell may ﬁnd a job faster. Another

explanation is that the vacancy referrals help job seekers pinpoint the most relevant

job openings in the market, without changing the targeted occupations and the search

area. Put diﬀerently, the JSA program may generate more vacancy information from

caseworkers, which helps the treated job seekers apply to the most relevant jobs earlier.

Table 7: Impact on diﬀerent types of vacancy referrals

2-digit occupation

Within

Outside

(1)

(2)

Assigned to program

Control mean

Observations

0.054

∗

(0.030)

1.106

26,538

0.050

∗

(0.027)

0.932

26,538

County of residence

Within

Outside

(3)

(4)

0.088

∗∗

(0.037)

1.670

26,538

0.015

(0.017)

0.368

26,538

Notes: The results are from a linear regression of each variable on a treatment indicator, weighted by the observed

intention to treat share. The sample includes the active PES oﬃces during the experiment period. The control

variables include the variables in Table 1. Standard errors in parentheses, control means in square brackets.

∗∗∗

p<0.01,

∗∗

p<0.05,

∗

p<0.1. ITT are intention-to-treat-eﬀects for individuals who were supposed to be randomized

to treatment at the active oﬃces. IV are instrumental-variables-estimates for those actually randomized to treatment

at the active oﬃces (given by the ﬁrst line in Panel A in Table 2).

We explore these explanations by studying eﬀects on diﬀerent types of vacancy

referrals. We exploit occupational and geographic information for each referral and

distinguish between referrals within and outside the jobseeker’s preferred occupation

(2 digit level), and referrals within and outside the county of residence. The results

Manning and Petrongolo (2017) show that the attractiveness of jobs decays with the distance to

the job. To what extent referrals from caseworkers can push job seekers to apply to more distant jobs

is an open question.

Electronic copy available at: https://ssrn.com/abstract=3515935

B 108 - 2021-22 - Endeligt svar på spørgsmål 1: Spm. om beregninger af effekten af digitale samtaler fra første dag på baggrund af de opdaterede regneprincipper, til beskæftigelsesministeren

from this exercise in Table 7 show that the distribution of referrals is unaﬀected by the

treatment: both referrals within and outside the occupation as well as referrals within

and outside the region increase by around 5%. It means that the treated received more

of the same, rather than new types of referrals. This supports the second explanation

above, that vacancy information—while leaving the direction of search unaﬀected—

streamlines the search process by pointing job seekers to the most relevant jobs already

from the start of the unemployment spell.

Comparing diﬀerent types of meetings

Our experiment uses three types of meetings between job seekers and caseworkers: face-

to-face, distance and group meetings. In this section, we present additional evidence

on the mechanisms behind the direct eﬀect by comparing these three meeting formats.

Since the type of meeting was randomized across oﬃces, we can make a credible com-

parison of the formats. To support this, we have conﬁrmed that we have balanced

groups for each type of meeting (not reported).

The regression results in Table 8 (ITT eﬀects) show that while face-to-face and dis-

tance meetings both signiﬁcantly increase exits out of unemployment, group meetings

appear to be less eﬀective. Overall, group meetings display smaller point estimates, and

for number of days in unemployment (1st quarter and 1st year) there is no signiﬁcant

impact. The smaller employment eﬀects for group meetings are consistent with the

results in previous studies (see, e.g., Maibom et al., 2017).

Above we presented evidence showing that vacancy referrals are key in explaining

the direct eﬀect of the job search assistance. One reason why group meetings perform

worse could be that caseworkers who provide support to many job seekers at the same

time do not have time to prepare and discuss suitable job openings with each job seeker.

Table 9 provides support to this interpretation. It shows that group meetings is the

Electronic copy available at: https://ssrn.com/abstract=3515935

B 108 - 2021-22 - Endeligt svar på spørgsmål 1: Spm. om beregninger af effekten af digitale samtaler fra første dag på baggrund af de opdaterede regneprincipper, til beskæftigelsesministeren

Table 8: Eﬀects of the JSA program on unemployment, by type of meeting

Face-to-face

meetings

ITT

(1)

Exit unemp. 1st quarter

0.043

∗∗∗

(0.009)

[0.329]

−2.003

∗∗∗

(0.478)

[75.92]

−6.780

∗∗∗

(2.247)

[204.4]

10,567

Distance

meetings

ITT

(2)

0.034

∗∗∗

(0.010)

[0.367]

−1.396

∗∗

(0.565)

[73.40]

−7.308

∗∗∗

(2.520)

[191.9]

8,259

Group

meetings

ITT

(3)

0.024

∗∗

(0.011)

[0.375]

−0.950

(0.582)

[73.30]

−3.413

(2.626)

[189.3]

7,712

Unemp. days 1st quarter

Unemp. days 1st year

Observations

Notes: ITT estimates from a linear regression of each variable on a treatment indicator, weighted by the observed

intention to treat share. The sample includes the active PES oﬃces during the experiment period. The control

variables include the variables in Table 1. Standard errors in parentheses, control means in square brackets.

∗∗∗

p<0.01,

∗∗

p<0.05,

∗

p<0.1.

only type of assistance that did not involve more vacancy referrals—there is no eﬀect

on caseworker actions nor on job seeker search behavior. In summary, we take this as

additional evidence that information about vacancies is the main mechanism behind

the eﬀects of the JSA program.

It seems likely that distance meetings will become an increasingly important ele-

ment of future JSA programs. Recent technological advancements make online com-

munication a convenient complement to more traditional ways of providing assistance,

especially for job seeker with long travel time to the local PES oﬃce. However, there

is still a lack of evidence on the eﬀectiveness of using new technologies when providing

JSA and other forms of labor market policies. Here, we have shed some light on this

question by showing that the direct eﬀect of JSA is independent of whether the support

from the caseworkers is given face-to-face or via distance meetings.

Electronic copy available at: https://ssrn.com/abstract=3515935

B 108 - 2021-22 - Endeligt svar på spørgsmål 1: Spm. om beregninger af effekten af digitale samtaler fra første dag på baggrund af de opdaterede regneprincipper, til beskæftigelsesministeren

Table 9: Type of meeting and information about vacancies

Caseworker action:

Vacancy referrals

Total

(1)

Face-to-face

meetings

Obs.

Distance

meetings

Obs.

Group

meetings

Obs.

0.194

∗∗

(0.079)

[1.750]

8,092

0.195

∗

(0.101)

[2.193]

5,982

−0.039

(0.082)

[1.752]

5,600

Suggested

(2)

0.144

∗∗

(0.070)

[1.572]

8,092

0.082

(0.091)

[2.042]

5,982

−0.019

(0.078)

[1.652]

5,600

Mandatory

(3)

0.050

(0.031)

[0.178]

8,092

0.113

∗∗∗

(0.033)

[0.151]

5,982

−0.020

(0.022)

[0.100]

5,600

Total

(4)

0.114

∗∗∗

(0.039)

[0.419]

4,908

0.151

∗∗∗

(0.055)

[0.584]

3,653

0.031

(0.038)

[0.382]

3,398

Search behavior:

Applications to

vacancy referrals

Suggested

(5)

0.077

∗∗

(0.036)

[0.383]

4,908

0.116

∗∗

(0.052)

[0.543]

3,653

0.020

(0.036)

[0.364]

3,398

Mandatory

(6)

0.037

∗∗∗

(0.012)

[0.036]

4,908

0.035

∗∗∗

(0.013)

[0.042]

3,653

0.012

(0.012)

[0.0186]

3,398

Notes: ITT estimates from a linear regression on a treatment indicator, weighted by the observed ITT-share. In

columns 1–3, the outcome variables are for the 1st quarter after registration. Outcome variables in columns 4–6

are the sum over the monthly activity reports in the ﬁrst quarter. Standard errors in parentheses, control means in

square brackets.

∗∗∗

p<0.01,

∗∗

p<0.05,

∗

p<0.1.

How do displacement eﬀects arise?

Our previous analyses showed that the JSA program gives rise to substantial displace-

ment eﬀects for the non-treated job seekers. This raises the question why displacement

exists. One explanation is that it arises in the labor market: if there are more job

seekers than vacant positions, targeted job search assistance can lead to a game of mu-

sical chairs where the non-treated job seekers end up last in line for the vacant jobs. A

second explanation is that resources are allocated from the non-treated to the treated,

so that the non-treated are oﬀered less assistance than in the baseline. From a pol-

icy perspective, it is crucial to distinguish between these two sources of displacement.

While expanding the JSA budget solves the problem with constrained PES resources,

it is much harder to come up with policies addressing displacement in the labor market.

This section adds to the literature by presenting evidence that discriminates between

Electronic copy available at: https://ssrn.com/abstract=3515935

B 108 - 2021-22 - Endeligt svar på spørgsmål 1: Spm. om beregninger af effekten af digitale samtaler fra første dag på baggrund af de opdaterede regneprincipper, til beskæftigelsesministeren

these two sources of displacement.

We ﬁrst investigate displacement through resource

constraints by exploiting the same detailed administrative data as above. Next, we look

directly at displacement in the labor market by estimating displacement eﬀects in tight

and slack labor markets.

Table 10 presents displacement eﬀect estimates, comparing the treated and the

non-treated at the active oﬃces with the non-treated at the non-active oﬃces, for the

same variables as in Section 5. For the non-treated, monitoring (Panel A), support and

training (Panel B), and vacancy information (Panel C) is at the same level as if the

experiment would not have taken place, all pointing in the direction of no displacement

of resources. Using data covering the universe of contacts between job seekers and

the PES, we also present evidence against displacement of meetings. Columns 1–3 of

Table 11 show that the increase in the number of meetings for the treated is not due

to less meetings for the non-treated at the active oﬃces.

We next examine if non-treated job seekers were assigned to diﬀerent types of case-

workers. To this end, we exploit information on caseload and tenure. For instance,

caseworkers involved in the program may have handed over cases to personnel outside

the program, in order to free up time for the extra meetings. If so, this would increase

the workload for caseworkers working with the baseline assistance, which may aﬀect

the quality of the services provided to the non-treated. In addition, the local oﬃces

may have allocated their most tenured caseworkers to the extra meetings. Both these

examples would imply a re-allocation of resources away from non-treated job seekers.

To measure caseload, we count the number of meetings per month and the number of

unique job seekers that the caseworker meets each month. Tenure is measured in days,

and we study both overall tenure at the PES and tenure within the local oﬃce where

the caseworker currently works.

The results reported in columns 4–7 of Table 11 con-

Cr´pon et al. (2013) and Gautier et al. (2018) take displacement in the labor market as given

without explicitly analyzing displacement of resources.

Since our data on meetings start in 2010 we count tenure from this year.

Electronic copy available at: https://ssrn.com/abstract=3515935

B 108 - 2021-22 - Endeligt svar på spørgsmål 1: Spm. om beregninger af effekten af digitale samtaler fra første dag på baggrund af de opdaterede regneprincipper, til beskæftigelsesministeren

Table 10: Displacement eﬀects of the JSA program on caseworker actions and job

seeker search behavior

Panel A:

Total

viola-

tions

(1)

Assigned to

program

In a program

area

Obs.

Panel B:

Has

action

plan

(1)

Assigned to

program

In a program

area

Observations

Panel C:

−0.015

∗∗

(0.006)

0.004

(0.009)

[0.760]

694,772

0.029

∗∗∗

(0.010)

0.009

(0.010)

[0.290]

247,714

Caseworker action:

Monitoring

Viola-

tions

contact

(2)

0.031

∗∗∗

(0.006)

0.003

(0.006)

[0.078]

247,714

Viola-

tions job

(3)

−0.000

(0.001)

−0.001

(0.002)

[0.010]

247,714

Search behavior:

Search intensity

Prob.

reporting

(5)

−0.006

(0.006)

0.012

∗

(0.006)

[0.562]

210,779

Total

activi-

ties

(6)

0.054

(0.270)

−0.207

(0.336)

[16.4]

135,961

Total

job appli-

cations

(7)

−0.001

(0.194)

−0.272

(0.197)

[8.87]

135,961

Caseworker action:

Support and training

Support

cat.

in plan

(2)

−0.002

(0.016)

0.017

(0.019)

[1.62]

525,055

Job search

supp. cat.

in plan

(3)

0.010

(0.009)

0.022

(0.014)

[0.360]

525,055

ALMP

parti-

cipant

(4)

0.004

(0.003)

0.005

(0.005)

[0.051]

694,772

Search behavior:

Search channels

Unsoli-

cited job

appl.

(5)

−0.136

(0.107)

−0.002

(0.114)

[2.85]

135,961

Other job-

enhancing

activities

(6)

0.208

(0.155)

0.089

(0.114)

[3.58]

135,961

Caseworker action:

Vacancy referrals

Total

(1)

Sugge-

sted

(2)

0.062

(0.049)

−0.022

(0.069)

[0.979]

694,772

Manda-

tory

(3)

0.040

(0.036)

0.001

(0.015)

[0.152]

694,772

Total

(5)

Search behavior:

Applications to

vacancy referralls

Sugge-

sted

(6)

0.069

∗∗

(0.027)

−0.016

(0.022)

[0.351]

135,961

Manda-

tory

(7)

0.028

(0.018)

−0.004

(0.006)

[0.045]

135,961

Assigned to

program

In a program

area

Obs.

0.102

(0.066)

−0.021

(0.072)

[1.13]

694,772

0.097

∗∗∗

(0.035)

−0.020

(0.023)

[0.397]

135,961

Notes: Regression of each outcome variable on an indicator for active PES oﬃce (“In a program area”), an indicator

for active PES oﬃce

intention to treat status is treated (“Assigned to program”) and indicator variables for

year, month and PES oﬃce. The sample includes all oﬃces: active oﬃces and non-active oﬃces, so the excluded

category is the non-active oﬃces. Standard errors clustered at the PES oﬃce level in parentheses, control means

in square brackets.

∗∗∗

p<0.01,

∗∗

p<0.05,

∗

p<0.1.

Electronic copy available at: https://ssrn.com/abstract=3515935

B 108 - 2021-22 - Endeligt svar på spørgsmål 1: Spm. om beregninger af effekten af digitale samtaler fra første dag på baggrund af de opdaterede regneprincipper, til beskæftigelsesministeren

ﬁrm that the JSA program did not lead to any displacement of resources: irrespective

of whether we study caseload (columns 4–5) or tenure (columns 6–7) it was not the

case that the non-treated at the active oﬃces received less resources.

Table 11: Displacement of meetings and services

Displacement of meetings

Meetings

quarter

(1)

Assigned to

program

In a program

area

Obs.

0.547

∗∗∗

(0.058)

−0.035

(0.054)

[3.472]

451,472

Physical

meetings

(2)

0.360

∗∗∗

(0.057)

−0.061

(0.041)

[2.617]

451,472

Distance

meetings

(3)

0.187

∗∗∗

(0.053)

0.026

(0.027)

[0.854]

451,472

Displacement of caseworkers

No.

meetings/

month

(4)

2.69

(2.15)

−0.163

(5.24)

[152.5]

451,406

No.

clients/

month

(5)

1.81

(1.62)

0.975

(3.29)

[103.2]

451,406

Tenure

at PES

in days

(6)

4.59

(9.54)

−20.5

(18.8)

[1,185]

451,406

Tenure at

local oﬃce

in days

(7)

0.004

(13.7)

−11.5

(25.2)

[943.9]

451,406

Notes: Regression of each outcome variable on an indicator for active PES oﬃce (“In a program area”), an

indicator for active PES oﬃce

intention to treat status is treated (“Assigned to program”) and indicator

variables for year, month and PES oﬃce. The sample includes all oﬃces: active oﬃces and non-active oﬃces, so

the excluded category is the non-active oﬃces. Standard errors clustered at the PES oﬃce level in parentheses,

control means in square brackets.

∗∗∗

p<0.01,

∗∗

p<0.05,

∗

p<0.1.

We now turn to evidence of displacement in the labor market. Cr´pon et al. (2013)

set up a theoretical model predicting that displacement is higher in weak labor markets,

and provide empirical support for this prediction. Table 12 reproduces this analysis

in our setting. We estimate our main model from column 5 in Table 3, but classify

all oﬃces according to whether the monthly local unemployment rate is above or be-

low median unemployment among the Swedish municipalities and interact this binary

variable with treatment status.

The results are striking. The displacement eﬀect is

considerably larger in high-unemployment labor markets than in low-unemployment

markets, but there is no signiﬁcant diﬀerences for the direct eﬀect.

One interpretation of these results is as follows. The direct eﬀect arises since the

treated job seekers gain from additional information about vacancies. As expected, the

vacancy information channel works under any labor market conditions, tight or slack.

We also include the monthly local unemployment to population ratio as an additional regressor.

It is the average unemployment rate facing the job seekers in the target population of the experiment

per oﬃce and month, based on their municipality of residence.

Electronic copy available at: https://ssrn.com/abstract=3515935

B 108 - 2021-22 - Endeligt svar på spørgsmål 1: Spm. om beregninger af effekten af digitale samtaler fra første dag på baggrund af de opdaterede regneprincipper, til beskæftigelsesministeren

Table 12: Displacement eﬀects of the JSA program in strong and weak labor markets

Exit

unemp.

1st quarter

(1)

Assigned to program×Below median unemployment rate

Assigned to program×Above median unemployment rate

In a program area×Below median unemployment rate

In a program area×Above median unemployment rate

Observations

0.027

∗∗∗

(0.009)

0.043

∗∗∗

(0.009)

−0.001

(0.008)

−0.030

∗∗

(0.011)

552,816

Unemp.

days

1st quarter

(2)

−1.416

∗∗

(0.560)

−1.588

∗∗∗

(0.592)

0.028

(0.468)

1.413

∗∗∗

(0.521)

552,816

Unemp.

days

1st year

(3)

−5.372

∗∗

(2.539)

−7.763

∗∗∗

(2.559)

−0.828

(2.506)

9.207

∗∗∗

(2.672)

552,816

Notes: Regression of each outcome variable on an indicator for treatment PES oﬃce (“In a program area”)

and an indicator for treatment PES oﬃce

intention to treat status is treated (“Assigned to program”), both

interacted with whether the unemployment level in the municipality of the local PES oﬃce was below (“Below

median unemployment rate”) or above (“Above median unemployment rate”) the median municipality, as well

as year dummies, month dummies, local PES dummies and the monthly unemployment rate in the PES oﬃce

municipality. Standard errors clustered at the PES oﬃce level in parentheses, control means in square brackets.

∗∗∗

p<0.01,

∗∗

p<0.05,

∗

p<0.1.

But, the harm for the non-treated job seekers is limited under good labor market

conditions, when access to alternative jobs is good. In contrast, when competition for

jobs increases, the non-treated are hurt by the fact that the pool of candidates that

ﬁrms chose from is larger.

Structural evaluation of the equilibrium eﬀects

So far we have examined equilibrium eﬀects of the experiment on exits from unem-

ployment, wages, vacancies and job applications. We now study the implications of a

full-scale roll-out of the program on labor market outcomes, government expenditures

and welfare. To this end, we estimate a Diamond-Mortensen-Pissarides (DMP) model

using the equilibrium eﬀect estimates from the experiment. To this end, we build on

the model by Gautier et al. (2018), but adjust it in accordance with the job search

policy evaluated in this paper. In the model, the workers (job seekers) choose the num-

ber of job applications they ﬁle and ﬁrms post vacancies. A key feature of the model

Electronic copy available at: https://ssrn.com/abstract=3515935

B 108 - 2021-22 - Endeligt svar på spørgsmål 1: Spm. om beregninger af effekten af digitale samtaler fra første dag på baggrund af de opdaterede regneprincipper, til beskæftigelsesministeren

is the endogenous matching function in which the success of an application depends

on the number of applications sent by other workers, which creates search congestion.

To capture the empirical results that the program participants ﬁnd jobs faster without

increasing the total number of job applications, the matching function allows for diﬀer-

ential success rates for participants and non-participants. The model is estimated by

indirect inference using the results from the experiment presented in previous sections.

Since we wish to capture the behavior of actual program participants, we use the IV

estimates of the program eﬀect for the participants. This is why the model refers to

participants and non-participants.

8.1

The model

The model deﬁnes an equilibrium in which search intensity, wages, unemployment

and labor market tightness are determined. It is a discrete-time model with ex ante

identical and risk neutral workers with the same productivity, who only diﬀer in whether

or not they participate in the JSA program (indexed by 0 or 1). The unemployed

receive beneﬁts

and the value of non-market time is

In each period, the worker

decides the number of job applications to ﬁle,

trading oﬀ job prospects and search

costs. For convenience, we assume that the search costs are quadratic in the number

of applications (γa

). If an application is successful the worker becomes employed,

otherwise (s)he has to apply again in the next period.

A key part of the model is the endogenous matching function that captures that

the success of an application depends on the number of applications sent by other

workers. Speciﬁcally, the matching function

m(a; a, θ)

is increasing in the number of

own job applications, decreasing in average search intensity of other workers,

, and

increasing in labor market tightness,

≡

v/u,

where

is the unemployment rate and

the vacancy rate. The matching function is derived below. Finally, let

be the discount

rate, and

E(w)

the value of being employed at the wage rate

Then, the value of

Electronic copy available at: https://ssrn.com/abstract=3515935

B 108 - 2021-22 - Endeligt svar på spørgsmål 1: Spm. om beregninger af effekten af digitale samtaler fra første dag på baggrund af de opdaterede regneprincipper, til beskæftigelsesministeren

unemployment for a non-participant is

= max

−

(a;

a, θ)

[E(w

)

−

]

a≥0

(4)

From the ﬁrst order condition we have that the optimal number of applications is:

E(w

)

−

∂m

;

a, θ)

2γ

∂a

∗

(5)

The participants receive more meetings with caseworkers through the JSA program.

We allow this to have three separate eﬀects. First, obtaining job search assistance may

reduce the participants’ search costs,

, in relation to the costs of non-participants,

Second, participating in the program costs non-market time. We normalize its value

to zero for participants and let it be

for the non-participants. In the estimations,

we do not restrict the sign of

even though our prior is that the value of non-market

time is lower for the participants. Third, we specify separate matching functions for

participants and non-participants under the assumption that the treatment may aﬀect

job-search eﬃciency and the success rate per job-application (see below). Thus, the

value of unemployment for a program participant is

= max

−

;

a, θ)

[E(w

)

−

]

≥0

(6)

and the ﬁrst order condition gives

E(w

)

−

∂m

;

a, θ)

2γ

∂a

∗

(7)

In equilibrium, the average number of applications in the market equals

τ a

∗

+ (1

−

∗

, where

is the share of program participants.

Note that this expression implicitly means that beneﬁts and search costs are realized at the end

of the period. This simpliﬁes the notation.

Electronic copy available at: https://ssrn.com/abstract=3515935

B 108 - 2021-22 - Endeligt svar på spørgsmål 1: Spm. om beregninger af effekten af digitale samtaler fra første dag på baggrund af de opdaterede regneprincipper, til beskæftigelsesministeren

The value of employment depends on the wage,

the exogenous job destruction

rate,

δ,

and the diﬀerence between the ﬂow values of employment and unemployment:

rE(w

) =

−

δ[E(w

)

−

] =

−

[E(w

)

−

]

where

τ U

+ (1

−

is the average utility.

(8)

Firms are also assumed to be identical. The value of a vacancy,

, for a ﬁrm is

determined by the vacancy cost,

, the probability of ﬁlling the vacancy,

the value diﬀerence between a ﬁlled,

J(w),

and an unﬁlled vacancy:

m(a

, a

, θ)

(J(w)

−

m(a

,θ)

and

−c

(9)

The value of a ﬁlled vacancy is given by

rJ(w)

−

δ(J(w)

−

(10)

where

is the value of output in each period. With free entry, the value of a vacancy

is zero in equilibrium. From (9), (10) and this zero-proﬁt condition we have

m(a

, a

;

τ, θ)

(r +

δ)c

∗

−

∗

(11)

which can be solved for the equilibrium value of tightness,

∗

, since the left-hand side

is decreasing in

and the right-hand side is increasing in

θ,

so that there is a unique

∗

that satisﬁes equation (11).

We now specify the matching functions for participants and non-participants, allow-

ing for diﬀerent search intensities and diﬀerent success rates per application for the two

groups. We use an urn-ball model, implying that a ﬁrm receiving many applications

Workers presumably realized that the JSA program was temporary. We therefore replace

in (8)

when estimating the model, but use

τ U

+ (1

−

in the policy simulations, assuming

that workers expect to be non-participants if they re-enter unemployment after losing their job.

Electronic copy available at: https://ssrn.com/abstract=3515935

B 108 - 2021-22 - Endeligt svar på spørgsmål 1: Spm. om beregninger af effekten af digitale samtaler fra første dag på baggrund af de opdaterede regneprincipper, til beskæftigelsesministeren

randomly selects one and declines the others. This creates search congestion, since the

probability that an application is accepted depends on the number of applications sent

by other workers. Note that the empirical analyses in earlier sections show that the

JSA program leads to more vacancy referrals and a higher job-ﬁnding rate, without

changing the total number of job applications. This implies a higher probability that

an application results in a job oﬀer for the participants than for the non-participants.

To incorporate this, we allow the job applications from the participants to have a

higher probability of being drawn from the pool of candidates than the applications

from non-participants.

Speciﬁcally, we introduce a parameter,

ω,

such that an application from a program

participant results in a job oﬀer with probability

ω+ωj

if the number of participant

competitors equals

and the number of non-participant competitors for the job is

For non-participants this probability is

1+ωj

By allowing both search costs and

the success rate per application to diﬀer, the model shows to what extent the higher

job-ﬁnding rate for the program participants is due to changes in the search costs, or

due to more suitable job applications (higher success rate per application).

If the number of workers and vacancies are suﬃciently large, the number of appli-

cations from participants and non-participants are approximately Poisson distributed

random variables with means

τ a

∗

/θ

and (1

−

∗

/θ,

respectively. Thus, for the non-

participants the probability that an application results in a job oﬀer is:

∞

1 +

ωj

and

) =

exp(−τ

∗

/θ)(τ a

∗

/θ)

(12)

where

) =

exp(−[1−τ ]a

∗

/θ)([1−τ

∗

/θ)

are the probabil-

ity of

applications from participants and

applications from other non-participants.

Electronic copy available at: https://ssrn.com/abstract=3515935

B 108 - 2021-22 - Endeligt svar på spørgsmål 1: Spm. om beregninger af effekten af digitale samtaler fra første dag på baggrund af de opdaterede regneprincipper, til beskæftigelsesministeren

For participants, we have

∞

ωj

(13)

Here, the parameter

captures other eﬀects of the JSA program on job search eﬃciency.

For example, the program may aﬀect the type of jobs the participants apply to, thereby

aﬀecting the share of vacancies for which a participant has a positive productivity. The

idea is that the ﬁrm ﬁrst selects a worker for the job and then learns whether the worker

is productive or not.

All this leads to the matching functions

;

, θ)

= 1−(1−ψ

)

and

;

, θ)

= 1

−

)

, and the aggregate matching function

m(a

, a

, τ, θ)

τ m

, a

, θ)

−

, a

, θ).

Wages are set in a Nash bargaining when workers and ﬁrms have met, with worker

bargaining power equal to

β.

Since participants and non-participants have diﬀerent

outside options,

and

, we allow their equilibrium wages to diﬀer. The bargaining

outcome is given by

∗

= arg max [E(w

)

−

]

[J(w

)

−

]

1−β

(14)

and the ﬁrst order condition gives

∗

−

β) w

δ U

−

(r +

δ)U

−

]

(15)

In equilibrium, inﬂow into and outﬂow from unemployment are equal, and, hence,

equilibrium unemployment is

τ m

∗

;

∗

, θ)

+ (1

−

∗

;

∗

, θ)

1 0

0 1

∗

(16)

The parameter

is included in the model presented in Gautier et al. (2018). We incorporate it in

our model for completeness.

Electronic copy available at: https://ssrn.com/abstract=3515935

B 108 - 2021-22 - Endeligt svar på spørgsmål 1: Spm. om beregninger af effekten af digitale samtaler fra første dag på baggrund af de opdaterede regneprincipper, til beskæftigelsesministeren

∗

The equilibrium can now be deﬁned as

∗

, a

∗

, θ

∗

, w

, u

∗

}

satisfying equations (5),

0 1

(7), (11), (15), and (16).

8.2

Estimation

We estimate the model using the actual share of program participants in the eligible

population. From Section 2, we have that the program participants corresponds to 62%

of those assigned to the program. With a fraction assigned to the program of about

0.5, the share of program participants during the experiment,

, is 0.31. Note that

this share is based on the participants and non-participants included in the experi-

ment. The non-participants also include workers in the stock of unemployed before the

experiment, and workers not in the target population (immigrants and workers with

repeated spells of unemployment). These groups were not included in the experiment,

and the analyses in Section 4.2 revealed no evidence of any displacement eﬀects for

these groups, indicating that the displacement mainly occurs within the target popu-

lation. The absence of displacement outside the experiment explains why we use the

share of participants in the target population. But, below we also report results from

sensitivity analyses using lower treatment shares, implicitly allowing for displacement

eﬀects for the non-target population.

Estimation is based on indirect inference using the reduced form estimates from the

experiment. The model estimates are then used to simulate a full-scale roll-out of the

program. The data moments used in the estimations, adapted to the monthly intervals

in the discrete-time model, are displayed in Table A-4 in the appendix. At the individual

level, we use estimates for the exit rate from unemployment (Table 3) and wages in

subsequent jobs (Table 4) for participants and non-participants. As already mentioned,

IV estimates are used since they reﬂect the program eﬀects for the participants. The

data moments also include the average re-employment rate from the PES data, the

average vacancy rate and unemployment rate facing the target population in each oﬃce

Electronic copy available at: https://ssrn.com/abstract=3515935

B 108 - 2021-22 - Endeligt svar på spørgsmål 1: Spm. om beregninger af effekten af digitale samtaler fra første dag på baggrund af de opdaterede regneprincipper, til beskæftigelsesministeren

based on their municipality of residence, and the average replacement rate (fraction of

previous income replaced by the unemployment insurance) from calculations by the

National Institute for Economic Research (2016). We also use information on the

number of job applications ﬁled by each worker, including both the average number of

job applications and eﬀects on the number of job applications for participants and non-

participants (see Table A-4 for details). The latter helps to capture if the treatment

is due to changed search costs or changed success rate per application. Finally, the

interest rate is set to

= 0.008, which is the monthly interest rate implied by an

annual rate of ten percent, and we normalize productivity to

= 1.

This gives nine unknown parameters to estimate in the model. Using indirect

inference, we minimize the sum of the diﬀerences between the data moments and the

corresponding model moments (see A-4 in the appendix for details). Each data moment

is weighted by the inverse of its variance, so that more precisely estimated moments

are given larger weight. Standard errors are computed using the delta method.

8.3

Government expenditure and welfare

Besides eﬀects on unemployment, wages and labor market tightness, we study gov-

ernment expenditure and welfare. The expenditures include unemployment insurance

beneﬁts,

and program costs,

, per worker. Then, government spending as a function

of the share of participants is

(τ ) =

δ(1

−

u)τ c

(17)

where

is the fraction of unemployed times the unemployment insurance beneﬁts,

and

δ(1

−

u)τ

is the fraction of new entrants into unemployment (job destruction rate,

δ,

times the employment rate, (1

−

u)).

To estimate the costs, we performed a detailed

time-use survey, which was sent out to all caseworkers involved in the program. Case-

Electronic copy available at: https://ssrn.com/abstract=3515935

B 108 - 2021-22 - Endeligt svar på spørgsmål 1: Spm. om beregninger af effekten af digitale samtaler fra første dag på baggrund af de opdaterede regneprincipper, til beskæftigelsesministeren

workers were asked to estimate the average time spent on a meeting, including time for

preparation before the meeting, time spent on the actual meeting as well as time spent

on documentation and other activities after the meeting. Taken together, caseworkers

spent on average 75 minutes per meeting and 225 minutes in total for the three ex-

tra meetings in the program,

corresponding to 2.5% of the monthly working hours.

Next, based on calculations made by the PES, we assume that the average caseworker

wage equals the average wage obtained by the workers in our target population. With

the monthly production,

normalized to one, the average wage is 0.96. We also add

overhead costs for premises, adminstration and managers, estimated to roughly 30%.

In total, the estimated program cost is 0.026 of monthly production.

The welfare eﬀects of the program summarize a number of distinct elements. First,

we have the resources spent on the program,

, assumed to be funded by a non-

distortionary tax to avoid complications introduced by tax incentives and their eﬀects

on job search behavior. Second, we have an eﬀect on output, (1

−

u)y,

of employment

changes induced by the program. Third, we have vacancy costs,

, which vary with

the number of open vacancies,

and the cost per vacancy,

. Fourth, both participants

and non-participants experience search costs as a function of

, γ

and the number of

applications. Finally, program participation implies a loss of non-market time,

total, welfare,

(τ ), is given by

−γ

∗2

−

∗2

(τ ) = (1

−

u)y

−

)

−

δ(1

−

u)τ c

−

(18)

8.4

Results

We now turn to the simulation results. Initially, Panel A of Table A-5 in the appendix

assesses the model ﬁt, displaying the diﬀerence between the moments implied by the

This takes costs associated with cancelled meetings into account since the survey also asked how

often meetings were cancelled and how much time that was lost due to cancelled meetings. These

events account for 3 of the total 75 minutes per meeting.

Electronic copy available at: https://ssrn.com/abstract=3515935

B 108 - 2021-22 - Endeligt svar på spørgsmål 1: Spm. om beregninger af effekten af digitale samtaler fra første dag på baggrund af de opdaterede regneprincipper, til beskæftigelsesministeren

model and the data moments. Generally, the ﬁt is good: the program eﬀects on the

job-ﬁnding rate and the wage rate all match perfectly, as do the aggregate statistics.

Panel B of Table A-5 in the appendix presents the model estimates. Let us comment

on some of the estimates. The estimate of

implies that the participants have a higher

probability to be drawn from the pool of candidates, leading to a higher success rate

per application. Participants also have higher search costs than non-participants (γ

). The higher success rate per application combined with the increased search costs

imply that participants apply to roughly the same number of jobs as non-participants.

This is consistent with the analyses of the mechanisms, which suggest that program

participants receive more vacancy referrals, leading to a higher job-ﬁnding rate even

though they do not apply for more jobs.

Figure 3 presents simulation results for diﬀerent shares of participants, including a

full-scale roll-out. Recall that the reduced form estimates show that the participants

ﬁnd jobs faster, and that the non-participants ﬁnd jobs at a lower rate due to dis-

placement of jobs. The simulation results are qualitatively similar, with a matching

rate almost four percentage points, or around 15%, higher for participants compared

to non-participants. We also see that the matching rate for both participants and

non-participants are decreasing in the treatment share, since a larger share of partic-

ipants creates more search congestion. Overall, however, the aggregate matching rate

increases with the share of program participants. Figure 3 also shows that participants

receive lower wages than non-participants, and that wages for both groups are decreas-

ing in the treatment share. The latter is because more participants create more search

congestion, which lowers the outside options of all workers. The higher matching rate

and the lower wages induce ﬁrms to create more vacancies, so that market tightness is

increasing in the share of participants.

Overall, the results in Figure 3 imply that unemployment decreases with the share of

program participants. Increasing the share of participants from 0 to 100 percent lowers

Electronic copy available at: https://ssrn.com/abstract=3515935

B 108 - 2021-22 - Endeligt svar på spørgsmål 1: Spm. om beregninger af effekten af digitale samtaler fra første dag på baggrund af de opdaterede regneprincipper, til beskæftigelsesministeren

the unemployment rate by around 0.2 percentage points, implying that the net eﬀect of

the JSA program on unemployment is positive. Next, column 1 in Table 13 summarizes

the impact on unemployment, and evaluates the impact on government expenditure

and welfare. We see that that the JSA program reduces government spending, and

that a full-scale roll-out of the program gives the lowest costs. Besides the case with

monthly productivity normalized to one, the change in government expenditures is also

calculated under the assumption that the monthly productivity is SEK 25,000 (average

wage rate for the target population). These calculations show that a full-scale roll-out

would decrease government expenditures by SEK 16 (≈ Euro 1.6 ) per worker, i.e. a

small eﬀect. This means that the decrease in beneﬁt payments as a result of the lower

unemployment rate is of roughly the same magnitude as the direct program costs.

Figure 3: Simulation results for diﬀerent shares of participants

.961

.32

Matching rate

Wage rate

Treated

Average

Non-treated

.28

.26

.957

.958

.959

.96

.24

Treated

Non-treated

Share treated

(a) Matching rate

.36

7.75

(b) Wages

.35

Unemployment rate (%)

Market tightness

.34

.33

.32

.31

7.55

7.6

7.65

7.7

Share treated

(d) Unemployment rate

We next study the welfare implications, which also take changes to search costs,

non-market time, and vacancy costs into account. The simulation results in column 1

Electronic copy available at: https://ssrn.com/abstract=3515935

B 108 - 2021-22 - Endeligt svar på spørgsmål 1: Spm. om beregninger af effekten af digitale samtaler fra første dag på baggrund af de opdaterede regneprincipper, til beskæftigelsesministeren

in Table 13 reveal that welfare is decreasing in the share of participants. Thus, the

positive eﬀect on production from the decreased unemployment rate cannot compensate

for the decreased value of non-market time (participants loose non-market time), the

direct program costs, and the increased vacancy costs induced by the increased vacancy

rate.

Table 13: Model predictions for the equilibrium search model

Main model

Face-to-

face and

distance

meetings

(2)

-0.163

-0.001

-18.5

-0.002

Sensitivity analyses

Delayed

vacancy

model

(3)

-0.321

-0.002

-42.0

0.002

Treatment

25%

(4)

0.004

0.000

4.8

-0.003

Treatment

20%

(5)

0.069

0.001

14.2

-0.004

(1)

Experiment (τ

= 0.31)

Unemployment (measured in %)

Government expenditure

Government expenditure (SEK)

Welfare

Full-scale roll-out

Unemployment (measured in %)

Government expenditure

Government expenditure (SEK)

Welfare

-0.048

-0.000

-1.9

-0.002

-0.228

-0.001

-16.1

-0.007

-0.599

-0.003

-69.4

-0.004

-0.988

-0.005

-128.2

0.006

-0.083

0.000

5.2

-0.009

0.094

0.001

27.7

-0.012

Note: Outcomes are normalized with the monthly output per worker set to 1.

8.5

Sensitivity analyses

We now present results from several sensitivity analyses. One result from the reduced

form analyses is that face-to-face and distance meetings increase the job-ﬁnding rate,

whereas the group meetings do not. The ﬁrst sensitivity analyses therefore excludes

group meetings. To this end, the data moments are re-estimated using only the oﬃces

with face-to-face and distance meetings. We also re-calculate the program costs; face-to-

face and distance meetings are more expensive than group meetings (0.031 of monthly

production compared to 0.026 for all three meeting types). In all other respects the

estimations are the same as before. The results from this exercise in column 2 of

Electronic copy available at: https://ssrn.com/abstract=3515935

B 108 - 2021-22 - Endeligt svar på spørgsmål 1: Spm. om beregninger af effekten af digitale samtaler fra første dag på baggrund af de opdaterede regneprincipper, til beskæftigelsesministeren

Table 13 indicate a larger reduction of the unemployment rate, and a more substantial

decrease of government spending than for the main analyses. However, welfare is still

decreasing in the share of participants. One reason is the larger loss of non-market time

in this model, i.e. the disutility from attending a face-to-face or a distance meeting

appear to be higher than for a group meeting.

Next, the equilibrium conditions presume that the labor market is in the new steady

state during the experiment, which includes the full demand side response with com-

plete adjustment of vacancies. However, ﬁrms may not have time to observe and react

to the new economic environment in the short run and. There was also some evidence

of such a pattern in the data, with vacancies unaﬀected during the earlier parts and

an increased vacancy rate towards the end of the experiment. One may therefore ex-

pect to see further adjustments of vacancies in the long run if the program was made

permanent. If so, the main analyses underestimate the eﬀect of the program. The

reason is that in the short run, any increased job ﬁnding for some participants leads to

search congestion with negative displacement eﬀects for both the non-participants and

other participants. In the long run, however, more job applications and lower wages

induce ﬁrms to open more vacancies, which increases the job ﬁnding rate for both

participants and non-participants, and this pushes down the unemployment rate. To

explore the implications of this, we estimate an alternative version of our model under

the assumption that job-search activity adjusts during the experiment, but that ﬁrms

do not adjust vacancies during the experiment.

This delayed vacancy approach assumes that the participants react to the program, and that both

participants and non-participants realize and respond to the fact that the return to an application

is lower during the experiment. However, we assume that ﬁrms do not observe that the average

wage and matching rates in the labor market have changed, and therefore post vacancies at the same

rate as before the experiment. Speciﬁcally, we assume that the market tightness remains at the pre-

experiment steady state level. We also assume that the wage bargaining is based on the outside options

of the worker, leading to diﬀerent wages for participants and non-participants. Formally, we estimate

the model under the assumption that the labor market is in a steady-state equilibrium before the

experiment. During the experiment, market tightness remains at the pre-experiment level. However,

the number of applications and wages adjust during the experiment and are given by equations (5),

(7) and (15). Under these assumptions, we re-estimate the model and simulate the new steady state,

allowing for the full demand side response with a complete adjustment of vacancies.

Electronic copy available at: https://ssrn.com/abstract=3515935

B 108 - 2021-22 - Endeligt svar på spørgsmål 1: Spm. om beregninger af effekten af digitale samtaler fra første dag på baggrund af de opdaterede regneprincipper, til beskæftigelsesministeren

As expected, this delayed vacancy approach, predicts a larger reduction of unem-

ployment and government expenditures (down by about SEK 85 per worker) than in

the main analyses. Also, interestingly, the welfare eﬀects are now reversed with posi-

tive welfare eﬀects compared to the negative welfare eﬀects found in the main analyses.

In fact, the highest level of welfare is now obtained with a full-scale roll-out of the

program, suggesting that the JSA program may also be welfare enhancing.

Finally, in the main analyses, the share of participants is set to the share in the target

population. However, if non-participants outside the experiment, such as the stock of

unemployed before the experiment and the inﬂow to the non-target population, also

are subject to displacement eﬀects, this implies a lower treatment share. We therefore

explore setting the share of participants to 0.25 and 0.20. The displacement eﬀects is

set to the same level as for the non-participants in the experiment, which means that we

extrapolate the displacement eﬀects for the non-participants in the target population

to the non-participants outside the target population. By construction, this leads to

less positive eﬀects. Table 13 shows that the share of participants in the main analyses

(31%) implies a non-negligible reduction of the unemployment rate (0.2 percentage

points) of a full-scale roll-out of the program. In contrast, a treatment share of 25%

implies a smaller decrease of the unemployment rate (0.08 percentage points), and

with 20% treated program participants, the unemployment rate increases with 0.09

percentage points).

Conclusions

This paper has evaluated direct and displacement eﬀects of job search assistance, using

a large-scale two-level randomized experiment. The JSA program more than doubled

the frequency of meetings with caseworkers at the local public employment oﬃce during

the ﬁrst quarter of unemployment. In line with the previous literature on job search

Electronic copy available at: https://ssrn.com/abstract=3515935

B 108 - 2021-22 - Endeligt svar på spørgsmål 1: Spm. om beregninger af effekten af digitale samtaler fra første dag på baggrund af de opdaterede regneprincipper, til beskæftigelsesministeren

assistance, we ﬁnd that treated job seekers exit unemployment faster than non-treated

job seekers. By exploring detailed data on caseworkers’ actions and job seekers’ search

behavior, we show that the driving mechanism behind the direct eﬀect is an increased

amount of vacancy referrals passed on from caseworkers to job seekers. This suggests

that caseworkers play an important role in bringing job seekers back to work, but also

that it is crucial

how

the extra assistance is designed. We also show that more informa-

tion about vacancies does not lead to broader search, but rather streamlines the search

process by pointing job seekers to the most relevant jobs early in the unemployment

spell.

By comparing diﬀerent meeting formats we ﬁnd additional support for the impor-

tance of vacancy referrals as the driving mechanism. The two meeting types that

involve more referrals—face-to-face meetings and distance meetings—are equally ef-

fective in bringing individuals back to work. In contrast, group meetings show no

increase in referrals and are also less eﬀective. A likely explanation is that it is diﬃcult

to prepare and discuss suitable vacancies during group meetings. Since technological

advancements make distance communication a convenient complement to traditional

face-to-face assistance, the fact that the distance meetings perform well is a highly

policy-relevant ﬁnding.

The experiment was explicitly designed to detect displacement. In addition to the

positive direct eﬀects for the treated, we indeed ﬁnd substantial displacement eﬀects for

the non-treated. We show that the displacement is not an artifact of the experiment due

to crowding out of resources. Instead the displacement is due to displacement of jobs,

as the competitive advantage for the treated due to the increased number of referrals

have negative eﬀects on the non-treated. This is supported by the fact that we see more

displacement in tight than in slack labor markets. It implies that JSA is more eﬃcient

under favorable labor market conditions. The fact that the JSA program is associated

with important equilibrium eﬀects is consistent with recent ﬁndings in the literature

Electronic copy available at: https://ssrn.com/abstract=3515935

B 108 - 2021-22 - Endeligt svar på spørgsmål 1: Spm. om beregninger af effekten af digitale samtaler fra første dag på baggrund af de opdaterede regneprincipper, til beskæftigelsesministeren

(Cr´pon et al., 2013; Gautier et al., 2018). Even if we ﬁnd substantial displacement

eﬀects, our assessment of the beneﬁts of the JSA programs is more positive than in

these recent studies. Overall, the program reduces unemployment since the positive

direct eﬀect on the exit rate outweigh the negative displacement eﬀect.

To trace out the equilibrium eﬀects of a full-scale roll-out of the program, we develop

and estimate an equilibrium search model. One result is that a full-scale implemen-

tation is associated with decreased unemployment, but because of program costs it

has negligible eﬀects on public spending. The impact on welfare is negative, however,

due to the time costs of participants, direct program costs, and vacancy costs from an

increased vacancy rate. We also show that the overall assessment of a full-scale roll-out

hinges on how broadly the displacement eﬀects hit and to what extent ﬁrms react to

the new equilibrium by creating more vacancies. One example is that the welfare eﬀect

reverses to positive when we allow for a delayed vacancy response, taking into account

that it may take time for ﬁrms to observe and react to the new market conditions

created by the JSA program.

Electronic copy available at: https://ssrn.com/abstract=3515935

B 108 - 2021-22 - Endeligt svar på spørgsmål 1: Spm. om beregninger af effekten af digitale samtaler fra første dag på baggrund af de opdaterede regneprincipper, til beskæftigelsesministeren

References

Albrecht, J., G.J. van den Berg, and S. Vroman,

“The aggregate labor market

eﬀects of the Swedish knowledge lift program,”

Review of Economic Dynamics,

2009,

12(1),

129–146.

Arni, P.,

“Opening the Blackbox: How Does Labor Market Policy Aﬀect the Job

Seekers Behavior? A Field Experiment,” mimeo, University of Lausanne 2015.

Ashenfelter, O., D. Ashmore, and O. Deschenes,

“Do unemployment insurance

recipients actively seek work? Evidence from randomized trials in four U.S. States,”

Journal of Econometrics,

2005,

125,

53–75.

Banerjee, A.V., E. Duﬂo, R. Glennerster, and D. Kothari,

“Improving im-

munisation coverage in rural India: clustered randomised controlled evaluation of

immunisation campaigns with and without incentives 2010; 340 :c2220,”

BMJ,

2010,

340. c2220.

Blundell, R., M. Costa Dias, C. Meghir, and J.V. van Reenen,

“Evaluating the

Employment Impact of a Mandatory Job Search Program,”

Journal of the European

Economic Association,

2004,

569–606.

Bollens, J and B. Cockx,

“Eﬀectiveness of a job vacancy referral scheme,”

IZA

Journal of Labor Policy,

2017, pp. 6–15.

Bonoli, G., R. Lalive, D. Oesch, N. Turtschi, A. von Ow, P. Arni, and

P. Parrotta,

“The impact of social networks on re-employment,” 2014. IZA Research

Report 60.

Card, D., J. Kluve, and A. Weber,

“Active labour market policy evaluations: A

Meta-Analysis,”

Economic Journal,

2010,

120,

F452–F477.

Electronic copy available at: https://ssrn.com/abstract=3515935

B 108 - 2021-22 - Endeligt svar på spørgsmål 1: Spm. om beregninger af effekten af digitale samtaler fra første dag på baggrund af de opdaterede regneprincipper, til beskæftigelsesministeren

, and

, “What Works? A Meta Analysis of Recent Active Labor Market Pro-

gram Evaluations,”

Journal of the European Economic Association,

2017, p. jvx028.

Cottier, L., Y. Fl¨ ckiger, P. Kempeneers, and R. Lalive,

“Does Job Search

Assistance Really Raise Employmen?,” Discussion Paper 11766, IZA 2018.

Cr´pon, B., E. Duﬂo, M. Gurgand, R. Rathelot, and P. Zamora,

“Do Labor

Market Policies have Displacement Eﬀects? Evidence from a Clustered Randomized

Experiment *,”

The Quarterly Journal of Economics,

2013,

128

(2), 531–580.

, T. Le Barbanchon, H. Naegele, R. Rathelot, and P. Zamora,

“What Works

for Young Disadvantaged Job Seekers: Evidence from a Randomized Experiment,”

mimeo, CREST; 2015.

Dahlberg, M. and A. Forslund,

“Direct Displacement Eﬀects of Labour Market

Programmes,”

The Scandinavian Journal of Economics,

2005,

107

(3), 475–494.

den Berg, G.J. Van and B. Van der Klaauw,

“Counseling And Monitoring Of

Unemployed Workers: Theory And Evidence From A Controlled Social Experiment,”

International Economic Review,

2006,

47,

895–936.

, B. Hofmann, and A. Uhlendorﬀ,

“Evaluating Vacancy Referrals and the Roles

of Sanctions and Sickness Absence,”

Economic Journal,

2019, p. accepted.

Dolton, P. and D. O’Neill,

“Unemployment Duration and the Restart Eﬀect: Some

Experimental Evidence,”

The Economic Journal,

1996,

106

(435), 387–400.

and

, “The Long-Run Eﬀects of Unemployment Monitoring and Work-Search

Programs: Experimental Evidence from the United Kingdom,”

Journal of Labor

Economics,

2002,

(2), 381–403.

Electronic copy available at: https://ssrn.com/abstract=3515935

B 108 - 2021-22 - Endeligt svar på spørgsmål 1: Spm. om beregninger af effekten af digitale samtaler fra første dag på baggrund af de opdaterede regneprincipper, til beskæftigelsesministeren

Engstr¨m, P., P. Hesselius, and B. Holmlund,

“Vacancy referrals, job search

and the duration of unemployment: A randomized experiment,”

LABOUR,

2012,

26,

419–435.

Ferracci, Marc, Grgory Jolivet, and Gerard J. van den Berg,

“Evidence of

Treatment Spillovers Within Markets,”

The Review of Economics and Statistics,

2014,

(5), 812–823.

Fougere, D., J. Pradel, and M. Roger,

“Does the public employment service aﬀect

search eﬀort and outcomes?,”

European Economic Review,

2009,

53,

846–869.

Gautier, P., P. Muller, B. van der Klauuw, M. Rosholm, and M. Svarer,

“Es-

timating Equilibrium Eﬀects of Job Search Assistance,”

Journal of Labor Economics,

2018,

(4), 1073–1125.

Gorter, C. and G.R.J. Kalb,

“Estimating the Eﬀect of Counseling and Monitoring

the Unemployed Using a Job Search Model,”

The Journal of Human Resources,

1996,

(3), 590–610.

Graversen, B.K. and J.C. van Ours,

“Activating unemployed workers works; Ex-

perimental evidence from Denmark,”

Economics Letters,

2008,

100

(2), 308–310.

and

, “How to help unemployed ﬁnd jobs quickly: Experimental evidence from

a mandatory activation program,”

Journal of Public Economics,

2008,

(10-11),

2020–2035.

H¨gglund, P.,

“Are there pre-programme eﬀects of active placement eﬀorts? Evidence

from a social experiment,”

Economics Letters,

2011,

112

(1), 91–93.

Lalive, R., C. Landais, and J. Zweim¨ ller,

“Market Externalities of Large Unem-

ployment Insurance Extension Programs,”

American Economic Review,

December

2015,

105

(12), 3564–96.

Electronic copy available at: https://ssrn.com/abstract=3515935

B 108 - 2021-22 - Endeligt svar på spørgsmål 1: Spm. om beregninger af effekten af digitale samtaler fra første dag på baggrund af de opdaterede regneprincipper, til beskæftigelsesministeren

Maibom, J., M. Rosholm, and M. Svarer,

“Experimental Evidence on the Eﬀects

of Early Meetings and Activation,”

The Scandinavian Journal of Economics,

2017,

119

(3), 541–570.

Manning, A. and B. Petrongolo,

“How Local Are Labor Markets? Evidence from

a Spatial Job Search Model,”

American Economic Review,

October 2017,

107

(10),

2877–2907.

Manoli, D.S., M. Michaelides, and Patel A.,

“Job-Search Assistance: Experi-

mental Evidence Using Administrative Tax Data,” 2018. NBER Working Papers

24422.

McConnell, S., K. Fortson, D. Rotz, P. Schochet, P. Burkander, L. Rosen-

berg, A. Mastri, and R. D’Amico,

“Providing Public Workforce Services to

Job Seekers: 15-Month Impact Findings on the WIA Adult and Dislocated Worker

Programs,” Report May 2016, Mathematica Policy Research 2016.

Meyer, B. D.,

“Lessons from the US unemployment insurance experiments,”

Journal

of Economic Literature,

1995,

33,

91–131.

Miguel, E. and M. Kremer,

“Worms: Identifying impacts on education and health

in the presence of treatment externalities,”

Econometrica,

2004,

(1), 159–217.

Pallais, A.,

“Ineﬃcient Hiring in Entry-Level Labor Markets,”

American Economic

Review,

November 2014,

104

(11), 3565–99.

van den Berg, G. and B. van der Klauuw,

“Counseling and monitoring of unem-

ployed workers: theory and evidence from a controlled social experiment,”

Interna-

tional Economic Review,

2006,

47,

895–936.

Electronic copy available at: https://ssrn.com/abstract=3515935

B 108 - 2021-22 - Endeligt svar på spørgsmål 1: Spm. om beregninger af effekten af digitale samtaler fra første dag på baggrund af de opdaterede regneprincipper, til beskæftigelsesministeren

Appendix A: Additional Tables and Figures

Table A-1: Sample statistics by treatment status

Variables

Assigned to

program

(1)

33.33

0.542

0.642

0.052

0.868

0.224

0.491

0.285

0.678

0.013

0.036

0.273

30.66

67.42

69.57

63.82

0.431

0.789

0.806

0.706

0.024

0.048

0.027

0.106

14,075

Program

participant

(2)

34.34

0.538

0.692

0.048

0.884

0.206

0.494

0.300

0.700

0.013

0.036

0.251

31.30

69.16

73.08

67.95

0.419

0.787

0.830

0.752

0.025

0.047

0.027

0.122

8,358

Attended at

least one meeting

(3)

35.37

0.544

0.760

0.035

0.920

0.172

0.512

0.316

0.716

0.015

0.033

0.236

31.78

70.44

76.43

73.18

0.400

0.779

0.833

0.793

0.029

0.042

0.027

0.142

3,183

Age

Male

Unemployment beneﬁts

Disabled

Matchable

Education level

Less than high school

High school

College

Place of birth

Sweden

Nordic countries

West Europe

Outside west Europe

Unemployment days

Year t–1

Year t–2

Year t–3

Year t–4

Unemployment spells

Year t–1

Year t–2

Year t–3

Year t–4

No. spells, last 4 yrs

Labor market education

Preparatory education

Labor market training

Subsidized employment

Observations

Notes: Summary statistics, weighted by the intention to treat share.

Electronic copy available at: https://ssrn.com/abstract=3515935

B 108 - 2021-22 - Endeligt svar på spørgsmål 1: Spm. om beregninger af effekten af digitale samtaler fra første dag på baggrund af de opdaterede regneprincipper, til beskæftigelsesministeren

Table A-2: Impact outside target population

Non-experiment periods 2015

Exit

Unemp.

unemp.

days

1st

1st year

quarter

(1)

(2)

(3)

In a program area

Control mean

Observations

−0.009

(0.008)

0.390

552,816

0.699

∗

(0.419)

73.78

552,816

1.585

(2.188)

187.0

552,816

Non-target population

Exit

Unemp.

unemp.

days

1st

1st year

quarter

(4)

(5)

(6)

0.011

(0.007)

0.535

367,778

−0.727

(0.581)

52.34

367,778

−3.049

(1.973)

162.7

367,778

Notes: Regression of each outcome variable on an indicator for active PES oﬃce (“In a program area”) during 2015.

The regressions include year dummies, month dummies, PES oﬃce dummies and the control variables in Table 1.

Standard errors in parentheses are clustered at the PES oﬃce level.

∗∗∗

p<0.01,

∗∗

p<0.05,

∗

p<0.1.

Electronic copy available at: https://ssrn.com/abstract=3515935

B 108 - 2021-22 - Endeligt svar på spørgsmål 1: Spm. om beregninger af effekten af digitale samtaler fra første dag på baggrund af de opdaterede regneprincipper, til beskæftigelsesministeren

Table A-3: Heterogenous eﬀects of the JSA program on unemployment

college

(1)

College

Born

west EU

(3)

0.029

∗∗∗

(0.007)

−0.013

∗

(0.007)

0.427

−3.939

∗∗

(1.809)

3.518

∗

(1.845)

173.1

413,279

Born

outside

west EU

(4)

0.049

∗∗∗

(0.012)

−0.018

∗∗

(0.008)

0.286

−11.692

∗∗∗

(3.827)

4.745

∗

(2.596)

225.9

139,537

Short

unemp.

history

(5)

0.031

∗∗∗

(0.007)

−0.016

∗∗

(0.007)

0.384

−3.429

∗∗

(1.536)

4.035

∗∗

(1.874)

187.9

276,408

Long

unemp.

history

(6)

0.037

∗∗∗

(0.009)

−0.014

∗

(0.008)

0.395

−8.694

∗∗∗

(1.858)

4.260

∗

(2.174)

186.2

276,408

Male

Female

(2)

0.024

∗∗∗

(0.009)

−0.016

∗

(0.008)

0.391

−5.462

∗∗

(2.079)

3.844

∗

(2.166)

181.1

149,982

(7)

0.044

∗∗∗

(0.008)

−0.025

∗∗∗

(0.008)

0.383

−7.937

∗∗∗

(1.972)

6.431

∗∗∗

(1.899)

189.5

303,807

(8)

0.022

∗∗∗

(0.007)

-0.005

(0.008)

0.398

−3.614

∗∗

(1.659)

1.546

(2.442)

183.9

249,009

Electronic copy available at: https://ssrn.com/abstract=3515935

Panel A: Exit unemp. 1st quarter

Assigned to program

In a program area

Control mean

Assigned to program

In a program area

Control mean

Observations

0.038

∗∗∗

(0.007)

−0.014

∗∗

(0.007)

0.389

−6.292

∗∗∗

(1.956)

4.075

∗

(2.161)

189.3

402,834

Panel B: Unemp. days 1st year

Notes: Regression of each outcome variable on an indicator for active PES oﬃce (“In a program area”) and an indicator for active PES

oﬃce

intention to treat status is treated (“Assigned to program”). The regressions include year dummies, month dummies, PES oﬃce

dummies and the control variables in Table 1. Short/long unemployment history is deﬁned as having below/above the median number of

unemployment days during the last four years prior to registration. Standard errors in parentheses are clustered at the PES oﬃce level.

∗∗∗

p<0.01,

∗∗

p<0.05,

∗

p<0.1.

B 108 - 2021-22 - Endeligt svar på spørgsmål 1: Spm. om beregninger af effekten af digitale samtaler fra første dag på baggrund af de opdaterede regneprincipper, til beskæftigelsesministeren

Table A-4: Reduced form estimates used in the equilibrium search model (moment conditions)

Data moment and inverse weight

Main model

Face-to-face and distance

Description

Value in the model

Panel A: Estimated program eﬀects

Exit rate, treated

0.039 (0.010

)

Exit rate, non-treated

-0.015 (0.006

)

Log wage, treated

-0.002 (0.008

)

0.065 (0.010

)

-0.017 (0.007

)

-0.006 (0.007

)

Table 3

Table 4

Table 5

Table 9

0.001 (0.005

)

0.029 (0.036

)

-0.308 (0.39

)

-0.284 (0.22

)

0.0002 (0.005

)

0.033 (0.033

)

-0.273 (0.35

)

-0.272 (0.20

)

−

|τ

))

−

|τ

= 0))

−

|τ

))

−

|τ

= 0))

|τ

=τ

−w

|τ

=τ

−w

|τ

Log wage, non-treated

Log vacancy rate

Job applications, treated

Job applications, non-treated

∗

|τ

=τ

−v

∗

|τ

∗

|τ

−

|τ

−

|τ

Panel B: Aggregate statistics

0.37

(0.003

)

0.37

(0.003

)

−

|τ

))

−

)(1

−

|τ

))

0.077 (0.0001)

0.019 (0.001)

0.60 (0.001)

3.45 (0.052

)

3.45 (0.052

)

0.60 (0.001)

Exit rate (three months)

Unemployment rate

∗

|τ

Vacancy rate

∗

|τ

Replacement rate

∗

|τ

Electronic copy available at: https://ssrn.com/abstract=3515935

Fraction of job seekers in

treated oﬃces who exit for

employment with three

months.

Unemployment rate treated

oﬃces.

Vacancy rate treated oﬃces.

Calculated using information

on vacancies at the municipal

level.

Average replacement rate in

Sweden (ref).

Average number of job

applications in treated oﬃces.

Job applications

|τ

Note: Empirical moments used in the estimation of the equilibrium search model. Panel A reports moments from the estimated program eﬀects with variances estimated

in the empirical models. The eﬀects on the number of job applications are re-scaled into IV estimate based on the estimates in Table 9. Moments for the face-to-face and

distance meeting model are estimated in the same way as the main model. Panel B reports moments in the form of aggregate statistics for the treated areas. For exit rate

and job applications variances across unit are used. Unemployment rate, vacancy rate and replacement rate do not have estimated variances. Here, the variances are set to

capture that the unemployment rate is measured precisely, whereas the vacancy rate and replacement rate are less precisely estimated.

B 108 - 2021-22 - Endeligt svar på spørgsmål 1: Spm. om beregninger af effekten af digitale samtaler fra første dag på baggrund af de opdaterede regneprincipper, til beskæftigelsesministeren

Table A-5: Model ﬁt and model estimates for the equilibrium search models

Panel A: Model ﬁt

Moment

Estimated program eﬀects

Exit rate, treated

Exit rate, non-treated

Log wage, treated

Log wage, non-treated

Log vacancy rate

Job applications, treated

Job applications, non-treated

Aggregate statistics

Exit rate (three months)

Unemployment rate

Vacancy rate

Replacement rate

Job applications

Panel B: Model estimates

Parameter

Fixed parameters

Treatment share

Discount rate

Productivity

Estimated parameters

Application cost, non-treated

Application cost, treated

Job destruction rate

Share of vacancies w. pos. surplus

UI beneﬁts

Bargaining power

Vacancy costs

Value of non-market time

Return to application, treated

Estimates

0.31

0.008

0.039

-0.015

-0.002

0.0002

0.033

-0.273

-0.272

Deviation from the moments

0.000

-0.000

0.000

-0.001

0.018

0.006

0.080

0.37

0.077

0.019

0.60

3.45

-0.000

0.006

-0.001

0.000

0.016

0.022

0.024

0.903

0.576

0.584

1.102

0.049

1.426

(0.008)

(0.011)

(0.003)

(0.154)

(0.093)

(0.196)

(1.171)

(0.047)

(0.855)

Note: The table summarizes the model ﬁt and present the model estimates.

Electronic copy available at: https://ssrn.com/abstract=3515935