Full text of Working Papers (Federal Reserve Bank of Chicago) : The Selection Effects of Part-Time Work : Experimental Evidence from a Large-Scale Recruitment Drive, Working Paper 2022-51

The full text on this page is automatically extracted from the file linked above and may contain errors and inconsistencies.

The Selection
Effects of
Part-Time Work:
Experimental Evidence
from a Large-Scale
Recruitment Drive
Hyuncheol Bryant Kim, Hyunseob Kim, and
John Y. Zhu
October 25, 2022
WP 2022-51
https://doi.org/10.21033/wp-2022-51

*Working papers are not edited, and all opinions are the responsibility
of the author(s). The views expressed do not necessarily reflect
the views of the Federal Reserve Bank of Chicago or the Federal
Reserve System.

THE SELECTION EFFECTS OF PART-TIME WORK:
EXPERIMENTAL EVIDENCE FROM A LARGE-SCALE
RECRUITMENT DRIVE∗
Hyuncheol Bryant Kim

Hyunseob Kim

John Y. Zhu†

October 25, 2022

Abstract
We implement a field experiment to examine how part-time work attracts applicants
with different quality and productivity levels than full-time work. In a large-scale
recruitment drive for a data-entry position in Ethiopia, either a part-time or fulltime job opportunity was randomly offered across villages. We find that the parttime work attracts a less qualified pool of applicants with a stronger preference for
short work hours, who in turn exhibit lower productivity, all relative to the fulltime work. Our preferred estimates show that this selection effect on productivity
may explain up to half of the typical part-time wage penalty. A simple conceptual
framework demonstrates that a lack of high quality potential applicants with a
strong preference for short work hours could explain the experimental evidence. The
results have implications for the selection effects of alternative work arrangements
and for the gender pay gap. JEL Codes: J22, J24, O15, M51
Keywords: Part-time work; alternative work arrangements; self-selection; labor productivity; wage-hour relation
∗

We thank Dan Aaronson, Gadi Barlevy, Murillo Campello, Syngjoo Choi, Will Cong, Adrian Corum, Camilo GarciaJimeno, Lukas Hensel, Lisa Kahn, Chul Young Kim, Seonghoon Kim, Jungmin Lee, Kyoungwoo Lee, Pauline Leung, Mike
Lovenheim, Zhuan Pei, Kiki Pop-Eleches, Dyah Pritadrajati, Yasuyuki Sawada, Kelly Shue, Elena Simintzi, David Slusky,
Margarita Tsoutsoura, and seminar and conference participants at Asian Development Bank, Asia Impact Evaluation Conference, Chicago Fed, Cornell University, Hong Kong Applied Economics Conference, Indiana University, 19th IZA/SOLE
Transatlantic Meeting for Labor Economists, Korea Empirical Applied Microeconomics Conference, Labor and Finance
Group Conference (WashU), NEUDC (BU), Seoul National University, and Yonsei University for helpful comments, and
Jee-Hun Choi, Vitor Costa, Sophie Croome, Daniel Gallego, Tingting Gu, Seollee Park, Siho Park, Seongheon Daniel Yoon,
Janna Yu, and Pengfei Zhang for excellent research assistance. We also thank Dechassa Abebe, Banchayew Asres, Bewuketu
Assefa, Jiwon Baek, Tizita Bayisa, Hyolim Kang, Jieun Kim, Minah Kim, Jiyeong Lee, Betelhem Muleta, Yong Hyun Nam,
Jeong Hyun Oh, Tembi Williams, Tae-Jun Yoon, and Soo Sun You for their excellent fieldwork, and Rahel Getachew,
Chulsoo Kim, Hongryang Moon at Myungsung Christian Medical Center for their support. This project was supported
by Africa Future Foundation, the Ministry of the Interior, Republic of Korea, and the Smith Family Business Initiative at
Cornell University. This project received IRB approval from Cornell University (protocol ID1604006319). This study can
be found in the AEA RCT Registry (AEARCTR-0001829). All errors are our own.
†
Hyuncheol Bryant Kim: Department of Economics, Hong Kong University of Science and Technology; email:
hbkim@ust.hk. Hyunseob Kim (corresponding author): Economic Research Department, Federal Reserve Bank of Chicago,
230 South LaSalle St., Chicago, IL 60604; phone: (312) 322-6131; email: hyunseob.kim@chi.frb.org. John Y. Zhu: Department of Economics, University of Kansas; email: johnzhuyiran@ku.edu.

1

Introduction
A growing fraction of the workforce is being employed under alternative work arrange-

ments (Mas and Pallais (2017); Katz and Krueger (2019); and Abraham et al. (2021)). A
key feature of these work arrangements is that the short or flexible work hours allow workers to more smoothly allocate their time between labor-market activities and non-market
activities, such as household work. This suggests that workers who choose alternative
work arrangements may differ from those choosing standard ones in their preferences,
skills, and productivity. In turn, self-selection into different work arrangements may account for part of the variation in aggregate productivity across arrangements and types
of occupations.
Part-time work is one of the most common alternative work arrangements (Mas and
Pallais (2020); and Goldin (2021)), accounting for a substantial fraction of employment
across developed and developing countries (e.g., Pagés et al. (2008); Dunn (2018); and
OECD (2020)). Part-time work is associated with considerable wage penalties, implying
that it may attract less productive workers than full-time work.1 Part-time work is also
associated with strong preferences for short working hours, however (e.g., proxied by the
number of young children), especially for women.2 This correlation implies that selection
into part-time work may be primarily driven by household responsibilities rather than by
worker productivity. However, little is known on whether and how self-selection on these
dimensions leads to a productivity difference between part-time and full-time workers.
In this paper, we use a randomized field experiment to provide credible estimates of
how part-time work attracts applicants with different quality and productivity relative
to full-time work.3 In a non-governmental organization’s (NGO) large-scale recruitment
1

See, e.g., Blank (1990); Aaronson and French (2004); and Manning and Petrongolo (2008). This
line of research finds a part-time wage penalty of about one-fifth to one-quarter of full-time wages. A
related literature finds generally positive relations between work hours and wages (e.g., Rosen (1976);
and Simpson (1986)).
2
See, e.g., Rosen (1976); Moffitt (1984); and Ermisch and Wright (1993).
3
The human capital literature posits that individuals’ performance on tasks (such as productivity) is

1

drive for a data-entry position in Ethiopia, either a part-time or full-time work opportunity
was randomly offered by village group—a cluster of several nearby villages. The part-time
and full-time jobs differed only in the hours required, four and eight hours per work day,
respectively. This difference was made clear in job ads that were distributed during a
census of households in the recruitment areas. Our experiment focuses on women, who
typically value the flexibility between labor-market and non-market activities more than
men (e.g., Blank (1990); and Wiswall and Zafar (2018)). Thus, the experiment offers an
ideal setting to examine the selection effects of part-time work.
We first analyze the types of applicants part-time work attracts compared to full-time
work by collecting information on each applicant’s skills related to data-entry work as
well as preferences for work hours from a survey and job aptitude tests. We find that
there is a “part-time quality gap”: Part-time applicants exhibit lower quality than fulltime applicants. For example, the applicants in the part-time pool exhibit lower scores
in data-entry and manual dexterity tests than those in the full-time pool. In addition,
part-time applicants have stronger preferences for working short hours than full-time
applicants.
Next, we analyze how part-time work affects labor productivity through selection by
inviting job applicants to an internship in which we measure worker-level labor productivity using data-entry speed. All job applicants were invited to the internship, which allows
us to measure the productivity difference between the part-time and full-time workers
that is due to selection when the firm (i.e., NGO) is assumed to apply different hypothetical performance cutoffs for regular employment. We find that the productivity gap,
as well as the quality gap, becomes more pronounced as the firm hypothetically applies
a more stringent hiring cutoff. For example, assuming that the firm hires the interns
determined by multiple factors such as cognitive skills, noncognitive skills, and personality traits, as well
as effort (e.g., Heckman, Stixrud, and Urzua (2006)). Heckman and Kautz (2012), in their survey of the
literature, illustrate the difficulty in disentangling the effect of skills and traits from effort on performance.
Thus, we use the term “quality” throughout this paper to capture these multiple aspects of ability as
well as effort. The productivity measures we employ capture the effect of both ability and effort.

2

with above-the-median overall productivity, the productivity of those hired through the
part-time arrangement is on average 0.46-standard deviation (or 13 percent) lower than
the productivity of those hired through the otherwise identical full-time arrangement.
This “part-time productivity gap” exists from the first day and persists throughout the
internship, implying that self-selection on stable characteristics (such as quality), rather
than differences in skill acquisition through the internship, drives the gap. In contrast,
if the firm is assumed to hire all interns, the productivity difference between part-time
and full-time recruited workers is smaller at 0.12 standard deviations (or 3 percent) and
insignificant.
We demonstrate that a standard self-selection model with a relative lack of potential
applicants in the population who both strongly prefer short work hours and have high
quality could explain the observed differences in quality between the full-time and parttime applicant pools.4 In the framework, a worker is offered a job opportunity – either
part-time or full-time – to which she applies when the payoff from the job offered exceeds
her outside option. An increase in a worker’s preference for working short hours increases
(decreases) the payoff from the part-time (full-time) job, and her outside option increases
in her quality. We show that high-quality applicants in the part-time (full-time) pool
have strong (weak) preferences for short hours, whereas low-quality applicants in the
two pools have similar preferences. Therefore, the selection effect of part-time work,
relative to full-time work, on the quality of hired workers depends on the firm’s ability
to distinguish quality and hire higher quality employees from the applicant pool. If the
firm cannot distinguish quality and hires randomly from the applicant pool, the average
quality of part-time employees is still lower than that of full-time employees. However,
the gap widens when the firm can distinguish quality and hires only the highest quality
4

We verify the assumption on the distribution of potential applicants by quality and short working
hour preference in both our own census data of the recruitment areas and representative samples of
workers across dozens of countries. Specifically, we use years of education and the number of children
living in the household as proxies for quality and preference for short work hours, respectively.

3

applicants, such as those above a threshold.
This paper relates to the growing literature that studies individuals’ selection into jobs
based on job attributes, ability, and preferences.5 Dohmen and Falk (2011), Dal Bó, Finan,
and Rossi (2013), Guiteras and Jack (2018), and Deserranno (2019) show how financial
incentives (such as piece rates and higher salary) attract workers with different productivity and prosociality. Ashraf et al. (2020) show that salient career incentives (compared
to public service motivation) attract more productive public health workers without sacrificing pro-social preference. Kim, Kim, and Kim (2020) show that career incentives
could attract higher-performing workers than wage incentives through self-selection. Our
paper is the first to provide experimental evidence on how part-time employment attracts
applicants with markedly different quality than full-time employment.6
Moreover, our finding that part-time work attracts less productive workers implies
that the wage penalties associated with short work hours are in part due to underlying
differences in worker productivity.7 In fact, the part-time productivity differentials we
estimate are of comparable magnitudes to typical part-time wage differentials in existing
research. More broadly, given that alternative work arrangements would generally attract
workers with high valuations of flexible work and non-work hours (e.g., Mas and Pallais
(2017); and Mas and Pallais (2020)), our results imply that the average worker recruited
through alternative arrangements might also be less productive than one recruited through
standard work arrangements.
Lastly, this paper adds to the current debate on the causes of, and solutions to, the
gender pay gap observed in both developed and developing countries (e.g., Morton et
al. (2014)). Mas and Pallais (2020) argue that the hours dimension of flexibility that
5

See Roy (1951) and Borjas (1987) for classical contributions to this literature.
Bick, Blandin, and Rogerson (2022) employ a structural estimation approach to show that part-time
(full-time) workers are more likely to be low (high)-productivity.
7
Existing papers examining the effect of part-time work on productivity largely rely on observational
data, hence do not distinguish the selection from other effects, and find mixed evidence. See, e.g., KünnNelen, De Grip, and Fouarge (2013); Garnero, Kampelmann, and Rycx (2014); and Devicienti, Grinza,
and Vannoni (2017).
6

4

we examine is particularly promising in explaining gender gaps in pay and employment.
Goldin (2021) argues that the positive wage-hour relation, exemplified by so-called “greedy
jobs”—high-paying jobs that require long work hours—are a key reason for the persistent
gender pay gap: Women who juggle work and family choose, more often than men, to
work short hours (i.e., part-time) thereby suffering a wage penalty (e.g., Juhn and McCue
(2017); and Kleven, Landais, and Søgaard (2019)). However, our findings suggest that
the positive relation between (hourly) wages and hours reflects, at least in part, positive
relations between productivity and work hours. Therefore, future discussions of the gender
pay gap should account for the productivity difference across work hours we identify.

2

Context and Experimental Design

2.1

Context: Recruitment Drive for a Data-Entry Position

In 2016, Africa Future Foundation (AFF), a non-governmental organization, sought to
hire up to 100 women for a data-entry position from the Holeta and Ejere areas of Ethiopia.
Holeta is an urban town of approximately 28,000 people, located about 31 miles west of
the capital, Addis Ababa. Ejere is a mostly rural district near Holeta with a population
of approximately 59,000. The data-entry industry in Ethiopia largely employs women.
Access to early childhood care and education, which would mitigate women’s preference
for short working hours, is generally limited in the recruitment areas. For instance, at
the time of the experiment, there were only three certified private kindergartens across
Holeta and Ejere. As a result, grandparents, other relatives, or personally hired nannies
typically provided childcare to working mothers. These characteristics of the industry
and region make it an ideal setting to examine the selection effects of part-time work
because the flexibility that part-time work offers would be an important consideration for
the potential applicants who are women.

5

The data-entry position for which AFF recruited involves reading documents that
contains census information of households in the areas and entering it as data fields on a
computer. Therefore, the job requires basic computer skills, clerical ability to read paper
surveys and input the information on a computer, fine motor skills to control hands and
fingers, and perseverance to perform tedious work. We measure the quality of applicants
on these dimensions using aptitude tests (see Section 3.2). The key job eligibility criterion
is to have a high school diploma. The level of education is generally high in the recruitment
areas with 60 percent and 38 percent of women holding high school diplomas in Holeta
and Ejere, respectively, while the corresponding number across Ethiopia is only 4 percent
(CSACE (2016)).

2.2

Experimental Design

Table I summarizes the stages in our experiment, reporting the number of women who
participated in each stage by the posted job. In May to July 2016, AFF conducted a census
of Holeta and Ejere, gathering information on 20,595 households with approximately
87,000 individuals in the areas. During the census, enumerators distributed job ads to
6,295 resident women with a secondary school diploma (or their family members in case
they were absent at the time of census visits). These eligible women represent potential
applicants in our experiment.
The part-time and full-time jobs were randomly assigned across 71 village groups—clusters
of several villages—with 35 and 36 village groups posting part-time and full-time jobs,
respectively.8 We randomized treatment at the village group level by combining nearby
8

The
experimental
design
and
the
outcome
variables
considered
in
this
study
are
pre-specified
in
the
pre-analysis
plan
at
the
AEA
RCT
Registry:
https://www.socialscienceregistry.org/trials/1829/history/12246.
Although the original study design included 81 village groups, 10 village groups in Ejere were excluded from the final study sample
due to safety concerns from political turmoils during which more than 500 people are estimated to
have been killed. See, e.g., https://www.theguardian.com/world/2016/oct/02/ethiopia-many-deadanti-government-protest-religious-festival. The original design also included long-term employment and
further randomization at the data-entry unit but AFF evacuated from the study areas for the same

6

villages with frequent contacts to minimize potential information spillovers between the
treatment and control villages.
The job ads we distributed (see Appendix Figure A1) make it clear that the part-time
and full-time jobs differ only in the hours required per day, four and eight hours. Other
aspects of the position such as the application requirements, task, and wage (per hour) are
identical. The monthly pay offered ranged from 1,000 to 1,250 (2,000 to 2,500) Ethiopian
Birrs for part-time (full-time) employees depending on their performance, which was in
line with pay at other data-entry firms in Ethiopia.9 Because there was no wage discount
for the part-time job offered, the quality and productivity gaps between part- and fulltime applicants observed in our experiment are likely lower bound estimates (in absolute
value), relative to a setting with a part-time wage discount.10
Among the 6,295 potential applicants, 456 individuals submitted a résumé and a copy
of their high school graduation exam report at the AFF office located in the Holeta city
center by August 2016. Those who submitted application materials were subsequently
asked to join a baseline job survey and to take aptitude tests, which were administered
at the AFF office in December 2016. We refer to those who both submitted application
materials and participated in the job survey and aptitude tests as “applicants” (N = 333).
Finally, AFF invited applicants to an internship program to assess their productivity
as data-entry workers. The AFF staff contacted each applicant by phone; if the person
was not available immediately, the staff made multiple contact attempts. 122 out of of 333
applicants took up the internship (referred to as “interns”) across August to December
2017. Interns were grouped into five waves, and those with higher scores in a data-entry
reason.
9
According to the authors’ market survey in 2016, a typical data-entry firm in Ethiopia paid the
average full-time worker a baseline wage of 80 Ethiopian Birrs (ETBs) per day (or 1,600 ETBs per
month), plus two ETBs per additional accurate entry over 30 entries per day as an incentive. 100 ETBs
was approximately US$3 as of the timing of the experiment.
10
In addition to the worker selection effects we focus on, fixed costs of employment could explain part
of the wage penalty (e.g., Rosen (1976)). Compensating differentials (e.g., Rosen (1986)) suggest that to
the extent that schedule flexibility that part-time work provides is valuable to workers, the part-time job
could offer lower wages conditional on productivity.

7

test (conducted as part of the aptitude tests) were invited earlier. The internship for
each wave consisted of 22 to 32 interns and lasted for three weeks. The program entailed
typing and data-entry tests as well as basic computer training (see Appendix Figure A2
for details). The interns were allowed to attend either the morning (9 a.m.–noon) or
afternoon (2 p.m.–5 p.m.) session with an identical program, which ensured that they
could participate regardless of their working hour preferences. They were paid a daily
wage of 30 ETBs and were told that tenured workers would be hired based on productivity
in the internship.
It is worth noting that AFF attempted to invite all job applicants to the internship, as
opposed to those with high measured ability only, such as top performers in the aptitude
tests. This feature allows us to gauge the productivity difference between part-time and
full-time workers that is due to selection when the firm (i.e., AFF) is assumed to apply
different hypothetical performance cutoffs to hire workers.

3

Data
Our primary data sources are the job survey, aptitude tests, administrative data col-

lected during the job application and internship, and the census of the recruitment areas.
The census data provide variables capturing basic demographic and socioeconomic status
and family structure, including age, education, employment status, and the number of
household members, including children.

3.1

Study Population and Randomization Balance

Appendix Table A1 presents randomization balance tests on individual, household,
and village characteristics for our sample of potential applicants. The table confirms that
the randomization was successful: Only one out of 20 characteristics differs significantly
at the 10 percent level between the village groups with part- and full-time job postings.
8

In addition to showing the balance, it provides useful information to understand labor
markets in the study areas. First, the fraction of potential applicants with post-secondary
education is 39 percent. Second, 19.5 percent of applicants were working in formal sectors and 13.2 percent were working for their family business. This low formal-sector
employment rate in the areas, even for those with high school education, implies that the
data-entry position should offer them an attractive labor-market opportunity. Third, the
average potential applicant’s household has 4.2 members among whom are 2.5 children.
Lastly, about one-third of the villages in the areas are in Holeta, the urban area, with the
rest in Ejere, the more rural area.
Further, Appendix Table A2 compares characteristics of job applicants with nonapplicants among the potential applicants. We find that those who are younger, more
educated, not married, have fewer children, and do not currently have an (official) job
are more likely to apply for the position. All of these differences are significant at the
1 percent level. The results show that outside option and family status are important
determinants of application decisions in general.

3.2

Measuring Applicant Quality and Work Hour Preference

We employ two types of applicant quality measures. First, we conducted job aptitude
tests that measure each applicant’s skills related to data-entry work. The most direct
measure of job-specific ability is data-entry speed, defined as the number of correct data
entries made within 15 minutes. We also measure applicants’ clerical and computation
abilities based on the O*NET Ability Profiler (O*NET Resource Center (2010)) and
manual dexterity based on the Bruininks-Oseretsky Test of Motor Proficiency, 2nd edition (BOT-2, Deitz, Kartin, and Kopp (2007)). The clerical ability test mainly involves
noticing if there are mistakes in the text and numbers. The computation test measures an
individual’s ability to apply arithmetic operations to calculate solutions to mathematical

9

problems. To measure manual dexterity, we counted how many small coins (out of 20)
the applicant moves using fingers from a table to a small box in 15 seconds. Second, we
employ years of education and whether the applicant currently works for an official job
as measures of general quality valued by the labor market (e.g., Dal Bó, Finan, and Rossi
(2013)).
For preference regarding work hours, we conducted surveys that directly ask the applicant’s preferences between (i) non-work, working part-time and full-time and (ii) family
and work, as well as how supportive the spouse is for the applicant’s work. We also collected information on the number of children living in the same household as a proxy for
preference for working short hours driven by child-rearing responsibilities (see, e.g., Rosen
(1976); Moffitt (1984); and Ermisch and Wright (1993)). The Data Appendix provides
details of the aptitude tests and survey modules we employ.
In the empirical analysis, we standardize these measures by subtracting the respective
mean and scaling by the standard deviation – as z -scores (Kling, Liebman, and Katz
(2007)). In addition, we stack the z -scores within the quality (preference) dimension
and analyze them as an overall measure of quality (preference) in a single pooled OLS
regression with standard errors clustered at the village group level. To mitigate the
influence of outliers, we winsorize data-entry speed and manual dexterity at the 1 percent
tails.

3.3

Measuring Labor Productivity

We employ two measures of labor productivity for the interns. First, we measure erroradjusted typing speed as the number of words the intern correctly entered per minute
using Mavis Beacon, a computer application designed for typing training. Each typing
task involves the intern typing in a series of words or sentences shown on the computer
screen for seven to 15 minutes. The interns performed two typing tasks a day over the

10

three-week internship period.
Second, we measure error-adjusted data-entry speed as the number of correctly entered
census data fields, scaled by the number of minutes spent.11 We gave all interns the same
set of census forms with identical information on a given day and asked them to type
in the information using the computer within 15 minutes. The interns performed the
data-entry task in the last two weeks of the internship, once a day in the second week and
twice a day in the third week. In the empirical analysis, we employ standardized measures
of labor productivity as z -scores. To mitigate the influence of outliers, we winsorize the
productivity measures at the 1 percent tails.

4

Conceptual Framework: Part-Time and Full-Time
Job Application and the Quality of Applicant Pools
We illustrate how offering a part-time job opportunity versus a full-time job oppor-

tunity affects the quality of the applicant pool through selection by modeling workers’
selection into jobs with differing hours.

4.1

Set Up
We consider a population of potential applicants, parameterized by two variables:

preference for short working hours (γ) and quality (θ). γ takes values in r0, 1s and measures
the strength of a worker’s preference for part-time work over full-time work, quoted in
hourly terms. The higher γ is, the more the worker prefers part-time work (four hours
per day) over full-time work (eight hours per day). θ takes values in r0, 1s and measures
11

We define a “correctly entered field” as a non-missing value in a census data field (such as a person’s
name) that is entered without an error or a missing value that is not supposed to be entered. All other
entries are considered incorrect.

11

both the worker’s quality and her hourly outside option value – for simplicity, we assume
they are perfectly correlated. The higher θ is, the greater is the worker’s quality and
hourly outside option value. Thus, the entire population of potential applicants can
be represented as a measure µ over the unit square r0, 1s ˆ r0, 1s. For example, for
0 ď a ă b ď 1 and 0 ď c ă d ď 1, µpra, bs ˆ rc, dsq is the measure of workers with γ
between a and b, and θ between c and d.
We then offer this population of potential applicants a job opportunity j. We consider two cases, when j “ P T is a part-time job and when j “ F T is a full-time job.
A worker’s hourly payoff from a part-time job is assumed to be

W P T pγ, θq “ w ` γ “ γ.

Here, w denotes the hourly wage, which we normalize to zero. Notice, the worker’s hourly
payoff from working a part-time job is her hourly wage plus an hourly benefit she derives
from the part-time job. For a worker with children, this benefit can be utility derived
from being able to spend more time with their children, or the cost savings from not
having to procure childcare. Or, it can simply represent preference for leisure. Naturally,
the benefit is increasing in γ.
A worker’s hourly payoff from a full-time job is assumed to be

W F T pγ, θq “ w ` p1 ´ γq “ 1 ´ γ.

Similar to before, the worker’s hourly payoff from working a full-time job is her hourly
wage plus an hourly benefit she derives from the full-time job. This benefit can be a
sense of pride from working full-time or could represent the cost savings from not having
to find, train for, and travel to multiple part-time jobs in order to fill up the work day.
Naturally, the benefit is decreasing in γ. The per-hour wage (w) is identical between the

12

part-time and full-time jobs, consistent with our experimental design.

4.2

Workers’ Application Decision, Distribution of Workers, and
the Quality of Applicant Pools
Let j P tP T, F T u be the job opportunity that is being offered. Since θ is a worker’s

hourly outside option value, a pγ, θq worker will apply for the job if and only if her hourly
payoff from having job j is weakly greater than θ. Let S j denote the subset of worker
types that apply. Then we have,

pγ, θq P S j ô W j pγ, θq ě θ

for j P tP T, F T u.

Within the unit square of all possible worker types, S P T consists of the set of points located
on or below the diagonal running from p0, 0q to p1, 1q, while S F T consists of the set of
points located on or below the diagonal running from p0, 1q to p1, 0q. For example, a worker
with type p0.75, 0.5q will apply when the job opportunity being offered is a part-time job
because W P T p0.75, 0.5q “ 0.75 ą 0.5. However, this same worker will not apply when the
job opportunity being offered is a full-time job, because W F T p0.75, 0.5q “ 0.25 ă 0.5. In
contrast, a worker with type p0.5, 0.25q will apply in both cases, and a worker with type
p0.5, 0.75q will not apply in either case.
S P T is the part-time applicant pool that corresponds to those women who applied
in the villages with the part-time job posting. S F T is the full-time applicant pool that
corresponds to those women who applied in the villages with the full-time job posting.
Obviously, the statistical properties of S P T and S F T depend on the statistical properties
of the population – i.e., µ. We make the following assumption about µ:
µ has a density. There is a parameter x P p0.5, 1q and a value l ą 0, such that
the density of µ on the subset, rx, 1s ˆ rx, 1s, of worker types is 0, while the density of µ
13

outside that subset is l.
See Figure I for a depiction of the unit square of worker types, applicant pools S P T
and S F T , and density µ. The specific functional form we have chosen is not crucial for our
results, and is made largely for computational tractability and conceptual clarity. What
matters is that there is a relative lack of workers who have both high quality and strong
preference for short working hours. The motivation for this assumption is as follows: We
posit that the preference for part-time work is driven by non-market responsibilities, such
as child-rearing, that also make acquiring high quality and outside options difficult. Thus,
we expect that workers who both strongly prefer part-time work and have high quality
and outside options are relatively rare.12
The empirical distribution of potential applicants in our recruitment areas along
proxies for these dimensions provides evidence for this assumption. Specifically, we employ
years of education as a proxy for worker quality and the number of children living in the
same household as a proxy for preference for short working hours. Figure II shows that
the density of potential applicants who have both higher levels of education and more
children living with them (i.e., those in the north-east corner) is particularly low relative
to the rest of population. Appendix Figure A3 shows similar distributional patterns across
24 African countries using the Demographic and Health Surveys (DHS) data on women.
Bick, Blandin, and Rogerson (2022) provide further empirical support by documenting a
negative correlation between worker productivity and preference for fewer working hours.
Our main theoretical result is that what kind of job opportunity is being offered –
part-time or full-time – affects the type of applicants that the job attracts, in particular,
the statistical properties of the applicants’ quality.
12

The assumption is consistent with negative correlations between women’s education and fertility
shown in existing research. The mechanisms include education delaying or reducing fertility (Keats (2018);
and Lavy and Zablotsky (2015)) and early pregnancy preventing mothers from having further educational
opportunities (Becker, Cinnirella, and Woessmann (2010)). We are agnostic about the direction of
causality for the purpose of the model predictions. Beyond schooling, see Jones and Long (1979) for
evidence that women with more children tend to receive less on-the-job training from the employer.

14

Proposition 1. The average quality of the part-time applicant pool S P T is less than the
average quality of the full-time applicant pool S F T .
Proof. The result can easily be seen geometrically from Figure I. Fix a quality level
θ̃ P r0, xq. Notice, the (marginal) measure of part-time and full-time applicants with
quality “ θ̃ is the same: l ¨ p1 ´ θ̃q. This means, below the quality level x, the distribution
of quality within S P T and S F T is identical. Above the quality level x, there are no
applicants in S P T , while there is a strictly positive measure of applicants in S F T . Thus,
the average quality of S P T is less than the average quality of S F T .

4.3

Firm’s Hiring from Applicant Pools and the Quality of Hired
Workers
Of course, the applicant pool is not the same as the pool of workers that are eventu-

ally hired by the firm. First, let us consider the extreme case when the firm cannot screen
θ or γ.13 Then the average quality of workers the firm hires when a job opportunity j is
being offered is exactly the average quality of the j-applicant pool. We now immediately
have the following result:
Corollary 1. Suppose the firm cannot screen θ or γ. Then, the average quality of the
hired workers is lower when the job opportunity being offered is part-time.
Next, let us consider the opposite extreme and suppose the firm can perfectly screen
θ. We posit that there is a quality cutoff θ˚ ă x such that a worker is worth hiring if
and only if her quality is at or above θ˚ . Consequently, the set of hired workers when
the job opportunity being offered is part-time is S P T pθ˚ q, defined to be the subset of S P T
13

We do not presume that the firm’s utility function depends on γ, only θ. However, since γ and θ are
correlated in the applicant pools, even if the firm cannot screen θ, if it could screen γ, it would do so.
For example, in the full-time applicant pool, it would try to hire those with the weakest preference for
short working hours, since they are more likely to have higher quality.

15

consisting of those workers with quality ě θ˚ . The set of hired workers when the job
opportunity being offered is full-time, S F T pθ˚ q, is defined analogously.
Proposition 2. The average quality of the hired workers when the job opportunity being
offered is part-time, S P T pθ˚ q, is less than the average quality of the hired workers when
the job opportunity being offered is full-time, S F T pθ˚ q. Moreover, the magnitude of the
difference – call it the average quality gap – is increasing in θ˚ P r0, xq.
Proof. The proof of the first part is virtually identical to the proof of Proposition 1. To
prove the second part, assume θ˚ ă x. The average quality of the hired workers when the
job opportunity being offered is full-time is
1 ` 2θ˚
,
3
which means that the average quality of this group increases at rate

2
3

with respect to θ˚ .

A simple calculation shows that the average quality of the hired workers when the job
opportunity being offered is part-time increases at a variable rate ă

2
3

with respect to θ˚ .

This implies the gap is increasing in θ˚ .

5

The Effect of Part-Time Work on Applicant Characteristics through Selection
We examine the effect of part-time employment on the applicant pool, relative to

otherwise identical full-time employment, by estimating the following equation on an
applicant pool:
yij “ α0 ` α1 P artij ` εij ,

(1)

where yij is a characteristic of applicant i in village group j measured in the job survey,
aptitude test, or census; P artij is an indicator equal to one if applicant i was in village
16

group j with the part-time job posting, and zero with the full-time job posting; and εij
is a random error clustered at the level of randomization, village groups. The coefficient
of interest is α1 , which captures the causal effect of part-time employment opportunities
on the applicant pool through selection.
To analyze the selection effect on applicant pools with differing ex ante likelihoods of
being hired, we estimate equation (1) on the following three applicant pools: (i) all applicants, (ii) applicants who participated in the internship with average performance greater
than or equal to the median intern’s (referred to as “above-the-median performance”), and
(iii) interns with average performance below the median.14 That is, we employ internship
participation and above-the-median performance as approximate criteria for AFF to hire
from the overall applicant pool.15
Table II presents the results of estimating equation (1). Column 3 of Panel A shows
that among all applicants, the part-time applicant pool has significantly lower average
quality than the full-time pool as measured by data-entry test score (significant at the
5 percent level) and manual dexterity (significant at the 10 percent level). The gaps in
data-entry test score and dexterity between the part- and full-time pools amount to -0.22
and -0.24 standard deviations (SDs), respectively. The differences in the other quality
measures however are economically smaller and insignificant. The difference in overall
applicant quality is -0.07 SDs yet insignificant.16
Importantly, column 6 in Panel A shows that the differences in quality measures
are larger in economic magnitude and largely significant when conditioning on above-the14

We employ the average performance in the internship as a proxy for the applicant’s quality in splitting
into the subsamples, given that internship participation is an important criterion for AFF’s hiring decision
and the internship performance likely represents a more direct and precise measure of job-specific skills
than the results of aptitude tests or other baseline proxies such as education.
15
The median words per minute (WPM) of the interns is 12. Karat et al. (1999) find that a group of
IBM employees in the US who are experienced computer users and native speakers of English exhibit an
average WPM of 33. AFF found applicants with below-the-median performance largely unemployable.
16
Given that our analysis focuses on the overall quality of applicants rather than individual quality
measures, we employ the “stacking” approach rather than adjust the p-value for each quality measure to
account for multiple hypotheses testing.

17

median internship performance: The part-time pool shows significantly lower quality (at
a 10 percent or less level) than the full-time pool as measured by six out of seven variables.
As a result, the part-time pool is lower by 0.44 SDs in overall quality than the full-time
pool with the difference being significant at the 1 percent level. In contrast, none of the
quality measures is significantly different between the part- and full-time pools among
interns with below-the-median performance, who are unlikely to be hired (column 9).
Also consistent with the pronounced quality difference among more hireable applicants,
Panel A of Appendix Figure A4 shows that the distribution of overall quality for part-time
applicants left-shifts relative to that for full-time applicants among the above-the-median
performers, but not among the below-the-median performers.
The results above imply that the part-time quality gap will be more pronounced
among hired workers, under the assumption that the firm hires applicants who are above a
common performance threshold, such as the median. While this threshold was considered
reasonable by AFF, firms in general may want to apply different cutoffs depending on
their production functions, labor demand, and other factors. Therefore, we generalize the
analysis by varying the hypothetical cutoff to hire from the applicant pool. Specifically,
we estimate equation (1) with overall quality as the dependent variable on a series of
subsamples consisting of all interns (i.e., hiring 100 percent of them) up to the top 35
percent in 5 percent increments in terms of average performance in the internship.17
Figure III shows that the magnitude of quality gap increases monotonically as a
more stringent quality cutoff is applied: The gap begins at -0.17 SDs when all interns are
assumed to be hired and expands to -0.30 SDs when the top 75 percent interns are assumed
to be hired. As a greater performance cutoff is applied, the productivity gap continues to
increase, reaching -0.54 SDs for the top 35 percent threshold. All of these estimates are
significantly different from zero at a 5 percent or less level. This result suggests that once
17

We stop at the top 35 percent subsample since the sample size becomes too small beyond this point
to allow for precise estimation (e.g., less than 20 interns in each of the part- and full-time pools).

18

the firm hires from applicants based on performance in pre-employment programs (such
as internships and tests), the quality gap between the part- and full-time employees will
be more pronounced than the gap that exists for underlying applicant pools.
Overall, these experimental results are consistent with the theoretical result that
the average quality of part-time applicants is lower than that of full-time applicants, and
this difference is driven by high-quality applicants.
We now turn to Panel B of Table II, which presents measures of preference for short
working hours. Not surprisingly, the part-time applicants prefer non-work or working parttime to working full-time and prefer family over work more than the full-time applicants
(first three measures). These differences range from 0.08 to 0.27 SDs across the three
applicant pools (columns 3, 6, and 9), although they are not precisely estimated. In
addition, the part-time applicants receive weaker spousal support for working (significant
at the 1 and 5 percent levels among all applicants and interns with above-the-median
performance, respectively) and have a larger number of children who live with them. The
difference between the part- and full-time applicants in overall preference for short work
hours is 0.17 SDs and significant at the 1 percent level (column 3). In addition, the overall
preference differs by 0.29 SDs among interns with above-the-median performance (column
6, significant at the 1 percent level) and by 0.16 SDs among those with below-the-median
performance (column 9, insignificant). Overall, the results in Panel B are consistent with
applicants self-selecting to a part-time or full-time job according to their preferences for
working hours.
Unlike the previous two panels, Panels C and D show that the part-time and fulltime applicant pools are little different in terms of other demographic and socioeconomic
variables, as well as motivations for choosing jobs.18 This finding suggests that our conceptual framework that parsimoniously features worker quality and preference for working
18

One exception is that part-time applicants with above-the-median internship performance have lower
compensation-related motivations for choosing jobs than the corresponding full-time applicants (significant at the 5 percent level).

19

hours likely captures key economic forces behind the selection effects of part-time versus
full-time work.

6

The Effect of Part-Time Work on Labor Productivity through Selection
The previous section shows that part-time job applicants have significantly lower

quality than full-time applicants, measured by job-specific skills as well as education and
official sector employment status, particularly among those with high internship performance. The findings imply that part-time workers would exhibit lower labor productivity
at work through selection, other things held constant. We test this implication by comparing the labor productivity of interns from the part-time and full-time applicant pools.
Specifically, we estimate the following equation on a sample of interns:

P roductivityijystl “ β0 ` β1 P artij ` δy ` µs ` λt ` υl ` εijystl ,

(2)

where P roductivityijystl represents the following labor productivity measures (indexed by
y): (i) error-adjusted typing speed and (ii) error-adjusted data-entry speed, with their
respective means subtracted and scaled by standard deviations, for intern i from village
group j in internship wave s on working day t in trial l (up to two on a given day); P artij
is an indicator equal to one if intern i was from village group j with the part-time job
posting, and zero with the full-time job posting; and δy , µs , λt , and υl are productivity
measure, internship wave, working day, and trial fixed effects. εijystl is an error term
clustered at the village group level.
We argue that the productivity difference between interns recruited through parttime and full-time opportunities, captured by β1 , is driven by self-selection of applicants.

20

A key identifying assumption is that there is no effect of the different job opportunities
on (data-entry) skill acquisition during the internship.19 We assess the plausibility of the
identifying assumption by examining internship attendance and productivity dynamics.20
Table III presents the results of estimating equation (2). Column 1 of Panel A shows
that typing and data-entry speeds are lower by 0.12 SDs for interns recruited through the
part-time opportunity than those recruited through the full-time opportunity, although
the difference is not significant. Column 2 shows a greater productivity difference of -0.41
SDs among the interns with above-the-median performance, significantly different from
zero at the 1 percent level. In contrast, column 3 shows that the productivity difference
is only 0.04 SDs and insignificant for the interns with below-the-median performance.21
In fact, Panel B of Appendix Figure A4 reveals that the productivity distribution of parttime recruited interns left-shifts relative to that of full-time recruited interns among the
above-the-median performers but not among the below-the-median performers. These
results are consistent with theoretical and empirical results above on the quality gap
between the part-time and full-time applicant pools, which we now show translates into
a productivity gap for interns.
Next, Figure IV plots daily mean standardized labor productivity over the course
of the three-week internship, separately for the part-time and full-time recruited interns,
adjusted for fixed effects in equation (2).22 Panel A makes it clear that the productivity
gap for interns with above-the-median performance exists from the first day and persists
19

Another identifying assumption is that P artij is orthogonal to εijystl , which is obtained via random
assignment.
20
The causal effect of actually working part- or full-time on productivity (due, e.g., to fatigue from
working long hours) is not present in our setting, given that all interns worked for the same amount of
time a day regardless of the job opportunity offered.
21
We also examine the heterogeneity of effects by productivity level through a quantile regression
version of equation (2). Appendix Table A3 shows that the productivity gap is generally increasing in
the percentile and becomes economically and statistically significant above the top decile, consistent with
the finding in Table III.
22
Specifically, we estimate a variant of equation (2) that replaces the P art indicator with the interactions between the indicators for whether part- and full-time jobs were posted and a set of indicators for
working days (from 1 through 15).

21

throughout the internship. Panel B shows that the productivity gap hovers around -0.4
SDs, which is significant on most work days including the first. In contrast, Panels C
and D show that there is virtually no productivity difference throughout among interns
with below-the-median performance. For both above- and below-the-median performing
interns, the labor productivity of part-time recruited interns increases over time at least
on par with the productivity of full-time recruits.23
These productivity dynamics cast doubt on the alternative explanation that interns
who have applied to the part-time job have weaker incentives to invest in skills during
the internship given lower returns on their investment once they are hired. First, the
alternative story cannot explain the significant initial productivity difference, which by
construction is unaffected by skills acquired in the internship.24 Second, the productivity
improvement over time for part-time recruited interns is not slower than the improvement
for full-time recruits, which appears inconsistent with part-time recruits having weaker
incentives to acquire skills. Further, we test whether internship attendance, an important human capital investment for the job and related careers, differs between the partand full-time recruited interns. Appendix Table A4 shows that the attendance rate is
not significantly different among the above-the-median performing interns, for which a
significant productivity gap exists. Therefore, the overall evidence does not support differences in skill acquisition as a main explanation for the productivity differential, hence
supporting the key identifying assumption.25
Analogous to the analysis of worker quality that varies hypothetical hiring cutoffs
in the previous section, Figure V presents the corresponding result for productivity. It
23

Appendix Figure A5 plots productivity trends by productivity measure (i.e., typing or data-entry
speed) and finds similar results with those from Figure IV.
24
Estimates in Panel B of Table III imply that the initial productivity difference between the part-time
and full-time recruited interns with above-the-median performance is -0.498 SDs (= ´0.509 + 0.011 ˆ 1
day), which is significant at the 1 percent level.
25
Our experimental setting does not allows us to observe labor productivity in the long run. Nonetheless, given the generally steep productivity increase we observe over the course of the internship, any
differences in skill acquisition would have shown up as differential productivity trends.

22

shows that, as for the quality gap, the magnitude of the productivity gap generally increases as the firm is assumed to apply a more stringent hiring cutoff. The differences are
significantly different from zero at a 5 percent or less level when the top 75 percent interns
or above are assumed to be hired. This similarity in results for quality and productivity
points toward the productivity gap being driven by worker quality.
As a final analysis, we examine the extent to which the productivity differences are
due to selection on observable measures of quality or preference for short work hours.
We estimate a variant of equation (2) that further includes the variables that capture
applicants’ (i) quality, (ii) preference for short working hours, and (iii) both (from Table
II). Appendix Table A5 presents the estimation results. We find that variables capturing
applicant quality explain most of the productivity difference due to offering a part-time
or full-time job opportunity, particularly for interns with above-the-median performance
(columns 5–8). For these interns, the quality measures explain 80 percent (= [´0.411 ´
p´0.081q “ ´0.330]/´0.411) of the raw productivity gap, whereas the preference measures
explain 19 percent (= [´0.411 ´ p´0.332q “ ´0.079]/´0.411q.26 This finding is consistent
with individuals’ quality differentials, rather than differences in work hour preference,
being a key source of productivity gaps between part-time and full-time applicants.

7

Conclusion and Discussion
How part-time and other alternative work arrangements affect employee selection

and workforce productivity are important questions, given their rising prevalence across
labor markets. We explore these questions by implementing a randomized field experiment
that provides a part-time or full-time data-entry job opportunity to women in Ethiopia.
26

In addition, we follow Gelbach (2016) and formally decompose the effect of offering part-time, relative
to full-time, employment opportunities on productivity that is explained by covariates capturing quality
and preference. We find that the former explains -0.314 SDs (significant at the 1 percent level) whereas
the latter explains -0.030 SDs (insignificant) among interns with above-the-median performance.

23

We also develop a conceptual framework for job application given worker quality and
preference for work hours to explain the mechanism underlying the selection effects of
part-time work.
The experimental results show that part-time work attracts lower-quality applicants
with a stronger preference for short work hours relative to full-time work. This “parttime quality gap” is more pronounced among applicants with top performance in the
internship, who are ex ante more likely to be hired by the firm. Our conceptual framework
demonstrates that the lack of potential applicants who have both a strong preference
for short working hours and high quality is key to this selection effect. The part-time
applicants also exhibit lower productivity as measured by data-entry speed during the
internship, which is again more pronounced for higher-performing and thus more hireable
interns.
These findings have several important implications for part-time work and, more
generally, alternative work arrangements. First, the wage penalty associated with working
short hours is, in part, explained by the lower average productivity of workers who selfselect to work shorter hours.27 Second, suppose our finding that there is a lack of high
quality workers who also highly value short work hours generalizes to a lack of high
quality workers who highly value alternative work arrangements (e.g., for flexibility in
work hours). Then our results suggest that workers recruited under other alternative
arrangements may also be less productive on average than those hired through standard
arrangements. Third, if quality is more evenly distributed across workers with different
preferences for short or flexible work hours, the self-selection effects of offering alternative
work arrangements would be weaker than the results in this paper.28 Investigating how
27

To shed light on the quantitative importance of the part-time productivity gap due to selection in
explaining the associated wage gap, we estimate a version of equation (2) that uses the log of labor
productivity as the dependent variable. We find a part-time productivity gap of about 14 log points
or 13 percent for interns with above-the-median performance, which is of comparable magnitude with a
typical part-time wage penalty of 20 to 25 percent.
28
For example, occupations with relatively low variability in productivity across workers are candidates
for those with muted selection effects of alternative work arrangements.

24

the selection effects of alternative arrangements, including part-time work, differ across
types of occupations appears an important avenue for future research.29
This paper also offers implications for the relation between the gender pay gap and
wage-work hour relation. The literature argues that wages that increase in hours are a key
reason for the persistent gender pay gap (see, e.g., Goldin (2014); Goldin and Katz (2016);
and Goldin (2021)) – when faced with an increased burden from the household, especially
the arrival of a young child, it is more often women than men who choose to work short
hours, thereby suffering a wage penalty (Juhn and McCue (2017); and Kleven, Landais,
and Søgaard (2019)). Our findings imply that the positive relation between (hourly) wages
and hours is due, in part, to firms rationally using longer hours as a selection mechanism
for more productive workers. Therefore, future discussions of the gender pay gap should
account for the productivity difference across work hours we identify.
Lastly, our findings speak to the efficacy of policies that reduce work hours at an
economy level, as have been implemented in both developed and developing economies
(see, e.g., Hunt (1999); Chemin and Wasmer (2009); and Park and Park (2019)). If only
a subset of firms in the economy were to introduce shorter working hours, there may be
a negative consequence for the adopting firms’ productivity due to the selection effect we
identify (e.g., productive workers may go to firms offering long-hour jobs). This negative
selection effect would dampen their incentives to offer jobs with shorter hours ex ante,
even when doing so would improve the welfare of both firms and workers (see Rebitzer
and Taylor (1995) for a similar result based on an efficient-wage model). Thus, policies
that reduce work hours across firms in the economy, as opposed to a subset of firms, may
be preferred to mitigate the potentially negative effect on productivity.30
29

Another topic for future research is whether varying the quantity (e.g., short) of hours worked, which
this paper focuses on, and varying the pattern of hours (e.g., flexible) have a similar (or different) effect
on the quality of workers. See Chen et al. (2019) for the valuation of flexible work using Uber drivers.
30
As of this writing, several governments are implementing or considering work hour reductions. Colombia will reduce weekly work hours from 48 to 42 essentially for all workers by 2026. California is considering
reducing work hours from 40 to 32 per week only for private-sector firms with more than 500 employees.
Meanwhile, Spain is piloting a 32-hour work week for volunteering firms. Our results suggest that the

25

References
Aaronson, Daniel and Eric French (2004). “The Effect of Part-Time Work on Wages:
Evidence from the Social Security Rules.” Journal of Labor Economics 22.2, 329–352.
Abraham, Katharine, John Haltiwanger, Kristin Sandusky, and James Spletzer (2021).
“Measuring the Gig Economy: Current Knowledge and Open Issues.” in: Measuring
and accounting for innovation in the 21st century, ed. Carol Corrado, Jonathan Haskel,
Javier Miranda, and Daniel Sichel. Chicago, IL: University of Chicago Press, 2021.
Ashraf, Nava, Oriana Bandiera, Edward Davenport, and Scott S. Lee (2020). “Losing
Prosociality in the Quest for Talent? Sorting, Selection, and Productivity in the Delivery of Public Services.” American Economic Review 110.5, 1355–1394.
Becker, Sascha, Francesco Cinnirella, and Ludger Woessmann (2010). “The Trade-off between Fertility and Education: Evidence from Before the Demographic Transition.”
Journal of Economic Growth 15.3, 177–204.
Bick, Alexander, Adam Blandin, and Richard Rogerson (2022). “Hours and Wages.” Quarterly Journal of Economics 137.3, 1901–1962.
Blank, Rebecca (1990). “Are Part-Time Jobs Bad Jobs?” in: A future of lousy jobs?
The changing structure of U.S. wages, ed. Gary Burtless. Washington, DC: Brookings
Institution, 1990.
Borjas, George J. (1987). “Self-Selection and the Earnings of Immigrants.” American
Economic Review 77.4, 531–553.
Center, O*NET Resource (2010). “O*NET Ability ProfilerT M Score Report.” https :
//www.onetcenter.org/AP.html.
Chemin, Matthieu and Etienne Wasmer (2009). “Using Alsace-Moselle Local Laws to
Build a Difference-in-Differences Estimation Strategy of the Employment Effects of
the 35-Hour Workweek Regulation in France.” Journal of Labor Economics 27.4, 487–
524.
Chen, M. Keith, Peter E. Rossi, Judith A. Chevalier, and Emily Oehlsen (2019). “The
Value of Flexible Work: Evidence from Uber Drivers.” Journal of Political Economy
127.6, 2735–2794.
CSACE, I (2016). “Ethiopia Demographic and Health Survey.” Addis Ababa, Ethiopia
and Rockville, Maryland, USA: CSA and ICF.
Dal Bó, Ernesto, Frederico Finan, and Martın A. Rossi (2013). “Strengthening State
Capabilities: The Role of Financial Incentives in the Call to Public Service.” Quarterly
Journal of Economics 128.3, 1169–1218.
negative selection effect on productivity of participating firms is more likely for the cases of California
and Spain than Colombia.

26

Deitz, Jean Crosetto, Deborah Kartin, and Kay Kopp (2007). “Review of the BruininksOseretsky Test of Motor Proficiency, (BOT-2).” Physical & occupational therapy in
pediatrics 27.4, 87–102.
Deserranno, Erika (2019). “Financial Incentives as Signals: Experimental Evidence from
the Recruitment of Village Promoters in Uganda.” American Economic Journal: Applied Economics 11.1, 277–317.
Devicienti, Francesco, Elena Grinza, and Davide Vannoni (2017). “The Impact of PartTime Work on Firm Total Factor Productivity: Evidence from Italy.” Industrial and
Corporate Change 27.2, 321–347.
Dohmen, Thomas and Armin Falk (2011). “Performance Pay and Multidimensional Sorting: Productivity, Preferences, and Gender.” American Economic Review 101.2, 556–
590.
Dunn, Megan (2018). “Who Chooses Part-Time Work and Why?” Monthly Labor Review.
https://doi.org/10.21916/mlr.2018.8.
Ermisch, John F. and Robert E. Wright (1993). “Wage Offers and Full-Time and PartTime Employment by British Women.” Journal of Human Resources 28.1, 111–113.
Garnero, Andrea, Stephan Kampelmann, and François Rycx (2014). “Part-Time Work,
Wages, and Productivity: Evidence from Belgian Matched Panel Data.” ILR Review
67.3, 926–954.
Gelbach, Jonah B. (2016). “When Do Covariates Matter? And Which Ones, and How
Much?” Journal of Labor Economics 34.2, 509–543.
Goldin, Claudia (2014). “A Grand Gender Convergence: Its Last Chapter.” American
Economic Review 104.4, 1091–1119.
— (2021). Career and Family: Women’s Century-Long Journey toward Equity. Princeton
University Press, 2021.
Goldin, Claudia and Lawrence F. Katz (2016). “A Most Egalitarian Profession: Pharmacy
and the Evolution of a Family-Friendly Occupation.” Journal of Labor Economics 34.3,
705–746.
Guiteras, Raymond P. and B. Kelsey Jack (2018). “Productivity in Piece-Rate Labor
Markets: Evidence from Rural Malawi.” Journal of Development Economics 131, 42–
61.
Heckman, James J. and Tim Kautz (2012). “Hard Evidence on Soft Skills.” Labour Economics 19.4, 451–464.
Heckman, James J., Jora Stixrud, and Sergio Urzua (2006). “The Effects of Cognitive and
Noncognitive Abilities on Labor Market Outcomes and Social Behavior.” Journal of
Labor Economics 24.3, 411–482.
Hunt, Jennifer (1999). “Has Work-Sharing Worked in Germany?” Quarterly Journal of
Economics 114.1, 117–148.

27

Jones, Ethel B. and James E. Long (1979). “Part-Week Work and Human Capital Investment by Married Women.” Journal of Human Resources 14.4, 563–578.
Juhn, Chinhui and Kristin McCue (2017). “Specialization Then and Now: Marriage, Children, and the Gender Earnings Gap across Cohorts.” Journal of Economic Perspectives
31.1, 183–204.
Karat, Clare-Marie, Christine Halverson, Daniel Horn, and John Karat (1999). “Patterns of Entry and Correction in Large Vocabulary Continuous Speech Recognition
Systems.” Proceedings of the SIGCHI conference on Human Factors in Computing
Systems. 1999, 568–575.
Katz, Lawrence F. and Alan B. Krueger (2019). “The Rise and Nature of Alternative
Work Arrangements in the United States, 1995–2015.” ILR Review 72.2, 382–416.
Keats, Anthony (2018). “Women’s Schooling, Fertility, and Child Health Outcomes: Evidence from Uganda’s Free Primary Education Program.” Journal of Development
Economics 135, 142–159.
Kim, Hyuncheol Bryant, Seonghoon Kim, and Thomas T. Kim (2020). “The Role of
Career and Wage Incentives in Labor Productivity: Evidence from a Two-Stage Field
Experiment in Malawi.” Review of Economics and Statistics 102.5, 839–851.
Kleven, Henrik, Camille Landais, and Jakob Egholt Søgaard (2019). “Children and Gender Inequality: Evidence from Denmark.” American Economic Journal: Applied Economics 11.4, 181–209.
Kling, Jeffrey R., Jeffrey B. Liebman, and Lawrence F. Katz (2007). “Experimental Analysis of Neighborhood Effects.” Econometrica 75.1, 83–119.
Künn-Nelen, Annemarie, Andries De Grip, and Didier Fouarge (2013). “Is Part-Time
Employment Beneficial for Firm Productivity?” ILR Review 66.5, 1172–1191.
Lavy, Victor and Alexander Zablotsky (2015). “Women’s Schooling and Fertility under
Low Female Labor Force Participation: Evidence from Mobility Restrictions in Israel.”
Journal of Public Economics 124, 105–121.
Manning, Alan and Barbara Petrongolo (2008). “The Part-Time Pay Penalty for Women
in Britain.” The Economic Journal 118.526, F28–F51.
Mas, Alexandre and Amanda Pallais (2017). “Valuing Alternative Work Arrangements.”
American Economic Review 107.12, 3722–3759.
— (2020). “Alternative Work Arrangements.” Annual Review of Economics 12, 631–658.
Moffitt, Robert (1984). “The Estimation of a Joint Wage-Hours Labor Supply Model.”
Journal of Labor Economics 2.4, 550–566.
Morton, Matthew, Jeni Klugman, Lucia Hanmer, and Dorothe Singer (2014). “Gender at
Work: A Companion to the World Development Report on Jobs.” Washington, DC:
World Bank.

28

OECD (2020). “Part-Time Employment Rate.” Organisation for Economic Co-operation
and Development. https://data.oecd.org/emp/part-time-employment-rate.htm.
Pagés, Carmen, Carlos Scartascini, Rita Funaro, Lucıa Madrigal, Suzanne Duryea, Juan
Camilo Chaparro, Marıa Victoria Rodrıguez-Pombo, Eduardo Lora, Pablo Sanguinetti,
William D Savedoff, NavarroJuan Carlos, Aimee Verdisco, Mariana Alfonso, and Andrew Powell (2008). Beyond Facts: Understanding Quality of Life. Inter-American
Development Bank, 2008.
Park, WooRam and Yoonsoo Park (2019). “When Less is More: The Impact of the Regulation on Standard Workweek on Labor Productivity in South Korea.” Journal of
Policy Analysis and Management 38.3, 681–705.
Rebitzer, James B. and Lowell J. Taylor (1995). “Do Labor Markets Provide Enough
Short-Hour Jobs? An Analysis of Work Hours and Work Incentives.” Economic Inquiry
33.2, 257–273.
Rosen, Harvey S. (1976). “Taxes in a Labor Supply Model with Joint Wage-Hours Determination.” Econometrica 44.3, 485–507.
Rosen, Sherwin (1986). “The Theory of Equalizing Differences.” in: Handbook of labor
economics, ed. Orley C. Ashenfelter and Richard Layard. N.H.: Elsevier, 1986.
Roy, A. D. (1951). “Some Thoughts on the Distribution of Earnings.” Oxford Economic
Papers 3.2, 135–146.
Simpson, Wayne (1986). “Analysis of Part-Time Pay in Canada.” Canadian Journal of
Economics 19.4, 798–807.
Wiswall, Matthew and Basit Zafar (2018). “Preference for the Workplace, Human Capital,
and Gender.” Quarterly Journal of Economics 133.1, 457–507.

29

Figure I:
Theoretical Applicant Pools for Part-Time and Full-Time Jobs in the Preference for Short
Working Hours-Quality Space

Notes: This figure presents the theoretical applicant pools for part-time (red triangle with the apex at the
right corner) and full-time (blue triangle with the apex at the left corner) jobs from potential applicants with
preferences for short working hours (x-axis) and quality (y-axis) in a unit square. It is assumed that the [x,1]
ˆ [x,1] square in the upper-right corner has the density of zero, whereas the rest of the population has the
density of l ą 0.

30

Figure II:
Distribution of Proxies for Preference for Short Working Hours and Quality, Holeta and
Ejere, Ethiopia

Notes: This figure presents the distribution of the number of children living in the household (x-axis, proxy for
preference for short working hours) and years of education (y-axis, proxy for quality) in rank for the population
of potential applicants in Holeta and Ejere, Ethiopia, the recruitment areas. The data are collected in the
census of the areas. Darker colors represent denser parts of the population.

31

Figure III:
Quality Difference between Part-Time and Full-Time Recruited Interns Conditional on
Hypothetical Hiring Cutoffs

Notes: This figure presents the average difference in overall standardized quality between part-time and
full-time recruited interns (red dots) and the 95 percent confidence intervals (blue bars), conditional on
hypothetical hiring cutoffs from 100 percent (i.e., hiring all interns) to the top 35 percent in the distribution
of average internship performance. Confidence intervals are calculated based on standard errors clustered at
the village group level.

32

Figure IV:
Labor Productivity of Part-Time and Full-Time Recruited Interns
Panel A.
Above-the-Median Interns

Panel B.
Above-the-Median Interns – Differences Between Part- and
Full-Time Recruited Interns, with Confidence Intervals

33

Panel C.
Below-the-Median Interns

Panel D.
Below-the-Median Interns – Differences Between Part- and
Full-Time Recruited Interns, with Confidence Intervals

Notes: This figure presents coefficient estimates from a variant of equation (2) that replaces the Part
indicator with the indicators for part-time and full-time recruited interns, interacted with indicators for
working days (from 1 through 15). Panels A and C (B and D) show standardized labor productivity trends
separately for part- and full-time recruited interns (average differences between the part- and full-time
recruited interns and the 95 percent confidence intervals) over working days. Confidence intervals are
calculated based on standard errors clustered at the village group level. Panels A and B (C and D) use
interns with above- (below-) the-median performance.
34

Figure V:
Labor Productivity Difference between Part-Time and Full-Time Recruited Interns
Conditional on Hypothetical Hiring Cutoffs

Notes: This figure presents the average difference in standardized labor productivity between part-time and
full-time recruited interns (red dots) and the 95 percent confidence intervals (blue bars), conditional on
hypothetical hiring cutoffs from 100 percent (i.e., hiring all interns) to top 35 percent in the distribution of
average internship performance. Confidence intervals are calculated based on standard errors clustered at the
village group level.

35

36

Experimental stage
(1)
Census taken (job ads distributed)
Submitted job application materials
Participated in job survey and aptitude tests
Participated in internship (in 2nd or 3rd week)

Participants
(2)
job-eligible women
preliminary job applicants
job applicants
interns

Number and percentage of job-eligible women
Part-time
Full-time
Total
(3)
(4)
(5)
(6)
(7)
3,202 100.0% 3,093 100.0%
6,295
230
7.2%
226
7.3%
456
162
5.1%
171
5.5%
333
61
1.9%
61
2.0%
122

(4) - (6)
p-value
(8)
0.92
0.68
0.90

Notes: This table presents the stages in the experiment and the number of participants by stage. Columns 4 and 6 show the fraction of
job-eligible women continuing over experiment stages in village groups where the part-time and full-time jobs were posted, respectively.
Column 8 shows p-values for differences in the fraction continuing between the village groups with part-time and full-time job postings.

May to July 2016
July to August 2016
December 2016
August to December 2017

Approximate timing

Table I: Experimental Stages

37

N
(1)

-0.146
-0.044
0.077
0.088

22.721
0.313
4.429
6.875
61
61
61
61
61
61

53
60
61
61

61
60
61
49
57
288

61
61
60
61
61
58
59
421

3.166
2.951
3.590
3.370
3.012
3.267

21.167
0.231
4.296
7.360

-0.159
-0.146
-0.182
-0.614
-1.423
-0.494

0.744
0.534
0.237
0.508
0.460
1.649
-0.006
0.581

0.110
-0.031
-0.024
-0.080
-0.316
-0.203**

1.075
0.063
0.233
0.568

0.122
0.121
0.273
0.721**
0.271
0.293***

-0.598**
-0.502**
-0.111
-0.632***
-0.567**
-0.331*
-0.409***
-0.445***

Above-the-median interns
FT
Mean diff.
N
applicants
(PT-FT)
(4)
(5)
(6)

61
61
61
61
61
61

51
60
61
60

59
60
61
52
57
289

61
61
60
60
60
54
59
415

3.252
2.987
3.516
3.340
2.804
3.176

22.036
0.303
4.559
5.909

0.072
-0.042
-0.066
-0.490
-1.276
-0.353

0.149
-0.131
-0.109
-0.174
0.000
1.061
-0.176
0.072

0.013
0.036
-0.024
0.087
0.122
0.098

0.747
-0.155
-0.225
0.572

0.154
0.118
0.081
0.332
0.065
0.160

-0.206
-0.109
0.068
0.054
0.073
0.366*
0.058
0.047

Below-the-median interns
FT
Mean diff.
N
applicants
(PT-FT)
(7)
(8)
(9)

Notes: Columns 2 and 3 (columns 5 and 6) [columns 8 and 9] show means for full-time applicants and mean differences between the part-time and full-time
applicant pools for all applicants (above-the-median performing interns) [below-the-median performing interns]. All variables in Panels A and B are standardized
by subtracting the respective mean and scaling by the standard deviation. Data-entry test score is the number of fields that the applicant enters correctly from
census forms in 15 minutes. Computer literacy is based on 12 questions about basic knowledge of computer hardware and software, such as Microsoft Windows
and Office. Years of education is the total number of years the applicant has been in school. Working in official sector = 1 if the applicant is employed in an
official sector. Number of children in household is the number of children who live in the same household. Age is in years. Married = 1 if the applicant is married.
Subjective health status ranges from 1 (“very bad”) to 5 (“very good”). Asset score is the number of the following items that a household owns: electricity, a
watch/clock, a television, a mobile phone, a landline phone, a refrigerator, a bed with a mattress, an electric mitad (grill), and a kerosene lamp. See the Data
Appendix for definitions of other variables. ***, **, and * denote the significance level at 1%, 5%, and 10%, respectively, based on standard errors clustered at the
village group level.

0.023
-0.006
0.007
0.033
-0.087
0.015

0.157
0.123
0.099
0.350***
0.129
0.167***

-0.077
-0.060
-0.048
-0.498
-1.259
-0.371

3.225
2.952
3.528
3.282
2.865
3.196

-0.221**
-0.030
0.016
-0.069
-0.223*
0.097
-0.048
-0.069

Mean diff.
(PT-FT)
(3)

0.103
0.015
-0.008
0.034
0.114
1.322
-0.170
0.193

All applicants
FT
applicants
(2)

Panel A. Measures of Quality (Standardized)
Data-entry test score
333
Clerical ability
333
Computation ability
330
Computer literacy
329
Manual dexterity
332
Years of education
311
Working in official sector
322
Overall quality (pooled)
2,290
Panel B. Measures of Preference for Short Work Hours (Standardized)
Preference for family to work
331
Preference for non-work, working part- to full-time
325
Preference for part-time to full-time work
330
(Reverse) Supportive spouse for work
280
Number of children in household
304
Overall preference for short working hours (pooled)
1,570
Panel C. Individual Characteristics
Age
286
Married
323
Subjective health status [1-5]
330
Asset score [1-10]
330
Panel D. Importance in Choosing Jobs [1-4]
Intrinsic motivation
333
Extrinsic motivation
333
Accomplishment
333
Status
333
Career progress
332
Compensation and benefits
332

Variable

Sample

Table II: Effects of Part-Time Work on the Applicant Pool through Selection

Table III: Effects of Part-Time Work on Labor Productivity through Selection
All interns
(1)

Above the median
(2)
Dep. Var.:
Productivity
Panel A: Without time trend
Part
-0.123
-0.411***
(0.089)
(0.100)
Constant
0.072
0.682***
(0.066)
(0.093)
Productivity measure fixed effects
Y
Y
Wave fixed effects
Y
Y
Work day fixed effects
Y
Y
Trial fixed effects
Y
Y
R2
0.500
0.513
N
4,821
2,511
Panel B: With time trend
Part
-0.341***
-0.509**
(0.112)
(0.209)
Day
0.141***
0.172***
(0.004)
(0.011)
Part ˆ Day
0.024**
0.011
(0.009)
(0.016)
Constant
-1.238***
-0.916***
(0.069)
(0.177)
Productivity measure fixed effects
Y
Y
Wave fixed effects
Y
Y
Trial fixed effects
Y
Y
R2
0.494
0.503
N
4,821
2,511

Below the median
(3)

0.036
(0.054)
-0.477***
(0.031)
Y
Y
Y
Y
0.543
2,310
-0.126
(0.089)
0.119***
(0.006)
0.017*
(0.009)
-1.579***
(0.042)
Y
Y
Y
0.529
2,310

Notes: Panel A of this table presents estimates from equation (2) on samples of all interns (column 1), above-the-median
performing interns (column 2), and below-the-median performing interns (column 3). Panel B presents estimates from a variant
of equation (2) that includes the P art indicator, the variable Day that represents the number of working days, and their
interaction term. The dependent variable is standardized error-adjusted typing speed or data-entry speed. P art = 1 (0) if the
intern is recruited in a village where the part-time (full-time) job was posted. Standard errors clustered at the village group
level are reported in parentheses. ***, **, and * denote the significance level at 1%, 5%, and 10%, respectively.

38

1

Panel A. Part-Time Job

Figure A1. Job Ads
Panel B. Full-Time Job

Appendix Figures and Tables (For Online Publication)

Figure A2. Internship Program and Schedule

Notes: This figure presents the detailed internship program and schedule for participants in the morning
session (from 9 a.m.–12 p.m.). Participants in the afternoon session (from 2 p.m.–5 p.m.) had an identical
program.

2

Figure A3. Distribution of Proxies for Preference for Short Working Hours and Quality,
24 African Countries
Panel A. All Residents

3

Panel B. Urban Residents

Notes: This figure presents the distribution of the number of children living in the household (x-axis, proxy
for preference for short working hours) and years of education (y-axis, proxy for quality) in rank for women
aged 20 and over for 24 African countries. Panels A and B are for all residents and urban residents,
respectively. The data are from the Demographic and Health Surveys (DHS).
4

Figure A4. Cumulative Distributions of Quality and Labor Productivity for Part-Time
and Full-Time Recruited Interns
Panel A. Quality

Panel B. Labor Productivity

Notes: This figure presents the cumulative distribution functions of overall standardized quality (Panel A)
and standardized labor productivity (Panel B) for interns performing above-the-median (left) and
below-the-median (right), separately for those recruited through part-time and full-time job postings.

5

Figure A5. Labor Productivity of Part-Time and Full-Time Recruited Interns by
Productivity Measure
Panel A. Above-the-Median Interns – Typing Speed

Panel B. Above-the-Median Interns – Data-Entry Speed

6

Panel C. Below-the-Median Interns – Typing Speed

Panel D. Below-the-Median Interns – Data-Entry Speed

Notes: This figure presents coefficient estimates from a variant of equation (2) that replaces the Part
indicator with the indicators for part-time and full-time recruited interns, interacted with indicators for
working days (from 1 through 15) by productivity measure. Panels A and C (B and D) use standardized
error-adjusted typing speed (data-entry speed) as a productivity measure. Panels A and B (C and D) use
interns with above- (below-) the-median performance.

7

Table A1. Characteristics of Study Population and Balance of Randomization
Variable

(1)
N
Panel A. Characteristics of Potential Applicants
Age
6,160
Married
6,167
Ethnicity
Amhara
6,234
Oromo
6,234
Language
Amharic
6,236
Oromigna
6,236
Religion
Orthodox
6,225
Protestant
6,225
Muslim
6,225
Post-secondary education
6,265
Working
within household
6,115
in official sector
6,078
Panel B. Household Characteristics
Number of household members
20,255
Asset score [1-10]
20,383
Number of children living in household
16,159
Having savings account
20,382
Receiving government subsidy
20,371
Panel C. Village Characteristics
Holeta (=1) vs. Ejere (=0)
233
Population
233
Gender ratio (F/M)
233

(2)
All

(3)
Part-Time

(4)
Full-Time

(5)
Diff. (PT-FT)

(6)
p-value

26.032
0.419

25.740
0.441

26.329
0.397

-0.588
0.044

0.339
0.160

0.202
0.735

0.177
0.754

0.227
0.716

-0.050
0.038

0.201
0.430

0.415
0.581

0.372
0.623

0.460
0.538

-0.088
0.085

0.235
0.256

0.694
0.251
0.021
0.389

0.660
0.275
0.026
0.376

0.729
0.226
0.016
0.402

-0.068
0.049
0.010
-0.026

0.205
0.312
0.176
0.516

0.132
0.195

0.090
0.193

0.175
0.196

-0.085*
-0.003

0.074
0.952

4.216
4.582
2.501
0.278
0.016

4.166
4.474
2.496
0.266
0.018

4.267
4.693
2.505
0.292
0.013

-0.101
-0.219
-0.009
-0.026
0.004

0.499
0.679
0.695
0.695
0.307

0.350
359.6
0.510

0.397
356.2
0.505

0.301
363.1
0.515

0.096
-6.817
-0.010

0.450
0.859
0.591

Notes: This table presents descriptive statistics on individual, household, and village characteristics for the population of
potential applicants in the recruitment areas. Columns 2, 3, 4, and 5 show means for all villages, villages with part-time and
full-time job postings, and the mean differences between the part-time and full-time villages. Column 6 shows the p-value for
the mean differences. Variables under Ethnicity, Language, and Religion = 1 if the applicant belongs to the ethnic group, is
able to use the language, and has the religion. Post-secondary education = 1 if the applicant has education at a post-secondary
level. Working within household (in official sector) = 1 if the applicant is employed within the household (in an official sector).
Number of household members is the number of individuals in the household. Having savings account = 1 if anyone in the
household has a savings account. Receiving government subsidy = 1 if anyone in the household receives a government subsidy.
Holeta = 1 if the viliage is in Holeta (= 0 if in Ejere). Population is the number of individuals enumerated in the census in the
village. Gender ratio (F/M) is the ratio between females and males in the village. See notes to Table 2 for the definitions of
other variables. The variables are collected in the census of the recruitment areas. * denotes the significance level at 10%.

8

Table A2. Comparison of Non-Applicants and Applicants
(1)
(2)
Non-applicants
N
Mean
Age
5,844
26.2
Married
5,844 0.426
Ever given birth
4,601 0.494
Number of children living in household 5,340 1.367
Working
5,848 0.299
Working in official sector
5,756 0.199
Post-secondary education
5,950 0.380
Asset score [1-10]
5,934 7.013
Supportive spouse for work
5,227 3.958

Variable / Sample

(3)
(4)
(5)
Applicants Difference
N Mean
(4)-(2)
316 23.1
-3.1***
323 0.291 -0.135***
276 0.330 -0.165***
304 0.819 -0.548***
324 0.182 -0.117***
322 0.118 -0.081***
315 0.556 0.175***
330 6.918
-0.095
280 4.259 0.301***

Notes: This table compares the mean characteristics of the non-applicants (columns 1 and 2) and applicants
(columns 3 and 4) among potential applicants in the recruitment areas. Column 5 shows the mean
differences between the applicants and non-applicants. Ever given birth = 1 if the worker has ever given
birth to child(ren). Post-secondary education = 1 if the worker has education at a post-secondary level.
Working = 1 if the worker is employed within the household or in an official sector. See notes to Table 2 for
the definitions of the other variables. The variables are collected in the census of the recruitment areas. ***
denotes the significance level at 1%.

9

10

OLS

(1)

0.05
0.010
(0.070)
Y
Y
Y
Y
0.469
4,821

(2)

0.1
-0.014
(0.072)
Y
Y
Y
Y
0.479
4,821

(3)

(4)
(5)
(6)
Productivity
Quantile regression
0.25
0.5
0.75
-0.046 -0.042 -0.064
(0.060) (0.058) (0.130)
Y
Y
Y
Y
Y
Y
Y
Y
Y
Y
Y
Y
0.484
0.494
0.494
4,821
4,821
4,821

(8)

0.9
0.95
-0.333 -0.575***
(0.210)
(0.180)
Y
Y
Y
Y
Y
Y
Y
Y
0.454
0.389
4,821
4,821

(7)

Notes: Column 1 reproduces the OLS estimates in column 1 of Table 3, Panel A. Columns 2 through 8 present quantile regression
estimates of equation (2) from quantiles 0.05 to 0.95 of the labor productivity distribution among the interns. The dependent variable is
standardized error-adjusted typing speed or data-entry speed. P art = 1 (0) if the intern is recruited in a village where the part-time
(full-time) job was posted. Standard errors clustered at the village group level are reported in parentheses. *** denotes the significance
level at 1%.

-0.123
(0.089)
Y
Productivity measure fixed effects
Wave fixed effects
Y
Work Day fixed effects
Y
Trial fixed effects
Y
2
R
0.500
N
4,821

Dep. Var.:
Estimates:
Quantile:
Part

Table A3. Quantile Regression of Labor Productivity on Part-Time Recruitment Status

Table A4. Internship Attendance by Part-Time Recruitment Status
All interns
(1)
(2)
Dep. Var.:
Part
Constant
Wave fixed effects
Work day fixed effects
Trial fixed effects

R2
N

-0.023
(0.024)
0.914***
(0.013)
N
N
N
0.002
3,538

-0.032
(0.024)
0.918***
(0.014)
Y
Y
Y
0.044
3,538

Above-the-median interns Below-the-median interns
(3)
(4)
(5)
(6)
1(Attend)
0.018
0.004
-0.077**
-0.075**
(0.032)
(0.036)
(0.037)
(0.030)
0.912***
0.920***
0.914***
0.913***
(0.023)
(0.023)
(0.013)
(0.016)
N
Y
N
Y
N
Y
N
Y
N
Y
N
Y
0.001
0.042
0.014
0.079
1,769
1,769
1,769
1,769

Notes: This table shows estimates of linear probability models that explain the intern’s attendance by the
part-time recruitment status. The dependent variable is an indicator = 1 if the intern attends in a given
work day-trial in the second or third week of the internship. P art = 1 (0) if the applicant is recruited in a
village where the part-time (full-time) job was posted. Standard errors clustered at the village group level
are reported in parentheses. *** and ** denote the significance level at 1% and 5%, respectively.

11

12

0.056
(0.055)
Y
Y
Y
Y
Y
0.627
4,890

-0.123
(0.089)

Y
Y
Y
Y
0.503
4,890

Y
Y
Y
Y
Y
0.528
4,890

-0.068
(0.088)

All interns
(2)
(3)
0.054
(0.054)
Y
Y
Y
Y
Y
Y
0.631
4,890

(4)

Above-the-median interns
(5)
(6)
(7)
(8)
Productivity
-0.411*** -0.081 -0.332*** -0.066
(0.100)
(0.094)
(0.118)
(0.091)
Y
Y
Y
Y
Y
Y
Y
Y
Y
Y
Y
Y
Y
Y
Y
Y
Y
Y
Y
Y
0.513
0.636
0.547
0.645
2,511
2,511
2,511
2,511

Y
Y
Y
Y
0.543
2,310

0.036
(0.054)

Y
Y
Y
Y
0.592
2,310

0.053
(0.048)
Y

Y
Y
Y
Y
Y
0.550
2,310

0.061
(0.053)

0.049
(0.050)
Y
Y
Y
Y
Y
Y
0.595
2,310

Below-the-median interns
(9)
(10)
(11)
(12)

Notes: Columns 1, 5, and 9 reproduce OLS estimates from Panel A of Table 3. Columns 2, 6, and 10 include measures of applicants’ quality shown in Panel A of
Table 2 as controls, including data-entry test score, clerical ability, computation ability, computer literacy, manual dexterity, years of education, and official sector
working status. Columns 3, 7, and 11 include variables that capture preference for short work hours shown in Panel B of Table 2 as controls, including preference
for family to work, preference for non-work, working part- to full-time, preference for part-time to full-time work, (reverse) supportive spouse for work, and
number of children in household. Columns 4, 8, and 12 include both the quality and work hour preference measures as controls. The dependent variable is
standardized error-adjusted typing speed or data-entry speed. P art = 1 (0) if the intern is recruited in a village where the part-time (full-time) job was posted.
Standard errors clustered at the village group level are reported in parentheses. *** denotes the significance level at 1%.

Controls for quality
Controls for work hour preference
Productivity measure fixed effects
Wave fixed effects
Work day fixed effects
Trial fixed effects
R2
N

Dep. Var.:
Part

(1)

Table A5. Estimates of Part-Time Productivity Gaps with Controls for Quality and Work Hour Preference

Data Appendix
B.1 Ability tests
O*NET Ability Profiler (O*NET score): clerical and computation ability tests
The O*NET Ability Profiler was originally developed by the US Department of Labor as
“a career exploration tool to help understand job seekers on their work skills” (O*NET Resource
Center 2010, 1). We use the clerical and computation ability tests of the Ability Profiler because
these skills are most relevant for the data-entry work.

(A) The clerical perception test measures an individual’s ability to see details in written
materials quickly and correctly. It involves noticing if there are mistakes in the text
and numbers, or if there are careless errors in working math problems (O*NET Resource
Center 2010, 2). The following is an example of the test questionnaire.

(B) The computation test measures an individual’s ability to apply arithmetic operations to
calculate solutions to mathematical problems. It consists of 20 questions. The following
is an example of the test questionnaire.

13

Bruininks-Oseretsky Test of Motor Proficiency, 2nd edition (BOT™-2)
The BOT™-2 was developed to measure various types of motor skills. It consists of eight
tasks: fine motor precision, fine motor integration, manual dexterity, bilateral coordination,
balance, running speed and agility, upper limb coordination, and strength. We use the manual
dexterity test, which is most relevant for the data-entry work. We asked survey participants
to transfer 20 small coins from a table to a small box in 15 seconds. Study participants could
try twice, and the higher number is the final score. The following image depicts the manual
dexterity test.

14

B.2 Measures of preference for short working hours
We measure the applicants’ preference for short working hours using three sets of survey
questions. The first measures the importance of family over work using ten survey questions.
We calculate a composite score of preference for family over work by subtracting the average
score for work from that for family. Scores range from ten to 50, and a higher score implies
stronger preferences for family (i.e., shorter hours).

Second, we measure preference for full-time, part-time, and no-work arrangements in five
life stages. To calculate a composite score, we assign one, two, and three for full-time, part-time
work, and no-work arrangements, respectively, and add scores for the five questions. A higher
score implies stronger preferences for working short hours.

15

Third, we measure preference for part-time work, relative to those for compensation and
work that they like. We assign one when individuals choose the arrangement with a part-time
option (B in Q509-1 and Q509-2), and zero otherwise. We calculate a composite score by adding
scores for the two questions. A higher score implies stronger preferences for part-time work.

B.3. Expectations toward work
Intrinsic motivation
Intrinsic motivation is an individual trait that captures whether the individual is motivated to
do things by intrinsic rewards such as his/her own desire to pursue goals or challenges. It is
the opposite of extrinsic motivation, described below. We measure intrinsic motivation using a
15-item scale (Amabile et al. 1994). All items were answered using a 4-point Likert scale format
ranging from strongly agree (1) to strongly disagree (4). We calculate an average score after
accounting for (any) reverse coding.

16

Extrinsic motivation
Extrinsic motivation is an individual trait that captures whether the individual is motivated to
act by external rewards, such as reputation and monetary rewards. We use a 15-item scale to
measure the level of motivation triggered by extrinsic values (Amabile et al. 1994). All items
were answered using a 4-point Likert scale format ranging from strongly agree (1) to strongly
disagree (4). We calculate an average score after accounting for (any) reverse coding.

17

Accomplishment and status seeking
These modules, developed by Barrick, Stewart, and Piotrowski (2002), measure different types of
motivation to work. The accomplishment-seeking module measures how much one cares about
achievement in work. The status-seeking module measures how much one cares about what
other people think of oneself and about one’s status relative to other members of the group.
All items were answered using a 4-point Likert scale format ranging from strongly agree (1) to
strongly disagree (4). We calculate an average score after accounting for (any) reverse coding.

18

Career progress concern
This module measures how much one cares about his/her career in the future. All items were
answered using a 4-point Likert scale format ranging from strongly disagree (1) to strongly agree
(4). We calculate an average score after accounting for (any) reverse coding.

19

Concerns for compensation and benefits
This module measures how much one cares about the compensation and benefits of jobs. All
items were answered using a 4-point Likert scale format ranging from strongly disagree (1) to
strongly agree (4). We calculate an average score after accounting for (any) reverse coding.

References for Data Appendix
Amabile, Teresa M., Karl G. Hill, Beth A. Hennessey, and Elizabeth M. Tighe. 1994.
“The Work Preference Inventory: Assessing Intrinsic and Extrinsic Motivational Orientations.”
Journal of Personality and Social Psychology 66.5: 950-967.
Barrick, Murray R., Greg L. Stewart, and Mike Piotrowski. 2002. “Personality and
Job Performance: Test of the Mediating Effects of Motivation among Sales Representatives.”
Journal of Applied Psychology 87.1: 43–51.
O*NET Resource Center. 2010.
https://www.onetcenter.org/AP.html.

O*NET Ability ProfilerT M Score Report.

20

Pp.

1–2.