1 s2.0 S0014292115001828 Main

European Economic Review 82 (2016) 212–230
Contents lists available at ScienceDirect
European Economic Review

journal homepage: www.elsevier.com/locate/eer
The labor market return to academic fraud

Astghik Mavisakalyan a, Juergen Meinecke b,n
a
Bankwest Curtin Economics Centre, Curtin Business School, Curtin University, GPO Box U1987, Perth WA 6845, Australia
b
Australian National University, Research School of Economics, HW Arndt Building 25a, Canberra, ACT 0200, Australia
a r t i c l e in f o abstract
Article history: Academic fraud by undergraduate students is pervasive, but should it be taken seriously as
Received 24 November 2014 an economic problem? Our research suggests so. Using a unique data set from the Cau-
Accepted 19 November 2015 casus, we estimate a large positive effect of academic fraud on the probability of
Available online 29 December 2015
employment. Econometrically, we deal with endogenous selection into academic fraud
JEL: and possible measurement error in the reporting of academic fraud using partial identi-
D82 fication techniques. The findings demonstrate that incentives to commit academic fraud
I21 are strong and point towards the potentially damaging consequences of academic fraud in
J21 broader settings.
J24
& 2015 Elsevier B.V. All rights reserved.
C21
Keywords:
Academic fraud
Labor market signaling
Selection on unobservables
Partial identification
1. Introduction
University students cheat, bribe and plagiarize. It happens everywhere and it happens a lot. Not even elite institutions
are immune to the problem (recently at Harvard University, 125 undergraduate students were accused of cheating in a take-
home final exam). Just over one third of people surveyed in the US by Transparency International perceive the education
system as corrupt or extremely corrupt (Transparency International, 2013). Although ubiquitous, academic fraud may not
necessarily be an economic problem. Our research, however, suggests otherwise.
What precisely do we mean when we talk about academic fraud? Abstractly, we refer to the actions undertaken by
university students with the intention to manipulate the quality signal that they send into the labor market.1 Simply put,
academic fraudsters pretend to be better types than they actually are. Academic fraud manifests in many different ways:
from cheating in tests and plagiarizing in essays to outright payment of bribes. Our prima facie research question is ‘What is
the causal effect of academic fraud on labor market outcomes?’
Economists will not be surprised that academic fraud is widespread: the incentives are strong. Opportunities to commit
academic fraud are numerous, the payoff is high relative to the cost and the perceived risk. Students want better grades to
n
Corresponding author.
E-mail addresses: astghik.mavisakalyan@curtin.edu.au (A. Mavisakalyan), juergen.meinecke@anu.edu.au (J. Meinecke).
1
More broadly, academic fraud is an umbrella term that spans a wide range of ethically, if not legally, dubious actions by various players in the
education market. Academic fraud can be committed by students, teachers, researchers, and educational institutions; it can happen at all levels of edu-
cation. Transparency International (2013) is full of examples of academic fraud of varying connotation.
http://dx.doi.org/10.1016/j.euroecorev.2015.11.005
0014-2921/& 2015 Elsevier B.V. All rights reserved.
A. Mavisakalyan, J. Meinecke / European Economic Review 82 (2016) 212–230 213
manipulate and distort the quality signal that they send into the labor market. Good grades are associated with better job
offers. This may be especially important to young persons who tend to have higher rates of time preference. Furthermore,
among young persons, university students tend to have the highest earnings potential and therefore face higher opportunity
cost of an adverse labor market outcome, perhaps raising the temptation for academic fraud.
Why is it important to study the effects of academic fraud? Academic fraud reduces total welfare of a society. While it
benefits those who commit it, academic fraud creates a cost that is borne by the honest; it creates negative externalities. The
analogy to the used car market is striking: fraudsters could be regarded as ‘lemons’ whose existence creates asymmetric
information. Akerlof (1970) asserts that in “a market in which goods are sold honestly or dishonestly;quality may be
represented, or it may be misrepresented. […] The cost of dishonesty, therefore lies not only in the amount by which the
purchaser is cheated; the cost also must include the loss incurred from driving legitimate business out of existence.”
Academic fraud weakens the quality signal that job market candidates, fraudsters and non-fraudsters alike, send to
potential employers. This should concern both non-fraudsters and employers. Spence (1973) presents his theory on job
market signaling in which the level of education signals productive capability. A critical assumption is, of course, that the
costs of signaling are inversely related to productive capability. The signaling described by Spence (1973) can be interpreted
as signaling at the extensive margin: people choose education levels to signal their types. The type of signaling we consider
is at the intensive margin: given that people are at university, they can choose to exert effort to achieve good grades. The
costs of signaling are therefore inversely related to productive capability. The possibility of academic fraud, however,
reduces the costs of sending a positive signal. The possibility of academic fraud will be particularly attractive for those
people who face high marginal costs of sending a positive signal (if they were to do it truthfully).2
Despite the gravity of the problem in the perception of people and the potential social costs, we are not aware of any
research in economics that attempts to empirically quantify the labor market returns to academic fraud. Lack of data and
endogeneity are likely the two main explanations. To narrow this research gap, we use a unique data set from the three
Caucasus countries of Armenia, Azerbaijan and Georgia that contains information on academic fraud, labor market out-
comes and a rich set of explanatory variables that enable us to proxy for, among other things, a person's ability, access to
corrupt networks, locus of control and risk preferences. In the corresponding survey, conducted in 2008, respondents were
asked if they have ever paid a bribe or given gifts with the intention to obtain a better grade.
From an econometric perspective, we are confronted with a binary outcome variable (employment) and a binary main
explanatory variable, academic fraud. Accordingly, our principal vehicle to estimate the effects of academic fraud on
employment is the probit model. However, there are two sources of endogeneity that can lead to biased estimates. First,
academic fraudsters may differ in unobserved ways to non-fraudsters. If, for example, academic fraudsters would tend to
attain better labor market outcomes even if they had not committed fraud, then our causal effect estimates may be, in the
‘worst case’, entirely spurious. A conventional way to deal with this possibility is to use an instrumental variable and a
bivariate probit model. In our application, however, we do not have available a persuasive instrumental variable and we are
reluctant to solely rely on functional form assumptions for identification of causal effects. (Altonji et al. (2005) explain the
drawbacks of the bivariate probit model when instruments are weak or absent.) Instead, we use a parametric partial
identification approach proposed by Altonji et al. (2005b). This approach acknowledges that unobserved heterogeneity can
never be fully and adequately accounted for (in particular in the absence of an instrumental variable). The main idea of this
approach is to study the ‘worst case’ scenario: how large would the amount of selection on unobservables need to be,
relative to the amount of selection on observables, to explain away the entire casual effect of academic fraud? This so-called
selection ratio can be quantified as we explain and pursue in Section 4.
The second source of endogeneity is measurement error in the main explanatory variable; some survey respondents may
not answer questions about academic fraud truthfully. Such misreporting can create bias in the causal effect estimate if
persons who do not answer truthfully in the survey differ in unobserved ways from persons who do answer truthfully. A
priori, it is not clear which direction this bias will point to. In Section 5.1, we use nonparametric partial identification
methods in a three-pronged approach. First, we use potential outcomes terminology to motivate the causal effect of aca-
demic fraud as a particular average treatment effect. Second, we need to be transparent and precise about the subpopulation
for which we partially identify that average treatment effect. Because true academic fraud status is unobserved, we cannot
identify an ATE for the population at large but only for the subset of people who respond truthfully (this subpopulation
consists of non-fraudsters who respond that they did not commit fraud as well as fraudsters who respond that they did
commit fraud). Third, we derive that, under a weak set of conditions, endogenous misreporting will bias the ATE estimate
downward. The strength of this analysis lies in its agnosticism with regard to endogenous misreporting: it does not matter
how many persons actually misreport and it does not matter what type of person tends to misreport.
In Section 5.2 we turn to simultaneously addressing the two econometric problems of selection on unobservables and
measurement error. We revisit the selection ratio of Section 4 and ask the hypothetical: what would the value of the
selection ratio be equal to, if we observed actual academic fraud instead of reported academic fraud? While in the absence of
the actual data we cannot point identify this hypothetical selection ratio, we are able to bound it from below. Conveniently,
we derive a bound that depends on the selection ratio that is based on reported (rather than actual) academic fraud status.
2
Also see Arrow (1973) and Stiglitz (1975) for related work as well as Riley (2001) for a review.
214 A. Mavisakalyan, J. Meinecke / European Economic Review 82 (2016) 212–230
Our main findings confirm that the effect of academic fraud on the probability of employment is positive and large.
Estimation results from the most extensive parametric model suggest marginal effects of 11.6 percentage points for males
and 9.2 percentage points for females. This implies that the employment probability is almost 15% higher for males and 17%
for females. These are economically significant effects. We show that the influence of unobserved heterogeneity would need
to be exceedingly large for these effects to disappear (although for females this possibility cannot be ruled out entirely).
Furthermore, misreporting of academic fraud does not pose a threat to these findings—if anything, it affords them an even
more conservative interpretation.
Our study can be linked to the large economics literature on corruption (see, for example, Olken and Pande, 2012;
Svensson, 2005 for reviews). Relevant for our work are the micro-level studies on the relationship between bribe-payment
and firm growth, since they, similarly to our own study, investigate the consequences of bribing for the outcomes of bribe-
payers. On the one hand, these studies document a negative effect of bribery on firm efficiency and growth (e.g., Kaufman
and Wei, 1999; Fisman and Svensson, 2007). On the other hand, there is empirical evidence supporting the idea that cor-
ruption is beneficial in places with bad institutions (e.g., Meon and Weill, 2010; Dreher and Gassebner, 2011). There is
relatively limited knowledge of the process of bribery by individuals. A study by Hunt (2007) shows that victims of mis-
fortune are more likely to bribe, and that consequently, bribery lowers the quality of life of individuals by compounding
other miseries brought on by misfortune.
The problem of corruption in education has been widely discussed in the education literature. This literature is largely
concerned with conceptualizations of the problem of corruption in the education sector, describing various types of cor-
ruption, their causes and consequences (e.g., Heyneman, 2004; Osipian, 2009). Some of these descriptive studies focus on
the countries of the South Caucasus (Janashia, 2004; Temple and Petrov, 2004). Others attempt to quantify and measure
corruption. Osipian (2007) proposes conceptual approaches to measuring corruption in higher education across countries;
Heyneman et al. (2008) empirically pursue the task of cross-country comparisons of corruption. Using data from Trans-
parency International on perceptions of corruption in education in 68 countries, their evidence suggests that perceptions of
corruption can adversely affect the payoff to education. However, Olken (2009) shows that perceptions of corruption contain
only a limited amount of information on the reality of corruption.
Two recent papers in the economics literature have studied corruption in the education sector; both papers focus on
academic fraud by educators. Borcan et al. (2014) study how an unexpected wage cut in the public services sector can lead to
corruption by public teachers. They show that when teachers wages drop suddenly, exam outcomes at public schools
improve relative to private schools. The authors attribute this selective improvement in exam outcomes to increased cor-
ruption by teachers at public schools. Jacob and Levitt (2003) study the corrupt behavior of teachers at Chicago public
schools. They develop an algorithm that detects teacher cheating (combining fluctuations in test scores with suspicious
answer patterns by students). The authors infer that serious cases of teacher cheating happen at about 4–5 percent of
elementary school classrooms every year. These corrupt patterns can be linked to the presence of accountability systems in
which test scores are used to punish or reward schools.
For the most part, however, research focuses primarily on two types of academic fraud: cheating and plagiarism (e.g.,
Bowers, 1964; Davis et al., 1992; Murdock and Anderman, 2006; Burrus et al., 2007; Whitley, 1998). These studies
mainly look at the drivers of academic fraud rather than its consequences. One exception is a study by Sims (1993)
which shows that those who admit to having engaged in academic dishonesty also admit to dishonesty later during
employment. Other research suggest that fraud is influenced by a number of characteristics of individuals including age,
gender, grade-point average (e.g., Bunn et al., 1992; Diekhoff et al., 1996; Newstead et al., 1996), as well as contextual
factors including peer influence, institutions' academic integrity programs and policies, and the perceived severity of
penalties for cheating (e.g., McCabe and Klebe Trevino, 1997; McCabe et al., 2001; Carrell et al., 2008). As McCabe and
Klebe Trevino (1993) note, the contextual variables are particularly relevant because they are open to administrative
influence.3
Some of the contextual variables identified in the literature are also important in driving the academic fraud practices in
the countries of our study. While there exists no systematic research on the importance of peer effects in the region, a study
by the OSCE (2010) based on a sample of 1200 university students in Armenia in 2010 found that 40 percent personally
knew someone who had been involved in academic fraud. Furthermore, the perceived severity of penalties for committing
fraud appears to be low. According to the same study, over 30 percent of students were confident that those involved in a
case will not be punished if detected because corruption is systemic. Temple and Petrov (2004) concord that in the states of
the former Soviet Union “university corruption is a particular facet of systemic corruption, being part of a wider culture of
corruption” (p.86). Another Armenian study by Aleksanyan (2012) concludes that the lack of corruption control mechanisms
and integrity campaigns in universities, among other things, is an important challenge in addressing the high levels of
corruption. Lastly, the low salaries of university teaching staff are an additional contextual factor of importance (OSCE, 2010;
Aleksanyan, 2012; Borcan et al., 2014).
While corruption in education is a broadly acknowledged problem, the example of the three Caucasus countries
nevertheless has idiosyncrasies that will ultimately limit its external validity. The particular manifestation of academic fraud
(paying bribes and giving gifts) together with its high degree of pervasiveness are peculiar and would likely be different
3
Gallant and Drinan (2006) and Lazear (2006) discuss institutional approaches to foster academic integrity and maximize learning.
somewhere else. Notwithstanding the idiosyncrasies of our setting, it offers many parallels with countries that are more
developed: the labor market is competitive, characterized by high labor force participation and employment levels;
opportunities to commit academic fraud abound; academic fraud is pervasive; the costs of committing academic fraud are
low; the risk of committing academic fraud is low; most importantly, while academic fraud may manifest in different ways,
it serves as a vehicle for the same purpose: to jump the queue and thereby mimic a better type. Given these parallels, the
findings here will, at the very least, provide a window into the potential negative consequences of academic fraud for
countries in which the problem is not yet regarded as one.
2. Data and descriptive statistics
We use data drawn from the annual Data Initiative (DI) survey (renamed into Caucasus Barometer in 2010) conducted by the
Caucasus Research Resource Centers (CRRC) with technical assistance from a group of international experts and financial support
from several multinational organizations such as the United Nations and the World Bank, in Armenia, Azerbaijan, and Georgia.
The DI was designed with the goal to “collect data on a wide variety of social, political, and economic indicators in the South
Caucasus region” (CRRC, 2007). Since 2007, the target population for the DI has consisted of all households in Armenia, Azer-
baijan, and Georgia with the survey methodology based on multistage stratified clustered sampling with probability proportional
to size (see CRRC, 2008 for further details). Sampling, questionnaire design, interviewers' training, field procedures, data coding
and archiving are harmonized across countries to ensure the comparability of collected data. The DI is one of the few high quality
nation-wide surveys conducted in the countries of South Caucasus and has been used in other published studies on the region
(Charles, 2010; Habibov and Afandi, 2009, 2011; Habibov, 2011, 2012; Roberts and Pollock, 2011).
The DI surveys are repeated cross sections and we use 2008 data (wave 5) because it is the only year during which questions
on academic fraud were included. The DI2008 data consists of 5900 observations in total of which 2082 are from Armenia, 2014
from Azerbaijan, and 1804 from Georgia. We restrict the sample to persons between the ages of 26 and 55 (to exclude students
and early retirees) and drop observations with missing outcome data. The final size of our sample is 3009 observations.
The DI2008 contains survey questions that ask for respondents’ academic fraud behavior. Based on that, we define the
variable FRAUD that takes on the value 1 if a respondent reported ever having paid bribes, presented gifts or used connections
with the intention to get desired grades and 0 otherwise. The use of household surveys to elicit direct responses on bribe
payment is viewed as one of the most reliable approaches to measuring bribing activity, particularly in developing countries
(see, for example Banerjee et al., 2013; Olken and Pande, 2012). First, there is relatively little stigma associated with bribery
because of its high prevalence and acceptance as a systemic problem (see Hunt, 2007). In Armenia, for example, nearly 65%
in the sample of 1528 adults surveyed by the CRRC in 2010 agreed with the statement that “citizens of Armenia consider
corruption as a fact of life” (CRRC, 2013). According to Temple and Petrov (2004), the routinization of corruption in higher
education in the states of the Former Soviet Union means that “secrecy is out of question”. Second, given the focus of anti-
corruption measures on officials, the fear of prosecution in reporting payment of bribes should be low. According to CRRC
(2013), within the sub-sample of respondents who self-identify as non-bribers (they would not pay a bribe even if asked to),
less than 5% mention their fear of punishment as an explanation.
Similar survey-based approaches to measuring bribing activity have been used in other studies on developing countries
(see, for example, Svensson, 2003; Hunt, 2007). Nevertheless, there are some possible drawbacks associated with the use of
the FRAUD variable in the context of this study. While we are able to identify academic fraudsters, we do not have information
on the details of the transaction, including amounts paid and exact timing.4
Table 1 contains definitions and descriptive statistics for all variables used in our baseline analysis. The table shows that
69% of people are in the labor force while 50% are actually employed. We define three ordinal education categories: EDUC I
consists of persons with at most ten years of formal education, comprising 39% of the sample; EDUC II consists of persons
with further education (for example, technical or incomplete university education), comprising 36% of the sample; and EDUC
III consists of persons with completed tertiary education, comprising 24% of the sample. About half of the population
is male.
The population is relatively evenly distributed across six age groups, the vast majority of people belong to ethnic and
linguistic majorities and most people live in urban areas. About 40% of the population is estimated to be in good or very
good health. Azerbaijan and Armenia together make up just above 80% of the sample, evenly split.
Table 2 takes a descriptive look at the effect of academic fraud. Restricting the sample to university graduates, the
table reports labor force participation rates as well as employment probabilities by gender. Looking first at the pro-
pensity of academic fraud, males are more likely to commit academic fraud than females (33% versus 21%). Next,
focusing on the labor force participation rates, while there is marked variation across genders, labor force participation
does not vary between non-fraudsters and fraudsters keeping gender constant. The labor force participation rate for
4
Pieces of information on these aspects of corruption in the education sector are provided in other sources. For example, based on the corruption
survey conducted by CRRC in Armenia in 2010, the average amount of extra contribution paid as a result of a contact with an education service provider
(kindergartens, schools, universities, etc.) was 21,118 Armenian drams (equivalent to around 55 US dollars), ranging from 1000 to 120,000 drams (CRRC,
2013). According to Aleksanyan (2012), at the university level, the main links in the ‘corruption chain’ include: admission exams, mid-term exams, final
graduation exams, master's degree exams and post-graduate exams.
Table 1
Descriptive statistics: sample means and standard deviations.
Variable Definition of variable Mean
IN LABOR FORCE 0–1 binary variable; equals 1 if 0.69

respondent is in the labor force (0.46)
EMPLOYED 0–1 binary variable; equals 1 if 0.50
respondent reported having a job (0.50)
MALE 0–1 binary variable; equals 1 if 0.50

respondent is male (0.50)
EDUC I 0–1 binary variable; equals 1 if 0.39
at most 10 years of formal education (0.49)
EDUC II 0–1 binary variable; equals 1 if 0.36
between 11 and 14 years of formal education (0.48)
EDUC III 0–1 binary variable; equals 1 if 0.24
respondent graduated with tertiary degree (0.43)
BASELINE CONTROLS:
AGE 26–30 0–1 binary variable; equals 1 if 0.18
respondent is aged 26–30 (0.38)
AGE 31–35 0-1 binary variable; equals 1 if 0.16
AGE 41–45 0-1 binary variable; equals 1 if 0.18
ETHNIC MAJORITY 0–1 binary variable; equals 1 if 0.94
respondent is from ethnic majority (0.24)
LINGUISTIC MAJORITY 0–1 binary variable; equals 1 if 0.96
respondent is from linguistic majority (0.20)
CAPITAL CITY 0–1 binary variable; equals 1 if 0.33
respondent lives in capital city (0.47)
OTHER URBAN 0-1 binary variable; equals 1 if 0.31
respondent lives in urban area (0.46)
RURAL 0–1 binary variable; equals 1 if 0.36
respondent lives in rural area (0.48)
GOOD HEALTH 0–1 binary variable; equals 1 if 0.39
respondent of good/very good health (0.49)
HOUSEHOLD SIZE Number of persons living in same 4.78
household (1.82)
AZERBAIJAN 0–1 binary variable; equals 1 if 0.42
respondent lives in Azerbaijan (0.49)
ARMENIA 0–1 binary variable; equals 1 if 0.41
respondent lives in Armenia (0.49)
GEORGIA 0–1 binary variable; equals 1 if 0.18
respondent lives in Georgia (0.38)
Note: Means representative of population; standard deviations in parentheses. N ¼ 3009.
both non-fraudsters and fraudsters in the sub-sample of males is 96%. Among females, this rate is 72% for non-
fraudsters as well as fraudsters.
In contrast, for the employment probability there is a considerable gap between non-fraudsters and fraudsters for both
males and females. Males who commit academic fraud have an employment probability that exceeds that of non-fraudsters
by 8 percentage points. For females that number equals 11 percentage points.
3. Probit estimation
3.1. Baseline specification
We undertake probit estimation of employment status on the academic fraud dummy variable. We estimate the model

Employedi ¼ 1 α Fraud þ β ðFraud MaleÞ
þ γ 0 þ γ 1 Male þγ 2 Educ II þγ 3 ðEduc II MaleÞ

þ γ 4 Educ III þ γ 5 ðEduc III MaleÞ þ W 0 γ 6 þ ui 4 0 ;
Table 2
Descriptive statistics: sample mean of Fraud and Labor Market Outcomes (by gender); subsample of University Educated.
Gender FRAUD IN LABOR FORCE EMPLOYED
FRAUD All FRAUD All
0 1 0 1
(1) (2) (3) (4) (5) (6) (7)
Males 0.33 0.96 0.96 0.96 0.78 0.86 0.81

(0.47) (0.19) (0.19) (0.19) (0.41) (0.35) (0.40)
Females 0.21 0.72 0.72 0.72 0.55 0.66 0.57
(0.41) (0.45) (0.45) (0.45) (0.50) (0.48) (0.50)
Note: Sample restricted to university graduates (Educ III¼1). Column (1) reports sample mean of fraud status conditional on gender; columns (2) and
(3) report sample mean of labor force participation conditional on gender and fraud status; column (4) reports sample mean conditional on gender;
columns (5) and (6) report sample mean of employment outcome conditional on gender and fraud status; column (7) reports unconditional sample mean.
Standard deviations in parentheses. N ¼ 770.
where W collects the baseline controls from Table 1 and the error term ui follows a standard normal distribution (giving rise
to the probit model). We estimate two specifications: parsimonious (restricting γ 6 ¼ 0) and baseline (γ 6 a 0).
Table 3 presents the estimation results for coefficients as well as marginal effects for the parsimonious and the baseline
specification. Consistent with the descriptive statistics in the previous section, the estimated coefficients for the parsimo-
nious model (reported in column (1)) and for the baseline specification (reported in column (4)) confirm a significant and
positive relationship between academic fraud and employment. The education dummies have the expected monotone
effect: the higher the level of education, the larger is the coefficient estimate; the education coefficient estimates for males
are lower than those for females. Males, individuals from ethnic and linguistic majority background and those in good
health are all more likely to be employed. Individuals in the youngest cohort are less likely to be employed in comparison to
the omitted age group of 36–40 year olds while those aged 31–35 and 41–45 are more likely to be employed.
To aid with the interpretation of the coefficients, we turn to the calculation of marginal effects for university educated
non-fraudsters. Marginal effects are calculated as a discrete change of a variable from 0 to 1 for a person with EDUC III¼1,
FRAUD ¼0 and MALE ¼1 (columns 2 and 5) or MALE ¼0 (columns 3 and 6); baseline controls are fixed at their sample means.
For example, denoting the cumulative distribution function of the standard normal distribution by ΦðÞ, the marginal effect
of academic fraud for males is calculated as
Φðα^ þ β^ þ γ^ 0 þ γ^ 1 þ γ^ 4 þ γ^ 5 þ w 0 γ^ 6 Þ Φð^γ 0 þ γ^ 1 þ γ^ 4 þ γ^ 5 þw 0 γ^ 6 Þ;
while for females the marginal effect is
Φðα^ þ γ^ 0 þ γ^ 4 þ w 0 γ^ 6 Þ Φð^γ 0 þ γ^ 4 þ w 0 γ^ 6 Þ:
Standard errors for all marginal effects are estimated using the delta method.
In the parsimonious specification, the marginal effect of changing the variable FRAUD from 0 to 1 for males is equal to
0.076; for females it is equal to 0.108. In the baseline specification, the marginal effect for males is equal to 0.077; for females
it is equal to 0.074. Effects of this magnitude are economically substantial. For both males and females, these effects are
larger than the effects of moving to ethnic or linguistic majority status, and at least as large as the effects of moving from
poor to good health. Furthermore, for males the marginal effect is especially large given the already high levels of
employment: in the absence of academic fraud, a university graduate has a predicted employment probability of 77.8%; by
committing academic fraud this number can be raised to 85.4%.
Other marginal effects of interest concern the education and gender variables. For example, looking at the estimates from
the baseline specification (the estimates from the parsimonious specification are similar), the marginal effect of a university
education for a person with the lowest education category (EDUC I) is equal to 20.1 (males) and 35.9 (females) percentage
points. Alternatively, the marginal effect of a secondary education is estimated to equal 6.5 (males) and 17.7 (females)
percentage points. One of the largest marginal effects is gender-related: the marginal effect of changing the dummy variable
MALE from zero to one is equal to 22.1 percentage points. (This number, by construction, is the same for males and females.)
3.2. Additional controls
We attempt to mitigate the influence of unobserved heterogeneity by adding proxy variables that could be correlated
with hitherto unexplained components of employment. We see three main categories of unobserved heterogeneity: ability,
institutional connectedness, and personality traits. First, omitted ability may lead to downward bias in the baseline esti-
mates (if persons of low ability are more likely to commit academic fraud). The type of omitted ability that we are concerned
about is conditional on having attained the highest education level: university educated individuals differ in terms of
Table 3
Results: probit coefficients and marginal effects.
Variable Parsimonious specification Baseline specification
Coefficients Marginal effects Coefficients Marginal effects
Males Females Males Females
(1) (2) (3) (4) (5) (6)
FRAUD 0.282nnn 0.076nn 0.108nnn 0.196nn 0.077nnn 0.074nn

(0.090) (0.034) (0.037) (0.088) (0.026) (0.033)
FRAUD MALE 0.009 0.102
(0.179) (0.203)
MALE 1.051nnn 0.233nnn 0.233nnn 1.066nnn 0.221nnn 0.221nnn
(0.109) 0.015 (0.015) (0.117) (0.017) (0.017)
EDUC II 0.452nnn 0.061n 0.168nnn 0.492nnn 0.065nn 0.177nnn
(0.055) (0.032) (0.027) (0.074) (0.029) (0.028)
EDUC III 0.965nnn 0.200nnn 0.349nnn 1.014nnn 0.201nnn 0.359nnn
(0.111) (0.042) (0.050) (0.083) (0.023) (0.029)
Educ II MALE 0.223nnn 0.246nnn
(0.053) (0.055)
Educ III MALE 0.395nnn 0.431nnn
(0.060) (0.068)
BASELINE CONTROLS:
AGE 26–30 0.235nnn 0.069nnn 0.091nnn
(0.038) (0.014) (0.015)
AGE 31–35 0.088nnn 0.026nnn 0.034nnn
(0.014) (0.006) (0.006)
AGE 41–45 0.228nnn 0.066nnn 0.088nnn
(0.024) (0.010) (0.010)
AGE 46–50 0.075 0.022 0.029
(0.082) (0.024) (0.032)
AGE 51–55 0.086 0.025 0.033
(0.097) (0.029) (0.038)
ETHNIC MAJORITY 0.146nnn 0.043nnn 0.057nnn
(0.038) (0.015) (0.015)
LINGUISTIC MAJORITY 0.108n 0.032nn 0.042n
(0.056) (0.014) (0.021)
OTHER URBAN 0.022 0.006 0.009
(0.048) (0.014) (0.018)
RURAL 0.001 0.000 0.000
(0.152) (0.044) (0.059)
GOOD HEALTH 0.187nnn 0.055nnn 0.072nnn
(0.066) (0.020) (0.025)
HH SIZE 0.028 0.008 0.011
(0.024) (0.007) (0.009)
ARMENIA 0.070nn 0.020nn 0.027nn
(0.028) (0.008) (0.011)
GEORGIA 0.270nnn 0.079nnn 0.104nnn
(0.041) (0.012) (0.016)
Pseudo-R2 0.241 0.271
Note: Marginal effects for a discrete change of corresponding variable from 0 to 1 for a person with EDUC III¼ 1, FRAUD ¼0 and MALE ¼1 (columns 2 and 5) or
MALE ¼ 0 (columns 3 and 6); baseline controls are fixed at sample means. Pseudo-R2 calculated as VarðÞ=ð1 d d
þ VarðÞÞ. Standard errors clustered at
country level;
n
Significance at 10 percent level.
nn
nnn
Significance at 1 percent level. N ¼3009.
unobserved skills and competencies that are likely to be of value in the labor market. Second, unobserved institutional
connectedness may lead to upward bias in the baseline estimates (if a person's connections aid in ‘finding’ a job). Third,
omitted personality traits may lead to upward bias in the baseline estimates (if such traits are valued in the job market).
Important examples of such personality traits include internal locus of control and rates of time preference.
To address the presence of unobserved ability, we exploit information on computer and foreign language proficiency.
Specifically, we have a dummy variable, COMPUTER PROFICIENCY, which is equal to 1 if a person has an advanced or intermediate
self-reported ability in the use of computers and 0 otherwise. Similarly, we have a dummy variable ENGLISH PROFICIENCY to
proxy for the respondent's English language ability. Proficiency in these two skills is not common among university
graduates: only 34% in the sample are computer proficient and only 29% are English proficient. To the degree that they
Table 4
Results: Probit marginal effects—extensions to baseline specification.
Variable Males Females
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10)
nnn nn nn nnn n nn nnn nnn nn

FRAUD 0.077 0.081 0.082 0.096 0.116 0.074 0.071 0.077 0.091 0.092nnn
(0.026) (0.036) (0.035) (0.036) (0.062) (0.033) (0.024) (0.026) (0.043) (0.029)
COMPUTER PROFICIENCY 0.166nnn 0.170nnn 0.205nnn 0.199nnn

(0.003) (0.012) (0.021) (0.022)
ENGLISH PROFICIENCY 0.025 0.026 0.031 0.031n
(0.020) (0.017) (0.024) (0.018)
FAMLOCGOV 0.038nn 0.036nn 0.050nnn 0.042nnn
(0.016) (0.014) (0.018) (0.015)
FAMNATGOV 0.009 0.013 0.012 0.016
(0.010) (0.010) (0.012) (0.011)
FAMPOLICE 0.027nnn 0.021nn 0.035nnn 0.025nn
(0.008) (0.011) (0.008) (0.011)
FAMINTORG 0.051n 0.044nn 0.066nn 0.051nn
(0.027) (0.022) (0.030) (0.024)
HIGH LOCUS OF CONTROL 0.072nn 0.070n 0.090nnn 0.083nn
(0.030) (0.038) (0.032) (0.040)
SMOKER 0.053nn 0.054 0.067n 0.064
(0.027) (0.037) (0.038) (0.045)
Other controls included Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes
Pseudo-R2 0.271 0.303 0.279 0.285 0.324 0.271 0.303 0.279 0.285 0.324
N 3009 2855 2892 2973 2718 3009 2855 2892 2973 2718
Note: Marginal effects for a discrete change of corresponding variable from 0 to 1 for a person with EDUC III¼ 1, FRAUD ¼ 0 and MALE ¼ 1 (columns 1–5) or
MALE ¼0 (columns 6–10); all other variables fixed at their sample means. Standard errors are clustered at country level. Pseudo-R2 calculated as
d
VarðÞ=ð1 d
þ VarðÞÞ; note that pseudo-R2 are identical for males and females because underlying probit model pools data for both.
n
nn
nnn
capture an unobserved skill that offers a comparative advantage in the labor market, we do regard both computer and
English proficiency to be valid proxies for ability. Columns (2) and (7) of Table 4 report the marginal effect estimates for
males and females after adding the variables COMPUTER PROFICIENCY and ENGLISH PROFICIENCY to the estimation. (Columns (1) and
(6) restate the baseline estimates.) While English proficiency is not a significant predictor of employment, the marginal
effect of computer proficiency is significant and large, which suggests that computer proficiency does indeed absorb some of
the effects of omitted ability. This is supported by two auxiliary estimations (not reported in the table) in which we add the
two proficiency variables separately: the marginal effect estimates of computer proficiency remain unchanged (for both
males and females) while for English proficiency they turn significant and jump to 0.083 (males) and 0.107 (females)
indicating that, when included separately, English proficiency assumes the role of partially controlling for omitted ability.
Overall, however, the inclusion of the ability proxies does not significantly alter the marginal effect of academic fraud. These
results suggest that either the two proxy variables are only weakly correlated with ability or that unobserved ability is not
an important omitted explanatory variable or both.
To proxy for institutional connectedness we add four dummy variables that capture the potential to use personal con-
nections for self-advantage (in particular, getting a job). The first, FAMLOCGOV, is a dummy variable, which is equal to 1 if the
respondent has family members who work for the local government and zero otherwise. The second, FAMNATGOV, is a
dummy variable, which is equal to 1 if the respondent has family members working for the national government and zero
otherwise. The third, FAMPOLICE, is a dummy variable, which is equal to 1 if family members work for the police and zero
otherwise and the last, FAMINTORG is a dummy variable, which is equal to 1 if family members work for a large international
organization or company. The results are reported in columns (3) and (8) of Table 4. The marginal effects of FAMLOCGOV,
FAMPOLICE and FAMINTORG are all positive and statistically and economically significant. The marginal effect estimates of FRAUD
are not affected by the inclusion of the four added regressors. As in the previous paragraph, we cannot corroborate
empirically that the original estimates may have been biased because of omission of unobserved variables.
Last, we attempt to control for omitted psychological traits and values of an individual. The first variable we add to the
estimation is self-reported locus of control. The DI2008 asks respondents about the level of control they feel having over
their own economic situation. Answers range from 1 (no control at all) to 10 (complete control) and can be categorized in
three categories: low locus of control, intermediate locus of control and high locus of control. It is plausible to assume that
employment and locus of control over economic situation are positively correlated. Individuals with an internal locus of
control tend to exert more effort in job search (Caliendo et al., 2010) and have a higher job arrival rate. Furthermore, existing
studies suggest a positive link between internal locus of control and human capital investment decisions (Coleman and
DeLeire, 2003; Barón and Cobb-Clark, 2010).5 This variable proxies for high personal rates of time preference. Human capital
theory suggests that persons with high rates of time preference invest less in education (Mincer, 1958; Becker, 1964); they
may also be more likely to commit academic fraud. Both variables, when included together, are significant (at varying levels)
in the estimation. Columns (4) and (9) of Table 4 show that, with the addition of these variables, the marginal effect of
academic fraud climbs to 9.6 percentage points for males and 9.1 percentage points for females. These large effects are
confirmed by the final specification in which we include the full set of additional variables; columns (5) and (10) of Table 4
contain the results. The marginal effect of academic fraud on employment is estimated to be 11.6 percentage points for
males and 9.2 percentage points for females.
The estimates presented here suggest that the marginal effects of academic fraud on employment are statistically sig-
nificant and economically large. But are they real or could they be spurious? We turn to a more rigorous analysis of
unobserved heterogeneity and its potential impact in the next section.
4. Selection on unobservables
Our premise in this section is that we can never adequately control for, or mitigate the influence of, unobserved het-
erogeneity. The proxy variables from Section 3.2 can, at best, only partially capture the effect of unobserved factors. Fur-
thermore, we do not have available a natural experiment or a persuasive instrumental variable which could aid in identi-
fying a causal effect.6 Instead, we follow a parametric partial identification approach suggested by Altonji et al. (2005b) in
which we take a closer look at the potential ‘harm’ that can be done by unobserved heterogeneity.
The problem with unobserved heterogeneity is, of course, that it can lead to bias in the probit estimates. In the ‘worst
case’, it is conceivable that the entire estimated marginal effect of academic fraud is a result of bias and therefore any causal
interpretation would be spurious. Here, the main idea is to study how strong selection on unobservables would need to be
(relative to selection on observables) to purge the probit estimate of any causal interpretation.
Formally, consider the following parametric specification of the model:

Employedi ¼ 1 αZ i þX 0i γ þui 4 0 ; ð1Þ
where Zi denotes the academic fraud dummy variable and the error term ui still follows a standard normal distribution.
Selection on unobservables is defined by
Suo ≔E½ui jZ i ¼ 1 E½ui jZ i ¼ 0; ð2Þ
which is the difference in means of ui between fraudsters and non-fraudsters. Using a standard probit model to estimate the
parameters of Eq. (1), we obtain an estimate α^ for α that, in the ‘worst case’ is entirely an artifact of selection on unob-
servables. Altonji et al. (2005b) establish the following link between the probability limit of the probit estimate α^ and the
amount of selection on unobservables:
plimðα^ Þ αþ πSuo ; ð3Þ
with
π≔VarðZ i Þ=VarðZ i Þ; ð4Þ
where π is estimable and larger than (but close to) one and Z i
is the error in the regression of Zi on Xi (so that represents Z i
the variation in Zi that is orthogonal to Xi). Under the ‘worst case’ hypothesis that the causal effect of academic fraud is zero,
i.e. α ¼ 0, Eq. (3) has the useful practical implication that the maximal amount of selection on unobservables can simply be
estimated by
max
S^ uo ≔π^ 1 α;
^ ð5Þ
where α^ is the probit estimate in Eq. (1) and π^ ≔VarðZ d Þ and VarðÞ
d i Þ=VarðZ d denotes a straightforward analog estimator (see
i
Manski, 1988). Furthermore, under the hypothesis that α ¼ 0, we can consistently estimate all coefficients in Eq. (1) enabling
us to calculate the amount of selection on observables that must have been present in this ‘worst case’ scenario. Selection on
observables is quantified by

E^ X 0i γ^ r jZ i ¼ 1 E^ X 0i γ^ r jZ i ¼ 0
S^ o ≔ ; ð6Þ
d X 0 γ^ r
Var i
5
Similar to Coleman and DeLeire (2003) and Cebi (2007) we treat locus of control as an exogenous variable. Some evidence that changes in locus of
control are not affected by labor market events is offered by Cobb-Clark and Schurer (2013).
6
A common approach for dealing with unobserved heterogeneity in a setting with a binary dependent variable and a binary endogenous variable is to
use a bivariate probit model to simultaneously estimate the joint likelihood of the dependent variable and the endogenous regressor. We do not pursue this
option because we do not have a persuasive exclusion restriction and we are reluctant to identify the causal effect entirely through functional form. Altonji
et al. (2005) point out serious flaws with the bivariate probit model in the absence of exclusion restrictions.
Table 5
Amount of selection on unobservables relative to amount of selection on observables.
S^ 0 α^ π^ max
Subsample Control variables S^ uo ^ max d i Þ θ^
4 VarðZ
^ S uo
θ≔
S^ o
(1) (2) (3) (4) (5) (6)
Males Baseline specification 0.127 0.297 1.309 0.227 1.785 1.579

Extensive specification 0.097 0.410 1.319 0.311 3.221 2.863
Females Baseline specification 1.896 0.196 1.309 0.150 0.079 0.053
Extensive specification 0.378 0.249 1.319 0.189 0.498 0.339
Note: Baseline specification refers to baseline probit estimation from Section 3.1; extensive specification refers to baseline specification plus the full list of
additional variables described in Section 3.2. Column (1) reports amount of selection on observables, which is calculated as
E^ ½X 0 γ^ r jZ i ¼ 1;Malei ¼ t E^ ½X 0 γ^ r jZ i ¼ 0;Malei ¼ t
where γ^ r is the probit estimate of γ in Eq. (1) in which α ¼ 0 and t A f0; 1g according to respective gender. Column
c ½X 0 γ^ r j;Malei ¼ t
Var
(2) reports the simple probit estimate of α in Eq. (1). Column (3) reports the estimated variance ratio defined in Eq. (4). Column (4) reports the maximum
amount of selection on unobservables defined in Eq. (5). Column (5) reports the selection ratio given by Definition 1. Column (6) reports the lower bound
on the hypothetical selection ratio given by Definition 3. N ¼3009 (baseline specification) and N ¼2718 (extensive specification).
^
where E½j is an analog estimator (see again Manski, 1988) and γ^ r are the estimates resulting from the estimation of Eq. (1)
while imposing the restriction α ¼ 0.
The statistic proposed by Altonji et al. (2005b) is now given by the following definition.
Definition 1 (Selection ratio). The selection ratio θ^ is defined as

^ max
^ S uo :
θ≔
S^ o
The selection ratio θ^ measures how large the amount of selection on unobservables would need to be relative to the
amount of selection on observables in order to explain away the entire causal effect of academic fraud. The actual com-
putation of the selection ratio θ^ is based on the following six steps:
(i) Obtain the estimate α^ of α in Eq. (1) using standard probit estimation without any constraints imposed. (The coefficient γ
will also be estimated but it is not needed.)
(ii) Obtain the estimate π^ in Eq. (4) by replacing population with sample variances.
max
(iii) Calculate S^ uo according to Eq. (5).
(iv) Obtain the estimate γ^ r of γ in Eq. (1) using standard probit estimation imposing the constraint α ¼ 0. Calculate linear
index X 0i γ^ r .
(v) Calculate S^ o according to Eq. (6).
(vi) Calculate the selection ratio θ^ according to Definition 1.
The probit estimations in steps (i) and (iv) include the regressors from Sections 3.1 and 3.2. To be precise, we compare the
amount of selection on unobservables to the amount of selection on observables for two particular probit specifications: the
baseline specification from Section 3.1 (with the baseline set of explanatory variables) and the extensive specification from
Section 3.2 (which adds the full set of proxy variables).
Table 5 reports the results for males and females separately. Looking first at males in the baseline specification, the
amount of selection on observables under the hypothesis of no causal effect of academic fraud is estimated to equal 0.127. At
the same time, the estimate of α is equal to 0.297, which, when adjusted for the value of π^ of 1.309 implies a maximum
amount of selection on unobservables of 0.227. These numbers mean that the maximum amount of selection on unob-
servables would need to be 1.785 times the amount of selection on observables to explain away the entire causal effect of
academic fraud. Similar calculations for the extensive specification reveal that the amount of selection on unobservables
would need to be more than three times as large as the amount of selection on observables to render the causal effect
economically irrelevant. In light of the rich set of explanatory variables that are included in both the baseline and extended
models that capture important first order effects (e.g., gender, age and health status) and aim at controlling for second order
effects (through correlations with unobserved factors), selection ratios of 179% and 322% appear exceedingly large.7
The results for females are not as clearly cut. The selection ratios are calculated to equal 0.079 (baseline specification) and
0.498 (extensive specification). These numbers do not permit the same strong inferences as in the case of males. The much
lower numbers for females find their root in the relatively large amount of selection on observables vis-à-vis that of males.
7
Altonji et al. (2005b), in assessing the effectiveness of Catholic schools, obtain similar numbers and declare them “highly unlikely”.
In an attempt to get a better understanding of this dichotomy between the genders, we turn to investigating the compo-
nents of Eq. (6). We find that the dichotomy is solely explained by the difference in the estimates of the numerator in that
equation. The numerator presents the difference in sample means of the linear index X 0i γ^ r between fraudsters and non-
fraudsters. In our estimations, that difference is equal to 0.005 for males while it is equal to 0.095 for females, almost 20
times as large (numbers from the baseline specification; extensive specification yields a qualitatively similar answer). This
gap is large, considering that both numbers are then divided by a sample variance of about 0.156 (males) and 0.135
(females), amplifying the numerators considerably.
An early indication that selection on observables is larger for females is already given by Table 3.8 Studying the change in
marginal effects of academic fraud between the parsimonious and the baseline models, the difference is relatively large for
females (drop from 10.8% to 7.4%) but virtually zero for males. This sensitivity of the marginal effect to the inclusion of
baseline regressors is suggestive of a relatively large amount of selection on observables for females. At the same time, the
pseudo-R2 increases from the parsimonious specification to the baseline specification by almost 12.5%, indicating that the
inclusion of baseline regressors does contribute to explaining variation in employment outcomes.
Summing up, males do almost not differ in observable characteristics between fraudsters and non-fraudsters. The
amount of selection on unobservables would need to be unusually large to be able to explain away a causal effect of
academic fraud. For females, the observable differences between fraudsters and non-fraudsters are more pronounced. The
amount of selection on unobservables need not be nearly as large as in the case of males to render the causal effect of
academic fraud irrelevant.
5. Measurement error
5.1. Nonparametric identification
It is likely that some survey respondents do not truthfully report their academic fraud status. We do not know how many
people misreport and we do not know what type of person tends to misreport. The analysis pursued here gets by without
making any assumptions on extent and type of misreporting. In particular, persons who misreport may differ in unobserved
ways to persons who do not misreport. Such endogenous misreporting creates measurement error in the main explanatory
variable, which can lead to bias in the estimates of the causal effect of academic fraud.
A priori, it is not clear which direction this bias will point to. But the mere possibility of bias from misreporting poses a
threat to our main findings from previous sections: that the effect of academic fraud on employment is large and positive (at
least for males). Using nonparametric partial identification methods see (Manski, 1989, 2007; Tamer, 2010 among others, for
details on partial identification), we derive that, if anything, misreporting creates downward bias in the estimates of the
effect of academic fraud on employment.
Before proceeding, we need to clarify some terminology. Define a person's actual academic fraud status by the dummy
variable Z~ i , taking on the value 1 if a person committed academic fraud and 0 otherwise. In contrast, define a person's
reported academic fraud status by the dummy variable Zi, taking on the value 1 if a person reports to have committed
academic fraud and 0 otherwise. Actual fraud status is unobserved, reported fraud status is observed and generally Z~ i a Z i ,
although in many cases they would be the same. We ignore other explanatory variables in this section for notational
simplicity.
A person's employment outcome is determined by actual academic fraud; it is not determined by reported academic
fraud. Then, Y i ðZ~ i ¼ 1Þ represents the employment outcome for person i if she commits academic fraud and Y i ðZ~ i ¼ 0Þ
represents the employment outcome if she does not commit academic fraud. Ideally, we would like to learn about the
individual treatment effect
Δi ≔Y i ðZ~ i ¼ 1Þ Y i ðZ~ i ¼ 0Þ:
This effect, however, is not identified because we do not observe Z~ i and, even if we did, the missing counterfactual for
person i will always be unknown. As is conventional, we focus on the average treatment effect instead. But the average
treatment effect E½Δi is also not identified because Z~ i is unobserved.
We can neither point identify nor partially identify E½Δi without making assumptions that we are not willing to make
(e.g., random selection into misreporting). Instead, we show that we are able to provide weak conditions to partially identify
a slightly modified average treatment effect, given by
Definition 2 (Average treatment effect).
Γ≔E½Δi jZ i ¼ Z~ i ¼ E½Y i ðZ~ i ¼ 1ÞjZ i ¼ Z~ i E½Y i ðZ~ i ¼ 0ÞjZ i ¼ Z~ i :
8
We thank an anonymous referee for pointing this out.
This is the average treatment effect for the subset of people who honestly report fraud status. While this constitutes a
succinctly defined subpopulation, it is also an unobserved group of people (because Z~ i is unobserved). This subpopulation of
honest respondents, however, is likely to be large; it comprises both non-fraudsters who report that they did not commit
fraud as well as fraudsters who report that they did commit fraud.
Using the average treatment effect Γ sidesteps the problem of the unobserved misreporting incidence: we do not know
how many people report untruthfully and we do not need to because we are only interested in the average treatment effect
for honest respondents. (Of course, we would prefer to know about the effect for dishonest respondents too, but this is
unattainable.) The practical relevance of the average treatment effect Γ rises with the fraction of honest respondents in the
population. The only drawback is that we do not know this fraction and need to assume that it is sufficiently large.
Before we can establish a lower bound for Γ we need to impose three conditions.
Condition 1. Mean outcomes are non-decreasing in the treatment conditional on Z~ i ¼ 1:
E½Y i ðZ~ i ¼ 0ÞjZ~ i ¼ 1 r E½Y i ðZ~ i ¼ 1ÞjZ~ i ¼ 1:
Condition 1 states that a person's expected employment status weakly benefits from academic fraud. Both expectations
are conditional on Z~ i ¼ 1, the subpopulation that does commit academic fraud. Given Z~ i ¼ 1, the right hand side is the mean
(factual) labor market outcome in the presence of academic fraud. The left hand side is the mean (counterfactual) labor
market outcome had person i not committed academic fraud instead. Condition 1 states that, in expectation, this coun-
terfactual outcome cannot be greater than the factual one. The condition is justified to the extent that academic fraud
improves a person's grades and, in turn, expected labor market outcomes.
Condition 2. People who do not commit academic fraud report truthfully: Z~ i ¼ 0 ) Z i ¼ 0:
This condition states that misreporting is asymmetric: only people who commit academic fraud will lie about it; people
who do not commit fraud have no reason to lie about their academic fraud status.
Condition 3. For p A f0; 1g, PrðZ~ i ¼ 1jY i ðZ~ i ¼ pÞ ¼ 1; Z i ¼ 0Þ ZPrðZ~ i ¼ 1jZ i ¼ 0Þ.
Both probabilities are conditional on Zi ¼0, the subpopulation that does not admit to academic fraud (they do not report
it). Given Zi ¼0, the right hand side is the probability of being a fraudster. In general, this probability will not be equal to one
because of misreporting. The left hand side is that same probability with the extra condition that Y i ðZ~ i ¼ pÞ ¼ 1. Condition 3
therefore states that among persons who do not admit to academic fraud, those who have above average labor market
outcomes are weakly more likely to have actually committed academic fraud than the average person who does not admit to
academic fraud. This condition is weak and has one principal justification: persons with above average labor market out-
comes have higher opportunity costs of reporting truthfully.
The main result is now given by the following proposition.
Proposition 1 (Partial identification of ATE). Under Conditions 1, 2 and 3, Γ ZE½Y i jZ i ¼ 1 E½Y i jZ i ¼ 0.
The proposition provides a partial identification result: it offers a lower bound on the average treatment effect of interest,
Γ. (Proof of Proposition 1 is in Appendix.) Intuitively, the lower bound is given by the difference in population averages of
employment outcomes between reported fraudsters and reported non-fraudsters. That difference in population averages, and
therefore the lower bound, can readily be estimated. In fact, in Sections 2 and 3 we already have provided two different
approaches to estimating the lower bound.
To see this, recall that Table 2 reports sample means of employment outcomes by academic fraud status and gender. The
difference in these sample means between fraudsters and non-fraudsters, of course, can be viewed as nonparametric lower
bound estimates controlling for gender. Comparing fraudsters to non-fraudsters (columns 5 and 6), the lower bound on the
right hand side of Proposition 1 is estimated to equal 0.08 for males and 0.11 for females. In the same vein, the results in
Table 4 can be viewed as parametric lower bound estimates. Under the probit model, the right hand side of Proposition 1 is
equal to the probit marginal effects; these are reported in the table. Using the parametric probit model instead, the lower
bound is estimated to equal 0.077 for males and 0.074 for females (baseline specification) and 0.116 for males and 0.092 for
females (extensive specification).
In closing, both nonparametric and parametric bound estimates suggest that the true causal effect of academic fraud
under misreporting remains large and positive. While not the only source of endogeneity in the estimations, misreporting of
academic fraud in itself does not pose a threat to the positive causal interpretation of the probit estimates.
5.2. Lower bound on the selection ratio
We now study the problems of selection on unobservables and measurement error simultaneously. In Section 4 we
calculate the selection ratio θ^ of Definition 1 ignoring the possibility that people may misreport their academic fraud status.
To remove the influence of misreporting, we investigate the following hypothetical: what would the selection ratio have
been equal to if we had observed actual academic fraud status?
In what follows, we show that this hypothetical selection ratio is bounded below by a modified version of the selection
^ We can therefore bound the hypothetical selection ratio for the subset of honest respondents by the feasible
ratio θ.
selection ratio that we have obtained earlier. We proceed by defining the hypothetical selection ratio.
Definition 3 (Hypothetical selection ratio). Let θ~ be the selection ratio for the subpopulation with Z i ¼ Z~ i .
A rigorous mathematical definition of the hypothetical selection ratio θ~ is part of the proof of Proposition 2 in Appendix.
The hypothetical selection ratio θ~ is infeasible because it presumes the availability of actual academic fraud status Z~ i when,
in practice, only reported academic fraud status Zi is observed. We now establish how the feasible selection ratio θ^ from
Definition 1 can be used to bound θ~ from below. We need one additional condition.
Condition 4.

PrðZ i ¼ 0Þ E½X 0i γ r jZ i ¼ 1; Z~ i ¼ 1 E½X 0i γ r jZ i ¼ 0; Z~ i ¼ 0 ZE½X 0i γ r jZ i ¼ 0; Z~ i ¼ 1 E½X 0i γ r jZ i ¼ 0; Z~ i ¼ 0:
Condition 4 is a weak restriction of selection on observables. Presume, for the sake of illustration, that PrðZ i ¼ 0Þ ¼ 1. Then
the condition states that dishonest fraudsters (i.e., those with Z i ¼ 0; Z~ i ¼ 1) do not select more on observables than honest
fraudsters (i.e., those with Z i ¼ 1; Z~ i ¼ 1). Intuitively, fraudsters who report non-truthfully do not have better observables
than fraudsters who report truthfully.
There exists empirical evidence in support of Condition 4. As part of the calculations for Table 5 in Section 4 we found
that selection on observables is positive (column (1) of the table). Positive selection on observables, as it turns out, is
equivalent to the following inequality:

PrðZ i ¼ 0Þ E½X 0i γ r jZ i ¼ 1; Z~ i ¼ 1 E½X 0i γ r jZ i ¼ 0; Z~ i ¼ 0
ZPrðZ i a Z~ i Þ E½X 0i γ r jZ i ¼ 0; Z~ i ¼ 1 E½X 0i γ r jZ i ¼ 0; Z~ i ¼ 0: ð7Þ
This inequality looks similar to Condition 4. While it is not sufficient for Condition 4, inequality (7) is nevertheless
necessary. The empirical observation of positive selection on observables therefore does lend some support to Condition 4.
(The equivalency between positive selection on observables and inequality (7) is established as part of the proof of
Proposition 2.)
The main partial identification result for θ~ is given by the next proposition.

~ Under Conditions 2 and 4, plim θ~ r 4 VarðZ i Þ plim θ^ .
Proposition 2 (Lower bound for θ).
~
The practical implication of Proposition 2 is that it suggests a feasible lower bound on the hypothetical selection ratio θ.
d i Þ θ^ and is reported in column (6) of Table 5. (Proof of Proposition 2 is in
That lower bound can be calculated by 4 VarðZ
Appendix.)
A corollary to Proposition 2 follows immediately:
d i Þ θ^ r θ.
Corollary 1. 4 VarðZ ^
This corollary shows that the numerical values from column (6) must be weakly smaller than the ones from column (5).
Corollary 1 has important implications for what can be learned here, over and above the results from Section 4, about
selection on unobservables. For females, the selection ratio θ^ reported in Section 4 is too low to make strong inferences. By
Corollary 1, the lower bound on the hypothetical selection ratio θ~ is even lower. We therefore are still not able to make
strong inferences for females.
For males, on the other hand, the lower bounds on θ~ remain informative. In fact, the results for the lower bound on θ~ still
support the key findings of Section 4. For males, the lower bounds are 1.579 (baseline specification) and 2.863 (extensive
model). With these numbers we still conclude that selection on unobservables would need to be exceedingly large to
explain away the entire causal effect of academic fraud. This is now shown to be true irrespective of whether one has
available data on reported or actual academic fraud.
6. Discussion
If the causal effect of academic fraud is positive, why does not everyone commit it? We offer two explanations: het-
erogeneity in risk aversion and heterogeneity in preferences for honesty. People could be risk averse. Note that our research
points toward a positive causal effect of academic fraud. Simply put, any person at university has the choice between the
safe option of not committing academic fraud and the costly and risky option of committing it. The safe option results—at
most—in a return of zero while the risky option comes with a strictly positive gain. Although the costs seem small, they are
positive and should be thought of as the outcome of a random draw with potentially severe consequences (for example, the
event of getting caught). Risk averse and risk neutral persons will therefore not always commit academic fraud simply
because the expected gains appear large. (Having said that, depending on the perceived costs, a risk averse person may, on
balancing perceived risk and return, decide to commit academic fraud.)
People could also have intrinsic preferences for honesty. Although classical economics likes to use the theoretical con-
struct of homo oeconomicus who, as Gneezy (2005) puts it, “acts selfishly and is unconcerned about the well-being of others”
and who would certainly not hesitate to commit academic fraud if the expected gain was positive, modern economics—
theoretical and empirical—considers the possibility that people are honest. People are honest for various reasons, not all of
them noble (e.g., Gibson et al., 2013; Fischbacher and Heusi, 2013; Pruckner and Sausgruber, 2013; Kartik, 2009; Vanberg,
2008; Bénabou and Tirole, 2006; Levitt, 2006; Haan and Kooreman, 2002; Dufwenberg and Gneezy, 2000). If honesty is
valued in itself then even people who otherwise are at the margin of committing academic fraud may choose not to do so
(this includes people who are risk loving or risk neutral).
7. Conclusion
Academic fraud is ubiquitous and, based on our findings, it does constitute a serious economic problem. The lessons from
the Caucasus are: academic fraud has pervaded incentives, distorts the allocation of human capital and gives fraudsters a
substantial unfair advantage. Our findings point towards the potentially damaging consequences of academic fraud in
broader settings. Recent events at academic institutions in more developed countries highlight the urgency of the problem
and the results presented here offer, at the very least, a note of caution.
The external validity of our results is, of course, limited. The particular manifestation of academic fraud in our setting
(paying of bribes and giving of gifts) are rather uncommon in developed countries where cheating and plagiarism are more
frequently encountered forms of academic fraud by students. Moreover, the systemic nature of corruption along with the
lack of formal control mechanisms in the region of our study means that this particular type of academic fraud is highly
prevalent. Nevertheless, to the extent that academic fraud in its various manifestations serves the same purpose (namely, to
jump the queue and mimic a better type) we would expect the findings of our work to be applicable to a wider range of
situations and settings.
Our paper is a first attempt at investigating the association between academic fraud and employment; many research
opportunities still exist and important and interesting research questions remain unanswered, among those: How risky is
academic fraud? How expensive is it and what can be ‘purchased’? Who chooses to commit academic fraud? What is the
effect of academic fraud on grades and on wages?
Acknowledgments
We are grateful to Gaurab Aryal, Yijuan Chen, Ivan Cherkashin, Dean Hyslop, Brian McCaig, Martine Mariotti and Yuji
Tamura for valuable comments. We thank the associate editor and two anonymous referees for improving the paper sub-
stantially. Seminar participants at the ANU Research School of Economics, the Social Policy Research Centre at the University
of New South Wales, the Melbourne Institute at the University of Melbourne, the Department of Economics at the University
of Auckland, the Department of Economics at Stellenbosch University, the Department of Economics at the University of
Pretoria and the 2014 Labour Econometrics Workshop at the University of Otago have provided helpful feedback. Lastly, we
are thankful to Giorgi Babunashvili and Nana Papiashvili from the Caucasus Research Resource Centers for technical
assistance. All errors are those of the authors.
Appendix A. Proofs
Proof of Proposition 1. By definition

Γ≔E½Y i ðZ~ i ¼ 1ÞjZ i ¼ Z~ i E½Y i ðZ~ i ¼ 0ÞjZ i ¼ Z~ i : ðA:1Þ
The first term on the right hand side can be broken down as follows:
E½Y i ðZ~ i ¼ 1ÞjZ i ¼ Z~ i ¼ E½Y i ðZ~ i ¼ 1ÞjZ i ¼ Z~ i ; Z~ i ¼ 1
¼ E½Y i ðZ~ i ¼ 1ÞjZ i ¼ 1; Z~ i ¼ 1
¼ E½Y i ðZ i ¼ 1ÞjZ i ¼ 1
¼ E½Y i jZ i ¼ 1;
where the first equality holds because outcomes are mean independent of actual fraud status (conditional on Z i ¼ Z~ i ), the
second equality holds by transitivity, the third equality follows from Condition 2 and the last equality results by definition.
The conditional expectation E½Y i ðZ~ i ¼ 1ÞjZ i ¼ Z~ i is therefore point identified.
Regarding the second term on the right hand side of Eq. (A.1), we obtain
E½Y i ðZ~ i ¼ 0ÞjZ i ¼ Z~ i
¼ E½Y i ðZ~ i ¼ 0ÞjZ i ¼ Z~ i ; Z~ i ¼ 0
¼ E½Y i ðZ~ i ¼ 0ÞjZ i ¼ 0; Z~ i ¼ 0

¼ ψ E½Y i ðZ~ i ¼ 0ÞjZ i ¼ 0; Z~ i ¼ 0 þ ð1 ψÞ E½Y i ðZ~ i ¼ 0ÞjZ i ¼ 0; Z~ i ¼ 0
r ψ E½Y i ðZ~ i ¼ 0ÞjZ i ¼ 0; Z~ i ¼ 0 þ ð1 ψÞ E½Y i ½Z~ i ¼ 0jZ i ¼ 0; Z~ i ¼ 1
r ψ E½Y i ðZ~ i ¼ 0ÞjZ i ¼ 0; Z~ i ¼ 0 þ ð1 ψÞ E½Y i ðZ~ i ¼ 1ÞjZ i ¼ 0; Z~ i ¼ 1
¼ ψ E½Y i jZ i ¼ 0; Z~ i ¼ 0 þ ð1 ψÞ E½Y i jZ i ¼ 0; Z~ i ¼ 1
¼ E½Y i jZ i ¼ 0;
where ψ≔PrðZ~ i ¼ 0jZ i ¼ 0Þ. The first equality holds because outcomes are mean independent of actual fraud status (con-
ditional on Z i ¼ Z~ i ), the second equation holds by transitivity, the first inequality is based on Lemma 1 (below), the second
inequality is based on Condition 1, the penultimate equality holds by definition and the final equality results from the law of
total probability.
Combining, the average treatment effect Γ from Eq. (A.1) is partially identified by
Γ Z E½Y i jZ i ¼ 1 E½Y i jZ i ¼ 0:
Lemma 1. Under Condition 3, E½Y i ðZ~ i ¼ pÞjZ i ¼ 0; Z~ i ¼ 0 r E½Y i ðZ~ i ¼ pÞjZ i ¼ 0; Z~ i ¼ 1 for p A f0; 1g.
Proof of Lemma 1. Under Condition 3 we have, for p A f0; 1g,

PrðZ~ i ¼ 1jY i ðZ~ i ¼ pÞ ¼ 1; Z i ¼ 0Þ
1r
PrðZ~ i ¼ 1jZ i ¼ 0Þ
!

1

¼ Pr Z~ i ¼ 1
Y i Z~ i ¼ p ¼ 1; Z i ¼ 0 1þ1
!

1 PrðZ~ ¼ 1jZ ¼ 0Þ
~
~ i i
¼ Pr Z i ¼ 1
Y i Z i ¼ p ¼ 1; Z i ¼ 0 þ1 :
It follows that
!

1 PrðZ~ ¼ 1jZ ¼ 0Þ
~
~ i i
Pr Z i ¼ 1
Y i Z i ¼ p ¼ 1; Z i ¼ 0
Z 1 PrðZ i ¼ 1jY i ðZ~ i ¼ pÞ ¼ 1; Z i ¼ 0Þ;
~
which can be rewritten as

PrðZ~ i ¼ 1jY i ðZ~ i ¼ pÞ ¼ 1; Z i ¼ 0Þ 1 PrðZ~ i ¼ 1jY i ðZ~ i ¼ pÞ ¼ 1; Z i ¼ 0Þ
Z ;
PrðZ~ i ¼ 1jZ i ¼ 0Þ 1 PrðZ~ i ¼ 1jZ i ¼ 0Þ
or equivalently as
PrðZ~ i ¼ 1jY i ðZ~ i ¼ pÞ ¼ 1; Z i ¼ 0Þ PrðZ~ i ¼ 0jY i ðZ~ i ¼ pÞ ¼ 1; Z i ¼ 0Þ
Z :
PrðZ~ i ¼ 1jZ i ¼ 0Þ PrðZ~ i ¼ 0jZ i ¼ 0Þ
Multiplying by PrðY i ðZ~ i ¼ pÞ ¼ 1jZ i ¼ 0Þ on both sides results in

PrðY ðZ~ ¼ pÞ ¼ 1jZ ¼ 0Þ

Pr Z~ i ¼ 1
Y i Z~ i ¼ p ¼ 1; Z i ¼ 0 i i i

PrðY ðZ~ ¼ pÞ ¼ 1jZ ¼ 0Þ

Z Pr Z~ i ¼ 0
Y i Z~ i ¼ p ¼ 1; Z i ¼ 0 i i i
:
The last inequality can be expressed as
PrðY i ðZ~ i ¼ pÞ ¼ 1jZ i ¼ 0; Z~ i ¼ 1Þ Z PrðY i ðZ~ i ¼ pÞ ¼ 1jZ i ¼ 0; Z~ i ¼ 0Þ;
which, when written in expectations, gives
E½Y i ðZ~ i ¼ pÞjZ i ¼ 0; Z~ i ¼ 1 ZE½Y i ðZ~ i ¼ pÞjZ i ¼ 0; Z~ i ¼ 0:
□
Proof of Proposition 2. We start by mathematically defining θ~ given in Definition 3. To do so, we first set out the analogs to
Eq. (1) through (6) for the case in which Z i ¼ Z~ i . These equations are given by
n ~ o
Employedi ¼ 1 αZ i þX 0i γ þwi 40 ðA:2Þ
S~ uo ≔E½wi jZ i ¼ 1; Z~ i ¼ 1 E½wi jZ i ¼ 0; Z~ i ¼ 0 ðA:3Þ
plimðα~ Þ αþ ϕS~ uo ðA:4Þ


ϕ≔VarðZ~ i jZ i ¼ Z~ i Þ=VarðZ~ i jZ i ¼ Z~ i Þ ðA:5Þ
1
S~ uo ≔ϕ~ α~
max
ðA:6Þ
h i h i
E^ X 0i γ^ r jZ i ¼ 1; Z~ i ¼ 1 E^ X 0i γ^ r jZ i ¼ 0; Z~ i ¼ 0
~S o ≔ : ðA:7Þ
d X 0 γ^ r
Var i
Notice that the two error terms wi (from Eq. (A.2)) and ui (from Eq. (1)) are related: ui ≔αðZ~ i Z i Þ þ wi . Furthermore, α~ is

the probit estimate of α in Eq. (A.2) for the subpopulation with Z i ¼ Z~ i ; Z~ i is the error in a regression of Z~ i on Xi; ϕ~ is the
estimate of ϕ using sample variances in place of population variances; and γ^ r is the probit coefficient estimate of γ imposing
the restriction α ¼ 0. The hypothetical selection ratio is defined by
~ max
~ S uo :
θ≔
S~ o
The overarching goal is to bound the probability limit of θ~ by that of θ^ from below. We begin by looking at the probability
limit of S~ uo under the null hypothesis that α ¼ 0 (which is the ‘worst case’ scenario for selection on unobservables).
max
max 1 1
plim S~ uo ¼ plim ϕ~ α~ ¼ plim ϕ~ plimðα~ Þ
h i h i
~ ~ E Z i wi jZ i ¼ Z~ i
~
1 E Z i wi jZ i ¼ Z i
¼
ϕ Var Z~ jZ ¼ Z~ Var Z i jZ i ¼ Z~ i
i i i
h i
E Z i wi jZ i ¼ Z~ i
¼ ; ðA:8Þ
Var Z i jZ i ¼ Z~ i
where the approximate equality results from a direct application of the omitted variables bias formula (in the linear model)
and permitting the same approximation as Altonji et al. (2005b); the final equality is based on Lemma 2 below. For S~ o we get
h i h i
E X 0i γ r jZ i ¼ 1; Z~ i ¼ 1 E X 0i γ r jZ i ¼ 0; Z~ i ¼ 0
plim S~ o ¼ : ðA:9Þ
VarðX 0i γ r Þ
Turning now to the components of θ, ^ we first study the probability limit of S^ max .
uo
max
^ 1
plim S uo ¼ plim π^ α^ ¼ plim π^ 1
plimðα^ Þ

1 E Z i ui E Z i ui
¼ ; ðA:10Þ
π Var Z i VarðZ i Þ
where again the approximate equality results from a direct application of the omitted variables bias formula and permitting
the same approximation as Altonji et al. (2005b). Taking a closer look at the covariance between Z i and ui, we get
E½Z i ui ¼ E½Z i wi ¼ E½Z i wi
¼ PrðZ i ¼ Z~ i Þ E½Z i wi jZ i ¼ Z~ i þ PrðZ i a Z~ i Þ E½Z i wi jZ i a Z~ i
¼ PrðZ i ¼ Z~ i Þ E½Z i wi jZ i ¼ Z~ i þ PrðZ i a Z~ i Þ E½Z i wi jZ i ¼ 0; Z~ i ¼ 1
¼ PrðZ i ¼ Z~ i Þ E½Z i wi jZ i ¼ Z~ i ; ðA:11Þ
where the first equality holds under the null hypothesis that α ¼ 0 (so that ui ¼ wi) and the second equality again follows by
Lemma 2 below. Multiplying by 4 VarðZ i Þ and combining Eqs. (A9) and (A10), we obtain
max
4 VarðZ i Þ plim S^ uo 4 PrðZ i ¼ Z~ i Þ E½Z i wi jZ i ¼ Z~ i
max
¼ 4 PrðZ i ¼ Z~ i Þ VarðZ i jZ i ¼ Z~ i Þ plim S~ uo
max
rPrðZ i ¼ Z~ i Þ plim S~ uo ; ðA:12Þ
where the equality holds by Eq. (A.8) and the inequality holds because VarðZ i jZ i ¼ Z~ i Þ r0:25 (the variance of a Bernoulli
random variable is bounded above by 0.25).
^ rplim θ~ it is sufficient that
To establish that 4 VarðZ i Þ plimðθÞ
max max
PrðZ i ¼ Z~ i Þ plim S~ uo plim S~ uo
r ≕plim θ~ ;
plim S^ o plim S~ o
^ That
The sufficiency follows because, by inequality (A.12), the left hand side is an upper bound on 4 VarðZ i Þ plimðθÞ.
sufficient condition is equivalent to

plim S^ o Z PrðZ i ¼ Z~ i Þ plim S~ o : ðA:13Þ
For the left hand side we get

1
plim S^ o ¼ E½X 0i γ r jZ i ¼ 1 E½X 0i γ r jZ i ¼ 0
VarðX 0i γ r Þ
1
¼ 0 r ~
0 r E½X i γ jZ i ¼ 1; Z i ¼ 1
VarðX i γ Þ

κ E½X 0i γ r jZ i ¼ 0; Z~ i ¼ 1 ð1 κÞ E½X 0i γ r jZ i ¼ 0; Z~ i ¼ 0
1
¼ 0 r ~ 0 r ~
0 r E½X i γ jZ i ¼ 1; Z i ¼ 1 E½X i γ jZ i ¼ 0; Z i ¼ 0
VarðX i γ Þ

þ κ E½X 0i γ r jZ i ¼ 0; Z~ i ¼ 0 E½X 0i γ r jZ i ¼ 0; Z~ i ¼ 1

¼ plim S~ o
κ
þ 0 r ~ 0 r ~
0 r E½X i γ jZ i ¼ 0; Z i ¼ 0 E½X i γ jZ i ¼ 0; Z i ¼ 1
VarðX i γ Þ

¼ Pr Z i ¼ Z~ i plim S~ o

þ 1 PrðZ i ¼ Z~ i Þ plim S~ o
κ
þ 0 r ~ 0 r ~
0 r E½X i γ jZ i ¼ 0; Z i ¼ 0 E½X i γ jZ i ¼ 0; Z i ¼ 1 ðA:14Þ
VarðX i γ Þ
where κ≔PrðZ~ i ¼ 1jZ i ¼ 0Þ and the second equality holds because, by Condition 2, PrðZ~ i ¼ 1jZ i ¼ 1Þ ¼ 1 and the fourth
equality holds by Eq. (A.9).
Given the result in Eq. (A.14), the sufficient condition in inequality (A.13) translates to

1 PrðZ i ¼ Z~ i Þ plim S~ o
κ
þ E½X 0i γ r jZ i ¼ 0; Z~ i ¼ 0 E½X 0i γ r jZ i ¼ 0; Z~ i ¼ 1 Z0:
VarðX 0i γ r Þ
Realizing that ð1 PrðZ i ¼ Z~ i ÞÞ=κ ¼ PrðZ i ¼ 0Þ and plugging in Eq. (A.9) we obtain

ZE½X 0i γ r jZ i ¼ 0; Z~ i ¼ 1 E½X 0i γ r jZ i ¼ 0; Z~ i ¼ 0;
which is Condition 4.
In the main text, in the context of discussing Condition 4, we claim that positive selection on observables is equivalent to

ZPrðZ i a Z~ i Þ E½X 0i γ r jZ i ¼ 0; Z~ i ¼ 1 E½X 0i γ r jZ i ¼ 0; Z~ i ¼ 0:
This follows immediately by combining Eq. (A.9) with Eq. (A.14) and recalling that κ≔PrðZ~ i ¼ 1jZ i ¼ 0Þ ¼
PrðZ i ¼ 0; Z~ i ¼ 1Þ=PrðZ i ¼ 0Þ ¼ PrðZ i a Z~ i Þ=PrðZ i ¼ 0Þ which again follows from Condition 2.

Lemma 2. E½Z i wi ¼ E½Z i wi and E½Z~ i wi ¼ E½Z~ i wi .
Proof of Lemma 2. Z i is the error term in the regression of Zi on Xi. Denote that regression by Z i ¼ τX i þZ i . The omission of a
constant term is without loss of generality. Then E½Z i wi ¼ E½ðZ i τX i Þwi ¼ E½Z i wi τE½X i wi ¼ E½Z i wi τE½X i E½wi jX i ¼ E½Z i wi

because E½wi jX i ¼ 0 by exogeneity of Xi. The proof for Z~ i follows analogously. □
Appendix B. Supplementary data
Supplementary data associated with this paper can be found in the online version at http://dx.doi.org/10.1016/j.euro
ecorev.2015.11.005.
References
Akerlof, G.A., 1970. The market for ‘Lemons’: quality uncertainty and the market mechanism. Q. J. Econ. 84, 488–500.
Aleksanyan H., 2012. Cases of Corruption and its Prevention in Armenia's Education System. Norwegian Institute of International Affairs.
Altonji, J.G., Elder, T.E., Taber, C.R., 2005. An evaluation of instrumental variable strategies for estimating the effects of catholic schooling. J. Hum. Resour. 40,
791–821.
Altonji, J.G., Elder, T.E., Taber, C.R., 2005b. Selection on observed and unobserved variables: assessing the effectiveness of catholic schools. J. Polit. Econ. 113,
151–184.
Arrow, K.J., 1973. Higher education as a filter. J. Public Econ. 2, 193–216.
Banerjee, A., Hanna, R., Mullainathan, S., 2013. Corruption. In: Gibbons, R., Roberts, J. (Eds.), The Handbook of Organizational Economics, Princeton Uni-
versity Press, New Jersey.
Barón, J.D., Cobb-Clark, D., 2010. Are Young People's Educational Outcomes Linked to their Sense of Control? IZA Discussion Paper, 4907.
Becker, G.S., 1964. Human Capital: A Theoretical and Empirical Analysis, With Special Reference to Education. Columbia University Press, New York.
Bénabou, R., Tirole, J., 2006. Incentives and prosocial behavior. Am. Econ. Rev. 96, 1652–1678.
Borcan, O., Lindahl, M., Mitrut, A., 2014. The impact of an unexpected wage cut on corruption: evidence from a “Xeroxed” exam. J. Public Econ. 120, 32–47.
Bowers, W.J., 1964. Student Dishonesty and Its Control in College. Bureau of Applied Social Research, Columbia University, New York.
Bunn, D.N., Caudill, S.B., Gropper, D.M., 1992. Crime in the classroom: an economic analysis of undergraduate student cheating behavior. J. Econ. Educ. 23,
197–207.
Burrus, R.T., McGoldrick, K., Schuhmann, P.W., 2007. Self-reports of student cheating: does a definition of cheating matter?. J. Econ. Educ. 38, 3–16.
Caliendo, M., Cobb-Clark, D., Uhlendorff, A., 2010. Locus of Control and Job Search Strategies. IZA Discussion Paper, 4750.
Carrell, S.E., Malmstrom, F.V., West, J.E., 2008. Peer effects in academic cheating. J. Hum. Resour. 43, 173–207.
Cebi, M., 2007. Locus of control and human capital investment revisited. J. Hum. Resour. 42, 919–932.
Charles, R., 2010. Religiosity and trust in religious: institutions tales from the South Caucasus (Armenia, Azerbaijan, and Georgia). Polit. Relig. 3, 228–261.
Cobb-Clark, D.A., Schurer, S., 2013. Two economists musings on the stability of locus of control. Econ. J. 123, F358–F400.
Coleman, M., DeLeire, T., 2003. an economic model of locus of control and the human capital investment decision. J. Hum. Resour. 38, 701–721.
CRRC, 2007. Data Initiative 2007 – Unified Sampling Methodology, Tech. Rep., Caucasus Research Resource Centres.
CRRC,2008. Data Initiative – 2008, Fieldwork Report, Caucasus Research Resource Centres.
CRRC, 2013. Armenia Corruption Surveys 2008–2010. 〈http://www.crrc.am/research-and-surveys/completed-projects/armenia-corruption-surveys-2008-
2010?lang ¼en.〉.
Davis, S.F., Grover, C.A., Becker, A.H., McGregor, L.N., 1992. Academic dishonesty: prevalence, determinants, techniques, and punishments. Teach. Psychol.
19, 16–20.
Diekhoff, G.M., LaBeff, E.E., Clark, R.E., Williams, L.E., Francis, B., Haines, V.J., 1996. College cheating: ten years later. Res. Higher Educ. 37, 487–502.
Dreher, A., Gassebner, M., 2011. Greasing the wheels? The impact of regulations and corruption on firm entry. Public Choice, 1–20.
Dufwenberg, M., Gneezy, U., 2000. Measuring beliefs in an experimental lost wallet game. Games Econ. Behav. 30, 163–182.
Fischbacher, U., Heusi, F., 2013. Lies in disguise–an experimental study on cheating. J. Eur. Econ. Assoc. 11, 525–547.
Fisman, R., Svensson, J., 2007. Are corruption and taxation really harmful to growth? Firm level evidence. J. Dev. Econ. 83, 63–75.
Gallant, T.B., Drinan, P., 2006. Organizational theory and student cheating: explanation, responses, and strategies. J. High. Educ. 77, 839–860.
Gibson, R., Tanner, C., Wagner, A.F., 2013. Preferences for truthfulness: heterogeneity among and within individuals. Am. Econ. Rev. 103, 532–548.
Gneezy, U., 2005. Deception: the role of consequences. Am. Econ. Rev. 95, 384–394.
Haan, M., Kooreman, P., 2002. Free riding and the provision of candy bars. J. Public Econ. 82, 277–291.
Habibov, N., 2012. Income inequality and its driving forces in transitional countries: evidence from Armenia, Azerbaijan and Georgia. J. Comp. Soc. Welf. 28,
209–221.
Habibov, N.N., 2011. Self-perceived social stratification in low-income transitional countries: examining the multi-country survey in Armenia, Azerbaijan,
and Georgia. Int. J. Soc. Econ. 38, 5–22.
Habibov, N.N., Afandi, E., 2009. Analysis of subjective wellbeing in low-income transitional countries: evidence from comparative national surveys in
Armenia, Azerbaijan and Georgia. J. Comp. Soc. Welf. 25, 203–219.
Habibov, N.N., Afandi, E.N., 2011. Self-rated health and social capital in transitional countries: multilevel analysis of comparative surveys in Armenia,
Azerbaijan, and Georgia. Soc. Sci. Med. 72, 1193–1204.
Heyneman, S.P., 2004. Education and corruption. Int. J. Educ. Dev. 24, 637–648.
Heyneman, S.P., Anderson, K.H., Nuraliyeva, N., 2008. The cost of corruption in higher education. Comp. Educ. Rev. 52, 1–25.
Hunt, J., 2007. How corruption hits people when they are down. J. Dev. Econ. 84, 574–589.
Jacob, B.A., Levitt, S.D., 2003. Rotten apples: an investigation of the prevalence and predictors of teacher cheating. Q. J. Econ. 118, 843–877.
Janashia, N., 2004. Corruption and higher education in Georgia. Int. High. Educ. 34, 10–120.
Kartik, N., 2009. Strategic communication with lying costs. Rev. Econ. Stud. 76, 1359–1395.
Kaufman, D., Wei, S.-J., 1999. Does Grease Money Speed up the Wheels of Commerce? National Bureau of Economic Research Working Paper, 7093.
Lazear, E.P., 2006. speeding terrorism, and teaching to the test. Q. J. Econ. 121, 1029–1061.
Levitt, S.D., 2006. White-collar crime writ small: a case study of bagels, donuts, and the honor system. Am. Econ. Rev. 96, 290–294.
Manski, C.F., 1988. Analog Estimation Methods in Econometrics. Chapman and Hall, London.
Manski, 1989. Anatomy of the selection problem. J. Hum. Resour. 24, 343–360.
Manski, 2007. Identification for Prediction and Decision. Harvard University Press, Cambridge, Massachusetts.
McCabe, D.L., Klebe Trevino, L., 1993. Academic dishonesty: honor codes and other contextual influences. J. High. Educ. 64, 522–538.
McCabe and Klebe Trevino, 1997. Individual and contextual influences on academic dishonesty: a multicampus investigation. Res. High. Educ. 38, 379–396.
McCabe, D.L., Klebe Trevino, L., Butterfield, K.D., 2001. Cheating in academic institutions: a decade of research. Ethics Behav. 11, 219–232.
Meon, P.-G., Weill, L., 2010. Is corruption an efficient grease? World Dev. 38, 244–259.
Mincer, J., 1958. Investment in human capital and personal income distribution. J. Polit. Econ. 66, 281–302.
Murdock, T.B., Anderman, E.M., 2006. Motivational perspectives on student cheating: toward an integrated model of academic dishonesty. Educ. Psychol.
41, 129–145.
Newstead, S.E., Franklyn-Stokes, A., Armstead, P., 1996. Individual differences in student cheating. J. Educ. Psychol. 88, 229–241.
Olken, B.A., 2009. Corruption perceptions vs. corruption reality. J. Public Econ. 93, 950–964.
Olken, B.A., Pande, R., 2012. Corruption in developing countries. Annu. Rev. Econ. 4, 479–509.
OSCE, 2010. Student Perception on Corruption in the Armenian Higher Education System.
Osipian, A.L., 2007. Corruption in higher education conceptual approaches and measurement techniques. Res. Comp. Int. Educ. 2, 313–332.
Osipian, 2009. Corruption Hierarchies in Higher Education in the Former Soviet Bloc. Int. J. Educ. Dev. 29, 321–330.
Pruckner, G.J., Sausgruber, R., 2013. Honesty on the streets: a field study on newspaper purchasing. J. Eur. Econ. Assoc. 11, 661–680.
Riley, J.G., 2001. Silver signals: twenty-five years of screening and signaling. J. Econ. Lit. 39, 432–478.
Roberts, K., Pollock, G., 2011. Class in the South Caucasus. Eur. Soc. 13, 757–782.
Sims, R.L., 1993. The relationship between academic dishonesty and unethical business practices. J. Educ. Bus. 68, 207–211.
Spence, M., 1973. Job market signaling. Q. J. Econ. 87, 355–374.
Stiglitz, J.E., 1975. The theory of ‘screening,’ education, and the distribution of income. Am. Econ. Rev. 65, 283–300.
Svensson, J., 2003. Who must pay bribes and how much? Evidence from a cross section of firms. Q. J. Econ. 118, 207–230.
Svensson, 2005. Eight questions about corruption. J. Econ. Perspect. 19, 19–42.
Tamer, E., 2010. Partial identification in econometrics. Annu. Rev. Econ., 167–195.
Temple, P., Petrov, G., 2004. Corruption in higher education: some findings from the states of the former Soviet Union. High. Educ. Manag. Policy 16, 83–99.
Transparency International, 2013. Global Corruption Report: Education.

Vanberg, C., 2008. Why do people keep their promises? An experimental test of two explanations. Econometrica 76, 1467–1480.
Whitley, B.E., 1998. Factors associated with cheating among college students: a review. Res. High. Educ. 39, 235–274.

1 s2.0 S0014292115001828 Main

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

1 s2.0 S0014292115001828 Main

Uploaded by

Copyright:

Available Formats

European Economic Review 82 (2016) 212–230

Contents lists available at ScienceDirect

European Economic Review

The labor market return to academic fraud

2. Data and descriptive statistics

Variable Deﬁnition of variable Mean

IN LABOR FORCE 0–1 binary variable; equals 1 if 0.69

MALE 0–1 binary variable; equals 1 if 0.50

Note: Means representative of population; standard deviations in parentheses. N ¼ 3009.

3.1. Baseline speciﬁcation

Gender FRAUD IN LABOR FORCE EMPLOYED

FRAUD All FRAUD All

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

Males 0.33 0.96 0.96 0.96 0.78 0.86 0.81

3.2. Additional controls

Variable Parsimonious speciﬁcation Baseline speciﬁcation

Coefﬁcients Marginal effects Coefﬁcients Marginal effects

Males Females Males Females

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

FRAUD 0.282nnn 0.076nn 0.108nnn 0.196nn 0.077nnn 0.074nn

Pseudo-R2 0.241 0.271

Variable Males Females

nnn nn nn nnn n nn nnn nnn nn

COMPUTER PROFICIENCY 0.166nnn 0.170nnn 0.205nnn 0.199nnn

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

Males Baseline speciﬁcation 0.127 0.297 1.309 0.227 1.785 1.579

Deﬁnition 1 (Selection ratio). The selection ratio θ^ is deﬁned as

5.1. Nonparametric identiﬁcation

Δi ≔Y i ðZ~ i ¼ 1Þ Y i ðZ~ i ¼ 0Þ:

Deﬁnition 2 (Average treatment effect).

Γ≔E½Δi jZ i ¼ Z~ i  ¼ E½Y i ðZ~ i ¼ 1ÞjZ i ¼ Z~ i  E½Y i ðZ~ i ¼ 0ÞjZ i ¼ Z~ i :

Condition 1. Mean outcomes are non-decreasing in the treatment conditional on Z~ i ¼ 1:

E½Y i ðZ~ i ¼ 0ÞjZ~ i ¼ 1 r E½Y i ðZ~ i ¼ 1ÞjZ~ i ¼ 1:

Condition 2. People who do not commit academic fraud report truthfully: Z~ i ¼ 0 ) Z i ¼ 0:

5.2. Lower bound on the selection ratio

Proof of Proposition 1. By deﬁnition

¼ E½Y i ðZ~ i ¼ 0ÞjZ i ¼ 0; Z~ i ¼ 0

Proof of Lemma 1. Under Condition 3 we have, for p A f0; 1g,

which can be rewritten as

S~ uo ≔E½wi jZ i ¼ 1; Z~ i ¼ 1 E½wi jZ i ¼ 0; Z~ i ¼ 0 ðA:3Þ

plimðα~ Þ  αþ ϕS~ uo ðA:4Þ

For the left hand side we get

Appendix B. Supplementary data

Transparency International, 2013. Global Corruption Report: Education.

You might also like

Γ≔E½Δi jZ i ¼ Z~ i ¼ E½Y i ðZ~ i ¼ 1ÞjZ i ¼ Z~ i E½Y i ðZ~ i ¼ 0ÞjZ i ¼ Z~ i :

E½Y i ðZ~ i ¼ 0ÞjZ~ i ¼ 1 r E½Y i ðZ~ i ¼ 1ÞjZ~ i ¼ 1:

¼ E½Y i ðZ~ i ¼ 0ÞjZ i ¼ 0; Z~ i ¼ 0

S~ uo ≔E½wi jZ i ¼ 1; Z~ i ¼ 1 E½wi jZ i ¼ 0; Z~ i ¼ 0 ðA:3Þ

plimðα~ Þ αþ ϕS~ uo ðA:4Þ