A brief literature overview

Three strands of academic literature are directly relevant for our study. First is the role of the recruiters’ expectations and interpretations of applicants’ observed characteristics, especially university identity. Second is the analysis of the relationship between the expected validity of educational credentials and labour market outcomes. Third is the discussion about methodological strategies that effectively allow for a credible identification of causal relations between the applicants’ observed characteristics and market outcomes.

Unescapably, recruiters face an informational problem when they assess applications, and thus act upon their beliefs and perceptions about the signals and information that are made available to them. The university identity conveys information to recruiters about applicants’ educational credentials acting as effective signals allowing them to cope with the information asymmetry. As one directly and easily observable trait, university identity often triggers stereotypes and preconceptions about productivity in the recruiters’ mind-sets (Reference Bosak, Eagly and DiekmanBosak et al., 2018). Thus, educational institutions are increasingly paying attention to the labour market success of their graduates as a means to boost their reputation among recruiters (Reference Rhiel, Saavedra, Urquiola, Hoxby and StangeRiehl et al., 2017). On this matter, several studies have stressed recruiters’ inability to make perfectly objective assessments of an applicant’s educational credentials, as they are only partial, imperfect signals of their skills (Reference ArrowArrow, 1973; Reference PhelpsPhelps, 1972; Reference SpenceSpence, 1973). Along these lines, Reference BaillyBailly (2008) proposed a reinterpretation of Spence’s classical theory about education signalling as a dynamic error-and-trial process of learning in which recruiters may have several interpretations of the educational signals conveyed by the applicants. In that sense, Bailly stresses that educational outputs should not be seen as substances that are always similarly interpreted, but rather as non-substantial components of a decision environment that may be differently valuated according to each recruiter’s beliefs.

These ideas strongly suggest avoiding a simplistic viewpoint to assess the role played by a university degree as a signal for recruiters – there is burgeoning literature on this matter. Reference Hastings, Neilson and ZimmermanHastings et al. (2015) found a highly heterogeneous premium schedule in Chile, varying along different types of college degrees. Degrees awarded by highly selective programmes and institutions are found to imply large labour market returns compared to other degrees. Reference Brotherhood, Ferreira and SantosBrotherhood et al. (2019) stress the importance of school quality differential among states in Brazil as drivers of unequal monetary returns in the labour market. While analysing productivity of colleges in Colombia in terms of their contribution to their graduates’ learning and future earnings, Rhiel et al. (2017) find that the most highly selective schools, and thus those providing the highest quality education, add the most to their students’ earnings in the labour market.

The aforementioned studies mainly rely on observed measures of school quality, either through a direct assessment of students’ abilities by means of entrance/exit college exams, or input-related indicators of school quality including expenditure per student or teacher–pupil ratio. Another vein of studies, however, focuses on the role of perceived school quality from the recruiter’s viewpoint. Reference Deming, Yuchtman and AbulafiDeming et al. (2016) conducted a field experiment about the perception of the value of postsecondary degrees in the United States, finding that applicants holding degrees awarded by online for-profit educational institutions are less likely to achieve success in the labour market. Reference MacLeod, Riehl and SaavedraMacLeod et al. (2017) also found evidence of perceived school quality premiums in Colombia, showing that school reputation, unlike years of schooling, is highly correlated with the evolution of graduates’ labour market earnings. Although MacLeod et al. use entry/exit exams as observed proxies of labour market school reputation, they originally propose a framework in which it is precisely the recruiters’ perceived reputation of educational institutions which may be at the origin of labour outcome differences.

One common, crucial issue in the aforementioned empirical studies is the application of adequately designed strategies to successfully identify the intended causal relations. One such methodological approach is a field experiment which is exempt of some core limitations of regression analyses, including over-saturation of the regression equation and the impossibility to effectively deal with unobservable confounders.

Two types of field experiment designs are commonly applied for collecting data in labour market studies such as the one that we undertake here. The first is an audit experiment, in which pairs of identical auditors, except for the characteristic that is tested as a source of differentials (such as race, skin colour, gender or educational credentials), engage in a personal interaction with recruiters to generate data. One handicap in this design that is often brought up in the literature (see, for example, Reference Bertrand, Duflo, Duflo and BanerjeeBertrand and Duflo, 2017) is that it relies on human interactions to generate information. One can hardly assume that people’s behaviour can be controlled to be identical for many obvious reasons. First is the simple fact that the auditors are different individuals and thus they will always have several distinctive personal traits that are unobservable prior to their interaction with recruiters. Second, as auditors know that they are part of an experiment, they may tend to react differently according to the evolution of their interaction with their counterparts.

A second type of design that overcomes the aforementioned limitations are correspondence studies, in which fictitious CVs are sent out to recruiters in order to generate information about their response, thus avoiding human interaction. This type of studies has been extensively applied in developed countries, such as the United States (Bertrand and Mullainathan, 2004), Sweden (Reference Carlsson and RoothCarlsson and Rooth, 2008), Germany (Reference Kaas and MangerKaas and Manger, 2010), Australia (Reference Booth, Leigh and VarganovaBooth et al., 2012), France ^{(Reference Duguet and PetitDuguet and Petit, 2005)} and the United Kingdom (Reference DrydakisDrydakis, 2015). There are only a few similar studies carried out in Latin America countries. Following this approach, Reference Bravo, Sanhueza and UrzuaBravo et al. (2008) analysed discrimination in Chile associated with names and places of residence. Reference Galarza and YamadaGalarza and Yamada (2014) studied discrimination in Peru associated with ethnic conditions and domestic standards of physical beauty.

Theoretical framework

We draw inspiration from Reference MacLeod, Riehl and SaavedraMacLeod et al. (2017) to set up the theoretical foundations and causal mechanisms of our study. Originally, MacLeod et al. posit a mechanism relating university reputation to the dynamic evolution of wages, and they operationalised it in a quasi-experimental setting using Colombian data. We limit ourselves to present here the essential aspects of their framework only to highlight that it is also useful to assess the nexus between university reputation and employability (not only wages), and that it allows also to establish clear hypotheses that can be tested in an experimental setting.

Universities build up labour market reputation based on their graduates’ performance. Let $R_u$ denote the labour market reputation of university $u$ , which is hard to perfectly observe. From the recruiter’s viewpoint, $R_u$ can be considered as a noisy group-measure of the work-relevant skills possessed by any graduate of university $u$ . If one denotes applicant’s $i$ stock of work-relevant skills as ${\theta }_i$ , then

(1) $R_u={\theta }_i+{\xi }_i$

where ${\xi }_i~N(0,{\sigma }_{\xi })$ represents the noise embedded in university reputation as a signal of the applicant’s skills. Thus, $R_u$ can be interpreted as the expected skill-stock of a random university $u$ graduate

(2) $R_u=E\left[{\theta }_i|i\in u\right]$

Assuming that skills are normally distributed among applicants, then the university’s reputation represents the mean of their graduates’ skill schedule

(3) ${\theta }_i~N\left(R_u,{\sigma }_{\theta }\right)$

where ${\rho }_{\theta }\equiv 1/{\sigma }_{\theta }$ denotes the precision of the university’s labour market reputation as a signal of its graduates’ skills.

This setting implies a crucial role for the applicants’ CVs and the information conveyed therein. Alongside the university’s identity, the CV provides a vital set of information for recruiters to make their hiring decisions. This set includes the applicant’s gender, their family and ethnic background, their work experience, hints of their socioeconomic position and many other personal characteristics that are carefully taken into consideration by the recruiters in their quest to successfully manage asymmetrical information. In this setting, once the CVs are sent out, the applicant’s success is entirely in hands of recruiters and the way in which they interpret the signals they are given in the CVs. This clearly evokes the importance of Reference BaillyBailly’s (2008) systems of beliefs in a setting of informational frictions induced by noisy signalling.

Let us denote the information set conveyed in applicant’s $i$ CV as $I_i$ , which has a finite overall precision, ${\rho }_I$ , as a signal of work-relevant abilities for the recruiter. The information set includes the university’s identity, which must not be confounded with the university’s labour market reputation (Reference MacLeod, Riehl and SaavedraMacLeod et al., 2017). The former is an observed factor, but the latter is an unobserved element resulting from the interpretation of the university’s identity by the recruiter (Reference BaillyBailly, 2008; Reference CaiCai, 2013).

Generally, let us now denote as $y_i$ a certain labour outcome for applicant $i$ . In their framework, Reference MacLeod, Riehl and SaavedraMacLeod et al. (2017) consider wages as the labour outcomes of interest, but it may be extended to employability, waiting-time for a positive response or any other material as well as non-material job-related characteristics. In our study, the outcome of interest is receiving a call-back, which, in this framework, is a noisy reflection of the recruiters expected work-relevant skill-stock, given the information included in the CV

(4) $y_i=E\left[{\theta }_i|I_i\right]+{ϵ}_i$

where ${ϵ}_i~N(0,{\sigma }_{ϵ})$ represents all other non-considered random factors affecting the outcome.

Assuming normality for all variables, it is possible to use Bayes’ rule to express the latter equation of labour outcome formation as

(5) $y_i={\pi }_R{R}_u+{\pi }_I{I}_i+{ϵ}_i$

where ${\pi }_R\equiv {\rho }_R/({\rho }_R+{\rho }_I)$ and ${\pi }_I\equiv {\rho }_I/({\rho }_R+{\rho }_I)$ . Thus, the outcome is a function of the applicant’s university reputation, the overall information included in their CV and other non-considered factors, which may be largely context-dependent. This means that the unconditionally expected outcome is a function of the applicant’s university reputation and other individually varying signals included in the CV and thus observed by the recruiter

(6) $E\left[y_i\right]={\pi }_R{R}_u+{\pi }_I{I}_i$

The precision of each element in Equation (6) as a signal of the applicant’s abilities is directly associated with its importance as a determining factor of the labour market outcome. We rule out the possibility of having ${\rho }_I=0$ and ${\rho }_R=0$ simultaneously, as this would mean that the outcome is confounded with a random error term, regardless of the CV’s content, which is unrealistic. It may be the case, however, that ${\rho }_R=0$ while ${\rho }_I\ne 0$ , which would imply the irrelevance of the university’s labour market reputation for recruiters’ decision-making. Similarly, it may be case that ${\rho }_R\ne 0$ while ${\rho }_I=0$ , representing a situation where the university’s labour market reputation completely determines the expected outcome, regardless of any other information embedded in the applicant’s CV.

We argue that this framework allows us to gauge the extent to which university reputation has a causal effect on labour market outcomes based on a correspondence experimental study. To capture the effect of university reputation, let us consider two applicants, $i$ and $j$ , for which we construct identical corresponding information sets $I_i$ and $I_j$ , except for the university identity. In this framework, this situation triggers different interpretations of the applicants’ CVs by the recruiters that are entirely due to distinct university reputations. Let us say that applicants $i$ and $j$ attended universities $u$ and $w$ , respectively. Thus

(7) ${\Delta }_{R}E\left[y\right]\equiv E\left[y_i\right]-E\left[y_j\right]={\pi }_R\left(R_u-R_w\right)$

In a correctly designed experimental setting, statistically significant differences in the expected outcomes can only be attributed to differences in the reputation of the applicants’ respective universities. Any evidence supporting that ${\Delta }_{R}E\left[y\right]\ne 0$ can be interpreted as causal evidence of labour market returns to university reputation, as measured by the considered outcome. We stress that the interpretation of this differential as an average impact of university reputation – or an average treatment effect (ATE) in traditional treatment effects literature – relies on the successful configuration of the complementary information conveyed in the applicant’s CVs. We now go on to explain how we set up the correspondence experiment to effectively estimate this impact.

Experiment design and procedures

Our experiment covers the urban areas of the three largest Bolivian cities, namely La Paz, Cochabamba and Santa Cruz, which concentrate 76% of the urban population. Our sample frame is composed of all formal firms operating in these cities that have a register as employers in the Ministry of Labour as of January 2017. This criterion excludes a very large number of single-owner firms, which represent 79% of the existing formal enterprises in these cities. As they often lack the status of a formal employer, an internship in this type of firm may hardly result in a subsequent formal job offer for the applicant, justifying their exclusion from our analysis. Thus, our sample frame comprises 46,095 firms with a valid and publicly known email address, of which we have drawn a random sample of 1424 firms, corresponding to a 2.5% sampling error. The firm sample composition is presented in Table 1.

Table 1. Distribution of the firm sample.

Source: Own.

Two key elements received particular attention prior to the execution of the correspondence study. First is the construction of a set of CVs specifically designed for the purpose of this study, and second is the definition of random groups of firms.

The applicants’ identities were carefully created to convey neutral information except for the university identity. For this, we take into account each of the following points, with particular attention when it comes to gender – which, of course, is the only element in the CV that we cannot set to be neutral:

• • The given and last names were carefully chosen to be some of the most common in the country, while also conveying neutral information about key socioeconomic characteristics, including ethnicity and family background.^{Footnote 3}

• • We indicated the possession of an undergraduate degree in Business Administration. We justify this choice because it is one of the most highly demanded by employers, it is offered by all the ECUB colleges and we argue that it helps us to avoid preference biases towards more specific or specialised professional training.

• • A set of skills and competencies was carefully selected in order to be adequately fitted to usual requirements by firms, while conveying abilities that can be considered common among the entry-level working population. The skill profiles were created using data from the World Bank’s Survey Towards Employability and Productivity collected in 2012, which is, to the best of our knowledge, the only survey that allows to measure the skills and competencies among the working population in the country. These skills and competencies include knowledge of accounting and finance, mathematics and statistics, good communication skills (writing, speaking and comprehensive reading), proficiency in the use of computer and Internet browsing and a basic level of writing and speaking in English as a foreign language.

• • We used data from the 2016 Demand Labour Survey collected by IDB to measure the recruiters’ perceptions about the quality of education in universities located in La Paz, Santa Cruz and Cochabamba (Reference Urquidi, Ergueta and ForondaUrquidi et al., 2020). In absence of an official assessment of university quality in the country, we use this information to define a ‘well-valued university’ as an institution that may facilitate the employability of its graduates in the formal sector as opposed to a ‘poorly valued university’. In separate analyses for each one of the three urban areas, the well-valued university is the one having the highest valuation and is reported to be known by more than 50% of the respondents. In turn, the poorly valued university is the one having the lowest valuation and is reported to be known by more than 50% of the respondents. Thus, well-valued universities are the ones enjoying the best reputation among recruiters in their respective locations, and the converse is true for the poorly valued universities.

• • Based on the vast existing evidence suggesting discrimination against female applicants (Reference Banerjee, Bertrand and DattaBanerjee et al., 2009; Reference Bertrand and MullainathanBertrand and Mullainathan, 2004), we chose to assign male identities to the main branch of the experiment seeking to maximise response rates. However, we acknowledge that this decision can pose grave identification threats, as we risk confounding the intended university reputation effects with gender-induced effects. In order to be able to show that our results are not capturing such gender-related confounding effects, we consider a third additional control experimental branch operationalised by the creation of a CV corresponding to a female applicant graduated from a well-valued university. We do not consider an identity corresponding to a female applicant graduated from a poorly valued university due to the small sample in our data. We choose to safeguard the power of our main experimental branch, while having enough data to confidently rule out gender-induced confounding effects.

Thus, three broad types of fictitious CVs were created:

• • Type A: Male applicant, graduated from a well-valued university.

• • Type B: Male applicant, graduated from a poorly valued university.

• • Type C: Female applicant, graduated from a well-valued university.

Types A and B are used in the main branch of the experiment as they signal male applicants with distinct university identities. Types A and C are the CVs used in the gender-control branch of the experiment. As we mentioned earlier, the university identity was assigned according to the results obtained by area (La Paz, Cochabamba and Santa Cruz), yielding a total of nine fictitious CVs specifically created for this study. Thus, our experiment also controls for regional variations.

Let us now focus on the strategy followed to create groups of randomly selected firms and thus operationalise the correspondence study, while making the most out of the sample of firms at hand. Following Reference Bertrand, Duflo, Duflo and BanerjeeBertrand and Duflo (2017), we reason that in an adequately designed experimental setting, which generates data that is effectively controlled for unobserved characteristics, decisions made by recruiters need to take place while simultaneously facing two generic types of fictitious CVs. Thus, our strategy consisted of forming random groups of firms in three stages: the first regarding the type of experiment (i.e. whether we are assessing the role of gender or that of university reputation), the second regarding the timeframe for sending out the applications and the third regarding the order in which the firm received the CVs.

At the first stage, we made sure that each firm received a pair of CVs according to the type of experiment in which they were randomly assigned:

• • Half of the firms (712) received Type A and Type B CVs, and thus their response allows us to gauge the effects of university reputation on the outcome. Hence, in the remainder of the document we refer to this part of the study as the main experiment branch.

• • Half of the firms (712) received Type A and Type C CVs, allowing to gauge any potential distortion on the outcome caused by our decision of assigning a male identity. Thus, we refer to this part of study as the gender-control branch. Since this a complementary experimental branch, we chose to assign well-valued universities to the applicants seeking to maximise the response rates for this additional branch.

At the second stage, each one of the groups formed in the first stage were randomly divided in two groups (each one comprising 356 firms). In each type of experiment, one of these subgroups received the CVs between January and February 2017, and the other subgroup received them between June and July 2017. These timeframes were chosen because they correspond to recess terms in the tertiary educational system, thus depicting periods in which firms usually expect to receive internship applications. By sending out CVs at two different points in time, we explicitly seek to control for potential seasonal confounding effects.

At the third and final stage, each of the subgroups formed at the second stage were randomly divided into two additional groups (each one comprising 178 firms), for which we have inverted the order in which the CVs were received. Through this procedure, we seek to avoid potential biases caused by any unobserved factor operating while the recruiters take a first glance at the fictitious CVs. In all cases, we made sure that each firm received the second CV within the next 48 hours after they received the first one.

Summary and concluding remarks

This study was motivated by the fact that the well-documented positive relationship between a rapid educational expansion and improvement of employability seems to be very weak, almost inexistent, in Bolivia during the last decade. This situation may have several explanations as it is related to fundamental characteristics of the labour market, such as culture, demographics, long-run productivity and so on. However, drawing inspiration from Reference BaillyBailly (2008) and Reference MacLeod, Riehl and SaavedraMacLeod et al. (2017), we have demonstrated that recruiters’ perceptions and interpretations of applicants’ observable signals play a crucial role in the transition from college to the workplace. Thus we made a case for informational frictions between applicants and recruiters as one plausible explanation of the above-mentioned disconnection, in the context of an increasingly heterogeneous labour force. Following Reference CaiCai (2013), our study is inscribed in scholarship about the role of university reputation on labour market outcomes, as we focus on the role of recruiters’ interpretation of educational credentials.

We generated data for this study by implementing a correspondence experiment in the three main Bolivian cities (La Paz, Cochabamba and Santa Cruz) on a sample of 1424 formal firms that are registered as employers in the Ministry of Labour. We sent pairs of fictitious CVs to each firm along with an application for an unpaid internship, hence capturing decisions made by firms in face of two applicants that are identical expect for two controlled characteristics. In the main branch of our experiment, half of the firms received CVs from a male well-valued and a poorly valued university, according to the 2016 IDB Labour Demand Survey data. We designed the experiment to have an auxiliary gender-control branch, in which CVs corresponding to a male and a female candidate from a well-valued university were sent out.

Our main finding – a quite novel one for the Bolivian case – is that university reputation premiums are quite large. Applicants from a well-valued university are 40% more likely to receive a positive response compared to applicants showing credentials form a poorly valued university – a 2.25 percentage point advantage over a baseline likelihood of 7.87%. This shows the extent to which informational frictions play a defining role in the communication process between applicants and recruiters. Perceived college quality heterogeneity from the recruiters’ viewpoint is thus proven to be an important source of disconnection between educational advancements, as measured by quantity of education and actual improvements in the workplace. University reputation and the perceived value of educational credentials need to be seriously taken into account in order to grasp a fuller picture about the transition from college to the job market in Bolivia.

This result is free of regional differences and seasonal confounding effects. Most notably, they are also free of gender-induced confounding effects. We find that both men and women are equally likely to receive a positive response (around 6.74%). The reason for this result may be manifold. First, it may be due to the fact that an unpaid internship implies nearly nil costs for the firm, but above all, that this situation is likely to arise under identical hiring costs for the firms. Second, our fictitious CVs signalled a degree in Business Administration, purporting to convey generic useful skills for firms in a wide array of economic activities. One may hardly defend that this profession is gender-oriented, providing another plausible explanation for our result. Additional studies are required to assess whether this is the case for other professions. Third, this result may partly be the reflection of the rapid inclusion of women in the tertiary education system for which we made a case earlier.

Of course, we do not wish to imply that gender discrimination is non-existent in the country overall. Rather, we wish to lay out some possible conditions in which discriminating behaviour in the labour market seems less likely to occur, based on an internally valid experimental setting. Indeed, discrimination may persist beyond the intended scope of our study.

We argue that several veins for future research in the country arise in light of this result. First, however, it is important to recognise that studies accounting for quantity of education only do not seem able to grasp the true complexity of the connection between education and labour outcomes in Bolivia. Thus, it may be useful to transcend traditional survey data and administrative registers in the quest to fully understand the matter at hand. We stress that one aspect requiring further attention by researchers is the formation of college reputation and its relation to school choice by students; we intend to pursue this. In this study, we have demonstrated that college reputation overpowers gender as a determining factor for recruiters’ decision-making in the formal sector. This result is valid in the main Bolivian cities when conveying a degree in a gender-neutral profession, namely Business Administration. We argue that this result may hardly be extended to other contexts, for which further data needs to be generated. We intend to focus our future research efforts in these directions, as we are convinced that many processes in the transition from college to the workplace remain uncovered. This certainly hinders public policies’ capabilities of improving people’s livelihoods and bringing about human development in the country.