Introduction

A growing number of students are experiencing stay-abroad periods during their tertiary education. Credit mobility is defined as a short-term mobility experience of up to one academic year in a foreign country for study or for an internship, during which a student gains credits that will be recognized upon their return to the home country to complete their degree (Junor & Usher, 2008; Teichler et al., 2011).Footnote 1 Student mobility is one of the components of transnational higher education with the most significant socio-economic, cultural, and political implications (Guruz, 2008).

Although government support for student mobility programs is not a recent phenomenon, incentives for mobility have expanded in recent years in terms of resources and people involved, and territories covered (Engberg et al., 2014; Guruz, 2008). For example, the total budget for the Erasmus + program, the largest and most reputed European credit mobility program, nearly doubled its financial resources from 2014–2020 to 2021–2027, with a total estimated investment of 26.2 billion euros for the latest period (European Commission, 2021). The program started in 1987 with only 3244 students and now involves more than 300,000 students yearly (European Commission, 2019). Temporal mobility experiences growth has been recorded across all regions globally, with North America and Western Europe as the favorite destinations welcoming almost half of all mobility students yearly.

It is already well established in the literature that international temporal mobility experiences benefit students. For instance, it has been shown that going abroad boosts student’s soft skills (Meya & Suntheim, 2014; European Commission, 2016), reputation (Engberg et al., 2014), career prospects (Di Pietro, 2013; Parey & Waldinger, 2011), acquisition of new skills (Sorrenti, 2017; Wang et al., 2019), and student performance (Contu et al., 2020; Gonzalez-Baixauli et al., 2018; Meya & Suntheim, 2014). However, despite the amount of work on the general impacts, little attention has been dedicated to exploring heterogeneity across mobility programs (Van Mol et al., 2021). For this reason, our work focuses on one of the dimensions differentiating international credit mobility programs, the temporal one.

Students can experience mobility in different moments of their academic careers and stay abroad for short or extended periods. We ask, (i) does the impact of student mobility on student performance vary across students traveling in different periods of their undergraduate program?; (ii) does the impact of student mobility on student performance vary across programs with different durations?

To answer those questions, we use unique data on more than ten thousand undergraduate students who graduated between 2010 and 2020 from one of the most internationalized Brazilian universities, the University of Campinas. The country choice is because, so far, most studies have focused on the impact of exchange programs using samples of European students, mainly from the Erasmus program (Contu et al., 2020; Czarnitzki et al., 2021; Di Pietro, 2013; European Commission, 2016; Gonzalez-Baixauli et al., 2018; Meya & Suntheim, 2014; Parey & Waldinger, 2011; Sorrenti, 2017; Wang et al., 2019). To the best of our knowledge, there is no study evaluating the impact of student mobility on academic performance in any Latin American country. Still, data reveal that Latin America and the Caribbean registered an increase of 40% in the number of tertiary students studying abroad from 2011 to 2018, behind only the Arab States (72%) and the Asia and Pacific region (51%) (UNESCO, 2021). Studying the impacts of student mobility in developing countries is extremely important, especially given the role of education in the development of those countries (Szirmai, 2015).

Brazil also constitutes a very suitable research context due to the process that the country has been experiencing recently. After a period of growth in the mobility phenomena, Brazil is experiencing a trend shift. Between 2000 and 2017, the population of Brazilian students studying abroad increased by more than 200%, going from 18.5 to 58.9 thousand students (UNESCO, 2021). The Science without Borders initiative, sponsored by the federal government between 2011 and 2015, granted more than 90 thousand international mobility scholarships, of which 79% were for undergraduate students (Brasil, 2016). Moreover, positive spillovers generated by the initiative, the so-called “Science without Borders effect,” boosted the number of scholarships even in areas not covered by the program (Manços, 2017; Granja & Carneiro, 2020). More recently, however, the growing trend slowed down. The change in the Brazilian federal administration and the economic and political crisis experienced by the country has resulted in severe budget cuts in the higher education system and the financial resources dedicated to international student mobility programs (Andrade, 2019; De Negri, 2021). According to a recent report from the Institute for Applied Economic Research, a national public institution supporting the Brazilian federal government’s public policies, federal investments fell about 37% between 2013 and 2020 (De Negri, 2021). The Ministry of Education suffered the most critical budget cut, and it is expected that this cut will directly impact the training of Brazilian researchers, both in Brazil and abroad (De Negri, 2021). Thus, it is crucial to investigate the impact of mobility programs to understand the consequences (if any) of such education budget cuts on students’ future.

By applying a combination of Propensity Score Matching and Difference in Differences, we explore the causal relationship between a mobility experience and students’ academic performances. This study offers empirical evidence on when and for how long students should go abroad, providing insights to policymakers engaged in maximizing the effects of mobility programs. This kind of analysis is of utmost importance, given the heterogeneity of mobility programs in the country and the varied potential outcomes depending on the type of mobility experience.

Recently, Van Mol et al. (2021) have recognized the importance of analyzing differences across mobility experiences. They distinguished between study levels when going abroad (Bachelor versus Master), the scope of the experience (study versus internship), and the destination country. Differently from them, we focus our attention on programs offering students the opportunity to go abroad at different moments during their studies and choose how long to stay. Moreover, while Van Mol et al. (2021) consider the impact of mobility on labor market returns, we look at the performance of the students when completing their studies upon return. The temporal parameters (time and duration of mobility) are variables that funding agencies and governments can adjust when designing or updating study programs.

This paper is structured as follows. First, it reviews previous studies about the impact of an exchange program on students. Second, it details the data and the methodology chosen for the analysis. Third, the paper presents and discusses the main results of the analysis. Last, the conclusions are presented.

International Student Mobility and Students’ Outcomes

Literature has extensively discussed the impact for students of participating in mobility programs during their studies (Roy et al., 2019). In reviewing the literature, we group those studies along five outcome dimensions: soft skills, reputation, career prospects, acquisition of new skills, and student academic performance.Footnote 2

Looking at the impact of international student mobility on soft skills, Meya and Suntheim (2014) review the literature on the field and list multiple benefits of studying abroad, namely: (i) positive impact on the development of students’ personalities and cross-cultural skills; (ii) transformation of these students into more independent, approachable and agreeable people; and (iii) increased acceptance of new cultures and new ways of working. Along the same line, a study by the European Commission (2016) about the impact of the Erasmus program on students’ personalities, skills, and careers found that an international mobility experience generated positive changes in students’ personalities, influencing characteristics considered valuable to employers.Footnote 3 According to the study, “the average change achieved in six months through the Erasmus program can be considered equivalent to a personality change that would normally happen over 4 years of life without Erasmus experience” (European Commission, 2016, p. 16).

Studying abroad also has a reputation effect on students. For instance, Engberg et al. (2014)Footnote 4 pointed out that receiving a mobility scholarship is already an advantage in itself. They argued that the award is usually seen as a proxy for academic excellence, which guarantees benefits in the labor market for those who obtained it. In addition, receiving high-quality training abroad and developing relationship networks could positively impact scholarship holders. The authors argue that having contact with another language and culture and expanding the beneficiaries’ worldview could also be translated into personal and professional advantages.

Other studies also showed that studying abroad has several benefits in terms of career prospects. For example, Di Pietro (2013) investigated how participation in study abroad programs during university impacted subsequent employment likelihood. By drawing on a sample of Italian graduates, the author found that the probability of being employed 3 years after graduation increased by about 22.9% points due to studying abroad. The effect was mainly driven by students from disadvantaged backgrounds (those with one or both parents with lower or upper secondary education). Amendola and Restaino (2017) explored data from a web survey on a cohort of students from the University of Salerno in the South of Italy who participated in the Erasmus program and found that students are generally motivated to go abroad because they believe in benefiting from a boost in their employability, with 61.87% of the surveyed students revealing that prospective employers perceived the mobility experience very positively during job interviews. Bryla (2015) leveraged a large-scale survey among Polish students who participated in mobility programs, finding that one-third attributed a very important role to the mobility experience in their professional career development over 5–6 years after their return. Moreover, the author found an association between mobility experiences and some characteristics of the employers. For instance, mobile students are more likely to be employed in companies with a higher level of internationalization. Also, in the same Polish context, Gajderowicz et al. (2012) found that employers perceive mobility as a signal of adaptiveness, motivation, and good learning skills. Employers prefer mobile students, and students who experienced a period abroad during their studies record a higher probability of finding a job and shorter search times than students who pursued their entire studies in Poland. Kratz and Netz (2018) found that facilitated access to job opportunities allows mobile students to obtain higher wage growth through employer changes. Additionally, the higher probability of working in large and multinational firms assures mobile students higher medium-term wages (Kratz & Netz, 2018). Waibel et al. (2018) explored heterogeneities among groups of individuals experiencing mobility. They found that those who benefited the most from mobility were those with the lowest propensity to study abroad, i.e., those from disadvantaged economic, social, and cultural groups. The positive effect of student mobility on early career occupational status is limited to graduates from generalist fields of study, while graduates from specialized fields have smooth access to the job market, regardless of their experiences in foreign countries. Netz and Grüttner (2020), when analyzing if the effect of studying abroad on graduates’ labor income varies across social groups in the German labor market, found that graduates from a high social origin benefit slightly more from international student mobility than those coming from a low social origin, concluding that student mobility tends to foster the reproduction of social inequalities in the labor market. In turn, Parey and Waldinger (2011) investigated the effect of studying abroad on international labor market mobility later in life for university graduates. Using a sample of five cross-sections of German students, they found that studying abroad increased the probability of working in a foreign country by about 15% points. They also found that the most disadvantaged students (those who were credit-constrained and had less educated parents) had the highest returns from studying abroad, showing the importance of focusing on those students to increase the return from exchange programs. However, not all studies converge in finding positive returns to mobility concerning students’ careers. For instance, Van Mol et al. (2021), having controlled for selectivity into student mobility, found that mobility does not impact early career outcomes, either in terms of wages or the time to find a job after graduation.

One way studying abroad can impact employability is by acquiring new skills, especially language skills. Sorrenti (2017) used a sample of Italian graduates from 2007 to 2010 and found that studying abroad was essential for foreign language acquisition. However, the author found a substantial heterogeneity across languages since higher effects happened for languages close to students’ native tongue, the latter being the languages less rewarded by the labor market in terms of wage premium. Similarly, Wang et al. (2019) evaluated the benefits of a yearlong study abroad program on developing linguistic and multicultural skills measured by their academic results (overall and on languages) before and after international mobility. They used a sample of students at a British university from 2008 to 2014 and found statistically positive effects of studying abroad on academic learning.

The closest branch of studies to ours investigates how participating in an international study program affects students’ academic performance. Meya and Suntheim (2014) investigated how studying abroad affects success at university, focusing on students from a German university between 2006 and 2011. They found that a brief study-related visit abroad significantly increased the final university grade. However, the grade increase was mainly driven by the mere transfer of grades obtained abroad. They also showed that studying abroad reduced the probability of finishing university within the standard period, suggesting that higher grades came at a cost. Another example is Contu et al. (2020), which investigated if exchange programs positively impacted the graduation bonus of students, focusing on those from the Erasmus program enrolled at an Italian university from 2015 to 2017. They found that the effect of international mobility on the graduation bonus was context-specific and depended on the faculty and the type of degree.

The majority of existing studies have found that students benefit from mobility programs concerning their academic performance. However, there is no full convergence of results. For instance, Gonzalez-Baixauli et al. (2018) analyzed a dataset of students from a Spanish university from 2001 to 2013 and found that, even though student mobility positively affected students’ grades, the impact was not homogeneous across mobility programs or geographical areas. They also found that the increase in grades partially vanished upon returning to their home university after the mobility period. On the other hand, Czarnitzki et al. (2021) focused on a sample of Belgian students from 2006 to 2010 and found that, on average, exchange students had a decrease of 7% in their final grade compared to non-mobile students. That effect was heterogeneous regarding the field of study, type of exchange, and the host institution. The authors stated that the negative effect could be due to a possible mismatch between the courses taken abroad and the home university curricula, leading to exchange students not learning the required content for upcoming courses, reducing their grades.

Our study adds to the work by Contu et al. (2020), Czarnitzki et al. (2021), Gonzalez-Baixauli et al. (2018), and Meya and Suntheim (2014) by focusing on credit mobility programs’ impact on student academic performance. It addresses a gap in the literature, which is the study of the temporal dimension of exchange programs (such as timing and duration), parameters that policymakers can adjust to increase the efficiency of those programs. Even though the academic literature already acknowledges the temporal dimension of exchange programs,Footnote 5 to the best of our knowledge, no studies asked whether there is a more appropriate moment or duration of a student mobility experience to increase students’ performance.

Data

Empirical Setting

Our sample comprises 11,432 students from the University of Campinas (UNICAMP), Brazil, from 2010 to 2020. UNICAMP is a well-known research-intensive university that stands out in the Brazilian higher education system. In 2019, it was among the best Brazilian universities evaluated by the Brazilian Ministry of Education (Brasil, 2020a). According to the Times Higher Education Latin America ranking, it was ranked third among Latin American universities in 2020 (Times Higher Education, 2020). The university is located in São Paulo state, the Brazilian state with the highest Gross Domestic Product in the country (Brasil, 2020b).

The choice for UNICAMP is because the university has broad experience with internationalization initiatives such as international cooperation and student mobility. Since its foundation in the 1960s, internationalization has been part of its primary institution strategy (Granja & Carneiro, 2020). The university is highly involved in the population of mobility programs in the country. For example, in the case of the Science without Borders program, UNICAMP placed itself in seventh place among the top 10 universities in terms of the number of students sent abroad (Brasil, 2016). Most universities ranked in this top 10 were large research-intensive public universities with similar characteristics to UNICAMP in terms of size and type (Schwartzman et al., 2021).Footnote 6

UNICAMP offers a varied range of exchange programs to its students, both at the undergraduate and postgraduate levels. Even though the selection criteria and the activities planned abroad are similar, programs have different natures and settings. For example, in addition to the mobility carried out via agreements with foreign institutions to exempt tuition fees (the majority aimed at undergraduate students), UNICAMP also participates in programs financed by either private or public agencies, such as the Santander private bank, the Association of Universities of the Montevideo Group (AUGM) and the Brazilian Ministry of Education.

Between 2010 and 2017, the university had more than 500 agreements with foreign institutions, covering more than 60 countries (Granja, 2018). A part of those agreements was fostered by the university’s participation in Science without Borders, a program created by the Brazilian federal government between 2011 and 2015. Additionally, some university courses, such as engineering, also offer the possibility of taking a double degree at foreign universities. The exchange duration varies depending on the university’s agreements with the host university and the external funding agency, usually lasting between one semester and 2 years.

Given its tradition of internationalization and the program variety, the number of UNICAMP students in mobility programs in the previous decade was elevated. Of the 11,432 students considered in this study, 1943 participated (at least once) in an institutional student mobility program (17% of the entire sample), while 9489 were in the non-treated (nonparticipants) group.Footnote 7

Variables

The main dependent variable of this paper is students’ academic performance, measured by the grades achieved in the university undergraduate program. Specifically, as an academic performance measure, we consider the standardized Performance Coefficient of the last semester students attended university. At UNICAMP, grades are calculated on a scale from 0 to 1, with 1 being the maximum grade. The grade for a semester is the average of the grades obtained in the course subjects taken during that semester, weighting by the course load (credits). The resulting aggregated grade is called Performance Coefficient. Since undergraduate courses and course subjects have different difficulty levels, all grades used in the analysis were standardized by course and year of admission at the university. The standardization strategy helps compare students from different cohorts and courses, and it is also widely used by UNICAMP in recruitment processes (for exchange scholarships, for instance) since it makes clear whether students’ grades fall below or above their cohort average.Footnote 8

Our final sample includes students who met one of the following criteria: (1) students who completed their courses; (2) students who abandoned university or did not renew their registration; and (3) students who were dismissed from the university (for instance, due to low grades or low progression). For students who met criteria 2 or 3, we considered the standardized Performance Coefficient of the last semester attended before quitting the university. We included them in our sample since the decision to drop a course is often the result of obtaining low grades, so excluding them might determine a selection problem. As a robustness check, we also run our analysis on the subsample of students who completed their courses (students satisfying the first criterion only).

Students who were still enrolled at the end of our observation period were not considered, as we aim to evaluate the impact of mobility on the overall student’s career, and those students do not have a final semester grade. Moreover, for the students who have not completed their study path, it is impossible to determine either the amount of time spent abroad or the participation in a mobility program if they go abroad later in their studies.

To ensure that each student was considered only once in the sample, only students registered for only one undergraduate course (who did not do more than one program at UNICAMP) were considered in the analysis. Moreover, due to the lack of complete information on non-regular students, only those who entered the university through the regular selection process (through an entrance exam) were considered.Footnote 9

Figure 1 shows the distribution of the grades for the last semester at the university for mobility students (also referred to from now on as the treatment group) and non-mobility students (non-treated or nonparticipants group). As we can observe, students who participated in international mobility programs had slightly higher final grades than the nonparticipants.Footnote 10 However, those differences cannot yet be attributed only to participation in mobility programs.

Fig. 1
figure 1

Dependent variable kernel density (mobility vs. non-mobility students). Data source Authors’ estimation from UNICAMP’s microdata

Full size image

Table 1 lists and describes all the variables included in our analysis. The rationale for choosing the independent variables is explained in detail when discussing the empirical strategy. Students’ academic, demographic, and socio-economic information was shared directly by the UNICAMP’s Academic Board and International Office after the approval of the Brazilian Research Ethics Committee.Footnote 11

Table 1 Variables description
Full size table

Table 2 shows the summary statistics for our sample of students. Not surprisingly, treated and non-treated students differ significantly in all baseline characteristics. Mobility students have, on average, better academic performance both before and during university. They also have, on average, higher incomes (55% were in the top 50th income percentile when entering university) than the students who do not participate in any institutional mobility program (45%). Moreover, mobility students have more educated parents than the non-mobility group (71% and 60%, respectively).

Table 2 Summary statistics of participants and nonparticipants
Full size table

There are also other differences regarding the composition of the groups. For example, females represent 46% of mobile students and 49% of non-mobile students. Black/brown/indigenous students are 11% of the mobility sample and 14% of the non-mobility one. Mobility students also have more previous internal mobility experience and are 1 year younger than nonparticipants when entering university. Those figures suggest self-selection in the sample, meaning that participants and nonparticipants in mobility programs would differ even without treatment (Caliendo & Kopeinig, 2008). The self-selection challenge is well-known in the study abroad literature (Kim & Lawrence, 2021; Meya & Suntheim, 2014) and will be discussed in the next section.

Empirical Strategy

To reduce the possible bias due to the selection of mobility programs, the methodology chosen for the analysis is a combination of Propensity Score Matching (PSM) and Difference in Differences (DiD). The sections below explain how both techniques were used in this study.

Searching for a Group of Potential Applicants

The final control group for our analysis was selected using Propensity Score Matching within the sample of all non-mobile students. Propensity Score Matching is a very flexible statistical technique used for impact evaluation that can be applied in the context of almost any program as long as there is a group of non-treated units (Gertler et al., 2016). It works by comparing treated and non-treated units with a similar probability (propensity score) of receiving a specific treatment (Caliendo & Kopeinig, 2008; Gertler et al., 2016). As stated by Netz and Grüttner (2020), PSM has become a very popular technique in the international student mobility literature for several reasons. One reason is that, unlike many regression techniques, it forces researchers to reflect upon the process of selection into international mobility by identifying the factors increasing the probability of experiencing mobility. A second reason is that it has the advantage of only comparing very similar treated and not treated individuals. Third, by presenting a non-parametric method of causal inference, it makes no assumptions about how variables are distributed and what the functional form of their relationships is.

To identify potential mobile students within the group of non-mobile students, we considered as relevant matching characteristics the following: students’ demographic and family characteristics, previous internal mobility experience, students’ academic performance, and access to study abroad scholarships. To ensure that none of the variables could be affected by having participated in mobility programs (therefore biasing our results) (Gertler et al., 2016), all variables included in the propensity score calculation are either time-invariant or measured before any mobility could occur.

We considered gender, age when entering university, and race/skin color as students' demographic characteristics. Those variables were added to account for any possible systematic differences between students with different demographic characteristics concerning their choice of going abroad and academic performance.

As family characteristics, we included the income per capita of their household before entering university and their parent’s education. Those two variables were added to account for students’ socio-economic background since students from higher-income families may be more likely to pursue part of their studies abroad (European Commission, 2016; Junor & Usher, 2008; Meya & Suntheim, 2014). Additionally, first-generation college students have many responsibilities that compete with the university for time and attention, such as working full-time or being married (Eveland, 2020; Warburton et al., 2001). Parents’ education was also added to account for social capital, as highly educated parents might support an exchange financially and by highlighting the benefits of learning about other countries, languages, and cultures (Di Pietro, 2019; Meya & Suntheim, 2014).

Previous internal mobility experience was added because such an experience might affect students’ final grades. For example, students who have already left their social environment once may be more likely to move to another country and spend more effort finding the perfect match regarding university and field of study (Meya & Suntheim, 2014).

As students’ academic performance, we added the grades in the first semester of university and grades in the entrance exam. Academic performance at the university is the most important criterion considered by UNICAMP to select exchange students. Grades in the entrance exam were also added to account for students’ pre-university academic ability, as students who apply for mobility programs may be academically more able than others. Thus, pre-university grades may predict university success and measure students’ commitment (Meya & Suntheim, 2014).

Finally, we also accounted for access to scholarships to go abroad. During 2011 and 2015, as already mentioned, the Brazilian government implemented a massive exchange program called Science without Borders, which sent more than 90 thousand Brazilians to study abroad (Brasil, 2016). Since the program offered more scholarships for students in selected areas (such as Biological Sciences, Health, Exact, Technological, and Earth Sciences) that entered university between 2010 and 2014, dummies to account for the year of admission and area of the course were added.

We predict the propensity score using a binary Probit linear probability modelFootnote 12 and report the results in Table 3. In the model, the dependent variable is a binary that takes the value of 1 if the student participated in an institutional mobility program in the period between 2010 and 2020 and 0 otherwise. As independent variables, we consider: grade in the first semester; student’s pre-university academic ability; income per capita of household before entering university; education of the parents; gender; race/skin color; age when entering university; previous internal mobility experience; year eligible for the SwB program; area eligible for the SwB program.

Table 3 Participation in student mobility programs (probit results)
Full size table

The results show that all variables, except for skin color and age, significantly impacted the probability of participating in a student mobility program. Higher grades in the entrance exam and in the first semester of university, high income per capita, more educated parents, previous internal mobility experience, and eligibility to the Science without Borders program are all associated with a positive effect on the conditional probability of being treated, holding all other regressors constant at their means. On the other hand, being female has a negative effect on the conditional probability of being in the treatment group.

After estimating the propensity scores for each unit of our sample, we then tested the balancing property of each observed covariate between the treatment and control groups, as well as the overall balance. The idea of checking the balance is to verify if there was a reduction in sampling bias achieved through matching.

The results presented in Table 4 indicate that there was indeed a reduction in the bias after matching. The first part of the table shows that the matching sufficiently balanced most observable covariates and reduced considerably initial differences of both treated and untreated. The second part of the table shows the results from comparing the joint significance of all matching variables in the Probit model. The Pseudo R2 of results after matching was much lower for the matched sample than for the unmatched one. Both the mean and the median of the absolute standardized bias have been reduced substantially. Additionally, Rubins’ B (the absolute standardized difference of the means of the linear index of the propensity score in the treated and non-treated group) and Rubin’s R (the ratio of treated to non-treated variances of the propensity score index) felt within the bounds suggested by Rubin (2001). Those results indicate that the samples became sufficiently balanced after matching.

Table 4 Balancing results before and after matching
Full size table

While propensity score matching can be a powerful tool, it relies on several assumptions to produce reliable results. The two main assumptions are discussed below.

Propensity Score Matching Assumptions

Conditional Independence (CI)

The Conditional Independence assumption (also called unconfoundedness or selection on observables) states that differences in outcomes between treated and comparison individuals with the same values for pre-treatment covariates are attributable to treatment (Caliendo & Kopeinig, 2008). The main challenge with the CI is that it is a very strong assumption and cannot be tested. Since it is crucial to match based on the characteristics that determine participation, it is essential to understand the criteria used for participant selection (Gertler et al., 2016).

In the case of our sample, we believe that the most important pre-treatment characteristics to determine participation in mobility programs were included in our model. At UNICAMP, the selection criteria for student mobility programs are overall well established, as mobility students must: (1) be a regular student at the university; (2) have completed between 25 and 85% of the course load at the time of application and attended at least two semesters in their undergraduate program; (3) have a ‘profile of excellence,’ based on good academic performance; (4) have the application approved by the course coordinator; (5) meet the requirements requested by the destination institution.

Criteria 1 and 2 were met for all students in the dataset, as all of them were regular, started university before 2018, and completed at least their first year at university. Criterion 3 was measured by the grade in the 1st year of university and the student’s pre-university academic ability (grades in the entrance exam). Criterion 4 was not directly observable, as there was no feasible way to know if the coordinator would have approved the application of a non-mobility student if they had asked for it. Therefore, we assume that the coordinator's approval was conditional on good academic performance. Criterion 5 varies from student mobility programs but usually relies on academic performance.

Since Criteria 4 and 5 were not directly observed in our dataset, we looked for other possible ‘hidden’ criteria that may have affected both participation and the outcome of interest by adding socio-economic and demographic variables in the model. Even if they were not directly considered in the selection process, they might still have affected students’ motivation to apply for an exchange program. They could also be related to student’s final grades. Besides, those characteristics could also have indirectly affected the course coordinator's approval (for instance, if there was any prejudice in the selection regarding skin color, gender, or socio-economic status). Finally, we also added two variables to account for eligibility to the Science without Borders program since those eligible students had more choices of scholarships and destination countries.

Common Support

The second assumption of PSM is called common support (or overlap). For Propensity Score Matching to produce estimates of a program’s impact for all treated observations, each treatment unit must be successfully matched to a non-treated unit (Gertler et al., 2016). In practice, however, it may be that for some treated individuals, there is no untreated with a similar propensity score (which is called lack of common support) (Gertler et al., 2016). The common support assumption says that persons with the same characteristics (X) have a positive probability (P) of being both participants and nonparticipants of the program (D) (Heckman et al., 1999). The assumption can be written as follows:

$$0 < P\left( {D = 1{|}X} \right) < 1$$

Several ways are suggested in the literature to validate this assumption. However, the most straightforward one is a visual analysis of the density distribution of the propensity score in both groups (Caliendo & Kopeinig, 2008). Figure 2 shows the distribution of the propensity scores for both the treatment and control groups in the sample. As expected, control units had their distribution of propensity scores more skewed to the right compared to the treated units. The graph shows that the common support assumption was satisfied, with 99.8% treated observations within the common support area.

Fig. 2
figure 2

Distribution of the propensity scores for treatment and control groups (Common Support Assumption). Data source Authors’ estimation from UNICAMP’s microdata

Full size image

Difference in Differences Estimation

Since baseline data on our outcome of interest (student performance) was available, we decided to combine the matching with a Difference in Differences estimation, a method that compares the changes in outcomes over time between treated and non-treated units (Gertler et al., 2016). The advantage of combining both methodologies is to reduce bias since the combination controls for observable differences between groups and solves the issue of any unobserved characteristic constant across time between both groups (Caliendo & Kopeinig, 2008; Gertler et al., 2016). This combination is useful as selecting a control group using PSM can only tackle observed selection into international student mobility, not dealing with selection bias occurring from unobserved heterogeneity between individuals going abroad and staying at home (Netz & Grüttner, 2020).

We explore the impact of student mobility programs on student academic performance as measured by the average treatment effect on the treated (ATT) students (those who benefited from a mobility program). The ATT for our main outcome variable before and after participation (\(\mathrm{\Delta Y}\)) can be formally specified as follows:

$$ATT= E\left({\Delta Y}^{T}|D=1\right)-E\left({\Delta Y}^{C}|D=0\right)$$

where \({Y}^{T}\) denotes the potential grades for the treated individuals; \({Y}^{C}\) denotes the potential grades for the non-treated individuals; D is a dummy variable for student mobility status; and E() denotes the mathematical expectation operator.

Our model is given by:

$${Y}_{it}={\beta }_{1}+{\beta }_{2}{treatment}_{i}+{\beta }_{3}{time}_{i}+ \gamma ({treatment}_{i}*{time}_{i})+ {X}_{i}+{\varepsilon }_{it}$$

where \({Y}_{it}\) stands for grades of student \(i\) at time \(t\); \(treatment\) is a dummy variable that takes the value of 1 if student \(i\) participated in a student mobility program; \(time\) is a dummy variable that takes the value of 1 at the end of the student’s \(i\) course; \(treatment*time\) is the interaction between the treatment variable and time; \({X}_{i}\) is a set of individual pre-treatment covariates of student \(i\) in time \(t = 0\); and \({\varepsilon }_{it}\) is the error term. \(\gamma\) is calculated by the model and represents the average treatment effect in a Difference in Difference estimation.

To combine DiD with PSM, the regression used weights derived from the propensity score,Footnote 13 and considered only the region of common support, i.e., where there is overlap in the propensity score distribution for both treated and non-treated students.

The combination of PSM and DiD is the best possible methodology that could be used in our setting. The rationale for using quasi-experimental methods for this analysis is mainly because doing an experimental framework (such as a Randomized Control Trial), where students are randomly assigned to study abroad (as in a lottery), was not feasible in our case. Moreover, since at UNICAMP there is no threshold at which students become automatically eligible to participate in student mobility, empirical strategies like regression discontinuity designs also cannot be applied. In fact, UNICAMP has several different mobility programs, and students are not restricted to only applying to one of them.

Results and Discussion

Impact of Mobility Programs on Academic Performance

Results from the Kernel-based propensity score matching difference in differences (Table 5) show that, overall, participation in international student mobility programs does not significantly increase students’ standardized final grades.

Table 5 Average treatment effect on the treated
Full size table

As stated in "International Student Mobility And Students’ Outcomes" section, most existing studies on the impact of academic mobility find that students benefit from mobility programs. However, there is no full convergence of results in the literature regarding the impact on grades. Researchers on this topic agree that the impact of a mobility program on students is context-specific and varies across mobility programs and students’ characteristics.

For that reason, in the next subsections, we investigate the possible heterogeneous impacts of student mobility programs on academic performance across different subgroups of students. Two main questions guide our analysis: (1) does the impact vary across students traveling in different periods of their undergraduate courses?; (2) does the impact vary across programs with different durations? Additionally, we also investigate possible economic and demographic heterogeneous effects and effects related to the destination region.

Does the Impact of Student Mobility on Student Performance Vary Across Students Traveling in Different Periods of Their Undergraduate Program?

To answer the first question, we disaggregate the effects of student mobility by three different types of students based on the time of the mobility experience (measured by the time elapsed between the starting year at the university and the year of the first mobility).

In Brazil, most undergraduate programs last for eight semesters (4 years), which may vary according to the schedule offered by the institution and upon request for an extension. Based on the structure of Brazilian undergraduate programs, we identify three types of students:

  • Type I: students who traveled at the beginning of their undergraduate studies. UNICAMP does not allow students to participate in international institutional mobility during their first year. Considering that just a few students traveled between the first and the second year (only 3%), those who attended university for one or 2 years before mobility were considered Type I;

  • Type II: students who traveled in the middle of their undergraduate studies (3 years after starting university);

  • Type III: students who traveled closer to the end of their undergraduate studies (more than 3 years after starting university).

Looking at the distribution of the students in our sample by the number of years before the first international mobility, most students at UNICAMP can be considered as Type II (46%), while 38% are Type I and 16% Type III.

Considering the above three student types, Table 6 reports the results from the kernel-based propensity score matching difference in differences analysis. While negative effects on grades are found for those who traveled at the beginning of university (− 0.05 points), positive and significant effects are found for students who traveled closer to the end of their courses (0.06 points).Footnote 14 Those results suggest that the time of mobility matters when it comes to increasing final grades.

Table 6 Average treatment effect on the treated by student type (students who traveled at the beginning of the university, in the middle or at the end of their courses)
Full size table

At UNICAMP, most of the grades obtained abroad are registered as proficiency, therefore, not incorporated into the student’s Performance Coefficient. This rule guarantees that differences in grades are due to changes in students’ performances and not due to different grading systems at the host institutions. With that in mind, a possible explanation for our results can be found in students’ behavior. Students in their first university years are still adapting to university life, taking more courses, learning about their courses’ challenges, and familiarizing themselves with their peers. By traveling at the beginning of their courses, students may suffer from a twofold adaptation challenge: adapting to university and a different country.

Moreover, traveling before being wholly integrated into their home universities may impose difficulties in re-entering the home education system when returning, impacting exam performance. On the contrary, those who travel closer to graduation are older and may have a more mature mindset. Those students are already more integrated into university life and most likely have a clearer idea of what they expect from their degrees, which may affect their grades positively. Currently, UNICAMP’s data does not allow testing of these mechanisms, and further research should address those aspects.

While the choice of the cutoffs for distinguishing the three types of students was based on the structure of undergraduate courses in Brazil, in "Changing cutoffs" section we report a sensitivity analysis of our results to our cutoff choice.

Does the Impact of Student Mobility on Student Performance Vary Across Programs with Different Durations?

To answer the second question, we disaggregated the effects by three different mobility types based on the duration of the mobility program (measured by the time elapsed between the starting and the ending date of the exchange period).Footnote 15 The thresholds were chosen based on the structure of the courses at UNICAMP, where the academic year is split into two academic semesters. Consequently, the majority of the academic activities in the university (such as internships, courses, and most exchange programs) are offered for at least one academic semester. We considered the following three types of students:

  • Type A: students who experienced short-term mobility (up to one semester);

  • Type B: students who experienced mid-term mobility (one semester to 1 year);

  • Type C: students who experienced long-term mobility (more than 1 year).

In our sample, 26% experienced short-term mobility, 27% stayed abroad for more than 1 year, while the remaining 47% experienced mid-term mobility.

Results from the estimations (Table 7) indicate that while international mobility positively and significantly impacted students who participated in programs lasting from one semester to 1 year, negative effects were associated with shorter periods abroad. That suggests that mobility duration also plays a role in academic performance. On average, students who participated in mid-term programs experienced an increase in their final grades of 0.08 points, while students spending shorter periods abroad had a decrease of 0.1 in their last semester grades.

Table 7 Average treatment effect on the treated by student type (students who stayed abroad for a short, mid-term, or long period)
Full size table

Those results may be explained by the fact that short-period stays can distract students since adapting to a new country and a different higher education system usually takes some time. Therefore, spending more time abroad gives students more chances to re-evaluate their relationship with their courses.

While more extended stays may be needed if students want the benefits of mobility programs to enrich their academic curriculum, there seems to be a threshold where students stop benefiting from mobility (after 1 year). The fact that long-term programs do not positively impact students’ grades could be related to the fact that students may face challenges in readjusting to their home universities after spending a long time abroad. However, additional research is still needed to test those hypotheses empirically.

Other Heterogeneous Effects: Economic/Demographic and Destination Country

In addition to the subgroups described above, we also disaggregated the analysis by some pre-treatment economic and demographic variables, such as gender, skin color/race, parent’s education, and income per capita (Table 8), and into region and language of the destination country (Table 9).Footnote 16

Table 8 Average treatment effect on the treated: economic and demographic heterogeneous effects
Full size table
Table 9 Average treatment effect on the treated: region of destination
Full size table

Our estimations suggested that, while there seem to be no differences between students coming from different economic and demographic settings, there are differences between students by destination countries. A positive impact on grades was found for students traveling to North America (the United States and Canada), Oceania (Australia and New Zealand), and English-speaking countries. In contrast, negative impacts were associated with students traveling to Portuguese-speaking countries (i.e., with the same language spoken in Brazil).

The discussion about the role of the country of destination and the selection of universities based on language skills is not new in the Brazilian literature on student mobility. For instance, in a study about the Science without Borders program at the University of Campinas, Granja and Carneiro (2020) mentioned the case of Portugal, saying that despite the preference of Brazilian students to study in Portuguese universities (at the earlier stages of the program one out of five fellows chose Portugal), public calls to the country were officially canceled in the following years, when it became clear to policymakers that students were choosing Portugal due to its language. That is because applying for an exchange program to go to Portugal usually does not require knowledge of another language other than Portuguese. In contrast, calls for countries where Portuguese is not the primary language typically require proof of language proficiency.

Even though our data does not allow us to test analytically if the observed country heterogeneity is explained by the language spoken, data on English proficiency at entry in the university programs seems to confirm that those students who chose a Portuguese-speaking language destination country are those students who had lower grades in English in the university admission exam (Fig. 3).Footnote 17 They also had slightly lower grades in the entrance exam, on average (Fig. 4), and lower income per capita when entering university (Fig. 5). We might assume that those students are either less committed or have had fewer opportunities to learn a second language. On the other hand, studying in English might result from strategic thinking, a willingness to invest extra effort, and an ambition to have a prestigious institution mentioned in the curricula. Further investigation, however, is still needed in that regard.

Fig. 3
figure 3

Distribution of English grades in the university entrance exam by the language of the destination country. Data source Authors’ estimation from UNICAMP’s microdata

Full size image
Fig. 4
figure 4

Distribution of general grades in the university entrance exam by the language of the destination country. Data source Authors’ estimation from UNICAMP’s microdata

Full size image
Fig. 5
figure 5

Distribution of income per capita when entering university by the language of the destination country. Data source Authors’ estimation from UNICAMP’s microdata

Full size image

Robustness Checks

Subsample Results

A possible concern that may arise in our analysis regards the internal validity of the results due to the sample selection since our sample included both students who completed their courses and those who abandoned university/were dismissed. The latter group was considered in the sample because dropping a course or being dismissed from the university may directly correlate with the student’s grades. Since students who graduated may differ from those who did not complete their courses, which could correlate both to the treatment assignment and students’ final grades, we ran a robustness check considering only the subsample of graduated students. Results are shown in Table 10.

Table 10 Average treatment effect on the treated robustness checks: subsample of students who completed their courses
Full size table

Results show that our results are overall robust to the sample selection. Considering the full subsample of students who completed their courses, participation in international student mobility programs does not significantly increase students' overall standardized final grades. However, the temporal dimension still plays a role in changing grades. While negative effects on grades are found for those who traveled at the beginning of university, positive and significant effects are found for students who traveled closer to the end of their courses.

We also find that the only students who benefit from mobility are those who experience mid-term mobility. Short-term mobility, as well as long-term mobility, are detrimental to students. Therefore, our main conclusions regarding the temporal dimension of mobility are consistent with the main findings reported previously in "Impact of mobility programs on academic performance" section. The only difference is that the negative sign of long-term mobility turns significant in the subsample of students who completed their courses, while it is insignificant in the original model.

Changing Cutoffs

Another concern that may arise in our analysis is the sensitivity of our results to the choice of cutoffs for the heterogeneity analysis, especially regarding the timing factor (period elapsed between the starting year at university and the year of the first mobility). To check robustness to different cutoffs, we recalculated the average treatment effect on the treated for different specifications. In the first specification, we grouped together the students who moved after 1 or 2 years after starting university, while those who traveled in the remaining years (3, 4 and 5) were grouped as a second category. In the second specification, students moving after 1, 2 and 3 years were grouped together, while students going abroad during their 4th and 5th year were considered as a separate group. Lastly, we calculated the impact for all years individually. Results from our estimation showed that changing the cutoffs did not affect our main conclusions. Overall, students traveling later in their courses benefit more from mobility, while those traveling closer to the beginning of their courses benefit less.

Conclusions

In this paper, we evaluate the impact of international student mobility programs on academic performance (measured by students’ grades), focusing on the temporal dimension of those programs. We address two main sub-questions: (1) Does the impact of student mobility on student performance vary across students traveling in different periods of their undergraduate courses?; and (2) Does the impact of student mobility on student performance vary across programs with different durations? To the best of our knowledge, this is the first paper to address the temporal dimension of the impact of student mobility on undergraduate students’ academic performance. It is also the first to focus on Brazil.

To address these research questions, we use microdata shared directly by the University of Campinas, one of Brazil’s most internationalized universities. The average treatment effects on the treated are calculated using Propensity Score Matching combined with Difference in Differences to minimize the selection problem.

Our results suggest that both the time of mobility and duration matter for student performance. While negative effects on grades are found for those students who traveled at the beginning of university, positive and significant effects are found for students who traveled closer to the end of their courses. Regarding duration, we found that mobility duration also plays an important role in academic performance. On average, while student mobility positively impacts students who participated in programs lasting from one semester to 1 year, negative effects are associated with shorter periods abroad.

Overall, our analysis presents empirical evidence that can be used to design international student mobility programs, providing insights to policymakers engaged in maximizing the effects of their programs. For example, focusing on 1-year programs and targeting students after their third year of university may be good strategies to enhance academic performance.

Our results also suggest that, while there seem to be no differences between students coming from different economic and demographic settings, there are differences between students by destination countries. However, additional research is still needed in that regard.

This study is not exempt from limitations. Regarding the strategy used, the matching between treated and not treated students can only be performed based on observed characteristics, requiring the strong assumption that no unobserved differences in the treatment and comparison groups are also associated with the outcomes of interest. We minimized this bias by adding different covariates to estimate the propensity score and the final model. The long time span and the detailed information shared by UNICAMP’s administration allowed for a robust matching. Furthermore, we also combined PSM with DiD to account for any unobserved characteristics that were constant over time.

Additionally, due to data constraints, it was not possible to analytically test the mechanisms behind the results of the heterogeneity analysis, in particular, the findings on the temporal dimension and destination region/language. As a future research agenda, we believe that understanding the processes behind the heterogeneity of results is key to providing improved recommendations for program design. For that, it would be valuable to have more detailed data on (a) the country and institution where the student traveled to; (b) students’ motivations for participating in an exchange program and for the choice of the destination university; (c) activities carried out abroad (including the list of courses taken at the host university and the received grades); (d) academic challenges that the students faced both during and after traveling; and (e) language proficiency in languages other than English immediately prior to traveling.

Finally, in this paper, we focus only on academic performance. Even though we believe that student academic performance is a valuable indicator of human capital, individual, institutional, and national outcomes should also be considered when designing an academic mobility program. Those factors include but are not limited to student employability, university improvement, and national development. Further research is needed to capture the effects of student mobility on those dimensions, both in the short and long run.