The Bright Futures Survey (http://brightfutures-project.com/technical-report/) is a multi-country survey of students enrolled in tertiary education in China, Germany, and the UK. The questionnaire was carried out in Mandarin Chinese, English, and German with about 8,000 students in all three countries. The fieldwork was conducted in 2017 and 2018, using different sampling strategies in Europe and China. After thorough research into how Chinese international students are sorted across universities in their chosen destinations (Cebolla-Boado et al., 2017), a two-stage sampling logic was adopted in Germany and the UK. Universities were first stratified into groups according to ranking and number of Chinese students enrolled in each institution to ensure that students from different types of universities were appropriately proportionally represented. Within each university selected, random samples of undergraduate and master's students of Chinese and native backgrounds were obtained and individually invited to answer the questionnaire online. For China, the sample was stratified to cover different provinces in the north, south, and east of the country and take into account university prestige.

Table 1 describes the sample sizes for each of the analytic groups in this paper: international Chinese students in tertiary education in the UK and Germany, Chinese students studying in China, and two control groups of British and German students that we use in the main descriptive analysis. Note that the samples of British and German students are not representative of the universe of tertiary students in the UK and Germany but only of native-born students matriculated in universities in which Chinese students are also matriculated.

The British sample of Chinese students is fully representative of the universe of Chinese students in the UK and implements the sampling approach successfully, covering the entire universe of British higher education institutions and Chinese students across them. Similarly, the British sample is representative of native-borns enrolled in those higher education institutions in which Chinese in our sample are studying. The German sample of Chinese students did not cover the entire universe of Chinese students in Germany, and as such cannot be considered fully representative, and the sample of native Germans is smaller than that of Chinese. We found, however, that the differences between the UK and Germany are unremarkable in terms of our interests and the independent variables we use in the following analyses. In sum, our comparisons between China and the UK are based on representative samples, while we present results using the German sample to increase the robustness of our results. Note that replicating the analyses we present here without the German sample provides identical substantive results.

Socioeconomic background is commonly considered when studying observable migrant selectivity. We use the father's occupation (e.g., a dummy combining professional, technical, and high-level administration vs. the rest) and education (whether the father is a university graduate) as socioeconomic background variables. However, our main interest in this paper is to delve into a less commonly studied aspect of migrant selectivity, namely, selection on unobserved characteristics. There is not a large research tradition investigating differences between migrants and non-migrants in terms of unobserved characteristics. Following our argumentation, we use four different aspects that look at agentic individuals as embedded in broader educational frameworks. Bright Futures Survey included four questions asking students if someone who “thinks up new ideas” (creative), someone who “makes their own decisions” (independent minded), someone who “looks for adventures and taking risks” (risk-taker), and someone who “values being successful” (achievement oriented) was “not at all like s/he,” “somewhat like s/he,” “neither like s/he nor unlike s/he,” “somewhat like s/he,” or “very much like s/he.” It is important to note that we understand the individual characteristics represented in each question as self-perceptions and representations, much shaped by broader educational scripts and frameworks, rather than inner and habitual personality traits as suggested in the psychological literature. Given the strong socialization role of education and broader societal expectations of self-development, however, it is possible that the gap between self-perceptions and habitual personality traits may well be narrow among the population we are focusing on.

The survey questions above are the main variables used in our empirical sections focusing on unobservable selectivity. In our analysis, with answers to each question recoded into dummies, taking the value of 1 for the first two categories of answers (“somewhat” and “very much like”) and 0 for the remaining three, we estimated separate models. In this exploratory stage, linear probability models and logistic regressions were estimated to capture the average answer given by students in all five of our analytic categories (Chinese in China, the UK, and Germany, and both groups of European students). The Appendix includes four sets of overlapping histograms (Figures A.1–A.4) in which the distribution of each of these variables (individual traits) is compared across groups (Chinese in China and the UK, Chinese in China and Germany, Chinese in the UK and British, and Chinese in Germany and Germans).

As a second step, we merged all four variables into a synthetic index of agentic individual characteristics (results of the principal component analysis are presented in the Appendix, Table A.1. Figure A.5, also in the Appendix, plots the distribution of the resulting index). This continuous variable is used as the explicandum of a doubly robust treatment effect model (Linden et al., 2016) using inverse probability weighting with regression adjustment (IPRWA), in which the treatment takes the value of 1 if the respondent is a Chinese student who migrated to Germany or the UK and 0 for Chinese students in China. The advantage of estimating IPRWA treatment effects is that one can first model selection into treatment (in our case migration status) and then match comparisons from treated and control groups to measure the average treatment effect that, in our analyses, corresponds to the net differences in our index of agentic individual characteristics for migrants and non-migrants. It is possible that students from highly educated families are more likely to be exposed to agentic individual ideals (as these ideals spread through education) and are thus more likely to migrate. In consequence, we expect family educational background to play an intermediary role in positive selection on observed characteristics of those Chinese students who go abroad. Furthermore, given that studying abroad brings financial costs, we expect that there is also selectivity on the basis of parental occupational background. In our treatment effects estimation, we use a probit regression to model selection into migration using father's education (1: university degree; 0: other) and father's occupation (1: professional, technical and high-level administration; 0: other). We also use a recalled proxy of ranking in high school (1: if the student reports being in the 5th percentile; 0 otherwise). These three variables, measuring selection on basic observables (i.e., social background and previous performance), which the literature argues to be relevant for educational migration, help us discount from the association between migration and individual characteristics. Our model also controls for the propensity to migrate among Chinese students in rural or urban settings in China (1: rural). Finally, note that our multivariate model further controls for student gender (1: female). The Appendix includes a table (Table A.2) reporting the basic description of all variables included in this analysis.

The first empirical results correspond to the differentiated effect of student groups on individual characteristics. This is done using unconditional linear probability models (LPM)². Figure 1 (obtained from models shown in Table A.3 in the Appendix) summarizes these results looking at average responses given across groups. Throughout the panels, it can be seen that differences across groups of students are, if anything, modest, not to say non-existent. The test for selectivity requires comparison of outcomes across Chinese students in the countries included in the analysis: China, Germany, and the UK. Doing so reveals that there are no major differences across respondents from this national origin by country of survey. While there are some signs of statistically significant difference between Chinese across groups, these are of a small substantive importance and do not consistently go in the direction the selectivity hypothesis would suggest. Compared to their national counterparts in origin, Chinese in the UK are 3 and 5% more inclined to identify themselves as independent minded and risk-taker, respectively; however, they also report that they are 5% less achievement oriented (in the case of Chinese in Germany, 12%). On the other hand, Chinese in the UK and Germany are equally inclined to say that they are as creative as Chinese in China, and we see no differences between Chinese in Germany and Chinese in China in the likelihood of reporting being independent minded and risk-taker.

Overall, our results fit better the agentic individual hypothesis. The hypothesis suggests that the distribution of individual characteristics across student nationalities and countries of survey should be similar, since transnationalization implies a diffusion process whereby the model becomes taken-for-granted independently of national contexts (Soysal Nuhoḡlu, 2015). Thus, we expect respondents' self-definitions to converge. This is indeed what we find as graphically summarized in Figure 1, obtained from unconditional LPM. While Europeans on the whole score slightly above Chinese respondents, the differences are small in size. In the first panel, around 80% of Europeans indicate that someone creative is “very much” or “somewhat” like them, while this figure is around 70–75% among Chinese students in all three countries. Similarly, a small gap between Chinese and Europeans also appears in the second panel; around 80–85% of Chinese and 90% of Europeans recognize themselves as independent minded. The third panel, where being a risk-taker is the object of interest, is the only one in which we find some differences between the student groups. While 45–50% of Chinese students in Europe and China fall into the risk-taker profile, a similar percentage to that displayed by German students, a higher percentage of British students, 65%, identify as such. While this is a considerable gap, the broad similarities between Chinese and German students are in line with the predictions of the agentic individual hypothesis. Finally, the fourth panel fully fits with the expectation that all students score similarly across national origins and countries of survey−70–80% of students in all categories claim to be achievement oriented. In other words, this preliminary and unconditional analysis suggests that there are no clear grounds for arguing that there is selection on unobservables among students in tertiary education who migrated to other countries to pursue their degrees. It is rather the opposite; a remarkable homogeneity in how students perceive themselves dominates, which suggests that conceptions of the self are rather transnationalized among individuals who have already made it into tertiary education. This finding points to the increasingly standardized nature of university students across higher education contexts.

The standardization of the agentic individual model among university students stands clear when we further disaggregate the analysis by gender. The Appendix includes a replication of these plots, splitting the sample by respondents' gender in order to discard the possibility of agentic individual characteristics being patterned differently across different groups of students in terms of gender (Figure A.6). The plots confirm that, unlike predictions of selectivity hypothesis, we do not find any systematic differences neither between Chinese students in different countries nor between Chinese and native-born students in European destinations in regard to agentic characteristics. Compellingly, male and female respondents do not differ in the importance they attribute to these characteristics when describing themselves.

After merging these different components of the agentic individuality into a synthetic dimension using principal component analysis (see Table A.1 in the Appendix), we confirm the remarkable similarities in the distributions of this factor across our analytical categories in Figure 2. This provides a more intuitive visual confirmation of our second working hypothesis. In the multivariate analysis that follows comparing Chinese students in China and Europe, we use this factor as our dependent variable of interest.

While our preliminary and unconditional conclusions already suggest that higher education leaves no room for migrant selectivity, it is necessary to discard systematic composition effects associated with migration in order to attain a more conclusive view on whether migration does indeed signify positive selection of individuals on unobservables. Isolating the effect of migration on any specific individual characteristic or behavior ideally requires longitudinal multi-sited data that link countries of origin and destination in order to identify the distinctive features of migrants and native-borns in origin and host societies. These distinctive features could be the result of the fact that individuals who migrate, even before they make the move, can be systematically different to those who prefer to stay. Alternatively, the systematic differences between migrants and native-borns in origin may result from the very nature of the formers' experience of migration. Detecting which of these two possibilities applies in our case is a complex task particularly with the kind of observational, cross-sectional data that we use. Nonetheless, treatment effects and other quasi-experimental research methods allow us to model differences between treatment and control groups (i.e., migrants vs. non-migrants, in our case), controlling for a list of relevant regressors and selection into treatment. Specifically, selection into treatment (migrating being the treatment and not doing so being the control) is essential in order to disentangle whether differences by treatment status are due to selection or to experiencing migration. Only after modeling selection can we adjust the regression using other covariates and thus credibly confirm whether being a migrant implies any difference in the individual characteristics with which we are concerned when compared with non-migrants. In the lack of longitudinal data, this modeling approach represents the best alternative in estimating our effect net of selection into migration, since it circumvents the most important limitation inherent to cross-sectional observational data, that is, the non-random allocation of migrant status among migrants and non-migrants. Furthermore, treatment effects with inverse probability weighting is a double robust estimation method, which implies that the estimators are unbiased if at least one of the equations is correctly specified (Funk et al., 2011).

Table 2 shows our results. Let us first focus on the probit regression predicting the treatment status (migration). Chinese students in Europe and China were included in the analytic sample. As the literature on international education suggests a positive selection of international students by social background, we used two relevant predictors modeling this family condition: father's occupation in professional, technical, and high-level administration, and father's highest level of education being a tertiary degree. Research focusing on brain drain in migration suggests international students being positively selected on school performance. Thus, in our treatment equation, we included high school results (being in the 5th percentile of class in high school). We also control for whether the student comes from a rural or an urban setting in China. Our probit equation confirms that there is positive selection into international education by father's education and occupation, although these estimates are far from implying that all international students are from privileged social origins (having a father with a university degree or a father in the highest occupational social class increases the likelihood of being a migrant around 50%). Similarly, having a successful high school performance merely increases this probability by 10%. Students from rural settings in China are less likely to engage in educational migration.

Our model estimates the “average treatment effect” (ATE) associated with being in the treated group compared to the control, net of selection and adjusted by a number of controls including parental occupation, parental education, and student gender. The results from this more demanding approach to estimating effects using observational data show that there are no grounds for claiming that educational migrants are positively selected on unobserved characteristics. The treatment group (Chinese migrants) scored on average 0.08 more in the synthetic score of agentic individuality than the control (Chinese non-migrants); however, this effect is not statistically significant. In other words, there is no sign of migrant selection in agentic individuality, when the empirical focus is on highly educated populations.

Our synthetic factor of agentic individual personality is consistent among analytic groups. We have re-estimated our principal component analyses for each of them separately with identical results to those here reported. We have also re-estimated our analysis using a different sample of Chinese educational migrants in Japan with very similar results to those discussed in this paper. Our multivariate results are stable controlling for other potentially relevant individual characteristics such as age, year of education (1st, 2nd, or 3rd years), and level of studies (master's vs. undergraduate degrees).

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.