Domestic brain circulation in China: Impact on publication, citation, collaboration and university prestige

scientists’ mobility is shaped by a complex interplay of personal, institutional, and systemic factors. At the individual level, researchers may relocate due to personal considerations, such as family circumstances or career ambitions(Appelt et al., 2015; Pellens, 2012). Professional factors—such as access to better research facilities, funding, or job opportunities—also play a critical role. Disciplinary cultures and publication expectations vary across fields and can influence the perceived benefits of mobility(Azoulay et al., 2017).

At the institutional and national levels, policies and incentives can significantly affect mobility decisions. For example, performance-based recruitment practices and internationalization strategies can shape how and where scientists choose to move (Hu et al., 2020; Liu & Hu, 2021). The “brain drain/brain gain” framework emphasizes how scientific mobility affects both sending and receiving countries in terms of innovation and capacity building (Jałowiecki & Gorzelak, 2004; Robertson, 2006).

Different types of mobility further complicate these dynamics. These include international moves across countries, domestic mobility within a nation, and intra-institutional transfers. Each form of movement carries unique motivations and consequences (Amelina, 2013; Chepurenko, 2015; Zweig et al., 2020). The “scientific diasporas” framework underscores the value of transnational networks formed through mobility, which facilitate knowledge diffusion across geographic and institutional boundaries (Séguin et al., 2006; Tejada Guerrero, 2012)

Mobility can significantly influence research productivity and impact. Moving to a new institution often provides access to improved research infrastructure, expanded collaboration networks, and diverse intellectual environments (Li & Tang, 2019; Yin & Zong, 2022). These factors may enhance academic output in the long term, although short-term disruptions are also common.

Measuring academic performance in the context of mobility remains challenging. Traditional metrics—such as publication counts, citation rates, and the h-index—may not fully capture the transitional phases in a mobile researcher’s career (Bornmann, 2014; Haunschild & Bornmann, 2023). Moreover, changes in institutional affiliation can obscure the continuity of scholarly contributions and collaborations (Jonkers & Tijssen, 2008; Petersen, 2018).

Mobility may affect not only the quantity but also the nature of the research. Shifts in coauthorship patterns and exposure to new disciplines can lead to innovative, interdisciplinary work. Conversely, temporary declines in output or recognition may occur as researchers adapt to new environments (Conchi & Michels, 2014; Waltman & van Eck, 2012).

Finally, academic performance itself can act as both a consequence and a driver of mobility. High-impact researchers are more likely to receive attractive offers, while institutions use performance indicators to evaluate and recruit talent (Costas et al., 2010; Liu & Hu, 2021). This mutual relationship creates a feedback loop between scholarly success and career movement.

The data used in this study were obtained from OpenAlex (Priem et al., 2022), which includes information on scientific publications worldwide and is widely used in research on the science of science (Gates & Barabási, 2023; Nishikawa-Pacher et al., 2022; Venturini et al., 2023). Given the cases of Chinese names as represented in the Pinyin and the common situations where multiple individuals share identical names, the cases where they are affiliated with the same institution, study similar topics, and cite each other’s work are rare. The OpenAlex uses a sophisticated algorithm to conduct the name disambiguation for scientists, which integrates the name strings, affiliations, citation networks, and external resources including Crossref, PubMed, ORCID, etc., and thus minimized the errors of putting scientists with identical names as one scientist.

We then defined the selection criteria for moved and unmoved scientists as follows. To investigate the impact of scientist mobility on their scientific performance and collaboration networks in China, we first selected all papers published by scientists affiliated with Chinese institutions in the OpenAlex and filtered out the “moved” scientists who conducted mobility from 2014 to 2017 in China. The migration of scientists occurs when they publish two consecutive articles at different institutions, and the year of publication of the last article in the former employer institution is regarded as the year of movement. To ensure we captured a meaningful period of work before and after the move, we required that each scientist had at least one publication both before and after the mobility year, and that their publication activity spans a multi-year window around the move (typically 2 years pre- and 2 years post-mobility, when data permitted). Scientists with multiple institutional transitions during the 2014-2017 period were excluded to isolate the effects of a single relocation. For the unmoved group, we selected scientists who remained affiliated with the same institution throughout 2014-2017. We applied the same inclusion criterion of minimum active publishing years within the window. This resulted in a total sample of 2,896 moved individuals. The “unmoved” group, which served as the contender of “moved” scientists, was made up of approximately 89,529 scientists who did not conduct mobility during this period.

To clarify our outcome measurement, we define post-mobility research performance over a standardized four-year observation window following the identified mobility event (occurring between 2014 and 2017 to ensure consistent coverage). Using OpenAlex-provided yearlypubs (yearly new publications) and yearlycits (yearly new citations), we sum the new citations for papers published during this four-year window. To avoid bias from simple productivity differences—where more papers naturally yield more citations—we normalize total new citations by the number of new publications in the same period, providing an average per-paper citation impact.

For the unmoved group, we used the same calculation method but applied a fixed observation window from 2015 to 2018 for all individuals, ensuring consistent temporal comparability. We extracted yearlypubs and yearlycits data for this period and computed the normalized average citation impact in the same manner as for the moved group. This logic applies equally to both prior and post observation periods, ensuring that all performance metrics are based on observed and validated time spans.

Importantly, to address data sparsity and variation in author career timelines, we implement a dynamic filtering approach: we include only time windows overlapping with the author’s active publishing years and adjust the measurement window length accordingly when a full four years is unavailable (e.g. for late-career movers). Cases with no overlap in the observation window were assigned zero citations and publications. This ensures that outcome measurements reflect real post-mobility activity while minimizing truncation bias or artificial inflation from unobserved years. Full details of this algorithm are provided in the SI and illustrated with an example code.

PSM is a statistical technique widely used in observational studies to control for selection bias (Liu & Hu, 2021; Rosenbaum & Rubin, 1983). PSM is especially useful when there is no way to randomly assign subjects to different groups, such as in a non-randomized clinical trial or retrospective cohort study (Austin, 2008; Luo et al., 2010). The PSM method involves estimating the probability of each subject being assigned to a particular group, based on a set of observed covariates, and then matching subjects in the treatment and control groups who have similar scores. This helps to ensure that the groups are comparable for the observed baseline characteristics, reducing the potential for confounding. PSM can be used with both binary and continuous outcomes and has been shown to yield reliable estimates of treatment effects when implemented correctly.

The primary objective of this methodology is to reduce selection bias and improve the estimation by matching moved and unmoved groups based on their propensity scores. These scores were estimated using logistic regression or probit regression, two widely acknowledged models for estimating the likelihood of treatment assignment. The covariates X considered for propensity score estimation included research age, discipline, academic performance measurements (number of publications, citations, and collaborators in four years before the year of mobility for “moved” scientists or 2015 for “unmoved” scientists), DFC university, university prestige and geographical locations (region and part) of employer university (see Figure 1 for the list), which were selected based on their relevance to the career development of scientists and were noted as X₁ to X₁₀.

Figure 1.

Visualization of PSM procedure. We collected all papers published in Chinese institutions from 2014 to 2017 to identify scientists’ mobility and selected scientists who worked in two or more institutions and whose mobility time fell between 2014 and 2017. The control group consisted of scientists with only one employer university during the observed time, resulting in a data set of about 100,000 people. Then we used PSM matching methods to find scientists with similar prior measures to the “moved” scientist and obtained 2,586 pairs of matched individuals between moved and unmoved groups. The matched premovement variables include (1) research age (i.e. the number of years since his/her first publication). (2) # Publications. (3) # Citation. (4) # Coauthors. (5) DFC University (i.e. the university is the Double First-Class or not) (6) University prestige (the percentile measured by the average citation per paper hosted by this university). (7) region (i.e. Beijing, Guangdong, etc.). (8) part (i.e. East China, North China, etc.). (9) discipline (i.e. Biology, Business, etc.). (10) Project 985/211 (the different Project types of the institutions, 2 for both “Project 985” and “Project 211”, 1 for only “Project 211”, and 0 for no project). The e(X) represents the estimated propensity score for a scientist with 10 covariates. Then PSM uses those scores to choose similar “moved” and “unmoved” individuals by the nearest neighbor algorithm.

The propensity score, denoted as e(X), is calculated based on the logistic regression model for each scientist participating in the matching process. By setting the outcome variable as a binary indicator of treatment status (1 for “moved,” 0 for “unmoved”), the logistic regression model estimates the probability of each scientist receiving treatment based on the selected covariates. The propensity score signifies the likelihood of a scientist receiving treatment, considering their covariate values.

To ensure a high degree of similarity between “moved” and “unmoved” scientists after the PSM process, we employed a single nearest-neighbor approach to identify a matched “unmoved” scientist for each “moved” scientist. This procedure resulted in matched pairs with the most similar scores. Ultimately, utilizing the PSM matching method, we established 2,896 pairs of “moved” and “unmoved” scientists (See Figure 1).

Ordered Logistic Regression, also known as ordered multinomial logistic regression, is a widely used statistical model for handling ordinal categorical outcomes (Fullerton, 2009; Issa & Kogan, 2014). In our study, the logit function transforms the linear combination into probabilities and is defined as: 1logit(pi)=ln(pi1−pi)=β0+β1 Publications +β2 Citations +β3 Cpllaborators +β4 Geo.$\operatorname{logit}\left(p_i\right)=\ln \left(\frac{p_i}{1-p_i}\right)=\beta_0+\beta_1$ Publications $+\beta_2$ Citations $+\beta_3$ Collaborators $+\beta_4$ Geo.

Here, p_i represents the probability of observation i belonging to a specific response category, i.e. rank_i =0, 1, 2. β₁ to β₄ are corresponding coefficients to estimate. Geo is a categorical variable controlling for geographical information, such as regional division or administrative division in China.

The publication records of scientists, including publication time, authors, host affiliations, and topics, were collected from OpenAlex (Priem et al., 2022; Wang et al., 2019), according to which we retrieved scientists who shifted their employer universities in China from 2014 to 2017. During this period, the flow of scientists’ migration was classified by the origin and destination regions across the southern, northern, and eastern regions in China. As shown in Figure 2A, scholars in the northern region tended to relocate within the same region. Universities are not evenly distributed across regions, specifically, universities in the northern region are predominantly concentrated in Beijing, while in the southern region, they are mainly located in Guangdong and Hong Kong. As for the eastern region, major institutions are concentrated in areas like Jiangsu, Zhejiang, and Shanghai. So, most of the scientific circulations in China are among these provinces, as shown in Figure 2B.

Figure 2.

Chinese scientific flows at the scale of regions and provinces. A. The Sankey graph shows the scientific labor flow happening between 2014 and 2017 among regions in China (followed by Chinese administrative divisions). The left side represents the source regions, while the right side shows the destination regions. The width of a band represents the frequency of the labor flow. B. The directed circular chart shows the scientific labor flow happening between 2014 and 2017 among the main provinces in China. Only the top 20 provinces or municipalities with the highest frequency of scientists’ mobility are displayed. The width of the arc represents the frequency.

In addition, we created scatter plots of provincial inflows vs. outflows and provincial inflows vs. the number of DFC institutions within the province. By combining Figure 3A and Figure 3B, we observed that provinces with higher inflow ratios also tend to have higher outflow volumes, indicating a proportional relationship between the inflows and outflows of scientists at the provincial level. Furthermore, in Figure 3B, it can be seen that the inflow ratios of provinces are also approximately proportional to the number of DFC universities within the province. These two scatter plots illustrate intuitively that there is a strong correlation between inflows, outflows, and the number of DFC institutions within the provinces. These findings preliminarily suggest that provinces with high inflows of scientists tend to also have high outflows. Additionally, the ratio of inflows to outflows is strongly correlated with the proportion of DFC universities within the province. Furthermore, the patterns of inflows and outflows at prestigious institutions, such as DFC, are synchronous with provincial-level inflows and outflows.

Figure 3.

Chinese scientific flow in the scale of regions and university prestige. A. The scatter plot of Inflow vs. Outflow in Chinese Regions. B. The scatter plot of Inflow vs. Number of DFC institutions in Chinese Regions. C. The pie chart depicts the proportion of transitions between other institutions and DFC institutions.

Beyond exploring the relationship between the number of provinces and DFC institutions, we also note that the prestige of institutions may influence the mobility of scientists. This inevitably raises the question: when scientists move from lower-prestige institutions to more prestigious ones, are they readily accepted?

First, based on the information we collected about DFC institutions, we aimed to highlight the actual data proportions of mobility between DFC institutions and other institutions in the flow data. As shown in Figure 3C, the data indicating movement from other institutions to DFC institutions account for approximately 50% of the total, suggesting that moving to more prestigious institutions is not an isolated occurrence within the migration data. Consequently, Section 4.7 will consider which academic factors are more closely related to scientists moving to higher prestigious institutions.

We used Propensity Score Matching (PSM), a common method in observational studies (Austin, 2009; Hill & Reiter, 2006; Huang et al., 2023; Leuven & Sianesi, 2003), to find a comparative “unmoved” group of scientists who did not move, for comparison with scientists who moved between academic institutions from 2014 to 2017 in China (See Figure 1 for the schematic representation). We took several academic measurements, such as publication counts, citation counts, number of collaborators, the research starting years, and the university prestige into consideration to ensure that the “unmoved” group closely resembles the “moved” group in these key characteristics. The violin plots on the left of Figures 4A–C demonstrate the matching balance of covariate distributions between the two groups, and Figure 4D also shows the proportion of prestigious institutions that are similar before moving. The covariate balance achieved through PSM suggests that the “moved” and “unmoved” groups are now more comparable in terms of the measured covariates.

Figure 4.

Comparisons of multiple measures between moved and unmoved groups of scientists. Panel A-C illustrate the distributions (in logarithmic scale) of the number of publications, number of citations, and number of collaborators, and Panel D the level of employer institution (as a percentage) for the “moved” and “unmoved” group of scientists before and after the year of mobility. Panel E displays the estimated coefficients of differences between the two experimental groups obtained using the t-test.

After relocating to a new employer university, scientists did not to witness an enhancement of their productivity in the first four years after settlement (Figure 4A). However, Figure 4E indicated that “moved” scientists built up more collaborative relationships during this period, which is similar to the viewpoint proposed by Jonkers and Tijssen (2008), Kato and Ando (2017), and Liu and Hu (2022). Further confirmed in Figure 4C, after movement occurred, there was a significant difference in the distribution of the number of collaborators between the two groups. Notably, the “moved” scientists had a significantly greater number of collaborators than the “unmoved” scientists. This finding suggests that job mobility can indeed facilitate scientific collaboration by providing opportunities to expand one’s collaboration network with new colleagues.

Switching to another employer university can promote collaborations by offering access to new resources and a community of scholars. However, this also brings challenges for the moved scientists, as it requires adjusting to a new scientific environment and life culture. Additionally, relocating may result in a decline in citation impact for researchers. As shown in Figure 4B and Figure 4E, the citation impact of scientists who moved and those who did not move were similar before the relocation year, with no significant difference, however, the citation impact of the “moved” scientists was lower than that of the “unmoved” scientists after moving, as reflected in the greater inconsistency of their citation impact distribution, particularly at the 25th, 50th, and 75th percentiles. Therefore, researchers should weigh the benefits and drawbacks of relocating to a new employed university and consider the potential influence on their academic prestige before making the decision.

In Figure 4D, we show the proportion of scientists belonging to the DFC universities in the two groups of scientists before and after the movement. Since the scientists in the unmoved group did not change their affiliation before and after the movement, we mainly focus on the variation of the proportion of scientists belonging to DFC universities. We noticed that the proportion of DFC universities decreased in the destination of “moved” scientists, indicating the fact that academic institutions prefer to hire faculty members who graduated from or worked in more prestigious universities. This result is consistent with previous findings that show systematic inequality and hierarchy in faculty hiring networks (Clauset et al., 2015). Similarly, in China, scholars who have been trained or have worked in top universities are often considered to possess a higher level of knowledge and expertise in their respective fields. This perception arises due to the availability of abundant resources and rigorous academic training typically offered by top-tier universities, factors that hold significant value for academic institutions (Hu et al., 2020; Kwok et al., 2011).

To further explore potential heterogeneity in mobility outcomes, we divided the scientists into two groups based on their career stage. Using the average academic age (4 years) in our dataset as the threshold, we classified researchers with no more than 4 years of academic age as early-career scientists, and those with more than 4 years as late-career scientists. For each subgroup, we conducted a series of t-tests to compare pre- and post-mobility changes in academic performance across four dimensions: number of publications, citations, collaborators, and institutional level (DFC affiliation).

Consistent with Figure 4E, both early- and late-career scientists experienced a drop in citation impact and a rise in the number of collaborators after mobility, reinforcing the general finding that mobility helps new collaborations but may disrupt citation continuity. Similarly, the proportion of scientists in DFC universities declined after mobility in both groups, echoing the institutional prestige patterns shown in Figure 4D.

However, a key difference appeared in publication output: early-career scientists showed increased productivity after moving, possibly due to better research conditions, stronger institutional support, and greater adaptability in the early stages of their careers. Mobility may also give more opportunities for younger scientists to access resources, form new collaborations, and gain visibility in broader academic networks, all of which can foster research growth. In contrast, late-career scientists showed decreased output, potentially reflecting transition costs, timeconsuming adjustments, and disruptions to proven research teams and agendas.

These findings, building on the patterns in Figures 4 and 5, demonstrate that mobility’s academic consequences are not uniform but vary significantly by career stage—highlighting the need for more targeted support and incentive mechanisms, especially for senior scientists who may face higher switching costs, and for early-career researchers to maximize the benefits of their mobility.

Figure 5.

Comparisons between moved and unmoved groups of scientists with short and long tenure. A. The estimated t-test coefficients of differences between the two experimental groups of scientists with short tenure. B. The estimated t-test coefficients of differences between the two experimental groups of scientists with long tenure.

In this section, we provide a partial elucidation of the robustness of the PSM process. First, we explained the balance of the matching process. Table 1 presents a concise statistical summary of the data before and after PSM matching. Overall, we observed that the two groups of scientists showed significant differences before matching, while disparities disappeared after matching. This shows that the enhanced balance in our matched sample bolsters the validity of our previous analyses and strengthens the reliability of the estimated treatment effects.

Secondly, we considered potential endogeneity issues that may arise. Although the PSM matching process can reduce bias and the influence of confounding variables in the study, we still could not completely rule out the possibility that the transition to a new institution may impose different requirements on scientists, leading to changes in academic performance, which is the endogeneity issue we were concerned about. Therefore, more robust tests are needed. We randomly reshuffled the two groups of scientists obtained after PSM matching 100 times, re-divided them into mobile and non-mobile groups, calculated the mean differences in academic performance for each pair, and plotted the distribution. As shown in Figure 6, it is evident that except for the variable “#Publications”, the original mean differences in other academic performance variables for the two groups are not within the range of the reshuffled distribution. It explains the significant effect of changes in these variables before and after movement. To some extent, this helps to alleviate the impact of endogeneity on the experiment, indicating that mobility indeed plays a significant role in altering scientists’ academic performance.

Figure 6

Distribution diagram of endogeneity test for variables. The figure displays the distribution of actual mean differences (black dashed lines) between the two groups and the mean differences generated from 100 random allocations separately. The variable names corresponding to the distribution maps A-D are consistent with those in Figure 4.

The preceding analysis has revealed a decrease in the proportion of scientists moving to higher prestigious institutions after migration, which resonates with the results presented in Figure 3 of Section 4.1. Both findings indicate a certain level of resistance for scientists transitioning to higher prestigious institutions. Building on these insights, we sought to further explore the patterns of scientists moving from lower prestigious to higher prestigious institutions. Such transitions not only reflect individual career aspirations but also shed light on the distribution of resources and opportunities within the academic community.

Here we applied regression models to identify the potential factors. To achieve this, a ternary variable, rank, was employed to capture the difference in the level of the scientists’ employer universities before and after the move. This variable has three categories: a value of 2 indicates a transition from a non-DFC university to a DFC university, a value of 1 indicates the flat moves, i.e. from a non-DFC university to another non-DFC university or from a DFC university to another DFC university, and a value of 0 indicates a shift from a DFC university to a non-DFC university. Subsequently, an ordered logistic regression analysis was conducted on the outcome variable rank, considering the scientists’ pre-movement academic productivity (Model 1), citations (Model 2), and the number of collaborators (Model 3) respectively, while controlling for the geographical region in China (Regression formula is (1) in Method section). The findings are displayed in Table 2, which suggests that a higher number of citations and collaborators prior to the movement is positively associated with the likelihood of scientists transitioning to a higher prestigious university. Additionally, the scientists’ number of pre-existing publications is insignificantly related to the chance of transitioning to a higher prestigious university.

To further examine how individual research performance relates to mobility across institutional tiers, we conducted both separate and joint estimations using ordered logistic regression models. The results, presented in Table 2, show that when the three variables—publication count, citation impact, and number of collaborators—are entered separately (Models 1-3), citation impact and collaboration intensity emerge as significant predictors of upward institutional mobility. In Model 4, where all three predictors are included simultaneously, the results remain consistent: citation impact and number of collaborators continue to show statistically significant and positive associations with mobility outcomes, while the effect of publication count becomes negative and marginally significant.

These findings corroborate the patterns illustrated in our simulation-based plots, which depict how the probability of institutional upward mobility varies with different levels of research performance (See Figure 7). The convergence between simulation and regression results highlights the central role of research quality and professional network strength in determining mobility opportunities. In particular, institutions tend to favor candidates who not only demonstrate highimpact scholarship but also maintain extensive collaborative ties, as these traits are often associated with long-term potential in research leadership, interdisciplinary engagement, and contribution to large-scale academic initiatives.

Figure 7.

Estimated probabilities of institutional mobility by scientific performance indicators, based on separate and combined regression models. The x-axes represent the number of publications (A, B), citations (C, D), and collaborators (E,F). Panels A, C, and E show marginal effects from separate regression models, while Panels B, D, and F present results from a combined model including all three indicators. The y-axis indicates the predicted probability of scientists moving from a non-DFC university to a DFC one (rank = 2), making a lateral move (rank = 1), or moving from a DFC to a non-DFC university (rank = 0).