The effect of social media information on college major choices in China: a difference-in-differences analysis.

The effect of social media information on college major choices in China: a difference-in-differences analysis - PMC Skip to main content An official website of the United States government Here's how you know Here's how you know Search Log in Primary site navigation Search Log in Search PMC Full-Text Archive Search in PMC Journal List User Guide As a library, NLM provides access to scientific literature. Inclusion in an NLM database does not imply endorsement of, or agreement with, the contents by NLM or the National Institutes of Health. Learn more: PMC Disclaimer | PMC Copyright Notice Sci Rep . 2026 Apr 14;16:17279. doi: 10.1038/s41598-026-47742-w The effect of social media information on college major choices in China: a difference-in-differences analysis Lijing Ma 1, 2 , Bo Liu 3, ✉ , Yilong Wu 1 , Qingsong Li 1, 2 , Chuanhua Zeng 1, 2 Author information Article notes Copyright and License information 1 School of Automobile and Transportation, Xihua University, Chengdu, 610039 China 2 Yibin Institute in Xihua University, Yibin, 644000 China 3 School of Marxism, Sichuan Technology and Business University, Chengdu, 611745 China ✉ Corresponding author. Received 2025 Sep 10; Accepted 2026 Apr 2; Collection date 2026. © The Author(s) 2026 Open Access This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, which permits any non-commercial use, sharing, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if you modified the licensed material. You do not have permission under this licence to share adapted material derived from this article or parts of it. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by-nc-nd/4.0/ . PMC Copyright notice PMCID: PMC13234011  PMID: 41981123 Abstract The choice of a college major in China is a high-stakes decision often plagued by information asymmetry. This study provides a large-scale causal analysis of how social media influencers impact these choices. Leveraging the “Zhang Xuefeng Phenomeno”, an event where a prominent influencer vehemently discouraged Journalism and Communication majors, as a natural experiment, we employ a difference-in-differences (DID) design on a comprehensive dataset of university admission ranks from 2017 to 2023, covering over 3.3 million observations. Our findings reveal that the influencer’s negative signal caused the admission rank for Journalism and Communication to worsen by a substantial 13.1% across all institutions and by 24.7% at elite Double First-Class Initiative universities. A significant asymmetric response is identified, as the negative impact is nearly twice the magnitude of the positive effect observed for the endorsed Computer Science major, suggesting that student decisions are driven more by loss aversion than by the pursuit of opportunity. The robustness of these estimates is confirmed using event-study approaches and placebo tests in untargeted fields such as Philosophy and History. Furthermore, an exploratory analysis offers suggestive evidence of positive spillover effects on related practical fields, such as Transportation. A triple-difference (DDD) analysis reveals this impact is significantly weaker in provinces with higher per capita GDP. This research provides robust evidence that social media influencers have become powerful new players in the educational landscape, capable of significantly influencing students’ intentions to choose a major. It also highlights novel challenges for policymakers in mitigating information asymmetry and optimizing human capital allocation to support the development of new quality productive forces. Subject terms: Education, Mathematics and computing Introduction The choice of a university major is a pivotal decision that shapes the career trajectories, income potential, and social mobility of young adults 1 – 3 . In China, this decision is situated within the high-stakes context of the National College Entrance Examination, the primary gateway to university for millions of students each year. The sheer scale of this system creates a formidable challenge: in 2024 alone, the student cohort included 4.90 million enrolled in regular bachelor’s programs, 109,600 in vocational bachelor’s programs, and 5.68 million in higher vocational programs 4 . This massive group of students confronts a vast landscape of 816 distinct undergraduate majors across 93 fields of study 5 . The university application and admissions process serves as the core mechanism for matching students to these programs and majors. However, navigating this complex matching problem under conditions of uncertainty and information asymmetry is a significant hurdle for millions of families. This urgent need has given rise to a massive college application consulting market, reaching 1.09 billion yuan in 2025 6 . The consulting market demand highlights how the decision-making process, while influenced by foundational theories of cultural capital and social habitus 7 , 8 , is fundamentally information-dependent 9 – 13 . Students and their families actively seek information on program characteristics, career prospects, and labor market outcomes from various sources. In recent years, there has been a dramatic transformation in the information landscape, which traditionally included official university materials and national discipline assessments, but has increasingly expanded to include third-party rankings and social media platforms. The provision of guidance on social media can be academically defined as a form of information intervention that aims to encourage students to make better choices 14 . The literature investigating these informational effects can be broadly divided into two streams. One well-established stream confirms the powerful influence of structured information on students’ major choices. This includes information from official sources, such as national discipline rankings 15 and universities’ promotional materials 16 , as well as findings from controlled experiments involving data on career earnings 17 . In contrast, a more nascent stream of research is beginning to explore the impact of unstructured, subjective information from social media, acknowledging that advice from online influencers now constitutes a significant public signal in the decision-making process 14 , 18 – 20 . While the influence of online personalities on students’ major choices is recognized, the large-scale causal effect of such an information shock remains empirically understudied. Leveraging a natural experiment, specifically the “Zhang Xuefeng Phenomenon”, this study conducts the first large-scale causal analysis of a social media information shock on college major choices in China, utilizing a comprehensive dataset of university admission rankings from 2017 to 2023. Our contribution is threefold. First, we shift the research focus from traditional, official information sources to the increasingly dominant yet understudied realm of social media. By analyzing the impact of a prominent influencer, we quantify the causal effect of this new form of public signal on high-stakes educational outcomes. Second, we explicitly test for and identify a significant asymmetry in how students respond to information. By separately analyzing the effects on majors that were discouraged and those that were endorsed by the influencer, we find that negative signals trigger a much stronger response. We frame this asymmetry through the lens of loss aversion from behavioral economics. Our results confirm that loss aversion is a primary driver in the high-stakes college major selection process, providing large-scale causal evidence for this behavioral tendency. Third, through a triple-difference analysis, we investigate the moderating role of provincial economic context, offering a more nuanced understanding of how local conditions mediate the influence of large-scale information shocks. Our findings provide insights for students, universities, and policymakers navigating the complexities of the digital age. Literature review The choice of an academic major is a pivotal decision made under significant uncertainty, and a substantial body of literature has examined how informational interventions can shape these high-stakes choices. This research can be broadly organized into two paradigms. The first, rooted in economic models of choice, focuses on how students respond to structured, often centralized, information. The second, drawing from communication and sociology, grapples with the recent emergence of a decentralized, peer-driven information ecosystem powered by social media. Traditional information channels and rational choice The first paradigm traditionally models students as rational agents who update their beliefs in response to new information about labor market returns, program quality, and personal aptitude. Within this framework, research has focused on two types of interventions. The first is institutional signaling, which includes direct university marketing, career counseling, and government-led policy reforms, such as changes to admissions systems. These studies confirm that official, top-down information can enhance students’ decision-making efficacy and effectively coordinate choices on a large scale 9 , 12 , 21 – 24 . Complementing this institutional perspective, a second line of inquiry uses experimental and econometric methods to explore how students’ subjective beliefs influence their decisions. Pioneering work by Arcidiacono 25 demonstrated the importance of anticipated earnings and self-assessed ability. Subsequent studies have refined this view, showing that while providing accurate earnings data can correct misperceptions and lead to better financial outcomes 10 , students’ non-pecuniary preferences, such as intrinsic interest and unobserved “tastes”, are remarkably resilient and often dominate pecuniary factors 26 , 27 . A key takeaway from this literature is that while students are responsive to information, their choices are filtered through a strong prism of pre-existing personal motivations. The rise of social media: a new information paradigm This traditional model, however, is increasingly challenged by the rise of social media, which represents not merely a new channel, but a paradigm shift in the information landscape. Unlike official sources, social media information is decentralized, interactive, and highly subjective. Here, influence is wielded not by institutions, but by peers and online personalities whose credibility is socially constructed 28 . Emerging evidence highlights the profound impact of this new ecosystem. The “Zhang Xuefeng effect,” for example, illustrates how a single influencer’s commentary can trigger nationwide shifts in application patterns, demonstrating a powerful new form of public signal 14 , 19 . Broader research confirms that media exposure, both traditional and digital, consistently shapes students’ career aspirations 18 , 20 . These phenomena are well-explained by communication theories like the Elaboration Likelihood Model (which emphasizes source credibility) and social comparison theory (which explains peer influence), as well as concepts like group polarization and echo chambers that describe how online environments can reinforce biases 29 – 32 . In the high-stakes context of China’s college admissions, these dynamics appear particularly potent, with studies indicating that informational interventions and transparency can reduce major selection mismatches, that online information can help mitigate poor choices, and that perceived social support can foster academic self-efficacy 15 – 17 , 33 . Research gaps Despite growing acknowledgment of social media’s role, the literature is marked by several critical gaps. First, while the influence of this new information channel is widely observed, much of the evidence remains correlational or descriptive, leaving a need for rigorous, large-scale analyses capable of isolating the causal effect of a social media information shock from underlying trends. Second, while some descriptive analyses allude to asymmetric responses to information, there is a lack of rigorous quantitative measurement of this phenomenon. The relative magnitudes of how students react to positive information (endorsements) versus negative information (warnings) have not been empirically estimated. Third, although the concept of regional differences is acknowledged, the impact of national-level information shocks is often treated as monolithic. There is a significant gap in quantitatively estimating how these effects are moderated by regional heterogeneity, such as varying economic conditions or local labor market structures across provinces. The “Zhang Xuefeng Phenomenon” as a natural experiment The new signaling role of social media The college admission process is fraught with information asymmetry. While traditional information interventions have been well-studied, the information landscape has been transformed by the rise of social media. Social media platforms have become primary venues for disseminating guidance, giving rise to a new class of “educational influencers”. Unlike traditional media, information from these influencers has two distinct features. First, it is amplified through interaction. Direct engagement between creators and users enhances receptiveness in a way that one-way media cannot. Second, it carries pronounced subjectivity. In pursuit of engagement, influencers often add strong personal interpretations to factual information, which can be rapidly magnified by internet algorithms. Despite this subjectivity, these widely disseminated views are not private advice but function as powerful public signals, similar in their reach to traditional media. Among the most prominent of these influencers is Zhang Xuefeng, a commentator known for his employment-focused and often provocative advice. His role as a powerful information intermediary set the stage for a nationwide natural experiment in 2023. Defining the natural experiment The public interest in educational influencers’ commentary is highly seasonal, consistently peaking during the May-June college application window each year. In May and June of 2023, immediately preceding the crucial college application window, Zhang’s remarks received extensive public attention and circulated widely on social media platforms. He strongly discouraged students from choosing Journalism and Communication majors, while simultaneously endorsing majors he perceived as yielding high labor-market returns, such as computer science. This advice became a topic of national conversation precisely when students were finalizing their decisions 34 . Zhang Xuefeng is the most followed educational content creators in China. To establish Zhang Xuefeng’s remarks as a credible information shock, we document the reach and timing of his influence across major Chinese social media platforms, as shown in Table 1 and Fig. 1 . His livestreams and short videos during this period accumulated tens of millions of views, with key videos receiving hundreds of thousands of likes and shares. Related hashtag topics trended nationally on multiple platforms, generating extensive mainstream media coverage. Table 1. Zhang Xuefeng’s influence across major Chinese social media platforms. Platform Influence Douyin Over 30 million followers, ranking among the top educational accounts nationwide. During the week of June 19-25, 2023, Zhang Xuefeng ranked on Douyin’s creator follower-growth chart, growing over 1.13 million new followers in a single week 35 . Kuaishou Approximately 20 million followers, with exclusive 2024 college counseling livestreams attracting 324 million cumulative views (from June 20 to July 7, 2024, data includes live broadcasts and replays) 36 . Weibo Millions of followers, with trending topics during peak discussion periods. Bilibili Hundreds of thousands of followers across multiple account matrices. Open in a new tab Fig. 1. Open in a new tab “Zhang Xuefeng” keyword heat index. Due to data availability limitations for the 2023 period, this figure shows data from August 2024 to August 2025 to illustrate the typical seasonal spike in public interest that occurs during the annual college application window. A core component of the s“Zhang Xuefeng phenomenon” is its emphasis on the labor-market returns of different majors. The majors that Zhang Xuefeng endorsed are correlated with significantly higher average salaries compared to those he discouraged. To illustrate this, we compiled publicly available data on the average starting salaries of Chinese college graduates in 2024 37 , 38 . As depicted in Fig. 2 , graduates in Software Engineering earned an average monthly salary of 7000 yuan–approximately 25% higher than their counterparts in Journalism, who earned 6000 yuan. Furthermore, empirical research underscores the significant role that major-specific skills and labor-market alignment play in shaping graduates’ earnings 1 . The pre-existing disparities in financial returns across fields provide crucial context for understanding why an information intervention centered on employability could exert such a profound impact on student decision-making. The public signal did not create a narrative out of thin air; rather, it amplified existing, quantifiable differences in the labor market, making this information more salient to students and parents during the high-stakes decision-making period. Fig. 2. Open in a new tab The labor-market returns of different majors. This event constitutes a valid natural experiment for several reasons: Exogeneity The timing and intensity of this information shock were exogenous, driven by online debate rather than any change in the underlying quality or pre-existing trends of the majors themselves. Clear treatment groups Zhang’s explicit discourage and endorse signals created well-defined treatment groups (e.g., Journalism) and control groups (majors not frequently discussed). Scale The national reach of the event allows for a robust analysis of its impact across different regions and institutions. Data and methods Data sources and sample construction A key distinction of our study is the construction of an unprecedentedly large and comprehensive dataset on college majors admission scores in China. We employed a manual data collection strategy, compiling the data from multiple public sources for the years 2017 to 2023. Major data sources include: (1) universities’ official admissions websites; (2) authorized third-party platforms (e.g., http://gaokao.chsi.com.cn , the official Gaokao information platform managed by the Ministry of Education). All data were cross-validated across multiple sources when available to ensure accuracy. The unit of analysis is a university-discipline-province-year-subject observation. Our primary outcome variable is the minimum admission rank, which we derived by matching the minimum admission score for each program to the official one-score-one-rank tables (yi fen yi duan biao). This rank-based measure is crucial for our analysis as it is standardized and comparable across different years and provinces, which have varying exam difficulties and numbers of test-takers. To address inconsistencies in how universities report admissions data, a primary processing step was the aggregation of specific majors into broader disciplines. Majors within the same discipline share similar characteristics, and they are likely to be affected similarly by public information shocks. Majors are aggregated into a consistent and comparable unit of analysis based on the official undergraduate major catalog from the Ministry of Education of China 5 , as shown in Table 2 . Furthermore, for a more focused analysis on elite institutions and to ensure comparability with existing literature, we extracted a specific subset from our main data containing 147 universities designated under the“Double First-Class Initiative”. Table 2. Undergraduate major categories of Chinese higher education. Discipline category Major category Philosophy Philosophy Economics Economics, Finance, Financial Studies, Economics and Trade Law Law, Political Science, Sociology, Ethnology, Marxist Theory, Public Security Education Education, Physical Education Literature Chinese Language and Literature, Foreign Languages and Literatures, Journalism and Communication History History Science Mathematics, Physics, Chemistry, Astronomy, Geographical Sciences, Atmospheric Sciences, Marine Sciences, Geophysics, Geology, Biological Sciences, Psychology, Statistics Engineering Mechanics, Mechanical, Instruments, Materials, Energy and Power, Electrical, Electronic Information, Automation, Computer Science, Civil Engineering, Hydraulic Engineering, Surveying and Mapping, Chemical and Pharmaceutical, Geology, Mining, Textile, Light Industry, Transportation, Marine Engineering, Aeronautics and Astronautics, Weaponry, Nuclear Engineering, Agricultural Engineering, Forestry Engineering, Environmental Science and Engineering, Biomedical Engineering, Food Science and Engineering, Architecture, Safety Science and Engineering, Bioengineering, Public Security Technology, Interdisciplinary Engineering Agronomy Plant Production, Nature Conservation and Environmental Ecology, Animal Production, Veterinary Medicine, Forestry, Aquaculture, Grassland Science Medicine Basic Medicine, Clinical Medicine, Stomatology, Public Health and Preventive Medicine, Traditional Chinese Medicine, Integrated Chinese and Western Medicine, Pharmacy, Traditional Chinese Pharmacy, Forensic Medicine, Medical Technology, Nursing Management Management Science and Engineering, Business Administration, Agricultural Economics and Management, Public Administration, Library, Information and Archives Management, Logistics Management and Engineering, Industrial Engineering, E-commerce, Tourism Management Arts Art Theory, Music and Dance, Drama and Film/Television Studies, Fine Arts, Design Open in a new tab The sheer scale of both our comprehensive dataset (3,317,803 observations) and the“Double First-Class Initiative” university subset (560,320 observations) provides the robust empirical foundation for this study’s claim as the first large-scale causal analysis of a social media information shock on college major choices in China. This broad coverage allows for a panoramic and more generalizable analysis of admission patterns nationwide. Research design To estimate the causal effect of the social media information shock, we employ a difference-in-differences (DID) research design. This quasi-experimental method is well-suited for this context as it allows us to isolate the impact of the event by comparing the change in outcomes for the affected majors (the treatment groups) before and after the event, relative to the change in outcomes for unaffected majors (the control group) over the same period. The treatment period (post-event) is defined as the year 2023, while the pre-treatment period spans from 2017 to 2022. Based on the content of the influencer’s advice, we define several distinct treatment groups for analysis: Discouraged major Journalism and Communication, which received a strong negative signal. Endorsed major Computer Science, which received strong positive signals. Spillover major Transportation, which serves as an exploratory case study to investigate potential spillover effects. It is worth noting that, to examine whether the influencer’s underlying employment-first ideology generated spillover effects beyond the explicitly named majors, we selected Transportation as an exploratory case study. This choice is motivated by two primary considerations. First, analyzing the broader Engineering category is methodologically problematic due to its extreme internal heterogeneity. While Zhang Xuefeng endorses practical engineering fields, he simultaneously and explicitly discourages certain engineering majors with historically poor employment prospects. An aggregated analysis of all engineering fields would therefore wash out these opposing effects. Second, Transportation represents a theoretically relevant test case. It is a practical, employment-oriented field that aligns closely with Zhang’s pragmatic philosophy, emphasizing job prospects, yet it was not explicitly targeted in his viral commentary. We therefore present the Transportation results as suggestive evidence of potential spillover effects. Variables To empirically implement the DID strategy outlined above, we define the following variables for our models. Dependent variable The primary outcome variable is , which is the natural logarithm of the minimum admission rank for major discipline m at university u in province p in year t for subject type s . Students within each province are ranked by their total exam scores. Universities admit applicants in descending order of their provincial ranks until the predetermined enrollment capacity for a specific major is filled. Thus, the cutoff rank reflects the provincial rank of the last (i.e., lowest-scoring) student admitted to that program. For instance, if a university program has a quota of 50 students and the 50th admitted student ranks 2000th in the province, the minimum admission rank for this program is 2000. Consequently, a higher numerical value of the rank (e.g., shifting from 2000 to 5000) implies that the major had to dip deeper into the applicant pool to fill its capacity, thereby indicating weaker admitted students and lower popularity. Independent variables For our primary DID specification, the key independent variable is the interaction term . Its coefficient, , measures the average causal effect of the information shock on the admission rank of the treated major discipline. is a treatment dummy that equals 1 if a major discipline m belongs to the treatment group in a given model, and 0 otherwise. This variable is defined separately for each analysis: as “Journalism and Communication”’ for the discouraged group, “Computer Science” for the endorsed group, and “Transportation” for the spillover analysis. is a dummy variable equal to 1 for the year 2023 and 0 for prior years (2017-2022). Empirical model To isolate the causal effect, we first estimate a standard fixed effects model and then present a more robust specification that includes high-dimensional fixed effects. Model 1: DID with five-way fixed effects This model includes a comprehensive set of separate fixed effects to control for time-invariant characteristics and common time trends: 1 where the fixed effects control for major-specific characteristics ( ), university-level prestige ( ), province-level differences ( ), year-specific shocks ( ), and subject-type differences ( ). Model 2: DID with high-dimensional fixed effects To account for the possibility that unobserved factors affecting admission ranks may vary at the intersection of different dimensions, our preferred specification replaces several individual fixed effects with high-dimensional interaction fixed effects: 2 where the province-by-year fixed effect ( ) is included to absorb province-specific time-varying shocks. This control is critical because factors such as exam difficulty, the number of test-takers, and major educational policies create significant variations that make admission ranks incomparable across years within the same province. The subject-by-year fixed effect ( ) accounts for differential trends and fluctuations across subjects (e.g., liberal arts, science, or comprehensive). The province-by-subject fixed effect ( ) is essential for addressing the structural changes within provinces during our sample period. As some provinces transitioned from a binary liberal arts/science system to a comprehensive one under the New Gaokao Reform, this term controls for the distinct, time-invariant characteristics of each subject track regime within a province, thus ensuring comparisons are made between comparable units. Model 3: event-study analysis for parallel trends test To formally test the parallel trends assumption underlying the DID design, we conduct an event-study analysis by estimating year-by-year treatment effects. Instead of a single treatment dummy, the treatment indicator is interacted with year dummies for each year except the reference year (2022): 3 where is an indicator variable for year t , and 2022 serves as the reference year. The coefficients for (pre-treatment years) should be statistically insignificant and close to zero if the parallel trends assumption holds. The coefficient captures the treatment effect in the post-event year. Model 4: robustness check with COVID-19 controls Our sample period (2017–2023) includes the COVID-19 pandemic (2020-2022), which may have exerted heterogeneous impacts on different majors. For instance, digital economy-related majors (e.g., Computer Science) may have become more popular during the pandemic, while traditional media majors (e.g., Journalism and Communication) may have been adversely affected by restricted in-person activities. To address this potential confounding factor, we augment Model 2 with COVID-19 control variables: 4 where is a dummy variable equal to 1 for years 2020–2022, indicates digital economy-related majors (Computer Science, Electronic Information, Automation), and indicates traditional media majors (Journalism and Communication, Arts). If the treatment effect remains robust after controlling for these pandemic-related heterogeneities, it provides stronger evidence that our main results are not driven by COVID-19 confounding. Model 5: DID with a threshold variable To explore the potential heterogeneity of the shock’s impact, we test whether the effect is moderated by provincial economic context. We extend our primary DID specification (Model 2) by incorporating a threshold variable based on provincial per capita GDP. This results in a triple-difference (DDD) model: 5 where is a dummy variable equal to 1 if the per capita GDP of province p is above the sample median, and 0 otherwise. The fixed effects structure remains identical to that in Model 2 to ensure a robust estimation. captures the average treatment effect of the information shock in the baseline group–provinces with lower per capita GDP. is the key coefficient of interest for the threshold analysis. It measures the difference in the treatment effect between high-GDP provinces and low-GDP provinces. A statistically significant would indicate that the impact of the influencer’s advice is indeed conditional on regional economic development. represents the net treatment effect for provinces with higher per capita GDP. Empirical results and analysis This section presents the empirical results of the information shock event. It begins with graphical evidence to visualize admission rank trends and to underscore the suitability of the empirical strategy. The analysis then proceeds to the core quantitative results from the DID model, which reveal a substantial and asymmetric impact of the information shock. Finally, this average effect is further dissected by exploring its heterogeneity across provinces with different economic contexts, employing a triple-difference (DDD) framework. Graphical evidence and descriptive trends Before proceeding to the causal analysis, we first examine descriptive trends in student preferences. Figure 3 displays the mean normalized admission ranks for eleven major disciplines at “Double First-Class Initiative” universities from 2017 to 2023. The Arts discipline is excluded from this visualization due to its distinct admission rules, which involve specialized artistic exams, making its rank data not directly comparable to other fields. The term “normalized admission rank” refers to the raw minimum admission rank for a specific major-university program divided by the total number of National College Entrance Examination takers in that province and year. This normalization is essential for ensuring comparability across observations, as the size of the candidate pool varies significantly by province and over time. The value plotted in the figure is the mean of these individual normalized ranks, aggregated at the discipline level for each year. It is calculated as follows: 6 where the mean is taken over all individual observations i (a unique major-university-province combination) belonging to a broader discipline d in year t . Fig. 3. Open in a new tab Comparison of admission rankings by discipline (Double First-Class Initiative). Higher value indicates a worse rank. We focus on the elite “Double First-Class Initiative” subsample for this initial visual analysis, as the wide variance in admission ranks across the full spectrum of institutions can obscure underlying trends. The data reveals pre-existing trends within the sample period, such as the rising popularity of Science and Engineering and the relative decline of disciplines like Literature. This finding highlights the necessity of using a difference-in-differences design to precisely isolate the net effect of the information shock from these confounding background trends. Figure 4 visualizes the admission rank trends for the treatment group (Journalism and Communication) and two control groups (other majors within the Literature discipline and a broader set of humanities and social science majors). While the pre-event trends (2017–2022) are not strictly parallel and exhibit some volatility, a stark divergence is evident in 2023. Following the event, the previous upward trajectory for both control groups halts, with their admission ranks plateauing in 2023. In sharp contrast, the admission rank for the treatment group diverges significantly, continuing to rise while the control groups’ ranks plateau. This provides strong evidence of a causal impact. Furthermore, the observed pre-event fluctuations underscore the necessity of our high-dimensional fixed effects model (Model 2), which is specifically designed to control for such complex, group-specific time trends that a simple visual inspection cannot fully account for, thus allowing for a more rigorous identification of the treatment effect. Fig. 4. Open in a new tab Comparison of admission rankings between Journalism and Communication and other majors (Double First-Class Initiative). Higher value indicates a worse rank. The average causal effects of the information shock The average causal effects of the information shock are estimated using two DID specifications, with results for the discouraged major, Journalism and Communication, presented in Table 3 . The analysis includes both a standard five-way fixed effects model (Model 1) and a more robust high-dimensional fixed effects specification (Model 2). Leveraging the large scale of the dataset, results are reported for both the full sample of all universities and the elite “Double First-Class Initiative” university subsample. Table 3. Casual effects on the discouraged major Journalism and Communication. Variables Model 1 Model 1 Model 2 Model 2 Treatment effect (0.013) (0.021) (0.013) (0.018) Observations 3317803 560620 3317803 560620 0.913 0.876 0.932 0.915 Major FE Yes Yes Yes Yes University FE Yes Yes Yes Yes Province FE Yes Yes Year FE Yes Yes Subject FE Yes Yes Province Year FE Yes Yes Province Subject FE Yes Yes Year Subject FE Yes Yes Open in a new tab Standard errors clustered by province in parentheses *** , , The results from the preferred specification, Model 2, reveal a substantial negative impact on the discouraged major. For the full sample, the coefficient on the interaction term is 0.123 and is highly significant ( ), as shown in Column 3 of Table 3 . This indicates that across all higher education institutions, the influencer’s negative commentary caused the admission rank for Journalism and Communication to worsen by approximately 13.1% ( ). This result, derived from over three million observations, provides broad and powerful evidence that a single public signal can significantly alter national enrollment patterns. The effect is even more pronounced for the elite “Double First-Class Initiative” university subsample (Column 4), where the estimated coefficient increases in magnitude to 0.221 ( 0.001). This implies that the admission rank of discouraged major at top-tier universities worsened by approximately 24.7% ( ). The larger effect size for elite universities suggests that applicants to these highly competitive institutions may be particularly sensitive to information regarding long-term career prospects. To provide a more concrete sense of magnitude, we translate these rank shifts into approximate student numbers. Our dataset contains 10316 distinct Journalism programs in 2022, which, assuming 10-15 students per program, corresponds to roughly 103000–155000 annual enrollments nationwide. The 13.1% effect for all universities suggests that approximately 13500–20300 students (or roughly 16900 using the midpoint estimate) who would have chosen Journalism opted for alternative majors in 2023. For elite universities, the 24.7% effect implies approximately 3800–5800 affected students. These data-driven estimates illustrate that the information shock had tangible effects on tens of thousands of individuals. The analysis reveals clear evidence of an asymmetric response to the influencer’s advice. Table 4 presents the corresponding results for the endorsed major, Computer Science, using the preferred Model 2 specification. A statistically significant positive effect is found for Computer Science in the full sample, with a coefficient of ( ), indicating that the endorsement caused the major’s admission rank to improve by 6.7%. This reveals a strong asymmetry, as the magnitude of the negative effect on Journalism and Communication (0.123) is nearly double that of the positive effect on Computer Science ( ). This finding aligns with the principle of loss aversion 39 . The ratio of the negative to positive effect is , which mirrors the widely documented loss aversion coefficient in behavioral economics 1.82–2.11) 40 . This suggests that in the high-stakes environment of college applications, students and their families implicitly weigh the potential loss of future career stability heavily against the gain of high salary opportunities, reacting almost twice as strongly to risk warnings. Table 4. Comparison of the casual effects on the discouraged, endorsed, and spillover majors. Variables Journalism and communication Computer science Transportation Model 2 Model 2 Model 2 Model 2 Model 2 Model 2 Treatment effect (0.013) (0.018) (0.020) (0.022) (0.025) (0.019) Observations 3317803 560620 3317803 560620 3317803 560620 0.932 0.915 0.932 0.915 0.932 0.915 Open in a new tab Standard errors clustered by province in parentheses. *** , ** , * Beyond the directly targeted majors, we conduct an exploratory analysis to examine potential spillover effects on related fields that align with the influencer’s pragmatic, employment-focused philosophy. The Transportation major serves as an illustrative case study. The results in Table 4 show a statistically significant positive effect for Transportation in the elite university sample, with a coefficient of ( ). This suggestive evidence indicates that the influencer’s message may have resonated beyond the explicitly mentioned majors, potentially affecting other practical, engineering-oriented fields perceived to have strong labor market outcomes. Parallel trends test and robustness checks To validate the causal interpretation of our findings, we conduct two critical robustness checks. First, the parallel trends assumption is tested with an event-study analysis. Second, potential confounding from the COVID-19 pandemic is addressed. Event-study analysis Figure 5 presents the results of the event-study analysis (Model 3), which estimates year-by-year treatment effects relative to the reference year 2022. The estimated coefficients for the pre-treatment years (2017–2021) are generally small in magnitude and mostly statistically insignificant at the 5% level, except for 2019. For 2017, 2018, 2020, and 2021, the point estimates hover around zero, with confidence intervals that encompass zero, supporting the parallel trends assumption for these years. The 2019 coefficient is ( p = 0.017), which is marginally significant. While this indicates some pre-existing deviation in that specific year, the overall pattern across the five pre-treatment years does not suggest a systematic divergence in trends prior to the event. Importantly, the treatment effect in 2023 is large, positive, and highly significant (0.102, ), representing a clear break from the pre-event pattern. This sharp divergence in 2023, following largely stable pre-trends, provides strong evidence supporting the causal impact of the information shock. Fig. 5. Open in a new tab Event-study plot testing parallel trends assumption. The coefficients represent the treatment effect on log(Rank) for each year relative to the reference year 2022. Error bars indicate 95% confidence intervals. The dashed red line marks the event year (2023). To further validate our identification strategy, we conduct placebo tests using History and Philosophy majors, two humanities fields that Zhang Xuefeng did not explicitly mention, as pseudo-treatment groups. Table 5 presents the placebo test results using our preferred Model 2 specification. We find no significant overall effect for History (coefficient = 0.031, p = 0.196) in the pooled DID specification, while Philosophy exhibits a marginally significant positive coefficient (coefficient = 0.161, p = 0.022). Table 5. Placebo tests on history and philosophy. Variables Journalism and communication History Philosophy Treatment effect 0.031 (0.013) (0.023) (0.066) Open in a new tab All models use Model 2 specification with high-dimensional fixed effects *** , ** , * To unpack these results, the event-study analysis (Fig. 6 ) reveals an important temporal pattern. For Philosophy, the significant pooled coefficient is largely an artifact of its pre-existing downward trend, as it exhibited significant negative coefficients during 2018–2020. Since a lower numerical rank indicates higher admission difficulty, these negative pre-treatment coefficients imply that Philosophy was significantly more competitive in 2018–2020 than in the 2022 reference year. Crucially, this pre-existing trend of decreasing competitiveness leading up to 2022 was followed by a statistically insignificant coefficient in the 2023 event year (0.012, p = 0.892). This pattern confirms that while Philosophy had an underlying trend of softened demand, there was no additional spurious shock attributable to our DID methodology in the treatment year. In contrast, History shows relatively stable pre-trends, confirming its validity as a control group. Although it exhibits a weakly significant positive deviation in 2023 (0.066, p = 0.032), potentially reflecting a broader, mild downward shift in applicant preferences for humanities in that year, its average treatment effect remains statistically insignificant. Fig. 6. Open in a new tab Event-study analysis for placebo tests. Robustness to COVID-19 confounding Table 6 presents the results of Model 4, which controls for COVID-19-related heterogeneities. The baseline DID coefficient without COVID controls is 0.123 (Column 1). After adding interaction terms between the COVID-19 period dummy and major-type indicators (Column 2), the DID coefficient increases slightly to 0.127, representing a change of only 3.1%. The COVID Digital interaction is negative and significant ( , p = 0.012), confirming that digital economy majors did experience a differential trend during the pandemic. The COVID Traditional interaction is small and insignificant (0.005, p = 0.631). Crucially, the treatment effect remains highly significant and of similar magnitude after controlling for pandemic-related factors. This stability demonstrates that our main findings are not driven by COVID-19 confounding and represent the genuine causal effect of the influencer’s information shock. Table 6. Robustness check with COVID-19 control variables. Variables Baseline + COVID controls Treatment effect (0.013) (0.018) COVID Digital majors – (0.007) COVID Traditional majors – 0.005 (0.010) Observations 3317389 3317389 0.932 0.932 Open in a new tab Standard errors clustered by province in parentheses *** , ** , * Heterogeneous effects across provinces While the average effects are substantial, the impact of the shock was not uniform across all regions. A further investigation explores whether the effect was moderated by provincial economic context. Figure 7 provides preliminary visual evidence of this heterogeneity. It contrasts the trend divergence in Jiangsu, a province with high per capita GDP, and Guangxi, a province with lower per capita GDP. The sharp upward turn for Journalism in 2023 appears more pronounced in Guangxi than in Jiangsu, suggesting that the negative impact may be weaker in more economically developed areas. Fig. 7. Open in a new tab Comparison of admission rankings between Journalism and Communication and other majors (different provinces). Higher value indicates a worse rank. To formally test and quantify this heterogeneity, a triple-difference (DDD) model (Model 5) is employed that interacts the DID term with a dummy variable indicating whether a province’s per capita GDP is above the median. The robustness of this analysis is strengthened by using a more refined control group consisting of other liberal arts majors, which ensures higher comparability. The results, presented in Table 7 , confirm the visual intuition from the figure and reveal significant regional disparities. Table 7. The heterogeneous effect of the information shock by provincial economic level. Variables Model 5 (0.009) − (0.014) Observations 492403 0.944 Threshold 77983.5 yuan Open in a new tab Standard errors clustered by province in parentheses *** , ** , * The baseline DID effect, which captures the impact in provinces with lower per capita GDP, is 0.067 and highly significant. This indicates that in these less developed regions, the event caused the admission rank of Journalism to worsen by a notable 6.9% ( ). The fact that this significant negative effect persists even with a more comparable control group serves as a powerful robustness check on the main finding. The triple-interaction term, the primary coefficient of interest in this model, is and is statistically significant at the 1% level. This negative and significant coefficient provides strong evidence that the adverse impact of the influencer’s advice was substantially weaker in provinces with higher per capita GDP. The net effect for these high-GDP provinces is the sum of the two coefficients ( ), implying a much smaller, though still negative, impact of approximately 2.4% ( ) in rank worsening. This finding suggests that students in more developed economic environments, who may have access to more diverse information channels and perceive a wider array of local job opportunities, exhibit greater resilience to such centralized negative signals. Discussion This study provides causal evidence of the powerful influence that social media influencers exert on the high-stakes educational choices of college applicants in China. The findings reveal that a single influencer’s viral commentary significantly altered national admission patterns, with the impact being substantial, asymmetric, and heterogeneous. This discussion interprets these key findings and considers their broader implications. The substantial magnitude of the effect, a 24.7% worsening in admission ranks for Journalism at elite universities, highlights the profound influence of this new information channel. The observed asymmetry, where the negative impact on the discouraged major was nearly double the positive impact on the endorsed major, strongly confirms the loss aversion framework. By quantifying this behavioral response at a national scale, our study contributes a novel educational context to the loss aversion literature. The 1.836 empirical ratio indicates that the influencer’s advice primarily acted as a powerful coordinating signal that amplified latent anxieties about future career precarity, i.e., a potential loss, which proved far more potent than the opportunity-highlighting function of the endorsement. The Transportation major spillover analysis should be viewed as exploratory. The choice of Transportation is influenced by both theoretical considerations, i.e., its alignment with the influencer’s pragmatic philosophy, and the research team’s professional background in transportation engineering. While the significant effect in the elite university subsample is suggestive, we cannot rule out selective reporting or chance findings. Future research should pre-specify potential spillover fields based on explicit theoretical criteria to provide more rigorous tests of how far such information shocks diffuse beyond directly mentioned majors. The heterogeneous impact across provinces raises an important question about the underlying mechanisms. There are two competing explanations for why the information shock had weaker effects in higher-GDP provinces. First, the economic buffer hypothesis suggests that students in wealthier provinces face lower economic pressure due to more diverse local labor markets and greater household financial resources, making them less sensitive to employment-related information shocks. Families in these regions may perceive a wider array of career opportunities and possess greater capacity to absorb uncertainty, reducing their responsiveness to risk-oriented signals about specific majors’ employment prospects. Second, the rational response hypothesis posits that families in wealthier provinces may be equally or even more concerned about economic returns to education, but possess more sophisticated information-processing capabilities and access to diverse information channels beyond social media. This informational advantage could enable them to critically evaluate and appropriately discount sensational online commentary, relying instead on more comprehensive evidence about labor market outcomes. We acknowledge that our current data do not permit formal testing of these mechanisms through direct measurement of mediating variables. While the province year fixed effects control rigorously for time-varying confounders such as concurrent economic shocks or policy changes, we cannot directly observe factors such as labor market diversity, internet penetration rates, or households’ information sources. Future research could more rigorously test these mechanisms by incorporating province-level structural data and conducting formal mediation analysis. Additionally, individual-level survey data on parental characteristics, information sources, and decision-making processes would provide more direct evidence on the mechanisms underlying the observed heterogeneity. Conclusions This study investigated the causal impact of a large-scale social media information shock on college major choices in China. Using a comprehensive national dataset of university admission ranks from 2017 to 2023, the analysis employed a difference-in-differences (DID) and a triple-difference (DDD) design to estimate the event’s effects. The findings demonstrate that the influencer’s commentary led to a substantial and statistically significant decline in the popularity of the discouraged major (Journalism and Communication), an effect that was considerably larger than the positive impact on the endorsed major (Computer Science). Furthermore, this impact was significantly weaker in provinces with higher economic development. In conclusion, this research provides strong evidence that social media influencers have emerged as new and powerful, yet unregulated, actors in the educational landscape, with the capacity to profoundly influence college major choice intentions and present novel challenges for students, institutions, and policymakers in the digital age. Author contributions Conceptualization: L.M., Q.L., B.L. Methodology: L.M., Y.W., B.L. Investigation: Y.W. Writing-original draft: L.M., Y.W. Writing-review & editing: L.M., Q.L., C.Z., B.L. Funding acquisition: C.Z. Supervision: Q.L. All authors reviewed the manuscript. Funding This work is supported by the Education and Teaching Reform Projects from Xihua University (Grant No. xcjz2025010 and Grant No. xjjg2025093). Data availability The dataset used and analyzed during the current study is available from the corresponding author on reasonable request. Declarations Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References 1. Hang, T. & Zhou, Y. Chinese university major decision and its effect on wages: Modeling interaction between major specificity and education-job relevancy using machine learning approaches. J. Chin. Political Sci. 30 , 475 (2025). [ Google Scholar ] 2. Quadlin, N. From major preferences to major choices: Gender and logics of major choice. Sociol. Educ. 93 , 91–109 (2020). [ Google Scholar ] 3. Dahl, G. B., Rooth, D.-O. & Stenberg, A. High school majors and future earnings. Am. Econ. J. Appl. Econ. 15 , 351–382 (2023). [ Google Scholar ] 4. Ministry of Education of the People’s Republic of China. Statistical bulletin on national education development in 2024. https://hudong.moe.gov.cn/jyb_sjzl/sjzl_fztjgb/202506/t20250611_1193760.html (2025). 5. Ministry of Education of the People’s Republic of China. Catalog of undergraduate majors in regular institutions of higher education. http://www.moe.gov.cn/srcsite/A08/moe_1034/s4930/202504/t20250422_1188239.html (2025). 6. iiMedia Research. Big data analysis report on China’s college-entrance examination application slot-filling and labor market employment prospects in 2025. https://www.iimedia.cn/c400/106344.html (2025). 7. Bourdieu, P., Passeron, J.-C. & Nice, R. Education, society and culture. Trans. Richard Nice 1 , 15–29 (1977). [ Google Scholar ] 8. Guo, C., Guo, M., Fang, C. & Hao, X. College major decision making behavior of urban and rural students under cultural capital impact in China. Sci. Rep. 15 , 913 (2025). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 9. Li, X. College Major Choices in China . Ph.D. thesis, University of California, Los Angeles (2022). 10. Wiswall, M. & Zafar, B. Determinants of college major choice: Identification using an information experiment. Rev. Econ. Stud. 82 , 791–824 (2015). [ Google Scholar ] 11. Piepenburg, J. G. & Fervers, L. Do students need more information to leave the beaten paths? The impact of a counseling intervention on high school students’ choice of major. Higher Educ. 84 , 321–341 (2022). [ Google Scholar ] 12. Ma, L., Li, X., Zhu, Q. & Ye, X. College-major choice to college-then-major choice: Experimental evidence from Chinese college admissions reforms. Econ. Educ. Rev. 94 , 102380 (2023). [ Google Scholar ] 13. Baker, R., Bettinger, E., Jacob, B. & Marinescu, I. The effect of labor market information on community college students’ major choice. Econ. Educ. Rev. 65 , 18–30 (2018). [ Google Scholar ] 14. Huo, Y. & Wang, Y. The “zhang xuefeng effect”: Information intervention and the college admission problem in China. Int. Stud. Econ. (2025). 15. Guo, C. & Wang, T. The influence of discipline evaluation participation and information disclosure on the quality of undergraduate enrollment. J. Higher Educ. 42 , 40–50 (2021). [ Google Scholar ] 16. Liu, S., Li, J. & Luo, X. Can online recruitment promotion improve the “blind selection” and “wrong selection” of students’ majors? Journal of National Academy of Education Administration 87–97 (2022). 17. Ding, Y., Du, L., Wu, Y., Yang, J. & Ye, X. The impact of informational interventions on college-major choice-evidence from a large-scale randomized experiment. China Econ. Q. 6 , 2239–2262 (2021). [ Google Scholar ] 18. Hoag, A., Grant, A. E. & Carpenter, S. Impact of media on major choice: Survey of communication undergraduates. Nacada J. 37 , 5–14 (2017). [ Google Scholar ] 19. Fan, Z., Zhao, X. & Wei, J. The public signal of we media and the college entrance examination will-decision application: A study based on “Zhang Xuefeng journalism event’’. J. East China Normal Univ. 43 , 58 (2025). [ Google Scholar ] 20. Goodwin, J., Finn, J., Dowling, M., Pearson, L. & Gough, R. Can media influence people’s career choices? A systematic review. Austr. J. Career Dev. 34 , 134–141 (2025). [ Google Scholar ] 21. Dwivedi, Y. K. et al. Setting the future of digital and social media marketing research: Perspectives and research propositions. Int. J. Inf. Manage. 59 , 102168 (2021). [ Google Scholar ] 22. Soares, J., Carvalho, C. & Silva, A. D. A systematic review on career interventions for university students: Framework, effectiveness, and outcomes. Aust. J. Career Dev. 31 , 81–92 (2022). [ Google Scholar ] 23. Carvalho, L., Mourão, L. & Freitas, C. Career counseling for college students: Assessment of an online and group intervention. J. Vocat. Behav. 140 , 103820 (2023). [ Google Scholar ] 24. Chen, Y. & Kesten, O. Chinese college admissions and school choice reforms: A theoretical analysis. J. Polit. Econ. 125 , 99–139 (2017). [ Google Scholar ] 25. Arcidiacono, P. Ability sorting and the returns to college major. J. Econ. 121 , 343–375 (2004). [ Google Scholar ] 26. Beffy, M., Fougere, D. & Maurel, A. Choosing the field of study in postsecondary education: Do expected earnings matter?. Rev. Econ. Stat. 94 , 334–347 (2012). [ Google Scholar ] 27. Gemici, A. & Wiswall, M. Evolution of gender differences in post-secondary human capital investments: College majors. Int. Econ. Rev. 55 , 23–56 (2014). [ Google Scholar ] 28. Qi, X. The influence of social media on Chinese students’ choice of private universities in china. In Proceedings of the 2023 6th International Conference on Computers in Management and Business , 164–169 (2023). 29. Petty, R. E. & Cacioppo, J. T. Communication and Persuasion: Central and Peripheral Routes to Attitude change (Springer, 2012). [ Google Scholar ] 30. Festinger, L. A theory of social comparison processes. Hum. Relat. 7 , 117–140 (1954). [ Google Scholar ] 31. Sunstein, C. The law of group polarization. J. Political Philos. (2002). 32. Garrett, R. K. Echo chambers online?: Politically motivated selective exposure among internet news users. J. Comput.-Mediat. Commun. 14 , 265–285 (2009). [ Google Scholar ] 33. Wang, Z. How perceived social support influences Chinese students’ intention to change majors: A chain mediation model with moderation. Front. Psychol. 16 , 1502128 (2025). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 34. Pengpai. Zhang xuefeng in trouble, who are really offended. https://www.thepaper.cn/newsDetail_forward_23696385 (2023). 35. Xinbang. Zhang xuefeng ranks first on douyin follower growth chart: 1.13 million new followers in one week. https://zhuanlan.zhihu.com/p/639968557 (2023). 36. Kuaishou Technology. 2024 kuaishou college entrance examination season data report. https://www.fxbaogao.com/report?id=4411859 (2024). 37. YicaiGlobal. Over half of china’s fresh graduates last year earned up to USD835 a month, report says. https://www.yicaiglobal.com/news/most-of-chinas-2024-college-graduates-earned-under-usd835-a-month-report-says (2024). 38. YicaiGlobal. Graduate earnings jump 53% in china over past decade, report says. https://www.yicaiglobal.com/news/china-college-graduate-earnings-surge-53-over-past-decade-report-says (2025). 39. Tversky, A. & Kahneman, D. Loss aversion in riskless choice: A reference-dependent model. Q. J. Econ. 106 , 1039–1061 (1991). [ Google Scholar ] 40. Brown, A. L., Imai, T., Vieider, F. M. & Camerer, C. F. Meta-analysis of empirical estimates of loss aversion. J. Econ. Liter. 62 , 485–516 (2024). [ Google Scholar ] Associated Data This section collects any data citations, data availability statements, or supplementary materials included in this article. Data Availability Statement The dataset used and analyzed during the current study is available from the corresponding author on reasonable request. Articles from Scientific Reports are provided here courtesy of Nature Publishing Group ACTIONS View on publisher site PDF (3.3 MB) Cite Collections Permalink RESOURCES Similar articles Cited by other articles Links to NCBI Databases Follow NCBI NCBI on X (formerly known as Twitter) NCBI on Facebook NCBI on LinkedIn NCBI on GitHub NCBI RSS feed Connect with NLM NLM on X (formerly known as Twitter) NLM on Facebook NLM on YouTube National Library of Medicine 8600 Rockville Pike Bethesda, MD 20894 Web Policies FOIA HHS Vulnerability Disclosure Help Accessibility Careers NLM NIH HHS USA.gov Back to Top

Related documents

• Beyond information seeking: a latent profile analysis of financial media repertoires and psychological motivations among Chinese users.#988619
• Self-Perceptions on Social Media vs. Offline Contrast With Those Perceived in Generalized but not Close Others#486104
• Self-Perceptions on Social Media vs. Offline Contrast With Those Perceived in Generalized but not Close Others#486062
• Self-Perceptions on Social Media vs. Offline Contrast With Those Perceived in Generalized but not Close Others#365369
• Social media use motives and academic performances: exploring the mediating role of daily usage duration.#399562