1. Introduction

Medical books serve as enduring pillars of knowledge dissemination in healthcare, education, and interdisciplinary practice—their value shaped by both professional necessity and historical continuity. Editors typically commission chapters from experienced authors, with pre- writing consensus on titles, frameworks, and guidelines ensuring rigor; this emphasis on quality, paired with the professional honor associated with contributing to medical literature, underscores why such books remain trusted resources for a broad audience (Kendirci, 2013). This demand is not transient but rooted in centuries of medical practice: historical studies on the dissemination of anatomical knowledge via Renaissance texts (McCall, 2022) illustrate that medical books have long been central to knowledge inheritance. For contemporary medical professionals, this legacy translates to sustained reading needs: reading remains critical for continuing education (Deshpande, 2003; Garcja Benjtez, 2018) and for engaging with interdisciplinary insights. Notably, modern medical practice increasingly intersects with psychology. This interdisciplinary overlap extends the relevance of medical books beyond traditional healthcare settings, attracting users from fields like psychology who seek to bridge medical and psychological knowledge. During disruptions like the COVID- 19 pandemic, these dynamics shift: pandemic discourse fragmentation (Vakoch et al., 2023) and temporary shifts to digital resources altered physical borrowing patterns, though the post-pandemic "return to routine" has renewed reliance on tangible books for many users.

Against this backdrop, non-medical universities present a unique lens to study medical book borrowing. Nanjing Normal University (NNU), for instance, lacks medical programs but observes consistent borrowing of R-category (medical/health) books—activity primarily driven by the School of Psychology, which exhibits significantly higher borrowing volumes of and more borrowers for R-category books than other departments. This reflects the interdisciplinary nature of modern information needs: psychology students and faculty seek medical books to explore intersections like behavioral medicine, narrative medicine (Pisanu et al., 2023), or mental health-climate change links (Kern de Castro & Reis, 2025), challenging the traditional "medical-centric" paradigm of health literacy research.

The borrowing of physical medical and pharmaceutical books is shaped by a multidimensional, context-adaptive framework of factors, encompassing users’ professional needs, book attributes, historical continuity of physical text value, and external interest spillovers. These determinants do not operate in isolation; instead, they interact dynamically—with events like public health crises or technological shifts reshaping their relative influence (e.g., pandemic-driven digital adoption temporarily weakening some physical book demand, while post- pandemic "return to routine" restored reliance on tangible resources). Centering on the following core research questions, this study takes non-medical disciplines as its analytical lens, redefines health literacy as a pedagogical and social construct, and thereby challenges the long-standing medical-centric paradigms in health literacy research:

Q1: How have the latent relationships among time-related (e.g., months), space-related (e.g., lending rooms), and interest-related (R1–R9) variables evolved across the periods 2016–2018, 2019–2021, and 2022–2024—specifically in terms of factor quantity, category refinement, variable loading intensity, and the stability/mobility of key variables’ factor affiliation?

Q2: What temporal trends characterize the main effect of Factor 1 (F1) on the logarithm of total book borrowings (LogTotal) across 2016–2018, 2019–2021, and 2022–2024, and to what extent do these trends reflect changes in the predictive power of F1 amid external contextual shifts (e.g., post-pandemic service model adjustments)?

Q3: How does the moderating effect of “Duration of Borrowing Years” on the relationship between F1 and LogTotal vary across 2016– 2018, 2019–2021, and 2022–2024—specifically in terms of the interaction term coefficient, R² change, conditional effect, and overall contribution to explaining LogTotal variance?

Q4: What evolutionary trajectory do the high-frequency title words in psychology and related disciplines exhibit across 2016–2018, 2019– 2021, and 2022–2024, and how do changes in their clustering patterns and dimension-explained variance reflect shifts in research themes and academic interests?

Q5: What are the directional and magnitude characteristics of the correlations between “Number of Borrowings,” “Collection Tenure,” and “Secondary Categories of R” for books spanning 2016–2024, and what do these correlations reveal about the influence of book age and category on borrowing behavior?

Q6: To what extent do the cumulative survival probability, median survival time, and critical failure nodes differ among the secondary categories of R books (e.g., R2, R7) during the borrowing cycle, and how might these differences relate to category-specific content attributes and audience demand stability?

Q7: How has the mean book borrowing volume of distinct reader types (Undergraduate Students, Graduate Students, Faculty & Staff) changed across 2016–2018, 2019–2021, and 2022–2024, and what do these trends imply about variations in resource demand and access preferences among reader groups amid external disruptions (e.g., pandemic, rise of e-resources)?

The study exhibits notable innovations in research perspective, analytical framework, and variable design, addressing gaps in understanding the dynamics of physical medical/pharmaceutical book borrowing. It divides the research period (2016–2024) into three distinct phases (pre-pandemic: 2016–2018; pandemic: 2019–2021; post-pandemic: 2022–2024) to systematically trace shifts in borrowing determinants— revealing a transition from "pandemic-driven complexity" to "post-pandemic simplicity" while highlighting the persistent impact of sustained user attention (a long-overlooked factor in borrowing behavior research). Unlike studies focusing solely on single factors (e.g., time or user demographics), it integrates three core dimensions (time: monthly borrowing fluctuations; space: physical library location preferences; interest: borrowing frequencies of medical subcategories) to holistically characterize borrowing behavior, capturing interconnected drivers of user decisions. It defines "sustained user attention" as "Duration of Borrowing Years" and uses this as a moderator—filling the gap of neglecting long-term user engagement in borrowing research. For non-medical universities (e.g., Nanjing Normal University, focused on teacher education), it consolidates fragmented R7 subcategories (e.g., R71 Obstetrics and Gynecology, R73 Oncology) into "R7 Clinical Specialized Medicine" to avoid data dispersion, ensuring statistical consistency while reflecting real-world low demand for granular medical subcategory borrowing. It combines quantitative (moderated regression, correlation analysis, survival analysis) and qualitative (title word visualization) techniques. PCA reduces multi-dimensional data complexity, while Voyant Tools-based title visualization intuitively reveals thematic clustering of borrowed books—linking statistical patterns to semantic content of materials.

The study holds multi-faceted significance across academic, practical, and societal/interdisciplinary realms. Academically, it addresses gaps in temporal dynamics research by dividing 2016–2024 into pre-pandemic, pandemic, and post-pandemic phases, revealing the shift in physical medical/pharmaceutical book borrowing determinants from "pandemic-driven complexity" to "post-pandemic simplicity"—a novel lens for studying disruption-induced user behavior changes. It also innovatively operationalizes "sustained user attention" as "Duration of Borrowing Years" (a moderator) to fill the void of neglecting long-term user engagement in prior studies, while advancing library and information science (LIS) research via a multi-dimensional, mixed-methods framework (integrating time, space, and interest data with PCA, survival analysis, and title visualization). Practically, it guides evidence-based library management for non-medical universities (e.g., Nanjing Normal University): it identifies high-demand subcategories and low-demand ones to optimize collection development, pinpoints key user groups (undergraduates, graduates) and spatial borrowing preferences to refine space/service allocation, and offers strategies (e.g., personalized recommendations) to reverse declining post-pandemic borrowing. Societally and interdisciplinarity, its subcategory borrowing data reflects shifts in public health awareness (e.g., R1 "Preventive Medicine" trends), highlights interdisciplinary medical knowledge demand (e.g., psychology students’ borrowing), and draws attention to access inequities (e.g., high engagement from faculty/staff) to inform inclusive service design, fostering public health literacy and knowledge equity.

2. Previous Research

2.1 Professional Demand: The Non-Negotiable Core Driver

Professional needs of medical practitioners and students stand as the foundational determinant of physical book borrowing, as these resources directly underpin clinical competence, academic progression, and lifelong learning. For newly qualified doctors, physical medical books act as "transitional tools" to bridge the gap between theoretical education and real-world clinical practice—borrowing behavior here is motivated not just by cost avoidance but by the need for trusted, portable references during ward rounds or emergency consultations (Dixit, 2004). Similarly, "must-read" core textbooks (e.g., standard anatomy or pharmacology texts) occupy an irreplaceable role in medical education: their authoritative content and structured layout make them prioritized borrowing choices, as users perceive them as non- negotiable for building professional competence (Nair, 2003). Junior faculty in specialized fields further reinforce stable demand for physical books. Discipline-specific guidebooks help these early-career professionals navigate academic department cultures, align with clinical protocols, and adapt to teaching responsibilities—creating consistent borrowing trends that persist even amid digital resource expansion (Gabbe, 1996). While professional demand is universally recognized as a core driver, existing studies primarily focus on "stable" professional contexts (e.g., pre-pandemic clinical practice). They rarely address how disruptive events (e.g., COVID-19’s shift to telemedicine) temporarily redefines "professional needs"—for instance, during lockdowns, practitioners may have prioritized digital resources for remote access, but post-pandemic, the return to in-person care reignited demand for physical books as "quick-reference tools" in clinics. This gap limits understanding of how professional demand adapts to crisis and recovery phases.

The inherent characteristics of physical medical books—format, content quality, and professional reputation—act as user choice filters, significantly influencing which titles are borrowed. Positive reviews from authoritative sources enhance a book’s visibility and credibility, positioning it as a "high-value" resource for users seeking rigor (Dar et al., 2022). Additionally, the evolution of textbook content and format in response to practice needs shapes borrowing trends: for internal medicine textbooks, historical shifts toward practical, core-focused content align with user preferences, explaining why certain titles rise to the top of borrowing lists (Rozman, 1998). For example, the pandemic increased demand for physical books with virus-specific content (e.g., quick-reference guides to COVID-19 treatment), even if those books lacked traditional "prestige" (e.g., peer-reviewed status). Existing studies rarely explore how crises redefine "valuable attributes"—a gap that hinders understanding of post-pandemic borrowing shifts.

External interest in health knowledge acts as a borrowing catalyst, indirectly driving demand for physical medical books. Journalistic texts covering International Classification of Diseases (ICD-10) disease categories (e.g., news articles on diabetes prevention, cancer treatment) foster broad public and professional interest in health topics, prompting users to seek deeper insights via physical medical books (Gonzalez- Garcia et al., 2020). For example, a news feature on emerging cardiovascular treatments may lead readers to borrow specialized cardiology textbooks from libraries to learn more—a "spillover effect" that links mainstream health media to academic resource use.

2.2 Medical/Pharmaceutical Book Reading Preferences and Influencing Factors

Reading preferences are shaped by individual, socioeconomic, and demographic factors, and they indirectly steer borrowing behavior by defining which resources users perceive as accessible or useful. These preferences are not fixed; they evolve with technological change, educational context, and even crises (e.g., pandemic-driven digital adaptation). While digital resources have gained prominence, physical books remain preferred for specific use cases, creating a context-dependent "dual use" pattern. Among healthcare professionals (e.g., radiologists, radiographers), 80% prefer physical books for learning—valuing their ability to support deep engagement (e.g., annotating chapters, cross-referencing sections)—even though 64% use portable devices for clinical tasks (Awais et al., 2019). Similarly, South Korean dermatology professionals reject grayscale e-readers for journal reading: only 3 of 31 respondents preferred digital formats, citing poor color reproduction (critical for analyzing skin conditions) as a barrier (Choi et al., 2014). U.S. health science library users further illustrate this contextuality: 55.4% use e-books for quick reference (valuing searchability), while 67.2% borrow print books for extended study (prioritizing readability and reduced eye strain) (Melssen, 2012). In contrast, U.S. medical students exhibit strong digital preference: 77% never borrow printed books, and 83% view online resources as "ideal"—suggesting generational or educational context (e.g., digital-native learning environments) drives this shift (Conway, 2017). The "physical vs. digital" debate in existing research often frames preferences as binary, rather than complementary. Post-pandemic data suggests users increasingly adopt "hybrid" approaches (e.g., using e-books for remote study, physical books for in-clinic reference), but few studies explore this middle ground.

2.3. Historical and Critical Reading: Contextualizing Physical Book Relevance

The historical evolution of medical literature and critical reading practices provide foundational context for understanding why physical books remain relevant to borrowing. While not direct determinants, these dimensions explain how users interact with physical texts and why they continue to prioritize them—even amid digital expansion. Medical literature has transformed across centuries, but physical texts have consistently served as knowledge legacy carriers. Western medieval medical texts (12th century onward) were produced for diverse readerships (students, practitioners, the public) in settings ranging from universities to royal courts—reflecting early patterns of resource specialization that mirror modern borrowing trends (e.g., students prioritizing textbooks, clinicians prioritizing reference guides) (Nicoud, 2024). Ancient codices (e.g., a 4th-century Greek-Coptic text integrating medical and magical rituals) further highlight physical books’ role in preserving pre-modern knowledge, even as disciplinary boundaries blurred (Zellmann-Rohrer & Love, 2022). Syriac translations of Galen’ s works similarly demonstrate physical texts’ role in cross-cultural knowledge exchange—transmitting Greek medical insights to Arabic regions, a function modern physical books continue to serve (e.g., preserving TCM knowledge in Chinese libraries) (Bhayro et al., 2013; Li et al., 2020; Tian & Liu, 2021).

Critical engagement with medical literature requires assessing study design, biases, and relevance—skills fostered through physical books. Healthcare professionals rely on physical texts for curated, peer-reviewed content, avoiding the quality variability of digital resources (Li et al., 2024). Medical education further integrates physical books into critical reading training: courses like "Literature: A Healing Art" use literary works to build empathy and analytical thinking, which students then apply to textbook case studies (Sirridge & Welch, 2003). Active training (e.g., group discussions of textbook chapters) also enhances skill retention: medical residents in participatory programs retained critical reading abilities seven months post-training, unlike those in lecture-based programs (Cobos-Aguilar et al., 1998). Medical humanities further bridge literary analysis and clinical practice, helping users contextualize textbook content with patient perspectives (Bolton, 2005).

Historical research focuses on "knowledge preservation" but rarely links this to modern borrowing. For example, do researchers borrow historical physical medical books to study knowledge evolution, and has this demand changed post-pandemic (e.g., increased interest in historical pandemic management texts)? This gap limits understanding of physical books’ "niche" role in contemporary borrowing. While critical reading research highlights physical books’ role, it rarely addresses how digital tools complement this. Post-pandemic, medical education increasingly uses hybrid models (e.g., digital annotations of physical texts), but few studies explore how this impacts critical reading and borrowing of physical books.

2.4. Intersection of Medicine/Pharmacy with Psychology: Indirectly Shaping Borrowing Demand

The intersection of medicine/pharmacy with psychology, sociology, and other disciplines shapes broader knowledge needs, indirectly influencing which physical books are borrowed. By defining the content users seek (e.g., texts integrating psychology and clinical care), these intersections create niche demand for specialized physical resources—even if they do not directly drive borrowing. Psychological frameworks integrated into clinical practice create demand for interdisciplinary physical books. In lifestyle medicine, hope theory elicits chronic disease behavior change, while exercise medicine uses behavioral psychology to address inactivity—driving interest in texts that synthesize these fields (Duncan et al., 2021; Bethell & Brodie, 2023). Cardiology similarly requires psychological input for atrial fibrillation (AF) management (e.g., improving patient adherence), increasing demand for books on integrated cardiac care (Sears et al., 2022). Palliative care further relies on trauma-informed care to reduce COVID-19 distress, creating need for texts on mental health competencies for HPM physicians (Brown et al., 2020; Podgurski et al., 2025). Psychological principles shape professional development needs, driving borrowing of specialized physical books. Forensic odontology practitioners need stress management training, increasing interest in texts on mental health for specialized fields (Sebastian et al., 2023). CME programs benefit from cognitive psychology integration, but current gaps create demand for books on CME design (Setia et al., 2024). Organizational psychology also reduces burnout via entrepreneurial orientation, driving interest in texts linking organizational behavior to healthcare (Kearney et al., 2020).

Specialized medical fields increasingly rely on interdisciplinary knowledge, creating niche demand for physical books that synthesize diverse disciplines — indirectly shaping borrowing trends. In psychoneuroimmunology (PNI), multi-omics techniques clarify links between psychological stress, the nervous system, and immune function, advancing mind-body medicine. This progress has heightened interest in texts that explain PNI methodologies and clinical applications, leading healthcare researchers and students to borrow physical books that bridge immunology, neuroscience, and psychology (Mengelkoch et al., 2023). Dermatology and psychology intersect in neurocosmetics, a field that targets the skin-brain axis to improve both skin health and emotional well-being. The integration of skin microbiome research and artificial intelligence in personalized skincare further expands this interdisciplinary focus, driving demand for physical books that translate complex scientific concepts into practical clinical guidance (Haykal et al., 2025). Mental health care also reflects this dynamic: religious beliefs (a psychological and cultural factor) reduce mental health service utilization due to stigma, prompting interest in texts that explore collaboration between psychiatry and religious leaders—creating a niche borrowing category for physical books on faith-based mental health care (Verduin & Tower, 2024).

Research on interdisciplinary demand focuses on "content need" but rarely addresses how crises like the pandemic amplified or diminished this demand. For example, during COVID-19, did the surge in mental health distress among patients and healthcare workers increase borrowing of PNI or neurocosmetics texts (e.g., for stress-related skin conditions)? Post-pandemic, did this demand stabilize or decline? Understanding these shifts is key to explaining how "complex" interdisciplinary borrowing during crises simplifies into more focused demand in recovery phases—but existing studies lack this temporal perspective.

2.5. Synthesis of Gaps and Alignment with the Present Study

The existing literature provides a robust foundation for understanding the determinants of physical medical/pharmaceutical book borrowing, reading preferences, and contextual factors (e.g., historical legacy, interdisciplinary needs). However, many studies focus on stable pre- pandemic contexts or isolated snapshots of behavior (e.g., pandemic-era digital adoption) but fail to examine longitudinal shifts across crisis and recovery phases. Book attribute research does not explore how pandemic-specific needs redefined "valuable" features, nor how these preferences simplified post-pandemic. Existing literature emphasizes static drivers (e.g., professional needs, socioeconomic status) but rarely addresses sustained user attention—the enduring interest in specific physical book categories (e.g., core textbooks, specialized research texts) that persists across crises and recovery. The present study addresses these gaps by: Examining longitudinal shifts in determinants from pandemic complexity (e.g., fragmented demand for digital and pandemic-specific physical texts) to post-pandemic simplicity (e.g., focused demand for core, trusted physical books). Highlighting sustained user attention as a persistent factor that stabilizes borrowing trends amid change. Unpacking hybrid format preferences to explain how post-pandemic borrowing becomes less complex, not less relevant. By addressing these gaps, the study advances understanding of physical medical/pharmaceutical book borrowing as a dynamic, context-adaptive behavior—one that evolves with crises but retains core continuity through sustained user attention.

3. Data and Method

3.1 Sample

The research sample consists of book borrowing records from the Library of Nanjing Normal University, covering three distinct time periods with the following specific data:

2016 - 2018: There are 1560 borrowers, involving 1355 books, and the total number of borrowings reaches 5231. 2019 - 2021: The number of borrowers decreases to 826, with 887 books involved, and the total borrowings are 2231. 2022 - 2024: There are 547 borrowers, 694 books involved, and the total borrowings amount to 1599.

These data collectively reflect the changes in the borrowing situation of the library's books over the years, providing a basis for analyzing the reading behavior trends of users during different periods.

In the Chinese Library Classification (abbreviated as "CLC"), the category R, representing "Medicine and Hygiene", has its secondary classifications divided according to different fields and research directions of medical disciplines. Specifically, R1 encompasses Preventive Medicine and Hygiene, covering areas like public health, disease prevention, environmental health, occupational health, and nutrition and food hygiene; R2 refers to Chinese Medicine, including traditional medical categories such as Chinese medicine, Chinese materia medica, acupuncture, tuina (Chinese massage), and ethnic medicine (e.g., Tibetan medicine, Mongolian medicine); R3 involves Basic Medicine, incorporating basic medical disciplines like human anatomy, histology and embryology, physiology, biochemistry, pathology, microbiology, and immunology; R4 denotes Clinical Medicine, containing basic theories and practices of diagnostic science, therapeutics, and various clinical specialties (e.g., common content in internal medicine, surgery, etc.); R5 is Internal Medicine, subdivided into specialties such as cardiovascular medicine, respiratory medicine, gastroenterology, hematology, endocrinology, and infectious diseases; R6 covers Surgery, including branches like general surgery, neurosurgery, thoracic and cardiac surgery, orthopedics, urology, and plastic surgery; R8 represents Special Medicine, encompassing medical fields in special areas such as military medicine, aerospace medicine, navigation medicine, diving medicine, and forensic medicine; and R9 involves Pharmacy, including pharmacy-related disciplines like medicinal chemistry, pharmaceutics, pharmacology, drug analysis, clinical pharmacy, and pharmaceutical administration.

It is noteworthy that there is no standalone secondary category designated as "R7" within the broader "R" category (encompassing Medicine and Hygiene). Instead, classifications commencing with "R7" correspond to a series of distinct secondary categories that collectively focus on specialized clinical disciplines. These include R71 (Obstetrics and Gynecology), R72 (Pediatrics), R73 (Oncology), R74 (Neurology and Psychiatry), R75 (Dermatology and Venereology), R76 (Otorhinolaryngology), R77 (Ophthalmology), R78 (Stomatology), and R79 (Foreign Ethnic Medicine). Given that Nanjing Normal University is centered on training pre-service teachers and does not offer medical-related majors, there is little need to subdivide medical books. If the various clinical specialized medicine categories under R7 were counted separately, it would lead to excessive data dispersion, which is not conducive to efficient analysis. For the purpose of statistical consistency and analytical coherence, the present study consolidates these specialized clinical subcategories under the umbrella term "R7 Clinical Specialized Medicine." coherence, the present study consolidates these specialized clinical subcategories under the umbrella term "R7 Clinical Specialized Medicine."

Figure 1. The circulation volume of the secondary classifications (R1 - R9) of R - type books

Figure 1 presents the circulation volume of the secondary classifications (R1 - R9) of R - type books. Different colors represent different reader groups (Undergraduate Students, Graduate Students, Faculty & Staff, and Others). The circulation volume of R7 is far higher than that of other classifications, making it the core category for the circulation of R - type books. There is a certain amount of circulation for R1 and R3, a small amount for R2, R4, and R9, and an extremely low volume for R5, R6, and R8. This indicates that the demand for R - type books is concentrated in specific secondary classifications. Undergraduate Students (blue) account for a high proportion in R1, R3, and R7, being the main reader group for these classifications. Graduate Students (orange) have relatively more circulations in R1, R3, R7, and a few other classifications, serving as an important force. Faculty & Staff (gray) and “Others” (yellow) make a relatively small overall contribution, with only sporadic circulations in individual classifications.

3.2 Research Design

3.2.1 Moderated Regression of Borrowing Behavior

This study centers on the book borrowing data of NNU Library spanning 2016–2024. The data matrix integrates multiple dimensions tied to readers and their borrowing behaviors, covering three main phases: 2016–2018, 2019–2021, and 2022–2024. Figure 2 visually presents the research design, illustrating the flow from data dimensions through PCA to the regression analysis framework, with clear demarcation of variable roles and data processing steps.

3.2.1.1  Core Dimensions of Data

The dataset encompasses three interrelated dimensions to characterize borrowing behaviors comprehensively:

Time dimension: Captured via monthly borrowing quantities (e.g., Jan, Feb, …) across the three periods. It tracks how borrowing frequencies fluctuate monthly within each time frame, reflecting temporal patterns in reader engagement.

Space dimension: Operationalized through variables like “Suyuan Location,” “Jingwen Location,” and “Compact Storage Area.” These denote physical - location - based borrowing preferences, accounting for spatial variability in access.

Interest dimension: Represented by variables R0–R9 (signifying borrowing frequencies for distinct interest - related book categories) and “Total Borrowing Volume,” which aggregates overall borrowing activity.

3.2.1.2  Analytical Workflow

Principal component analysis (PCA): Before regression modeling, Principal Component Analysis (PCA) condenses the multi - dimensional data. It extracts key components F1 (TSI/TS), F2 (TSI/TI), F3 (TS/SI)) et al., where TSI likely denotes a composite index integrating Time, Space, and Interest dimensions. These components reduce data complexity while preserving critical variance, serving as summarized indices for subsequent analysis. While varimax rotation is simple and easy to implement, its rigid assumption of orthogonal factors makes it ill-suited for the interrelated, complex nature of library loan data. Optimal oblique rotation, by allowing factor correlation, enhancing interpretability, preserving critical data variance, and adapting to dynamic scenarios, produces more accurate, actionable factor models. The purpose of this study is to leverage data to improve user experience, optimize collections, and refine service strategies; therefore, optimal oblique rotation is a more powerful and contextually appropriate choice.

Regression analysis framework:

Moderator: Duration of Borrowing Years. For each reader, count whether they have borrowed books in Category R each year (0 for non- borrowing, 1 for borrowing). Then, aggregate the "Duration of Borrowing Years" by three time periods: 2016-2018, 2019-2021, and 2022- 2024. This indicator reflects readers' sustained attention to books in Category R.

Independent Variables: F1 derived from original multi - dimensional data (post - PCA transformation), it captures the combined effects of Time, Space, and Interest patterns.

Covariates: Gender (1 for male, 0 for female), Reader Type (0 for other, 1 for undergraduate student, 2 for graduate student, 3 for faculty and staff), Department (1 for Psychology School, 0 for Other) and Renewal act (1 for renewal, 0 for non-renewal) as covariates, controlling for potential confounding effects on borrowing behaviors (e.g., differences in borrowing habits between genders or academic roles). Although there is no medical school, the School of Psychology has the highest number of R-category book borrowings and borrowers at NNU, which reflects the interdisciplinary nature of the discipline. Therefore, it is assigned a value of 1, while other departments are assigned a value of 0.

Dependent Variable: LogTotal (presumably a log - transformed total borrowing volume) is the outcome to be explained, linking it to the structured independent variables and covariates. The logarithmic transformation of total borrowings into LogTotal serves three key purposes. First, it mitigates the impact of extreme values (e.g., abnormally high borrowing volumes of popular books) and reduces data skewness, making the distribution of borrowing data closer to a normal distribution—this is critical for meeting the normality assumption of linear regression models, thereby improving the reliability of parameter estimates. Second, it converts multiplicative relationships in the original borrowing data into additive ones, simplifying the interpretation of model coefficients (e.g., a 1-unit change in the predictor corresponds to a percentage change in total borrowings, rather than an absolute change). Third, it stabilizes the variance of borrowing data across different periods or groups, avoiding heteroscedasticity issues that could distort the model’s statistical inference, and ensuring consistent explanatory power of the model for both low and high borrowing volume samples.

This design integrates time - series, spatial, and interest - based borrowing data. Leveraging PCA for dimensionality reduction, it constructs a regression framework with clearly defined variable roles—moderator, independent variables, covariates, and dependent variable—to dissect factors driving library book - borrowing behaviors from 2016 to 2024.

3.2.2 Title Words Visualization

As the core condensed carrier of a text, a title directly reflects the core theme, research focus, or information focus of the content through its contained vocabulary. Conducting word frequency and clustering analyses on titles is essentially a process of exploring the thematic structure and information patterns of a text corpus by means of data-driven methods. Based on the semantic similarity or co-occurrence relationship of words in titles, clustering analysis groups titles with similar themes into the same cluster. Its core value lies in transforming dispersed vocabulary into a structured thematic group, thereby revealing the inherent logical connections within the text corpus.

The visualization is generated using Voyant Tools, a web - based platform for digital text analysis. Its creation hinges on a sequence of interconnected analytical steps. Translate the titles of R books borrowed by readers into English and enter them into the text box. Initially, the input digital texts undergo preprocessing, which includes tokenization to break the text into individual words and word frequency counting. Subsequently, semantic vectorization techniques, such as word embedding or topic modeling, are employed to transform these words into high - dimensional semantic vectors. These vectors capture the semantic relationships between words, with words of similar meaning positioned closer in the vector space. To enable visualization in a two - dimensional plane, dimensionality reduction methods are utilized. The X and Y coordinates correspond to the reduced principal components, so the position of each word in the 2D space reflects its semantic associations, with semantically similar words clustering together. The fill color is mapped to the third principal component for further semantic differentiation. Word size is typically related to word frequency, with larger points representing more frequently occurring words. The visualization provides an intuitive spatial representation of the semantic relationships and thematic clustering within the R book titles, facilitating the identification of key topics and semantic structures.

3.2.3  Correlation Analysis

The study conducts a correlation analysis of R book data from 2016 to 2024 via SPSS, incorporating bootstrap validation. The dataset includes 2920 books with three key variables:

Number of Borrowings: Total times a book was borrowed.

The length of time since being added to the collection: Temporal metric of book presence in the library.

Secondary Categories of R: Categorical classification of books.

The analytical process begins with correlation analysis setup: in SPSS, navigate to "Analyze" → "Correlate" → "Bivariate," move the three variables into the "Variables" box, select Pearson correlation, enable 2-tailed significance testing, and check "Flag significant correlations" to highlight results significant at the 0.01 level (denoted by **). To enhance result robustness, bootstrap resampling is applied through the following configuration: go to "Analyze" → "Correlate" → "Bivariate" → "Bootstrap," select "Perform bootstrap" to activate resampling, and set the specifications as simple random sampling for the sampling method, 1000 samples (balancing accuracy and computation time), a 95.0% confidence interval level (standard for statistical inference), and percentile as the confidence interval type (for nonparametric interval estimation).

3.2.4 Survival Analysis of R Books

For all R-category books that readers borrow between 2016 and 2024, researcher codes a yearly borrowing status for each book: assign 1 if the book is borrowed in a given year and 0 if it is not. Using these yearly codes, researcher aggregates the "Length of Borrowing Years" for each R-category book. This metric quantifies the sustained popularity of individual R-category books and serves as the "Time" variable in subsequent survival analysis. When multiple records correspond to the same category and require frequency statistics, case weighting is a prerequisite. This process involves navigating to "Data" – "Weight Cases", transferring the variable "Number of Borrowings" to the "Frequency Variable" box, and confirming the operation by clicking "OK". The analysis is initiated by selecting "Analyze" – "Survival Analysis" – "Life Table" from the menu options. The variable "Length of Borrowing Years" is moved to the "Time" box to define the time metric. The variable "Whether Borrowed in 2024" is transferred to the "Status" box, followed by clicking "Define Event" to specify the value corresponding to the occurrence of borrowing (e.g., 1 for borrowed, 0 for not borrowed) based on its coding scheme. For group comparisons, the variable "Secondary Categories of R" is placed in the "Factor" field. In the "Display Time Intervals" section, a suitable time range (0–9) and step size (1) are set according to the actual range of "survival years" and the specific requirements of the analysis. Once all parameters are configured, the results are generated to facilitate the survival analysis of the books.

4. Results

4.1 The Extracted Factors

Table 1 reveals the dynamic adjustments of latent relationships among time-, space-, and interest-related variables across different periods. The expansion of factor quantity (from 3 to 6) and the refinement of categories (from 3 to 4 unique combinations) suggest that the relationships between time, space, and interest variables have become more complex and differentiated since 2022. This may be driven by external changes (e.g., post-pandemic adjustments to service models, shifts in user demand) that have increased the independence and specificity of individual variable dimensions.

Table 1. Factor loadings of time-, space-, and interest-related variables by periods (2016–2024)

a. For oblique factor rotation (e.g., Promax), the phenomenon that the factor loading of "July (1.012)" in Factor 2 of the 2016-2018 period exceeds 1 is directly related to the core characteristic of oblique rotation—allowing correlations between latent factors. This constitutes a key difference distinguishing it from orthogonal rotation (e.g., Varimax, where factor loadings are strictly constrained within the range of [-1, 1]).

b. Factor loadings greater than 0.71 indicate a significant influence; those ranging from 0.32 to 0.71 represent a moderate influence; and when factor loadings are less than 0.32, the association is generally considered non-significant.

Factor loadings (a measure of the strength of association between a variable and a factor) and the composition of variables within each factor vary across periods, particularly in terms of loading significance (refer to Note b: >0.71 = significant; 0.32–0.71 = moderate; <0.32 = non-significant) and variable type concentration (see Table 1):

4.1.1 Time-Related Variables (Months)

2016–2018: Time variables are scattered across all 3 factors, with only 2 variables showing significant loadings: February (0.899, F1) and September (0.881, F1). Most others (e.g., December: 0.526, F2; March: 0.406, F3) have moderate loadings, indicating weak to moderate association with their respective factors.

2019–2021: Time variables are concentrated in F1 and F3, with 2 variables showing significant loadings: July (0.836, F1) and September (0.843, F3). June (0.803, F2) also reaches significant loading, reflecting a strengthened association between time variables and factors.

2022–2024: Time variables are spread across 5 factors (F1, F2, F4, F5, F6), with 5 variables showing significant loadings: July (0.997, F1), August (0.988, F1), April (0.864, F4), June (0.803, F4), and R2 (0.979, F5, note: R2 is treated as a time-related proxy variable here). The average loading intensity increases (e.g., July’s loading rises from 0.836 in 2019–2021 to 0.997), indicating that time variables have become more strongly associated with specific factors and play a more critical role in defining factor characteristics.

4.1.2 Space-Related Variables (Lending Rooms)

2016–2018: 2 space variables are included (Suiyuan Chinese Book Lending Room: 0.654, F1; Jingwen Chinese Book Lending Room: 0.698, F3), both with moderate loadings—no significant associations.

2019–2021: 2 space variables are included (Jingwen Chinese Book Lending Room: 0.495, F1; Bio-geography Book Lending Room: 0.933, F2). Notably, Bio-geography Book Lending Room shows a significant loading (0.933), becoming the core variable defining F2.

2022–2024: 3 space variables are included (Suiyuan Chinese Book Lending Room: 0.390, F1; Jingwen Chinese Book Lending Room: 0.651, F2; Suiyuan Hong Kong-Taiwan Book Lending Room: 0.907, F6). Suiyuan Hong Kong-Taiwan Book Lending Room (newly added) has a significant loading (0.907), while existing space variables maintain moderate loadings. The addition of a new space variable and the significant loading of the new variable reflect the expanded coverage and enhanced explanatory power of space-related factors.

4.1.3 Interest-Related Variables (R1–R9)

2016–2018: 3 interest variables are included (R3: 0.805, F1; R7: 0.485, F2; R9: 0.828, F3). R3 and R9 have significant loadings, making them core variables for F1 and F3, respectively.

2019–2021: 4 interest variables are included (R5: 0.800, F1; R1: 0.526, F1; R2: 0.496, F1; R7: 0.456, F3). R5 shows a significant loading (0.800), while others have moderate loadings—interest variables become more concentrated in F1.

2022–2024: 4 interest variables are included (R1: 0.851, F2; R3: 0.749, F3; R4: 0.518, F4; R9: 0.480, F3). R1 (0.851) and R3 (0.749) have significant loadings, and interest variables are split across F2, F3, and F4—no longer concentrated in a single factor, indicating more differentiated interest preferences linked to different time/space contexts.

4.1.4. Shifts in Key Variables’ Factor Affiliation: Stability vs. Mobility

Key variables (those with high loadings or consistent presence across periods) show different patterns of factor affiliation, reflecting changes in their latent associations with time/space/interest dimensions:

R9: Appears in all three periods, consistently associated with space-related factors: F3 (2016–2018: 0.828), F2 (2019–2021: 0.923), F3 (2022– 2024: 0.480). Although its loading decreases in 2022–2024, its affiliation with space-related factors remains unchanged, indicating a stable latent relationship between R9 (interest) and space variables.

September: Appears in 2016–2018 (F1: 0.881) and 2019–2021 (F3: 0.843), and re-emerges in 2022–2024 (F5: 0.350). It is consistently associated with time-related factors, reflecting September’s stable role as a representative time variable across periods.

July: Shifts across factors: F2 (2016–2018: 1.012, note: exceeds 1 due to oblique rotation), F1 (2019–2021: 0.836), F1 (2022–2024: 0.997). Its loading intensity increases over time, and it shifts from a "Time-Interest" factor (F2) to a "Time-Space-Interest" (F1) and then "Time-Space" (F1) factor—indicating that July’s latent association has shifted from interest-centric to a combination of time and space, possibly due to seasonal changes in user behavior (e.g., summer vacation adjustments to space usage).

Suiyuan Chinese Book Lending Room: Shifts across factors: F1 (2016–2018: 0.654), F3 (2019–2021: 0.736), F1 (2022–2024: 0.390). It moves from a "Time-Space-Interest" factor to a "Time-Space" factor, and its loading decreases in 2022–2024—suggesting a weakened association with time/interest variables, possibly due to reduced usage frequency or functional adjustments of this lending room.

4.2 Linear Moderation Effect Analysis

4.2.1 Comparison of Basic Information and Model Fit Across Three Periods

Across three periods (2016–2019, 2019–2021, 2022–2024), the comparison and interpretation of linear moderated regression analysis results rely on Model 1 of the SPSS PROCESS macro. The dependent variable Y is defined as LogTotal (logarithm of total book borrowings), the focal predictor X is F1, and the moderator W is Duration of Borrowing Years. Sample size decreased progressively; the 2022–2024 period accounted for only 35.1% of the first period. Table 2 presents the regression coefficients from the moderated regression analysis. The 2016- 2018 and 2019-2021 periods use an identical set of 6 covariates. In contrast, the 2022-2024 period expands the covariate set to 9 by adding three new variables: F4, F5, and F6.

The R² value, which measures how much of the variation in LogTotal is explained by the model, varies significantly across periods. The 2016-2018 period has the highest R² at 0.7195, indicating an excellent fit between the model and the data. The 2019-2021 period experiences a sharp drop in R² to 0.5681, a decrease of approximately 15 percentage points. The 2022-2024 period rebounds to an R² of 0.6872, which is about 12 percentage points higher than 2019-2021 but still 3 percentage points lower than 2016-2018. This fluctuation suggests that the relationship between the predictors and LogTotal was strongest in pre-pandemic period, weakened during the outbreak of the pandemic, and partially recovered in the post-pandemic period.

All three models are globally significant (p<0.001), meaning the combined set of predictors has a significant effect on LogTotal. However, the F-value, which quantifies the strength of this overall significance, decreases steadily. The 2016-2018 period has the highest F-value at 441.8283, indicating an extremely strong joint explanatory effect of the predictors. The 2019-2021 period’s F-value drops to 119.2659, less than one-third of the earlier value, showing a marked weakening in the combined predictive power of the variables. The 2022-2024 period’ s F-value further falls to 97.7680, the lowest among the three. This downward trend in F-value is closely linked to two factors: the shrinking sample size (smaller samples tend to reduce the precision of parameter estimates, lowering the F-value) and the increase in covariates in 2022-2024 (adding more variables can dilute the joint explanatory power of the core predictors, even if the new covariates themselves are significant).

4.2.2 Comparison of the Main Effects of Predictor Variable (F1) Across Three Periods

The main effect of F1 (the focal predictor variable) on LogTotal (the outcome variable) exhibits a clear and consistent downward trend across the three periods (2016-2018, 2019-2021, and 2022-2024), though it remains statistically significant (p<0.001) in all phases. In 2016- 2018, F1 has the strongest main effect on LogTotal: its coefficient reaches 0.4772. By 2019-2021, this effect weakens sharply—the coefficient drops to 0.2730, a 42.8% reduction from the earlier period. The downward trend continues into 2022-2024: the coefficient decreases further to 0.2373, a 13.1% decline compared to 2019-2021. Overall, the main effect of F1 in 2022-2024 is only about half (49.7%) of what it is in 2016-2018, which indicates a substantial long-term weakening of F1’s ability to drive changes in LogTotal. The relevant parameters are shown in Table 2.

Table 2. Regression coefficients

4.2.3 Comparison of Moderating Effects (F1 × Duration)

The moderating effect is the core of the model. It reflects the moderating role of "Duration" (W) on the path from F1 (X) to LogTotal (Y) through three key indicators: the interaction term coefficient (Int_1), the R² change (R²-chng), and the conditional effect. Across the three periods, the interaction term (F1 × Duration) remains statistically significant (p<0.001), but its strength and contribution to explaining LogTotal variation differ markedly (see Table 2):

2016-2018: The moderating effect is the strongest. The interaction term coefficient is -0.1515 (significantly negative) with a small standard error of 0.0071, leading to a high absolute t-value of 21.4341. The R² change reaches 0.0831, meaning the interaction term alone explains 8.31% of the variation in LogTotal. Additionally, the F-value for the interaction term is as high as 459.4216, confirming that the moderating role of "Duration" makes the largest contribution to the model.

2019-2021: The moderating effect weakens significantly. The absolute value of the interaction term coefficient drops to 0.0862, a 43.1% decrease from 2016-2018. Correspondingly, the R² change falls to 0.0521 (a 37.3% reduction), and the F-value plummets to 98.3842. These changes indicate that the moderating contribution of "Duration" is halved compared to the earlier period.

2022-2024: The moderating effect shows a pattern of "coefficient reversal in magnitude but declining contribution". Although the absolute value of the interaction term coefficient rises to 0.1766 (a 104.9% increase from 2019-2021), the R² change drops to only 0.0303—merely 36.5% of the 2016-2018 level. The F-value further decreases to 51.7562, confirming that the actual moderating contribution of "Duration" continues to weaken, despite the larger absolute coefficient.

The 16th, 50th, and 84th percentiles of "Duration" are all 1.0000 across the three periods, allowing for direct comparison of the conditional effects:

2016-2018: The conditional effect is the strongest. The effect value is 0.3257 with a small standard error of 0.0124, resulting in a high t- value of 26.3569 (p<0.001). The 95% confidence interval [0.3015, 0.3500] is far from zero and narrow, indicating that when "Duration" = 1 (1 year), F1 exerts a significant and strongest positive effect on LogTotal.

2019-2021: The conditional effect declines significantly. The effect value drops to 0.1869 (42.6% decrease from 2016-2018) with a slightly increased standard error of 0.0141. The t-value falls to 13.2430, and the 95% confidence interval [0.1592, 0.2146] shifts downward— confirming that the positive effect of F1 on LogTotal (when "Duration" = 1) is halved.

2022-2024: The conditional effect drops to the lowest level. The effect value further decreases to 0.0607, a 67.5% reduction from 2019- 2021 and only 18.6% of the 2016-2018 level. Although the effect remains significant (t=6.3691, p<0.001) with a 95% confidence interval [0.0420, 0.0795] that excludes zero, the magnitude of the positive effect is drastically weakened.

While "Duration" consistently exerts a negative moderating effect on the F1→LogTotal path (as evidenced by the significantly negative interaction term across all periods), its moderating strength and practical contribution show a clear downward trend over time. This is most evident in the continuous decline of R² change and the drastic weakening of conditional effects—indicating that the ability of "Duration" to regulate the relationship between F1 and LogTotal becomes increasingly limited across the three periods.

4.2.4 Comparison of Covariate Effects

The significance of covariates reflects differences in the impact of control variables across periods. The relevant parameters are shown in Table 2. Consistently significant covariates (p<0.05):

F2: It is significantly positive across all three periods (2016-2018: 0.0754; 2019-2021: 0.0761; 2022-2024: 0.0799). The coefficient is stable with a slight upward trend, indicating that its positive impact on LogTotal persists and strengthens marginally over time.

F3: It is significantly positive in 2016-2018 (0.1057) and 2019-2021 (0.1381). In 2022-2024, it remains significant but its coefficient drops by 52% (to 0.0666), showing that its positive impact first strengthens and then weakens.

Renewal: It is significantly positive across all three periods (2016-2018: 0.0395; 2019-2021: 0.0475; 2022-2024: 0.0464). The coefficient is stable, demonstrating a consistent positive impact on LogTotal.

Department: It is significantly positive across all three periods (2016-2018: 0.0324; 2019-2021: 0.0784; 2022-2024: 0.0942). The coefficient doubles in magnitude each period, indicating a substantial increase in its positive impact on LogTotal.

Period-specifically significant covariates:

Reader: It is only significantly positive in 2016-2018 (0.0361, p<0.001) and becomes non-significant in 2019-2021 and 2022-2024 (p>0.05). This suggests that the impact of "Reader" on LogTotal exists only in the early period.

Gender: It is non-significant across all three periods (p>0.05) and has no significant impact on LogTotal. Additionally, its effect size is close to 0 (ranging from 0.0049 to 0.0098).

Newly Added Covariates (F4, F5, F6; only included in 2022-2024): All three are significantly positive (F4: 0.1100; F5: 0.0966; F6: 0.0934, p<0.001), making them important positive predictors of LogTotal in this period.

4.3 Title Word Clustering

In the period of 2016–2018, the high - frequency title words show certain clustering characteristics. From the relevant visualization (see Figure 5), words related to “therapy”, “counseling”, “family practice”, and “anatomy” tend to cluster together. This indicates that the research or book content in this period may focus on basic medical and psychological counseling services. For example, “therapy” and “counseling” are core concepts in psychological intervention, and their clustering suggests a relatively concentrated theme in the field of psychological and basic medical services. The dimension explanation shows that Dimension 1 (X - Axis) accounts for 30.53% and Dimension 2 (Y - Axis) accounts for 10.47% of the total association, which provides a certain spatial distribution basis for the clustering of these words.

Figure 5. The top 60 most frequent title words (2016–2018)

During 2019–2021 (see Figure 6), the high - frequency title words present a more diverse clustering pattern. Words such as “art”, “theories”, “psychological counseling”, and “trauma therapy” form clusters. This reflects that the content may expand to the combination of art and psychology, as well as in - depth exploration of psychological trauma - related fields. The dimension explanation is Dimension 1 (X - Axis): 21.17% and Dimension 2 (Y - Axis): 17.96%, which shows a different spatial distribution structure of title words compared to 2016–2018, and also indicates a change in the thematic focus of the content, with more attention to theoretical and specialized psychological fields.

Figure 6. The top 60 most frequent title words (2019–2021)

For 2022–2024 (see Figure 7), the clustering of high - frequency title words shows new features. Words like “counseling practice”, “psychological theory”, “anxiety”, and “modern” are clustered. This implies that the content may be more inclined to the practical application of psychological counseling and the exploration of modern psychological problems such as anxiety. The dimension explanation is Dimension 1 (X - Axis): 21.97% and Dimension 2 (Y - Axis): 15.45%, which provides a unique spatial framework for the clustering of these words, indicating that in this period, the theme has shifted to a more practical and modern - oriented psychological research and application field.

Figure 7. The top 60 most frequent title words (2019–2021)

In summary, for the field of psychology and its related disciplines, the high-frequency title words from 2016–2018 to 2022–2024 exhibit a distinct evolutionary trajectory: initially focusing on basic medical and psychological services, shifting to the integration of art and psychology, and ultimately advancing toward modern practical applications of psychology. Correspondingly, changes in word clustering patterns consistently reflect the continuous evolution of research and content themes within these fields. Each time period is characterized by a unique clustering mode, which is closely correlated with the dimension-explained variance that accounts for the total associations among title words. Three core trends emerge from this analysis: first, semantic convergence, evidenced by increasingly integrated clusters that mirror the interdisciplinary development of the field; second, thematic shifts, transitioning from broad thematic categorization in the early stage to a growing focus on applied and practical topics in the later stage, accompanied by the emergence of new sub-themes (e.g., sports medicine, modern therapeutic approaches); and third, dynamic changes in dimension-explained variance, which reflect the evolving semantic relationships between title words over time. Collectively, the visualized clustering results demonstrate a clear transition from broad, discrete clusters to integrated, application-oriented groupings—one that accurately mirrors the evolving research priorities and academic interests within the psychology field and its related domains.

4.4 Correlations Analysis

The model for 2920 books spanning 2016–2024 employs bootstrap specifications to assess the reliability and stability of statistical estimates. The sampling method is specified as "Simple," indicating that bootstrap samples are drawn randomly with replacement from the original dataset of 2920 books. The number of bootstrap samples is set to 1000. This is a commonly adopted threshold in bootstrap analyses, as it provides a sufficient number of replications to stabilize the estimated sampling distributions. Table 3 presents the correlations among three variables—Number of Borrowings, Collection Tenure (The length of time since being added to the collection), and Secondary Categories of R.

Table 3. Correlations (2016-2024)

**. Correlation is significant at the 0.01 level (2-tailed).

c. Bootstrap results are based on 1000 bootstrap samples

There is a statistically significant negative correlation between Number of Borrowings and Collection Tenure (Pearson correlation = - 0.069, p = 0.000). This indicates that newer books (with less time since being added to the collection) tend to be borrowed slightly more frequently, though the correlation coefficient is small, suggesting a weak relationship. The 95% confidence interval (-0.098 to -0.041) confirms the significance of this negative association. Number of Borrowings and Secondary Categories of R show a statistically significant positive correlation (Pearson correlation = 0.129, p = 0.000). This implies that certain secondary categories within the R class are associated with higher borrowing rates, though the correlation is weak. The 95% confidence interval (0.094 to 0.162) supports the reliability of this positive trend, indicating that category-specific preferences influence borrowing behavior. There is no significant correlation between Collection Tenure and Secondary Categories of R (Pearson correlation = -0.018, p = 0.341). The 95% confidence interval (-0.054 to 0.019) includes zero, confirming that the age of books in the collection does not systematically relate to their secondary R categories. While all correlations are weak in magnitude, the significant associations highlight that newer books are marginally more borrowed, and specific R secondary categories are linked to higher borrowing frequencies. The lack of correlation between collection time and categories suggests that age does not bias the distribution of R-type books across subcategories.

4.5 Survival Analysis of R Books

This survival function plot in Figure 8 clearly demonstrates the variation pattern of the "cumulative survival probability" of samples in the Secondary Categories of R with the Length of Borrowing Years through stepwise-decreasing survival curves: the overall survival probability declines over time, with the early stage (0–2 years) being a high-incidence period for failures; there are significant differences in survival resilience and the concentrated failure time across different Secondary Categories of R. Specifically, certain categories (e.g., R7, represented by the blue line) maintain a relatively high survival probability over an extended period—their survival probability remains above 0.6 at the 4-year mark and only decreases significantly around the 8-year mark. This indicates that samples in such Secondary Categories of R are more likely to "survive continuously" during the borrowing process (i.e., maintain the borrowing status for a longer duration). In contrast, for R2 (represented by the green line), the survival probability drops to below 0.2 around the 2-year mark, showing extremely rapid "failure" (e.g., detachment from the continuous borrowing status), which reflects a distinct disadvantage in the competition for borrowing duration. Additionally, the time points at which the survival probability of different categories first decreases significantly vary (e.g., approximately 4 years for R1 and 3 years for R3). This variation indicates that there are differences in the "critical failure nodes" of samples corresponding to each secondary category during the borrowing cycle, which may be associated with factors such as the inherent content attributes of the categories and the stability of audience demand.

Figure 8. R book survival function plot

The median survival times vary notably among the secondary categories of R. R7 exhibits the longest median survival time at 4.14 units, indicating that books in this category tend to remain in circulation or take the longest to experience the target event compared to other subcategories. In contrast, R6 has the shortest median survival time at 1.67 units, suggesting that books in R6 are more likely to experience the event (such as being borrowed for the first time or removed) more quickly. Among the remaining categories, the median survival times are relatively clustered: R4 (3.25) and R5 (2.98) have longer durations than the mid-range values, while R3 (2.89), R2 (2.87), R8 (2.87), R9 (2.65), and R1 (2.61) fall within a narrower range, with minimal differences between them. This clustering suggests that most R subcategories share similar survival patterns, with R7 and R6 emerging as outliers driving the overall variability.

Based on the data on mean borrowing volume, sample size (N), and standard deviation (Std. Deviation) of different reader types across three periods (2016–2018, 2019–2021, 2022–2024) in Table 4, the total mean borrowing volume across all readers shows a "downward then slight upward" trajectory: it decreases from 3.41 books in 2016–2018 to a low of 2.70 books in 2019–2021 (a decrease of approximately 20.8%), and then rebounds slightly to 2.92 books in 2022–2024 (an increase of about 8.1% compared to 2019–2021). This trend may reflect external impacts (e.g., changes in resource access methods, such as the rise of e-books, or social events affecting offline borrowing) in the middle period (2019–2021), followed by a partial recovery of borrowing demand in the later period. While the mean borrowing volume partially recovered in the later period, the overall scale of the borrowing group continued to shrink. Faculty & Staff have the highest borrowing volume: their mean borrowing volume remains above 4.6 books in all three periods (4.75 in 2016–2018, 4.67 in 2019–2021, 4.82 in 2022–2024), showing strong stability and even a slight increase in the later period. This is likely related to their academic research and teaching needs, which require sustained access to physical resources. Graduate Students rank second: their mean borrowing volume decreases from 4.17 books (2016–2018) to 3.04 books (2019–2021) and then rises slightly to 3.11 books (2022–2024). The fluctuation aligns with the total trend but remains higher than that of undergraduate students, reflecting their higher demand for specialized resources for thesis writing and research. Undergraduate Students have the lowest borrowing volume: their mean borrowing volume drops from 2.74 books (2016–2018) to 2.21 books (2019–2021) and then rebounds to 2.47 books (2022–2024). The relatively low and fluctuating borrowing volume may be due to their more diverse learning resource channels (e.g., online courses, e-textbooks) and less dependent on physical books. The "0 Other" reader group (with a mean borrowing volume of 4.00 books and N=2) only appears in 2016–2018 and is absent in the subsequent two periods. This may be due to changes in reader classification standards (e.g., merging "Other" into other categories) or a significant reduction in non-core reader groups (e.g., external visitors, alumni) accessing physical resources, leading to their exclusion from the statistical sample.

The total standard deviation of borrowing volume decreases from 5.24 (2016–2018) to 3.89 (2019–2021) and further to 3.62 (2022–2024). This indicates that the differences in borrowing volume among readers are gradually narrowing—while some readers may have reduced their borrowing frequency, the "high-frequency borrowers" and "low-frequency borrowers" groups are becoming less distinct, and the overall borrowing behavior is more concentrated. Faculty & Staff and Graduate Students consistently show high standard deviations: for example, the standard deviation of Faculty & Staff increases from 4.55 (2016–2018) to 6.39 (2022–2024), and that of Graduate Students decreases from 7.05 (2016–2018) to 3.87 (2022–2024) but remains higher than that of undergraduate students. This suggests that within these groups, there are significant differences in borrowing needs—some individuals (e.g., senior researchers, doctoral students) may still borrow large quantities of books, while others rely more on digital resources, leading to uneven borrowing volume. Undergraduate Students have relatively low standard deviations (3.15, 2.69, 2.64 in the three periods), indicating more homogeneous borrowing behavior among this group, with no extreme high or low borrowing volume.

Table 4. Changes in borrowing volume metrics by reader type (2016–2024)