Dynamic identity management in online health communities: How illness stage moderates the effects of anonymity — A longitudinal panel data analysis

Abstract

Objective

This study examines how anonymity influences self-disclosure and social feedback across illness stages, and whether these effects vary with disease progression. To theorize these dynamics, we propose a Dynamic Identity Management (DIM) framework to visualize how anonymity shapes self-disclosure and social feedback, and whether these effects vary with disease progression.

Methods

We collected posts from the “Cancer” super-topic on Zhihu between January 2018 and January 2024 using Python web crawlers. Through semantic embedding (Sentence-BERT) and clustering (similarity threshold=0.8), we linked anonymous posts to their real-name authors, constructing a six-year longitudinal panel dataset of 1,998 posts from 215 users. User and year fixed-effects models were employed to estimate the causal effects of anonymity on disclosure motivation, emotional expression, information breadth and sensitivity, and social feedback (likes and comments), with illness stage (initial diagnosis, mid-stage, terminal) as a moderator.

Results

Anonymity significantly increased intrinsic disclosure motivation (β=−0.318, p<.01) and negative emotional expression (β=−0.196, p<.05), but did not affect social feedback or factual information disclosure. Both initial and terminal stages directly increased social feedback (likes and comments, p<.001) and negative emotional expression (p<.001) compared to the mid-stage. Illness stage significantly moderated anonymity’s effects: the negative effect of anonymity on comments was stronger during initial (β=−0.334, p<.05) and terminal stages (β=−0.313, p<.05); its facilitative effect on negative emotional expression was amplified in both critical stages (initial: β=−0.258, p<.05; terminal: β=−0.265, p<.05); and its positive effect on intrinsic motivation was attenuated during these high-stakes phases (initial: β=0.297, p<.05; terminal: β=0.357, p<.05).

Conclusion

Anonymity selectively facilitates intrinsic motivation and negative emotional expression but does not increase social feedback or factual information disclosure. Users dynamically manage identity across the illness stage: in high-risk stages (initial diagnosis and terminal), they prioritize emotional release through anonymity, even though this reduces social feedback relative to real-name posting; in the low-risk mid-stage, they use anonymity for intrinsically motivated disclosure without losing social feedback. This pattern supports the DIM framework’s proposition: anonymity is not a fixed personal attribute; its effects depend on the illness stage. By specifying this dynamic mechanism, the DIM framework advances understanding of anonymity in OHCs and offers actionable insights for designing stage-sensitive health support systems.

Introduction

China faces a high cancer burden. Incidence and mortality rates remain persistently high. ¹ According to global cancer statistics, China accounts for nearly a quarter of new cancer cases and cancer deaths worldwide each year. This heavy disease burden creates urgent needs: patients and their families require information, emotional support, and peer connection.

The growth of internet use has expanded online health communities (OHCs). These communities including patient forums, social media groups, and Q&A sites—provide spaces for individuals affected by cancer to interact. By removing geographical and time barriers, OHCs help address gaps in cancer care that arise from limited clinical consultation time and geographic dispersion of specialized services. ²

Two types of behavior shape user participation in OHCs: self-disclosure and social feedback. Self-disclosure refers to sharing illness experiences, treatment processes, and emotion. ³ Patients may describe their diagnosis, report side effects of chemotherapy, or express fears about prognosis. Social feedback refers to receiving responses from other users, ⁴ such as likes, comments, or direct replies. In this study, self-disclosure refers to the sharing of illness experiences, treatment processes, and emotions, while social feedback refers to likes and comments received from other users.These responses can provide informational support (e.g., advice on managing symptoms) or emotional support (e.g., expressions of empathy). Together, these interactions form the basis of peer support in OHCs, helping to reduce information asymmetry, alleviate psychological distress, and improve well-being. ⁵

Within OHCs, cancer patients and caregivers face a fundamental tension. On one hand, they need social connection. The uncertainty and emotional impact of cancer create a strong desire to share experiences and seek support from others who understand. Fear of death, anxiety about treatment outcomes, and feelings of isolation are common, and research shows that social support can improve coping and quality of life for cancer patients. On the other hand, they fear exposure. Cancer is a stigmatized condition in many societies, often associated with blame, shame, or assumptions about mortality. Disclosing personal medical information may invite unwanted attention, social judgment, or even discrimination. Using real names also increases privacy risks, as posts can be traced back to individuals and may affect employment, insurance, or personal relationships. ⁶ Each act of communication therefore involves weighing benefits against risks ⁷ : benefits include receiving support, while risks include loss of privacy.

Many platforms offer anonymity as a way to manage this tension. Users can choose to post without revealing their real identity, which may reduce the perceived risks of disclosure. ⁸ Zhihu provides a useful setting for studying anonymity in health contexts. Zhihu is a Chinese Q&A platform with approximately 100 million monthly active users. It allows users to post under their real name or anonymously within the same account. When posting anonymously, the user’s profile information is hidden, and the post is labeled as from “Anonymous User.” Users can switch between these two states across different posts. ⁹ This design makes it possible to observe how the same individual uses anonymity at some times and real names at others, and this within-person variation helps identify causal effects.^10,11

Existing research on anonymity in health contexts has a key limitation: it tends to treat identity choice as static. Some studies compare anonymous users with real-name users, assuming individuals consistently prefer one mode, while others examine anonymity in experimental settings. While valuable, these approaches cannot explain why the same user switches between identities over time. ¹² This static perspective views identity as a relatively stable background variable rather than a resource users can actively manage.^13–15

Yet the cancer journey is not uniform at diagnosis, patients urgently need information and support; during treatment, they focus on documenting and sharing experiences; at the terminal stage, concerns shift to life meaning and emotional expression. Different stages bring different needs and alter how patients weigh privacy against social connection. This calls for a framework that explicitly incorporates time and context.

To address this gap, we propose a Dynamic Identity Management (DIM) framework, which conceptualizes anonymity as a strategic resource that users deploy flexibly as their situation changes. The framework integrates four theories: the Social Identity Model of Deindividuation Effects (SIDE) explains how anonymity shifts focus to group identity, reshaping the interaction environment; Privacy Calculus Theory captures how users weigh benefits against risks when deciding what to disclose; Social Penetration Theory describes how relational needs drive disclosure from superficial to intimate levels; and Situational Strength Theory serves as an overarching lens, explaining how these mechanisms are amplified or attenuated across different illness stages. In short, SIDE explains environmental shifts, Privacy Calculus explains decision-making, Social Penetration explains relationship deepening, and Situational Strength explains stage-specific variations—together, they form a dynamic, contextualized framework for understanding identity management in online health communities.

Using data from Zhihu’s “Cancer” super-topic, this study asks three questions:

RQ1:How does anonymity affect users’ social feedback (likes and comments) and self-disclosure behaviors (motivation, emotion, and information disclosure)?

RQ2:Does illness stage (initial diagnosis, mid-stage, terminal stage) moderate these effects? Do the effects of anonymity differ depending on where a patient is in their cancer journey?

RQ3:What do these patterns reveal about how users manage identity across the illness stage? Do they reveal systematic strategies—for example, using real names when seeking information and anonymity when expressing emotions—and do these strategies change as needs shift?

By answering these questions, this study aims to move beyond static conceptions of anonymity and provide a more nuanced understanding of how cancer patients and caregivers navigate the tension between connection and exposure in online communities.Understanding these dynamics can inform the design of OHC platforms that better support users at different illness stages. The findings may also inform the design of OHC platforms, suggesting ways to support users at different stages of illness with appropriate identity options. ¹⁶

Literature review and research framework

Research on self-disclosure and social feedback in online health communities has yielded valuable insights from multiple perspectives, including technological affordances, individual motivations, and social norms. However, a significant theoretical blind spot persists: the dynamic management of user identity. Prevailing scholarship predominantly adheres to a static paradigm, framing the choice between anonymity and real-name participation as a fixed user attribute or a stable technical preference rather than as a strategic and adaptive behavior. This static perspective overlooks a central reality of health crises: for patients, identity management is a dynamic process of psychosocial adaptation that continually evolves with illness progression, shifting psychological needs, and changing social roles. It cannot explain why an individual switches identity states over time or reveal the strategic reasoning behind such shifts.

To address this theoretical gap, we propose a Dynamic Identity Management (DIM) framework, which conceptualizes anonymity as a strategic resource that users deploy flexibly as their situation changes. The framework integrates three complementary theoretical perspectives(The Social Identity Model of Deindividuation Effects (SIDE),Privacy Calculus Theory,Social Penetration Theory)—each explaining a distinct mechanism through which anonymity influences user behavior—and employs Situational Strength Theory as an overarching meta-framework to specify how these mechanisms are amplified or attenuated across different illness stages.

Deindividuation theory was first proposed by Festinger et al., ¹⁷ suggesting that anonymity in groups reduces self-awareness and makes individuals more susceptible to group norms. Zimbardo ¹⁸ later reinforced this through experimental studies, establishing anonymity as a key mechanism in deindividuation. A major theoretical advance came with the Social Identity Model of Deindividuation Effects (SIDE), developed by Spears and Lea. ¹⁹ The SIDE model posits that anonymity weakens individual identity cues, causing behaviour to align more closely with perceived group prototypes and norms. In health communities, this implies that the salience of a “patient” social identity can be heightened in anonymous settings, making community norms such as “encouraging authenticity” and “mutual support” more influential. ²⁰

Researchers have further extended the SIDE model to digital contexts, showing that anonymity in computer-mediated communication can enhance individuals’ tendency to categorise themselves at the group level, making social norms more salient and triggering conformity. ²¹ In the DIM framework, SIDE explains how anonymity reshapes the social environment it influences the likelihood of receiving social feedback (likes and comments) and the authenticity of emotional expression by altering the perceived audience and normative expectations. ²²

Although existing literature often broadly concludes that “anonymity leads to deindividuation,” we argue for a necessary distinction: its effects differ across two behavioural pathways, social feedback and self-disclosure. Research indicates that anonymity can weaken social bonds and reduce specific interpersonal motivations for others to provide feedback.^23,24 Consequently, anonymity may suppress social feedback, such as likes and comments. However, anonymity also reduces social constraints, which can facilitate deeper and more authentic self-disclosure. ²⁵ This contradictory effect reveals the complexity of anonymity’s role. Based on the SIDE model, we propose:

- H1a₁:Anonymous posts receive fewer likes than real-name posts.

- H1a₂:Anonymous posts receive fewer comments than real-name posts.

For emotional expression, anonymity works primarily by reducing impression-management concerns and reinforcing group norms. In real-name settings, users often feel pressured by social desirability to present a positive self-image. When anonymous, freed from such pressures, their emotional expression is more likely to align with community norms that accept negative emotions. Therefore, we propose:

- H1b₂:Anonymous posts contain more negative emotional content than real-name posts.

Privacy Calculus Theory originated in the seminal work of Laufer and Wolfe ²⁶ on privacy decision-making and was formally articulated by Culnan and Armstrong. ²⁷ The theory conceptualises information disclosure as a rational decision-making process based on a risk-benefit trade-off. Individuals are more likely to disclose when perceived benefits, such as social support, emotional catharsis, or self-clarification outweigh perceived risks, such as privacy leakage, social discrimination, or identity theft.

Within this framework, anonymity directly influences disclosure behaviour by reducing potential social, professional, and psychological risks, thereby altering the user’s decision calculus.^28–31 Privacy Calculus Theory provides the micro-level decision-making foundation for our hypotheses on disclosure motivation, especially intrinsic motivation. We refine this view by proposing that the effect of anonymity on disclosure exhibits “dimensional heterogeneity.” Specifically, anonymity exerts a “purifying” effect on motivation: when external risks are substantially lowered in an anonymous setting, the underlying drivers of disclosure are more likely to reflect genuine psychological needs rather than externally influenced strategic choices. Accordingly, we propose:

- H1b₁: Anonymous posts are more likely to be intrinsically motivated than real-name posts.

Building on this, we further posit an “emancipatory” effect of anonymity on information content: low-risk environments empower users to address topics that might be considered taboo in real-name settings, thereby expanding the breadth and depth of information disclosure. ³² Accordingly, we propose:

- H1b₄: Anonymous posts disclose a greater breadth of personal information than real-name posts.

- H1b₅: Anonymous posts disclose more sensitive personal information than real-name posts.

However, regarding rational expression, its main effect lies in shaping the assessment of social risks and privacy costs. Thus, anonymity primarily influences the willingness to disclose emotional and private information. It does not significantly affect rational expression. ³³ Rational expression relies heavily on cognitive resources (e.g., disease analysis, treatment data sharing, and logical reasoning). Its extent is governed more by the user’s knowledge structure, cognitive habits, and available cognitive resources. ³⁴ Based on this, we propose:

- H1b₃: The level of rational expression in anonymous posts does not differ significantly from that in real-name posts.

Social Penetration Theory, developed by Altman and Taylor, ³⁵ explains how interpersonal relationships deepen from superficial to intimate through bidirectional self-disclosure. The theory posits that information exchange typically follows an outward-to-inward stage, beginning with broad, low-sensitivity peripheral topics and gradually progressing toward more unique and private core content.

In the context of a health crisis, this penetration process is driven by intense, stage-specific relational needs. Illness stage serves as a key situational variable reflecting the intensity of these needs. During initial diagnosis and terminal stages, individuals grapple with profound uncertainty, death anxiety, and existential concerns, causing their need for deep, immediate social support to peak. ³⁶ This functions as a “relationship accelerator”, powerfully propelling the social penetration process. Related research further demonstrates that self-disclosure behaviours in OHCs exhibit dynamic fluctuations, often increasing during major health changes but receding during stable periods.^37–39

The social penetration process primarily drives emotionally oriented disclosure. In contrast, rational expression depends more directly on individual resources inherent to the user, including knowledge structure and cognitive habits. ³⁴ This fundamental characteristic makes rational expression less susceptible to direct influence from external situational variables.

Across different disease stages, a significant conflict arises between the patient’s emotional state and available cognitive resources. During initial diagnosis and terminal stages, the demand for emotional disclosure increases markedly. However, intense emotional fluctuations, such as shock, panic, or grief can lead to cognitive overload, which systematically impedes deep, rational expression. ⁴⁰ In contrast, during stable phases of illness, the emotional burden lessens, and patients typically possess more sufficient cognitive capacity to construct rational content. Therefore, the influence of disease stage on rational expression operates through the mediating variable of “emotion,” which interferes with cognitive resources. Based on this, we propose the following hypotheses:

- H2a₁: Posts published during the initial and terminal illness stages receive more likes than those published during the mid-stage.

- H2a₂: Posts published during the initial and terminal illness stages receive more comments than those published during the mid-stage.

- H2b₁: Posts published during the terminal illness stage are more likely to be intrinsically motivated than those published during the mid-stage.

- H2b₂: Posts published during the initial and terminal illness stages express more negative emotion than those published during the mid-stage.

- H2b₃: Illness stage has no significant effect on the level of rational expression in posts.

- H2b₄: Posts published during the initial and terminal illness stages disclose a greater breadth of personal information than those published during the mid-stage.

- H2b₅: Posts published during the initial and terminal illness stages disclose more sensitive personal information than those published during the mid-stage.

While SIDE, Privacy Calculus, and Social Penetration each illuminate a distinct pathway through which anonymity affects user behaviour, they do not, by themselves, explain why these effects vary systematically across the illness stage. To address this gap, we introduce Situational Strength Theory ⁴¹ as an overarching meta-framework that specifies the boundary conditions under which the mechanisms proposed by the three core theories are amplified or attenuated.

Situational Strength Theory, originally developed by Meyer and colleagues (2010), posits that situations vary in their “strength” along four dimensions: clarity (the degree to which cues about expected behaviour are unambiguous), consistency (the extent to which cues are compatible with each other), constraints (the degree to which external forces limit freedom of choice), and consequences (the significance of behavioural outcomes for the individual). In strong situations, behaviour is more heavily influenced by external cues and norms, and individual differences play a smaller role; in weak situations, personal dispositions and motivations have greater latitude to shape behaviour.

In the context of OHCs, we conceptualise illness stage as a natural manifestation of situational strength. The initial diagnosis and terminal stages constitute strong situations: they are characterised by high uncertainty, intense emotional arousal, urgent needs for information and support, and severe potential consequences of disclosure (e.g., stigma, discrimination, or unwanted social attention). These features heighten the clarity, consistency, and consequentiality of the situation, thereby magnifying the effects of anonymity on certain behaviours while suppressing others. In contrast, the mid-stage (treatment, remission, deterioration) represents a weaker situation: the illness stage is more predictable, emotional intensity subsides, and the immediate stakes of disclosure diminish, allowing users’ intrinsic motivations to emerge more freely.

Situational Strength Theory integrates the three core theories and generates specific moderation hypotheses:

Reconfiguring SIDE for Social Feedback. In strong situations, the community’s need for credible information and efficient support intensifies. This amplifies the credibility discount applied to anonymous posters. It also exacerbates the cognitive resource competition among members, who then prioritise engaging with lower-risk, real-name sources. Thus, the negative effect of anonymity on social feedback is strengthened. We therefore propose:

- H3a₁: The negative effect of anonymity on like count will be stronger during the initial and terminal illness stages.

- H3a₂: The negative effect of anonymity on comment count will be stronger during the initial and terminal illness stages.

Reconfiguring Privacy Calculus for Self-Disclosure. The high stakes of strong situations sharply increase the psychological benefits of emotional disclosure. Anonymity’s “safe isolation” thus synergises with this heightened need, strengthening its role as a channel for negative emotion. However, these same high stakes may elevate extrinsic, instrumental needs above intrinsic ones, thereby weakening anonymity’s typical “purifying” effect on motivation. We therefore propose:

- H3b₁: The positive effect of anonymity on intrinsically motivated disclosure will be weakened during the initial and terminal illness stages.

- H3b₂: The negative effect of anonymity on the expression of positive emotion (i.e., the increase in negative emotional expression) will be strengthened during the initial and terminal illness stages.

Stable Boundaries of the Anonymity Effect. Research has identified a stable boundary to the anonymity effect. A study during the 2022 mpox outbreak found no significant difference in the reporting of sensitive behavioural information between patients completing anonymous versus non-anonymous questionnaires. ⁴² This suggests that in high-risk contexts, the drivers of disclosure decisions become more complex, and the presumed “protective shield” of anonymity does not necessarily translate into increased disclosure behaviour. Consequently, we argue that the null effect of anonymity on rational expression is robust and remains unaffected by the strength of the situational context. We therefore predict that the presumed “liberating” effect of anonymity on information disclosure will not be activated or strengthened, even in high-intensity situations. This leads to the following hypotheses:

- H3b₃: Illness stage does not moderate the effect of anonymity on rational expression.

- H3b₄: The positive effect of anonymity on the breadth of information disclosure will not be significantly stronger in the initial and terminal disease stages.

- H3b₅: The positive effect of anonymity on the sensitivity of information disclosed will not be significantly stronger in the initial and terminal disease stages.

Figure 1 presents the Dynamic Identity Management (DIM) theoretical framework. The framework posits that anonymity influences two broad categories of user behaviour—social feedback (likes and comments) and self-disclosure (motivation, emotion, rational expression, information breadth, and information sensitivity)—through three distinct mechanisms:

- SIDE:anonymity shapes the social environment, affecting how users perceive community norms and how others respond to their posts.

- Privacy Calculus: anonymity alters the risk-benefit calculus, influencing users’ disclosure decisions.

- Social Penetration: illness stage, as a driver of relational needs, directly affects the depth and breadth of disclosure.

Figure 1.

Dynamic identity management (DIM) theoretical framework.

These mechanisms are not static; they are systematically moderated by illness stage, which functions as a situational strength variable. In strong situations (initial diagnosis and terminal stages), the effects of anonymity on social feedback and emotional expression are amplified, while its effect on intrinsic motivation is attenuated. In the weaker situation (mid-stage), anonymity serves a purer function of enabling intrinsic expression without social cost. Table 1.

Table 1.

Summarizes all research hypotheses.

H1: Main effects of anonymity
H1a: Anonymous posts receive fewer likes than real-name posts	H1a1: Anonymous posts receive fewer likes than real-name posts.
	H1a2: Anonymous posts receive fewer comments than real-name posts.
H1b: Effects of Anonymity on Self-Disclosure	H1b1: Anonymous posts are more likely to be intrinsically motivated than real-name posts.
	H1b2: Anonymous posts contain more negative emotional content than real-name posts.
	H1b3: The level of rational expression does not differ significantly between anonymous and real-name posts.
	H1b4: Anonymous posts disclose a greater breadth of personal information than real-name posts.
	H1b5: Anonymous posts disclose more sensitive personal information than real-name posts.
H2: Main Effects of Illness Stage
H2a: Effects of Illness Stage on Social Feedback	H2a1: Posts published during initial and terminal stages receive more likes than those published during the mid-stage.
	H2a2: Posts published during initial and terminal stages receive more comments than those published during the mid-stage.
H2b: Effects of Illness Stage on Self-Disclosure	H2b1: Posts published during the terminal stage are more likely to be intrinsically motivated than those published during the mid-stage.
	H2b2: Posts published during initial and terminal stages express more negative emotion than those published during the mid-stage.
	H2b3: Illness stage has no significant effect on the level of rational expression in posts.
	H2b4: Posts published during initial and terminal stages disclose a greater breadth of personal information than those published during the mid-stage.
	H2b5: Posts published during initial and terminal stages disclose more sensitive personal information than those published during the mid-stage.
H3: Moderating Effects of Illness Stage
H3a: Moderating Effects on Social Feedback	H3a1: Illness stage moderates the effect of anonymity on like count. The negative effect of anonymity will be stronger during initial and terminal stages.
	H3a2: Illness stage moderates the effect of anonymity on comment count. The negative effect of anonymity will be stronger during initial and terminal stages.
H3b: Moderating Effects on Self-Disclosure	H3b1: The positive effect of anonymity on intrinsically motivated disclosure will be weakened during the initial and terminal illness stages.
	H3b2: The negative effect of anonymity on the expression of positive emotion will be strengthened during the initial and terminal illness stages.
	H3b3: Illness stage does not moderate the effect of anonymity on rational expression.
	H3b4: The positive effect of anonymity on the breadth of information disclosure will not be significantly stronger in the initial and terminal disease stages.
	H3b5: The positive effect of anonymity on the sensitivity of information disclosed will not be significantly stronger in the initial and terminal disease stages.

Figure 1 maps this structure: anonymity and illness stage enter as independent variables, the three mechanisms serve as explanatory pathways, and the two categories of user behaviour are the outcomes. Illness stage functions both as a direct predictor and as a moderator of the anonymity–behaviour relationship.

This integrated perspective leads to three groups of hypotheses. First, SIDE informs the main effects of anonymity on social feedback (H1a) and emotional expression (H1b₂): by weakening interpersonal cues and heightening group norms, anonymity may reduce feedback from others while facilitating the release of negative emotions. Second, Privacy Calculus Theory underpins the main effects of anonymity on self-disclosure (H1b₁, H1b₄, H1b₅): by lowering perceived social risk, anonymity shifts the risk–benefit calculus toward intrinsic motivation and broader, more sensitive disclosure. It also predicts the null effect on rational expression (H1b₃), which depends on cognitive resources rather than risk assessment. Third, Social Penetration Theory drives the main effects of illness stage (H2): the intensity of relational needs varies across stages, affecting both social feedback received and self-disclosure behaviors. Finally, Situational Strength Theory generates all moderation hypotheses (H3): the high-stakes conditions of initial diagnosis and terminal stages amplify the effects of anonymity on social feedback and emotional expression, while attenuating its effect on intrinsic motivation.

Method

Data source

This was a longitudinal panel data analysis conducted from the “Cancer” super-topic on Zhihu platform, a leading Chinese Q&A community with approximately 100 million monthly active users. The “Cancer” super-topic is a dedicated online community where cancer patients, survivors, and caregivers share experiences, seek advice, and provide mutual support. Zhihu’s key feature—allowing users to post under their real name or anonymously within the same account, enables within-person comparisons over time, making it possible to observe how the same individual uses anonymity differently across various illness stages.Figure 2 illustrates the overall research design and workflow.

Figure 2.

Research design and data analysis flowchart.

We collected posts published between January 1,2018, and January 1,2024, using Python web crawlers. The initial dataset contained 26,895 posts, including 3,651 anonymous posts and 23,244 real-name posts. After removing posts with undeterminable illness stages, we retained all anonymous posts for matching.

To link anonymous posts to their real-name authors, we used a semantic embedding approach. We generated 384-dimensional embeddings for each anonymous post using a pretrained Sentence-BERT model (MiniLM-L6-v2), then constructed a cosine similarity matrix and performed clustering with a threshold of 0.8. This resulted in 369 anonymous user clusters, each representing posts presumably by the same anonymous user.

For real-name users, we retained those with at least three posts, yielding 2,134 users. We aggregated all posts from each anonymous cluster into one text file and all posts from each real-name user into another text file, then generated embeddings for these aggregated files. We computed cosine similarity between each anonymous cluster and each real-name user, applying a one-to-one matching rule requiring similarity≥0.8. This process successfully matched 215 users.

To validate the matching threshold, we constructed a validation set using only real-name posts, including 373 users and 3,057 posts. For each post in this validation set, we treated it as if it were anonymous, generated its semantic embedding using the same Sentence-BERT model, and applied the identical clustering and matching procedure to link it back to a real-name user. This allowed us to calculate precision and false match rate (FMR) at different thresholds. As shown in Figure 3, the threshold of 0.8 achieved an optimal balance between matching accuracy and the risk of erroneous links, with precision reaching 80.87% and FMR at 3.55%. This level of precision is considered robust for the purposes of this study, as it ensures that the vast majority of matched pairs are correct while maintaining a sufficient sample size for longitudinal analysis.

Figure 3.

Similarity threshold validation for anonymous user identity matching.

The final sample consists of 215 users and 1,998 posts. Each user contributed at least one anonymous and one real-name post, with the number of posts per user ranging from 2 to 42 (median=7). The posts span 14 years from 2011 to 2024. Users switched between anonymous and real-name identity 916 times in total, averaging 4.3 switches per user. Among the 215 users, 158 (73.5%) have posts in all three illness stages. Table 2 summarizes the panel structure of the sample.

Table 2.

Panel structure characteristics of the sample.

Panel characteristic	Value
Number of unique users	215
Total observations (posts)	1,998
Average observations per user	9.3
Median observations per user	7
Range of observations per user	2 – 42
Time span	2011 – 2024
Number of years	14
Users with both anonymous and real-name posts	215 (100%)
Total identity switches (anonymous ↔ real-name)	916
Average switches per user	4.3
Users with posts in all three illness stages	158 (73.5%)

Note. Data computed from the original dataset. A “switch” is defined as a change in identity status between consecutive posts by the same user.

Variables

We measured two categories of user behavior: social feedback and self-disclosure. Social feedback reflects the social support a user receives from the community, operationalized as the number of likes and comments a post receives. Likes represent passive endorsement or emotional agreement, while comments represent active responses such as advice, empathy, or questions. Both variables were log-transformed (natural logarithm of count plus one) to reduce skewness. Self-disclosure reflects the user’s own expression of needs and experiences, measured across multiple dimensions: motivation, emotion, rational expression, information breadth, and information sensitivity.

Disclosure motivation captures the underlying purpose of posting. Using keyword matching and regular expressions, we classified posts into four categories—recording, emotional venting, social sharing, and seeking help—then collapsed these into a binary variable: 0 for intrinsic motivation (recording and emotional venting) and 1 for extrinsic motivation (social sharing and seeking help). Intrinsic motivation reflects internal needs such as documenting one’s journey or releasing emotions, while extrinsic motivation reflects external goals such as seeking information or connecting with others.

Emotional expression captures the affective tone of each post. Based on the SIDE model’s prediction that anonymity reduces impression management pressures, we measured emotional valence using a BiLSTM sentiment analysis model trained on 3,000 manually labeled posts (Cohen’s κ=0.85). Figure 4 presents the model’s performance, which achieved an accuracy of 87.5% and an AUC of 93.3% on the test set. Each post was classified as negative (0) or positive (1). Negative emotions include fear, pain, despair, and grief; positive emotions include hope, gratitude, and relief.

Figure 4.

BiLSTM sentiment analysis model AUC, accuracy, and loss.

Rational expression reflects fact-based, logical discourse in contrast to emotional or subjective expression. Using LDA topic modeling, we identified five topics that reflect a gradient from highly objective to deeply subjective content. Table 3 presents the representative keywords for each topic. Two independent coders rated each topic on a rational-emotional continuum from–2(most rational) to +2(most emotional) (ICC=0.78), and each post was assigned the score of its primary topic.

Table 3.

LDA topic categories and representative keywords.

Health factors	Family relations	Disease& suffering	Survival & value	Treatment & research
Lifestyle	Mother	Selling Property	Worth It	Immunotherapy
Heredity	Father	Fate/Destiny	Travel	Clinical Trials
Food	Grandmother (Maternal)	Do One’s Best, Accept the Outcome	Cherish	Organ Failure
Smoking	Grandmother (Paternal)	Miracle	Accompany	Misdiagnosis
Family History	Girlfriend	Painkillers	Do One’s Best	Prognosis
Staying Up Late	Workplace	Recuperation/Tonic	Time	Mutation
Environmental Pollution	School	Try	Children	Guidelines
Society	Colleagues	Hard to Endure	Live On	Immunity
Carcinogenic	Public	Pain Relief Medication	Liberation	Medication
Harmful Habits	Grandmother (Maternal)-another	Ascites	Euthanasia	Genetic Mutation

Information breadth captures how many types of personal information a user discloses. Using a hybrid rule-based and NLP framework, we extracted information across six dimensions: location, occupation, gender, marital status, education, and hobbies. Information breadth is the count of dimensions disclosed in a post (range0–6), with higher scores indicating that the user reveals more facets of their personal life.

Information sensitivity reflects how sensitive the disclosed information is. We calculated inverse frequency weights for each dimension based on how rarely they were disclosed in the community, assuming that less frequently disclosed information implies higher sensitivity. The resulting weights are: occupation(41.01%), education(26.48%), hobbies(13.92%), marital status(12.54%), location(3.41%), and gender(2.64%). A post’s sensitivity score is the sum of weights of the dimensions it discloses, with higher scores indicating more sensitive disclosure. Occupation and education are considered highly sensitive because they are closely tied to social identity and may carry long-term privacy risks, while location and gender are considered less sensitive because they are more commonly shared.

Independent variables. The key independent variable is anonymity, a binary indicator (1=anonymous post, 0=real-name post). Illness stage is represented by two dummy variables: stage_initial (1=initial diagnosis stage, 0 otherwise) and stage_terminal (1=terminal stage,0otherwise), with mid-stage (treatment/remission, deterioration) as the baseline. Stage classification used a medically grounded dictionary with keyword matching and priority logic (death>deterioration>treatment>diagnosis>remission>survivor). Manual validation of a random 10% sample (≈200 posts) yielded over 90% consistency.

We also constructed interaction terms between anonymity and illness stage: anonymous×stage_initial and anonymous×stage_terminal. Year fixed effects were included to control for common time trends, with dummy variables for each year from 2019 to 2024 (2018 and before as the reference). Table 4 presents descriptive statistics for all key variables. The sample includes 1,998 posts from 215 users. Among the independent variables, 57% of posts are anonymous. The three illness stages are evenly distributed: initial 33%, terminal 33%, mid-stage 34%. Table 5 summarizes all variable definitions and measurement methods.

Table 4.

Descriptive statistics of key dependent and independent variables.

Variable	N	Mean	SD	Min	Max	Description
Dependent Variables
Likes (log)	1,263	2.07	1.91	0	7.97	Natural log of (likes + 1)
Comments (log)	1,072	1.90	1.62	0	6.70	Natural log of (comments + 1)
Rationality	1,998	-0.27	1.42	-2	2	Score from LDA (-2=rational, +2=emotional)
Information Breadth	1,998	1.26	1.03	0	6	Number of information dimensions disclosed
Information Sensitivity	1,998	0.08	0.13	0	1.00	Weighted sensitivity score
Emotion (1=positive)	1,998	0.55	0.50	0	1	1 = positive, 0 = negative
Motivation (1=extrinsic)	1,998	0.22	0.41	0	1	1 = extrinsic, 0 = intrinsic
Independent Variables
Anonymous (1=yes)	1,998	0.57	0.50	0	1	1 = anonymous post
Stage_Initial	1,998	0.33	0.47	0	1	1 = initial diagnosis stage
Stage_Terminal	1,998	0.33	0.47	0	1	1 = terminal stage
Stage_Mid (reference)	1,998	0.34	0.47	0	1	1 = mid-stage

Note. For binary variables, the mean represents the proportion of observations coded as 1. Log transformations were applied to Likes and Comments to correct for skewness. To eliminate the influence of extreme values, we removed the top and bottom 1% of likes and comments. The final effective sample sizes are 1,263 for likes and 1,072 for comments.

Table 5.

Variable definitions, measurement methods, and data sources.

Variable category	Variable name	Measurement method	Data source/Model
Dependent Variables	Likes	Natural logarithm of (number of likes+1)	Platform Scraping
	Comments	Natural logarithm of (number of comments+1)	Platform Scraping
	Rationality	Score based on LDA topic model, mapping topics to rationality-emotional continuum (-2 to 2)	LDA Topic Model
	Information Breadth	Count of disclosed categories (0-6) across six dimensions	NLP Rule Extraction
	Information Sensitivity	Sum of inverse frequency weights assigned to six types of disclosed information	Weighted Calculation
	Emotion	0=Negative,1=Positive	BiLSTM Sentiment Analysis
	Disclosure Motivation	0=Intrinsic,1=Extrinsic	Keyword Matching
Independent Variables	Anonymous	0=Real Name,1=Anonymous	Platform Identifier
	Stage_Initial	1=Diagnosis Stage,0=Other	Keyword Matching
	Stage_Terminal	1=Death or Survival Stage,0=Other	Keyword Matching
	Anonymous × Initial	Interaction term:Anonymity×Initial Stage	Calculated
	Anonymous × Terminal	Interaction term:Anonymity×Terminal Stage	Calculated

Note. The mid-stage of illness (Treatment & Remission, Deterioration) serves as the baseline group.

Statistical models

We employ two-way fixed effects models to estimate the causal effects of anonymity while controlling for time-invariant individual heterogeneity. For continuous dependent variables (likes, comments, rationality, information breadth, information sensitivity), we employed two-way fixed effects models including both user fixed effects (α_i) and year fixed effects (δ _t ) to control for time-invariant individual heterogeneity and common time trends.

The analysis uses a panel dataset containing 1,998 text posts from 215 users at different cancer stages, comprising 1,138 anonymous and 860 real-name posts. The distribution across initial, mid, and terminal stages is approximately balanced (approaching a 1:1:1 ratio), indicating a well-structured dataset for analysis.

The measured indicators cover two primary aspects of user behavior. First, user feedback behaviors include liking and commenting. Second, user self-disclosure behaviors are assessed across three dimensions: content (emotional expression and rational expression), depth (breadth of personal information disclosure and information sensitivity), and motivation (purpose of self-disclosure). Specifically, we classified recording and emotional venting as intrinsic disclosure motivations, while social sharing and seeking advice/help were categorized as extrinsic motivations. For illness stage classification, the diagnosis stage corresponds to the initial stage; treatment and remission and deterioration stages constitute the mid-stage; and death and survivor stages form the terminal stage.

Model 1 (Panel Fixed Effects Model, for continuous variables):

$Y_{it}={\alpha }_i+{\delta }_i+{\beta }_1·{Anonymous\hspace{0.17em}Effect}_{it}+{\beta }_2·{Stage\hspace{0.17em}Initial}_{it}+{\beta }_3·{Stage\hspace{0.17em}Terminal}_{it}+{\beta }_4·\left({Anonymous\_Effect}_{it}·{Stage\hspace{0.17em}Initial}_{it}\right)+{\beta }_5·\left({Anonymous\_Effect}_{it}·{Stage\hspace{0.17em}Terminal}_{it}\right)+{\epsilon }_{it}$

Model 2 (Conditional Logit Model, for binary variables):

$Logit\left[P\left(Y_{\_it}=1\right)\right]={\alpha }_{-i}+{\delta }_i+{\beta }_1·{Anonymous}_{\_}{Effect}_{-it}+{\beta }_2·{Stage}_{-}{Initial}_{-it}+{\beta }_3·{Stage}_{\_}{Terminal}_{-it}+{\beta }_4·\left({Anonymous}_{-}{Effect}_{-it}·{Stage}_{\_}{Inital}_{-it}\right)+{\beta }_5·\left({Anonymous}_{-}{Effect}_{-it}·{Stage}_{-}{Terminal}_{-it}\right)+{\epsilon }_{-it}$

We employ fixed effects panel models for empirical analysis. Individual fixed effects (α_i) are incorporated to control for time-invariant user-level heterogeneity. Parameter estimation uses within-transformation for continuous dependent variables and conditional logit regression for binary outcomes, with standard errors clustered at the individual level to account for within-user correlation.

The dependent variables encompass two primary categories of behavioral indicators. Feedback behaviors are measured through the number of likes and comments. Self-disclosure behaviors are operationalized through multiple dimensions: the motivational dimension (captured by the binary variable disclosure motive distinguishing between intrinsic and extrinsic motivation) and the content dimension (comprising emotional expression, rational expression, disclosure breadth, and disclosure sensitivity).

The core independent variables include post anonymity (Anonymous_{i t} ), a binary indicator equal to 1 when user i posts anonymously at time t and 0 for real-name posts, and illness development stage. Disease progression is captured through two dummy variables with the mid-stage (encompassing treatment, remission, and deterioration) serving as the baseline category: StageInitial_i for the initial diagnosis stage and StageTerminal_i for the terminal stage.

In terms of coefficient interpretation, β ₁ represents the average main effect of anonymity across all illness stages, indicating the average change in the dependent variable when posting anonymously compared with real-name posting. The coefficients β ₂ and β ₃ reflect the structural differences of the initial and terminal stages, respectively, compared with the mid-stage baseline. The interaction term coefficients β ₄ and β ₅ represent the moderating effect of illness stage on anonymity’s effect, indicating the additional change in anonymity’s marginal effect during the initial and terminal stages. Consequently, the total effect of anonymity is β ₁ +β ₄ during the initial stage and β ₁ +β ₅ during the terminal stage. Testing the statistical significance of these interaction terms enables an in-depth examination of how anonymity’s influence mechanisms evolve throughout disease progression.

Result

We employed two-way fixed effects models with user and year fixed effects to estimate the effects of anonymity and illness stage on user behaviors. The longitudinal panel design tracked 215 users over a six-year period (2011-2024), capturing 916 identity switches and enabling within-person comparisons across illness stages. Table 6 presents the main effects of anonymity (H1), Table 7 presents the main effects of illness stage (H2), and Table 8 presents the interaction effects between anonymity and illness stage (H3). Table 11 provides a complete summary of all hypotheses, analytical methods, key coefficients, and whether each hypothesis was supported. All models include year fixed effects (2019–2024, with pre-2019 as baseline), and standard errors are clustered at the user level.

Table 6.

Main effects of anonymity (H1).

Dependent variable	Coefficient	Std. Error	p-value	Model fit
Likes (log)	-0.0310	0.1032	0.764	R2 = 0.053
Comments (log)	-0.0485	0.1089	0.656	R2 = 0.070
Rationality	-0.0868	0.0650	0.182	R2 = 0.008
Information Breadth	-0.0092	0.0399	0.817	R2 = 0.017
Information Sensitivity	-0.0063	0.0053	0.231	R2 = 0.005
Emotion (1=positive)	-0.1963*	0.0926	<0.05	Log-likelihood = -971.27
Motivation (1=extrinsic)	-0.3183**	0.1096	<0.01	Log-likelihood = -635.26

Note. ***p<0.001, **p<0.01, *p<0.05. For continuous variables, coefficients are from linear fixed effects models with user and year fixed effects. For binary variables, coefficients are from conditional logit models with user and year fixed effects.R² represents within-unit (user) variance explained by the fixed effects models. Standard errors clustered at user level.

Table 7.

Main effects of illness stage (H2) (baseline = mid-stage).

Dependent variable	Stage_Initial			Stage_Terminal
	Coefficient	Std. Error	p-value	Coefficient	Std. Error	p-value
Likes (log)	0.5200***	0.1348	<0.001	0.7947***	0.1354	<0.001
Comments (log)	0.6555***	0.1380	<0.001	0.6560***	0.1347	<0.001
Rationality	-0.1526	0.0808	0.059	-0.0975	0.0885	0.271
Information Breadth	0.1681**	0.0551	<0.01	0.1977***	0.0587	<0.001
Information Sensitivity	0.0095	0.0066	0.153	0.0060	0.0071	0.393
Emotion (1=positive)	-0.5349***	0.1285	<0.001	-0.5368***	0.1275	<0.001
Motivation (1=extrinsic)	-0.2120	0.1527	0.165	-0.7761***	0.1677	<0.001

Note. Model fit indices are the same as in Table 6 for each dependent variable. See Table 6 notes for model specifications.

Table 8.

Interaction effects (Anonymity × Illness Stage) (H3).

Dependent variable	Anonymous × initial			Anonymous × terminal
	Coefficient	Std. Error	p-value	Coefficient	Std. Error	p-value
Likes (log)	-0.1785	0.1258	0.156	-0.3844**	0.1379	<0.01
Comments (log)	-0.3336*	0.1393	<0.05	-0.3133*	0.1581	<0.05
Rationality	0.0559	0.0859	0.516	0.1126	0.0853	0.187
Information Breadth	-0.0453	0.0521	0.385	-0.1029	0.0560	0.067
Information Sensitivity	0.0023	0.0072	0.749	0.0073	0.0078	0.347
Emotion (1=positive)	-0.2581*	0.1245	<0.05	-0.2651*	0.1247	<0.05
Motivation (1=extrinsic)	0.2967*	0.1485	<0.05	0.3571*	0.1620	<0.05

Note. Model fit indices are the same as in Table6 for each dependent variable. See Table 6 notes for model specifications.

Main effects of anonymity (H1)

Table 6 reports the coefficients for anonymity across all dependent variables. Anonymity significantly increased intrinsic disclosure motivation (β=−0.318, p<.01), supporting H1b₁ and indicating that anonymous posts are more driven by internal needs such as personal recording or emotional venting rather than seeking external support. Anonymity also significantly increased the expression of negative emotions (β=−0.196, p<.05), supporting H1b₂ and suggesting that anonymous posts contain more negative emotional content.

However, anonymity did not significantly affect the number of likes (β=−0.031, p=.764) or comments (β=−0.049, p=.656); thus, H1a₁ and H1a₂ were not supported, indicating that anonymous and real-name posts receive similar levels of social feedback. Consistent with H1b₃, anonymity had no significant effect on rational expression (β=−0.087, p=.182). H1b₄ and H1b₅ were not supported, as anonymity did not increase information breadth (β=−0.009, p=.817) or information sensitivity (β=−0.006, p=.231).

These findings show that anonymity has selective effects: it increases intrinsic motivation and negative emotional expression but does not affect social feedback, rational expression, or factual information disclosure.

Main effects of illness stage (H2)

Table 7 reports the coefficients for the initial and terminal stages relative to the mid-stage baseline. Compared to the mid-stage, posts during the initial stage received significantly more likes (β=0.520, p<.001) and comments (β=0.656, p<.001), and posts during the terminal stage also received significantly more likes (β=0.795, p<.001) and comments (β=0.656, p<.001). These findings support H2a₁and H2a₂, indicating that critical stages attract more social feedback from the community.

For self-disclosure, the terminal stage showed significantly stronger intrinsic motivation (β=−0.776, p<.001), supporting H2b₁, while the initial stage showed a negative but non-significant effect (β=−0.212, p=.165). Both initial (β=−0.535, p<.001) and terminal stages (β=−0.537, p<.001) exhibited significantly more negative emotions, supporting H2b₂. Information breadth increased significantly in both critical stages (initial:β=0.168, p<.01; terminal:β=0.198, p<.001), supporting H2b₄.

Illness stage did not significantly affect information sensitivity (initial:β=0.010, p=.153; terminal:β=0.006, p=.393); thus H2b₅ was not supported. Consistent with H2b₃, illness stage had no significant effect on rational expression (initial:β=−0.153, p=.059; terminal:β=−0.098, p=.271).

These patterns confirm that illness stage shapes user behavior: critical stages increase social feedback, negative emotions, and information breadth, while leaving information sensitivity and rational expression unchanged.

Interaction effects (anonymity × illness stage) (H3)

Table 8 reports the interaction terms between anonymity and illness stage, revealing how the effects of anonymity change across the cancer journey. For likes, the negative effect of anonymity was significantly stronger during the terminal stage (β=−0.384, p<.01) but not during the initial stage (β=−0.179, p=.156), partially supporting H3a₁ (significant for terminal stage only). For comments, the negative effect of anonymity was significantly stronger during both initial (β=−0.334, p<.05) and terminal stages (β=−0.313, p<.05), supporting H3a₂.

Anonymity’s effect on negative emotional expression was significantly strengthened during both critical stages (initial:β=−0.258, p<.05; terminal:β=−0.265, p<.05), supporting H3b₂. Anonymity’s positive effect on intrinsic motivation (i.e., reducing extrinsic motivation) was significantly weakened during both critical stages (initial:β=0.297, p<05; terminal:β=0.357, p<.05), supporting H3b₁.

Consistent with H3b₃, H3b₄, and H3b₅, illness stage did not significantly moderate the effects on rational expression (initial:β=0.056, p=.516; terminal:β=0.113, p=.187), information breadth (initial:β=−0.045, p=.385; terminal:β=−0.103, p=.067), or information sensitivity (initial:β=0.002, p=.749; terminal:β=0.007, p=.347).

Together, these interaction effects show that in critical stages, anonymity reduces social feedback more strongly but also enables more emotional expression, while its effect on intrinsic motivation weakens. This pattern reveals that anonymity is not a fixed tool but a resource whose consequences depend on the user’s situation.

Dynamic identity management strategies across illness stages

To understand how users dynamically manage identity as their illness progresses, we integrated the main effects of anonymity (Tables 6 and 7) with the significant interaction effects (Table 8) to calculate the total effect of anonymity for each illness stage. Table 9 presents these total effects numerically, and Figure 5 provides a heatmap visualization of how anonymity’s influence shifts across the illness stage.

Table 9.

Total effects of anonymity by illness stage.

Dependent variable	Initial stage	Mid-stage	Terminal stage
Likes (log)	-0.2095	-0.0310	-0.4154
Comments (log)	-0.3821	-0.0485	-0.3618
Rationality	-0.0309	-0.0868	0.0258
Information Breadth	-0.0545	-0.0092	-0.1121
Information Sensitivity	-0.0040	-0.0063	0.0010
Emotion (1=positive)	-0.4543	-0.1963	-0.4614
Motivation (1=extrinsic)	-0.0216	-0.3183	0.0389

Note. ***p<0.001, **p<0.01, *p<0.05. Total effect = main effect of anonymity + corresponding interaction term (Anonymous × Stage). For binary variables, coefficients are from conditional logit models.

Figure 5.

Heatmap of Total Effects of Anonymity Across Illness Stages.

Two key patterns emerge from Figure 5: temporal dynamics (how the same behavior changes across stages) and stage-specific configurations(how different behaviors within the same stage contrast with each other). These patterns, detailed in the following sections, reveal the strategic nature of dynamic identity management.

Initial Stage:Real-Name for Information Seeking, Anonymity for Emotional Expression

In the initial diagnosis stage, patients face high uncertainty and urgent information needs. The interaction between anonymity and initial stage had a significant negative effect on comments (β=−0.334, p<.05, Table8). As shown in Table 9, the total effect of anonymity on comments in the initial stage was −0.382 (moderate red in Figure 5). This negative effect confirms that anonymity reduces interaction precisely when users most need informational support. Therefore, users tend to post with real names to obtain more feedback from the community.

At the same time, anonymity significantly increased negative emotional expression during the initial stage (anonymity×initial stage:β=−0.258, p<.05, Table8), with a total effect of −0.454 (dark red in Figure 5). This indicates that anonymous posts express substantially more fear and anxiety triggered by diagnosis. Expressing these emotions under a real name may feel too vulnerable, so anonymity provides a safe space for emotional release.

real-name posts serve information seeking, while anonymous posts serve emotional expression. Users accept reduced interaction on anonymous posts because the primary goal is emotional release, not conversation. Viewed across the illness stage (Figure 5), the initial stage marks the beginning of a U-shaped pattern (high-low-high pattern) for emotional expression: anonymity’s effect is strong at this stage (−0.454), will weaken during the mid-stage, and re-intensify at the terminal stage—a stage that mirrors the intensity of patients’ emotional needs.

Mid-Stage: Anonymity Supports Intrinsic Motivation and Emotional Expression

In the mid-stage (treatment, remission, deterioration), the illness stabilizes and social urgency decreases. The main effects of anonymity on likes and comments were not significant (p>.6, Table6), indicating that anonymous posting does not reduce interaction during this phase. Consistent with this, Figure 5 shows effects near zero for likes and comments.

Anonymity significantly increased intrinsic motivation during the mid-stage (main effect β=−0.318, p<.01, Table6). Anonymous posts are more driven by internal purposes such as personal recording or emotional venting rather than seeking external support. This is the only variable showing a blue effect in Figure 5, representing a genuine shift toward intrinsic expression.

Anonymity also significantly increased negative emotional expression during the mid-stage (main effect β=−0.196, p<.05, Table6). Even as the crisis recedes, cancer continues to evoke difficult emotions, and anonymity continues to provide a safe channel for expressing them.

Importantly, in the mid-stage, users can express emotions through anonymity without sacrificing social interaction. The anonymity penalty disappears, allowing anonymity to serve its purest function: creating space for authentic self-expression without social cost. This stage represents the trough of the U-shaped curve for emotional expression (Figure 5), where anonymity’s effect weakens to −0.196. Conversely, it is the peak of the inverted U-shaped curve for disclosure motivation (−0.318), marking the moment when anonymity most purely serves intrinsic, self-directed expression.

Terminal Stage: Real-Name for Meaning-Making, Anonymity for Emotional Management

In the terminal stage (end-of-life, survival), patients face existential concerns and the need to construct life meaning. The interaction between anonymity and terminal stage had a significant negative effect on likes (β=−0.384, p<.01,Table8), with a total effect of −0.415 (darkest red in Figure 5). Anonymous posts receive substantially fewer likes, which represent emotional validation. Consequently, users adopt real names to obtain more social connection.

However, anonymity’s facilitative effect on negative emotional expression was also strongest in the terminal stage (interaction β=−0.265, p<.05,Table8), with a total effect of −0.461 (darkest red in Figure 5). Even when users choose real names for meaning-making, they still rely on anonymity for the most private and painful emotions such as fear of death and grief.

Regarding motivation, the anonymity × terminal stage interaction was significantly positive (β=0.357, p<.05, Table8). This shifted the total effect of motivation from negative (−0.318 in mid-stage) to near zero (β=0.039 in terminal stage, near-white in Figure 5). The effect of anonymity on intrinsic motivation weakens, as even anonymous posts become more externally focused, likely driven by a need for connection at life’s end.

Users employ real-name posts for life narrative and social connection, while using anonymity as a supplementary channel for private emotional management. Within-stage contrast (Figure 5) reveals the polarized configuration of the terminal stage: the left side (social feedback) shows the deepest red (strongest negative effect), while the right side (emotional expression) also shows the deepest red (strongest positive effect). This visually captures the core trade-off—users accept the highest social visibility cost in exchange for the safest space to release their most profound emotions. Viewed temporally, this stage completes the U-shaped curve for emotional expression (−0.461) and marks the dissipation of anonymity’s effect on intrinsic motivation (near-white,0.039).

From the initial to the mid-stage, anonymity’s negative effect on comments decreased from −0.38 to near zero, indicating that the social cost of anonymity disappears as the illness stabilizes. At the same time, anonymity’s positive effect on intrinsic motivation reached its strongest in the mid-stage (−0.318), showing that anonymity shifts from being primarily an emotional safety space to a tool for self-recording and intrinsic expression.

From the mid-stage to the terminal stage, anonymity’s negative effect on likes increased sharply from near zero to −0.42, as the need for emotional validation re-emerges. Meanwhile, anonymity’s effect on negative emotional expression deepened from −0.20 to −0.46, indicating that even as users seek connection through real names, they still need anonymous space for processing the deepest emotions.

Users dynamically adjust their identity strategies across the illness stage. In the initial stage, real names serve information seeking and anonymity serves emotional release. In the mid-stage, anonymity supports intrinsic motivation and emotional expression without social cost. In the terminal stage, real names support meaning-making while anonymity continues to serve private emotional management. As Figure 5 illustrates, these shifts manifest as a U-shaped stage for emotional expression, an inverted U-shaped stage for disclosure motivation, and a progressively intensifying penalty for social feedback from the mid-stage to the terminal stage. These changes are driven by stage-specific needs—information, introspection, and existential meaning—and reveal that anonymity is not a fixed choice, but a strategic resource whose function adapts as needs shift.

Heterogeneity analysis

Subgroup analyses reveal that the effects of anonymity vary systematically by gender, education, and marital status (see Table 10), demonstrating that dynamic identity management is not only a temporal process but also socially situated.

Table 10.

Heterogeneity analysis results.

Grouping variable	Outcome variable	Subgroup	Effect size	p-value	Anonymous mean	Real name mean	Effect direction
Gender	Emotion (Pos/Neg)	Male	-0.2246***	<0.001	0.4142	0.6387	Negative
Gender	Disclosure Motivation	Male	0.0842**	<0.01	0.1785	0.0942	Positive
Marital Status	Emotion (Pos/Neg)	Married	-0.1981***	<0.001	0.4355	0.6335	Negative
Marital Status	Emotion (Pos/Neg)	In Relationship	-0.3724***	<0.001	0.3276	0.7	Negative
Marital Status	Disclosure Motivation	Married	0.1004**	<0.01	0.1935	0.0932	Positive
Marital Status	Likes (log)	Married	-0.4997**	<0.01	2.583	3.0828	Negative
Education Level	Emotion (Pos/Neg)	High School	-0.1817**	<0.01	0.4183	0.6	Negative
Education Level	Emotion (Pos/Neg)	Primary School	-0.3701*	<0.05	0.3571	0.7273	Negative
Education Level	Information Breadth	Associate Degree	-2.75*	<0.05	2.25	5	Negative
Education Level	Likes (log)	High School	-0.4708*	<0.05	2.656	3.1268	Negative
Education Level	Emotion (Pos/Neg)	Doctoral Degree	-0.3064*	<0.05	0.3778	0.6842	Negative

The term “effect direction” refers to anonymity versus real-name identification.

Gender. Anonymity significantly increased negative emotional expression among male users (β=−0.225, p<.001) but not among female users, and enhanced intrinsic motivation for male users (β=0.084, p<.01). This pattern suggests that male users treat anonymous spaces as informational refuges for self-regulated emotion management, consistent with offline social norms that constrain male emotional expression. The absence of significant effects for female users may indicate that they already have greater latitude for emotional expression in real-name contexts, or that they employ different coping strategies.

Education Level. Anonymity’s effect on negative emotional expression was significant across all education levels (high school:β=−0.182, p<.01; primary school: β=−0.370, p<.05; doctoral:β=−0.306, p<.05), indicating a universal “emotional safe haven” effect that transcends educational attainment. This equalizing function of anonymity is particularly important for less-educated users who may have limited access to professional psychological support offline. However, for users with associate degrees, anonymity significantly reduced information disclosure breadth (β=−2.75, p<.05), suggesting heightened sensitivity to privacy boundaries within this group—a finding that warrants further investigation.

Marital Status. For married users and those in relationships, anonymity significantly increased negative emotional expression (married:β=−0.198, p<.001; in relationship: β=−0.372, p<.001) and enhanced intrinsic motivation for married users (β=0.100, p<.01), while reducing social feedback (likes:β=−0.500, p<.01). This pattern indicates that when real-life intimate relationships impose constraints on emotional disclosure—perhaps to avoid burdening partners—anonymous online spaces become crucial alternative channels for unfiltered expression and inward-focused emotional management.

These heterogeneity findings underscore that users’ strategic identity choices emerge from the interplay among individual psychological needs, illness contexts, and socio-structural positions. They also highlight the need for tailored platform designs that accommodate diverse user groups.

The models include year fixed effects (2019–2024, with pre-2019 as baseline). Many year dummies are statistically significant, particularly for likes and comments (e.g., year_2023 for likes: β=1.165, p<0.001), indicating increasing platform engagement over time. However, these time trends do not alter the core findings regarding anonymity and illness stage, as the main coefficients remain robust after controlling for year effects.

As summarized in Table11, all hypotheses proposed in Table1 have been empirically tested. Of the 25 hypothesis tests, 17 were supported (significant in the predicted direction), 4 items are not supported(Significance in prediction direction is not significant), 1 were partially supported, and 5 null hypotheses were confirmed as predicted.

Table 11.

Summary of hypotheses, analytical methods, and key findings.

Hypothesis	Dependent variable	Method	Key finding (coefficient)	p-value	Supported?
H1a1	Likes (log)	FE linear	−0.031	0.764	No
H1a2	Comments (log)	FE linear	−0.049	0.656	No
H1b1	Motivation (extrinsic)	Conditional logit	−0.318**	<0.01	Yes
H1b2	Emotion (positive)	Conditional logit	−0.196*	<0.05	Yes
H1b3	Rationality	FE linear	−0.087	0.182	Yes (null)
H1b4	Information Breadth	FE linear	−0.009	0.817	No
H1b5	Information Sensitivity	FE linear	−0.006	0.231	No
H2a1	Likes – Initial	FE linear	0.520***	<0.001	Yes
H2a1	Likes – Terminal	FE linear	0.795***	<0.001	Yes
H2a2	Comments – Initial	FE linear	0.656***	<0.001	Yes
H2a2	Comments – Terminal	FE linear	0.656***	<0.001	Yes
H2b1	Motivation – Terminal	Conditional logit	−0.776***	<0.001	Yes
H2b2	Emotion – Initial	Conditional logit	−0.535***	<0.001	Yes
H2b2	Emotion – Terminal	Conditional logit	−0.537***	<0.001	Yes
H2b3	Rationality – Initial	FE linear	−0.153	0.059	Yes (null)
H2b3	Rationality – Terminal	FE linear	−0.098	0.271	Yes (null)
H2b4	Breadth – Initial	FE linear	0.168**	<0.01	Yes
H2b4	Breadth – Terminal	FE linear	0.198***	<0.001	Yes
H2b5	Sensitivity – Initial	FE linear	0.010	0.153	No
H2b5	Sensitivity – Terminal	FE linear	0.006	0.393	No
H3a1	Likes × Terminal	FE linear	−0.384**	<0.01	Partial
H3a2	Comments × Initial	FE linear	−0.334*	<0.05	Yes
H3a2	Comments × Terminal	FE linear	−0.313*	<0.05	Yes
H3b1	Motivation × Initial	Conditional logit	0.297*	<0.05	Yes
H3b1	Motivation × Terminal	Conditional logit	0.357*	<0.05	Yes
H3b2	Emotion × Initial	Conditional logit	−0.258*	<0.05	Yes
H3b2	Emotion × Terminal	Conditional logit	−0.265*	<0.05	Yes
H3b3	Rationality × Stage	FE linear	n.s.	>0.05	Yes
H3b4	Breadth × Stage	FE linear	n.s.	>0.05	Yes
H3b5	Sensitivity × Stage	FE linear	n.s.	>0.05	Yes

Note. FE = fixed effects; n.s. = not significant (p > .05). For binary outcomes (emotion, motivation), coefficients are from conditional logit models.

Discussion

This study proposed the Dynamic Identity Management (DIM) framework to explain how individuals adapt their use of anonymity as their illness progresses. The framework’s central premise is that the effects of anonymity depend on the illness stage, rather than being determined solely by stable individual preferences. The findings support this premise. Across the illness trajectory, users shifted their identity strategies in response to changing needs: in high-risk stages (initial diagnosis and terminal), they used anonymity for emotional release, even though this reduced social feedback relative to real-name posting; in the low-risk mid-stage, they used anonymity for intrinsically motivated disclosure without losing social feedback. This pattern of stage-contingent identity switching cannot be explained by static models that treat anonymity as a stable user preference. The DIM framework, by specifying illness stage as a moderator grounded in situational strength, accounts for when and why the same individual switches between anonymous and real-name posting over time.

This study examined how anonymity affects self-disclosure and social feedback in online health communities, and how illness stage moderates these effects. Using longitudinal panel data from Zhihu, we found that anonymity increases intrinsic disclosure motivation and negative emotional expression but does not affect likes, comments, or factual information disclosure. These findings are consistent with Privacy Calculus Theory and the SIDE model: anonymity reduces perceived social risk and creates a safe space for emotional expression.^26,27 However, anonymity does not change how users share identity-related information, suggesting that normative boundaries around such information persist even under anonymity. ¹⁶

This “motivation-behavior dissociation” challenges the view that anonymity uniformly facilitates all disclosure^28,32 and aligns with Nissenbaum’s theory of contextual integrity, ¹⁶ which posits that online health communities have norms that limit identity information sharing even under anonymity. While previous research treats anonymity as a fixed user attribute,^5,8 we show that users switch between anonymous and real-name identities strategically over time.

Illness stage moderates these effects. During initial diagnosis and terminal stages, anonymity’s negative effect on comments becomes stronger, and its negative effect on likes strengthens in the terminal stage. At the same time, anonymity’s effect on negative emotional expression is amplified in both critical stages. This pattern indicates that users trade social feedback for emotional safety when facing high-stakes health situations.

The total effects of anonymity by illness stage confirm this pattern: for emotional expression, anonymity’s effect is larger in critical stages than in the mid-stage, while for disclosure motivation, anonymity’s effect is strongest in the mid-stage and weakens in critical stages. These results demonstrate that users adjust their use of anonymity based on their illness stage.

Applying Situational Strength Theory, ⁴¹ we find that illness stage alters anonymity’s effects, advancing research beyond simple main-effect models. While previous studies note that health context matters,^36,43 few show how it changes anonymity’s effects. Our findings refine the SIDE model: in high-stakes situations, community members discount anonymous posters more heavily, yet users rely more on anonymity for emotional expression. The significant effects of illness stage on emotion support Social Penetration Theory,^35,36 while the non-significant effect of initial stage on motivation suggests early disclosure is more information-driven. ³⁷

We also identify boundary conditions for anonymity effects: while anonymity influences emotional and motivational disclosure, it does not affect rational knowledge sharing or identity revelation—a dimensional heterogeneity overlooked in prior work.^25,31 This pattern aligns with dual-process models, ⁴⁰ where emotional disclosure responds automatically to safety cues, while informational disclosure depends on cognitive resources and social utility.^33,34

Heterogeneity analysis shows that these effects vary by user characteristics. Anonymity has a stronger effect on emotional expression for male users, less educated users, and married users, suggesting that offline social roles influence how people use anonymity online.^23,44 Specifically, male users use anonymity to express emotions, consistent with gender norms that discourage male emotional expression offline. ²³ Less educated users benefit more from anonymity’s emotional release, potentially compensating for limited access to professional support. ⁴⁵ Married users may use anonymity to avoid burdening partners with negative emotions. ⁴⁶

Taken together, these findings reveal a clear pattern of dynamic identity management: users strategically shift their identity use as their needs change across the illness stage. In high-risk phases, they prioritize emotional safety over social connection; in the low-risk mid-stage, they leverage anonymity for intrinsic expression without social cost.

Conclusion

This study proposes a Dynamic Identity Management (DIM) framework to visualize how anonymity shapes self-disclosure and social feedback across illness stages, and provides empirical validation using panel data from 215 Zhihu users. The findings show that anonymity increases intrinsic motivation and negative emotional expression but does not affect social feedback or factual information disclosure. Illness stage moderates these effects: in initial and terminal stages, users accept fewer likes and comments in exchange for safe emotional expression. Heterogeneity analysis further reveals that these patterns vary by gender, education, and marital status, demonstrating that users strategically manage anonymity to navigate the tension between connection and exposure. This stage-contingent pattern supports the DIM framework’s central proposition: the effects of anonymity depend on the illness stage.

The study advances OHC research in four key ways. First, it shifts the conceptualization of anonymity from a static attribute to a dynamic resource that users adapt across contexts. Second, it demonstrates how situational factors, specifically illness stage, reconfigure the effects of anonymity on disclosure and feedback. Third, it identifies boundary conditions by showing that anonymity selectively influences emotional and motivational dimensions while leaving rational and factual disclosure unchanged. Fourth, it uncovers social group differences, revealing that offline roles (gender, education, marital status) shape how individuals use anonymity online. These findings carry practical implications for platform design: OHCs should offer stage-sensitive anonymity options, provide separate features for emotional and informational support, and develop tailored approaches for diverse user groups.

This study has several limitations. First, the data come from a single platform (Zhihu), whose user base is more urban and educated than the general cancer population, potentially limiting generalizability. Second, textual data cannot capture unobservable psychological motivations; future research could complement this approach with surveys or interviews. Third, while user fixed effects control for time-invariant heterogeneity, time-varying unobserved factors may still bias estimates. Fourth, findings may not extend to less stigmatized health conditions, where the trade-off between connection and exposure could differ.

Future research should examine these dynamics across different platforms, cultures, and disease types to further validate and extend the DIM framework. Longitudinal studies could examine whether identity use influences long-term mental health outcomes, and experimental designs could evaluate whether stage-sensitive features improve user well-being. Developing validated measures of Privacy Calculus and situational awareness would also enable more precise testing of the DIM framework’s underlying mechanisms.