The research team, led by Wen Deng of Zhongnan University of Economics and Law together with Zijun Mao and Yi Yang, analyzed 126 fertility-policy posts published by People’s Daily on Weibo and the 196,218 user comments attached to them between January 2013 and June 2025. People’s Daily, the Communist Party’s central mouthpiece, joined Weibo in 2012 and had acquired over 158 million followers by July 2026, making its account a quintessential venue where official communication meets user commentary. Each post and its aggregated comments formed one paired communication event, yielding 252 documents totaling more than 4.5 million words. The corpus was organized into three non-overlapping policy phases: the selective two-child reform period with 18 events, the universal two-child policy period with 75 events, and the three-child and pronatalist-support period with 33 events.

The methodological core combined Latent Dirichlet Allocation, a hierarchical Bayesian topic model that represents each document as a probability distribution over latent topics, with network agenda-setting analysis. The team estimated candidate models with two to twelve topics, each solution run across five fixed random seeds, and selected the four-topic solution using an equally weighted composite of semantic coherence and model fit. The four topics were interpreted as substantive policy attributes: Policy Change and Childrearing Burdens, anchored by terms such as children, policy, second child, unable to afford, education, and housing prices; Women’s Employment and Work-Family Responsibilities, built around women, enterprises, maternity leave, discrimination, and wages; Gender Preference and Family Norms, covering son preference, bride price, rural areas, and sex ratio; and Maternity Care and Childbirth Support, dominated by painless delivery, hospital, anesthesiologist, and anesthesia. Because media and user texts were pooled to train a common model, the topic probabilities estimated for both agendas occupied the same latent topic space, a prerequisite for direct comparison.

To trace temporal precedence, the researchers applied cross-lagged correlation analysis anchored to the Rozelle-Campbell baseline. Rather than using fixed daily or weekly intervals, which would have produced numerous empty observations given the irregular rhythm of policy communication, the analysis ran on an event clock: a one-event lag represented the transition from one communication episode to the next chronologically ordered episode. The intervals between the 125 adjacent-event pairs ranged from 0 to 560 days, with a median of 7 days and a mean of 36.23 days, a positively skewed distribution driven by a small number of long gaps. For each adjacent pair, five Pearson correlations were computed, including the media-to-user and user-to-media event-lagged correlations, and each directional coefficient was evaluated against a baseline incorporating contemporaneous media-user correspondence and within-agenda continuity. A directional association was classified as exceeding the baseline when the difference was greater than 0.01.

The results complicate any binary account of top-down propaganda versus bottom-up public opinion. In the first two policy periods, media-preceding-only patterns were descriptively more frequent than user-preceding-only patterns, but most adjacent event pairs showed neither direction exceeding the baseline: 58.8 percent in Period 1 and 52.7 percent in Period 2. In Period 3, the picture reversed. The proportion classified as neither dropped to 28.1 percent, user-preceding-only patterns rose to 28.1 percent, and media-preceding-only and bidirectional patterns each accounted for 21.9 percent. It was the only period in which user-preceding-only patterns outnumbered media-preceding-only patterns, a finding consistent with a potentially more reciprocal communication environment as policy moved from relaxation to active pronatalist mobilization.

Topic-level distributions revealed equally nuanced asymmetries. Women’s Employment and Work-Family Responsibilities showed the smallest distributional difference between the two agendas, with a Kolmogorov-Smirnov statistic of 0.198 and a small positive mean difference of 0.017 indicating a slight media advantage. Policy Change and Childrearing Burdens received greater average salience in the user agenda, with a negative mean difference of 0.047. The largest divergence occurred for Gender Preference and Family Norms, with a KS statistic of 0.675, even though its mean salience was nearly identical in both agendas. This combination is theoretically important: similar average attention does not imply that the two agendas emphasize a topic during the same communication events. Maternity Care and Childbirth Support likewise diverged substantially at the event level despite a modest media advantage in mean salience.

The network analyses added a second layer of insight. Descriptive topic joint-intensity networks, in which edge weights were calculated as the mean product of event-level topic probabilities, showed that both agendas underwent a broadly similar thematic transition. In Period 1, both were strongly centered on Policy Change and Childrearing Burdens, with mean probabilities of 0.705 in the media agenda and 0.766 in the user agenda, and the strongest connection in both networks linked policy change to maternity care. By Period 2, the thematic center shifted toward women’s employment, and by Period 3 the strongest edge in both networks connected Policy Change and Childrearing Burdens to Women’s Employment and Work-Family Responsibilities, with joint intensities of 0.078 in the media network and 0.096 in the user network. Gender Preference and Family Norms remained peripheral throughout.

Yet aggregate thematic similarity did not translate into stronger event-level structural correspondence. Using Quadratic Assignment Procedure correlations, which compare two 126-by-126 event-similarity matrices built from cosine similarities between four-topic probability vectors while accounting for the non-independence of network ties, the team found a strong positive correlation across the full sample, r = 0.641. The period-specific pattern was non-monotonic: correspondence was moderately strong in Period 1, r = 0.551, peaked in Period 2, r = 0.687, and weakened substantially in Period 3, r = 0.365, although it remained statistically significant. By Period 3, the two agendas displayed highly comparable aggregate thematic centers, but they organized individual communication events in increasingly dissimilar ways. The authors caution that such correspondence could reflect shared responses to the same policy announcements rather than documented institutional feedback, since neither QAP nor event-lagged analysis can distinguish reciprocal adjustment from common-context mechanisms.

The gender analysis revealed a user agenda that was demographically and thematically heterogeneous. Among users with valid gender information, 38.79 percent were male and 61.21 percent female, and the female share rose from 52.01 percent in Period 1 to 67.39 percent in Period 3. Using a weighted dictionary-mapping procedure with terms assigned weights of 3, 2, or 1 for core, related, and background expressions, the team found that male users consistently devoted a larger share of attention to Policy Change and Childrearing Burdens, whereas female users showed greater attention to Women’s Employment and Work-Family Responsibilities and higher mapping shares for Maternity Care and Childbirth Support across all three periods. Attention to Gender Preference and Family Norms varied by both gender and period, with female users ahead in the first two periods and male users surpassing them in the third. These descriptive patterns align with the gendered distribution of employment, caregiving, and bodily consequences associated with fertility policy.

The study’s implications extend beyond communication theory into public health practice. The four topics identify distinct conditions through which users understand and evaluate fertility policy: the material feasibility of raising children, the employment and caregiving consequences of childbearing, the normative environment of reproductive decisions, and access to obstetric services. The authors argue that communication focused primarily on changes in fertility rules may leave material concerns about childcare, education, housing, and family costs insufficiently addressed, and that public health communicators should monitor event-specific user concerns and provide timely, service-oriented information rather than relying exclusively on generalized pronatalist messaging. The findings also carry caveats: the data came from a single platform and one state-media account, user comments were conditioned by exposure to the corresponding posts, and platform moderation, which has intensified as population decline was incorporated into China’s national security framework in 2024, means the measured user agenda represents surviving and observable discourse rather than the full range of public attitudes. Temporal precedence, the authors stress, denotes predictive association, not causal influence. What the thirteen-year record does establish is that agenda-setting power on China’s fertility policy is not a fixed property of either institutional media or the user, but a dynamic relationship whose direction, structure, and gendered emphasis shift across successive communication events and policy stages.

Subject of Research: Science News

Article Title: Changing predictive asymmetry: network agenda correspondence between state media and the user on China's fertility policy (2013–2025)

Article References: Deng, W., Mao, Z., & Yang, Y. (2026). Changing predictive asymmetry: network agenda correspondence between state media and the user on China's fertility policy (2013–2025). Frontiers in Public Health, 14, Article 1943400. https://doi.org/10.3389/fpubh.2026.1943400

Image Credits: AI Generated

DOI: 10.3389/fpubh.2026.1943400

Keywords: agenda-setting, China fertility policy, Weibo, network analysis, social media, public health, Changing, predictive, asymmetry, network, agenda, correspondence