Test for multicollinearity of variables

Figure 2 shows the correlation matrix of the variables, with darker shades indicating stronger correlations. The analysis shows that all coefficients are within an acceptable range, peaking at 0.72 between ‘posts’ and ‘reads’ on NP WeChat. The VIF test confirms that there is no problem with multicollinearity, as the highest VIF value is 2.32, well below the threshold of 10. This ensures the stability of the subsequent regression analysis.

Fig. 2 Heat map of the matrix of correlation coefficients between variables. Full size image

Baseline regression results

Our empirical analysis employs a two-way fixed effects model to examine the direct impact of policies, news, and academic conferences on public interaction behaviors, controlling for WeChat posts, consumption indices, account duration, and media types. Table 3 presents the regression outcomes. Each additional policy correlates with a 1.50% rise in ‘likes’ (p < 0.05) and a 1.40% increase in ‘watching’ (p < 0.10), affirming H1a and H1b, but not H1c. Additional news items boost ‘likes’ by 2.60% (p < 0.10), confirming H2a but not H2b or H2c. Conversely, each academic conference is linked to a 7.80% decrease in ‘likes’ (p < 0.05) and a 7.10% drop in ‘watching’ (p 0.05 for bs_2, indicating a positive mediating effect. The positive mediating effect is 2.69 times larger than the direct effect, which causes the direct effect of news on ‘watching’ to become insignificant. In the case of ‘reads’, both the Sobel p-value and bs_1 are less than 0.05, while bs_2 p = 0.167, indicating a positive mediating effect. The mediating effect is 1.59 times larger than the direct effect, making the direct effect of news on “reading” insignificant. This confirms that WeChat posts mediate the effect of news on NP WeChat interactions, thereby supporting hypotheses H6a, H6b, and H6c. However, the mediation of policies and academic conferences by posts was not supported, leading to the rejection of hypotheses H5a, H5b, H5c, H7a, H7b, and H7c.

Moderating effects of the administrative level

We examined the moderating influence of the NP administrative level on the relationship between policy, news, academic conferences, and public interaction behaviour using the following model:

$$\:\begin{array}{c}{Y}_{it}={\beta\:}_{0}+{\beta\:}_{1}{Policy}_{it}+{\beta\:}_{2}{News}_{it}+{\beta\:}_{3}{Conference}_{it}\\\:+{\beta\:}_{4}\left({Level}_{it}*{Policy}_{it})+{\beta\:}_{5}({Level}_{it}*{News}_{it}\right)\\\:+{\beta\:}_{6}\left({Level}_{it}*{Conference}_{it}\right)+\gamma\:{Controls}_{it}+{\alpha\:}_{i}+{\lambda\:}_{t}+{\epsilon}_{it}\end{array}$$ (4)

Fig. 4 Path diagram visualizing the moderating effects of the NP administrative level. Full size image

Where \(\:{Level}_{it}\) represents the administrative level of NP \(\:i\) at time \(\:t\), and \(\:{\beta\:}_{4}\), \(\:{\beta\:}_{5}\), \(\:{\beta\:}_{6}\) are the coefficients of the interaction terms for policy, news, and academic conferences, respectively. The results in Table 7 indicate a significant positive interaction effect between NP levels and policies for ‘reads’, significant at the 5% level (β = 0.72), confirming H8c, while H8a and H8b were not supported. This suggests that higher NP levels enhance information dissemination on WeChat. Conversely, the interaction terms for news exhibit significant negative coefficients at the 1%, 5%, and 1% levels, implying that higher NP levels reduce the positive impact of news on interactions. This contradicts H9a, H9b, and H9c. Regarding academic conferences, the interaction effects were not significant, leaving H10a, H10b, and H10c unconfirmed. Figure 4 presents a path diagram visualising the moderating effects of the NP administrative level on public interaction behaviour.

Table 7 Moderating effects: administrative level. Full size table

Heterogeneity analysis: regional and developmental heterogeneity

Given the regional differences, NP social media interactions may vary due to factors such as online attention66. To explore this, the five NPs were divided into eastern and western groups based on their geographical location. The Three River Source NP and Chengdu Giant Panda NP are located in the west, while the remaining three NPs are located in the east. After excluding 58 unique observations, a multivariate analysis of variance (MANOVA), with p-values all approaching 0, revealed significant regional differences in the impact of environmental factors (such as policies, news, and academic conferences) on public interaction behaviour.

As shown in Table 8, the direct and moderating effects of policies and news differ by region. In the eastern group, policy significantly increases ‘likes’ (p < 0.01), ‘watching’ (p < 0.05), and ‘reads’ (p < 0.10), while in the west, no significant effects were observed. The interaction between administrative level and policies in the east has a positive effect on ‘likes’ (β = 0.34, p< 0.05), while in the west, the effect on ‘reads’ is significantly negative (β=−0.07, p < 0.10), suggesting that policy localisation plays a role in shaping public interaction behaviour67. The interaction between administrative level and news in the east negatively affects public interaction behaviour (p < 0.01), while no effect is found in the west. This suggests that the NP administrative level does not moderate the relationship between the academic environment and public interaction, reflecting the academic independence in NP research fields. Figure 5 presents the path diagram visualising the direct and moderated effects by region.

Fig. 5 Path diagram visualizing the direct and moderated effects by region. (a) Path diagram for western region; (b) Path diagram for eastern region. Full size image

Regarding the temporal dimension, the data were split into two periods: before (prior to October 2021) and after the establishment of the NP. Two observations were dropped due to uniqueness. The MANOVA revealed significant differences in the impact of policy, news and academic conferences across these periods. The regression results in Table 9 show that news significantly increased ‘likes’ (p < 0.05), ‘watching’ (p < 0.10) and ‘reads’ (p< 0.05) in the post-establishment period, while no significant effects were found in the pre-establishment period. Academic conferences had a negative effect on ‘likes’ (p < 0.05) and ‘watching’ (p < 0.05) in the pre-establishment period, with no effect in the post-establishment period, indicating the evolution of NP policy across developmental stages68.

In the moderated effects model, the interaction of administrative level with policy positively influenced ‘reads’ (β = 0.07, p < 0.05), while its effect on ‘watching’ in the pre-establishment period was significantly negative (β = −0.43, p < 0.10). The interaction between administrative level and news negatively affected public interaction behaviour in the pre-establishment period (p < 0.01), with no significant effect on ‘likes’ and ‘watching’ after the NP establishment. The lack of a moderating effect of administrative level between academic conferences and WeChat interactions highlights the academic autonomy within NP research fields. Figure 6 visualises the direct and moderating effects by temporal dimension.

Fig. 6 Path diagram visualizing the direct and moderate effects by region. (a) Path diagram for the pre-establishment period; (b) Path diagram for the post-establishment period. Full size image

Table 8 Heterogeneity analyses: regressions of area heterogeneity groupings. Full size table