*5.2. Relationship Establishment*

In order to show the formation process of the relationship in the consumer advice network more intuitively, we present the evolution of the entire network through a network diagram. We divided the evolution of the network into four stages, starting from the initial stage of network establishment (December 2010) and then presenting it every three years. Although this is only a rough display, some simple conclusions can be drawn from Figure 3 (only nodes with an in-degree greater than 10 are shown): firstly, the network shows a dynamic change process as new nodes continue to join the network and new relationships are formed; and secondly, the formation of network relationships is extremely heterogeneous, with some nodes establishing more relationships and others very few.

**Figure 3.** Evolution of consumer advice network.

We also conducted frequency statistics on the in-degree of network members. Figure 4 shows the frequency distribution. We found that most members have less than 10 relationships and fewer members have more than 1000 relationships. Therefore, only a few members established a large number of relationships and became opinion leaders in the consumer advice network. Our research question asks why these members become opinion leaders.

**Figure 4.** Frequency distribution of the in-degree.

#### *5.3. Regression Analysis*

This paper uses IBM SPSS Statistics software (version 21) to estimate the coefficients of each variable using the "analysis-survival function-Cox regression" path. The main coefficients were estimated as shown in Table 5.



Notes: Sample *n* = 1356, \* *p* < 0.05, \*\* *p* < 0.01, and \*\*\* *p* < 0.001. The data analysis was conducted using IBM SPSS 21.0 software.

> In Model 1, we first considered the effect of content contribution. From the regression results, it can be seen that the content contribution is positively correlated with the establishment of the relationship(B = 0.090, *p* < 0.001). That is, hypothesis H1, "The more content a member contributes, the more other members will establish a relationship with him/her and the more likely he/she is to become an opinion leader", is supported.

> From the perspective of network structure, the establishment of a relationship has the characteristics of preferential attachment (B = 0.170, *p* < 0.001). That is, hypothesis H4, "The more followers a member has, the more other members will establish a relationship with him/her and the more likely he/she is to become an opinion leader", is supported.

However, when the number of fans (in-degree) is squared, the coefficient becomes 0, which shows that this effect is not linear but rather diminishes the margin.

In addition, the homophily among members also has a positive impact on the establishment of relationships(B = 0.150, *p* < 0.01). Therefore, hypothesis H5, "In consumer advice networks, homophily positively affects the establishment of network relationships", is supported.

From the perspective of relationship norms, the stronger the structural equivalence between members, the more likely it is to establish relationships, which reflects the role of overlap in network structures in promoting the establishment of network relationships (B = 0.941, *p* < 0.01). Additionally, its effect is far greater than that of other factors. That is, hypothesis H6, "In consumer advice networks, structural equivalence positively affects the establishment of network relationships ", is supported.

Hypotheses H4, H5, and H6 are consistent with the findings of previous studies. In traditional social networks, collaborative networks, and offline networks, preferential attachment mechanisms [86], structural equivalence [87], and homophily [88–90] play very important roles in the establishment of network relationships, in the development of cooperation, and in the formation of alliances.

However, hypothesis H7, "In consumer advice networks, reciprocity positively affects the establishment of network relationships", is not supported in this study. This conclusion is different from acquaintance social networks. In traditional acquaintance social networks, reciprocity has a strong positive impact on the establishment of network relationships. The reason is that in the consumer advisory network, members mainly engage in information acquisition and exchange. Members pay attention to opinion leaders but they do not draw the attention of opinion leaders to them. This shows that the exchange of information does not have the feature of reciprocity.

In Model 2, we further discussed the influence of behavioral activities on the establishment of relationships. The social interaction of members in the network has a positive effect on the establishment of relationships (B = 0.060, *p* < 0.001). Therefore, Hypothesis H2,"The more actively a member interacts online, the more other members will establish a relationship with him/her and the more likely he/she is to become an opinion leader", is supported. In addition, the total helpfulness of the content contributed by influencers has a positive effect on the establishment of relationships (B = 0.020, *p* < 0.005). Hypothesis H3, "The more helpful content a member contributes, the more other members will establish a relationship with him/her and the more likely he/she is to become an opinion leader", is supported.

The results of hypothesis H2 and hypothesis H3 were consistent with the results of previous studies. For example, Li suggested that in social blogs with word-of-mouth (WoM) marketing, an opinion leader, who is normally more interconnected and has a higher social standing, can deliver product information, provide recommendations, give personal comments, and supplement professional knowledge, which helps companies to promote their products [42].

In Model 3, we tested the robustness of the above analytical results by including the control variables in the regression model. The results from model 3 show that the conclusions obtained in model 1 and model 2 are valid in model 3, thus the results of the model are robust.

#### *5.4. Comparative Analysis*

To further explore the relationship between influencers' online behavioral activities, the network structure, and opinion leader formation, a comparative analysis of opinion leaders and non-opinion leaders was conducted.

Firstly, the influencers in consumer advice networks were classified according to the criteria proposed by the existing literature. Regarding the classification of opinion leaders in social networks, Trusov et al. stated that the top 20% of influential network nodes in the network are opinion leaders [91], while Goldenberg et al. classified them based on the degree criterion. In this study, the influencers who ranked in the top 20% in terms of the number of followers (i.e., in-degree)

the mean values of variables and Figure 5 shows the results of the

opinion leaders and

 **6.** Mean **Level SurvivalTime Out- Degree In-Degree Contribution Interaction Helpfulness** Opinion leaders Mean 6.60 102.11 98.25 1392.85 421.85 692.90 1610.20 Non-opinion leaders Mean 6.38 78.45 128.25 183.09 179.46 221.63 498.79 ANOVA F = 1.379 *p* = 0.243 F = 61.721 *p* < 0.001 F = 0.392 *p* = 0.533 F = 277.217 *p* < 0.001 F = 18.914 *p* < 0.001 F = 44.796 *p* < 0.001 F = 28.224 *p* < 0.001 Overall Mean 6.42 83.76 122.25 425.04 227.94 315.88 721.07 Max 7.00 132.00 1000.00 2680.00 1524.00 1810.00 6624.00 Min 4.00 34.00 1.00 56.00 7.00 8.00 23.00 SD0.7718.52191.17565.81242.27338.26944.89

**Table**valueofeachvariable.

 leaders.

non-opinion

 were set as opinion leaders [92]. Table 6 shows

comparison

 between

Notes: the analysis of variance was done by IBM SPSS 21.0 software.

**Figure 5.** Comparison between opinion leaders and non-opinion leaders.

As seen in Figure 5, there are differences in the behavioral activities and network structure between opinion leaders and non-opinion leaders; however, are there statistically significant differences? To answer this question, we used IBM SPSS Statistics software (version 21) to perform an analysis of variance on the two groups. In terms of the level, there was no significant difference between opinion leaders (MOL = 6.60) and non-opinion leaders (MN-OL = 6.38; F = 1.379, *p* = 0.243). However, there was a significant difference in the survival time of the two groups (MOL = 102.11, MN-OL = 78.45; F = 61.721, *p* < 0.001). From the regression analysis, we found that level and survival time did not affect the establishment of the relationship. However, the above results show that the survival time of opinion leaders was significantly longer than that of non-opinion leaders.

In terms of the out-degree, there was no significant difference between opinion leaders (MOL = 98.25) and non-opinion leaders (MN-OL = 128.25; F = 0.392, *p* = 0.533). There was a significant difference in the number of fans (i.e., in-degree) between the two groups (MOL = 1392.85, MN-OL = 183.09; F = 277.217, *p* < 0.001).

In terms of behavioral activities, there was a significant difference in the amount of content contributed by the two groups (MOL = 421.85, MN-OL = 179.46; F = 18.914, *p* < 0.001). There was a significant difference in the number of comments received by the two groups (MOL = 692.90, MN-OL = 221.63) (F = 44.796, *p* < 0.001).

There was also a significant difference in the number of likes obtained by the two groups (MOL = 1610.20, MN-OL = 498.79; F = 28.224, *p* < 0.001). The above results show thatin order to become an opinion leader in a consumer advice network, it is essential to actively contribute content and participate in interaction.
