We begin by analyzing the relationship between the default outcome and borrowers’ choice to disclose their microblog account. We use a logit regression model first, and then utilize the PSM technique and instrument variable regressions to address endogeneity concerns. We estimate:
$$ logit(Default)=\alpha +{\beta}_1 Microblog\_ disclosed+{\beta}_2 Controls+\varepsilon $$
(1)
The results show that the microblog disclosure is negatively related to the default probability (coefficient is −0.748), and is significant at the 0.01 level. However, the data are unbalanced in covariates between the group who discloses their microblog and those who do not. The unbalanced data weakens the reliability of the results of the regression model (Imbens and Wooldridge 2009). Therefore, we utilize PSM to adjust for the differences in covariates. The results after PSM support Hypothesis 1. We find significant differences in the default rate between treated and control groups.
Difference-in-difference (DID) model
Although the result of the logistic model shows that disclosure of a microblog account is a predictor of default probability, it does not identify the underlying cause: is it because the borrowers are afraid of social stigma costs? We use a DID model to identify the cause. In April of 2013, the P2P platform launched a marketing campaign to encourage borrowers to disclose their microblog accounts. We estimate the effect of the campaign on the default probability of a loan whose borrower disclosed their social media account. The estimated model is:
$$ \operatorname{l}n\left(\frac{P\left( Defaul{t}_{it}=1\right)}{1-P\left( Defaul{t}_{it}=1\right)}\right)=\alpha +{\beta}_1 Mb\_ disclose{d}_i+{\beta}_2Cm{p}_{it}+{\beta}_3 Mb\_ disclose{d}_i\times Cm{p}_{it}+{\beta}_4 Control{s}_i+{\varepsilon}_{it} $$
(2)
The dummy variable Mb_disclosed equals 1 if the borrower of a loan has disclosed his microblog, otherwise it equals 0. The dummy variable Cmp is a time variable, which takes the value of 0 or 1 for periods prior to or after the disclosure campaign. Controls represent a vector of loan characteristics, such as loan amount, interest rate, and lending period. The main parameter of interest is β3. The result of β3 is negative (−0.652) and significant (p < 0.01), suggesting that this campaign negatively influenced the default probability of the loans whose borrowers have disclosed their social media account. One possible reason is that these borrowers care about social stigma costs. Borrowers may worry the P2P lending company could use their microblog account as an outlet to spread the word if a default occurs, which would increase their social stigma costs. With this worry in mind, they are less likely to default after the disclosure campaign.
The effect of microblog behavior on default probability
We select borrowers who disclose their microblog accounts on the P2P lending site, and collect microblog metrics (e.g. #Followers, #Friends, #Fans and #Microblogs) from their profile pages on sina.com. This combined data sample includes 5239 listings.
We use a logit model to estimate the default probability of the effect of the microblog metrics on default likelihood for borrowers who have disclosed his microblog.
$$ logit(Default)=\alpha +{\beta}_1 Microblog\_ Metrics+{\beta}_2 Controls+\varepsilon $$
(3)
Because of the large variance and scale of the microblog metric variables, we use their natural logs in the model.
We first analyze the effect of #Followers and #Microblogs, respectively. The independent variable in both models is negatively related to the default probability at the 0.01 significance level. The results demonstrate that the larger the scope of the social network a borrower has on a social media site, the less likely they are to default on a loan; the more engagement a borrower has with his social media site, the less likely he is to default. Both Hypotheses 2a and 2b are supported.
We next examine the effect of two different types of social network, that is, friends and fans. For a borrower, both friends and fans on the microblog site are sources of social capital. Either friends or fans knowing about a borrower’s default can damage his social image and cause a social stigma cost; therefore, both #Friends and #Fans influence the borrower’s default likelihood. However, as previous studies have demonstrated, close friends have a stronger behavioral effect on each other than strangers do (Bond et al. 2012; Christakis and Fowler 2013). We therefore expect the effect of #Friends on borrowers’ default likelihood to be more intensive than that of #Fans. Our result shows that #Friends and #Fans are both negatively related to the default probability with p < 0.01, but the coefficient of #Friends (−0.153) is almost double to that of #Fans (−0.079). The results indicate that although both variables are predictors of default likelihood, #Friends is a stronger signal than #Fans.
We also consider that a borrower having a large #Followers is more likely to have a healthy financial situation as an influential person. Therefore, their low default probability may be due to their financial well-being instead of avoiding costs in social capital. We include an additional term to represent a borrower’s influence, which can also be regarded as a proxy for financial position. From the Sina microblog site, we received not only data showing how many followers a borrower has (e.g., #Followers), but also data showing how many people the borrower is following (e.g., #Following). It is reasonable to assume that #Followers of influential borrowers is always greater than #Following. Therefore, we created a dummy variable “Influential,” whose value equals 1 when #Followers is greater than #Following, and otherwise, equals 0. The result shows that #Followers remain significant while Influential is not significant.