Page 117 - 臺大管理論叢第32卷第1期

NTU Management Review Vol. 32 No. 1 Apr. 2022

Table 5). The three determinants included (1) three comment metrics (i.e., volume, positive
comments, and negative comments), (2) discrete emotions (i.e., excited, amused, praising,
ridiculing, complaining, and disappointed), and (3) streamers' characteristics (i.e., gender,
facial attractiveness, sociable style, persuasive style, and comical style) and behaviors (i.e.,

chatting with viewers, sharing food features, responding to viewers' questions, sharing
cooking, and tasting food processes). The comment metrics and discrete emotional features
of the viewers' comments are qualitative variables and treated with dummy variables. This

8
study adopts the package rjags for Bayesian data analysis. The parameter estimation of
the HB model uses Gibbs sampling with over 10,000 MCMC iterations, of which the first
10,000 iterations are discarded and the last 10,000 iterations are used to form estimates of
the posterior distribution of the HB model parameters.
The HB model is shown to offer a less restrictive method of pooling to obtain an

aggregate measure of the influence of the explanatory variables (Allenby, Jen, and Leone,
1996). One of the reasons is because the HB model assumes that the coefficients for each
live stream come from a common distribution. The beta is normally distributed, and the

-1
-2
parameters a and a of the prior distribution of the σ are given as a = a = 0.01. Ω
0
1
0
1
is distributed in the Wishart distribution: W (γ,R), and it is assumed that the diagonal
k
elements of R are all equal to 1, the rest are equal to 0, and parameter γ is set to equal to
1. Heterogeneity in response to the explanatory variables is observed when pooling data
across live streaming. Equation 2 serves to shrink the posterior coefficient estimates for
each live stream toward a hyperparameter. In addition, we have to assess the convergence
of MCMC simulations toward the posterior distribution. This study uses the Heidelbergr
and Welch (H-W) diagnostic to detect the model. This method calculates a test statistic

(based on the Cramér-von Mises test statistic) to accept or reject the null hypothesis that a
Markov chain is from a stationary distribution. The results show that the chain passed the
H-W diagnostic.

8 Package “rjags” is a vetted package that provides an interface from R to the JAGS library for
Bayesian data analysis and has been cited by many studies (e.g., Coblentz, Rosenblatt, and Novak,
2017; Ramírez-Hassan and Montoya-Blandón, 2020). JAGS uses Markov Chain Monte Carlo
(MCMC) to generate a sequence of dependent samples from the posterior distribution of the
parameters.

109