Page 116 - 臺大管理論叢第32卷第1期

Internet Celebrity Economy: Exploring the Value of Viewers’ Comment Features and Live Streamers’
Marketing Strategies in Forecasting Revenue

where d is a P × 1 (P = 5) matrix, which are dummy variables that denote the presence of
i
streamer heterogeneity (i.e., gender, physical attractiveness, and hosting style). Γ is rep-
resented by a k × P matrix that measures the relationship between viewers’ gift-sending
behavior structure and streamer heterogenity. Ω is the positive definite covariance matrix

of k × k.
The model allows for heterogeneity in the mean vector and covariance matrix of the
normal distribution and thus reflects heterogeneous behaviors in terms of levels (β ) and
i
2
variability (σ ). The HB model with normal distribution assumes that the mean vector β is
i
normally distributed across viewers and employed an inverted gamma distribution for σ ,
2
-2
that is, σ ~Gamma(a , a ).
1
0
The core of the HB model accounts for the uncertainty of the sample (Chiang, Chib,
and Narasimhan, 1999). Accounting for uncertainty is critical whenever data limitations
exist that lead to imprecise inferences about any aspects of behavior (Allenby et al., 2005).
HB models are a combination of two things: (1) a model written in a hierarchical form
that is (2) estimated using Bayesian methods (Allenby et al., 2005). We can analyze the

marketing data using one model for within-unit analysis and another model for across-
unit analysis (Allenby et al., 2005). The within-unit model could be used to describe
the viewers' behaviors (e.g., viewers' commenting or gift sending) over time, while the
across-unit analysis could be used to describe the heterogeneity (e.g., the streamer's
outward beauty) of the units (Jen and Chen, 2007; Rossi, Gilula, and Allenby, 2001).

The sub-models are combined to form the hierarchical model, and Bayes theorem is used
to integrate the pieces and account for all the uncertainty that is present (Allenby et al.,
2005). Furthermore, with the development of computational methods, Markov Chain

Monte Carlo (MCMC) method replaces the past complex analysis required to implement
Bayes' theorem (Haugh, 2017).

5. Estimation and Results

This section reports the estimation results for 10 live-streaming data and investigates
the performance of the proposed HB model. Specifically, we use the HB model to examine
the effects of comment features and the behaviors of streamers on the gift-sending

behavior of viewers (see Table 4) and the cross-level effects of streamer heterogeneity (see

108