Modeling Longitudinal E-commerce Sales Responses with Attribute-Level Consumer Review Scores Using Text Mining

Online product reviews in review web sites, blogs, and social networks have made significant impact on consumers’ online purchasing behaviors (Lee and Youn 2009). A growing stream of research has begun to investigate how online product reviews can influence product sales (e.g., Liu 2006, Godes and Mayzlin 2009), yet the results so far are mixed. Two major problems can be identified in existing research. First, many researchers conduct their analyses in a cross-sectional design, which cannot control for unobserved differences in product quality (Duan and Whinston 2008). The other problem is that most studies only used the average ratings of reviews in their analysis (e.g., 3 or 4 stars) without taking into account the actual content of the reviews, e.g., what the reviewers say about the performance of various product attributes. This simplification approach may be limiting as scholars have shown that consumers in fact respond to the actual content of reviews, not merely the summary statistics (Chevalier and Mayzlin 2006).An accurate understanding of consumers’ reviews about product attributes has important implications for e-commerce. For instance, firms can pinpoint the specific product attributes that can influence sales outcomes so as to act accordingly (e.g., remedy attributes that receive lousy reviews). In addition, by isolating the longitudinal effects of product reviews on sales, firms can appreciate better the lingering effects of consumer reviews and adopt a systematic approach in managing the impact of such reviews.Using a new text mining tool, the study intends to fill a void in the online product reviews literature by modeling the longitudinal effect of attribute-level consumer review scores on online sales responses. Specifically, we shall extend and make improvements on extant product review mining tools (e.g, Hu and Liu 2004) to fit the Chinese Internet market. Our study will consist of two parts. In part I, we shall apply a newly developed text mining tool to extract consumers’ attribute-level product rating scores for selected product categories in a major Chinese e-tailer. Second, we shall use econometric modeling to relate these rating scores to lagged sales variables. The performance of our model will also be compared to those of benchmark models.