Understanding the effects of different review features on purchase probability

ABSTRACT

The role of electronic word-of-mouth (eWOM) has been recognized by marketers and academics, but little research has examined the impact of eWOM on purchase behavior. Building on dual-process models of persuasion, this study aims to disentangle the effect of different online review features (i.e. argument quality, review valence, review helpfulness, message sidedness, source credibility and reviewer recommendation). Using product reviews and purchase data from an online retailer website, we investigate the financial impact of online product reviews on purchase decisions. The results demonstrate the persuasive power of different review features that are derived from dual-process models of information processing. Managerial implications on how advertisers and companies should design and manage online product reviews are offered.

KEYWORDS:

• eWOM
• online product reviews
• dual-process model
• heuristic-systematic model
• purchase behavior

Acknowledgments

The authors thank the IMC Medill Spiegel Digital & Database Research Center for granting access to the data for this study.

Disclosure statement

No potential conflict of interest was reported by the authors.

Notes

1. We assessed multicollinearity using GVIFs proposed by Fox and Monette (1992). GVIFs are an appropriate way to assess multicollinearity in models with categorical predictors and polynomials. When there is multicollinearity, slope estimates are unbiased but have inflated standard errors (i.e. the variance of the estimates is inflated), but standard errors are also determined by sample size, which is very large in our case.

2. The model does not overfit the data. Our logistic regression model is estimated on 413,666 observations and has 25 parameters. Perhaps, the greatest risk to overfitting comes from the dummy variables, but we have large samples (valence has 63,979, 28,253, and 322,434 values; review recommendation has 29,274, 64,170 and 320,222 values; category has 18,542, 145,620, 29,515, 112,842, 17,670, 62,341, and 27,136; seasonality has 8224, 142,476, 111,599, 108,494, and 43,873 observations). To make sure that we are not overfitting, we use five-fold cross validation by assigning a random value 1–5 to each case (Kuhn and Johnson 2013, 69–70). We then estimated our model five times, each time leaving out one part, and applied the model to the left-out part. The value of AUC was computed on the held-out values. The original AUC value was 0.6146. The five-fold CV value is 0.6138, which differs by 0.0008. We also tried 10-fold cross validation and got the same AUC value to four decimal places. Thus, there is no evidence for overfitting.

3. For a predictor having a single degree of freedom, GVIF equals VIF (equaling 1/(1 − R²) from a model predicting X_j from the other predictors in the model). Let X₁ be a block of predictors (e.g. multiple dummies for a categorical predictor), X₂ be a block of the remaining predictors in the model, and X be all predictors (bind X₁ and X₂ to product X). Let R₁ be the correlation matrix of X₁, R₂ the correlation matrix for X₂ and R be the correlation matrix for X. Then, GVIF = det(R₁)*det(R₂)/det(R). A benefit of GVIF is that it is invariant to the choice of the baseline value (e.g. negative for valence), or category (requiring six dummies).

Additional information

Notes on contributors

Su Jung Kim

Su Jung Kim is an assistant professor in the Greenlee School of Journalism and Communication at Iowa State University. She received her PhD degree in the Media, Technology, and Society (MTS) program from Northwestern University and was a post-doctoral research associate at the Medill IMC Spiegel Digital & Database Research Center at Northwestern University before joining Iowa State University. Her research centers on the use of big data in communication, advertising, and marketing. Her research interests include cross-platform media use and its effects, social media and electronic word-of-mouth (e-WOM), and mobile customer engagement. Her current projects examine how online product reviews influence the perception and behavior of consumers.

Ewa Maslowska

Ewa Maslowska is an assistant professor in the Amsterdam School of Communication Research, University of Amsterdam, where she also earned her PhD degree in persuasive communication program. She completed a post doc in the Medill IMC Spiegel Digital & Database Research Center, Northwestern University. Her research interests center around consumer behavior, advertising and digital consumer environments. She conducts experimental and data-driven research into the dynamics of consumer engagement.

Edward C. Malthouse

Edward C. Malthouse is the Theodore R and Annie Laurrie Sills professor of integrated marketing communication, and industrial engineering and management science at Northwestern University. He is the research director for the Medill IMC Spiegel Digital & Database Research Center. He was the co-editor of the Journal of Interactive Marketing between 2005 and 2011. He earned his PhD degree in computational statistics from Northwestern University and completed a postdoc at the Kellogg marketing department. His research interests center on customer engagement and experiences, digital social and Mobile Media, Big Data, customer lifetime value models, Predictive Analytics, unsupervised learning, and integrated marketing communication.