Effects of Multi-Aspect Online Reviews with Unobserved Confounders: Estimation and Implication

Online review systems are the primary means through which many businesses seek to build the brand and spread their messages. Prior research studying the effects of online reviews has been mainly focused on a single numerical cause, e.g., ratings or sentiment scores. We argue that such notions of causes entail three key limitations: they solely consider the effects of single numerical causes and ignore different effects of multiple aspects -- e.g., Food, Service -- embedded in the textual reviews; they assume the absence of hidden confounders in observational studies, e.g., consumers' personal preferences; and they overlook the indirect effects of numerical causes that can potentially cancel out the effect of textual reviews on business revenue. We thereby propose an alternative perspective to this single-cause-based effect estimation of online reviews: in the presence of hidden confounders, we consider multi-aspect textual reviews, particularly, their total effects on business revenue and direct effects with the numerical cause -- ratings -- being the mediator. We draw on recent advances in machine learning and causal inference to together estimate the hidden confounders and causal effects. We present empirical evaluations using real-world examples to discuss the importance and implications of differentiating the multi-aspect effects in strategizing business operations.