Travel reviews mirror institutional trust: visitors from high-corruption home countries write less positive reviews, whereas those from low-corruption countries show stronger emotional reactions to surprises — a pattern that replicates across cities, implying sentiment models and platform rankings risk systematic bias if they ignore origin-country context.

Main Finding

Tourists’ home-country perceived corruption systematically shapes the emotional valence expressed in online hotel reviews: higher home-country corruption is associated with lower positive sentiment overall, while tourists from low-corruption countries react more strongly (both positively and negatively) to discrepancies between experience and expectations. These patterns replicate in a robustness sample from a different city.

Key Points

• The study integrates expectancy–disconfirmation and institutional trust theories to explain cross-national differences in sentiment expression in service evaluations.
• Primary evidence comes from 56,260 Booking.com hotel reviews in New York City; a robustness check uses 242,541 Booking.com reviews from Dubai.
• Textual emotional valence (sentiment) is the outcome; a country-level perceived-corruption indicator is included as a contextual moderator of experience–expectation effects.
• Main empirical results:
  • Higher perceived corruption in tourists’ home countries → lower positive sentiment in reviews.
  • Tourists from low-corruption countries exhibit stronger emotional responses (amplified positive or negative sentiment) to experience–expectation discrepancies.
• Robustness analyses across cities reproduce the broader directional pattern, suggesting generalizability across settings.

Data & Methods

• Data sources: Booking.com textual hotel reviews (56,260 reviews — New York City; 242,541 reviews — Dubai for robustness).
• Outcome measure: sentiment analysis of review text to derive emotional valence.
• Key contextual variable: country-level perceived corruption (used as moderator of the relationship between experience–expectation disconfirmation and sentiment).
• Analytical approach: statistical moderation tests examining how corruption at the home-country level conditions the link between disconfirmation (experience minus expectation) and textual sentiment; robustness checks across an independent city sample to assess external validity.
• (Note: the summary reports methods at the level described in the study; specific model specifications, controls, and estimation details are reported in the original paper.)

Implications for AI Economics

• Model specification and feature design:
  • Macro-institutional variables (e.g., perceived corruption) are meaningful predictors of user-generated textual signals and should be considered as features in econometric and machine-learning models that use review text.
  • Sentiment models and downstream demand/quality estimators trained on pooled cross-country data may be biased if they ignore institutional heterogeneity.
• Platform algorithms and benchmarking:
  • Review aggregation, ranking, and recommendation algorithms should account for country-origin effects to avoid systematic misinterpretation of sentiment (e.g., underweighting reviews from high-corruption-origin users that may be less positively worded).
  • Cross-country comparability of reputation scores may require normalization or context-aware calibration.
• Market inference and policy:
  • Marketers and platform analysts using sentiment as a proxy for satisfaction or willingness-to-pay should adjust for institutional background to improve targeting, pricing, and performance measurement.
  • Policymakers and platform regulators interpreting platform metrics across national user bases should be cautious about raw sentiment comparisons.
• Fairness and generalizability:
  • Ignoring macro-institutional conditioning risks algorithmic bias against user groups defined by home-country characteristics; incorporating such contextual factors can improve fairness and predictive accuracy.
• Research directions for AI economics:
  • Incorporate institutional-context features into demand-forecasting and recommendation systems; evaluate how doing so affects revenue optimization and welfare.
  • Study whether and how platform design (e.g., prompts, review templates) can mitigate institutionally driven differences in emotional expression.
  • Investigate causal pathways and whether institutional effects operate via expectations, trust, expression norms, or posting thresholds—important for structural models of review-based signaling.