Studies of emotion-driven online shopping report very strong effects. One survey finds that feeling emotionally "in sync" with a live-stream host predicts positive arousal with a standardized coefficient of 0.84. Another finds that enjoyment is the strongest driver of the urge to buy. Almost all of these numbers come from one-time questionnaires of 100 to 200 young respondents in a single country, and often a single city. Separately, cross-market recommender systems learn from tens of millions of real Amazon interactions across eight or more countries, but they contain no emotional variables at all. Nobody knows whether the emotional effects in the surveys are real at scale, whether they have the same size in Germany as in Japan or Mexico, or how much data a new market needs before its local estimate beats simply borrowing the global one.
This note proposes to answer those questions directly. We use a large multi-country e-commerce log as a known "population" and repeatedly draw survey-sized samples from each market. That shows how often small, single-market studies overstate an emotional effect or get its sign wrong. We then test a simple remedy: estimate the emotional effect in each market by partial pooling. Each market's estimate is pulled toward a global estimate, and the pull is weaker the more local data that market has. We compare this remedy against the strongest cross-market recommenders. Finally, we use the measured sample-size thresholds to design a properly powered multi-country replication of the survey findings, linked to real purchase and return records. The replication shares only summary statistics between countries, the same approach used in federated brain-imaging studies.
Core Claim
Our claim has two parts. The first is that emotional-trigger effects estimated from roughly 150 users in one market are, when filtered by statistical significance, systematically inflated. This is the "winner's curse": in a small study, only the estimates that noise happened to push upward clear the significance bar. We expect these estimates to vary across markets more than their reported standard errors suggest. The second is that the market-specific part of the emotional effect is real but small. It should therefore be learned with an explicit amount of shrinkage toward the global effect, set by how much data each market has, and it should not be learned as a free per-market parameter. If this holds, it explains a puzzle already in the literature. Cross-Market Product Recommendation found that its meta-learned model (FOREC) lost its advantage to the simpler NMF++ when target markets were very small. We predict a measurable crossover sample size below which market-specific adaptation of emotion-related signals causes negative transfer, meaning it makes predictions worse than borrowing from other markets.
Where This Sits
The survey literature motivates the work, and it is also the object we audit. Quantitative Research on the Impact Mechanism of Emotional Marketing in Brand Live Streaming fits regressions on 186 Chinese respondents. It reports a "feel-good" path in which emotional resonance raises positive arousal, and a "safety" path in which emotionally good interaction raises trust. Study on Impulse Shopping Intention of Consumers in Live Broadcast Rooms finds, on 171 respondents aged 18–23, that emotional attachment acts entirely through perceived value. Now or Never studies 202 Gen Z shoppers in Hanoi and finds that enjoyment drives the urge to buy while visual design does not. Investigating the Impacts of Recommendation Agents on Impulsive Purchase Behaviour finds, on 157 respondents, that satisfaction with recommendations drives impulse purchases. All four rely on self-report, measure the predictor and the outcome in the same questionnaire, use one region, and include no power analysis. We do not extend them with another survey of the same kind. We instead measure how far studies of their size and design can be trusted.
The small-sample critique comes from neuroimaging. A functional connectome phenotyping dataset including cognitive state and personality measures documents how small cohorts produced contradictory brain–behavior findings. Federated analysis in COINSTAC pooled about 2,000 adolescents from India and Europe by exchanging only local summary statistics, and showed the result was identical to central pooling. That paper supplies the method for our replication phase. We borrow its logic because path models of the kind these surveys use can be fitted exactly from each site's covariance matrix, so no raw consumer data needs to cross a border.
The cross-market recommender literature supplies both the data and the main competitors. FOREC, from the XMarket paper, meta-learns a shared model and fine-tunes a small head for each market. Market-Aware Models for Efficient Cross-Market Recommendation replaces meta-learning with one learned vector per market that multiplies item embeddings. Dual prototype attentive graph network for cross-market recommendation (DGRE) separates user behavior patterns shared across markets from market-specific item prototypes. None of the three uses emotional signals, and none reports how its per-market adjustment behaves as the target market shrinks toward survey scale. DGRE's two findings, that behavior clusters predict better than nationality but market-specific item prototypes give the largest gain, are exactly the tension that partial pooling is meant to resolve.
Research Program
The first component is an affect measure for items. We use item metadata and review text from XMarket, available through its benchmark paper, Cross-Market Product Recommendation. For each interaction we compute the emotional intensity and hedonic framing of the product's content and of the reviews posted before that interaction. This stands in for the stimulus-side "emotional resonance" of the survey models. We score text with a multilingual affect encoder. Enhancing Multimodal Affective Analysis with Learned Live Comment Features constrains this step: its gains were much smaller on English data than on Chinese data because of a shift in genre. We therefore test whether the scorer measures the same thing across languages before comparing any market to another. Concretely, we check scorer agreement on items matched across markets by ASIN, the shared Amazon product identifier. Without this check, an apparent cultural difference could be a measurement artifact.
We use two outcomes. The first is adoption, meaning whether an exposed user interacts. The second is a regret proxy: a low rating given after an affect-heavy exposure, relative to the same product's rating in other markets. The regret proxy responds to the gap highlighted in ‘Conformity’ Or ‘Consumption Trap’?, which reports roughly 30% post-purchase regret, and to the 28–32% return rates cited in the anchor-factors study. We are explicit that XMarket records reviewed purchases, not impulse purchases. This phase therefore tests the statistics of affect effects across markets, not impulsivity as such.
The second component is the subsampling audit. In each of the eight Electronics markets, we first estimate the affect effect on the full market to serve as the reference value. We then draw thousands of subsamples of users at N between 100 and 5,000, a range that includes the survey sizes. For each N we record four quantities. The first is the exaggeration ratio among estimates that pass p < 0.05. The second is the rate at which the sign is wrong. The third is how often the survey-typical design of one market and one young cohort ranks markets correctly. The fourth is how often a result from one market "replicates" in another. The audit produces empirical curves of trustworthiness against sample size, grounded in real behavioral data from several countries rather than in assumed effect distributions.
The third component is a partially pooled affect model. We attach the affect score to a recommender backbone and give each market $m$ its own affect coefficient $\beta_m$. These coefficients are drawn around a global value $\beta_0$ with between-market spread $\tau$. This yields the familiar shrinkage estimate
$$\hat\beta_m = w_m \tilde\beta_m + (1-w_m)\hat\beta_0, \qquad w_m = \frac{\tau^2}{\tau^2 + \sigma_m^2},$$
where $\tilde\beta_m$ is the estimate from market $m$ alone and $\sigma_m^2$ is its sampling variance, which falls roughly as $1/n_m$ with the market's number of users $n_m$. The form follows from a normal hierarchical model. It predicts a crossover near $n_m \approx \sigma^2/\tau^2$, where $\sigma^2$ is the per-user noise variance, i.e. the point at which local and global evidence carry equal weight. We fit $\tau$ by empirical Bayes, which estimates it from the spread of the market estimates themselves. The prediction to test is that the measured crossover in recommendation quality matches this ratio.
The fourth component is a powered, federated replication. We use the phase-one curves to set per-country sample sizes. We then field the dual-pathway live-stream questionnaire from the brand live-streaming study in at least four countries, linked to consented records of purchases and returns within 30 days. Each site shares only covariance matrices. We fit a multi-group structural model and test whether the measures mean the same thing in each country before comparing paths. The model predicts actual returns, not just stated intention.
Baselines
The naive reference points are a single-market model trained on target-market data only and a pooled model that ignores market identity. For the survey arm, the naive reference is the published coefficients themselves. These include emotional resonance to arousal of 0.838 (R² = 0.702) and interaction quality to trust of 0.843, from Quantitative Research on the Impact Mechanism. The replication "beats" these if it gives precise estimates that differ from them beyond their confidence intervals, or if it confirms them with power above 0.9 in every country.
The first strong recommender competitor is the Market-Aware model, MA-NMF, from Market-Aware Models for Efficient Cross-Market Recommendation. It is the strongest efficient baseline. It beat meta-learning in 85% of source–target pairs and in 7 of 8 markets under global training, using about 15% of the training time. Its learned market vector is the closest existing analogue of our per-market coefficient, but it has no explicit shrinkage. The second strong competitor is FOREC, which improved nDCG@10 over NMF by 24% on average and by up to 50% Cross-Market Product Recommendation. The third is DGRE, which was best in 34 of 39 scenarios Dual prototype attentive graph network. We also report NMF++, because it is the model that overtook FOREC on very small markets. The file gives no absolute XMarket nDCG@10 values for these models, and we found no reported result on Lacuna for any affect-augmented cross-market recommender on XMarket. We will therefore reproduce all baselines ourselves on identical splits. We use the XMRec ensemble's NDCG@10 of 0.7393 on market t1, from A Practical Two-stage Ranking Framework, only to check that our backbone is competitive on triple-only data. That dataset has no text, so affect cannot enter it.
For fairness, every model gets the same affect features, the same backbone embedding size, the same target-market subsample at each N, and the same hyperparameter search budget per market. The key ablation compares MA-NMF with the affect feature concatenated against our model. Both see the same inputs, and they differ only in whether market adaptation of the affect effect is explicitly shrunk. Our method wins if it improves nDCG@10 and calibration of the regret proxy over MA-NMF-plus-affect below the crossover N, matches it above, and places the crossover where the shrinkage formula predicts.
Why not just…?
Why not just add a sentiment feature to the Market-Aware model? This is the real threat. The per-market multiplicative vector in MA-NMF already gives each market a small amount of freedom, and with regularization it might pool implicitly. The head-to-head ablation above settles it. If MA-NMF-plus-affect matches us at every N, the conclusion becomes that explicit shrinkage is unnecessary, and the subsampling audit remains the main contribution. We would report that plainly.
Why not just run a much larger survey in more countries? Size alone does not remove the bias that comes from measuring the predictor and the outcome in the same questionnaire. Size also does not say how large each country's sample must be. That depends on the between-market spread $\tau$, which only the multi-market log audit can estimate before fielding. Phase one exists to make phase four efficient rather than a guess.
Why not just correct the published coefficients analytically for the winner's curse? Analytic corrections require assuming a distribution of true effects and a selection process. Our audit measures these directly on real multi-country data. The correction also cannot capture sign flips across markets, which is the part of the problem that comes from limited geographic scope.
Why not just use FOREC-style fine-tuning per market and skip affect? FOREC's own authors saw it lose to NMF++ on tiny markets, which is the failure we target. Dropping the affect signal also would not address whether emotional triggers are worth modeling at all. That question is settled by comparing each model with and without affect, at every sample size.
Decisive Checks
We see four results that would decide the question. First, if subsamples of about 150 users in XMarket markets show exaggeration ratios near 1 and cross-market sign agreement above 90%, the winner's-curse concern is overstated and the survey coefficients deserve more trust. Second, if the measured crossover in recommendation quality departs substantially from $\sigma^2/\tau^2$, the hierarchical explanation of FOREC's small-market failure is wrong. Third, if the affect scorer disagrees with itself on items matched across markets, cross-market differences in affect effects cannot be interpreted, and we would restrict claims to markets that share a language. Fourth, the federated replication either reproduces a strong resonance-to-arousal path that predicts actual 30-day returns in every country, or it shows the published coefficient of 0.84 to be a product of small, single-city, self-reported designs.
References
- Cross-Market Product Recommendation
- Quantitative Research on the Impact Mechanism of Emotional Marketing in Brand Live Streaming on Consumers' Impulsive Purchase Intentions
- Study on Impulse Shopping Intention of Consumers in Live Broadcast Rooms based on Anchor Factors and Perceived Value
- Now or Never: Exploring Factors Influencing Online Impulsive Buying Behavior on Fashion Brand Websites among Generation Z Consumers
- Investigating the Impacts of Recommendation Agents on Impulsive Purchase Behaviour
- A functional connectome phenotyping dataset including cognitive state and personality measures
- Federated Analysis in COINSTAC Reveals Functional Network Connectivity and Spectral Links to Smoking and Alcohol Consumption in Nearly 2,000 Adolescent Brains
- Market-Aware Models for Efficient Cross-Market Recommendation
- Dual Prototype Attentive Graph Network for Cross-Market Recommendation
- Enhancing Multimodal Affective Analysis with Learned Live Comment Features
- ‘Conformity’ Or ‘Consumption Trap’? -- Exploring The Relationship Between Personalized Recommendation Marketing Mode and Consumers' Impulsive Consumption in Tik Tok Mall
- A Practical Two-stage Ranking Framework for Cross-market Recommendation