Who Pays the Hidden Fee? Auditing Shrouded-Price Burden Under Spatial Competition

Many online prices come in two parts: a headline price shown while browsing, and fees or taxes that appear only at checkout. People respond much less to the second part. In eBay data, buyers reacted to sales tax only about one-third as strongly as to the posted price (Consumer behavior in online shopping is affected by sales tax). Sellers therefore have a reason to move money from the visible price into the hidden fee. Competition does not obviously stop this, because competition happens mostly on the number customers compare. This proposal asks who ends up paying those hidden fees. It also asks why standard audits would get the answer wrong.

The core idea fits in two sentences. A simple model of competing sellers predicts that hidden-fee burden piles up on customers who are both inattentive and "captive," meaning they have no nearby alternative seller. Captivity is tied to geography and age structure, and inattention can vary within broad census race categories. As a result, an audit that compares crude averages across four coarse groups can misattribute or miss the burden. The proposal builds the model, derives the pattern that separates hidden-fee extraction from ordinary cost differences, and tests on semi-synthetic and real audit data whether standardized, fine-grained auditing recovers the true burden where the usual tools do not.

Core Claim

The claim is that the share of the final bill collected as a late-revealed fee is a diagnostic quantity, separate from the total price. If customers fully notice fees, splitting a price into base plus fee changes nothing. Only the total matters, and a genuine cost difference (rural delivery is expensive) raises the total without changing the split. If customers under-weight fees, sellers move money into the fee wherever competition for that customer is weak. The fee share then rises with captivity and with the local share of low-attention customers. This gives a testable signature that cost-based pricing cannot produce.

Two measurement problems sit on top of that signature, and current tools handle neither. The first is composition. Captive markets are disproportionately older and rural, so crude regional comparisons confuse age structure with captivity. This is the same "aging paradox" that direct age-standardization resolves in epidemiology, where raw digestive-disease cases rose over 74% while standardized rates fell (Wang et al.). The second is subgroup masking. Within-group variation in model performance can exceed between-group variation (Coarse race data conceals disparities in clinical risk score performance). Constraints on coarse groups also leave intersectional subgroups up to 3.5 times more unfair (Preventing Fairness Gerrymandering). We expect the same masking for fee burden.

Model

We use a Hotelling duopoly. In this standard model, two firms sit at the ends of a line, and customers at position $x \in [0,1]$ pay a "travel" cost $t$ per unit of distance to buy from the farther firm. The cost can stand for physical distance, delivery time, or brand mismatch. Each firm $i$ posts a base price $p_i$ and a checkout fee $f_i$. A customer of attention type $\theta \in (0,1]$ chooses using perceived utility

$$u_i(x,\theta) = v - p_i - \theta f_i - t\,|x - x_i|,$$

but actually pays $p_i + f_i$. The welfare loss from inattention is therefore $(1-\theta) f_i$. Fees are disciplined by checkout abandonment, which rises with the fee revealed at the final screen; this is the drop-off margin Einav et al. measure. The population is a mixture over $\theta$ whose composition varies along the line. Firms may also differ in how much protected niche they hold, which extends the market with endpoints and an outside option used in the coopetition models.

Existing spatial models in the file share one assumption: they all set $\theta = 1$. This holds for the federated-learning competition papers (Coopetition in Heterogeneous Cross-Silo Federated Learning, Federated Learning in Competitive EV Charging Market) and for selective data sharing on a Hotelling line (Pareto-Improving Data-Sharing✱). We take the line structure from them and relax that assumption.

Hidden-fee competition with inattentive consumers has been studied in behavioral industrial organization outside this file, so the bare "salience in Hotelling" result is not the contribution. The contribution is the auditing consequence. The model predicts where along the captivity–attention space the burden lands. That prediction becomes a stratification for standardization and a search target for the subgroup auditor.

Research Program

The first component is analytical. We will characterize the equilibrium fee share as a function of captivity and local attention mix. We will show conditions under which the fee share is invariant to cost shocks but rises with captivity. That contrast is the identification argument. We calibrate $\theta \approx 1/3$ for tax-like fees from Einav et al. We use the personalized-disclosure experiment as evidence that $\theta$ is movable: switching rose from 17% to 28% when costs were made explicit (Comparison Friction). This motivates a disclosure counterfactual.

The second component is a measurement pipeline. It estimates a standardized hidden-fee burden: the directly standardized average of $(1-\hat\theta)f$ or of the fee share, using strata defined by age band crossed with a captivity index. The weights come from a fixed reference population. Standardization reuses the epidemiological method directly, with captivity as the extra stratum that standard age-only weights omit. The file flags that omission as the "latent covariance" failure.

Demographics are usually unobserved, so granular group membership is imputed. We extend BIRDiE (Estimating Racial Disparities When Race is Not Observed) from four census categories to six Asian subgroups. The extension uses the Wikidata name tables (Enabling disaggregation of Asian American subgroups), which so far have been used only with BISG-style weighting. BISG-style weighting is exactly what McCartan et al. show biases disparities toward zero.

Finally, a Kearns-style auditor searches over subgroups defined by imputed granular group, age, and captivity for the subgroup with the largest standardized burden. This adapts the method to a continuous economic outcome rather than classification error.

Baselines

The naive reference point is a crude average fee burden per coarse census group with no standardization. It corresponds to how disparities in e-commerce behavior are usually reported with coarse race buckets (Evaluating Amazon effects).

The regression baseline is ordinary least squares on fee burden with additive age, income, race, and state dummies, following Berke et al. It represents the standard econometric control strategy.

The strongest existing estimator of disparity under unobserved race is BIRDiE at coarse granularity. On North Carolina voter records it estimated a 48.5-point Black–White registration gap against a true 54.6 points. Standard BISG estimated 24.1 points (McCartan et al.). BISG weighting with the Wikidata subgroup tables is also run, as the current granular practice.

The strongest subgroup-search baseline is the Kearns et al. auditor applied to unstandardized fee burden over the same attributes. The only number reported for it in the file is the 3.5× intersectional gap on Communities and Crime. That number concerns classification fairness, not prices. No reported result was found on Lacuna for any method on fee or shrouded-price burden.

All methods receive the same semi-synthetic populations, the same audit data, the same imputation inputs, and the same hyperparameter search budget. The auditor class and iteration budget are shared between the plain and standardized versions.

Two ablations isolate the key ingredient: standardization by age alone without the captivity stratum, and granular imputation without standardization. Beating the baselines means two things in semi-synthetic data, where truth is known: lower absolute error in recovered subgroup burden, and correct identification of the most-burdened subgroup more often. On real audit data, it means the fee-share signature appearing where the model predicts and not where only costs differ.

Why not just…?

Why not just run a flexible regression instead of standardization? A fully saturated regression on age-by-captivity cells is mathematically close to direct standardization, so the estimator itself is not the point. What matters is choosing captivity as a stratum, and that choice comes from the model. The real threat is additive dummies, which cannot represent the attention-times-captivity interaction the model predicts. The OLS baseline and the age-only ablation test this. If a saturated regression matches our pipeline, we will report the gain as coming from the stratification and not from the standardization machinery.

Why not just point the Kearns auditor at fee burden and skip the economic model? This is the most serious alternative. An unguided auditor will find the largest fee gaps, but in real data those gaps are dominated by genuine cost differences such as rural delivery. It cannot tell exploitation from cost pass-through. The model contributes the fee-share signature: a split that moves with captivity while totals move with cost. The semi-synthetic experiment injects cost-driven and attention-driven disparities separately and checks whether each auditor flags the right one. If the plain auditor on fee share, rather than on fee level, does as well, the model's remaining value is only in telling us to use the share. We would say so.

Why not just measure attention per subgroup directly with checkout experiments? That would be ideal. The Medicare letter experiment shows attention can be measured and shifted. But individual checkout logs with granular demographics are not available to outside auditors. Attention alone also does not give equilibrium burden, which depends on captivity. Our disclosure counterfactual in the model plays the role of the Kling et al. intervention.

Proposed Experiments

The semi-synthetic experiment builds populations from the North Carolina voter file, which records real self-reported race, age, and geocoded address. Within Asian voters, granular subgroup labels are sampled from Wikidata surname posteriors and treated as ground truth. Captivity is computed from the geographic density of competing sellers or fulfillment points. Purchases and fees are then generated from the solved equilibrium under two disparity sources: attention varying across granular subgroups, and costs varying across geography. We vary how strongly attention correlates with age and location.

The decisive check is whether coarse crude averages and coarse BIRDiE understate or reverse the burden for specific subgroups, and whether the standardized, granular, model-guided auditor recovers it. A further test is whether the cost-only injection produces no fee-share signal while the attention injection does.

The real-data experiment extends the distributed price audit of Detecting price and search discrimination on the internet, which measured displayed prices from many vantage points. Our version carries automated sessions through to the final checkout screen for delivery-fee-bearing products across many zip codes. It records base price, fees, and taxes. The fee share is regressed on captivity, holding total cost proxies fixed. The main prediction is a positive captivity slope on fee share and a weaker or zero slope on the base price. Zip-level age structure is then standardized out to test whether crude rural-versus-urban fee gaps survive.

Individual-level subgroup claims cannot be made from posted-price audits, which vary by location and not by person. They are restricted to the semi-synthetic arm. The crowdsourced Amazon histories serve only as a check that inferred captivity correlates with observed online substitution.