Auditing Geographic Bias in Housing Ad Delivery: A Conceptual Protocol for Spatial Exposure Analysis

1. Introduction and research question

Housing advertisements increasingly circulate through digital systems that can direct messages toward selected consumer segments. In housing markets, the practical concern is that the same campaign may receive unequal geographic exposure across places that are similar on observable housing and market characteristics. The research question of this protocol is therefore narrow and operational: after adjustment for measured place-based covariates, does housing-ad delivery vary systematically across geographic units, and can that residual variation be distinguished from observable market structure and campaign design? [1] [2]

This manuscript does not claim empirical findings. It proposes an audit design for a bounded problem: geographic bias in delivery, not bias in prices, click-through rates, or housing values themselves. The design is informed by prior work on targeted advertising, where delivery responds to consumer search and platform segmentation, and by real-estate studies that show the value of fine-grained spatial covariates in housing analysis. A controlled audit is appropriate because it permits explicit comparators and falsifiable expectations without requiring access to platform source code. [1] [3] [2]

2. Evidence scope and conceptual foundations

The evidence base for this protocol is intentionally narrow and mixed. From advertising research, a theoretical model of targeted advertising shows that ad intensity can vary across consumer segments as firms respond to information conditions and strategic competition. A randomized experiment on pharmaceutical advertisements shows that controlled ad variants can be compared under explicit treatment and control conditions to estimate the effect of disclosure content on perceived information and stated intentions. These studies do not address housing ads directly, and they do not by themselves establish a method for spatial auditing. They support a more limited methodological premise used here: ad exposure and ad responses can be treated as analyzable outcomes under clearly specified comparators. [1] [3]

The housing literature contributes three domain-specific lessons. First, inter-city housing relationships are spatially heterogeneous and may display hierarchical ripple patterns rather than uniform transmission across places. Second, fine-scale spatial feature construction is feasible when communities, coordinates, accessibility measures, environmental attributes, and population context are assembled into a common dataset. Third, valuation research often defines geographic exposure explicitly for some domains, especially climate and hazard risk, using mapped exposure or distance-based proxies. Taken together, these lessons support an audit design built on interpretable spatial covariates rather than opaque latent representations alone. [4] [2] [5]

The protocol therefore treats housing ad delivery as a spatial allocation problem with observable covariates and residual unexplained variation. The audit does not assume that geographic differences are automatically discriminatory; instead, it tests whether differences remain after controlling for expected correlates of relevance and inventory. That distinction matters because housing markets are geographically uneven, and a valid audit must separate ordinary market segmentation from delivery distortions. [4] [2]

3. Audit unit, variables, and operational definitions

The unit of analysis is a geographic cell, defined as either a residential community, census tract, postal zone, or regular grid cell, depending on the granularity available in the audit dataset. The preferred unit is the smallest level for which both housing-market covariates and ad-delivery measures can be linked reliably. For each unit g and campaign c, the core outcome is realized exposure E_gc, defined as the number of ad impressions, unique users reached, or ad exposures per eligible user in that unit during the observation window. [2]

A corresponding expected exposure, denoted mu_gc, is estimated on the same outcome scale as E_gc. Covariates may include housing density, a price-level proxy, tenure mix, population density, accessibility to major services or transit, listing density, and hazard variables when they plausibly affect legitimate campaign relevance. The covariate set should be pre-specified and restricted to variables that could reasonably influence lawful geographic targeting. The primary residual is r_gc = E_gc - mu_gc. The primary standardized residual is the Pearson residual z_gc = (E_gc - mu_gc)/sqrt(max(mu_gc, 1)), which remains defined when fitted exposure is small. If the outcome is modeled with a log link, mu_gc refers to the back-transformed fitted mean on the count or rate scale, not the linear predictor. Positive residuals indicate delivery above expectation; negative residuals indicate under-delivery. [2] [5]

The key explanatory classes are defined as follows. Geographic position captures absolute and relative location, including urban core status, distance to major transport nodes, and adjacency to high-density housing clusters. Housing structure captures unit size, building age, tenure composition, and price segment. Market context captures local competition, inventory turnover, and observed listing density if available. Platform outcome variables include impression count, reach, frequency, click-through, and cost metrics when accessible. If only partial ad metrics are available, the audit can still proceed with exposure counts and relative share of delivery across places. [2] [5]

4. Proposed methodology: algorithm and decision sequence

The audit proceeds in six steps. Step 1 defines the spatial frame and observation window. Step 2 maps each ad impression or reach event to a geographic unit using geocoded user location, delivery zip code, or market-area assignment. Step 3 builds an expected-delivery model using only pre-specified covariates. Step 4 computes residual delivery by geographic unit. Step 5 tests whether residuals are spatially patterned, concentrated, or associated with protected or policy-relevant place attributes. Step 6 performs falsification checks and sensitivity analyses. [2] [5]

An exact operational rule is required for auditability. Default rule: audit each campaign at three nested spatial scales in a fixed order—grid cell, postal zone, and tract or community—and treat the tract or community result as primary unless fewer than 30 analyzable units remain, in which case move to the next coarser scale. A unit is analyzable only if the eligibility denominator is observed and the fitted expected exposure mu_gc is at least 1. For a prespecified place-characteristic score S_g, rank analyzable units and define the top and bottom quartiles; ties at quartile boundaries are broken at random with a fixed recorded seed. Estimate one-to-one nearest-neighbor matching without replacement between top- and bottom-quartile units using Mahalanobis distance on log population density, log listing density, and log housing-stock count, with exact matching on campaign and region where feasible. Accept the matched sample only if the absolute standardized mean difference for each matching variable is at most 0.1. Let Delta_c be the difference between the median Pearson residual in the matched top-quartile units and the matched bottom-quartile units. Flag the campaign for geographic review when Delta_c exceeds 0.5 in absolute value and a permutation test that shuffles quartile labels within matched pairs yields a two-sided p-value of 0.05 or less. If more than one spatial scale is analyzable, the primary determination is made at the primary scale and the remaining scales are reported as sensitivity analyses. [2]

The expected-delivery model should be aligned with the residual definition. The default specification is a hierarchical Poisson log-link model for counts, E_gc ~ Poisson(mu_gc), with log(mu_gc) = beta_0 + beta_1 X_1gc + ... + beta_p X_pgc + alpha_region + gamma_campaign + log(N_gc), where N_gc is the eligible-user or opportunity denominator entered as an offset when available. Fitted values mu_gc are then used directly in r_gc and z_gc. If overdispersion is substantial under a preregistered diagnostic such as Pearson dispersion greater than 1.5, replace the Poisson model with a negative-binomial model. If more than half of analyzable units have zero exposure, fit a two-part fallback: first model Pr(E_gc > 0) with logistic regression, then model E_gc conditional on E_gc > 0 with a truncated count model, and define mu_gc as the product of the fitted probability and fitted positive count. More flexible benchmarks may be added for sensitivity analysis, but the count model remains primary because its fitted mean and residuals are directly interpretable. [2]

The comparator set is essential. The main comparator is not a synthetic absence of advertising, but a matched-place benchmark: units with similar housing stock, accessibility, and market conditions should receive similar exposure in expectation. A secondary comparator is within-campaign variation across places exposed to the same creative and budget. Where disclosure controls are possible, a randomized creative condition can be appended, using the logic of controlled ad exposure from experimental advertising research. In that design, the content remains constant while only a disclosure, location cue, or targeting label varies, allowing the investigator to separate content effects from spatial allocation effects. [3]

5. Comparator logic and outcomes

The primary outcome measure is spatial inequality in ad delivery. Suitable measures include the Gini coefficient of exposure across units, the coefficient of variation of exposure rates, and the share of impressions captured by the top decile of units. A second outcome is spatial clustering of residuals, estimated through Moran-type statistics or local cluster maps. A third outcome is directional imbalance: whether under-delivery is systematically concentrated in places with lower income, different tenure profiles, or distinct housing forms once structural market variables are controlled. [4] [2]

The protocol distinguishes between bias in delivery and legitimate targeting. If a platform systematically targets areas with higher observed housing search intensity, higher listing density, or stronger likely conversion, that pattern may be explained by market relevance rather than bias. The audit therefore tests whether delivery residuals are associated with place features that should not, in principle, determine ad visibility once the campaign goal is fixed. For example, if a housing campaign aims at broad rental awareness, a strong residual gradient by neighborhood income or spatial centrality would warrant scrutiny; if the campaign aims at a very specific local submarket, narrower concentration may be expected. [1] [2]

The housing valuation literature is relevant here because it shows how location and accessibility are ordinarily entered into models of real-estate value. That literature supports using interpretable geographic covariates rather than vague spatial descriptors. However, the goal here is different: the question is not whether place raises price, but whether place receives delivery beyond what price-related relevance would predict. This distinction reduces the risk of mistaking market geography for platform bias. [2] [5]

6. Validation design, falsification tests, and failure cases

Validation should be prospective and pre-registered. First, before any campaign is audited, fix the covariate set, spatial scales, matching variables, dispersion diagnostic, and flagging threshold; tuning is then limited to a stated sensitivity grid around the default threshold for |Delta_c| and the minimum fitted exposure rule. Second, assess internal predictive adequacy with blocked cross-validation across geographic units or campaign-time folds, using prediction error on the count scale because mu_gc is defined on that scale. Third, include negative-control campaigns whose formal targeting settings do not use geographic criteria, if such campaigns are available. Fourth, use placebo procedures to construct null reference distributions: permute the place-characteristic score labels within matched strata, and separately jitter unit centroids within a pre-specified buffer that does not cross the study boundary. The default buffer is one half of the median nearest-neighbor distance between unit centroids; wider and narrower buffers are reported only as sensitivity analyses. [3] [2]

A supplementary falsification design can adapt the logic of randomized advertisement comparison without overstating what that literature establishes. Hold the creative, budget, and bidding strategy fixed while varying only a neutral disclosure element or location-related label that does not change the substantive housing offer. Random assignment of that ad variant can then test whether observed differences arise from message processing rather than from the platform's spatial allocation of impressions. If the randomized variant changes post-exposure responses but leaves the spatial residual pattern materially unchanged, the evidence is more consistent with a content effect than with geographic delivery bias. If the residual pattern shifts under otherwise fixed delivery settings, that pattern warrants closer audit of the serving system. [3]

The main failure cases are predictable. Spatial proxies may be too coarse to capture actual audience geography; a neighborhood label can mask within-area heterogeneity. Observed covariates may omit variables that legitimately shape relevance, such as local rental regulations or transient demand shocks. Platform auctions may modify delivery in response to user behavior faster than the audit window can observe. Finally, ad delivery data may be affected by sparse counts, especially when campaigns are small or geographically narrow, producing unstable residuals. These conditions do not invalidate the audit, but they limit generalization and require reporting of uncertainty, missingness, and model dependence. [2] [5]

7. Discussion: relation to prior methods and boundary conditions

Compared with prior housing-price network studies, this protocol shifts the analytic target from price spillovers to ad exposure allocation. The difference is substantive: spillover studies infer how prices move across cities, whereas this audit asks whether advertising systems distribute visibility unevenly across places. The relevance of spatial spillover work lies in its demonstration that housing phenomena are geographically structured and can exhibit hierarchical core-periphery organization, a reminder that a spatial null of uniformity is implausible in housing settings. Still, the current protocol does not infer price contagion; it uses the spatial logic of the housing literature to justify place-based auditing. [4]

Compared with valuation studies, the proposed method uses a narrower outcome and a different inferential target. Hedonic work organizes explanatory variables around structural housing characteristics, neighborhood context, and accessibility, and climate-risk valuation studies often identify exposure with hazard maps or distance-based proxies. This protocol borrows that discipline of explicit covariate construction. It does not import the valuation objective itself. The question here is whether observed delivery exceeds or falls below an expected benchmark after those place variables are accounted for, not whether those variables raise or lower property value. [2] [5]

The protocol also differs from the randomized experiment on price disclosure in pharmaceutical ads. That experiment manipulated ad content and measured consumer perceptions and intended behaviors. In this manuscript, randomization is optional and auxiliary rather than central; the main target is spatial allocation. However, the experimental logic remains useful because it shows how controlled stimuli can clarify causality when the mechanism is uncertain. The boundary condition is that ad disclosure or creative variation can identify response to messaging, but not by itself explain geographic delivery bias unless paired with spatial assignment data. [3]

A final boundary condition concerns generalization. The protocol is most appropriate for markets where housing ads are served through identifiable digital channels and where location can be measured at least approximately. It is less informative when the platform obscures geolocation, when campaigns are too small for stable spatial inference, or when the market is dominated by local offline advertising. These limits are not defects of the idea; they define the domain in which a geographic audit is technically meaningful. [1] [2]

8. Conclusion

This manuscript establishes a conceptual and methodological basis for auditing geographic bias in housing ad delivery. The central proposal is to treat ad exposure as a spatially distributed outcome and to compare realized delivery with an expected benchmark built from observable housing-market and place characteristics. The protocol specifies the audit unit, variables, comparator, operational rule, and validation sequence, and it makes clear that evidence of residual geographic concentration is not equivalent to proof of discriminatory intent. [2] [5]

What remains untested is empirical implementation: the choice of platform data source, the stability of the expected-delivery model across cities, and the extent to which residual spatial patterns persist under stronger controls. Those questions require a full audit study. The present contribution is narrower but precise: it offers a falsifiable framework that can be used to determine whether housing ad delivery is spatially aligned with market structure or whether it departs from that structure in ways that deserve further scrutiny. [1] [3] [4]

References

1. Dan, Wu; Jiang, Zhao (2026). The complementary-competitive interplay of consumer information search and experience sharing: Implications for targeted advertising and firm performance. PLOS One. 21(10), e0340849. doi:10.1371/journal.pone.0340849.
2. Lin, Wenbo; Wu, Kaichuang; Zhang, Xiaolu; Gu, Xiaoling; Jiang, Ming; Wang, Jun; Han, Feng (2026). A comprehensive dataset for uncovering housing price drivers in emerging cities: Nanning as a case study. Data in Brief. 66, 112884. doi:10.1016/j.dib.2026.112884.
3. Hollin, Ilene L.; Ball, Jennifer Gerard (2022). Does price disclosure in pharmaceutical advertising result in price transparency? Evidence from a randomized experiment. Exploratory Research in Clinical and Social Pharmacy. 8, 100180. doi:10.1016/j.rcsop.2022.100180.
4. Qiu, Lu; Jiao, Yanzhe; Dong, Gege; Cui, Guangcan (2026). Finite-Length Spatiotemporal Modelling for Housing Price Network Spillovers. Entropy. 28(5), 537. doi:10.3390/e28050537.
5. Piseddu, Tommaso; Vanhuyse, Fedra (2026). Existing evidence on the impact of climate risk on real estate valuations: a systematic map. Environmental Evidence. 15(1), 8. doi:10.1186/s13750-026-00389-6.