UQO Working Paper No. 5· v2.0· November 2026· 91 pages

Airbnb and Residential Rents in Quebec: Evidence and Identification Limits from a Listing-Level Cross-Section

Real Estate & Urban Economics

On SSRN

One more Airbnb listing within 500 m is associated with asking rents 0.0043 log points higher across Quebec neighbourhoods, but within city–neighbourhood cells the estimate is 0.0008 with a standard error of 0.0008: the cross-section cannot sign the within-neighbourhood relationship.

The paper links 7,907 rental listings to 3,349 Airbnb listings by distance and measures what a cross-sectional hedonic design can and cannot establish, with spatially robust standard errors, minimum detectable effects, Oster bounds, placebo exposures and a reliability check. It does not identify a causal effect: the city-level association is consistent with the sorting of listings into high-rent neighbourhoods and with measurement error in a capped scrape whose raw files are lost.

Simon-Pierre Boucher — contact@spboucher.ai

Paper PDF (revised)Read online SSRN Code & data repository

0.0043 (s.e. 0.0007)

log-point difference in monthly asking rent per additional Airbnb listing within 500 m, with city fixed effects and standard errors clustered by postal-code neighbourhood; one standard deviation of exposure (13.06 listings) corresponds to a rent about 5.62% higher. An association, not a causal effect

0.0008 [−0.0008, 0.0023]

the same coefficient within city-by-neighbourhood cells (standard error 0.0008, t = 0.93): the 95% interval contains both zero and the city-level estimate

0.0023

minimum detectable effect per listing at 80% power within cells, of the same order as, or larger than, quasi-experimental magnitudes; the design can reject neither zero nor effects of that size

0.0053

city-fixed-effects coefficient of a placebo count of listings displayed above $1,000 per night, plausibly long-stay rather than short-term rentals; it is at least as large as that of the listings priced below the threshold

δ = 0.09

Oster's δ at which the within-cell coefficient would be zero; the bias-adjusted coefficient is negative, so the within-neighbourhood estimate is not a lower bound

≈ 3×

standard errors that account for spatial dependence relative to the heteroskedasticity-robust errors usually reported

−0.155 → 0.007

mean paired R² difference between OLS and a random forest, from random folds to folds that hold out whole neighbourhoods: spatial blocking removes most, though not all, of the tree-based advantage

Abstract

Does short-term rental activity raise residential rents? This paper examines the question with a January 2026 cross-section of 3,349 Airbnb listings and 7,907 Realtor.ca rental listings in Quebec, linked by great-circle distance. With city fixed effects, one additional Airbnb listing within 500 m is associated with asking rents 0.0043 log points higher (standard error 0.0007 clustered by postal-code neighbourhood). Within city–neighbourhood cells the coefficient is 0.0008 (standard error 0.0008; 95% interval [-0.0008, 0.0023]), and the minimum detectable effect of 0.0023 per listing is of the same order as quasi-experimental magnitudes. A placebo count of listings displayed above $1,000 per night, which are plausibly not short-term rentals, reproduces the city-level association; Oster’s δ for a zero coefficient is 0.09; split-half reliability of the count is 0.959. The within-neighbourhood sign is therefore unidentified in this cross-section, and the city-level association is consistent with sorting and with measurement error.

JEL: R21, R31, R52, L83, C21

Key Results

First listing-level, distance-matched analysis for Quebec

Individual rental dwellings are matched by great-circle distance to individual Airbnb listings, giving exposure counts within 250 m, 500 m, 1 km and 2 km and hedonic rent equations with city, neighbourhood and city-by-neighbourhood fixed effects.

The identification problem is quantified rather than acknowledged

The paper reports the quantities that decide whether the within-neighbourhood estimate is informative: confidence intervals, minimum detectable effects, Oster bounds, placebo exposures and measurement reliability. The answer is that it is not informative.

Spatially robust inference for a buffer-based exposure

Heteroskedasticity-robust, cluster-robust, Conley spatial HAC and spatial block-bootstrap standard errors are compared; spatial lag and spatial error models give a total impact of 0.0045 (standard error 0.0007), the same size as the city-fixed-effects estimate.

An audit of the exposure measure

Against public Inside Airbnb snapshots, the scraped file covers 8.2% of the Montreal stock, yet the two counts are strongly correlated at the level of rental listings (0.94); split-half reliability of 0.959 rules out random thinning but not the systematic omission of unpriced and uncollected listings.

A specification curve and a blocked machine-learning benchmark

240 specifications cross radius, functional form, fixed effects, controls and sample under three standard-error types; under spatially blocked cross-validation the linear model and the tree-based learners perform about equally (0.47 against 0.46 and 0.45).

Peer review

This paper went through a multi-referee review: 6 referee reports and 83 review items. The version presented on this page is the revised manuscript. Scores are on a 0–100 scale and are a diagnostic of the manuscript, not an acceptance probability.

Score after revision

61.1/100

53.9 before review · +7.2

Review dated November 4, 2026

Review decision: claims not supported as written → reframed and revised
Scientific integrity: PASS
Reproducibility: R2partially reproduced
Maturity: Bone revision cycle from submission
Referee reports: 6
Review items: 83

Scores by dimension

Before reviewAfter revision

Originality & contribution51.2 → 52.0
Scientific rigor56.3 → 64.0
Identification & methodology52.7 → 60.0
Data & reproducibility47.5 → 58.0
Robustness & validation56.8 → 68.0
Writing71.0 → 70.0
Figures & tables57.7 → 72.0
Literature positioning52.0 → 66.0

Revised manuscript91 pages11 figures34 tables15 equations79 references35,259 words

What the review changed

Peer review found that the evidence did not support the earlier reading of the two estimates as upper and lower bounds on a causal effect. The revised paper concludes that the within-neighbourhood sign is unidentified in this cross-section and that the city-level association is consistent with sorting and measurement error; several secondary claims were withdrawn and the title was changed.

City-level and within-neighbourhood estimates read as upper and lower bounds on the causal effect → neither estimate bounds it; the within-neighbourhood sign is unidentified (0.0008, standard error 0.0008, minimum detectable effect 0.0023)
“Any Airbnb presence lowers rents within neighbourhoods” → not supported by the tests reported; withdrawn
“The association is larger at the upper quantiles” → the data do not reject equality across quantiles; withdrawn
“The machine-learning advantage disappears when neighbourhoods are unseen” → most, not all, of it disappears (mean paired R² difference from about −0.155 to 0.007)
Sample: 3,456 Airbnb and 8,303 rental listings → 3,349 and 7,907 after a bounding box and de-duplication
New diagnostics added: a placebo count of listings above $1,000 that reproduces the city-level association (0.0053), Oster δ of 0.09, split-half reliability of 0.959, and a coverage check showing the file holds 8.2% of the Montreal stock

Limitations stated in the paper

A single cross-section cannot separate an effect of Airbnb on rents from reverse causality or unobserved neighbourhood characteristics; a causal estimate requires panel data or a policy shock.
The exposure count comes from a capped, undated scrape whose raw file is lost: it excludes 1,544 listings without a displayed price, covers about 8.2% of the Montreal stock and includes listings that are plausibly monthly rentals.
Rents are asking rents on Realtor.ca, which over-represents agent-listed, newer and higher-priced units; the within-neighbourhood estimate is in practice a Montreal estimate.
Tests proposed during the review (restricting to listings far from neighbourhood boundaries, an external placebo exposure) were not run, and the first cleaning steps of the pipeline are not reproducible.

Read online (interactive reader)Revised PDF Revised branch on spbgit (review-20261103)

Data

Airbnb listings, Quebec

A capped, undated scrape of 5,000 listings collected around January 2026, of which 3,456 carried a displayed price; after a bounding box and de-duplication, 3,349 listings. The price field is bimodal and partly records multi-night totals, with 26.4% of listings displayed above $1,000.

Residential rental listings

Realtor.ca rental listings extracted on 8 January 2026: 8,303 listings with a monthly asking rent and valid coordinates, 7,907 after removing 396 duplicate records, in 253 municipalities; 37.9% share their coordinates with another listing.

Exposure and neighbourhood variables

Counts of Airbnb listings within four radii of each rental listing, with ring and margin decompositions; postal codes give forward sortation areas and city-by-area cells for the fixed effects.

Inside Airbnb snapshots for the coverage check

Public snapshots for Montreal (14 September 2026, 10,717 listings) and Quebec City (17 September 2026, 2,588 listings), about eight months later than the paper's file; 55.4% of Montreal listings require 28 nights or more.

Methodology

Hedonic rent equations with nested location fixed effects

Log monthly asking rent regressed on the 500 m Airbnb count, bedrooms, bathrooms and building type, with city fixed effects and then city-by-neighbourhood cell fixed effects.

Inference under spatial dependence

Heteroskedasticity-robust, cluster-robust, Conley spatial HAC and spatial block-bootstrap standard errors, with confidence intervals and minimum detectable effects for the within-cell estimate.

Is the within-neighbourhood estimate informative?

Coefficient-stability calculations following Oster, placebo exposure counts built from listings that are plausibly not short-term rentals, and a split-half measurement of the reliability of the exposure count.

Spatial models, quantiles and robustness

Spatial autoregressive and spatial error models with HAC inference, quantile regressions with a block bootstrap, alternative radii and exposure measures, subsamples, leave-one-city-out and a 240-specification curve.

Machine-learning benchmark and Airbnb price model

OLS, LASSO, Elastic Net, random forest and gradient boosting compared under random and neighbourhood-blocked cross-validation over 10 seeds, plus a hedonic model of Airbnb displayed prices.

Reproducibility

The two raw scraped files are no longer available, so the cleaning steps that produced the committed data sets cannot be re-executed.

The five cleaned data sets (3,456 priced Airbnb listings and 8,303 rental listings with their spatial merges) are committed in Parquet format with SHA-256 checksums and redistributed for replication only.

A single entry point regenerates every table, figure and inline number of the paper from the committed data sets (about thirty minutes on a laptop), with regeneration tests, a logged run and pinned package versions.

The Inside Airbnb files used for the coverage check are not redistributed; their URLs, download dates and checksums are recorded.

Cite this paper

Boucher, S.-P. (2026). Airbnb and Residential Rents in Quebec: Evidence and Identification Limits from a Listing-Level Cross-Section. UQO Working Paper No. 5, Université du Québec en Outaouais.

@techreport{boucher2026wp5,
  author      = {Simon-Pierre Boucher},
  title       = {Airbnb and Residential Rents in Quebec: Evidence and Identification Limits from a Listing-Level Cross-Section},
  institution = {Université du Québec en Outaouais, Département des sciences administratives},
  number      = {5},
  type        = {UQO Working Paper},
  year        = {2026},
  month       = {nov},
  url         = {https://www.spboucher.ai/research/wp5},
}