Does Airbnb raise rents in Toronto, and are the city's rules working?

Some variables had to be invented, not downloaded. For every single rental I computed a walkability score, a transit score, a custom "tourist appeal" score built from Google ratings of nearby attractions, and the real door-to-door public transit time to the airport, combining street maps with actual TTC, GO, and UP Express timetables. That last one becomes the hero of chapter 4.

The spatial work also powered the custom accessibility measures: I used street networks and transit timetables to compute how long each rental's real transit trip to the airport takes (door to door, with transfers) for thousands of addresses at once. In a separate project, the same skill set surfaced 20 years of gentrification patterns across Toronto.

Controls were step one, and they end on a twist: with all 19 in place, the ordinary regression finds no significant Airbnb effect at all. Step two is the clever part. A formal check (the Durbin–Wu–Hausman test, below) shows that naive null is biased, because something invisible still drives both Airbnb activity and rents. So I borrowed a technique from economics (an instrumental-variable design) and built the instrument myself: public transit time to the airport.

The logic: a quick train to the airport matters a lot to tourists, so Airbnbs cluster where airport access is good. But long-term renters rarely pay extra rent just to reach the airport faster, especially once the model already accounts for general transit quality, walkability, and distance to downtown. That one-sidedness lets the model isolate the part of Airbnb activity driven by tourism, and measure what that does to rents. And the data agree with the logic: when the location variables entered the rent model, every one of them priced in significantly except transit time to the airport.

And then I checked the machinery: formal tests confirmed the correction was needed (a Durbin–Wu–Hausman test showed the naive model really was biased), that the lever is statistically strong (first-stage F-statistic above the conventional threshold of 10), and that the result survives a stricter confidence test (an Anderson–Rubin interval that is wider, yet its floor actually rises, from about +0.07% to +0.09% per listing, so even the most cautious reading stays above zero). The estimate: each additional entire-home Airbnb per 1,000 homes is linked to a 0.42% rent increase, statistically significant across all 6,165 rental listings (p < 0.05).

For technical readers, the full specification. Both stages run at the listing level. Stage one predicts each tract's Airbnb density from the airport instrument; stage two prices rents using the predicted density in place of the observed one:

One more wrinkle: the average hides a pattern. A citywide estimate quietly assumes Airbnb presses on every neighbourhood with the same force. The thesis put that assumption to a direct test, letting the per-listing effect vary with how owner-occupied each census tract is. The test found real variation: the effect strengthens as the owner-occupancy ratio rises, and the pattern repeats whether the ratio comes from the 2011, 2016, or 2021 census. In the headline 2016 model, the link is statistically clear in tracts covering roughly three quarters of the city.

That direction is a genuine surprise. A prominent U.S. study (Barron and colleagues, 2021) predicts the opposite, reasoning that low-ownership areas hold more convertible rental stock and should feel Airbnb hardest. The thesis offers a Toronto-sized explanation: market thickness. In high-ownership tracts the rental market is thin, with few rental units and few nearby substitutes, so when Airbnb tightens supply, renters have nowhere to step sideways and the pressure shows up in rent. In low-ownership tracts the market is thick enough to absorb more of the blow. Conversion of homes into Airbnbs still happens there; its rent effect is diluted by all the alternatives around it.

For technical readers: the moderation model interacts Airbnb density with the tract's owner-occupancy ratio.

Two details for careful readers. The gaps in that chart were tested formally rather than eyeballed: per additional listing, the outside class carries a 0.58-point larger rent effect than the regulated class, and, within the outside class, unlicensed listings carry 0.78 points more than licensed ones, with both differences statistically significant. And when each class gets the same city-wide weighting that produced the 5.62% average, the association grows: rents about 10.35% higher for the outside class's density, 11.90% for its unlicensed core, since the stronger per-listing effect more than offsets the smaller listing count.

The verdict: the by-law works, but only inside its own fence. Rather than returning homes to the long-term market, most hosts stepped just over the 28-day line, kept charging nightly-style rates, and kept the pressure on rents. For policymakers, the message is concrete: the decisive question is what to do about the two-thirds of the market standing outside the fence.

Where this stands in 2026: a follow-up study on this site revisits the loophole with fresh data. In the June 2026 snapshot, 58.7% of entire homes sit past the 28-day line, priced like housing at asking rents far above CMHC benchmarks. Read the follow-up →

That habit (original data work, careful methods, plain-language delivery) is what the SOAN ENGAGE Conference recognized with its Methodological Excellence Award (April 2026), citing "original dataset creation, complex data analysis methods, and clear communication of findings."

It's the same approach I bring to briefing notes, dashboards, and stakeholder presentations in my current research roles, and the approach I'd bring to your team.