The answer is that Auckland’s pre-reform zoned capacity was largely theoretical. Much of it required redevelopment that was not economically feasible at prevailing prices and construction costs. Their upzoning worked because it created feasible capacity: projects that developers could profitably build right away. The payoff from this increase in housing supply was a decrease in rents compared to other cities.

More generally, zoning constrains housing supply when developers want to build at a higher density than the zoning allows. For example, if zoning allows only detached houses where developers would otherwise build six-storey apartments, then zoning is preventing supply from keeping up with demand. Upzoning relaxes the zoning constraint and increases the supply of housing. Since increasing supply leads to lower prices, upzoning is a key policy option for improving housing affordability.

Upzoning reduces land costs and raises land prices

Upzoning increases the supply of high-density land, for example, by converting land zoned for detached houses into apartment-zoned land.1 So to understand upzoning, we need to look at its effects on the land market. Upzoning has seemingly contradictory effects: it reduces land costs per home, while increasing the price of land per parcel.

Upzoning a parcel of land reduces land costs by reducing the physical amount of land required to build a home. When zoning allows one house per parcel, you need to buy a whole parcel to get one home. But if the zoning allowed six apartments on a parcel, you now only need to pay for one-sixth of the parcel. So upzoning allows you to save on land costs by splitting the cost of land over more homes.

But doesn’t upzoning increase the value of the upzoned parcel itself? Most likely, yes. An apartment building generates more revenue than a single house, so developers are willing to pay more for apartment-zoned land. But this is consistent with a decrease in land costs: even though the parcel is more valuable when zoned for higher density, the land cost per home still decreases.2

For example, suppose that a vacant parcel zoned for a single-family house has a land price of $1M; when we upzone for a six-unit apartment building, developers bid more, so the land price rises to $3M. But the land cost is now split across six units, so the land cost per home falls from $1M / 1 = $1M per house to $3M / 6 = $0.5M per apartment.3 As we’ll see below, this is what happens in practice.

Upzoning also has the effect of reducing the price of existing high-density land. Imagine you own a vacant parcel of land zoned for apartments. When zoning restricts apartments across the city, developers are prevented from building, so they bid up the price of apartment-zoned land. This adds a scarcity premium to your land, making it artificially expensive. So when another street is upzoned for apartments, developers build them and increase the housing supply; this lowers rents, which reduces how much developers are able to bid for your land.4

So upzoning reduces land costs per home and has opposing effects on land prices, depending on which land we’re talking about. It increases the price of the upzoned parcel, by allowing more-valuable uses for which developers will pay more. And it decreases the price of existing high-density land, by increasing housing supply, reducing rents, and reducing developers’ budget for land.

When is zoning a constraint on housing supply?

How do we know when zoning is a constraint on the supply of housing? The simplest test is to change it and see what happens: if we upzone a parcel to allow higher density, do developers build more homes there? If so, then zoning was a constraint on supply. If upzoning doesn’t lead to any long-run changes in behavior (holding other factors constant), then it wasn’t a constraint.

Land values are another signal of restrictive zoning. Is land zoned for high-density more expensive than land zoned for low-density (controlling for location)? This gap is the scarcity premium created by restrictive zoning, and upzoning should decrease it. We can also look at whether upzoning a parcel increases its land price: are developers willing to pay more for land with more valuable uses? If so, then zoning is making high-density land artificially scarce.5

Another test of constrained zoning is whether new housing is built right up to the density limit; if zoning limits density, developers will build near the maximum allowed. More practical signs of restrictive zoning include developers requesting variances or rezonings, and NIMBY activists fighting to prevent upzoning: if zoning didn’t have any effect, they wouldn’t care if it was changed.

What would a city with unconstrained zoning look like? Developers building single-family houses on apartment-zoned land. When there’s enough upzoned land that developers can meet all of the demand for apartments, they will use the remaining land to build houses. As long as new projects in an apartment-zoned area are consistently built as apartments, we know that we don’t have enough land zoned for apartments. Alternatively, zoning is unconstrained when upzoning one additional parcel does not increase its land value, because developers are not willing to pay more for high-density land; the scarcity premium is zero. This also implies that low- and high-density land prices should be equal (controlling for location), since there’s no scarcity premium on high-density land.

When zoning constrains development in the city, it pushes new housing out into the suburbs. This forces people to either double-up with roommates in the city, or live in the suburbs with a long commute. Upzoning to allow more density in the city reverses this, allowing more people to live near their job and family. Hence, restrictive zoning may not change the metro-wide stock of housing, but it does distort supply by preventing people from living in desirable locations.

Zoned capacity: feasible or infeasible?

Returning to Auckland, if the 2015 zoning allowed developers to build 350,000 additional homes, why did they upzone to allow more? ‘Zoned capacity’ is the difference between the total housing stock theoretically allowed under current zoning and the existing housing stock. In economic terms, this is the housing supply that developers would produce if prices were infinitely high.6 But housing that will be built when rents reach $7000 is simply not relevant. What we care about is capacity that can be built right now.

The reason zoned capacity can be meaningless is redevelopment costs. Suppose a parcel has an existing 3-storey apartment and we upzone to allow for a 5-storey building. According to zoned capacity, there are now two floors of housing waiting to be built. But this requires paying demolition costs; a developer would tear down a 3-storey building to rebuild at five storeys only when prices are very high. The upzoning didn’t create feasible zoned capacity that can be profitably built at current prices.

Hence, a large zoned capacity does not imply that developers will produce a large quantity of housing at current prices. The 350,000 units of zoned capacity in Auckland were not all feasible to build. Their upzoning led to a construction boom that reduced rents because it created feasible zoned capacity.

So the theoretically-allowed housing stock is not relevant for policy. Instead, we need to ask: how much housing will be produced at what rent level? And we should evaluate a proposed upzoning based on how much feasible zoned capacity it adds.

Evidence from large upzonings

It’s important to study large upzonings that increased density by a substantial amount, since this allows us to test whether upzoning has an effect. If you take a tiny dose of medicine, you can’t learn whether the medicine actually works. Similarly, a small upzoning that barely changes zoning laws or is counteracted by poison pills (regulations that undermine feasibility) is not informative. We want a large policy change that will generate statistically-detectable effects. Luckily, we have two case studies of large upzonings: Auckland and São Paulo.

But first, we need to define one acronym: floor area ratio, or FAR, is the ratio of a building’s total floor area to the area of the land it sits on. At FAR=1, a 1-storey building could cover the entire lot, or a 2-storey building could cover half of the lot. For a building that covers half of the parcel, an extra 0.5 FAR allows you to build one storey taller.

Auckland

Auckland is the largest city in New Zealand. In 2016, it upzoned three quarters of its residential land, tripling the city’s zoned capacity, and raising the average maximum FAR by +0.24 from 0.78 to 1.02 (a 31% increase).7 The Auckland Unitary Plan (AUP) changed the zoning from primarily single-family houses to higher-density zones allowing triplexes, townhomes, and mid-rise apartments.8 Importantly, the upzoning increased both the number of allowed units as well as the maximum density; many zoning reforms fail by allowing more units per parcel without increasing density, which means the homes must be smaller and hence lower quality.

The four zones under AUP are9:

• Terrace Housing and Apartment Buildings (THA): 5-7 storeys, 2.5 FAR, no dwelling limit.

• Mixed Housing Urban (MHU): 3 storeys, 1.35 FAR, 3 dwellings.

• Mixed Housing Suburban (MHS): 2 storeys, 0.8 FAR, 3 dwellings.

• Single House (SH): 2 storeys, 0.7 FAR, 1 dwelling.

Was zoning a binding constraint in Auckland? Yes. First, as we’ll see below, upzoning led to a construction boom; if zoning was not binding, there would be no effect on supply. Second, the price of land in upzoned areas increased relative to non-upzoned areas.10 This implies that developers were willing to pay more for land that allowed higher-density housing. If zoning was not a constraint and extra density had no value, then land prices would not change.

Did upzoning reduce land costs? We have direct evidence from Cooper, Greenaway-McGrevy, and Jones (2025), who calculate land cost per square meter of floorspace as land prices divided by floorspace; see their Figure 3c plotted below. Land rezoned to the apartment category (THA) had high and increasing prices before 2016. When the reform was implemented, upzoned land gained valuable new development rights and land prices rose. At the same time, the maximum FAR increased from 0.7 to 2.5, so the allowed floorspace rose by a factor of 3.6. On net, the effect was a sharp decrease in land cost per square meter.11

What happened to housing supply? Greenaway-McGrevy and Phillips (2023) (working paper, published) studies Auckland’s upzoning using housing permits, the number of homes approved for construction. They compare upzoned areas, which allowed townhomes and apartments, to non-upzoned areas that remained limited to detached houses. If permits increased more in upzoned areas relative to non-upzoned areas, then we can attribute that extra construction as an effect of the upzoning.

The upzoning worked as predicted: most of the increased permits came from attached dwellings (e.g. townhomes and apartments) rather than detached houses (Fig. 3), and they were mostly located inside the urban core rather than outside (Fig. 4). The increase in permits in the highest-density zone came entirely from attached dwellings, while the lower-density zones saw a mix of attached and detached. This shows that high-density zones are used for higher-density housing.

Overall, upzoning led to a big increase in housing supply in upzoned areas, as shown in the graph above. This indicates that restrictive zoning was preventing people from living near transit and jobs downtown.12 Hence, the effect of upzoning is to increase housing supply in the places where people want to live.

We also expect that upzoning enabled people to have more personal space, e.g., moving out on their own instead of living with parents or roommates. Future research should study where the residents in upzoning-induced housing came from. Did they move in from the suburbs, or were they already living doubled-up in Auckland?

Increased supply in upzoned areas raises the question: did upzoning merely reallocate new construction to more desirable locations, or did it also increase the overall housing stock?13 The authors show that most of the new construction was a net increase, with some new housing being displaced from the non-upzoned areas to the upzoned areas.14 So upzoning didn’t just reallocate homes to better locations, it also increased the housing stock. This suggests that upzoning enabled in-migration from outside of the Auckland metro area, and allowed for household formation, with roommates undoubling and adult children moving out on their own.15

Greenaway-McGrevy (2026) also studies construction in Auckland, this time using the synthetic control method. This approach creates a composite version of Auckland that was never upzoned, which we can use as a control group for the real Auckland.16 So to get the effect of upzoning, we simply compare synthetic-Auckland to real-Auckland (see Fig. 6b, shown below). With this approach, Greenaway-McGrevy estimates that upzoning increased permits by about 52,000 (or 87%) over seven years, relative to the control group.17 So the latest evidence implies an 8.2% increase in the net housing stock.18

Since upzoning caused a large increase in permits, these two papers give strong evidence that zoning was a binding constraint on housing supply in Auckland. And for such a large increase in supply, we expect to find a decrease in housing rents. This is studied in Greenaway-McGrevy and So (2024) (working paper), using data on quality-adjusted rents over 2000-2022.19 They also use the synthetic control method, constructing a synthetic Auckland from other cities in New Zealand. In preliminary results, they find that after six years, upzoning decreased rents by 22% relative to the control group. In other words, without upzoning, rents would be 28% higher than they actually are.20 Auckland rents still rose in absolute terms after 2016, so the reform wasn’t a panacea; but rents would have increased even more without upzoning.

Hence, Auckland shows us how upzoning works. Since zoning was a constraint, allowing more homes led to increased housing supply, reduced land costs per home, and ultimately led to lower rents. Putting the causal results together, we get an elasticity of rents to housing stock of -3.6: a 1% increase in housing stock reduces rents by 3.6%.21 We can use this elasticity to infer that if upzoning had increased the housing stock by 12% over six years (instead of the actual 6.9%), then rents would have stayed flat.22

São Paulo

São Paulo is the largest city in Brazil, with a population of 11 million and a metro population of 22 million. Anagol, Ferreira, and Rexer (2023) (working paper) studies the effects of São Paulo’s 2016 zoning reform, which changed allowed density on a block-by-block basis. Allowed density increased on half of city blocks, driven by increased FAR along public transit routes, though some central areas were downzoned.23 On average, the maximum allowed FAR increased by 35% (or 0.55 FAR points, from 1.55 to 2.09). The authors use the fine-grained zoning map to test how upzoning affects permitting, construction, and prices. As we’ll see, upzoning increased housing supply, so zoning was a binding constraint.24 But even though the São Paulo upzoning was larger than Auckland’s in percentage terms, it actually had a smaller effect on housing supply.

To study the effects of upzoning on permitting and construction, the authors compare blocks on opposite sides of zoning boundaries. On the upzoned side, FAR is increased and developers can build at higher density; on the other side, the allowed density was unchanged or decreased. The key idea is that these blocks are very close to each other, so they are similar in many ways, sharing the same neighborhood, amenities, jobs, and transit; comparing nearby blocks helps isolate the effect of upzoning. Then the authors ask: after the 2016 reform, did permitting and construction increase more on the upzoned side compared to the non-upzoned side?25

Anagol et al. find that a 1-point increase in FAR leads to a 28 to 63% increase in multifamily permits filed, relative to control blocks. Approved permits increased even more, by 47 to 88%; this suggests that upzoning increased the approval rate. They measure new supply using property listings, and find that upzoning led to 2.1 additional listings (or a 10.3% increase) in treatment blocks relative to control blocks. Hence, upzoning led to more permits being filed and approved, and more homes being constructed and listed for sale.26

To study the effect of upzoning on prices, we need a different method. Prices are determined by supply and demand across the entire market, so increased supply in one block will reduce prices in the neighboring block as well, making it unsuitable as a control group. To address this, the authors use entire neighborhoods to measure separate housing markets; they divide São Paulo into 329 neighborhoods, each with roughly 10,000 households. The assumption here is that increases in housing supply in one neighborhood do not affect prices in other neighborhoods.27

Using neighborhood-level data, we can compare the change in quality-adjusted prices from 2019 to 2022 to the change in average FAR due to upzoning. They find that a 1-unit increase in FAR is associated with a 5.7 percentage point decrease in housing price growth.28 Their Figure 6 (see below) shows that neighborhoods with increased density have lower prices (i.e., negative price growth), and that larger density increases correspond to larger decreases in price growth.

The authors also estimate a structural model, which allows them to run a ten year simulation. They estimate that the 2016 reform would lead to a 1.9% increase in the housing stock over ten years (47,000 new units over a base of 2.45M), compared to the status quo. This leads to a 0.5% decrease in housing prices, a result that is partly driven by modelling assumptions.29

They also evaluate a counterfactual ‘double upzoning’ reform that increases average FAR from 1.55 to 3.49; this is a 125% increase in allowed density compared to the actual 35% increase under the 2016 reform. The ‘double upzoning’ would lead to a 27.5% increase in housing stock (678,000 new units) and a 7.5% decrease in prices. Hence, the effect of upzoning on housing supply is strongly nonlinear: a 3.6x larger increase in allowed density leads to a 14x larger increase in housing stock.30

Discussion

In Auckland, a 31% increase in allowed FAR (from 0.78 to 1.02) led to an 8.2% increase in housing stock over seven years. São Paulo increased FAR by 35% (from 1.55 to 2.09), but we expect only a 1.9% increase in housing stock over ten years. What’s going on? One possible explanation is that São Paulo’s nominal increase in FAR was undermined by poison pills like height limits, shadow regulations, or inclusionary zoning.

A different explanation is that Auckland’s zoning was more stringent, and that São Paulo had a smaller supply effect because its upzoning actually eliminated the scarcity premium (the gap in land price between similar high- and low-density parcels). With data on land prices, we could test whether the scarcity premium before upzoning was larger in Auckland than in São Paulo, and whether it was reduced to zero in the latter. But the bigger supply effect from the ‘double upzoning’ cuts against this explanation: further upzoning would have an effect only if zoning remained a constraint.

A better explanation is based on redevelopment costs. Auckland’s existing housing stock was mostly detached houses, while São Paulo had more apartments. Since houses are cheaper to demolish and redevelop than apartment buildings, applying a similar density increase to both cities led to much bigger effects in Auckland. Similarly, the stronger effect from ‘double upzoning’ suggests that higher allowed densities are needed to overcome redevelopment costs in São Paulo–they didn’t upzone enough.

So after adjusting for redevelopment costs, we can say that Auckland had a larger effective upzoning, since it led to a bigger increase in feasible zoned capacity: housing supply that can profitably be built at current prices. In contrast, São Paulo seems to have added more infeasible zoned capacity that is currently unprofitable to build.31 This shows that not all upzonings are equal, because not all zoned capacity is equal. When existing buildings are costly to redevelop, increasing allowed FAR may not translate to an increase in feasible capacity.

Other upzonings

Maltman and Greenaway-McGrevy (2025) studies another upzoning in New Zealand, this time in Lower Hutt, a municipality in the Wellington metro area. Lower Hutt implemented a sequence of reforms over 2016-2023: reducing and then removing parking minimums, a limited upzoning allowing small apartments and townhomes, and a widespread upzoning allowing larger apartments and townhomes. Overall, FAR increased by +1.04 or 141%.32 The authors use a synthetic control method, and report a 10-22% increase in metro-wide housing permits after six years, or a 0.9%-1.4% increase in the net housing stock, accounting for construction being displaced from non-upzoned areas.33 Looking at quality-adjusted rents, they find that upzoning reduced rents by 17% compared to the synthetic control group.

Rollet (2025) (working paper) studies a sequence of upzonings in New York City. These mostly occurred over 2005-2010 during the Bloomberg administration, and upzoned 6% of parcels. On average, upzoning increased allowed floorspace by 61%.34 Comparing earlier- and later-upzoned parcels, Rollet shows that a 1-point increase in maximum FAR is associated with a 0.09 increase in built floorspace after ten years. Again, this is explained by redevelopment costs: since NYC is already built-up, redevelopment occurs slowly due to fixed costs from demolition.35 As a result, redevelopment is more likely in low-density neighborhoods where prices are high, since these have low demolition costs and high returns. Rollet’s study is not designed to estimate the effect of supply on rents, since rents are determined by the metro-level housing supply, and he doesn’t have a control group for metro NYC.36

Büchler and Lutz (2024) studies a sequence of upzonings in the Canton of Zurich over 1995-2020. Most municipalities relaxed their zoning, with the typical reform increasing allowed floorspace by 31%. To estimate the effect of upzoning on housing supply, they compare earlier-upzoned areas to later-upzoned areas.37 The authors report that, after ten years, upzoning increases both floorspace and the number of housing units by 9%.38 They find larger effects for larger upzonings and where zoning was binding.39 As with Rollet, this study is not designed for estimating an effect on rents, since rents are determined by supply and demand at the metro level, and they don’t have a control group for metro Zurich.

These examples reinforce the lesson that upzoning works when it creates economically feasible housing supply, with enough new density to overcome redevelopment costs. Some limited zoning reforms have had limited effects. Chicago introduced discretionary transit-oriented zoning in 2013 and 2015, but did not actually upzone to allow higher density (i.e., by-right zoning). Minneapolis allowed three units per parcel in 2020, but without increasing allowed floorspace, so the new homes had to be smaller. These reforms did not create feasible zoned capacity.

Conclusion

Zoning has become a constraint on housing supply in cities around the world, making upzoning a key policy tool for improving housing affordability. We can detect when zoning is a binding constraint by looking for a scarcity premium in the price of high-density land, such as land in an apartment zone being more expensive than otherwise similar land in a house zone. When zoning is a constraint, upzoning increases land values; we’ll know that zoning has stopped being a constraint when high-density land is no longer scarce and upzoning has no uplift effect on the parcel’s land value. In this world, apartment-zoned land will sell for the same price as otherwise similar house-zoned land, and developers will build houses in apartment zones. Cities should continue upzoning until the zoning constraint stops binding.

And we have research from England, Switzerland, and the US backing this up. What happens when the same demand increase hits areas that allow vs. block new housing? In places where supply cannot keep up with demand, these studies find that housing prices grow faster, they become gentrified, and displacement risk increases. This is what the standard supply and demand model predicts.

But constructing new apartment buildings could also improve neighborhood amenities and attract richer residents. If more supply brings nicer stores and cafes, then it will induce a secondary increase in demand. And if this “induced demand” effect is larger than the supply increase, then rents could rise overall. On this view, rents and displacement risk are lower when we respond to a demand increase by blocking new density.

We can empirically test for induced demand. If new buildings bring amenities that make the neighborhood more desirable, we should see higher rents and displacement risk nearby. However, multiple studies report that adding new apartment buildings reduces rents and displacement in the same neighborhood. Moreover, these papers show that naive correlations between construction and rents do not capture causation, because developers tend to build in already-gentrifying areas.

Demand increases lead to gentrification and displacement

There are several papers studying the effects of increases in housing demand, looking at effects on prices, rents, and evictions. The key is to compare the same demand increase across areas with more- and less-constrained housing supply. Since supply constraints prevent new homes from absorbing the increase in demand, we should see a larger increase in prices and rents in areas with constrained supply.

Hilber and Vermeulen (2016) (working paper, published) tests whether increases in housing demand have a larger effect on prices when supply is more constrained. They use panel data on 353 Local Planning Authorities (LPAs) in England over 1974–2008, with price data on detached houses and townhomes. To measure supply constraints, they look at how often the local government rejects residential projects with 10+ homes (the ‘refusal rate’). One problem is that the refusal rate can be a biased proxy: if developers know an application will be rejected, they won’t apply in the first place; so observed refusals could understate how much a local government restricts supply. The authors address this by using a 2002 permitting reform as a change in actual restrictiveness.

Measuring housing demand is also tricky. Average income is often used as a proxy, but observed incomes are partly determined by whether an area allows enough housing so that new workers can move in. (We say that income is ‘endogenous’ to supply constraints.) Instead, Hilber and Vermeulen use a demand measure based on a city’s industry mix in 1971 and national industry employment growth afterwards. If your city had industries that boomed nationally, we’d predict higher housing demand. This ‘Bartik demand shifter’ allows us to define an increase in housing demand that is consistent across cities, regardless of actual supply constraints.1

The main results are from a regression of log prices on the interaction of the demand shifter and supply constraints. They find that a demand increase raises prices much more in cities that block new housing. Specifically, a 1 standard deviation increase in supply constraints raises the price-demand elasticity by around 0.7 (see their Table 5). In other words, when we make supply more constrained, the effect of a demand shift on housing prices becomes stronger. To give an intuitive interpretation, they use a counterfactual exercise: if regulatory supply constraints were removed entirely, housing prices would fall by about 16% (see Table 7).

Buchler et al. (2025) (working paper, published) repeats the same exercise using data on prices and rents in Switzerland over 2010-2021. They measure supply constraints using an index of density restrictions based on assessments of local officials, and use a Bartik shifter as the change in housing demand. As with Hilber and Vermeulen, they find that supply constraints amplify the effect of demand growth on prices and rents. For a given demand shift, price growth is 8% higher and rent growth is 13% higher when density is restricted at the 75th percentile compared to the 25th percentile. In an earlier study on US metros, Saks (2008) (working paper, published) also finds that demand increases lead to bigger increases in housing prices when supply is more constrained.

Asquith (2025) (working paper) studies the effect of housing demand increases on evictions and the stock of rent-controlled housing in San Francisco over 2004-2013. He uses the rollout of private shuttles run by tech companies as a demand shock, where a new shuttle stop in a neighborhood drives increased local demand. For example, Google employees moved to San Francisco after their IPO made them rich enough to buy a home in the city and take the shuttle to the office in Mountain View.

In preliminary results, he finds that the demand shock caused around 40 duplex landlords per year to evict their tenants and withdraw their rentals from the rent-controlled housing stock. This occurs even despite regulations strongly disincentivizing such withdrawals: under the Ellis Act, landlords have to give up two years of rental income and face ten years of restrictions before they can exit the rental market.2 Basically, renting or selling to tech workers was actually so lucrative for landlords that it was worth paying these regulatory costs.

Since San Francisco builds very little housing, we can interpret these results as showing what happens when housing demand increases and supply is prevented from keeping up. This is low-density gentrification, where rich people move into the neighborhood without an increase in density. If housing supply was more elastic, demand from new tech workers would have been absorbed, instead of spilling over into the existing rental market. Hence, Asquith shows how evictions and displacement are caused by restrictions on housing supply.3

Future research

Economists find it obvious that demand increases cause higher prices, so there are few studies on this question. But since laypeople are skeptical that housing supply matters, it would be helpful to have more research showing what happens after an increase in housing demand. A US study could extend Hilber and Vermeulen by using the Baum-Snow and Han (2024) supply elasticities to show that a demand shift leads to larger price increases in low-elasticity neighborhoods where zoning prevents supply from keeping up. Another approach is to show the ‘demand cascade’, where increases in high-end demand cause richer locals to downgrade, thereby increasing demand in the low-end market. For example, use a Bartik shifter for college-educated workers (compare Edlund et al. 2022) and show that prices and rents rise for both high- and low-end housing (or if using age to measure quality, for both new and old homes). With data on residential addresses, we could study the effect on displacement (evictions or moving to a poorer neighborhood) and gentrification (in-migration from a richer neighborhood). And with data on families, we could show that higher demand leads to larger households via doubling-up.

Increased supply causes gentrification without displacement

If supply constraints lead to higher rents and displacement in the face of growing demand, what happens when we allow supply to keep up with demand? A recent literature studies the local effects of new housing supply. These papers use the construction of new apartment buildings as the treatment variable, and test for spillover effects on rents in other buildings in the same neighborhood.4 As we’ll see below, developers build in areas where demand was already rising. So just comparing neighborhoods with vs. without a new building will give us a biased answer. To estimate the causal effect of new construction, we need a proper control group that follows the same trends as the neighborhood with the new building. Moreover, we should think of these papers as studying what happens when new apartment buildings are constructed in the context of rising demand.

The new buildings increase the supply of homes, which has a negative effect on rents. However, new buildings don’t just change the housing stock, they also change the neighborhood. If a new building improves local amenities and makes the neighborhood more desirable, then there is an induced-demand effect that increases local rents (while reducing demand and rents elsewhere).5 On the other hand, higher density leads to more congestion and can make the neighborhood less desirable (a deterred-demand effect). We estimate the net effect of the supply and amenity channels, and the evidence shows that the supply effect dominates: even if neighborhoods do become nicer, we still see a net decrease in rents.6

Asquith, Mast, and Reed (2023) (working paper, published) is the key paper in this literature. They study the local effects of large market-rate rental buildings, using data from 11 cities over 2010-2019.7 To capture amenity effects, they focus on ‘pioneer’ buildings, defined as the first new building in a neighborhood since 2010. They also limit the sample to low-income areas, meaning census tracts with median household income below the metro median. They use Zillow rental data from 2013-2018 and Infutor migration data from 2011-2017.

The paper uses multiple empirical approaches based on a difference-in-differences strategy, which compares a treatment and control group before and after a treatment takes effect. Their most persuasive method compares neighborhoods with new buildings completed in 2015-16 (treatment group) to neighborhoods with buildings completed in 2019 (control group). Both groups have a new building at some point, so we are making an apples-to-apples comparison. Note that they are comparing the treatment and control groups over the period (2013-2018) before the control neighborhoods have a building completion. They define the treatment area as a 250m (820ft) ring around the new building, capturing an area of 1-2 city blocks. The median new building increases the rental stock within 250m by 37%, so they are capturing a large supply shock.

The main results are in the Figure 2 event study (see below, ‘Near-Near Specification’), with the coefficients showing the effect relative to the control group. They find an immediate drop in rents in the year of completion, staying around 6% lower for three years. The flat pre-period coefficients (to the left of year 0) support the ‘parallel trends’ assumption that the treatment and control groups follow similar trajectories before the building is completed. Across all specifications, they find a 5-7% decrease in rent, translating to $100-$159 monthly.8

They also find consistent effects on in-migration. Their Figure 4 shows that in-migrants to the neighborhood with the new building are more likely to move from poorer neighborhoods, which makes sense if rents are falling nearby.9 Hence, even in low-income neighborhoods where new buildings may induce demand by improving amenities, the net effect of new market-rate housing is to decrease rents.10

Li (2022) (working paper, published) does a similar exercise, studying the effect of new residential high-rises in New York City over 2003-2013. Her sample includes high-rises with 7+ storeys, with data on 916 completions. She looks at the spillover effect on rents of existing residential buildings within a 500ft radius.11 Her difference-in-differences method compares neighborhoods that have a new high-rise completed nearby, with variation in timing of completion. The average building increases the total housing stock nearby by 9.4%.12

Her main result is that a 10% increase in housing units nearby reduces rents by 1% (see Table 2). The event study in Figure 5 shows that this is not driven by differential trends in treated neighborhoods before completion, and that the effect becomes more negative over time. Li also finds suggestive evidence that new high-rises improve amenities, with an increase in new restaurants. So even if improved amenities lead to increased demand, the effect is outweighed by the increase in supply, since we observe a decrease in rents.

Rollet (2025) (working paper) extends Li’s dataset and runs the same test in New York City over 2003-2021. Rollet makes a few different modelling choices: he uses floorspace while Li uses housing units; his sample is based on new residential buildings with a large percentage and absolute increase in floorspace13, resulting in 290 completions; finally, Rollet defines an observation as a 500ft neighborhood disk around a new building, instead of using parcel-level data. He also improves on Li (2022) by using a modern event study estimator that avoids using already-treated units in the control group. In preliminary results, Rollet finds that building completions lead to a steady decline in nearby rents, which continues 10 years afterwards. He reports that a 10% increase in floorspace reduces rents by 4.2%. This effect is much larger than Li’s, and can be explained by differences in the study designs.14

Pennington (2025) (working paper) is another recent paper studying the local effects of supply increases. She uses serious building fires in San Francisco over 2013-2017 as a source of new construction, and estimates the effects of new buildings on rents, displacement, and gentrification. She measures rents using Craigslist data and uses Infutor address data to define moves. Displacement is defined as moving to a zipcode with at least 10% lower median income, while gentrification means higher-income newcomers moving into the neighborhood. Hence, gentrification can occur without displacement. In preliminary results, Pennington finds that new market-rate construction reduces rents and displacement and increases gentrification.15

Finally, Brunåker et al. (2024) (working paper) use population register data from Sweden over 1991-2017 to study the effect of new market-rate buildings in poor neighborhoods on gentrification and displacement. They focus on large co-ops (equivalent to condos) with more than 100 residents; following Asquith et al., they restrict the sample to ‘pioneer’ buildings, with no similar buildings in the same neighborhood in the previous 11 years. This gives 76 new buildings in neighborhoods with bottom-quartile income. They use a stacked difference-in-differences method, selecting control observations in the same municipality with similar income rank. As with Pennington, they find gentrification without displacement: rich people move into the new buildings, but there is no reduction in the rental stock and no out-migration from the neighborhood.

Developers build in already-gentrifying neighborhoods

Asquith, Mast, and Reed show that there is a strong selection effect for the location of new buildings: developers build in neighborhoods that are already gentrifying. In Table 2, they compare neighborhoods with versus without new housing over 2000-2010, the pre-construction period. The areas with new construction after 2010 had higher income growth (11% vs -1%), higher growth in college-graduate share (12% vs 5%), and slightly higher rent growth (19% vs 16%). Hence, a simple comparison of new construction and rents will show a positive correlation, because developers choose to build in neighborhoods with rising rents. But as we’ve seen, a proper causal analysis finds that new supply reduces rents.

Li also shows that developers choose to build in areas with rising rents. When we compare treated parcels (with a new high-rise nearby) to control parcels, we see a pre-trend in rents prior to completion of the new building (see her Figure 4). This shows that rents were already rising in the treated neighborhood before the new building was completed. Hence, we have strong evidence against the naive view that infers causation from the correlation between new construction and higher rents.

Metro-wide effects

Beyond local effects of new supply, we also have evidence that increasing housing supply reduces rents across the entire metro area. Mense (2025) (published) uses variation in new construction driven by wintry weather in Germany. He reports that a 1% increase in new completions lowers average rents by 0.19%.16 He studies heterogeneity by housing quality, finding a 0.13% rent decrease in the bottom decile of quality, and a 0.28% decrease in the top decile. Hence, housing supply decreases rents at the metro level, with an additional local effect at the neighborhood level.

How much density is needed to offset displacement?

Even if new housing supply reduces rents nearby, it may still cause displacement directly if construction requires tearing down cheap old homes and evicting renters. So while zero-displacement projects are unambiguously good, projects that displace renters through redevelopment are less clear-cut. This raises the question: what is the level of densification needed for these opposing effects to cancel out? That is, when tearing down one old home, how much do we need to densify to reduce rents enough to offset the loss of the old home? We want to know the breakeven level of densification.

Vacancy chain studies offer one answer to this question. They find that each new market-rate home frees up 0.6 homes in below-median-income neighborhoods. So the new building needs to be at least 1/0.6 = 1.67x bigger (in number of units) than the old building to maintain the stock of cheap homes. Note that the offsetting occurs at the metro level, not at the neighborhood level.

Nathanson (2025) (working paper) does a stricter version of this exercise, where we redevelop a home in the bottom decile of quality, replacing it with top-decile homes. He asks how much we need to densify to maintain the price of bottom-decile housing, with the price effect of new top-decile supply offsetting the decrease in the bottom-decile stock. He finds a breakeven threshold of 5x, so the new building needs to have at least five times more units. Note that the vacancy chain multiplier is from looking at below-median instead of bottom-decile, so naturally that threshold is smaller.

Instead of forcing developers to provide subsidized homes with no compensation (known as unfunded IZ), a better approach is to give them something in exchange (funded IZ). This ‘payment’ to developers could be a property tax exemption, reduced parking requirements, permit fast-tracking, or a density bonus (allowing more homes than under the current zoning). We say an IZ policy is calibrated when the value of the payment is at least equal to the cost of the subsidized homes, to avoid imposing a tax on new housing. The first three examples represent a pure cost: we have to give up tax revenue, reduce street parking spaces per resident, and delay permits for other projects. These forms of payment are economically equivalent to raising taxes and paying the developer in cash.

But a density bonus is special. When zoning is suboptimal and constrains development, new density creates surplus land value that we can use to fund redistribution. Since apartments are more lucrative than a single house, upzoning increases the option value of a parcel, raising its land value. Density-bonus IZ, when properly calibrated, funds the new subsidized homes using this surplus land value.1 Separately, the new market-rate housing is a benefit in itself. The more market housing we have, the more we reduce market-rate prices (through vacancy chains); this allows more people to find housing in the market, thereby reducing the need for subsidized housing in the first place.2 More density makes it easier to achieve our redistributive goal, because market-rate housing is complementary to below-market housing.3

In contrast, when IZ is funded by a property tax exemption (with no density bonus), we have to give up tax revenue to get subsidized homes. But when we allow more density, we actually benefit from higher market-rate supply and lower prices (and more property tax revenue). Housing is a benefit, not a cost. It would be one thing if we had to increase market-rate prices in order to get subsidized homes. But through density bonuses we can get both: lower market-rate prices and more subsidized housing (funded by surplus land value). There’s no tradeoff, relative to the baseline of suboptimal restrictive zoning.

If density-bonus IZ can create subsidized housing out of nothing, then what is the optimal number of IZ projects? Suppose you’re a politician who cares only about the amount of subsidized housing. In this case, you should max out on IZ apartment buildings to maximize the number of subsidized homes.4 If you instead use density bonuses sporadically (instead of systematically covering the region with IZ apartments), that’s like finding a pile of hundred dollar bills on the sidewalk and picking up only one or two.5 In practice, we don’t see any cities engaged in IZ-maxxing. This suggests that local governments are not in fact subsidized-housing maximizers, and are willing to trade off a substantial number of subsidized homes to protect neighborhood character or to avoid higher infrastructure costs.

If upzoning via density-bonus IZ creates a surplus, why not just do pure upzoning? Because new density raises land values, and in principle, this land value can be captured for public benefit. The simplest method is using a land value tax, but properly-calibrated IZ achieves the same goal. To capture all of the surplus, we have to keep upzoning until we reach the optimal zoning, where zoning regulations are not a binding constraint on development. So IZ is a transitionary policy, to be used only as we move to the optimal zoning. Cities that use IZ as a way to prevent further upzoning are undermining overall welfare.6 Again, IZ is funded out of surplus land value, and uncalibrated IZ is harmful because it imposes a tax on new housing.7 If politicians are unable to commit to a calibrated IZ policy, it may be best to avoid using it altogether.

Varieties of inclusionary zoning

There are many different ways to implement IZ.8 The main factors are the share of subsidized homes and the level of rent reduction. An IZ policy that requires 20% of units to rent at 50% of market rents means the developer gets 90% of the rental income of a market-rate building (80% of units at 100% market rents + 20% of units at 50% market rents), on top of administrative costs like verifying tenant incomes. IZ can be voluntary or mandatory, and unfunded or funded. It may apply at a threshold in building size, such as apartments with 20 units.9 Developers can provide subsidized homes on- or off-site10, or may make an in-lieu cash payment instead of providing in-kind subsidized homes. This variation in policy details makes it difficult to study IZ using panel regressions, since the average effect across such different programs is uninformative.

Soltas (2024) (published, working paper) studies a voluntary property tax exemption IZ program in New York City. He estimates a marginal cost of $1.6M per subsidized unit, driven by developer participation in expensive neighborhoods like Manhattan. This suggests evaluating IZ using a cash benchmark: is the program better than simply buying homes and renting them out at subsidized rates? In this case, the program seems poorly targeted by geography. Of course homes in Manhattan will be more expensive than in the Bronx!

Krimmel and Wang (2024) (published, working paper) studies a mandatory density-bonus IZ program in Seattle. They find that developers avoided the upzoned policy region and chose to build in other neighborhoods. This implies the program was not calibrated: the IZ tax on developers was larger than the payment (of higher density), so developers declined to participate. Similarly, Kestelman (2025) (working paper) shows that a voluntary density-bonus IZ program in Los Angeles did not increase the housing stock relative to neighboring areas. Again, comparing to a cash benchmark is helpful: if your bid for a home is below the market price, of course the seller will reject your offer.

Evidence from structural and simulation models

Rollet (2025) (working paper) builds a parcel-level general equilibrium model of housing supply and demand in New York City. The model accounts for the costs of redeveloping existing buildings, with redevelopment more likely to occur where prices are high and density is low. Using the estimated model, Rollet performs a policy experiment to evaluate the effects of IZ. The baseline scenario simulates NYC’s growth until 2060, assuming 13% growth in housing floorspace and 8% population growth, and holding zoning fixed at 2019 levels. The IZ policy is mandatory and unfunded, and requires 20% of floorspace to be rented at 50% of market-rate rents.

Rollet finds that floorspace increases by 13% in the baseline scenario. Under IZ, floorspace decreases by 1% relative to the baseline. So IZ reduces housing supply, because the IZ tax makes some projects infeasible.11 This implies an increase in market-rate rents (not shown in the current draft), which in turn increases the need for subsidized housing. The ‘exchange rate’ is 1.6 square feet of market-rate housing lost for each square foot of below-market-rate housing gained.12 Rollet doesn’t report the effect of IZ on aggregate welfare, but it is possibly welfare-reducing, because the cost of higher market rents outweigh the benefits of new subsidized homes.13

Rollet also considers a transit-oriented upzoning policy, where density is increased to FAR=6 within 0.25 miles of a transit station and to FAR=4 within 0.5 miles; this is a 59% increase in allowed density citywide. Upzoning increases built floorspace by 16% over the baseline. Combining upzoning and IZ produces a relative increase of 12%, again demonstrating how the IZ tax reduces supply. In this case, the density bonus from upzoning allows more projects to be feasible, but total supply is still lower than under pure upzoning. Again, we want to see the welfare analysis here, since we need to compare the benefit of more subsidized homes against the cost of fewer market-rate homes and higher market rents.

Lebret et al. (2025) (working paper) also builds a parcel-level model of housing in New York City, but using housing units instead of floorspace. They compare upzoning and unfunded mandatory IZ policies to a baseline after ten years. The IZ policy requires 30% of units to be rented at 50% of market rents (so landlords get 85% of rental income). Similar to Rollet (2025), they find that IZ adds subsidized units at the cost of losing market-rate units, and reduces total supply on net (see their Fig. 5A). The IZ policy has an exchange rate of 1.2 market-rate units lost for each subsidized unit gained.14

The upzoning policy increases allowed density by 25% uniformly across the city. Upzoning increases market-rate supply by 36% compared to baseline.15 A combined upzoning + IZ policy leads to a relative increase in total supply of 27%, driven by a reduction in market-rate supply and an even-larger increase in subsidized units.16 So as with Rollet, they find that the total number of housing units is lower under IZ + upzoning than under pure upzoning. But the count of total units is not the final story, since we care about aggregate welfare.

Phillips (2024) (published) studies IZ using simulations of Los Angeles’ Transit Oriented Communities (TOC) program, which is a voluntary density-bonus IZ policy adopted in 2017. The TOC program allows higher density near transit stops, as well as other incentives like reduced parking requirements. TOC has four tiers, with the fourth tier providing the largest density bonus of 55% higher residential FAR. The IZ requirement varies, but in practice involves developers renting homes to extremely-low-income (ELI) households, defined as 30% of area median income. The fourth tier has an IZ requirement of 11% of units rented to ELI households.

Phillips uses the Terner Housing Policy Simulator to compare the effects of different IZ requirements. The Simulator first assigns an optimal housing type to each parcel, and then applies a real estate pro forma to determine the probability of development over ten years. We can calculate the effect of a policy on the expected number of homes by multiplying the probability of development by the number of homes for each parcel, and aggregating across parcels.

The main results are reported in Phillips’ Figure 1, shown below. Higher IZ requirements (ELI units per building) consistently reduce market-rate supply (yellow line), while having an inverted-U relationship with subsidized ELI homes. The negative effect of IZ on market-rate units indicates that the IZ tax outweighs the density bonus, making projects infeasible even at low IZ requirements.17 Hence, the TOC density bonus is miscalibrated, and does not provide enough funding to developers to compensate for the IZ tax.

The ELI units follow a Laffer curve, with maximum ELI production of roughly 50,000 units at 25% IZ. At higher IZ requirements, we get fewer subsidized ELI homes, because so many projects are made infeasible by the IZ tax. Note that IZ has diminishing returns: requiring 16% IZ translates to only 7800 fewer ELI units compared to 25% IZ. More importantly, the cost of IZ is dramatic: at 25% IZ, the exchange rate is 5.1 market-rate units lost for each ELI unit gained. This exchange rate is minimized at 16% IZ, at 4.3 market-rate units lost per ELI unit.

Phillips does not explicitly model the effect of lower market-rate supply on rents. Instead, he calculates the increase in market rents needed to cancel out the monetary subsidy to ELI units. For example, at 16% IZ, the rent discount on ELI units is $1.41B. Since 85% of renters pay market rents, the loss in housing supply need only cause a small increase in rents to raise aggregate market rental costs by $1.41B. Specifically, the 16% IZ policy has zero financial benefits if the reduction in supply increases rents annually by 0.8% (over a 4% baseline).18 This shows how easily an IZ policy can be self-defeating.

Phillips shows that 16% IZ + TOC produces more total housing than a no-IZ, no-upzoning scenario (260k vs 245k units). But this is still less than a no-IZ upzoning scenario (400k). To identify the optimal policy, we need a model capable of doing welfare analysis, since the utility benefits of subsidized homes are plausibly larger than the benefits of market-rate homes. It is possible that pure upzoning maximizes aggregate social welfare, by reducing market-rate rents and the need for subsidized housing.

Other issues

IZ means the government is delegating the task of providing subsidized housing to private developers. On the face of it, this outsourcing of the welfare state is odd and likely suboptimal. Developers do not have a comparative advantage in implementing social policy, and may even abuse their power, for example, allocating subsidized homes as a form of compensation to their junior employees. Cook et al. (2025) (working paper) develops a model where subsidized homes are rationed by private developers; they find a small effect of developers favoring higher-income applicants. Moreover, IZ creates barriers to entry, favoring large-scale developers who can navigate the bureaucracy. Allowing in-lieu cash payments helps smaller developers avoid costly and time-consuming negotiations. A simpler approach would use a land value tax to fund nonprofit-run subsidized housing.

Unfunded IZ funds subsidized housing through a narrow tax base: developers and (via higher housing prices) purchasers of new market-rate housing. This is not optimal from a tax theory perspective. Moreover, if subsidized housing is a social good, benefitting the entire community, then it should be funded by the community as a whole. In contrast, density-bonus IZ pays for the subsidized homes out of the land value surplus created by upzoning, rather than by taxing developers and new residents.

One misconception about miscalibrated IZ is that the market-rate homes in an IZ apartment building cross-subsidize the non-market homes. But developers cannot set above-market rents, because renters would simply find cheaper rents elsewhere. Instead, it is the market-rate homes across the entire city that do the cross-subsidizing. In other words, the pass-through of the IZ tax is market-wide rather than at the building level. A miscalibrated IZ policy (where the IZ tax exceeds the payment to developers) raises rents market-wide, by killing marginal projects and reducing supply. In practice, developers in a growing city can wait for rents to rise so that the IZ policy becomes calibrated; but note that this is equivalent to IZ causing higher rents.

Density-bonus IZ creates a surplus because zoning is suboptimal and a binding constraint on development. So as we upzone and reduce the scarcity of apartment-zoned land, the magnitude of the surplus decreases. Hence, the number of subsidized homes we can fund per upzoned parcel decreases as the city becomes fully upzoned. Note that this is a sign of success: under the optimal zoning, upzoning creates no land value surplus. Moreover, the benefit is captured through lower housing prices, which allows more people to afford market-rate housing.

An open question is how well local governments can calibrate the IZ tax and developer payment, to avoid reducing supply and raising prices. In principle, cities can capture the upzoning surplus during the transition to the optimal zoning. This becomes more complicated when we consider a dynamic model, where developers anticipate future upzonings and delay construction while waiting for IZ requirements to be reduced (to maintain calibration at a smaller surplus). Political economy factors also matter, for example, with cities having perverse incentives to downzone so their IZ policy has more leverage. A land value tax would avoid such problems.

Luckily, homes do not depreciate that quickly, and older homes do stay on the market as a housing option. This is good news for lower-income families, because new homes with new materials and construction techniques tend to be more expensive, and older homes with wear and tear tend to be more affordable.

The process of today’s expensive new housing becoming tomorrow’s affordable housing is known as filtering. More specifically, filtering refers to how a home moves across the income distribution as it ages and deteriorates. We say that a home filters down if the next occupant has a lower income than the initial occupant. When the next occupant is richer, the home filters up.

Filtering is a race between depreciation and quality-adjusted price growth.1 If quality-adjusted prices increase faster than a home depreciates, then actual prices rise. Assuming that richer people buy more expensive homes, this implies that the new occupant will have higher income than the original occupant. That is, the home has filtered upwards. In contrast, if quality-adjusted prices grow slower than depreciation, then actual prices fall, and the new occupant will be poorer. The home has filtered downwards.

The filtering rate is an indicator for the overall housing market, alongside prices, doubling-up, and homelessness. When housing supply is adequate, price growth is low and homes filter downwards. When supply is constrained, richer people move into lower-quality housing, and homes filter upwards. Hence, we should interpret the filtering rate as a policy choice, since it is determined by housing policies.

Filtering is closely related to vacancy chains. Depreciation creates the affordable older homes that are freed up in the later rounds of a vacancy chain. But while depreciation can take decades to make a new home become affordable, vacancy chains free up old depreciated housing quickly—within one year. Note that filtering follows a specific home over time, and vacancy chains follow a vacancy across the housing stock.

Filtering rates are useful for long-range forecasting. If we want to know the number of affordable homes in 20 years, then we need to know whether and how quickly expensive new homes age into low-price homes over time.

Income and rent decreases with building age

We can see how filtering works by looking at correlations between building age and occupant incomes or rents. Myers and Park (2020) use US census and ACS data to plot the share of below-median-income households living in rental buildings, separately by building vintage (construction year); see their Figure 4b below. As expected, when we compare across vintages, older buildings have more lower-income residents. Moreover, for each vintage the share of lower-income residents is increasing with time (and age) until 2011, then starts to decrease. This shows that filtering depends on the broader housing market: when housing supply is adequate, homes filter down; but when we face a shortage, homes filter up. Filtering is an outcome of housing supply and demand.

This CMHC report on rental apartments estimates how much rents decrease relative to a new building (see their Figure 2 below). As you would expect, older apartments have lower rents compared to new apartments, with the full effect occurring within 20 years. Note that this relative effect holds even if rents are increasing overall, because people have higher willingness-to-pay for newer buildings with new materials and construction techniques.2 It would be useful to see this graph split by metro area, where we expect a flatter graph in supply-constrained cities like Vancouver and Toronto, as demand increases bid up rents of older buildings.

Estimates of the filtering rate

While theoretical studies of filtering date back at least to Sweeney (1974), the first paper to directly estimate filtering rates was Rosenthal (2014) (published), using data from the American Housing Survey (AHS) over 1985-2011. He uses a repeat-income model, where we compare incomes for successive occupants of the same home. Filtering is measured as the change in occupant income at turnover.3 This is a direct measure of filtering, whereas previous estimates were indirect, using hedonic regressions of rents on age.4 Note that this method assumes that home characteristics (aside from depreciation) do not change between occupants.

Rosenthal’s main result is that housing filtered down at a rate of 1.9% per year over the study period. This implies that an arriving occupant of a 50-year-old home will have an income 60% lower than an occupant of a new home. Rental apartments filter down faster than owner-occupied homes, and owned homes transition to the rental market, which amplifies the overall filtering rate.

I plot his Figure 1a below, showing inflation-adjusted relative income against age for rental homes. The y-axis is the percentage difference in occupant income for a given home age compared to a new home. Here the filtering rate is the slope of the curve. For this time period, rentals clearly filtered down, with arriving occupants having lower income than initial occupants. (The graph for owner-occupied homes also shows down-filtering, though less pronounced.) The curve increases after age 50, probably due to survivor bias, as higher-quality homes are more likely to avoid demolition and stay in the dataset.

Rosenthal also uses simulations to test for heterogeneity in filtering across census regions, and shows that filtering rates are smaller when price growth is higher, notably in the New England and Pacific regions. Overall, we can say that over 1985-2011, housing supply was not overly constrained, so supply was able to keep up with demand, and homes were occupied by lower-income families as they aged.

But Liu et al. (2022) (published, working paper) shows that this conclusion does not hold after 2011. They study owner-occupied homes with Freddie Mac-funded mortgages over 1993-2018, and extend Rosenthal (2014) by testing for heterogeneity in filtering rates across geography and over time. They estimate a filtering rate around zero over 2012-2018, suggesting that supply constraints started to bind during this period, raising prices and inducing buyers of new homes to instead compete for the stock of old homes. When the housing market changed, the indicator changed.

Liu et al. also find that filtering rates vary widely across metro areas, with up-filtering in LA and DC and down-filtering in Detroit and Chicago (see their Figure 2 below). As the filtering rate is an indicator for the housing market, this is consistent with regional housing markets performing differently. Restrictive zoning in LA means that supply can’t keep up with demand, so newcomers tend to be richer than current residents.

Lastly, they investigate within-metro heterogeneity in filtering rates (see their Figure 6a below). For instance, Washington DC has neighborhoods with strong down-filtering and neighborhoods with strong up-filtering. This illustrates how housing markets can vary across neighborhoods within the same city.

Spader (2025) (published, working paper) also extends Rosenthal (2014), this time using the same AHS data over a longer time period (1985-2013 and the new 2015-2021 sample). As with Liu et al., Spader finds down-filtering over 1985-2013, but up-filtering over 2015-2021. Again, this is consistent with a worsening housing shortage starting after the Great Recession.

Kindstrom and Liang (2024) (working paper) studies filtering in Sweden, using register data on all residents over 1990-2017. They don’t use a repeat-income model, but instead study how a building’s average income (relative to the national average) varies with building age. They find that, on average, a building’s relative income decreases by 14% (from 1.12 to 0.96) over 50 years. This is much smaller than the income decay of 60% in Rosenthal (2014), which is partly explained by lower inequality in Sweden.

Kindstrom and Liang report several other findings. First, adjusting for renovations leads to a much steeper filtering function, with relative income decreasing by 24% (from 1.12 to about 0.85) over 50 years. Second, it takes about 30 years for buildings to reach a representative income distribution, with equal shares of residents by income quartiles. Third, income decays faster in rental apartments compared to owned apartments, which in turn have faster income decay than detached houses.5

Hansen and Rambaldi (2025) (working paper) study filtering using data from the census of home transactions in Victoria, Australia over 2011-2016. They also use a different method than the repeat-income model, since they have data on buyer and seller income from the same home sale. This allows them to measure contemporaneous relative income as the ratio of buyer to seller income.6 They define the marginal filtering rate as the change in relative income with respect to age, and find up-filtering over this period, suggesting that supply constraints are binding.

Hansen and Rambaldi directly estimate the effect of supply constraints on filtering rates, using the share of development applications refused by local councils and a 2014 policy change partially removing council discretion. They find that higher refusal rates lead to more up-filtering, and that setting the refusal rate to zero implies a negative filtering rate (i.e., down-filtering). Since supply constraints drive price growth, removing those constraints would reduce prices and result in buyers having lower incomes than sellers.

Policy relevance

Rosenthal frames his paper as addressing the debate over housing vouchers vs. construction subsidies (e.g., LIHTC). If homes filter down to low-income families, Rosenthal argues, then we can provide subsidized housing by offering vouchers that families can use to access these low-price homes, as opposed to directly providing non-market subsidized housing.

This framing is odd. We don’t need to know filtering rates to judge whether vouchers will fully cover housing costs; we can simply look at current market rents. If rents are low enough, then a large-scale voucher program is feasible. However, the filtering rate would be relevant for forecasting the stock of low-income housing to evaluate a voucher program over the long-term.

Misinterpretations

The existence of up-filtering is not evidence that housing markets don’t work, as some claim. Instead, it can simply mean that markets are constrained by zoning regulations, and that we need to remove those constraints. Remember that filtering is just another housing market indicator; saying that up-filtering is evidence against markets is like saying that high prices are evidence against markets.

Spader makes this mistake: "In these areas [with up-filtering], supplying affordable housing units may require lower reliance on market-rate units and greater reliance on production of new housing units through LIHTC, inclusionary zoning, and other strategies for producing new affordable units." (p.20) Again, up-filtering just means there was a housing shortage. The natural remedy to a shortage is to increase supply, i.e, upzoning to add more market-rate homes.

Similarly, it’s equally mistaken to think that new homes automatically filter down and become affordable. If supply is constrained, up-filtering occurs instead. Liu et al. write that “filtering rates are equilibrium outcomes rather than fixed deep parameters of an economy.” (p.24) It speaks volumes that they felt the need to include this sentence!

Rosenthal’s title is "Are Private Markets and Filtering a Viable Source of Low-Income Housing?" There are two problems here. First, filtering is an outcome, not a mechanism, so it’s misleading to describe it as a source of low-income housing. This is like asking if prices are a viable source of low-income housing. But prices are the outcome, not a mechanism. Second, it neglects context. Markets may provide low-income housing in some locations and time periods, but not others. A better title would be: "Were Private Markets a Viable Source of Low-Income Housing Over 1985-2011"?

With this framing, the question now becomes: how does increasing the supply of High-quality affect prices of High- and Low-quality homes? If the High and Low submarkets are disconnected from each other, then we expect a shift in High-quality supply to reduce High-quality prices, but have no effect on Low-quality prices. In contrast, if the submarkets are connected, then increasing High-quality supply will reduce the prices of both High- and Low-quality homes.1 Hence, the key empirical question is whether housing submarkets are interconnected.

One key mechanism connecting submarkets is a vacancy chain. When person A moves into a new home H0, they vacate their previous unit H1; person B moves into H1 and vacates their original home H2; then person C moves into H2, freeing up their home for someone else, and so on. This occurs because housing is a durable good: when you're done using a home, someone else can re-use it. So if a vacancy chain initiated by an expensive new building reaches cheaper homes in poorer neighborhoods, then the High and Low submarkets are in fact connected, and increasing High-quality supply will reduce Low-quality prices. The policy implication is that simply allowing developers to build market-rate housing can improve housing affordability for everyone.

Vacancy chains occur when locals upgrade their housing situation, moving into a new building and vacating their previous unit. For example, think of young professionals getting a higher-paid job. The chain ends when the previous unit is not vacated, as when an adult child moves out of their parents’ home. And when someone moves into a new building from out-of-town, they still create a vacancy, but in their original city.2

Note that vacancy chains can occur quickly, within one year. For example, consider people A, B, and C living in homes H1, H2, and H3 in 2023. In 2024, a new home H0 is built and A moves in, leading B to move into H1 and C to move into H2; all of the moves take place during 2024.3 Over time, the housing allocation looks like this:

• 2023: (A,H1), (B,H2), (C,H3)

• 2024: (A,H0), (B,H1), (C,H2)

These are short-run effects, as opposed to depreciation, where it takes decades for a new home to become cheaper as it ages and becomes lower-quality.4

Quantifying the vacancy chain effect

Lansing et al. (1969) was the first to study vacancy chains in housing. They showed that, in later rounds of a chain, lower-income families are more common (Table 6) and home prices are lower (Graph 1). A recent literature updates these results using administrative data.

Bratu et al. (2023) (published, working paper) studies vacancy chains in Finland using register data on all residents over 2009-2019. They look at chains created by new multi-family buildings built near the Helsinki central business district. Because they can link individuals to housing units, they can trace out the exact vacancy chain.

In the first round of the chain, 20% of movers into a new building have incomes below the median. But by the fourth round, 50% of movers have below-median income (see their Figure 3a below). This shows that new market-rate buildings that are unaffordable for the poor do indirectly free up other housing that is affordable. The fraction of chains with a link in a below-median-income neighborhood is 66%; the other 34% of chains end in above-median-income neighborhoods. This means that building 100 new market-rate units will free up 66 units in below-median-income neighborhoods.

They also study the vacancy chains created by rent-controlled social housing. As you'd expect, these chains reach below-median-income neighborhoods more quickly, since the first movers are more likely to be low-income.

Mast (2023) (published, working paper) was the first paper in the recent economics literature on vacancy chains, using US data. Because his data is at the building level, tracing the chain is a bit fuzzier. But the results are similar: after six rounds, 40% of new residents are from below-median-income neighborhoods. Mast runs a simulation to compare to the no-construction scenario, where the new building was not built (in contrast, Bratu et al. compare to before the new building was built). He finds that building 100 units of new market-rate housing frees up 45-70 units in below-median-income neighborhoods. (Since Mast has neighborhood-level income data, it could be the case that richer people are moving out of poorer neighborhoods; with household-level data, Bratu et al. are able to compare both approaches, showing that Mast’s results are slightly overstated.)

Kindström and Liang (2024) (working paper) repeats the vacancy chain analysis for Sweden, using register data on all residents over 1990-2017, and tells the same story: while people moving into the new building are richer, the poorest quartile becomes the largest group of in-movers (27%) by the third round of vacancies (see their Figure 5 below). Even when looking at chains initiated by above-median-income buildings, the poorest quartile makes up 30% of movers by round 6 (see their Figure 7). So even new buildings that are aimed at rich people create vacancy chains that include poor people.

Fang et al. (2025) (working paper) studies vacancy chains created by a single 512-unit condo tower in Honolulu. They follow movers using residential histories from a private data company, and use property assessment data to test whether the vacated homes are cheaper. Their data coverage is imperfect, identifying about half of the occupants in the new building. The building has 202 market-rate and 310 income-restricted units, which allows them to test for differences by housing type.

The main finding is that vacated homes are much cheaper than the new condo units. Their Figure 5 shows that the assessed price per square foot was $1152 in the new units, but only $715 and $642 in the homes freed up in rounds 1 and 2 of the chain. This effect is stronger for chains initiated by income-restricted units, consistent with Bratu et al.’s results for new social housing. The new units are also smaller than the vacated homes, indicating that the new residents are both upgrading and downsizing.

Sources of chain breaking: household formation, in-migration, measurement

Chains can end for several reasons. First, when a new household forms, the mover’s original unit may not be vacated, so the chain ends. This occurs when roommates split up, or when an adult child moves away from their parents. (Note that the chain could continue even when the origin unit is not fully vacated. For example, one roommate may move out and vacate a bedroom, allowing a new roommate to move in. The papers above are not able to account for this, which biases chain length downwards.)

Second, the vacated unit may be in another city, as occurs when the new resident is a migrant. These papers treat in-migration as ending the chain, but in reality the chain continues in the migrant’s origin city. Hence, these descriptive results understate the causal effects of new housing.5

Third, chains can end because the origin unit is vacant during the measurement period. For example, if A moves out of H1 and into H0 in December 2023 (vacating H1) and B moves into H1 in January 2024, the chain will end because H1 was vacant in December 2023. It is not clear how Bratu et al. and Kindstrom and Liang address this.

Figure A7 (below) in Bratu et al. decomposes the sources of chain terminations. Vacancies are the largest factor in the first round, while 50% of all chains break in round 5 due to in-migration. Future papers should report these statistics to allow cross-country comparisons.

Open questions

It would be informative to understand how vacancy chains vary across cities. For example, are chains more likely to reach below-median-income neighborhoods when young professionals make up a bigger share of the population (so that moves are from local upgrading rather than in-migration)? Are chains shorter when a city has higher inequality (so moves across neighborhoods are less likely)? Are chains longer when initiated by apartments vs. single-family homes, or rentals vs. condos? How many rental homes are freed up by a new condo, and vice versa?

If vacancy chains effectively add homes in poorer neighborhoods, how much can city governments rely on providing new market-rate housing as a way of meeting ‘affordable’ housing targets?6 Similarly, how should governments account for old homes freed up by the construction of new subsidized housing? Can we use data on housing prices to show directly that vacancy chains free up cheap housing? If new housing involves tearing down cheap old homes, how much does density need to increase to maintain the stock of cheap homes (via vacancy chain effects)?

Applications of vacancy chains

Mense (2025) (published, working paper) studies the effects of housing supply on rents in Germany, using delays in housing construction caused by rain and snow. He finds that a 1% increase in new supply reduces rents by 0.19% on average. This effect holds across the distribution of housing quality: rents fall by 0.13% in the first decile of quality and by 0.28% in the ninth decile. This is consistent with the quality submarkets being connected by vacancy chains: new high-quality supply enables move-ins and move-outs that ultimately reach the lowest-quality homes. Mense cannot trace out the vacancy chains, since he does not have data on individual address histories. But he does build a structural model based on the vacancy chain mechanism, where new supply initiates a vacancy chain and renters move across quality levels. This mechanism is a natural explanation for why adding high-quality homes decreases rents over the entire quality spectrum.7

Both supply and demand shocks propagate through interconnected submarkets

The other way housing submarkets are connected is by a reverse vacancy chain, or a 'demand cascade': when a rich person moves into a city and housing supply is constrained, they outbid an existing resident for High-quality housing. In turn, that existing resident downgrades to a Low-quality home, where they outbid someone else, who now has to double up (e.g., move in with family), become homeless, or move away. When an increase in High-quality demand is not absorbed by an increase in High-quality supply, it cascades down the quality submarkets. Hence, the connectedness of submarkets works for both supply increases and demand increases.

Appendix

Here’s how we can model a vacancy chain using supply and demand. There are two submarkets, for High- and Low-quality homes. Shifting S_H right (and moving along D_H, which decreases P_H) causes a shift left in D_L. See here for details.