econ referee report 2

NBER WORKING PAPER SERIES

CREATIVE DESTRUCTION: BARRIERS TO URBAN GROWTH AND THE GREAT BOSTON FIRE OF 1872

Richard Hornbeck Daniel Keniston

Working Paper 20467 http://www.nber.org/papers/w20467

NATIONAL BUREAU OF ECONOMIC RESEARCH 1050 Massachusetts Avenue

Cambridge, MA 02138 September 2014

For comments and suggestions, we thank Nava Ashraf, Leah Bouston, Bill Collins, Brad DeLong, Bob Ellickson, Ed Glaeser, Claudia Goldin, Tim Guinnane, Matthew Kahn, Larry Katz, Michael Kremer, Naomi Lamoreaux, Gary Libecap, Sendhil Mullainathan, Trevor O'Grady, Chris Udry, John Wallis, Bill Wheaton, and seminar participants at EIEF, George Washington, Harvard, MIT, NBER, Pittsburgh, UCLA, and Yale. For financial support, we thank Harvard's Taubman Center for State and Local Government and Harvard's Warburg Fund. For excellent research assistance, we thank Louis Amira, James Feigenbaum, Jan Kozak, Michael Olijnyk, Joseph Root, Sophie Wang, Alex Weckenman, and Kevin Wu, in addition to many others for their data entry work. The views expressed herein are those of the authors and do not necessarily reflect the views of the National Bureau of Economic Research.

NBER working papers are circulated for discussion and comment purposes. They have not been peer- reviewed or been subject to the review by the NBER Board of Directors that accompanies official NBER publications.

© 2014 by Richard Hornbeck and Daniel Keniston. All rights reserved. Short sections of text, not to exceed two paragraphs, may be quoted without explicit permission provided that full credit, including © notice, is given to the source.

Creative Destruction: Barriers to Urban Growth and the Great Boston Fire of 1872 Richard Hornbeck and Daniel Keniston NBER Working Paper No. 20467 September 2014 JEL No. H23,K11,N31,N91,O18,R3

ABSTRACT

Historical city growth, in the United States and worldwide, has required remarkable transformation of outdated durable buildings. Private land-use decisions may generate inefficiencies, however, due to externalities and various rigidities. This paper analyzes new plot-level data in the aftermath of the Great Boston Fire of 1872, estimating substantial economic gains from the created opportunity for widespread reconstruction. An important mechanism appears to be positive externalities from neighbors' reconstruction. Strikingly, gains from this opportunity for urban redevelopment were sufficiently large that increases in land values were comparable to the previous value of all buildings burned.

Richard Hornbeck Department of Economics Harvard University 232 Littauer Center Cambridge, MA 02138 and NBER [email protected]

Daniel Keniston Yale University P. O. Box 208269 New Haven, CT 06520-8269 and NBER [email protected]

Urban growth and economic growth are closely associated, and yet rapid economic growth

may outpace the capacity of cities to evolve. Durable buildings create natural rigidities, and

barriers to land assembly can impose large economic losses (Libecap and Lueck, 2011; Brooks

and Lutz, 2012). Land owners’ construction choices may not internalize their building’s

effects on neighbors, and the resulting lower quality structures in turn discourage future

investment.

These challenges of urban redevelopment have come up repeatedly in American history,

as modern metropolises grew from small cities. That American cities have faced these im-

pediments, however, does not imply particular effectiveness in overcoming them. Indeed,

in the aftermath of major city fires of the 19th and early 20th centuries, contemporaries

speculated that these initially calamitous events would generate benefits through the oppor-

tunity for major reconstruction (Rosen, 1986). This intriguing hypothesis, if true, suggests

inefficiencies in even wealthy urban areas.

By contrast, much of the focus on “urban renewal” has been directed at poorer urban

areas. While policy interventions in poorer neighborhoods are widely studied and contro-

versial (Jacobs, 1961; Anderson, 1964; Wilson, 1966; Collins and Shester, 2013), frictions

in real estate markets may cause even wealthy urban areas to not reach their potential for

economic development. Indeed, we examine how rapid economic growth can itself generate

substantial inefficiencies.

This paper analyzes the Great Boston Fire of 1872, with a focus on gaining insights

into the magnitudes and sources of inefficiencies in urban growth. We examine whether

the Fire created real benefits and, if so, through what channels. This historical setting

provides an opportunity to observe private landowners’ responses to the opportunity for

reconstruction during a period of rapid urban growth, avoiding challenges of the modern

period in which tighter land-use regulations and government reconstruction efforts often

obscure market incentives. The government had little role in the reconstruction of 1873

Boston, prior to zoning regulations or stronger building codes in Boston (Rosen, 1986; Fischel,

2004).

We develop a dynamic model of urban growth, which illustrates the conditions under

which widespread urban destruction generates some benefits. In our benchmark case of no

cross-plot externalities, the Fire might appear partly beneficial, as destroyed buildings are

replaced with new more valuable buildings, but the destruction generates no real economic

benefits. In the presence of neighborhood externalities, however, reconstruction after the Fire

exhibits a multiplier effect that generates economic gains. This extended model provides a

number of testable predictions that we take to the data: increases in land values in the

burned area and nearby unburned areas; increases in building values in the burned area for

1

even the most high-end buildings, and increases over time in nearby unburned areas; greater

increases in building values following the Great Fire than following individual building fires;

and no increase in land value following individual building fires.

The empirical analysis uses a new detailed plot-level dataset, covering all plots in the

burned area and surrounding areas in 1867, 1872, 1873, 1882, and 1894. Our digitization

of city tax assessment records provides data on each plot’s value of land, value of building,

size, owner name, and occupant characteristics. We begin by estimating impacts on plots

in the burned area, relative to plots in the unburned area, and then allow the impacts to

vary by distance to the burned area. The empirical specifications control flexibly for changes

associated with pre-Fire characteristics. We mainly consider impacts of the Fire on average

plot outcomes, but also use quantile regressions to examine changes in the distribution

of outcomes predicted by the model. Using data on individual building fires that occurred

around this period, drawn from Boston fire department records, we also compare the impacts

of individual building fires to the Great Fire.

The striking initial result is that land values increased immediately in the burned area,

relative to the unburned area. These estimates imply economically substantial gains from

the opportunity for widespread reconstruction, as individual landowners previously had the

opportunity to replace their own building. Land values continued to be higher through 1882

and, consistent with the model, had reversed by 1894.

Further, land values increased immediately in nearby unburned areas, relative to further

unburned areas. The nearest unburned areas received an increase in land value similar to the

burned area, and the estimated impact declines until leveling at around 1400 feet. Assuming

no impact beyond that point, the implied total impact on land values is comparable to the

total value of buildings burned in the Fire. Any increase in land value is consistent with

inefficiencies being lessened by the Fire, regardless of whether those gains exceed the direct

losses from the Fire, but the value of burned buildings provides a natural benchmark for the

economically substantial magnitude of the impacts. We are unable to quantify all spillover

effects at the city level, which might have positive and negative components, but increased

land values imply at least large local gains from the opportunity for urban redevelopment.

Building values increased substantially in the burned area, following reconstruction, and

converged over time. These impacts were greatest at the lowest quantiles of building values,

reflecting replacement of the worst building stock, but building values increased even at

the highest quantiles. Seen through the lens of the model, these results suggest that even

the most recently constructed (and, therefore, the highest value) buildings were replaced

with discreetly better buildings, consistent with neighborhood externalities. Likewise, in

nearby unburned areas, estimated increases over time in building values are consistent with

2

neighborhood spillover effects moving forward the time of optimal building replacement.

The great extent of the Fire appears central to its impacts, and perhaps the starkest

indication of this phenomenon is seen in the comparison between the Great Fire’s impacts

and the impacts of individual building fires around this period. Building values increased

following single building fires, but building values increased by more following the Great Fire.

Further, while land values increased following the Great Fire, burned plots’ land values were

unchanged following an individual building fire. These estimates are again consistent with

the Great Fire generating some multiplier effect, whether due to neighborhood externalities

or some other mechanism.

Tax assessment data provide characteristics for all plots, though a main concern is

whether assessed values accurately reflect market conditions. We collected supplemental

data from Boston’s Registry of Deeds on plot sales, and show that assessed values align

closely with the available sales data in the burned and unburned areas both before and after

the Great Fire. A particular concern is whether the fire caused a mechanical increase in

the assessment of land value in the burned area, such as in the assessment of vacant plots.

There is generally no assessed land value premium for vacant plots, however, and the impacts

on land values persist into 1882 when burned buildings have been replaced. Further, land

values increase immediately in nearby unburned areas that are not vacant after the Fire. By

contrast, following individual building fires, there is no mechanical increase in burned plots’

land value. Assessors appear to effectively separate building value and land value, consistent

with their instructions, margin notes, and the great variation in the fraction of total assessed

value that is assigned to buildings.

We also consider several additional potential mechanisms through which the Fire might

generate economic gains. The Fire might provide an opportunity for industrial firms to

change locations and improve the efficiency of their agglomeration, though we do not find

systematic increases in industrial agglomeration. The Fire created an opportunity to improve

public infrastructure in the burned area, though there were only moderate changes in the

road network and water pipes. The Fire may have caused changes in the composition of

residential and commercial occupants, which generate spillover effects along with changes in

building quality.

The Fire might encourage plot assembly, by discouraging hold-up and reducing transac-

tions costs, and thereby generate economic gains. We estimate only small increases in plot

size, however, and only after adjusting for declines in plot size associated with road widening.

These changes accompanied only a small decline in the number of unique landowners in the

burned area, implying that large landowners could not buy up plots to coordinate reconstruc-

tion on a larger scale. Estimated increases in plot sizes would only explain a small portion

3

of increased land values in the burned area, and none of the increase in nearby unburned

areas, assuming a large premium to land assembly equal to that estimated for modern Los

Angeles (Brooks and Lutz, 2012) or more primitively in our data. Our interpretation is that

the removal of durable buildings did little to reduce the frictions preventing land assembly,

rather than there being no frictions in land assembly. This interpretation is consistent with

estimated land frictions in modern Los Angeles (Brooks and Lutz, 2012), and substantial

land rigidities in even rural areas (Libecap and Lueck, 2011).

As an epilogue, we estimate modern plot-level differences in the burned area. There is

some indication of positive impacts on the total value of land and buildings, though estimates

are sensitive to the empirical specification. Our model and research design are not geared

to understand very long-run dynamics, compared to city-level studies of convergence after

destruction across broader geographic areas (Davis and Weinstein, 2002; Miguel and Roland,

2011). By focusing on plot-level data within Boston, we analyze the more short-run and

medium-run dynamics and the mechanisms generating economic gains from reconstruction.

While we examine only one city, we use plot-level data to exploit micro-level variation and

the statistical inference for the main results is robust to correcting for spatial correlation

across plots.

The Fire itself is not a policy proposal, but the Fire’s impacts are indicative of substantial

inefficiencies in even wealthy urban areas. Indeed, the implied magnitude of inefficiencies

is even larger because even widespread reconstruction after the Fire is not predicted to

obtain first-best land-use in the presence of neighborhood externalities. Our main interpre-

tation, emphasizing neighborhood externalities, is consistent with research on neighborhood

spillovers from rent control (Sims, 2007; Autor, Palmer and Pathak, 2014), home foreclosures

(Campbell, Giglio and Pathak, 2011; Hartley, 2010; Mian, Sufi and Trebbi, 2014), gentrifica-

tion (Guerrieri, Hartley and Hurst, 2013; Ioannides, 2003), and targeted housing investments

(Schwartz et al., 2006; Rossi-Hansberg, Sarte and Owen, 2010). City governments might cor-

rect these externalities by subsidizing investments with positive spillovers, taxing investments

with negative spillovers, and removing their own regulatory impediments to redevelopment.

Various policies have been developed to address similar frictions, including eminent domain,

building codes, and zoning regulations, though these policies’ application may be ineffective

or counterproductive (Munch, 1976; Chen and Yeh, 2013; Turner, Haughwout and van der

Klaauw, 2014).

Overall, impacts of the Boston Fire demonstrate substantial economic gains from urban

redevelopment. Our focus on wealthy areas, which had been growing rapidly, is complemen-

tary to research on the economic impacts of large-scale urban renewal in poor or declining

areas. Indeed, it is the growth process itself, combined with the fixed costs of building re-

4

placement, that generates the inefficiencies that are partially alleviated by the opportunity

for simultaneous reconstruction.1 The results suggest there can be substantial barriers to

urban growth amidst an otherwise prosperous economy.

I Historical Background

I.A Great Fires in the United States

Urban fires were a more common occurrence in the 19th and early 20th century United

States. Dangerous heating and lighting methods led to frequent small fires amongst densely-

located fire-prone buildings (Wermiel, 2000). Individual building fires exacted a substantial

toll and, constrained only by primitive firefighting technologies, sometimes spread through

central business districts completely destroying all buildings in a wide area.

Historians and contemporaries generally describe rapid recovery after major city fires,

and even the potential for short-run losses to generate long-run gains (Rosen, 1986). Recon-

struction was primarily managed by the private sector, though governments of burned cities

considered improvements to public infrastructure. Political obstacles largely prevented the

implementation of more ambitious proposals, however, as Rosen (1986) highlights follow-

ing Great Fires in Boston, Chicago, and Baltimore. Following the San Francisco Fire (and

Earthquake), estimates around the burned boundary find increases in residential density

(Siodla, 2013) and firm relocation (Siodla, 2014).

I.B The 1872 Great Fire of Boston

In November 1872, a small fire spread through a large section of Boston’s business district,

eventually destroying 776 buildings over 65 acres of the downtown Boston area (Figure 1).2

Boston firefighters were unable to stop the fire quickly, before it spread, due partly to sickness

amongst the fire department’s horses that prevented the rapid deployment of equipment to

the burning area (Fire Commission, 1873). The Fire burned for 22 hours, eventually stopping

with the arrival of massive firefighting resources from surrounding areas. The Fire killed 20

people and caused approximately $75 million in damages, or 11% of the total assessed value

of all Boston real estate and personal property (Frothingham, 1873).

In anticipation of the empirical analysis, a natural question concerns the endogeneity of

which plots burned. The Fire began in the south-central part of the burned region and spread

out and to the North, toward somewhat more valuable parts of the downtown area. Extensive

investigations and hearings following the Fire provide no accounts of the fire department

1In our model, areas with declining real estate demand would decline further after widespread destruc- tion. Particular functional forms for neighborhood externalities could generate multiple equilibria, whereby widespread destruction could generate gains in both declining and growing areas.

2Figure 1 also shows the location of individual land plots in our main sample, which we discuss below.

5

protecting areas differentially, which were all fairly high value at the time (Fire Commission,

1873; Fowler, 1873). Wide roads provided a natural barrier to the fire spreading, though

the fire sometimes crossed wide roads and sometimes ended within a block.3 In practice,

the empirical analysis will include controls for pre-Fire plot characteristics that allow for

differential changes over subsequent periods.

In anticipation of the theoretical framework, we note that the Fire occurred following a

period of rapid growth in Boston real estate values (Appendix Figure 1). Boston real estate

values declined later in the 1870’s, during the national “Long Depression,” but subsequently

resumed their upward growth.4

The Fire prompted substantial inflows of private sector capital to fund reconstruction,

given the strong demand for real estate investment in Boston. Boston capital markets were

well-integrated at this time, both domestically and internationally, so we will assume perfect

capital markets in the model. Insurance payouts also partly funded reconstruction, though

these payouts were often less than property owners were due.5 Insurance payouts should not

impact optimal land-use, in the presence of perfect capital markets, though we do explore

in the empirical analysis whether landowners disproportionately exited the burned area.6

Reconstruction was privately managed and generally unconstrained by government regu-

lation, as this was prior to zoning regulations. Although the weeks immediately after the fire

saw calls for government action to coordinate reconstruction, the ultimate role of the city

government in post-fire reconstruction was very limited. The city purchased some land to

widen and extend downtown roads, though landowners’ opposition limited more-ambitious

proposals to modify the road network. Similarly, calls for a strong building code were un-

dermined by lobbying from building contractors, and the ultimate legislation was weak and

substantially rescinded in 1873 (Rosen, 1986). The city also widened underground water

mains and installed new fire hydrants, though overall improvements in public goods fell

short of their initial potential.

Interestingly, public reaction to the Fire was generally optimistic following the initial

shock. On the one year anniversary of the Fire, the Evening Transcript wrote: “occurrences

3We do not observe systematic differences in 1872 in land value and building value across the Fire boundary, using our data and restricting the sample to plots within 100 feet of the Fire boundary.

4We converted these valuations to constant 1872 dollars using the David-Solar CPI (Lindert and Sutch, 2006).

5Insurance coverage was worth three-fourths of total fire damages, but many insurance companies were bankrupted by the Fire. As a result, total insurance payouts covered closer to half of of the total damages (Fowler, 1873).

6In practice, some landowners may have been liquidity-constrained after the fire destroyed their property and the collateral needed to raise more capital. We would have been interested in testing this hypothesis more fully, though we have been unable to link particular plots to their insurance underwriter and the fraction paid out on the insurance policies.

6

calamitous in their first effects sometimes result in important material good.... That great

fire ... furnished the opportunity for rebuilding the metropolis at its very center of operation

on a comprehensive scale” (cited in Rosen, 1986). Newspapers and other contemporaries

noted that buildings in the burned area were often better after reconstruction. These ob-

served impacts need not imply any economic gains from the Fire, however, and we formalize

this intuition below in our benchmark model. We then present an extended model, with

neighborhood externalities, that highlights how the Fire might indeed result in important

material good.

II Dynamic Model of Urban Growth

II.A Benchmark Model with Durable Buildings

Our benchmark model clarifies conditions under which the Fire may only appear to generate

economic benefits. We consider the decisions of landowners choosing when to replace their

building, but who experience no spillover impacts from nearby plots. This benchmark model

formalizes our null hypothesis, in which the Fire does not generate any economic benefits.

We assume that each landowner owns one plot, and that all landowners and plots are

homogeneous. Landowners construct a sequence of durable buildings to maximize the net

present value of rents from their plot, which are assumed to depend solely on the quality

of their building (q) and the city’s overall productivity (ωt). In each period, a building of

physical quality q generates rent of r (q,ωt). In particular, we assume that the marginal

return to building quality is increasing in city productivity: ∂2r (q,ωt) /∂q∂ωt > 0.

We focus on the case in which city productivity is growing over time, which increases the

return to building quality and encourages landowners to construct higher quality buildings.

The predicted impacts of a Great Fire would differ in a city with declining productivity.7

For clarity, we assume that landowners may only completely replace their old building

with a new building of quality q ′

by paying a convex cost c ( q ′)

.8 In particular, we assume

that buildings cannot be renovated and that buildings do not depreciate. These two as-

sumptions make the model’s predictions more apparent, but do not qualitatively change the

predictions.9 As a matter of notation, we assume that building construction is instanta-

7Notably, the failure of declining cities to recover after disasters is not inconsistent with our predictions; indeed, we would predict that widespread destruction would hasten the decline of cities otherwise declining. Only in historically poor cities, with currently increasing returns to real estate investment, would we predict that destruction generates increased building quality.

8The assumption of convex costs guarantees an interior solution. 9The model’s predictions are similar if buildings depreciate, or if buildings can be renovated to some

higher quality at a cost that is greater than the costs of constructing new buildings with those two levels of quality (e.g., with fixed costs to renovation). As depreciation rates increase, and renovation becomes relatively cheaper, the dynamic optimization problem simply reverts to an effectively static optimization problem that negates the purpose of this theoretical framework.

7

neous.10

Building construction is a forward-looking dynamic optimization problem, in which each

landowner considers the optimal time to replace a building. Landowners do not replace a

building when it would generate higher static rents; rather, landowners solve for the optimal

replacement policy incorporating the option value of retaining antiquated but still profitable

real estate. This intuition is captured by the following Bellman equation, which reflects the

landowner’s value of owning a building of quality q when the city has productivity ωt (and

includes the option to rebuild):

V (q,ω) = max

{ r (q,ω) + δE [V (q,ω′)] r (q∗,ω) + δE [V (q∗,ω′)] − c (q∗)

where q∗ maximizes r (q,ωt) +βE [V (q,ω′)]−c (q). That is, q∗ represents the optimal quality building to construct if the landowner chooses to construct a new building.

The landowner faces a tradeoff between two choices: (1) receiving rent r (q,ωt) and

continuing with the old building of quality q; and (2) paying a lump sum cost c (q∗) to

construct a higher-quality building, receiving higher rents, and continuing with the new

building of quality q∗. Landowners’ expectations reflect an exogenous fixed growth rate

for city productivity, in addition to an exogenous probability d that the building will be

destroyed.11

The random destruction of buildings, with some probability d, provides a mechanism

to consider the impacts of an individual building fire. Notably, in this case of exogenous

building destruction between periods, the landowner will choose to rebuild in the next period

at quality q∗.

Landowners’ optimal construction decisions involve periods of no activity and occa-

sional quality upgrades. Given that city productivity is increasing, landowners over-build

for contemporaneous conditions and then wait for city productivity to increase before re-

placing their then-obsolete building. To illustrate the equilibrium building growth paths,

we assume r (q,ω) takes the Cobb-Douglas form qαωβ (α ≥ 0, β ≥ 0, α + β ≤ 1), with c (q) = cqγ (c > 0, γ > 1).12 We generate a sample of 3000 buildings and simulate the

model until it reaches steady-state, i.e., until the growth rate of the distribution of buildings

stabilizes.

Figure 2, Panel A, graphs the steady-state evolution of the building distribution. The

10Equivalently, foregone rents could be included in the cost of construction. 11Owners’ beliefs about future valuations can be written as E [V (q,ω′)] = (1 − d) V (q,ω′) + d ·V (0,ω′) if

the owner does not rebuild, and E [V (q∗,ω′)] = (1 − d) V (q∗,ω′) + d · V (0,ω′) if the owner does rebuild. 12In our quantitative simulations we set δ = 0.9, α = β = .5, γ = 2, and c = 5. The probability of

exogenous destruction (d) is set to 0.01, and the growth rate is set to 0.06.

8

thick central line shows the mean of log building quality, which grows at a constant rate in

steady state along with the constant growth in city productivity. There are discrete jumps,

however, in the growth paths of individual buildings. Newly constructed buildings are the

highest quality buildings for one period, before being surpassed by more-recently constructed

buildings. The upper thin line denotes the maximum of log building quality in steady state,

which reflects the optimal building to construct when constructing a new building in that

period (whether by choice or because the building was exogenously destroyed).13 Surviving

buildings are endogenously replaced once city productivity increases sufficiently, and this

minimum threshold in log building quality is represented by the lower thin line.

One example building growth path, shown as a dashed line in Panel A, reflects periods

of endogenous reconstruction and exogenous destruction. In period 0, the building is exoge-

nously destroyed and is reconstructed at a higher quality level. The building remains at this

quality level as city productivity grows, until in period 42 the landowner finds it optimal to

finally tear down the building and replace it with a substantially higher quality building.

This building happens to be exogenously destroyed a few periods later, and is rebuilt to only

slightly higher quality.

Figure 2, Panel B, graphs the steady-state evolution of the building distribution for a city

that experiences a “Great Fire” in period 0 that destroys half of the buildings. Outcomes for

the burned buildings are shown using dashed lines, and outcomes for the unburned buildings

are shown using solid lines. The Fire induces all landowners in the burned area to reconstruct

their building at the current optimal quality, which raises average building quality. Further,

the Fire compresses the distribution of building qualities in the burned area around the

maximum: burned buildings are rebuilt to the same quality as newly reconstructed buildings

in unburned areas, such that there is no impact at the highest quantiles of the distribution

of building values. The Fire’s impacts on building quality are greatest toward the bottom

of the distribution, where the entire stock of older buildings is cleared out.

In this benchmark model, the Fire does not affect landowners in unburned areas. Over

time, landowners in unburned areas choose to replace their buildings and landowners in

the burned area delay further replacement, such that the distribution of building qualities

converges. Notably, convergence is slower for the bottom of the distribution. As a result,

the average quality of unburned buildings will surpass the average quality of rebuilt burned

buildings for some periods and then oscillate until random building destruction induces long-

run convergence.14

13Note that the optimal new building is “over-built,” as its quality is higher than the optimal quality if there were no expected future growth in city productivity.

14The model generates a sharp reversal, as landowners in the burned area choose to replace a large number of surviving buildings reconstructed after the Fire, though this dynamic would be smoother with

9

While the burned area might appear more-developed shortly after the Fire, there are

no economic gains from the Fire in this benchmark model. All landowners could choose

to replace their buildings in period 0 in the absence of the Fire, but the large majority of

landowners instead prefer to postpone reconstruction. There would be economic gains from

forcing individual landowners to reconstruct buildings if there were positive externalities

from reconstruction, however, which we explore in the next section.

This benchmark model yields five main testable predictions:

1. The Fire does not increase plot land values, which reflect the option value from each

land plot, V (0,ωt).

2. The Fire increases average building values in the burned area, following reconstruction,

which then converge to average building values in unburned areas.

3. The Fire’s impact on building values is decreasing in the quantile of building value,

and is zero at the highest quantiles.

4. The Fire has the same impact on building values as individual building fires.

5. Building values and land values are unaffected in unburned areas.

II.B Extended Model with Neighborhood Externalities

We now extend the benchmark model, allowing for building rents to increase in the quality

of nearby buildings. These spillover effects generate externalities, given assumptions that

land ownership is fractured and perfect contracts are unavailable.15 In this extended model,

the Fire generates economic gains that may partially or fully offset the direct losses from the

destruction, in contrast to our null hypothesis of the benchmark model in which the Fire

generates no economic gains.

Consider a modified building rent function of r (q,Q,ωt), where Q is a vector of nearby

buildings’ qualities with mean Q̄. We assume that the number of surrounding buildings is

sufficiently large that landowners take Q as given, such that neighborhood spillovers repre-

sent a pure externality. In particular, higher building quality generates positive externalities,

as building rents are increasing in the quality of nearby buildings (∂r (q,Q,wt) /∂Q̄ > 0).

Further, the return to building quality is increasing in the quality of nearby buildings

(∂2r (q,Q,wt) /∂q∂Q̄ > 0).

some random shocks to the incentives for reconstruction. 15We assume that landowners are unable to contract with all neighbors to receive payments in proportion

to the magnitude of spillover (i.e., that the associated transaction costs are prohibitive).

10

In equilibrium, the landowner’s value of owning a building of quality q when the city has

productivity ωt is now given by:

V (q,Q,ωt) = max

{ r (q,Q,ωt) + βE [V (q,Q′,p′)] r (q∗,Q,ωt) + βE [V (q∗,Q′,p′)] − c (q∗)

where q∗ maximizes r (q,Q,ωt) + βE [V (q,Q′,p′)] − c (q) and the vector Q reflects optimal building quality decisions of nearby landowners.

An individual building fire continues to have the same impacts as in the benchmark model.

Following an individual building fire, with no change to the quality of nearby buildings, the

burned building is reconstructed to quality q∗.16

The Great Fire, however, creates a positive multiplier effect because owners of burned

properties take into consideration the simultaneous construction of many surrounding higher-

quality buildings. This encourages even higher building qualities due to the assumption of

complementarity, and higher overall rents due to the assumption of positive neighborhood

spillovers. In nearby unburned areas, landowners immediately benefit from higher building

qualities in the burned area and a set of neighbors who would otherwise have waited to

upgrade do so immediately. Over time, landowners in nearby unburned areas also choose to

reconstruct their buildings sooner and to a higher quality level due to increases in nearby

buildings’ quality. In this manner, the impacts of a Great Fire spread through the city.

Landowners’ construction decisions are not completely efficient after the Fire, as the

spillover effects are not internalized, but the Fire temporarily reduces the magnitude of inef-

ficiency. Prior to the Fire, there is a disperse distribution of building qualities that includes

some particularly low-quality buildings. Since landowners consider the whole distribution

of neighbors when reconstructing properties, new buildings are lower quality than if all

other buildings were also replaced. The Fire transforms this sequential-move game into a

simultaneous-move game, and in a growing city landowners’ best responses are to construct

buildings of yet higher quality.

We focus on a single equilibrium case, in which non-increasing returns to quality cause the

Fire’s impacts to fade over time as city productivity increases and all buildings are replaced.

Indeed, in some later periods, burned areas are relatively disadvantaged because of the large

concentration of then-obsolete buildings constructed in the immediate aftermath of the Fire.

By contrast, for particular functional forms of neighborhood spillovers, the Fire could have

16We assume that one building makes a trivial contribution to the overall vector of neighborhood buildings. In principle, the earlier-than-expected reconstruction of that one burned building has some small unexpected benefit to nearby landowners and encourages them to reconstruct their buildings sooner. This small increase in the expected future quality of neighboring buildings would encourage the burned building to be rebuilt to slightly higher quality.

11

persistent impacts due to multiple equilibria.17

To illustrate these effects of the Fire, we extend the earlier numerical simulation to in-

clude neighborhood spillover effects. We modify the benchmark rent function, dividing the

productivity term into the effect of city-wide productivity (ωt) and the impact of neighbor-

hood building quality (Q) : r (q,Q,ω) = qα(Qηω1−η) β .18 We assume that the average quality

of neighboring buildings summarizes the spillover effects from neighbors: if building i has N

neighbors, Qi = (∑N

n=1 qn

) /N. Otherwise, the simulation is the same as for the benchmark

model. In the steady state, there is a constant rate of growth in neighborhood productivity

(Qηω1−η).19

Figure 3, Panel A, shows the changes in building quality after a Great Fire. The dashed

lines show changes for the burned area, and the solid lines represent changes had there

been no Fire. The presence of value spillovers creates a multiplier effect from simultaneous

reconstruction that causes buildings’ quality in the burned area to rise temporarily above

that of the best buildings had there been no Fire. The Fire’s impacts are again greatest

toward the bottom of the distribution but, in contrast the baseline, continue to positively

impact values at even the highest quantiles. Note that these effects would not apply to an

individual building fire, where the milder predictions of the baseline model continue to hold.

Over time, as in the benchmark model, building quality converges with oscillation to the

same steady state had there been no Fire.

The Fire now affects landowners in unburned areas that are close enough to the Fire

to experience changes in Q due to post-Fire reconstruction. Figure 3, Panel B, shows the

growth path for nearby unburned areas, and the solid lines continue to represent changes

had there been no Fire.20 The Fire causes landowners in nearby unburned areas to upgrade

their buildings sooner, due to reconstructed higher-quality buildings in the burned area.

Indeed, the Fire’s impacts would gradually spread through the city as landowners reconstruct

buildings to reflect their neighbors’ higher quality. These geographic spillover effects within

17If building quality externalities exhibit increasing returns over some range of qualities, then burned areas could have been trapped in an inferior equilibrium in which landowners do not invest in high-quality buildings because nearby low-quality buildings lower the return to building quality. The Fire could then enable landowners to coordinate reconstruction at higher equilibrium building values, in which landowners find it optimal to construct high-quality buildings when nearby buildings are high quality. Of course, in the case of multiple equilibria, the Fire could also prompt a transition into the inferior equilibrium.

18In our simulation, we set η = 0.8 (and continue to set α = β = 0.5). 19One technical challenge concerns owners’ beliefs about the transitional dynamics immediately after the

Fire. For simplicity, we assume that owners expect productivity and neighboring building quality to grow at the same rate after the Fire as prior to the Fire. These beliefs are correct in the long-run, and the main numerical results are not sensitive to alternative beliefs during this period of transition. In particular, model predictions are qualitatively robust to the opposite, and overly pessimistic, assumption that neighboring building quality will cease to grow entirely after the Fire.

20We simulate a nearby area in which plots receive 1/2 the Q̄ spillovers of plots with all burned neighbors.

12

the city complicate an analysis of the Fire’s aggregate impacts, as even the comparison group

is affected by the treatment, and we return to this issue in a later section of the empirical

analysis.

Predicted impacts on land values are of particular interest. In the model, a natural

definition of land value is the option value from owning a plot with no building: V (0,wt). 21

There is no distribution of land values because plots are homogeneous, so we show changes

in the value of land for each plot. Figure 4 shows the value of land in the benchmark model

where the Fire has no impact on land value (lower black line). For the extended model with

neighborhood spillovers, the upper red line shows increased land values in the burned area.

The middle blue line shows smaller increases in land value for nearby unburned areas.22

Land values converge over time, and even fall below the benchmark level due to the aging

of buildings reconstructed immediately after the Fire.

The extended model with neighborhood externalities yields seven main testable predic-

tions, of which five differ from predictions of the benchmark model:

1. The Fire increases plot land values in the burned area. Land values converge over time

in the burned area, and may even fall below land values in unburned areas.

2. The Fire increases land values in nearby unburned areas.

3. As in the benchmark model: the Fire increases average building values in the burned

area, following reconstruction, which then converge to average building values in un-

burned areas.

4. The Fire’s impact on building values is decreasing in the quantile of building value, as

in the benchmark model, but there are temporary impacts at the highest quantiles.

5. The Fire increases building values in nearby unburned areas.

6. The Fire has a greater impact on building values than individual building fires.

7. As in the benchmark model: individual building fires have no impact on land values.

II.C Additional Potential Mechanisms

There are several other potential channels through which a Great Fire might impact urban

growth. We discuss informally some of these channels below, highlighting some potential

21Note that this value equals the value of owning a building of quality q that would be chosen for replace-

ment (i.e., a “tear down” building): V (0,wt) = V ( q,wt

) .

22Plots on the boundary of the burned area experience the same impact on land value, regardless of whether burned status. Plots with a smaller fraction of burned neighbors experience correspondingly smaller spillovers.

13

empirical implications of each mechanism, and return to these channels in a later section of

the empirical analysis.

Land Assembly and Ownership Concentration. We have assumed that post-Fire redevel-

opment occurs within fixed land plots, though the Fire might have impacts through land

assembly. Land assembly, or the combination of plots, allows the construction of larger

buildings and might create more value per-unit of land when there are otherwise rigidities

preventing land assembly.

There are two main reasons why the Fire might increase plot sizes in the burned area.

First, the Fire might reduce transaction costs resulting from hold-up or other aspects of

bargaining between plot buyers and sellers.23 Second, local heterogeneity in building quality

may discourage otherwise profitable plot consolidation even within a single owner’s neigh-

boring holdings.24 By destroying all buildings in an area, the Fire coordinates the timing of

new construction and lowers the cost of land assembly.

The Fire might also concentrate land ownership, thereby improving the coordination of

urban development. If plots are assembled across owners, there would be a natural decline in

the number of owners. Further, the Fire might increase the concentration of land ownership

by reducing hold-up, as above, even if landowners’ goal is more to internalize neighborhood

externalities than to assemble plots.

We will examine changes in plot sizes, in addition to the potential land value premium

associated with observed changes in plot sizes. We will also examine whether there are

increases in ownership concentration, perhaps that reflect many small landowners selling out

in the aftermath of the Fire.

Business Agglomeration. The Fire may also impact urban development by improving the

efficiency of firms’ location decisions. Whereas firms often must make sequential location

decisions, the Fire may allow firms to move simultaneously into a more-productive spatial

distribution. Firms have a variety of reasons to locate near similar firms or firms producing

inputs or complementary goods.25 The size and location of industrial clusters may drift from

the optimum over time, however, as the city develops and new technologies are introduced.

The Fire might increase industrial agglomeration, or otherwise improve the efficiency of firm

23The bargaining power of some landowners may decline after a fire: their outside option has worsened because they cannot live in the building or continue to operate a business without substantial reconstruction costs, and some may lack liquidity and become impatient (e.g., if they are less-wealthy or less-diversified). The Fire also reduces imperfect information about the value of burned plots, as there is no uncertainty regarding building value.

24When reconstructing an older building, the nearby newer buildings may be prohibitively costly to tear down early to build one larger building.

25Optimal industry locations can reduce transportation costs, attract customers interested in cross- shopping, signal competitive prices, allow monitoring of competitors, or encourage learning and productivity gains.

14

locations, by reducing moving costs and creating an opportunity for improved cross-firm

coordination. We will examine whether the Fire increases industrial agglomeration in the

burned area.

Residential Sorting. Similarly, the Fire might impact residential sorting along with re-

placement of the building capital stock (see, e.g., Brueckner and Rosenthal, 2009). If resident

characteristics generate spillover effects, and residents’ location is fairly persistent in the ab-

sence of the Fire, then these spillover effects could become capitalized into land values and

influence land-use.

We will examine whether the Fire was associated with changes in residents’ characteris-

tics. In general, if resident characteristics are correlated with building characteristics then

we would consider building spillover effects to be some potential combination of their phys-

ical and human components. Changes in occupant characteristics provide an additional

mechanism through which building reconstruction generates spillover effects.

Infrastructure Investment. The Fire may benefit landowners by creating a unique oppor-

tunity for improvements in roads and other infrastructure. First, the absence of buildings

lowers the costs of land acquisition and construction. Second, post-disaster solidarity may

strengthen political will for public goods improvements.

We document the implemented changes in infrastructure and landowners’ apparent sup-

port or opposition. We then consider whether the estimated changes in land values are

consistent with fixed infrastructure investments, or whether land values converge over time.

While infrastructure investments are not exogenous, within the burned area, we also consider

whether areas are affected differently with differential exposure to infrastructure changes.

III Data Construction

III.A Annual Tax Assessment Records

Historically, the City of Boston sent tax assessors to each building to collect information

for annual real estate and personal property taxes. The Boston Archives contain these

handwritten ledgers from 1822 to 1944, typed records until 1974, and then digitized data.

Tax assessors recorded information for each building unit, each commercial establish-

ment, and each residential occupant. For each building unit, data include: street name and

number, assessed value of the building, assessed value of the land, plot size, and name of

the building owner. For each commercial occupant, data include: detailed industry, value

of business capital, and proprietor name. For each residential occupant, data include the

value of personal possessions and the name and occupation of all males aged 20 or older.

We collected all of these variables, aside from commercial proprietors’ names and residential

15

occupants names and occupation.26 We digitized data for 1867, 1872, 1873, 1882, and 1894

covering all plots in the area burned during the 1872 Fire (which occurred after that year’s

tax assessment) and all plots in surrounding downtown areas.27

III.B Plot-Level City Maps

The assessment data contain addresses, but not geographical proximity to the Fire. We

generated this measure by plotting each assessment entry on high resolution scans of the

plot-level Sanborn and Bromely fire insurance maps of Boston in 1867, 1873, 1883, and 1895

(Sanborn Map Company, 1867-1895; G.W. Bromley and Co, 1883). These maps indicate

the location of each building and its street address (Appendix Figure 2), and often indicate

the plot’s square footage and owner name, which were used in matching the assessed plots

to their geographic location. We “georeferenced” these historical maps to a contemporary

digital map of Boston, defining each map in geographic space. Figure 1 maps the location

of digitized land plots in 1867, and we limit the sample to land plots in this same region in

each subsequent year (Appendix Figure 3). There are 31,000 land plots in our main sample,

pooling across all five years.28

Once the debate over street widening had been resolved, the City of Boston produced a

detailed map of the burned area that shows the plot-level outline of the fire and the area

of land to be taken from all plots affected by road widening (Appendix Figure 4). As with

the fire insurance maps, we georeferenced images of this map to create a GIS polygon of the

burned area and flagged all plots that lost area due to road widening.

By combining these resources we can identify the geographic location of each plot and

calculate whether that plot is in the burned area and its distance from the burned area.

Further, we can effectively analyze a panel dataset of fixed geographic locations despite

potential changes over time in street addresses and plot boundaries. This allows us to match

plots to their pre-Fire outcomes by city block or match plots to their nearest corresponding

plot prior to the Fire. Within a year, we can also create measures of geographic industrial

agglomeration and adjust for spatial correlation in the error term.

26We collected commercial proprietors’ industry, but did not collect residents’ occupation because the land itself is used for housing and our analysis is focused on land-use. We also did not collect the names of commercial proprietors and residential occupants, as we would not be able to track individuals moving in/out of the neighborhoods analyzed.

27Through selective double-entry and back-checks, we have found initial entry and data cleaning to pro- duce highly accurate data. Tax assessors totaled the numeric data at the bottom of each page (capital, possessions, land value, building value, plot size), which was used to validate the sum of entered data. The data include what are now the West End, North End, Financial District, Downtown Crossing, Leather Dis- trict, Chinatown, and Fort Point. The data exclude more residential areas in what are now the South End, Back Bay, and Beacon Hill.

28We exclude wharfs, which are somewhat unusual in that the land area itself is endogenous. The estimated impacts on land and building value are similar, or somewhat higher, when including plots from wharfs.

16

III.C Validity of Tax Assessment Records

A main concern with tax assessment data is whether assessed values accurately reflect eco-

nomic conditions. Assessors were instructed to assign market values to land and buildings,

separately, and then also provide the total value. At that time, as in the modern period,

properties were first assessed and then the tax rate was chosen to obtain the level of tax rev-

enue targeted by the Boston City government (Fowler, 1873). Tax assessment ledger notes

contain some references to disputed property valuations and sales valuations of that building

or comparable units.

We collected supplemental data, from Boston’s Registry of Deeds, to test the relationship

between assessed values and the available data on property sales. We searched our assessment

database for cases in which plots had changed owner names between 1867 and 1894, but

retained the same street address and area in square feet. We then searched Boston’s Registry

of Deeds to confirm that a property sale had taken place, and obtained the sale price from

the property’s original deed of sale. This search yielded 72 preserved deeds for property sales

outside the burned area and 16 property sales inside the burned area.

Appendix Figure 5 shows the relationship between properties’ assessed value and sales

value, along with the 45-degree line. Assessed values align closely with the available sales

data in the burned and unburned areas, both before and after the Fire. Appendix Table 1

reports the average difference between assessed values and sales values, broken out by before

and after the Fire in the burned and unburned areas. The estimated difference-in-difference

estimate is small and statistically insignificant, although imprecisely estimated due to the

small sample. Indeed, a main advantage of the tax assessment data is in providing valuations

for all plots, both increasing power and avoiding selection bias in which plots are sold.

A related concern is whether assessors effectively provided separate valuations of land

and buildings. The tax assessment ledgers contain some margin notes that indicate land

assessments being calculated by multiplying plot size by an indicated value per-foot. Note,

however, that the same per-foot valuation is not mechanically applied to all nearby plots

(e.g., due to differences in street access and side of block). Assessors then appear to add an

assessment of the building’s value to obtain the recorded total value, and there exists much

heterogeneity in the fraction of total value assigned to buildings.

In our discussion of the empirical results, we will explore ways in which the results may

or may not be consistent with potential biases from the use of tax assessment data.

III.D Individual Building Fires

We have also obtained a sample of individual building fires, drawing on archived records of

the Boston Fire Department. These records contain the address of every fire to which the

17

department responded, as well as the owner of the building and an estimate of damages.

We digitized these records, from 1866 to 1891, and merged them to our georeferenced tax

assessment data. Using tax assessment data, we can then estimate impacts of idiosyncratic

building fires on building values and land values, and compare these estimates to the impacts

of the Great Fire.

Our goal is to obtain a sample of idiosyncratic fires that completely destroyed the building,

comparable to damage in the Great Fire. Fire Department records do not consistently note

the level of destruction, however, so we focus on fires with building damages greater than

$5000 or those with less damage for which the record specifically mentions that the building

was “totally destroyed.” This procedure naturally skews our sample toward more-valuable

buildings, but we use our tax assessment data to control for these buildings’ characteristics

prior to their idiosyncratic fire. We impose two further conditions to highlight the comparison

between individual fires and the Great Fire. First, we exclude single building fires that

occurred within the burned area after the Great Fire. Second, we exclude all fires that are

noted as having been caused by arson or were suspected to be arson. Our remaining sample

contains 109 major single building fires to compare with the Great Fire.

IV Empirical Methodology

The main empirical analysis compares changes in the burned area to changes in unburned

areas, and then separates the analysis by distance to the Fire boundary. Our data cover

all land plots in the sample region in each sample year, but one technical issue is that

there exists no direct link between every plot and its corresponding plot in other years. We

circumvent this problem by estimating changes in fixed geographic areas, given that we know

the location of each plot.

Our initial empirical specification estimates differences between the burned area and the

unburned area in each year, relative to differences between the burned area and the unburned

area in 1872 (just prior to the Fire). We regress outcome Y for plot i in year t on year fixed

effects (αt), an indicator variable for whether the plot is within the burned area (IFirei ), and interactions between the burned area indicator variable and indicators for each year (other

than 1872):

Yit = αt + ρIFirei + β1867I Fire i × I

1867 t(1)

+ β1873IFirei × I 1873 t + β1882I

Fire i × I

1882 t + β1894I

Fire i × I

1894 t + εit.

The estimated coefficient β1873 reports the change from 1872 to 1873 in the burned area,

relative to the change in unburned areas. The identification assumption is that plots in the

burned area would have changed the same as plots in the unburned area, on average, in the

18

absence of the Fire. In practice, we relax this assumption by including additional controls

that may be associated with differential changes.

In our main specifications, we control for differential changes associated with plot char-

acteristics prior to the Fire. While we cannot match each plot in later years to its own

characteristics prior to the Fire, we can predict that plot’s pre-Fire characteristics based on

its precise geographic location. As a first approximation, we assign each plot the average

pre-Fire values over all plots within its same fixed city block in 1867 and 1872. As a closer

approximation, we assign each plot the characteristics of the nearest plot in 1867 and 1872.

In practice, this “nearest neighbor” is very often that same plot in the earlier years.29 We

estimate a final specification including controls for both the nearest plot’s value and the

mean block value, as both may be independently predictive. That full empirical specifica-

tion is similar to Equation 1, but includes interactions between year fixed effects and plot i’s

predicted characteristics from 1867 and 1872 based on its block average (Ȳ blocki1867 and Ȳ block i1872 )

and based on its nearest neighbor (Ȳ neari1867 and Ȳ near i1872):

30

Yit = αt + ηtȲ block i1867 + γtȲ

block i1872 + µtȲ

near i1867 + κtȲ

near i1872(2)

+ β1873IFirei × I 1873 t + β1882I

Fire i × I

1882 t + β1894I

Fire i × I

1894 t + εit.

The estimated coefficient β1873 continues to report the relative change in the burned area from

1872 to 1873, but adjusting for the possibility that initially-different plots might have changed

differently, over each time interval, even in the absence of the Fire. Pre-Fire plot characteris-

tics generally appear to summarize the relevant cross-sectional variation that might predict

differential changes. For example, there is little additional predictive power from including

additional year-interacted controls for distance to the Old State House, which land value

data confirm is a marker of the historical center of the central business district.

For some specifications, we restrict the main sample of 31,000 plots to the 11,000 plots

that are within 1000 feet of the burned area or were themselves burned. This focuses the

empirical analysis on initially more-similar areas. There are important limitations in focusing

on areas near the boundary of the burned area, however, as nearby unburned areas may be

indirectly affected by the Fire. These spillover effects would bias estimates of changes in the

burned area, and this bias is exacerbated by focusing on areas closer to the burned boundary.

Spillover effects from the burned area are of direct interest, and we also examine how

29For the few cases in which the closest plot has missing or zero values, such as if the building was under construction, we substitute data from the closest plot with non-zero values. In a few cases when the block- level building value average is zero (e.g., due to construction), we set the log value equal to zero and include an indicator variable for those plots.

30Note that the inclusion of nearest neighbor controls in 1867 and 1872 absorbs the “main effect” of the Fire (ρIFirei ) and the relative change from 1867 to 1872 (β1867I

Fire i × I

1867 t ).

19

proximity to the burned boundary impacts unburned areas. We begin with a nonparametric

estimate of relative changes by distance to the burned boundary, and then parameterize

this relationship. While we are unable to estimate aggregate city-wide impacts, given some

potential impact on all plots, we can observe whether the spatial spillover effect appears to

dissipate within some observed distance from the burned boundary.

Two additional empirical details are worth noting. First, the regressions are weighted

by plot size. Consider the case in which two smaller plots combine into one: the one plot

continues to report land value for the same area covered by two plots previously, and weight-

ing the analysis by plot size ensures this fixed geographic area is handled comparably over

time.31 Second, the standard errors are clustered by block to adjust for serial correlation

and within-block spatial correlation. We also explore estimating Conley standard errors,

which allow for more continuous spatial correlation within periods. We introduce additional

empirical specifications and details as they are used.

V Summary Statistics and Baseline Differences in the Burned Area

Table 1 reports summary statistics for plots’ land value and building value prior to the Fire.32

On average, in 1872, plots were higher value in the burned area (column 1) than in the

unburned area (column 2). Column 3 reports this estimated difference, in logs, and indeed

the overall cross-sectional differences in 1872 are substantial and statistically significant. Plot

values had changed more similarly from 1867 to 1872, however, in the burned and unburned

areas (column 4). Land value declined from 1867 to 1872, on average, for plots in the burned

area relative to plots in the unburned area (at an annual rate of 3.5%).

The empirical methodology focuses on comparing changes after 1872 in the burned area,

relative to changes in the unburned area. This research design avoids bias from fixed dif-

ferences in the burned and unburned areas, though it is still a concern that initial cross-

sectional differences might predict differential changes after 1872. For this reason, we focus

on empirical specifications that control for differential changes associated with plots’ pre-

Fire characteristics in 1867 and 1872. We have greater confidence in the results when the

estimates are less sensitive to inclusion of these controls.

When restricting the sample to plots within 1000 feet of the Fire boundary, plot values in

1872 are more similar in the burned area (column 1) and restricted unburned area (column

5). Column 6 reports this estimated difference, in logs, and there is no longer a substantial

or statistically significant difference in land value. The cross-sectional difference in building

31Otherwise, areas experiencing plot consolidation would mechanically receive less weight and there would be a shift in the composition of the area analyzed. In addition, weighting by plot size recovers the average effect per square foot, which is used in calculating the total impact in the burned area.

32Plot values per square foot are weighted by plot size to align exactly with the subsequent empirical analysis.

20

value is smaller, but remains statistically significant. Column 7 reports changes from 1867

to 1872 in the burned area, relative to the restricted unburned area, which are similar to the

overall relative changes (column 4). We report results based on this restricted comparison

group, though geographic spillover effects might be greater, and continue to control for plots’

pre-Fire characteristics in 1867 and 1872.

VI Main Results

VI.A Impacts on Land Value

Table 2 reports estimated impacts on plot land values in the burned area, relative to plots

in unburned areas. Column 1 reports estimates from our initial specification: land values

relatively declined from 1867 to 1872 in the burned area, increased sharply from 1872 to

1873, remained similar from 1873 to 1882, and by 1894 had declined below 1872 levels

relative to the unburned area. Column 2 reports similar changes after the Fire, controlling

for plots’ average block land value prior to the Fire (in 1867 and 1872). Column 3 reports

similar estimates controlling for pre-Fire values of the nearest plot, which absorbs all pre-Fire

variation. Column 4 controls for pre-Fire values of the nearest plot and block averages. We

prefer these controls to projecting directly the negative annual trend of 3.5% (in column 1),

as asset values in principle should not exhibit large predictable changes.33 From an ex post

perspective, however, initial differences may predict differential changes and so we control

for pre-Fire characteristics. The estimated impacts after the Fire are robust to controlling

for predicted plot characteristics in 1867 and 1872.34

Estimated increases in land values from 1872 to 1873, of roughly 15% to 18%, capitalize

substantial economic gains from the opportunity for widespread reconstruction. Increased

land values are consistent with the extended model with neighborhood spillovers, rather than

the benchmark model in which land values are unchanged. Higher land values largely persist

through 1882, suggesting that initial increases are not an artifact of tax assessment in the

immediate aftermath of the Fire. Land values declined relatively in the burned area by 1894,

which may reflect predictions of the model that these areas face future rigidities in replacing

an entire cohort of obsolete buildings constructed just after the Fire. Estimates from later

periods may be spurious, however, as the identification assumption becomes more tenuous

in later periods.

Estimated increases in land value are smaller when restricting the sample to plots within

33Land values may exhibit some predictable changes along with predicted changes in location fundamentals, as in the extended model, but these changes are smoothed due to land values capitalizing the net present value of rents associated with any expected changes.

34We have also explored using kernel regressions to predict plot characteristics, as an intermediate case between block controls and neighbor controls, and the estimates are robust to that approach.

21

1000 feet of the burned boundary (Table 2, columns 5 – 8), though this could be due to

spillover effects on nearby unburned areas. Indeed, the extended model predicts the Fire

will increase land values in nearby areas.

Figure 5 shows estimated changes in land value from 1872 to 1873, grouped by plots’ dis-

tance to the burned boundary. The burned area is to the left of the dashed line, represented

by negative distances, and the unburned area is to the right and grouped into bins of 100

feet. The estimated coefficients are relative to the omitted category of plots more than 2900

feet from the burned boundary, and the vertical lines represent 95% confidence intervals.35

The empirical specification controls for plots’ pre-Fire outcomes, corresponding to the speci-

fication in column 4 of Table 2. Estimates from Table 2 are essentially the average difference

between points to the left and the right of the dashed line.

Land values increased in nearby unburned areas, and by a similar magnitude as the

increase in land values throughout the burned area. Increases in land value become smaller

as distance to the burned boundary increases, and appear to level off around 1500 feet. By

contrast, Appendix Figure 5 (Panel A) shows that distance to the Fire boundary was not

associated with systematic changes in land value from 1867 to 1872. The relative decline in

the burned area, corresponding to column 1 of Table 2, was driven by some interior burned

areas having higher land values in 1867.

Positive spillover effects generate two reasons why the Fire’s total impact on land value

would be understated by relative changes in the burned area (e.g., estimates from Table

2). First, the relative comparison understates the aggregate impact in burned areas because

nearby unburned areas are also affected. Second, the impacts on nearby unburned areas

should also be included in the aggregate impacts of the Fire. These problems could be

overcome, however, if we assumed that further unburned areas are unaffected by the Fire.

A within-city analysis is fundamentally limited in its ability to calculate city-wide impacts,

but we can bound the Fire’s impacts if spillovers are positive on net.

In principle, there could be negative spillover effects due to displacement of economic

activity within Boston, which is not reflected in our model. For example, if the overall

demand for some economic activity is fixed and increased activity is drawn into the burned

area, then this comes at the expense of unburned areas. The downtown Boston economy was

sufficiently integrated with the Greater Boston area, and even the world economy, that we

suspect there is less scope for negative spillovers through displacement of economic activity.

Furthermore, the ultimate relative decline of land values by 1894 suggests that any impacts

on economic activity were, as predicted by the model, of a temporary nature. The potential

35For the interior of the burned region, plots more than 400 feet from the burned boundary are grouped together.

22

for negative spillovers, however, is an important caveat to the estimated total impact.

We estimate the total impact on land value, subject to caveats, by parameterizing the

spatial relationship seen in Figure 5. We begin by modeling the Fire’s impact with a con-

tinuous linear function: constant within the burned area, decreasing linearly with distance

outside the burned area (distit), and then zero after some distance cutoff (c):

Yit = β0 + β1 max

{ c − distit

c , 0

} + µtȲ

near i + ηtȲ

block i + εit.(3)

This functional form is consistent with the non-parametric results in Figure 5, although a

completely non-parametric analysis provides no formal quantitative estimate of where the

spillovers end. The Appendix (Section A) discusses some details of the estimation and

explores the results’ robustness to alternative functional forms.36 Defining Yit as the log

of land value per square foot, we substitute this piecewise linear function for the indicator

function denoting the burned area, and simultaneously estimate the fire effect β1 and distance

cutoff c that best fits the data.

These estimation results provide a predicted impact of the Fire on each plot’s land value,

which depends on its distance to the burned area. Based on the value of each plot in 1873,

we calculate the implied rise in land value due to the Fire. We then sum these impacts across

all plots to obtain an estimated total impact on land value, and convert all dollar amounts

to 1872 dollars using the David-Solar CPI (Lindert and Sutch, 2006).

Table 3, Panel A, reports these estimates for an estimated spillover cutoff of 1,394 feet

from the burned area (column 1). The Fire is estimated to have increased land values by $5.5

million in the burned area (column 2), and by $9.7 million in the unburned area (column

3). The percent impact is greater in the burned area, but the level impact is greater in the

unburned area because many more plots are affected. The estimated total impact is $15.2

million (column 4), or 1.17 times the 1872 value of buildings in the burned area (column 5).

To give a sense of robustness, Panels B and C report estimated impacts when assuming the

distance cutoff to be 1149 feet or 1639 feet (i.e., the 95% confidence interval for the estimated

distance cutoff).37

The total impact on land values is comparable to the value of buildings burned, and

may have been even greater.38 This is not to imply that the Fire itself was value-enhancing,

36The estimated functional shapes are visually similar to the piecewise linear function (Appendix Figure 9), and the implied total impacts are slightly larger than under piecewise linear specification.

37Note that the standard errors in Panel A, columns 2 to 5, do take into account the uncertainty in the estimated distance cutoff.

38We also suspect that assessed building values overstate their “true economic value,” as buildings are never assessed at close to zero value even when they are a “tear down” and due for replacement.

23

as the actual damages included lost property and goods and were estimated to be at least

$75 million. The value of burned buildings provides a natural point of comparison, but

the impact on land values does not need to exceed the value of buildings for there to be

substantial inefficiencies. Indeed, the null hypothesis from the benchmark model is that the

Fire has no impact on land values. For the Fire to have any substantial impact on land

values, there must have been substantial inefficiencies in urban development.

In the immediate aftermath of the Fire, however, there may be something unusual about

tax assessments and the city economy in general. Table 3, Panels D – F, presents analogous

estimates of the total impact on land value in 1882. Estimates indicate that the Fire continues

to exert a strong impact on land values. Appendix Figure 7 shows estimated changes in land

value from 1872 to 1882 (Panel A) and from 1872 to 1894 (Panel B), grouped by plots’

distance to the burned boundary. By 1882, land values remain higher in the burned area

and have increased in nearby unburned areas relative to further unburned areas. Land

values are lower in the burned region by 1894, though still higher in the nearby unburned

area, which may reflect expectations that the burned area will become dragged down by a

large cohort of increasingly obsolete buildings.

VI.B Impacts on Building Value

Table 4 reports estimated impacts on building values in the burned area, relative to unburned

areas. Building values declined immediately with the destruction from the Fire.39 Building

values became substantially higher by 1882, however, and had partially converged by 1894.

These results are similar when restricting the sample to within 1000 feet of the burned area,

though the magnitudes are smaller in a manner consistent with positive spillover effects on

nearby areas.

Figure 6 shows estimated changes in building value, grouped by plots’ distance to the

burned boundary. From 1872 to 1873, building values decline in the burned area and are

mostly unchanged in the unburned area (Panel A). Note that some buildings just inside the

“burned area” appear not to have been completely burned, or may have simply been repaired

or reconstructed quickly. Buildings outside of the burned area do not appear to have been

damaged by the Fire.

Building values increased substantially throughout the burned area from 1872 to 1882

(Figure 6, Panel B). Nearby unburned buildings also appear to have increased in value,

relative to further unburned buildings. This geographic spillover effect appears to level off

quickly, around 300 feet, although there is some downward trend at much further distances.

By 1894, there appear to be impacts on nearby unburned areas that decline in geographic

39While vacant plots are excluded from analysis of the log value of buildings, many buildings were assessed in a partially-constructed state in the spring of 1873.

24

distance up to 1500-2000 feet (Panel C). Note that these estimates are conditional on plots’

predicted building value prior to the Fire, and there were not clear relative changes in these

areas’ building value from 1867 to 1872 (Appendix Figure 6, Panel B). While reconstruction

of burned buildings might have contributed to temporarily higher building costs in Boston,

by 1882 these nearby unburned areas had also been systematically upgraded.

The increase in building values in the burned area is consistent with both the benchmark

and extended models, but the apparent rise in building values in nearby unburned areas

is indicative of the spillovers present in the extended model. The data suggest nearby

landowners upgraded their buildings sooner to complement the increased quality of buildings

in the burned area, and the distance of geographic spillovers appears to spread from 1882

to 1894 as further landowners react to upgrading in nearby unburned areas. By contrast,

in the benchmark model, landowners in all unburned areas do not change their building

construction decisions.

Impacts on the distribution of building values in the burned area may also be indicative

of neighborhood spillover effects. While both the benchmark model and extended model

predict increases in building quality, the extended model is associated with a multiplier

effect that increases building quality at even the highest quantiles. We estimate quantile

regressions that are analogous to the mean impacts reported in column 1 of Table 4.40

Figure 7, Panel A, shows estimated changes from 1872 to 1882, by quantile, for build-

ing values in the burned area. Building values increased significantly at even the highest

quantiles, which is consistent with even the highest-quality buildings being replaced with

buildings of discreetly higher quality due to the Fire’s multiplier effect generated by neigh-

borhood spillovers. Building values increase much more at the lowest quantiles, which reflects

the removal of less-valuable buildings and a compression in the bottom of the distribution.

This result is consistent with both models, but illustrates a mechanism through which neigh-

borhood quality increases. By contrast, there is a relatively consistent effect across quantiles

of plot land value (Appendix Figure 8).

Figure 7, Panel B, shows that building values had converged by 1894 for all but the

lowest quantiles. Both models predict that convergence would be slowest for the bottom of

the distribution, and these estimates highlight that the moderate convergence reported in

Table 4 is a combination of slow and fast convergence at different points in the distribution.

There is even some indication of lower building quality in the burned area at higher quantiles.

Thus, while average building values continue to remain higher in 1894 (Table 4), the decline

40Estimates are similar from conditional quantile regressions, including controls for pre-Fire characteristics, but the interpretation of the conditional quantile results is less clear: the theory predicts that the worst buildings are upgraded the most, rather than the worst buildings conditional on their previous nearest neighbor value.

25

seen in land values by 1894 may be more forward-looking about growth potential in future

periods (Table 3). In the coming periods, the burned area will possess an aging stock of

buildings, built in the immediate aftermath of the Fire, that may reduce relative rents.

VI.C Robustness to Spatial Correlation and Weighting

One natural question concerns potential spatial correlation in plots’ land value and building

value, which might cause the empirical analysis to overstate the statistical precision of the

estimates. Our main empirical specifications allow for spatial correlation within blocks,

but we now consider allowing for spatial correlation across plots that declines linearly in

geographic distance up to some distance cutoff (Conley, 1999). Appendix Table 2 presents

the estimated impacts on land value and building value, along with standard errors that

assume different distance cutoffs. The estimated coefficients correspond exactly to those in

Table 2 and Table 4, and the estimated standard errors generally rise and then decline with

further distance cutoffs. The statistical precision remains similar, however, and the main

results remain statistically significant.

The main specifications are weighted by plot size, for reasons discussed above, but we

also consider whether the estimates are sensitive to this specification choice. Appendix Table

3 reports similar estimated impacts of the Fire on land value and building value, with and

without controls, as in the main results reported in Table 2 and Table 4.

VI.D The Great Fire vs. Individual Building Fires

We next compare the impacts of the Fire to impacts of individual building fires, which pro-

vides a natural test of whether the Fire has an additional multiplier effect from widespread

reconstruction. Both the benchmark model and extended model predict that individually-

burned buildings will be replaced with higher quality buildings, but in the benchmark model

this increase in quality is the same if a Great Fire destroys all buildings in an area. By con-

trast, the extended model predicts larger increases in building value due to the simultaneous

reconstruction of neighboring buildings to higher quality.

We extend the previous estimating equation to include both the impact of the Great Fire

and the impacts of individual building fires. For a direct comparison with the 1872 Great

Fire’s impacts in 1873, 1882, and 1894, we analyze the impacts of individual building fires

after approximately 1 year, 10 years, and 22 years have passed since the individual building

fire. To estimate individual fire effects after time interval τ, we assign the indicator IIFi equal to 1 if the plot experienced an individual building fire and Iτit equal to 1 for individual fire data approximately τ years prior to a round of digitized assessment data. 41 For plot i in

41Since very few individual fires occurred exactly 1 year, 10 years, or 22 years prior to a round of digitized assessments, we consider individual fires that occurred within a 2-year window of this target. For example, we

26

year t, the interaction of these two indicator variables defines whether that plot experienced

an individual building fire τ years ago (IIFi ×Iτit). The full estimating equation then becomes:

Yit = αt + ηtȲ block i1867 + ηtȲ

block i1872 + µtȲ

near i1867 + µtȲ

near i1872(4)

+ β1873IFirei × I 1873 t + β1882I

Fire i × I

1882 t + β1894I

Fire i × I

1894 t

+ δ1IIFi × I 1 t + δ10I

IF i × I

10 t + δ22I

IF i × I

22 t + εit.

The estimated coefficient δ1 represents the 1-year impact from an individual building fire,

and can be compared to the estimated impact of the Fire in 1873 (β1873). Similarly, δ10 and

δ22 can be compared to β1882 and β1894, respectively.

Table 5 reports estimated impacts of both the Great Fire and individual building fires.42

Building values are higher 10 years and 22 years after individual fires, but the increase in

building values are smaller than increases in building value after the Great Fire (columns 1

and 2). These differences are mostly statistically significant, consistent with the extended

model’s prediction of a multiplier effect following the Great Fire.

We do not know the magnitude of selection bias associated with individual building fires,

i.e., whether these buildings would have experienced differential changes in building values.

We suspect that the bias is positive, however, as older buildings might be at greater risk of

catching on fire and these older buildings may be due for upgrades. The analysis controls for

buildings’ characteristics in 1867 and 1872, however, which partly addresses these concerns.

There was no immediate increase in land value following individual building fires, in

contrast to the immediate increase in land value following the Great Fire (Table 5, columns

3 and 4). One concern with the main analysis, discussed above, is that the Great Fire

might cause a spurious increase in the assessment of land value (e.g., if assessors value land

higher for vacant plots). The absence of higher land values immediately following individual

building fires suggests that fires are not mechanically associated with increased assessment

of land value. There is generally no assessed land value premium for vacant plots,43 and land

values also increased immediately in nearby unburned areas that are not vacant after the

Great Fire. Assessors appear to effectively separate building value and land value, consistent

estimate 10-year effects on plots that experienced individual building fires between 1870 and 1874 (using 1882 tax assessment data) or between 1882 and 1886 (using 1894 tax assessment data). We then control for when the individual fire occurred in this 2-year window. The individual fire indicator Iτit = 1 if

∣∣t − tIFi − τ∣∣ < 2. To control for when the fire occurred within this 2-year window, we interact IIFi × I

τ it with t − t

IF i − τ and

report the impact of IIFi × I τ it when t − t

IF i = τ.

42We control for pre-Fire plot values in all specifications, since by restricting focus to individual building fires with damages greater than $5000 there is a mechanical association between individual building fires and higher-value plots.

43We do not estimate a substantial or statistically significant difference in the log value of land per square foot for vacant plots, compared to non-vacant plots within 100 feet.

27

with their instructions, margin notes, and the great variation in the fraction of total assessed

value that is assigned to buildings.44

Overall, it appears that the Fire’s inherent largeness is fundamental to its impacts on land

values and building values. We attribute much of this impact to spillover effects from the

widespread reconstruction and upgrading of nearby plots. In the next section, we consider

some additional mechanisms through which the Fire may impact urban redevelopment.

VII Additional Potential Mechanisms

VII.A Infrastructure Investment

Along with the opportunity for private landowners to reconstruct buildings, the Boston

city government had an opportunity to improve public infrastructure in the burned area.

Government plans were largely limited by resistance from landowners, but there were some

moderate improvements to the road network.

Changes in the road network are certainly not exogenous, but we begin by considering

whether plots on non-widened roads experienced increases in land value and building value.

Appendix Table 4 reports these results, which show similar increases in land value and

building value in these burned areas relative to unburned areas.

We have emphasized the gain to landowners from widespread reconstruction of neighbor-

ing buildings, but in principle these gains could be due to any new amenity in the burned

area that also generates spillover effects. While we emphasize the amenity created by higher-

quality nearby buildings, these impacts are difficult to distinguish formally from another

amenity such as higher-quality roads, wider water mains, or new fire hydrants. Whereas

these changes in infrastructure are potentially long-lasting, however, we see convergence in

land values and building values that appear more to reflect temporary upgrades to the build-

ing capital stock. Immediate increases in land value could reflect changes seen in the road

network, though this is also somewhat inconsistent with landowners’ coordinated resistance

to changes in the road network despite compensation paid for lands used (Rosen, 1986).

Given these considerations, and the limited effectiveness of changes in the road network in

reducing traffic problems (Rosen, 1986), we do not expect this to be the primary source of

the Fire’s impacts.

44On average, building value makes up 37% of the combined value of buildings and land. In considering variation across plots in the fraction of total value assigned to buildings, the standard deviation across all plots and years is 19 percentage points. Conditional on block-by-year effects, which explain 49% of the variation in the fraction of total plot value assigned to the building, the standard deviation across plot residuals is 13 percentage points. Thus, even within a block and year, there remains substantial variation in the fraction of total assessed value that is assigned to a plot’s building or land.

28

VII.B Land Assembly

The Fire may also impact land values by lowering the cost of land assembly, i.e., combining

land plots into larger units. There may be returns to scale in plot size, and yet various

rigidities might prevent the assembly of plots into larger units that would increase the value

of land per square foot (see, e.g., Brooks and Lutz, 2012).45

Table 6 reports estimated impacts on log plot size in the burned area, relative to unburned

areas, based on the same estimating equations as before. In the main sample (columns 1 and

2) and the restricted sample within 1000 feet (columns 3 and 4), there is little immediate

change in average plot size from 1872 to 1873. There is some indication of higher plot sizes

in later periods, following reconstruction, which is more consistent with the returns to land

assembly increasing with neighborhood quality. If the immediate increase in land values

were driven by declines in the cost of land assembly, then we should expect to see greater

land assembly in 1873.

Some plots in the burned area were made smaller, however, due to road widening. Table

6, columns 5 to 8, report estimates that exclude plots of land in the burned area that were

subject to road widening. Excluding those plots directly impacted by road widening, there

are small increases in plot size from 1872 to 1873 and larger increases in later periods.46

Quantitatively, the observed increases in plot size would not explain the estimated in-

creases in land value. Focusing on the increase in plot sizes for areas without road widening,

and assuming that the doubling of plot size provides a land premium of 40% (Brooks and

Lutz, 2012), the observed increases in plot size would generate approximately a 2% increase

in land value. In addition, this premium would be offset by areas losing plot size along with

road widening.

If we assume there are returns to land assembly, then it is interesting to consider the

absence of substantial land assembly immediately after the Fire. The Fire provided an

opportunity to assemble land without the need to coordinate on demolition of neighboring

buildings, which suggests that rigidities in land assembly are more related to hold-up and

transactions costs associated with the land itself. This interpretation is consistent with the

importance of rigidities even in rural agricultural areas (Libecap and Lueck, 2011).

45Indeed, in our setting, we find that log land value per square foot is positively correlated with log plot size in the cross-section prior to the Fire (coefficient of 0.481, standard error of 0.038). We do not have an appropriate empirical setting to estimate plausibly a causal relationship between plot size and land value, however, and do not claim that these estimates are causal. Indeed, the estimated cross-sectional premium declines substantially when we control for distance and direction from the State House as a proxy for the central business district (coefficient of 0.172, standard error of 0.022).

46Quantile regressions indicate that these average impacts are driven by a decline in the number of small plots.

29

VII.C Ownership Concentration

The Fire might also lead to a concentration of land ownership in the burned area. Landown-

ers might combine their own existing plots, buy-out neighbors to combine plots, or buy-out

neighbors without combining plots to better coordinate redevelopment. Ownership might

also concentrate in the burned area if some landowners were liquidity-constrained and in-

duced to sell by the Fire.

Appendix Table 5 reports some basic statistics on the number of unique landowners in

the burned area and unburned areas, over time.47 There was a general decline in the number

of unique owners over time, and a more rapid decline from 1872 to 1873 (columns 1-4).

The magnitudes are fairly small, however, and 8 of the 19 owners that exited were a direct

consequence of road changes eliminating their landholdings in the burned area.

A similar exercise shows changes over time in the number of plots in the burned area

and unburned areas (Appendix Table 5, columns 5-8). There were general declines in the

number of plots over time, with a more rapid decline from 1872 to 1873. The magnitudes

are also small, however, and 20 of the 61 plots eliminated were a direct consequence of road

changes.

Overall, there were some small relative declines in the number of landowners. There is no

indication, however, of small landowners in the burned area systematically selling off their

properties. Landownership remained highly fractured, and there were few mechanisms for

landowners to internalize their spillover effects on neighbors. Despite the Fire, we expect that

reconstruction was still well below efficient levels of quality due to the inability to internalize

spillover effects on nearby areas.

VII.D Business Agglomeration

The Fire may have allowed business owners to locate more efficiently, thereby increasing

productivity in the burned area. We focus on whether industries took advantage of potential

vacancies to agglomerate more closely, which is one of the more standard desirable features

in firms’ location decisions. We calculate a measure of spatial agglomeration (Ripley’s L

function) for the 18 industries that had more than three establishments inside and outside

the burned area in all sample years. We then consider how these industry-level statistics

changed relatively in the burned area.

The L function provides a normalized measure of the number of same-industry estab-

lishments within a radius r of each establishment, relative to the number of establishments

47Measuring the number of unique owners is challenging, due to multiple alternative spellings and own- ership vehicles (trusts, associations, partnerships, etc.) under which a single individual might register land ownership. We have attempted to reconcile as many of these as possible through manual matching; never- theless, ownership names remain noisy.

30

that would be expected under perfect spatial randomness (following Ripley, 1977). The Ap-

pendix (Section B) provides some additional details, but values of Lib > 0 are associated

with greater agglomeration, whereas negative values signify a more uniform dispersion than

would occur given a random distribution of points. We calculate Lib (r) for three radius

values (50, 100, 200) for 18 industries in 1867, 1872, 1882, and 1894.48

Appendix Table 6 presents these estimates of agglomeration, by industry, for the burned

area and unburned areas.49 Most industries display some clustering, but is no systematic

increase in industry agglomeration in the burned area, relative to the unburned area, from

1872 to 1882 (column 8) or from 1872 to 1894 (column 9). Industries appear to become

somewhat less agglomerated over time, especially the more common industries.

Appendix Table 7 reports estimated impacts on industry agglomeration in the burned

area, relative to unburned areas. The estimating equations are analogous to before, but use

the calculated Lib (r) values to characterize the degree of agglomeration in each industry and

year, and weighting each observations by the total number of sample establishments in its

industry-year.50 There is no indication of increased agglomeration in the burned area, and

some indication of a decline in industrial agglomeration when controlling for industries’ level

of agglomeration in 1867 and 1872 (columns 2, 4, 6).

These results do not immediately support the hypothesis that changes in business location

are driving the observed increases in land values, as much of the existing literature has

argued that industry agglomeration is productivity enhancing. The literature has primarily

examined the equilibrium relationship between clustering and productivity, however, rather

than the transitional dynamics. It is possible that certain industries had become overly

clustered prior to the Fire, and the increases in dispersion were associated with efficiency

gains.

VII.E Occupant Sorting

The Fire may induce differential sorting of residents and commercial establishments, along

with changes in building quality. The spillover effects we estimate may work both through

the direct effects of building quality as well as through the characteristics of the occupants

of higher quality buildings.

We begin by considering the number of commercial and residential occupants, which

we measure as the number of assessed occupants per 1000 square feet.51 Appendix Table

8, columns 1 and 2, report increases in the number of commercial occupants following the

48We exclude 1873 when many buildings were unoccupied at the time of assessment in the burned area. 49For this Appendix Table, the distance radius is set to 100 feet. 50The results are robust to unweighted regressions. 51Tax assessment data report the number of commercial establishments and the number of male residents

over 20 years of age.

31

initial decline in the immediate aftermath of the Fire. By contrast, columns 3 and 4 report

declines in the number of residential occupants. Overall, there was a temporary decline in

the total number of occupants. These results suggest the fire caused a shift in neighborhood

composition towards commercial uses, which may have been one channel through which

positive spillovers operated.52

An analysis of occupants’ capital value is greatly complicated by censoring, which is

unfortunately inherent to the assessment of taxable property. Property was only taxed, and

therefore assessed, for occupants whose income was greater than $1000. We have an extreme

censoring problem – even the median is censored – as many occupants had lower incomes and

so their value of capital is unobserved. Appendix Table 8, columns 5 – 8, report estimated

impacts on log capital per square foot when assigning a value of 500 to these missing values.53

There is some indication of increased capital values of commercial establishments (column 6),

but only after controlling for pre-Fire values and there is a differential trend in commercial

capital value from 1867 to 1872. In contrast to the analysis of land value, there is no sense

in which capital values in 1872 would already capitalize expected changes after 1872. There

is also some indication of higher residential capital value after the Fire (column 7), but not

after controlling for pre-Fire values.

In the end, there may be changes in building occupancy and capital investment that are

one channel through which building reconstruction generates economic gains and influences

neighbors. The estimates are sensitive to the empirical specification, however. In addition,

it is ultimately the replacement of buildings that drives changes in occupancy patterns (as

in Brueckner and Rosenthal, 2009). Thus, our interpretation generally focuses on spillovers

from higher quality buildings, with the understanding that these spillovers may operate in

part through “higher quality” occupants.

VIII Epilogue: Estimated Impacts in 2012

As an epilogue, we consider whether the burned area differs from unburned areas in the

modern period. We use data on Boston property values from plot assessments in 2012, which

are assessed at market value.54 Separate valuations for land and buildings are unavailable

for condominiums, which make up a substantial portion of the downtown Boston area, so

we are limited to analyzing the total value of plots.

52These results should be interpreted with caution, however, as there was a pre-Fire trend towards lower residential density. In contrast to the analysis of land value, there is no sense in which 1872 residential density already capitalizes future changes in residential density.

53Capital values of 500 are among the lower common values, and the estimates are similar when assigning values of 50 or 100 that are among the lowest values observed.

54For details on assessment methodology, see: http://www.cityofboston.gov/assessing/assessedvalues.asp. We assigned plot locations by merging on plot ID to the Boston parcels map: http://boston.maps.arcgis.com.

32

Appendix Table 9 reports changes from 1872 to 2012 in the burned area, relative to

changes in the unburned area. There is no statistically significant difference in the basic

specification (column 1), but the burned area appears to become substantially more valuable

conditional on controls for plots’ pre-Fire characteristics (column 2). The influence of pre-

Fire controls is somewhat surprising, as we expected plot characteristics in 1867 and 1872

to have little predictive power in 2012 data. The estimates are smaller, and statistically

insignificant, when limiting the sample to areas within 1000 feet of the burned boundary.

In principle, there may continue to be spillover effects on unburned areas, though a variety

of confounding factors would be difficult to control for over such a long period of time.

There is no indication that the burned area was disadvantaged over the long-run, though

we suggest caution in interpreting these results as evidence of long-run gains. The most

plausible channel for long-run impacts is perhaps persistent changes in plot boundaries, but

we found only small impacts on plot size. Given cross-sectional differences between the

burned area and unburned areas, the identification assumption of parallel trends becomes

increasingly tenuous over longer periods of time.

IX Conclusion

Following the 1872 Great Fire of Boston, burned plots and nearby unburned plots experi-

enced substantial increases in land value. These increased land values capitalize substantial

economic gains, which we attribute to neighborhood spillover effects from the simultaneous

reconstruction of many nearby buildings. Changes in building values, by distance to the Fire

boundary and by quantile, are consistent with predicted impacts of neighborhood spillover

effects on building reconstruction. By contrast, individual building fires had no impact on

land value and generally smaller impacts on building value. While our data provide vari-

ous indications of cross-plot spillover effects, there is less evidence for substantial impacts

through increased plot sizes, increased urban density, or increased industry agglomeration.

The Fire temporarily mitigated substantial inefficiencies in urban growth, as individual

landowners fail to internalize cross-plot externalities. Indeed, in this case, burned buildings’

value was entirely offset by increased land values. While “urban renewal” is almost always

associated with poorer urban areas, these results indicate the potential for substantial land-

use inefficiencies in even wealthy and growing cities. Our within-Boston empirical analysis

is unable to quantify all spillover effects at the city level, though positive spillover impacts

by distance to the Fire boundary appear to dissipate within the sample region. Even if some

spillover effects are negative, increased land values imply at least large relative gains in the

burned area from the opportunity for simultaneous reconstruction at higher levels of building

quality.

33

The historical growth of American cities generally occurred in the absence of major city

fires, but the example of Boston highlights the potential magnitude of inefficiency within

a generally successful urban area. The Boston Fire provides a clear exogenous shock, and

the opportunity to collect a rich dataset covering private reconstruction responses in the

absence of modern regulatory constraints. Our empirical analysis focuses on plot-level dif-

ferences within Boston, in contrast to long-run economic analyses of city-level impacts after

destruction. While all data are drawn from one city, the statistical inference is robust to

correcting for spatial correlation across plots.

Our focus on wealthy areas of Boston is complementary to research on the economic

impacts of large-scale urban renewal in poor or declining areas. The Fire’s impacts are in-

dicative of substantial inefficiencies in even wealthy areas, and suggests that the study of

such frictions in today’s rapidly growing cities – both in the United States and in rapidly

urbanizing nations of the developing world – is an important area of research. The ability of

the Great Fire to generate substantial gains through simultaneous reconstruction also sug-

gests that policy interventions might achieve some of the same goals without the associated

destruction.

34

References

Anderson, Martin. 1964. The Federal Bulldozer: A Critical Analysis of Urban Renewal, 1949-1962. MIT Press.

Autor, David H., Christopher J. Palmer, and Parag A. Pathak. 2014. “Housing Market Spillovers: Evidence from the End of Rent Control in Cambridge, Massachusetts.” Journal of Political Economy, 122(3): 661–717.

Boston Archives. 1822-1918. Boston Tax Records. City of Boston. Available at the Office of City Clerk Archives and Records Management Division, 201 Rivermoor St., West Roxbury, MA 02132.

Brooks, Leah, and Byron Lutz. 2012. “From Today’s City to Tomorrow’s City: An Empirical Investigation of Urban Land Assembly.” George Washington University. Mimeo.

Brueckner, Jan K., and Stuart S. Rosenthal. 2009. “Gentrification and Neighborhood Housing Cycles: Will America’s Future Downtowns Be Rich?” The Review of Economics and Statistics, 91(4): 725–743.

Campbell, John Y., Stefano Giglio, and Parag Pathak. 2011. “Forced Sales and House Prices.” American Economic Review, 101(5): 2108–31.

Chen, Daniel L., and Susan Yeh. 2013. “The Impact of Government Power to Expro- priate on Economic Growth and Inequality.” Mimeo.

Collins, William J., and Katharine L. Shester. 2013. “Slum Clearance and Urban Re- newal in the United States.” American Economic Journal: Applied Economics, 5(1): 239– 273.

Conley, T.G. 1999. “GMM Estimation with Cross Sectional Dependence.” Journal of Econometrics, 92(1): 1–45.

Davis, Donald R., and David E. Weinstein. 2002. “Bones, Bombs, and Break Points: The Geography of Economic Activity.” American Economic Review, 92(5): 1269–1289.

Fire Commission. 1873. Report of the Commissioners Appointed to Investigate the Cause and Management of the Great Fire in Boston. Rockwell & Churchill, City Printers. http://quod.lib.umich.edu/m/moa/AAR8660.0001.001/1.

Fischel, William A. 2004. “An Economic History of Zoning and a Cure for Its Exclusionary Effects.” Urban Studies, 41(2): 317–340.

Fowler, William Worthington. 1873. Fighting Fire. Dustin, Gilman and Co.

Frothingham, Frank E. 1873. The Boston Fire. Lee and Shepard, Publishers.

Guerrieri, Veronica, Daniel Hartley, and Erik Hurst. 2013. “Endogenous Gentrifica- tion and Housing Price Dynamics.” Journal of Public Economics, 100: 45–60.

35

G.W. Bromley and Co. 1883. Atlases of the City of Boston. Vol. 1. http://dome.mit.edu/handle/1721.3/47999.

Hartley, Daniel. 2010. “The Effect of Foreclosures on Nearby Housing Prices: Supply or Disamenity?” Federal Reserve Bank of Cleveland. Working Paper.

Ioannides, Yannis M. 2003. “Interactive Property Valuations.” Journal of Urban Eco- nomics, 53(1): 145–170.

Jacobs, Jane. 1961. The Death and Life of Great American Cities. Random House.

Libecap, Gary D., and Dean Lueck. 2011. “The Demarcation of Land and the Role of Coordinating Property Institutions.” Journal of Political Economy, 119(3): 426 – 467.

Lindert, Peter H., and Richard Sutch. 2006. “Consumer Price Indexes, for All Items: 1774-2003.” Historical Statistics of the United States, Earliest Times to the Present: Mil- lennial Edition, Table Cc1–2.

Mian, Atif R., Amir Sufi, and Francesc Trebbi. 2014. “Foreclosures, House Prices and the Real Economy.” University of Chicago Booth School of Business. Working Paper.

Miguel, Edward, and Grard Roland. 2011. “The Long-Run Impact of Bombing Viet- nam.” Journal of Development Economics, 96(1): 1–15.

Munch, Patricia. 1976. “An Economic Analysis of Eminent Domain.” Journal of Political Economy, 84(3): 473–498.

Ripley, Brian D. 1977. “Modelling Spatial Patterns.” Journal of the Royal Statistical Society. Series B (Methodological), 172–212.

Rosen, Christine. 1986. The Limits of Power: Great Fires and the Process of City Growth in America. Cambridge University Press.

Rossi-Hansberg, Esteban, Pierre-Daniel Sarte, and Raymond Owen. 2010. “Hous- ing Externalities.” Journal of Political Economy, 118(3): 485–535.

Sanborn Map Company. 1867-1895. Sanborn Map Company, Boston. http://sanborn.umi.com.

Schwartz, Amy Ellen, Ingrid Gould Ellen, Ioan Voicu, and Michael H Schill. 2006. “The External Effects of Place-Based Subsidized Housing.” Regional Science and Urban Economics, 36(6): 679–707.

Sims, David P. 2007. “Out of Control: What Can We Learn from the End of Massachusetts Rent Control?” Journal of Urban Economics, 61(1): 129–151.

Siodla, James. 2013. “Razing San Francisco: The 1906 disaster as a natural experiment in urban redevelopment.” Colby College. Mimeo.

36

Siodla, James. 2014. “Making the Move: The Impact of the 1906 San Francisco Disaster on Firm Relocations.” Colby College. Mimeo.

Turner, Matthew A., Andrew Haughwout, and Wilbert van der Klaauw. 2014. “Land Use Regulation and Welfare.” Econometrica, 82(4): 1341–1403.

Wermiel, Sara E. 2000. The Death and Life of Great American Cities. The Johns Hopkins University Press.

Wilson, James Q. 1966. Urban Renewal: The Record and the Controversy. MIT Press.

37

38   

Figure 1. Historical Downtown Boston, the Burned Area, and Sample Plot Locations

Notes: The shaded red area was burned during the 1872 Great Fire of Boston, and this is overlaid on the land mass of downtown Boston in 1867 (Sanborn Map Company). Small black points denote each geo-located plot in our main sample for 1867.

39   

Figure 2. Simulated Steady-State Evolution of the Building Quality Distribution Panel A. Benchmark Model

Panel B. Benchmark Model: Unburned and Burned Areas after a Great Fire in Period 0

Notes: Panel A shows the simulated evolution of the steady-state in our benchmark model. The thick central line shows the mean of log building quality, and the upper/lower thin lines show the max/min of the building quality distribution. The thin dashed line shows the evolution of a particular example building. Panel B shows the simulated evolution of the steady-state in our benchmark model, following a Great Fire in period 0. The black lines represent the Unburned Area (same as in Panel A), and the red lines represent the Burned Area.

40   

Figure 3. Simulated Steady-State Evolution of the Building Quality Distribution Panel A. Extended Model: Burned Area after a Great Fire in Period 0

Panel B. Extended Model: Nearby Unburned Area after a Great Fire in Period 0

Notes: For our extended model with neighborhood externalities, Panel A shows the Burned Area following a Great Fire in period 0 (red lines) and the steady-state in the absence of a Fire (black lines). Panel B shows the steady-state in Nearby Unburned Areas (blue lines) and the steady-state in the absence of a Fire (black lines).

41   

Figure 4. Simulated Steady-State Evolution of Land Value after a Great Fire

Notes: Following a Great Fire in period 0, the upper red line shows the simulated evolution of land value in the Burned Area in our extended model with neighborhood externalities. The middle blue line shows the evolution of land value in a Nearby Unburned Area in our extended model. The lower black line shows the simulated evolution of land value in the Burned Area (or Unburned Area) in our benchmark model, which is the same as if there had been no Fire.

42   

Figure 5. Estimated Changes in Land Value from 1872 to 1873, by Distance to the Fire Boundary (in Feet)

Notes: For the indicated distance from the boundary of the burned area, each circle reports the estimated change in land value from 1872 to 1873 (and the vertical lines reflect 95% confidence intervals). The omitted category is plots more than 2900 feet from the burned area. Negative distances reflect areas within the burned area, and burned plots more than 400 feet from the Fire boundary are grouped together. The empirical specification includes controls for plots’ predicted land value in 1867 and 1872 based on block average and nearest neighbor.

-. 1

0 .1

.2 .3

.4 E

st im

a te

d C

oe ff

ic ie

n t

an d

9 5%

C o

nf id

en ce

I n

te rv

al

-500 0 500 1000 1500 2000 2500 3000 Distance in Feet from the Burned Area Boundary

43   

Figure 6. Estimated Impacts on Building Value, by Distance to the Fire Boundary (in Feet) Panel A. Estimated Impacts in 1873

Panel B. Estimated Impacts in 1882

Panel C. Estimated Impacts in 1894

Notes: For the indicated distance from the boundary of the burned area, each circle reports the estimated change in land value from 1872 to the indicated year (by Panel). The vertical lines reflect 95% confidence intervals. The omitted category is plots more than 2900 feet from the burned area. Negative distances reflect areas within the burned area, and burned plots more than 400 feet from the Fire boundary are grouped together. The empirical specification includes controls for plots’ predicted land value in 1867 and 1872 based on block average and nearest neighbor.

-3 -2

-1 0

E st

im at

ed C

oe ffi

ci en

t a nd

9 5%

C on

fid en

ce In

te rv

al

-500 0 500 1000 1500 2000 2500 3000 Distance in Feet from the Burned Area Boundary

-. 5

0 .5

1 1

.5 2

E st

im at

ed C

oe ffi

ci en

t a n

d 9

5% C

on fid

en ce

I nt

er va

l

-500 0 500 1000 1500 2000 2500 3000 Distance in Feet from the Burned Area Boundary

-. 5

0 .5

1 1.

5 2

E st

im at

ed C

oe ffi

ci en

t a nd

9 5%

C on

fid en

ce In

te rv

al

-500 0 500 1000 1500 2000 2500 3000 Distance in Feet from the Burned Area Boundary

44   

Figure 7. Estimated Impacts on Building Value in the Burned Area, by Quantile

Panel A. Estimated Quantile Effects in 1882

Panel B. Estimated Quantile Effects in 1894

Notes: Each Panel reports estimated impacts on the distribution of log building value, for that year relative to 1872.

-. 5

0 .5

1 1.

5 E

st im

at ed

C oe

ffi ci

en t a

nd 9

5% C

on fid

en ce

In te

rv al

0 20 40 60 80 100 Quantile

-. 5

0 .5

1 1.

5 E

st im

at ed

C oe

ffi ci

en t a

nd 9

5% C

on fid

en ce

In te

rv al

0 20 40 60 80 100 Quantile

Table 1. Plot Values in 1872, and Differences in the Burned Area

Burned Area Unburned Area

Difference in 1872:

(1) - (2)

Difference in Changes:

1867 to 1872 Restricted

Unburned Area

Difference in 1872:

(1) - (3)

Difference in Changes:

1867 to 1872

(1) (2) (3) (4) (5) (6) (7)

Land Value $13.95 $8.44 0.774*** -0.174*** $14.69 0.087 -0.165***

per Square Foot (6.77) (8.83) (0.084) (0.041) (11.32) (0.093) (0.045)

Building Value $7.48 $4.14 0.733*** 0.182 $5.94 0.246** 0.136

per Square Foot (4.35) (3.77) (0.097) (0.120) (4.75) (0.106) (0.127)

Number of Plots 580 6013 1837

Total Plot Area 1,724,877 10,642,991 3,753,481

Differences in Logs: Differences in Logs:

Burned vs. Unburned Burned vs. Restricted Unburned

Notes: For the indicated outcome variable, columns 1 and 2 report the average value across plots in the burned area and unburned area, respectively. Column 3 reports the estimated log difference in 1872 for plots in the burned area, relative to plots in the unburned area. Column 4 reports this estimated log difference in changes from 1867 to 1872, i.e., the difference in the burned area in 1872 (relative to the unburned area in 1872) relative to the difference in the burned area in 1867 (relative to the unburned area in 1867). Columns 5 to 7 correspond to columns 2 to 4, but for a restricted sample of plots within 1000 feet of the Fire boundary. All means and regressions are weighted by plot size. Robust standard errors clustered by block are reported in parentheses: *** indicates statistical significance at the 1% level, ** at the 5% level and * at the 10% level.

45

Table 2. Estimated Impact on Land Values in Burned Area, Relative to 1872

(1) (2) (3) (4) (5) (6) (7) (8)

1867 x Burned 0.174*** 0.019 - - 0.165*** 0.016 - -

(0.041) (0.013) ( ) ( ) (0.045) (0.014) ( ) ( )

1872 x Burned 0 0 0 0 0 0 0 0

( ) ( ) ( ) ( ) ( ) ( ) ( ) ( )

1873 x Burned 0.149*** 0.169*** 0.168*** 0.172*** 0.125*** 0.124*** 0.131*** 0.133***

(0.020) (0.020) (0.017) (0.018) (0.023) (0.022) (0.020) (0.021)

1882 x Burned 0.157*** 0.137*** 0.139*** 0.144*** 0.059 0.073 0.052 0.083*

(0.043) (0.044) (0.040) (0.042) (0.049) (0.049) (0.044) (0.046)

1894 x Burned -0.102* -0.147** -0.172*** -0.145** -0.250*** -0.196*** -0.234*** -0.188**

(0.056) (0.061) (0.056) (0.060) (0.069) (0.073) (0.067) (0.073)

Controls:

Year Fixed Effects X X X X X X X X

Year FE x Pre-Fire Block Average X X X X

Year FE x Pre-Fire Neighbor Value X X X X

R-squared 0.153 0.797 0.934 0.938 0.116 0.689 0.885 0.888 Number of Plots 31302 31302 31302 31302 11367 11367 11367 11367

Log Value of Land per Square Foot

Full Sample Restricted Sample

Notes: For all specifications, the outcome variable is the log value of land per square foot. From estimating equation 1 in the text, column 1 reports the estimated difference between plots in the burned area and plots in the unburned area in the indicated year, relative to the omitted year of 1872. From estimating equation 2 in the text, columns 2 to 4 include controls for plots' predicted characteristics prior to the Fire, based on their block and/or nearest neighbor (which is generally that same plot in the earlier year). Columns 5 to 8 correspond to columns 1 to 4, but for the restricted sample of plots within 1000 feet of the Fire boundary. The regressions are weighted by plot size. Robust standard errors clustered by block are reported in parentheses: *** indicates statistical significance at the 1% level, ** at the 5% level and * at the 10% level.

46

Table 3. Estimated Total Impact of Fire on Land Values in 1873 and 1882

Distance Burned Unburned Total Ratio of (4) to

Cutoff Area Area Impact Burned Building Value

(1) (2) (3) (4) (5)

In 1873:

Panel A. Estimated Cutoff 1,394 5,545 9,666 15,211 1.18

(125) (536) (1,150) (1,632) (0.13)

Panel B. 1149 Foot Cutoff 1,149 5,305 8,039 13,343 1.03

( ) (502) (728) (1,229) (0.10)

Panel C. 1639 Foot Cutoff 1,639 5,735 11,133 16,869 1.31

( ) (535) (994) (1,529) (0.12)

In 1882:

Panel D. Estimated Cutoff 1,412 7,236 12,561 19,797 1.53

(189) (1,313) (2,562) (3,781) (0.29)

Panel E. 1040 Foot Cutoff 1,040 6,749 9,408 16,157 1.25

( ) (1,287) (1,714) (3,001) (0.23)

Panel F. 1784 Foot Cutoff 1,784 7,376 14,894 22,270 1.73 ( ) (1,328) (2,572) (3,899) (0.30)

Impact in $1000's of 1872 Dollars:

Notes: Panels A to C consider the total effect on land value in 1873, adjusted to 1872 dollars using the David-Solar CPI (Lindert and Sutch 2006). From estimating equation 3 in the text, we constrain the impact of the Fire to be constant within the Burned Area, declining linearly in the Unburned Area until some distance cutoff, and then zero after that distance cutoff (i.e., fitting the points in Figure 5). Column 1, panel A, reports the estimated distance cutoff after which geographic spillover effects are zero. Column 1, panels B and C, report alternative assumed distance cutoffs. Column 2 reports the estimated total impact of the Fire on land value in the Burned Area, Column 3 reports the estimated total impact of the Fire on land value in the Unburned Area, and Column 3 reports the estimated total impact of the Fire in all areas. Column 5 reports the ratio of the estimates in Column 4 to the total 1872 value of buildings in the Burned Area. Panels D to E report analagous estimates, but for the impact on land value in 1882 (converted to 1872 dollars). Robust standard errors clustered by block are reported in parentheses.

47

Table 4. Estimated Impact on Building Values in Burned Area, Relative to 1872

(1) (2) (3) (4) (5) (6) (7) (8)

1867 x Burned -0.182 -0.053 - - -0.136 -0.043 - -

(0.120) (0.052) ( ) ( ) (0.127) (0.059) ( ) ( )

1872 x Burned 0 0 0 0 0 0 0 0

( ) ( ) ( ) ( ) ( ) ( ) ( ) ( )

1873 x Burned -1.803*** -1.881*** -1.961*** -2.016*** -1.800*** -1.890*** -1.965*** -2.011***

(0.161) (0.160) (0.167) (0.168) (0.161) (0.160) (0.167) (0.171)

1882 x Burned 0.401*** 0.478*** 0.637*** 0.511*** 0.357*** 0.402*** 0.493*** 0.441***

(0.067) (0.069) (0.056) (0.055) (0.070) (0.066) (0.050) (0.049)

1894 x Burned 0.174** 0.371*** 0.546*** 0.410*** 0.090 0.203** 0.274*** 0.246***

(0.078) (0.087) (0.068) (0.080) (0.089) (0.081) (0.066) (0.069)

Controls:

Year Fixed Effects X X X X X X X X

Year FE x Pre-Fire Block Average X X X X

Year FE x Pre-Fire Neighbor Value X X X X

R-squared 0.108 0.474 0.775 0.788 0.163 0.467 0.735 0.743 Number of Plots 30198 30198 30198 30198 10595 10595 10595 10595

Log Value of Building per Square Foot

Restricted SampleFull Sample

Notes: For all specifications, the outcome variable is the log value of building per square foot. From estimating equation 1 in the text, column 1 reports the estimated difference between plots in the burned area and plots in the unburned area in the indicated year, relative to the omitted year of 1872. From estimating equation 2 in the text, columns 2 to 4 include controls for plots' predicted characteristics prior to the Fire, based on their block and/or nearest neighbor (which is generally that same plot in the earlier year). Columns 5 to 8 correspond to columns 1 to 4, but for the restricted sample of plots within 1000 feet of the Fire boundary. The regressions are weighted by plot size. Robust standard errors clustered by block are reported in parentheses: *** indicates statistical significance at the 1% level, ** at the 5% level and * at the 10% level.

48

Table 5. Estimated Impact of Fire: Great Fire vs. Individual Fires

Full Sample Restricted Sample Full Sample Restricted Sample (1) (2) (3) (4)

1873 x Burned -1.950*** -1.944*** 0.170*** 0.129***

(0.173) (0.178) (0.018) (0.022)

1882 x Burned 0.514*** 0.445*** 0.142*** 0.080*

(0.059) (0.053) (0.042) (0.046)

1894 x Burned 0.413*** 0.247*** -0.156*** -0.200***

(0.083) (0.072) (0.060) (0.072)

~7 Months After Individual Fire -0.127 -0.005 -0.054 -0.019

(0.131) (0.028) (0.062) (0.042)

~10 Years After Individual Fire 0.346** 0.128* 0.084 -0.008

(0.152) (0.068) (0.102) (0.156)

~22 Years After Individual Fire 0.012 -0.013 -0.210 -0.205

(0.085) (0.083) (0.269) (0.298)

Test of Equality of Individual Fire and Great Fire Effects (p-value):

~7 Month Interval 0.000 0.000 0.001 0.002

~ 10 Year Interval 0.299 0.000 0.606 0.600

~ 22 Year Interval 0.000 0.003 0.848 0.988

Controls:

Year Fixed Effects X X X X

Year FE x Pre-Fire Block Average X X X X

Year FE x Pre-Fire Neighbor Value X X X X

R-squared 0.788 0.744 0.938 0.889 Number of Plots 30128 10525 31219 11284

Log Value of Building per Sqr. Ft. Log Value of Land per Sqr. Ft.

Notes: The reported estimates are from equation 4 in the text, which jointly estimates the impact of the 1872 Great Fire and the impact of individual building fires. The first three rows report the estimated impacts of the 1872 Great Fire in 1873, 1882, and 1894 (corresponding to estimates in Table 2 and Table 4), and the second three rows report the impact of individual building fires after approximately 1 year, 10 years, and 22 years. Below, we report the statistical significance of the difference between the Great Fire impact and the corresponding individual fire impact. Columns 1 and 2 report impacts on building value, corresponding to columns 4 and 8 of Table 4. Columns 3 and 4 report impacts on land value, corresponding to columns 4 and 8 of Table 2. Note that this sample excludes plots in the 1872 Burned Area that also experienced individual building fires, as well as individual building fires that were suspected to be arson. Robust standard errors clustered by block are reported in parentheses: *** indicates statistical significance at the 1% level, ** at the 5% level and * at the 10% level.

49

Table 6. Estimated Impact on Plot Sizes in Burned Area, Relative to 1872

(1) (2) (3) (4) (5) (6) (7) (8)

1867 x Burned -0.069 - -0.063 - -0.094 - -0.089 -

(0.043) ( ) (0.043) ( ) (0.064) ( ) (0.064) ( )

1872 x Burned 0 0 0 0 0 0 0 0

( ) ( ) ( ) ( ) ( ) ( ) ( ) ( )

1873 x Burned 0.006 -0.014 -0.001 -0.021 0.061*** 0.050** 0.055** 0.040

(0.023) (0.025) (0.024) (0.026) (0.021) (0.023) (0.023) (0.025)

1882 x Burned 0.090* 0.067** 0.094* 0.057 0.156*** 0.137*** 0.160*** 0.126***

(0.046) (0.033) (0.047) (0.038) (0.055) (0.036) (0.056) (0.040)

1894 x Burned 0.088* 0.029 0.023 0.011 0.165*** 0.091** 0.100 0.067

(0.051) (0.036) (0.057) (0.041) (0.061) (0.044) (0.066) (0.045)

Controls:

Year Fixed Effects X X X X X X X X

Year FE x Pre-Fire Block Average X X X X

Year FE x Pre-Fire Neighbor Value X X X X

R-squared 0.058 0.818 0.074 0.805 0.039 0.819 0.056 0.811 Number of Plots 31353 31353 11381 11381 30340 30340 10368 10368

Notes: For all specifications, the outcome variable is the log number of square feet per plot. From estimating equation 1 in the text, column 1 reports the estimated difference between plots in the burned area and plots in the unburned area in the indicated year, relative to the omitted year of 1872. From estimating equation 2 in the text, column 2 include controls for plots' predicted characteristics prior to the Fire, based on their block and nearest neighbor (which is generally that same plot in the earlier year). Columns 3 and 4 correspond to columns 1 and 2, but for the restricted sample of plots within 1000 feet of the Fire boundary. Columns 5 to 8 correspond to columns 1 to 4, but excluding plots that had land taken for the widening of roads (Appendix Figure 4). The regressions are unweighted. Robust standard errors clustered by block are reported in parentheses: *** indicates statistical significance at the 1% level, ** at the 5% level and * at the 10% level.

Log Plot Size

Plots Unaffected by Road Widening

All Plots Restricted Sample All Plots Restricted Sample

50

X Appendix

This Appendix has four sections. In the first section, we provide additional details on the

estimation of the Fire’s total impact on land value and alternative functional forms for

parameterizing geographic spillover effects. The second section provides additional details

on our measurement of within-industry agglomeration using Ripley’s L function. The third

section contains the Appendix Figures, and the fourth section contains the Appendix Tables.

X.A Estimating the Total Impact of the Great Fire on Land Value

In our baseline specification of the total impact of the Great Fire on land value, we model

impacts on both the burned area and nearby buildings using a piecewise linear function.

Using the log of land value per square foot in 1873 as the outcome variable Yit, we estimate

the following function via non-linear least squares:

Yit = β0 + β1 max

{ c − distit

c , 0

} + µtȲ

near i + ηtȲ

block i + εit,(5)

where distit is the distance of point i from the burned area. 55 Here, β̂1 represents the

estimated effect of the fire, and ĉ the estimated cut-off point beyond which the fire has no

further effects.

We calculate each plot’s predicted log land value per square foot Ŷit using coefficients

estimated from equation 5 above, as well as the value without the fire: Ỹit = β̂0 + µ̂tŶi1867 +

η̂tŶi1872. The impact of the fire on land values, combining the burned area and the spillover

zone, is then:

Vt = ∑

distit<c

plotsizeit

( exp

( Ŷit

) − exp

( Ỹit

)) As a robustness check in the estimation of the spillover effects, we experimented with

three alternative formulas for the spillover function:

1. A variant allowing the spillover to be non-linear:

Yit = β0 + β1 max

{( c − distit

c

)γ , 0

} + µtŶi1867 + ηtŶi1872 + εit

2. An asymptotic variant with no cut-off:

Yit = β0 + β1

( 1

1 + β2distit

)γ + µtŶi1867 + ηtŶi1872 + εit

55This distance is zero for points within the burned area.

1

3. A polynomial with no cut-off:

Yit = β0 + 4∑

n=1

βndist n it + µtŶi1867 + ηtŶi1872 + εit

Appendix Figure 9 shows these four estimated functions, including the baseline linear

model, which are all fairly similar in approximating the relationship apparent in Figure

5. The three non-linear specifications estimate the mean in the burned area to be slightly

greater than the baseline model, and generally show the spillover effects continuing into

more distant parts of the city. Divergent properties of polynomials are visible past 3000 feet,

where the third function turns back upwards, as only 6.3% of the sample lies beyond 3000

feet from the burned area.

The first and second alternative models allow us to generate alternative estimates of

the Fire’s total impact, as they include an estimate of where spillover effects end. The

first model estimates a total impact of $16 million (with a standard error of $3.8 million),

which is only slightly larger than our baseline estimate. Estimates are less similar with the

second alternative formula, as we must assume the Fire’s spillover effects disappear only

when distance goes to infinity, and the estimated impact is $124 million. Identification

of the the second alternative model is tenuous, however, as the within-sample functional

form is used to project impacts on distances far out of sample. Our baseline estimates are

more conservative, assuming that the spillover effects go to zero at some cutoff within the

sample region. Within the sample region, all four functional forms provide a broadly similar

parameterization of the basic relationship seen in Figure 5.

X.B Measuring Agglomeration

We measure within-industry agglomeration using Ripley’s L function. For industry i with

Nib establishments in an area b with square footage Ab, let λib be the sample estimate of the

density of establishments per square foot: λib = Nib/Ab. The value of Lib for radius r, is

then given by:

Lib (r) =

√√√√λ−1ib Ni∑ k=1

Nib∑ j=1,j 6=k

I [d (k,j) < r] /πNi − r,(6)

where I [d (k,j) < r] is an indicator function equal to one if firms k and j are within distance r of each other. Higher values of Lib are associated with greater agglomeration. A

value of Lib equal to −r is associated with complete dispersion (i.e, no establishments in

2

industry i have other establishments from industry i with r feet). A value of Lib equal to√ Ab/π − r is associated with complete agglomeration (i.e., all establishments in industry i

are within r feet of each other).

To mitigate “edge effects,” we do not consider firms within r feet of the sample boundary

in the outer summation, indexed by k, in equation 6. These firms near the boundary are

included as potentially being part of clusters of firms near the non-boundary firms and are

included in the j-indexed inner summation. Similarly, firms across the boundary of the

burned area are counted as potentially being part of the cluster of firms on the other size of

the boundary. Edges of the sample area that intersect with the ocean or the Boston Common

(a large park) are not counted as boundaries since firms near these edges chose to locate in

spots where the potential for agglomeration was naturally limited.

3

4   

Appendix Figure 1. Total Assessed Value of Boston Real Estate and Personal Property

Notes: The upper line reflects the total assessed value of real estate, and the lower line reflects the total assessed value of property from the City of Boston’s assessment record books (Boston Tax Records). All values are converted to constant 1872 dollars using the David-Solar CPI (Lindert and Sutch 2006). The vertical line denotes the year of the Boston Fire.

5   

Appendix Figure 2. Plot-level Map with Detailed Information and Geo-Located Points

Notes: Detailed plot-level maps, such as the one above, are georeferenced to the Boston-wide map. These detailed maps often provide the plots’ square footage and owner name. The overlaid red dots correspond to each plot and are assigned to particular tax assessment records.

6   

Appendix Figure 3. Sample Plot Locations in Each Subsequent Year

Panel A. Plot Locations in 1872

Panel C. Plot Locations in 1882

Panel B. Plot Locations in 1873

Panel D. Plot Locations in 1894

Notes: As in Figure 7, each point reflects one geo-located plot in our main sample for the indicated year. These points are overlaid on a map of Boston in 1867 and the area burned in 1872 (as in Figure 1).

7   

Appendix Figure 4. Post-Fire Changes in Boston Roads

Notes: This map of the burned district indicates areas (shaded in pink) that were purchased by the City for road widening and the creation of Post Office Square.

8   

Appendix Figure 5. Plot Assessed Value vs. Plot Sale Price

Notes: Log Assessed Value is plotted against Log Sale Price for a sample of 88 plots: 16 plots in the burned area (black) and 72 plots outside the burned area (gray). Plot observations are hollow diamond shapes when observed before the Boston Fire, and solid diamonds when observed after the Fire. Log Assessed Value comes from our tax assessment database, and Log Sale Price is from Boston’s Registry of Deeds. Plots are shown against the 45-degree line.

9   

Appendix Figure 6. Estimated Differences in Values in 1867, Relative to 1872, by Distance to the Fire Boundary (in Feet) Panel A. Estimated Differences in Land Value

Panel B. Estimated Differences in Building Value

Notes: For the indicated distance from the boundary of the burned area, each circle reports the estimated impact on value in 1867 relative to 1872 (e.g., positive coefficients represent a decline from 1867 to 1872). Panel A presents estimates for the log value of land per square foot, and Panel B presents estimates for the log value of building per square foot. The specification does not include controls for plots’ pre-Fire outcomes. The omitted category is plots more than 2900 feet from the burned area. Negative distances reflect areas within the burned area, and burned plots more than 400 feet from the Fire boundary are grouped together.

-. 4

-. 2

0 .2

.4 E

st im

at ed

C oe

ffi ci

en t

an d

95 %

C o

nf id

en ce

I nt

er va

l

-500 0 500 1000 1500 2000 2500 3000 Distance in Feet from the Burned Area Boundary

-1 .5

-1 -.

5 0

.5 E

st im

at ed

C oe

ffi ci

en t

an d

95 %

C o

nf id

en ce

I nt

er va

l

-500 0 500 1000 1500 2000 2500 3000 Distance in Feet from the Burned Area Boundary

10   

Appendix Figure 7. Estimated Changes in Land Value by Distance to the Fire Boundary (in Feet) Panel A. Estimated Impact in 1882, Relative to 1872

Panel B. Estimated Impact in 1894, Relative to 1872

Notes: For the indicated distance from the boundary of the burned area, Panel A reports estimated changes from 1872 to 1882 and Panel B reports estimated changes from 1872 to 1894 (each circle reports the point estimate and the vertical lines reflect 95% confidence intervals). The omitted category is plots more than 2900 feet from the burned area. Negative distances reflect areas within the burned area, and burned plots more than 400 feet from the Fire boundary are grouped together. The empirical specification includes controls for plots’ predicted land value in 1867 and 1872 based on block average and nearest neighbor.

-. 4

-. 2

0 .2

.4 .6

.8 E

st im

at ed

C oe

ffi ci

en t

an d

95 %

C o

nf id

en ce

I nt

er va

l

-500 0 500 1000 1500 2000 2500 3000 Distance in Feet from the Burned Area Boundary

-. 4

-. 2

0 .2

.4 .6

.8 E

st im

at ed

C oe

ffi ci

en t

an d

95 %

C o

nf id

en ce

I nt

er va

l

-500 0 500 1000 1500 2000 2500 3000 Distance in Feet from the Burned Area Boundary

11   

Appendix Figure 8. Estimated Impacts on Land Value in the Burned Area, by Quantile Panel A. Estimated Quantile Effect in 1873

Panel B. Estimated Quantile Effect in 1882

Panel C. Estimated Quantile Effect in 1894

Notes: Each Panel reports estimated impacts on the distribution of log land value, for that year relative to 1872.

-. 8

-. 6

-. 4

-. 2

0 .2

.4 E

st im

at ed

C oe

ffi ci

en t a

nd 9

5% C

on fid

en ce

In te

rv al

0 20 40 60 80 100 Quantile

-. 8

-. 6

-. 4

-. 2

0 .2

.4 E

st im

at ed

C oe

ffi ci

en t a

nd 9

5% C

on fid

en ce

In te

rv al

0 20 40 60 80 100 Quantile

-. 8

-. 6

-. 4

-. 2

0 .2

.4 E

st im

at ed

C oe

ffi ci

en t a

nd 9

5% C

on fid

en ce

In te

rv al

0 20 40 60 80 100 Quantile

12   

Appendix Figure 9. Functional Forms for Estimating Total Impact on Land Value

Notes: The baseline model shows the estimated functional form, based on equation 3 in the text, which parameterizes the results shown in Figure 5. Alternative models 1 to 3 report alternative estimated functional forms, as described in Section A of the Appendix.

.1 .2

.3 .4

.5

-1000 0 1000 2000 3000 4000 Distance from burned area

Baseline Model Alternative 1 Alternative 2 Alternative 3

Appendix Table 1. Average Log Sale Value Minus Log Assessed Value After Fire: Before Fire: Difference:

1882 and 1894 1867 and 1872 (2) - (1) (2) (1) (3)

Burned Area -0.042 0.083 -0.125

[0.297] [0.162] (0.119)

Unburned Area -0.143 0.030 -0.173

[0.631] [0.312] (0.124)

Difference 0.102 0.054 0.048 (0.151) (0.078) (0.169)

Notes: Based on data from Boston's Registry of Deeds, matched to our tax assessment database, cells report the average log difference in sale price and assessed value of plots (sale price - assessed value). Column 1 reports estimates from after the Fire (in 1882 and 1894), and Column 2 reports estimates from before the Fire (in 1867 and 1872). Row 1 reports estimates in the Burned Area, and Row 2 reports estimates in the Unburned Area. Standard deviations are reported in brackets. Row 3 reports the difference in the Burned Area, relative to the Unburned Area; and Column 3 reports the difference after the Fire, relative to before the Fire. Column 3, row 3, reports the difference-in-difference estimate. The sample includes 72 plots in the Unburned Area, and 16 plots in the Burned Area. Robust standard errors are reported in parentheses: *** indicates statistical significance at the 1% level, ** at the 5% level and * at the 10% level.

13

Appendix Table 2. Main Results with Conley Standard Errors at Varying Cutoffs

Full Sample Sample Full Sample Sample

(1) (2) (3) (4)

1873 x Burned 0.172 0.133 -2.016 -2.011

Clustered by Block (0.018) (0.021) (0.168) (0.171)

250 foot cutoff (0.019) (0.020) (0.164) (0.167)

750 foot cutoff (0.021) (0.025) (0.257) (0.259)

1,250 foot cutoff (0.022) (0.027) (0.247) (0.247)

1,750 foot cutoff (0.018) (0.022) (0.208) (0.209)

1882 x Burned 0.144 0.083 0.511 0.441

Clustered by Block (0.042) (0.046) (0.055) (0.049)

250 foot cutoff (0.039) (0.041) (0.059) (0.058)

750 foot cutoff (0.058) (0.061) (0.065) (0.062)

1,250 foot cutoff (0.064) (0.068) (0.051) (0.048)

1,750 foot cutoff (0.058) (0.064) (0.045) (0.039)

1894 x Burned -0.145 -0.188 0.410 0.246

Clustered by Block (0.060) (0.073) (0.080) (0.069)

250 foot cutoff (0.054) (0.062) (0.076) (0.071)

750 foot cutoff (0.094) (0.115) (0.096) (0.095)

1,250 foot cutoff (0.112) (0.133) (0.081) (0.079)

1,750 foot cutoff (0.109) (0.133) (0.081) (0.066)

Controls:

Year Fixed Effects X X X X

Year FE x Pre-Fire Block Average X X X X

Year FE x Pre-Fire Neighbor Value X X X X

R-squared 0.987 0.991 0.902 0.934 Number of Plots 31302 11367 30198 10595

Land Value Building Value

Log Value per Square Foot

Notes: The reported coefficients correspond exactly to those reported in Table 2 and Table 4: column 1 corresponds to Table 2, column 4; column 2 corresponds to Table 2, column 8; column 3 corresponds to Table 4, column 4; and column 4 corresponds to Table 4, column 8. For each coefficient, alternative standard errors are reported based different assumed distance cutoffs in the estimation of Conley standard errors (Conley 1999): 250 feet, 750 feet, 1,250 feet, and 1,750 feet. As a basis of comparison, we also report our main standard errors that are clustered by block.

14

Appendix Table 3. Main Results, Unweighted Specifications

Full Restricted Full Restricted

Sample Sample Sample Sample (1) (2) (3) (4)

1873 x Burned 0.192*** 0.152*** -1.693*** -1.695***

(0.020) (0.021) (0.158) (0.166)

1882 x Burned 0.147*** 0.091* 0.543*** 0.494***

(0.048) (0.052) (0.058) (0.051)

1894 x Burned -0.116* -0.102 0.480*** 0.377***

(0.064) (0.074) (0.064) (0.060)

Controls:

Year Fixed Effects X X X X

Year FE x Pre-Fire Block Average X X X X

Year FE x Pre-Fire Neighbor Value X X X X

R-squared 0.944 0.904 0.806 0.771 Number of Plots 31302 11367 30198 10595

Log Value per Square Foot

Land Value Building Value

Notes: The reported specifications correspond to those reported in Table 2 and Table 4, but not weighting the regressions by plot size. Robust standard errors clustered by block are reported in parentheses: *** indicates statistical significance at the 1% level, ** at the 5% level and * at the 10% level.

15

Appendix Table 4. Main Results Excluding Plots With Road Widening

Full Restricted Full Restricted

Sample Sample Sample Sample (1) (2) (3) (4)

1873 x Burned 0.148*** 0.108*** -1.852*** -1.841***

(0.019) (0.023) (0.202) (0.211)

1882 x Burned 0.100** 0.040 0.439*** 0.374***

(0.046) (0.048) (0.051) (0.047)

1894 x Burned -0.192*** -0.239*** 0.353*** 0.178**

(0.067) (0.078) (0.097) (0.081)

Controls:

Year Fixed Effects X X X X

Year FE x Pre-Fire Block Average X X X X

Year FE x Pre-Fire Neighbor Value X X X X

R-squared 0.937 0.890 0.784 0.739 Number of Plots 30289 10354 29320 9717

Land Value Building Value

Log Value per Square Foot

Notes: The reported specifications correspond to those reported in Table 2 and Table 4, but the sample excludes plots that lost land for road widening (Appendix Figure 4). Robust standard errors clustered by block are reported in parentheses: *** indicates statistical significance at the 1% level, ** at the 5% level and * at the 10% level.

16

Appendix Table 5. Number of Unique Owners and Number of Plots, by Burned and Unburned Areas

Burned Unburned Burned Unburned Burned Unburned Burned Unburned

(1) (2) (3) (4) (5) (6) (7) (8)

Panel A. Full Sample

1867 402 3,534 620 6,120

1872 367 3,390 -1.74 -0.81 580 6,013 -1.29 -0.35

1873 346 3,401 -5.72 0.32 519 5,970 -10.52 -0.72

1882 322 3,287 -0.77 -0.37 486 5,504 -0.71 -0.87

1894 309 3,097 -0.34 -0.48 465 5,076 -0.36 -0.65

2012 112 1,964 -0.64 -0.52

Panel B. Restricted Sample

1867 402 1261 620 1911

1872 367 1160 -1.74 -1.60 580 1837 -1.29 -0.77

1873 346 1177 -5.72 1.47 519 1808 -10.52 -1.58

1882 322 1108 -0.77 -0.65 486 1693 -0.71 -0.71

1894 309 971 -0.34 -1.03 465 1462 -0.36 -1.14

2012 112 439 -0.64 -0.59

Number of Plots Annual Percent ChangeNumber of Owners Annual Percent Change

Notes: Columns 1 and 2 report the number of unique owner names in the burned area and unburned area, respectively. Columns 3 and 4 report the annual percent change from the period before in the number of unique owners. Columns 5 and 6 report the number of individual land plots in the burned area and unburned area, and columns 7 and 8 report the annual percent change in this number from the period before. Note that 8 of the 19 owner decline between 1872 and 1873, and 20 of the 61 plots eliminated between 1872 and 1873, were a direct consequence of road changes (Appendix Figure 4).

17

Appendix Table 6. Industry-by-Industry Changes in Agglomeration (Ripley's L Function, 100 foot radius)

Obs. 1872 1882 1894 1872 1882 1894 1872 to 1882 1872 to 1894 Industry (1) (2) (3) (4) (5) (6) (7) (8) (9)

Shoes 297 215 143 189 229 555 376 -398 -173

Leather 159 171 178 185 222 1167 1100 -937 -864

Clothes 112 93 154 166 153 134 238 80 -13

Liquors 110 224 287 -100 199 202 185 60 -310

Dry Goods 107 108 101 -100 276 380 283 -111 -214

Hats 107 169 412 204 200 233 231 210 3

Tailor 88 311 457 142 211 350 295 7 -254

Machinery 50 248 130 28 498 331 223 49 54

Hardware 48 62 118 221 422 276 254 203 327

Jewelry 48 571 648 703 373 553 638 -103 -133

Printer 48 78 99 105 283 197 221 107 89

Fancy Goods 46 140 -100 318 161 -100 414 21 -75

Teams 45 26 -11 24 210 329 -100 -156 308

Kitchen Goods 37 87 216 -100 181 500 289 -190 -295

Cigars 35 318 318 -100 98 235 188 -137 -509

Paper 34 145 169 111 351 115 219 260 98

Clothing Accessories 18 152 412 142 627 264 289 624 328 Cotton 13 165 71 -100 -100 659 -100 -853 -265

Notes: For the 18 most common identifiable industries, column 1 reports the number of times that industry is observed in 1872. Columns 2 to 4 report agglomeration index values for that industry in the burned area in 1872, 1882, and 1890. Higher values correspond to greater agglomeration: these values are generated by Ripley's L function with a distance radius of 100 feet, and refer to Section B of the Appendix for details. Columns 5 to 7 report estimates for the unburned area in 1872, 1882, and 1894. Column 8 reports the change from 1872 to 1882 in the burned area, relative to the change in the unburned area; Column 9 reports the change from 1872 to 1894 in the burned area, relative to the change in the unburned area.

Clustering Index

Burned Area Unburned Area

Difference-in-Difference

Burned vs. Unburned

18

Appendix Table 7. Estimated Impacts on Industrial Agglomeration, Relative to 1872

(1) (2) (3) (4) (5) (6)

1867 x Burned -4.9 - -35.3 - -69.0 -

(65.7) ( ) (62.1) ( ) (87.0) ( )

1872 x Burned 0 0 0 0 0 0

( ) ( ) ( ) ( ) ( ) ( )

1882 x Burned -23.9 -62.5 -160.8 -148.6* -236.6 -156.1*

(69.4) (38.3) (115.3) (82.1) (177.8) (87.4)

1894 x Burned -33.6 -106.4*** -161.5* -187.5** -209.6 -194.2*

(41.7) (31.4) (84.1) (87.7) (155.6) (96.5)

Controls:

Year Fixed Effects X X X X X X

Year FE x Industry L Value in 1867 X X X

Year FE x Industry L Value in 1872 X X X

R-squared 0.199 0.68 0.136 0.433 0.114 0.431 Industry-by-Year Observations 144 144 144 144 144 144

Ripley's L Function

Radius = 50 ft. Radius = 100 ft. Radius = 200 ft.

Notes: For these estimates, the unit of observation is an industry-year pair in the burned area or unburned area. For each industry-year, its level of agglomeration is calculated using Ripley's L Function for a distance radius of 50 feet for columns 1 and 2, 100 feet for columns 3 and 4 (as shown in Appendix Table 6), or 200 feet for columns 5 and 6. Refer to Section B of the Appendix for details on this formula. As in the main estimating equations, each column then reports differences in the burned area relative to the unburned area for the indicated year, relative to differences in 1872. Columns 2, 4, and 6 include controls for that industry's level of agglomeration in 1867 and 1872. Standard errors are reported in parentheses: *** indicates statistical significance at the 1% level, ** at the 5% level and * at the 10% level.

19

Appendix Table 8. Estimated Impacts on Occupant Density and Value of Capital, Relative to 1872

(1) (2) (3) (4) (5) (6) (7) (8)

1867 x Burned 0.146** - 0.362*** - -0.616*** - -0.029 -

(0.067) ( ) (0.067) ( ) (0.218) ( ) (0.054) ( )

1872 x Burned 0 0 0 0 0 0 0 0

( ) ( ) ( ) ( ) ( ) ( ) ( ) ( )

1873 x Burned -0.294*** -0.368*** -0.327*** -0.202*** -4.438*** -3.752*** -0.073 -0.174***

(0.043) (0.052) (0.064) (0.069) (0.225) (0.218) (0.051) (0.053)

1882 x Burned 0.268*** 0.251*** -0.404*** -0.328*** 0.047 1.253*** 0.258*** -0.070

(0.066) (0.075) (0.072) (0.088) (0.199) (0.185) (0.078) (0.094)

1894 x Burned 0.340*** 0.289*** -0.283*** -0.158 -0.331 0.997*** 0.209** -0.246**

(0.066) (0.077) (0.080) (0.097) (0.225) (0.203) (0.090) (0.113)

Controls:

Year Fixed Effects X X X X X X X X

Year FE x Pre-Fire Block Average X X X X

Year FE x Pre-Fire Neighbor Value X X X X

R-squared 0.02 0.534 0.053 0.559 0.121 0.65 0.044 0.681 Number of Plots 31353 31353 31353 31353 31353 31353 31353 31353

Log Value of Capital per Square Foot (Assigning 500 to Zero Values of Capital)

Commercial Residential

Notes: In columns 1 to 4, the outcome variable is the number of assessed occupants per 1,000 square feet (commercial occupants for columns 1 and 2, and residential occupants for columns 3 and 4). In columns 5 to 8, the outcome variable is the log value of capital per square foot. The value of capital is censored for many observations, as discussed in the text, and we assign a capital value of 500 to all missing values (after summing across all occupants in that plot). For all observations, we then divide by the square footage and take logs. The estimating equations are otherwise as before. From estimating equation 1 in the text, the odd columns report the estimated difference between plots in the burned area and plots in the unburned area in the indicated year, relative to the omitted year of 1872. From estimating equation 2 in the text, the even columns include controls for plots' predicted characteristics prior to the Fire, based on their block and nearest neighbor (which is generally that same plot in the earlier year). The regressions are weighted by plot size. Robust standard errors clustered by block are reported in parentheses: *** indicates statistical significance at the 1% level, ** at the 5% level and * at the 10% level.

Number of Assessed Occupants per 1,000 Square Feet

Commercial Residential

20

Appendix Table 9. Estimated Impact on Land and Building Value in 2012

(1) (2) (3) (4)

1867 x Burned 0.081 - 0.069 -

(0.050) ( ) (0.054) ( )

1872 x Burned 0 0 0 0

( ) ( ) ( ) ( )

2012 x Burned 0.123 0.569*** 0.108 0.266

(0.217) (0.207) (0.233) (0.209)

Controls:

Year Fixed Effects X X X X

Year FE x Pre-Fire Block Average X X

Year FE x Pre-Fire Neighbor Value X X

R-squared 0.842 0.928 0.863 0.932 Number of Plots 15382 15382 5491 5491

Full Sample Restricted Sample

Log Total Value per Square Foot

Notes: For all specifications, the outcome variable is the log total value of land and buildings per square foot. From estimating equation 1 in the text, column 1 reports the estimated difference between plots in the burned area and plots in the unburned area in the indicated year, relative to the omitted year of 1872. From estimating equation 2 in the text, column 2 includes controls for plots' predicted characteristics prior to the Fire, based on their block and nearest neighbor. Columns 3 and 4 correspond to columns 1 and 2, but for the restricted sample of plots within 1000 feet of the Fire boundary. The regressions are weighted by plot size. Robust standard errors clustered by block are reported in parentheses: *** indicates statistical significance at the 1% level, ** at the 5% level and * at the 10% level.

21