University of Southern California Dissertations and Theses

Location choice and the costs of climate change

(USC Thesis Other)

Location choice and the costs of climate change

PDF

Download a page range

Download transcript

Copy asset link

Request this asset

Transcript (if available)

Content Location Choice and the Costs of Climate Change
by
Qi Qi Amanda Ang
A Dissertation Presented to the
FACULTY OF THE USC GRADUATE SCHOOL
UNIVERSITY OF SOUTHERN CALIFORNIA
In Partial Fulfillment of the
Requirements for the Degree
DOCTOR OF PHILOSOPHY
(ECONOMICS)
May 2024
Copyright 2024 Qi Qi Amanda Ang

Acknowledgements
I would like to thank my family, who have made huge sacrifices in order for me to study
in the United States. To my advisors - Matt, Paulina and Andrii, thank you for your patience and accompaniment throughout these years. To my friends and family at Westfield
Residence and Aster Study Center and my friends in the PhD program, especially Jingyi
Fang and Amy Mahler, I couldn’t have done this without you.

Table of Contents
Acknowledgements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
List of Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
List of Figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
Abstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
Chapter 1: Population Growth and Wildfire Mitigation . . . . . . . . . . . . . . . 9
1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.2 Population Density and Wildfire Ignitions . . . . . . . . . . . . . . . . . . . 14
1.2.1 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
1.3 Model of Internal City Structure . . . . . . . . . . . . . . . . . . . . . . . . . 22
1.3.1 Working Households . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
1.3.2 Retired Households . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
1.3.3 Firms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
1.3.4 Developers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
1.3.5 Market Clearing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
1.3.6 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
1.3.7 Quantification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
1.3.8 Model Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
1.3.9 Estimating Fréchet Elasticity . . . . . . . . . . . . . . . . . . . . . . . 32
3

1.3.10 Estimating Agglomeration Externalities . . . . . . . . . . . . . . . . 33
1.3.11 Model Inversion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
1.4 Decomposing Climate- and Population-Driven Wildfires . . . . . . . . . . . 35
1.5 Corrective Tax . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
1.6 Building Restrictions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
1.7 Social Cost of Wildfires . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
1.8 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
Chapter 2: Social Cost of Flooding . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
2.2 Dataset . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
2.3 Measuring the impact of urban flooding on travel times . . . . . . . . . . . 54
2.4 Estimating damages . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
2.4.1 Robustness Checks . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
Chapter 3: Amenities in QSE Models . . . . . . . . . . . . . . . . . . . . . . . . . . 67
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
3.2 Residential Amenity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
3.3 Dataset . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
3.4 Empirical Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
3.5 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
4

List of Tables
1.1 Share of Ignitions and Burned Area by Cause (2000-2018) . . . . . . . . . . 15
1.2 Check Exogeneity of Industry Shares w.r.t. Wildfire Ignitions . . . . . . . . 19
1.3 Estimated Relationship Between Wildfire Ignitions and Population Density 20
1.4 External and Estimated Parameters . . . . . . . . . . . . . . . . . . . . . . . 32
1.5 Estimation of Gravity Equation . . . . . . . . . . . . . . . . . . . . . . . . . 33
1.6 Cost of Wildfires Through Expected PM 2.5 Exposure and Model Welfare . 42
2.1 São Paulo Household Survey Descriptive Statistics . . . . . . . . . . . . . . 49
2.2 Queried Trips Descriptive Statistics . . . . . . . . . . . . . . . . . . . . . . . 51
2.3 Effect of flooding on trip duration without accounting for non-random
flood reporting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
2.4 Effect of accumulated rain on flood duration . . . . . . . . . . . . . . . . . . 57
2.5 Effect of predicted flood duration on trip duration . . . . . . . . . . . . . . 58
2.6 Estimated Average Added Travel Time . . . . . . . . . . . . . . . . . . . . . 59
2.7 Estimated annual damages (2017 USD) . . . . . . . . . . . . . . . . . . . . . 60
2.8 Probability of taking a trip on a day with flooding Pooled observations from
2012 and 2017 travel survey . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
2.9 Probability of taking a trip in an hour with flooding Pooled observations
from 2012 and 2017 travel survey . . . . . . . . . . . . . . . . . . . . . . . . 63
3.1 Correlation Coefficients . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
3.2 R Sq. Values (Ranked in Descending Order) . . . . . . . . . . . . . . . . . . 75
3.3 Regression Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
5

3.4 R Sq. Decomposition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
6

List of Figures
1.1 Predicted Wildfire Risk Reduced Form Results . . . . . . . . . . . . . . . . 18
1.2 Projected Changes in Av. Temperature . . . . . . . . . . . . . . . . . . . . . 36
1.3 Corrective Tax . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
2.1 Dataset Construction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
2.2 Comparing Survey and Queried Trips . . . . . . . . . . . . . . . . . . . . . . 52
2.3 Repeated Observation of One Survey Trip . . . . . . . . . . . . . . . . . . . 53
2.4 Alternative Routes Generated Using Google Directions API . . . . . . . . . 65
2.5 Difference between Re-routing and Flood Effect . . . . . . . . . . . . . . . . 66
3.1 Calibrated Residential Amenities (Los Angeles Metropolitan Area) . . . . . 70
7

Abstract
This dissertation comprises three essays examining household location choice and the
cost of climate change. The first essay identifies an inverted U-shaped relationship between population density and wildfires in the US. I conduct a case study of Los Angeles County, and find that risk-proportional taxation is an effective means for addressing
higher wildfire risk. The second essay assesses the economic impact of flood-induced
travel delays in São Paulo, Brazil, estimating an annual cost of $125 million USD, equivalent to 0.06 % of the city’s GDP. The third essay utilizes a spatial equilibrium model to
analyze local residential amenity measures, informing infrastructure investment strategies and managed retreat considerations in areas susceptible to natural disasters.
8

Chapter 1
Population Growth and Wildfire Mitigation
1.1 Introduction
Wildfires are enormously costly events in the US and around the world, both in terms
of capital damages as well as the cost to human health from exposure to wildfire smoke
(Burke et al., 2021; Wang et al., 2020; Hsiao, 2021). Increased residential density in the
wildland-urban interface has been put forward as a possible explanation for an increase
in wildfire ignitions (Radeloff et al., 2018). Between 2000 and 2018, human-started wildfires accounted for 84.9 % of all wildfires and 37 % of total area burned (US Forest Service).
In this paper, I ask: What is the causal impact of population density on wildfires? How
does wildfire risk change as households move in response to climate change? What are
the impacts of policies which try to move people away from the firezone?
The answer to the first question is not empirically straightforward, since the determinants of residential choices could be correlated with unobserved factors that also affect
wildfire risk. The ideal experiment in answering this question would involve randomly
assigning more people to areas with same baseline wildfire risk and evaluating differences in wildfire outcomes between the two locations. In my empirical strategy, I use
9

plausibly exogenous productivity shocks as a predictor of local population density. I find
that the relationship between wildfires and population density has an inverted U-shape,
where places of intermediate density have the highest incidence of wildfire ignitions. In
the contiguous US, the likelihood of wildfire peaks at 27 % at a population density of
113 people per squared kilometer and in the state of California, the likelihood of wildfire
peaks at 32 % at a population density of 2,885 people per squared kilometer.
The incidence of wildfire risk due to population density is an example of an environmental externality. Individuals select residence location based on private benefits and
costs; they do not account for their contribution to wildfire risk. Such choices elevate the
social costs linked to wildfires, such as higher ignition probabilities and increased smoke
exposure. One potential policy to mitigate wildfire costs is reducing population density
in high-risk areas. However, many metropolitan areas in the US face the twin problems
of lack of affordable housing and increasing wildfire risk (Moretti, 2017). Ospital (2023)
finds that relaxing land use regulation will induce people to move away from risky areas
and thereby reduce the present discounted cost of wildfires by 10 percent. However, if
human activity is also a cause of wildfire ignitions, then moving people to places of lower
wildfire risk might result in more ignitions in those locations.
In this paper, I capture the tension between the forces which determine housing supply and the role of population density in producing wildfires using a quantitative spatial
model (Ahlfeldt et al., 2015). As people move into or out of a location, wages and housing
prices change; that in turn will affect the number of people in the location until we reach
an equilibrium.
10

I make a methodological contribution to this class of spatial models by incorporating different types of households who might make location choices differently and are differentially affected by wildfires. Unlike the standard model outlined in Ahlfeldt et al. (2015),
which focuses solely on commuting households within a city, my model encompasses
both commuting and retired households. This is important in the state of California since
there are currently 6 million individuals aged 65 and above (15 % of population) and retired households are expected to grow three times as fast as overall population by 2060,
with some counties experiencing more than 150 % growth in retired households1
. Commuting households seek attractive locations with good labor market access. In contrast,
retired households, don’t commute, receive a fixed income and are attracted to different
kinds of amenities. Their vulnerability to environmental shocks makes them significant
in environmental economics literature. Moreover, we expect their share of total population to increase in the coming decades and so it becomes crucial to understand their
preferences and location choices.
This model allows me to generate counterfactual distributions of population density under different climate scenarios and policy environments. Combining these predictions
with my wildfire risk equation produces location-specific wildfire probability predictions that account for the effects of density. This approach allows us to separate wildfires
caused solely by temperature changes from those driven by population density.
Wildfire damage encompasses various components, including expenses related to fire
suppression, structural destruction, and exposure to PM 2.5 particles. In this study, I explore the repercussions of ignitions occurring in particularly sensitive locations. I establish a connection between expected wildfire ignitions and the location-specific potential
for emitting PM 2.5. Utilizing data from the Missoula Fire Lab Emissions Inventory, I
1Facts of California’s Elderly - Data and Reports | CA Dept. of Aging, Accessed on 10/27/2023.
11

calculate the potential PM 2.5 emissions for each location in the event of a fire. For each
counterfactual scenario, I am able to generate a counterfactual distribution of expected
PM 2.5 emissions. Finally, I use a pollution transport model (InMAP) to estimate the
costs of these emissions as they spread from the initial location. These estimates serve as
a conservative lower bound for the damages attributable to wildfires. Currently, PM 2.5
exposure is not factored into the overall assessment of losses caused by natural disasters2
.
This paper speaks to the literature on spatial exposure to environmental shocks. The closest paper is Ospital (2023). Ospital (2023) builds a detailed quantitative spatial model in
order to address how zoning regulations drive people into risky locations and estimates
the welfare improvement associated with relaxing zoning restrictions. The main contribution of my paper is accounting for the relationship between population density and
wildfire risk, whereas Ospital (2023) takes wildfire risk as a disamenity for households
but which are not directly affected by them. Ospital (2023) finds that there is a 10 % welfare improvement associated with relaxing zoning restrictions and moving households
away from risky locations. I evaluate a scenario where housing supply is fixed in highrisk locations and unconstrained elsewhere. I find that this does not improve wildfire
outcomes because of migration into the city. However, combining a tax which is proportional to wildfire risk and building restrictions mitigates the effect of population growth
on the cost of wildfires.
The existing literature on spatial exposure to natural hazards focuses on flood risk or
sea level rise (Balboni, 2019; Cruz and Rossi-Hansberg, 2021; Desmet et al., 2018). This
study, however, underscores a characteristic of wildfires which sets them apart from other
natural disasters. Unlike many other calamities where human activity can adapt or mitigate damage after the event, the incidence of wildfires is directly influenced by human
2Billion Dollar Disasters | Calculating the Costs | National Centers for Environmental Information
(NCEI), Link here, Accessed on 10/07/2023.
12

actions. From an urban economic perspective, we can think of wildfire risk as an environmental cost of urban sprawl (Kahn, 2000; Bento et al., 2005; Glaeser and Kahn, 2010;
Mangum, 2017) or as an endogenous disamenity (Almagro and Domínguez-Iino, 2022;
Khanna et al., 2020). In this paper, households increase wildfire risk as they move into
places with low to intermediate population density and each location’s riskiness depends
on the number of households living in that location.
Finally, I contribute to the literature on the economics of wildfires in addressing the
urban economic forces which drive households to locate in fire risk zones. Baylis and
Boomhower (2023) quantify the implicit subsidy received by households in low-density,
fire-prone areas, given that the expenses for fire suppression are covered by federal or
state entities rather than the households themselves. I integrate these cost estimates with
the externalities of population density on wildfire risk, and examine the implications of
levying a fire suppression tax on households. Conditional on locating in a fireprone area,
mandated fire protection from building codes lower the likelihood of property destruction due to fire (Baylis and Boomhower, 2022). My focus is on the choice of residing
in these risky areas and exploring policy alternatives that could encourage relocation.
Borgschulte et al. (2022) quantify the decline in productivity due to elevated PM 2.5 from
wildfire smoke. My research investigates how shifts in population density contribute to
PM 2.5 emissions by influencing the frequency of wildfire occurrences.
This paper is structured as follows. In Section 1.2 I develop the main wildfire function
which I use to link population density and wildfire probability. I then lay out the theoretical model which connects local characteristics, labor and housing markets to household
location choices (Section 1.3). I use this model to evaluate how fire risk changes in the following counterfactual scenarios: (1) household migration in response to climate change
(Section 1.4) (2) corrective tax that would recover the cost of fire suppression (Section 1.5)
13

(3) restricting housing supply in high risk locations (Section 1.6). I calculate the costs associated with these counterfactual scenarios in terms of PM 2.5 emissions (Section 1.7).
Section 1.8 concludes.
1.2 Population Density and Wildfire Ignitions
Human-induced fires significantly influence wildfire outcomes, as shown in Table 1.1. An
increasing number of households live in high fire risk locations, exacerbating the wildfire risk (Syphard et al., 2007; Radeloff et al., 2018; Keeley et al., 2021). Figure ?? shows
population growth in the most fire-prone areas. I aim to quantify the causal effect of
population density on wildfires to create a "wildfire production function". This function
should include causal estimates and have strong predictive power, because I will then use
it to predict how wildfire probabilities change in response to counterfactual changes in
density.
14

US CA
Share Share Share Share
Ignitions Burned Area Ignitions Burned Area
Causes
Human 84.9 % 37.3 % 86.3 % 67.4 %
Natural 15.1 % 62.7 % 13.6 % 32.6 %
Human Causes
Open Burning 40.7 % 19.5 % 18.0 % 6.4 %
Arson 22.5 % 25.2 % 19.4 % 15.9 %
Equipment & Vehicle Use 13.7 % 24.8 % 34.6 % 35.3 %
Recreation 6.9 % 12.2 % 8.6 % 20.0 %
Misuse of fire by a minor 4.4 % 1.1 % 7.1 % 2.0 %
Smoking 4.3 % 1.9 % 6.7 % 0.4 %
Power Generation 2.7 % 8.9 % 3.4 % 12.2 %
Table 1.1: Share of Ignitions and Burned Area by Cause (2000-2018)
Notes: Data on wildfire ignitions and burned area by cause are taken from Short (2021). Human causes
of wildfires are defined by the US National Wildfire Coordinating Group. I exclude fires for which the
cause has been classified by the US Forest Service as “missing data / not specified / undetermined.” These
fires make up 8 % of all fires recorded in this dataset. In the first two columns, I calculate each share out
of all the identified natural and human-caused wildfires in the US, and in the third and fourth columns I
calculate each share out of all the identified natural and human-caused wildfires in the state of California.
In the bottom panel, I calculate each share out of all human-caused wildfires in the US / in the state of
California.
Simply regressing the number of ignitions on population density is problematic due to
endogeneity and reverse causality: households prefer areas with natural amenities which
are correlated with fire risk, and the incidence of fires themselves may influence subsequent neighborhood migration patterns. I obtain causal estimates by including PUMA
15

fixed effects3 and using a shift-share variable to instrument for population density. I get
predictive power by allowing the relationship between density and wildfires to have a
different slope based on baseline wildfire risk categories.
For grid cell i in year t,
Wildfiresit =β0 + β1Densityit + β2Density2
it + β3 HotDaysit + β
′
SeasonalPrecipit + β
′ WHPi
+ β
′ WHPi × Densityit + β
′ WHPi × Density2
it + λP UMAi
+ γt + ϵit
(1.1)
We might be concerned that wildfire ignitions also drive residential location decisions,
or that location decisions and wildfires are both related to an unobserved omitted variation. I use a shift-share instrumental variable approach to isolate the variation in residential density from local productivity shocks. In order for the estimate to be causal, the
shift-share variable must be highly predictive of residential density and it must not affect
ignitions through any other mechanism except that of residential density. I control for
all time-variant factors within a PUMA and only look at how changes in population in a
given PUMA across time affect the number of ignitions in that PUMA.
I construct the shift-share variable using grid-level share of employment by industry
(NHGIS IPUMS) in a baseline year 2000 and changes in national employment by industry taken from the County Business Patterns dataset (US Census). I exclude location j’s
3PUMA fixed effects are based off model locations used in Parkhomenko and Delventhal (2023). The
contiguous US is divided into 4504 locations based on PUMAs and counties. This allows comparability
between large metropolitan areas and smaller rural counties.
16

county from constructing the growth variable so as to avoid violating the exclusion restriction. I construct the instrumental variable as follows, for grid cell i, industry k and
year t:
Densityit = β0 + β1
X
k
Shareik × Growthkt + ϵit (1.2)
We might be concerned that employment shares themselves affect wildfires because certain industries e.g. construction or agriculture are more likely to be located in places
with high fire risk or produce output that is more likely to lead to ignitions. I regress
employment shares on ignitions and find that, conditional on being in a given PUMA,
employment shares do not have a statistically significant relationship with ignitions (see
Table 1.2).
Because ignitions are rare outcomes and the data contains many zeros, I use PseudoPoisson Maximum Likelihood to estimate equation (1) and find that a 1 % increase in
population density is associated with a 0.704 % increase in ignitions. In order to construct the wildfire production function I use in my model, I run the regression separately
for the state of California, and I find that a 1 % increase in population density is associated with a 1.319 % increase in wildfire risk. These results indicate an inverted U-shape
relationship between population density and wildfires. For the contiguous US, the predicted probability of a wildfire occurring is 26.6 % before the inflection point, reaching a
peak at 27.0 % and decreases to 26.9 % after the inflection point. In the state of California, the probability of a wildfire occurring is 30.3 % before the inflection point, reaching
a peak at 32.9 %, and decreases to 31.7 % after the inflection point.
In order to execute the IV with the PPML estimator, I use a control function approach as
recommended by Lin and Wooldridge (2019). I initially estimate equation (2) to generate
predicted residuals, which are then included along with the endogenous variable in the
17

second-stage regression. Table 1.3 shows the OLS and IV results. Figure 1.1 shows predicted and observed wildfire outcomes for the whole US and the state of California. I use
the coefficients in column (4) in following sections for generating the expected number
of wildfires associated with counterfactual distributions of population. Figure ?? shows
the change in wildfire probability associated with a uniform 10 % increase in population
density across the whole state of California. In Section 1.3, I show how we can use a
model to generate counterfactual distributions of population density which we can then
use to generate counterfactual predictions of wildfire probability.
Figure 1.1: Predicted Wildfire Risk
Reduced Form Results
Notes: This figure shows the predicted number of wildfire events from the estimation of (3) and (4) in
Table 1.3. Results were aggregated using binned log population density for clarity. Panels A and C show
observed and predicted wildfire events for the whole sample (contiguous US) and for the state of California.
Panels B and C show the predicted number of wildfire events for high- and low-risk locations. In the main
estimation, we have 5 wildfire risk categories. In this plot, I group categories 1 and 2 into the “low risk”
group and categories 3 and 4 into the “high risk” group.
18

Outcome: No. of Ignitions (2000)
PPML PPML
(1) (2)
Industry shares
Agriculture 0.627 -0.730
(0.502) (0.500)
Construction 2.807∗∗∗ 1.384∗∗
(0.529) (0.495)
Manufacturing 0.475 0.693
(0.455) (0.420)
Wholesale Trade -0.484 -0.422
(0.675) (0.624)
Retail Trade 2.179∗∗∗ 0.740
(0.494) (0.460)
Transportation 1.779∗∗∗ 0.673
(0.536) (0.523)
Information -4.741∗∗∗ -0.207
(0.852) (0.769)
Finance, Insurance, Real Estate -3.218∗∗∗ -1.302∗
(0.590) (0.545)
Professional Services -2.938∗∗∗ -0.574
(0.567) (0.512)
Education, Health, Social Services 0.343 -0.0323
(0.471) (0.431)
Arts, Entertainment -0.691 -0.361
(0.514) (0.479)
PUMA FE × ✓
WHP ✓ ✓
Days above 90 deg. ✓ ✓
Seasonal precip. ✓ ✓
N 94462 43459
pseudo R
2 0.189 0.198
Standard errors in parentheses
∗ p < 0.05, ∗∗ p < 0.01, ∗∗∗ p < 0.001
Table 1.2: Check Exogeneity of Industry Shares w.r.t. Wildfire Ignitions
Notes: Standard errors are clustered at the PUMA level. Pseudo-R squared is reported for PPML regressions. Observations where the outcome is always 0 and the independent variable is always greater than 0
are dropped in (2) because the estimate does not exist.
19

Outcome: Number of Ignitions
US US US CA
OLS PPML PPML IV PPML IV
(1) (2) (3) (4)
Density (log) 0.00166 0.736*** 0.704*** 1.319***
(0.00434) (0.188) (0.193) (0.281)
Density (log) sq. 0.000114 -0.0783*** -0.0808*** -0.104***
(0.000396) (0.0184) (0.0176) (0.0253)
Baseline Fire Risk × Density (log)
(WHP = 1) × Density (log) 0.00161 -0.327 -0.341 -0.824**
(0.00433) (0.188) (0.194) (0.309)
(WHP = 2) × Density (log) 0.00733 -0.434* -0.413* -1.152***
(0.00449) (0.188) (0.193) (0.281)
(WHP = 3) × Density (log) 0.0121* -0.508** -0.458* -1.033***
(0.00487) (0.188) (0.193) (0.280)
(WHP = 4) × Density (log) 0.00420 -0.568** -0.517** -1.047***
(0.00478) (0.188) (0.193) (0.283)
Baseline Fire Risk ✓ ✓ ✓ ✓
Baseline Fire Risk × Density (log) sq. ✓ ✓ ✓ ✓
Days Above 90 deg. ✓ ✓ ✓ ✓
Seasonal Precip. ✓ ✓ ✓ ✓
Year FE ✓ ✓ ✓ ✓
PUMA FE ✓ ✓ ✓ ✓
N 7202220 7023090 7023090 557865
R
2 0.103 0.218 0.218 0.156
Standard errors in parentheses * p < 0.05, ** p < 0.01, *** p < 0.001
Table 1.3: Estimated Relationship Between Wildfire Ignitions and Population Density
Notes: Standard errors are clustered at the PUMA level. Pseudo-R squared is reported for PPML regressions. The number of observations changes between OLS and PPML because PPML omits observations
where fixed effects are not identified. See Appendix for results when these observations are included. I use
a control function approach and include the first stage residuals from equation (2) in specifications (3) and
(4) (Lin and Wooldridge, 2019). I use a inverse hyperbolic sine transformation for population density as an
equivalent to the log transformation since the dataset contains many zeros.
20

1.2.1 Data
Locations. Each observation in the reduced-form estimation is a 4 km x 4 km grid cell
covering the contiguous US, sourced from the UOregon PRISM dataset. I control for regional variations in population growth and wildfire risk using PUMA fixed effects, as
outlined by Parkhomenko and Delventhal (2023). These fixed effects divide the contiguous US into 4,504 model locations based on US Census PUMAs and counties, ensuring
comparability by avoiding juxtapositions of small rural counties with large metropolitan
areas.
Wildfires. Wildfire counts are taken from the US Forest Service’s Spatial Wildfire Occurrence dataset (Short, 2021). The dataset includes geographic coordinates for the ignition
location of each wildfire event, burn area, start date, date of containment and a classification of the cause of the fire. Baseline wildfire risk is taken from the US Forest Service
Wildfire Hazard Potential (WHP) dataset (Dillon et al., 2023). WHP is generated using a
fire simulation model that takes in data on fuels and vegetation as well as occurrences of
past fires. WHP can be expressed as an index where values range from 0 to 150, or divided into five categories: very low, low, moderate, high, and very high. For the purposes
of this paper, I group "high" risk and "very high" risk locations in the same category since
there are very few "very high" risk locations overall. Areas with higher WHP are more
likely to experience “torching, crowning, and other forms of extreme fire behavior under
conducive weather conditions” based on landscape data in 2014.
Population density. Population density is taken from the Worldpop dataset. Worldpop
takes in census data and produces fine-scale annual population data at the 1 km grid cell
level using a random forest model.
21

Temperature and precipitation. Daily maximum temperature and monthly total precipitation are taken from the UOregon PRISM dataset.
Shift-share variable. The shift-share instrumental variable is constructed using blockgrouplevel industry shares (14 industry groups) from the year 2000 and national industry-level
growth rates taken from the US Census County Business Patterns dataset.
1.3 Model of Internal City Structure
In the previous section, I developed a function which takes population density and weather
variables as inputs and gives us a prediction for the number of wildfire events in a given
location. Since wildfire risk depends on local population density, I am interested in how
wildfire risk evolves as local population changes. I use a quantitative spatial model to
generate counterfactual distributions of local population. Using a structural model allows us to account for general equilibrium forces which affect within-city migration: A
shock to the city induces a change in the probability that a household will live in a given
location. However, as households move across the city, local housing prices and wages
change which also affect the number of households living in that area.
Within this framework, I think of wildfire risk as an “endogenous environmental disamenity” which evolves as people make their location decisions within the city. The
current version of my model does not include a “feedback loop” between wildfire outcomes and location choices. In Ospital (2023)’s model, household residential amenity
is negatively correlated with the exogenous probability of wildfire occurring and in the
housing market, 10 % of the housing stock is lost when a wildfire occurs. McCoy and
Walsh (2018) suggest that fireprone properties become less desirable 2-3 years after an
event, but prices return to the baseline afterward. Borgschulte et al. (2022) suggest that
22

households do not move in response to exposure to PM 2.5 from wildfires. Issler et al.
(2020) find that 95 % of homes within an area burned by a wildfire are rebuilt. However,
this remains an open question in the literature and is worth evaluating in light of how
future fires may not resemble past fires.
For this model, I take Los Angeles County as my setting. Los Angeles County is one of the
most fireprone areas in the United States and also part of the second largest metropolitan
area in the US. I take the model from Ahlfeldt et al. (2015) as my starting point and extend
it to include both commuting and non-commuting households. I am interested in how
retired households might affect wildfire risk differently from working households for the
following reasons: (1) retired households are more likely to locate in high-risk locations
(Wibbenmeyer and Robertson, 2022) (2) retired households have different preferences for
local amenities compared to working households (Komissarova, 2022) (3) retired households are more vulnerable to environmental change compared to working households
and (4) retired households are not constrained by having to choose a suitable workplace
location. This reasons suggest that retired households make their location decisions in a
different way compared to working households and thus, might have a different impact
on wildfire outcomes. Furthermore, the population of people aged 65 and older is now
16.8 % of the total US population 4 and is projected to become 23 % of total US population in 2060 5
, making this a policy-relevant group of people to study. In Los Angeles
county, population above 65 is projected to grow to 37 % by the year 2060 (Hauer, 2019).
First, I construct a model of the city of Los Angeles using the U.S. Census LODES dataset,
which provides residence-workplace paired population counts. I can infer the amenity
42020 Census: 1 in 6 People in the United States Were 65 and Older, Source: https://www.census.
gov/library/stories/2023/05/2020-census-united-states-older-population-grew.html Accessed
on 10/03/2023.
5Demographic Turning Points for the United States: Population Projections for 2020 to 2060 (US
Census). Source: https://www.census.gov/content/dam/Census/library/publications/2020/demo/
p25-1144.pdf Accessed on 10/03/2023.
23

value of locations based on observed commuting patterns and housing costs: If people are willing to commute long distances or pay high housing prices in order to live
in a given location, the location has a high amenity value. After calibrating the model
with either observed data or parameters from the literature (see Table 1.4), I modify the
model inputs to generate relevant counterfactual scenarios.I obtain counterfactual population density distributions from the model’s equilibrium outcomes. These are linked to
location-specific wildfire probabilities using the "wildfire production function" derived
from the California-specific regression in Section 1.2. Utilizing historical emissions data,
I estimate the expected PM 2.5 emissions for each potential wildfire location. A pollution
transport model is then employed to assess wildfire-related mortality costs due to PM 2.5
exposure. Additionally, I quantify welfare changes by summing expected utility across all
city locations within the model’s framework. In the model, households choose where to
live (and work) by maximizing utility. Utility depends on local amenities, housing prices,
and for commuting households, wages and commuting costs. Working households choose
a residence and a workplace location whereas retired households only choose a residence
location. Working households receive a wage based on their workplace location whereas
retired households receive an exogenously determined level of income. I account for
location-specific residential agglomeration externalities that are type-specific. As more
working households move into a location, it becomes more attractive to other working
households and similarly for retired households. Prices and wages are determined by labor and housing market clearing. Households can migrate into and out of the city based
on how the expected utility of living in Los Angeles compares with the expected utility
of living in the rest of the US6
.
6Using the results from Hornbeck and Moretti (2018), I calibrate the LA-specific migration elasticity
such that a 1 % improvement in productivity increases the working population of LA by 4 %. Given the
agglomeration forces I use in my model, my estimate migration elasticity is 1.55. I assume that the expected
utility of living elsewhere in the US remains fixed across all counterfactual scenarios. In the model, 1
% higher average wage in the city leads to a 0.9 % lower ratio between retired and working households
(Komissarova, 2022).
24

1.3.1 Working Households
Households maximize utility by choosing consumption of a tradable final good c and a
non-tradable good (housing, h). I use a Cobb-Douglas utility function u where γ is the
share of utility that is determined by housing.
u(c, h) =
c
1 − γ
!1−γ

h
γ
!γ
Working households choose a residence and workplace location to maximize utility subject to a budget constraint. Solving the household’s maximization problem gives us the
following expression for indirect utility:
Vij =
XiEjwj
q
γ
i
exp(κtij)
where Xi
is a residential amenity, Ej
is the workplace amenity, wj
is wages associated
with workplacel location j and qi
is housing price. exp(κtij) is a measure of the cost of
commuting from i to j, including κ which governs the disutility of commuting and travel
time tij.
Each household receives a Fréchet-distributed preference shock ϵ. The distribution of location preferences depends on a scale parameter θ which governs how spread out households are across locations. I describe how θ is estimated in Section 1.3.7. Given the
Fréchet distribution of preferences F(ϵ) = exp(−ϵ
−θ
), we can express the probability that
a worker lives in i and works in j as follows:
πij =
V
θ
ij
P
i
′
j
′ V
θ
i
′
j
′
25

We can then obtain the number of working households living in i (NRi
) by summing
across all j locations and the number of households working in j (N Ej
) by summing
across all i locations.
NRi =
X
j
πij N Ej =
X
i
πij
1.3.2 Retired Households
Retired households choose a residence location i based on retired household specific
amenities X
R, exogenously determined income yj and housing prices qi
. Intuitively, retired households are like working households if tij = 0 (no commuting) and if all workplace locations give them the same level of income.Utility maximization gives us the following indirect utility function for retired households:
V
R
ij =
X
R
i
yj
q
γ
i
I assume both retired and working households receive preference shocks from the same
Fr’echet distribution. Given the Fréchet distribution of preferences F(ϵ) = exp(−ϵ
−θ
), we
can express the probability that a worker lives in i and given income level yj as follows:
π
R
ij =
(V
R
ij )
θ
P
i
′
j
′ (V
R
i
′
j
′)
θ
Assuming that each household receives the same income yj = y, the number of retired
households living in i (NRR
i
) is given by:
NRR
i =
X
j
π
R
ij =
(V
R
ij )
θ
P
i
′ (V
R
i
′ )
θ
26

1.3.3 Firms
Given local productivity Aj
, firms produce a final good Yj using labor N Ej and floorspace
HEj
, using the following production function:
Yj = Aj N Eα
j HE1−α
j
where α is the share of labor used in production. Firms are perfectly competitive and the
good is costlessly traded across locations. Firms choose labor and floorspace to maximize
profits. Together with the zero-profit condition, this gives us the following expression for
equilibrium wages:
wj = α A
1
α
j

1 − α
qj
! 1−α
α
1.3.4 Developers
Given construction productivity φi
, real estate developers produce floorspace H using
the final good K and land L using the following production function:
Hi = K
1−ηi
i
(φiLi
)
ηi
where ηi
is the share of land used in floorspace production. Developers are perfectly competitive. Each location has a maximum amount of available land Λi
, developers choose
land for floorspace production such that Li ≤ Λi
. Equilibrium supply of floorspace is
given by the following expression:
Hi = φi
(1 − ηi
)
1−ηi
ηi q
1−ηi
ηi
i
Li
where 1−ηi
ηi
is the housing supply elasticity and qi
is the equilibrium price of floorspace.
27

1.3.5 Market Clearing
Given model parameters {α,γ,κ,θ, ηi
} and exogenous location fundamentals {tij,Xi
,XR
i
,Ej
,Aj
,φi
},
an equilibrium is given by {NRi
,NRR
i
,N Ej
,Hi
,qi
,wj
} such that (1) labor market clears (2)
floorspace market clears (3) utility maximization holds for working and retired households and (4) profit-maximization and the zero-profit condition are satisfied for firms
and real estate developers.
The labor market clears when the number of workers in location j is equal to the number
of residents commuting to that location:
N Ej =
X
i
πij|i NRi =
X
i
Ej
(wj
/ exp(κtij))θ
P
s Es
(ws
/ exp(κtis))θ
NRi
where πij|i
is the probability of commuting to j conditional on living in i. Retired households indirectly affect wages through competition for housing. If a location is attractive
to both working and non-working households, some working households may not be able
to live in that location and thus have less access to labor market opportunities available
at that location.
The floorspace market clears when demand for floorspace is equal to supply of floorspace.
Residential floorspace demand HRi depends on wages in working locations, floorspace
prices and the number of households living in each location.
HRi =
γ (
P
j πij wj + π
R
ij yj
)
qi
28

Both working and retired households spend a constant share of their income γ on housing. Commercial floorspace HEj depends on local productivity, floorspace prices and the
number of people employed in that location.
HEj =

(1 − α)Aj
qj
!
1
α
N Ej
The market clears when total floorspace demand HRi +HEi
is equal to floorspace supply
Hi
, where developers choose the maximum amount of land available Li = Λi
.
HRi + HEi =
γ (
P
j πij wj + π
R
ij yj
)
qi
+

(1 − α)Ai
qi
! 1
α
N Ei = Hi = K
1−ηi
i
(φiΛi
)
ηi
1.3.6 Data
Commuting flows. I obtain the matrix of commuting flows between all census tracts in
LA county from the LEHD Origin-Destination Employment Statistics (LODES) database.
I take the average values between years 2012-2016 as a baseline measure of commuting
flows in LA county.
Commuting times. Time taken to commute between pairs of census tracts is taken
from Delventhal et al. (2022). Data is obtained using the Census Transportation Planning Products dataset.
Retired Households. Tract-level retired population is taken from IPUMS NHGIS (ACS
2006-2010, 2015-2019). I define as ‘retired’ people those who are above 65 years of age
and who are reported as “not in the labor force” in the American Community Survey. To
calibrate the ratio of total income of retired to working households, I compare the sum of
total personal income and retirement income of retired households with the total wage
income of working households.
29

Wages. I use blockgroup-level data from IPUMS NHGIS (ACS 2012-2016) and construct
a wage index following Parkhomenko and Delventhal (2023). First I regress the log of
median self-reported income per blockgroup b in census tract i on a set of observable
characteristics:
logIncomeb = β0 +
X
a
β2,a Age Binsa,b + β3 Share of Female Workersb +
X
r
β4,r Race Sharesr,b
+
X
k
β5,k Industry Sharesk,b +
X
o
β6,o Occupation Shareso,b + µi + ϵb
I then generate the tract-level wage index by taking the constant βˆ
0 and the tract-level
fixed effect µi
.
wi = βˆ
0 + µi
In the case where self-reported wage data are not available for a census tract, I use the
average wage index value of its neighboring census tracts.
Floorspace prices. I use tract-level data from IPUMS NHGIS (ACS 2012-2016) and construct an index of floorspace prices following Parkhomenko and Delventhal (2023). I
regress the log of median self-reported rent per census tract on a set of observable characteristics:
logRenti = β0 + β1 Roomsi + β2 Year Builti + β3 Number of Unitsi + ιi
where Roomsi
is the median number of rooms per dwelling, Year Builti
is the median
year built across all dwellings, and Number of Unitsi
is the modal number of units per
30

dwelling for each census tract i. I then generate the tract-level rent index by taking the
constant and the tract-level residual.
qi = βˆ
0 + ιˆi
In the case where self-reported rent data are not available for a census tract, I use the
average rent index value of its neighboring census tracts.
1.3.7 Quantification
I calibrate the model such that model-generated residential and employment counts are
equal to observed residential and employment counts using a standard procedure in the
urban economics literature (see Appendix). I take data from ACS 2015-2019 as the benchmark economy. I calibrate residential amenities separately for working and retired households. I focus my investigation on Los Angeles County and define model locations as the
set of 2,341 census tracts within the county7
.
1.3.8 Model Parameters
Table 1.4 shows the parameters used in the model. I describe how I estimate the Frechét
elasticity θ in Section 1.3.9 and the elasticities governing agglomeration externalities in
Section 1.3.10. From the model, housing elasticity is given by the expression 1−ηi
ηi
. I use
this expression to calculate the implied value of ηi
, land share of housing production
using the estimated census tract level housing supply elasticities from Baum-Snow and
Han (2019).
7According to the TIGER LINE shapefile dataset for LA county, there are 2,346 census tracts for LA in
total, I exclude 5 of them with anomalous geographic features e.g. offshore islands, coastline.
31

Parameter Description Value Source
α Labor share in production 0.82 Valentinyi and Herrendorf (2008)
γ Consumption share of housing 0.24 Davis and Ortalo-Magné (2011)
θ Fréchet elasticity parameter 6.85 Estimated - see section 1.3.9
κ Commuting elasticity 0.01259 Severen (2021)
1−ηi
ηi
Housing supply elasticity Various Baum-Snow and Han (2019)
χ Residential agglomeration externality 0.1276 Estimated - see section 1.3.10
(Working Households)
χ
R Residential agglomeration externality 0.1432 Estimated - see section 1.3.10
(Retired Households)
λ Employment agglomeration externality 0.086 Heblich et al. (2020)
Table 1.4: External and Estimated Parameters
1.3.9 Estimating Fréchet Elasticity
In order to estimate the Fréchet scale parameter θ, we can take the log of the commuting
probability πij:
logπij = −θκtij + X¯
i + E¯
j
where
X¯
i =


Xi
q
γ
i


θ
E¯
j = (Ejwj
)
θ
Using commuting flows from LODES 2015-2019 and travel times from the Census Tract
Planning Products dataset, I estimate the following gravity equation using Pseudo-Poisson
Maximum Likelihood, where X¯
i
is an origin fixed effect, E¯
j
is a destination fixed effect and
ζij is an error term.
logπij = −θκtij + X¯
i + E¯
j + ζij
Using the estimated coefficient on tij and the externally determined value of κ from Severen (2021), I can back out the implied value of θ: θκˆ = 0.0863, κ = 0.01259 =⇒ θ = 6.85.
32

I use estimates from Severen (2021) because he estimates commuting costs which are specific to Los Angeles.
πij
tij -0.0863***
(0.000170)
Origin FE Y
Destination FE Y
N 5,470,920
pseudo R
2 0.460
Standard errors in parentheses
* p < 0.05, ** p < 0.01, *** p < 0.001
Table 1.5: Estimation of Gravity Equation
1.3.10 Estimating Agglomeration Externalities
In my model, each location’s amenity value has a fundamental component that does
not depend on population density and a component that changes as population density
changes. The fundamental component depends on temperature and precipitation, among
other factors, which can change over time. Locations become more attractive to working
households as more working households move there. Locations become more attractive
to retired households as more retired households move there. Locations become more
productive the more people work there. Each of these agglomeration forces is governed
by parameters χ,χR and λ respectively.
Xi = x¯i

NRi
Λi
!χ
X
R
i = x¯
R

NRR
i
Λi
!χ
R
Aj = a¯j

N Ej
Λj
!λ
where x,¯ x¯
R and a¯ represent the location fundamentals that are exogenous to population.
I use temporal variation in population density to estimate the agglomeration elasticity
33

parameters χ and χ
R. I calibrate amenities X and X
R for two time periods using ACS
2006-2010 and ACS 2015-2019. I estimate the following equations to get estimated values χ = 0.1276 and χ
R = 0.1433. I include temperature and precipitation to use in Section
1.4 when I investigate how climate migration leads to differential wildfire outcomes.
logXit = β0 + β1Densityit + β2Tempit + β3Precipit + λi + ϵit
logX
R
it = β0 + β1Retired Densityit + β2Tempit + β3Precipit + λi + ϵit
1.3.11 Model Inversion
Following the approach of Ahlfeldt et al. (2015), I calibrate the model as follows:
Workplace Amenities Using observed wage indices, I solve for the workplace amenities
which are associated with labor market clearing using observed values of labor demand
and supply.
Labor demand N Ej =
X
i
πij|i NRi =
X
i
(Ej wj
/eκtij )
θ
P
s
(Es ws
/eκtis)
θ
NRi Labor supply
Residential Amenities Using the origin and destination fixed effects from the gravity
equation, and wage and rent indices from ACS 2015-2019, I calibrate residential and
workplace amenities for working households such that model generated residential and
workplace employment (NRi and N Ej
) are equal to observed residential and workplace
employment from LODES. I run a separate gravity equation for retired households in
two periods 2006-2010 and 2015-2019 and use a location fixed effect to get a value for
residential amenities for retired households and conduct a similar calibration such that
model generated retired population (NRR
i
) is equal to observed retired population from
34

ACS 2015-2019. Local Productivity Using calibrated wage and observed rent indices, I
generate the calibrated value of local productivity Aj using the following equation:
Aj =
wj
α
α
qj
1 − α
1−α
Land Productivity Using calibrated wage and observed rent indices from ACS 2015-
2019, I calibrate local land productivity φi using the following equation:
φi = η (HRi + HEj
)

1
η
η
1
(1 − η)
1−η
! 1
1−η
q
−
1
1−η
where HRi + HEj
is the total demand for floorspace generated using observed data from
ACS 2015-2019 and LODES.
1.4 Decomposing Climate- and Population-Driven Wildfires
In this section, I investigate how wildfire risk changes in response to climate change.
First, I consider a scenario where wildfire risk is not a function of population density. I
simulate this by keeping population density at baseline levels and allowing temperature
and precipitation to change according to projections for 2060 (RCP 4.5, NOAA CMIP5). I
use 2060 data because wildfires fluctuate from year-to-year in a cyclical pattern and how
severe wildfire impacts might be in the short run depends on which year you are using as
a comparison. Figure 1.2 shows the projected changes in average temperature.
35

Figure 1.2: Projected Changes in Av. Temperature
Next, I consider a scenario where wildfire risk is a function of population density. I simulate this by accounting for household migration in response to climate change. If households dislike extreme temperatures, fewer people may want to live in locations which become hotter as a result of climate change. If wildfire risk depends on population density,
hotter places may become less risky as people move out, but places with milder climates
may become more risky as people move in. Albouy et al. (2016) find that the average
American is willing to pay 1-3 percent of their income to avoid the forecasted effects of
climate change, and they are much more sensitive to extreme heat as opposed to extreme
cold. Cruz and Rossi-Hansberg (2021) incorporate temperature as a component of residential amenities and local productivity and estimate that a 1 % increase in temperature
will allow reduce amenities by 5 % and productivity by 15 % in the hottest places in the
world.
36

In my setting, I set aside the effect of climate on local productivity and focus on the
effect of climate on residential amenities. In my model, a 1 degree increase in average
temperature in LA county reduces working household residential amenity by 1.1 % and
retired household residential amenity by 3.9 %. In 2060, the average temperature in Los
Angeles is projected to increase by 1.4 degrees and annual precipitation will increase and
become more variable. I use these projections to generate a counterfactual set of residential amenities, assuming the current relationship between climate and amenities holds
in the climate change scenario. I use these counterfactual amenities and solve for a new
spatial equilibrium. Households move away from the hotter areas of LA, with the effect
being stronger for retired households compared to working households. I then input this
spatial equilibrium into my main wildfires function to get the predicted wildfire probability associated with climate migration. I examine the impact on wildfire probabilities
when the proportion of retired households shifts from a baseline of 16 % to a projected
37 % by 2060. While such households tend to avoid the most hazardous and warmest
locations in LA, they contribute to elevated wildfire risk in areas with moderate fire vulnerability.
1.5 Corrective Tax
In the previous section, I examined how households might want to avoid risky locations
even in the absence of government regulation since people do not like living in places
with excess heat. However, this reflects a long-run equilibrium. In the short run, there is
an urgent need to reduce damages from wildfires. In this section, I apply a corrective tax
which moves people out of risky locations.
37

Baylis and Boomhower (2023) point out that households in high fire risk locations are
implicitly receiving a subsidy to live there, since fire suppression costs are funded at the
federal or state level and households do not take these costs into account when making
their location decisions. I simulate a scenario where a local government decides to tax
people who live in high risk areas in order to recover the cost of fire suppression. This
is a corrective, but not a Pigouvian tax, since the tax does not account for how fire risk
changes as people move into or out of a location. I use expected protection costs generated by Baylis and Boomhower (2023) and express it as a share of median income in each
location to get a location-specific wildfire tax. Figure 1.3 shows the degree of taxation in
each location.
Figure 1.3: Corrective Tax
Notes: This figure shows the corrective tax I am using in this counterfactual scenario. The tax defined
based on the expected fire suppression cost (Baylis and Boomhower, 2023) in each location as a share of
median income. The median location receives an 8 % income tax and a total of 873 out of 2341 locations
receive a nonzero tax.
38

In the model, I modify the indirect utility function for retired and working households
by adding the tax (1−τi
) such that households now tradeoff between the attractiveness of
a location, the labor market access it offers and the cost of living in a high risk location.
Vij =
Xi
(1 − τi
)Ejwj
q
γ
i
exp(κtij)
V
R
ij =
X
R
i
(1 − τi
) yj
q
γ
i
I compare a tax which is proportional to expected fire suppression cost with a flat tax.
This simulates a proposed improvement in the political economy of fire suppression:
Currently, fire suppression is funded through federal / state expenditures and people
living in high-risk locations do not fully internalize the cost of living there. Would fire
outcomes improve if people paid a local tax that was proportional to fire risk? We can
also think of this counterfactual scenario as simulating what would happen if insurance
pricing more closely reflected the expected loss from wildfires. Taxes move people out of
the highest risk location into lower risk locations. However, a demographic change where
a higher share of the population is retired still leads to higher wildfire probabilities since
the tax is not changing to account for the marginal effect of households moving into
locations just outside the highest risk locations in LA.
1.6 Building Restrictions
In this section, I simulate a scenario where housing supply in the highest risk locations is
fixed and housing supply in unconstrained in all other locations in the city. All locations
which receive a non-zero tax in the previous section have fixed housing supply. Since
households can migrate into and out of the city, building restrictions do not reduce fire
risk in the highest risk zones, but more people move into the city center. (Figure ??) In
this scenario, 86,077 more working households and 17,391 more retired households move
39

into the city. I then consider a scenario where we combine a tax and building restrictions
in the high-risk zones and find that wildfire risk is reduced.
1.7 Social Cost of Wildfires
Major components of the cost of wildfires include fire suppression costs, structural damage and exposure to PM 2.5. In 2020, wildfire damages amounted to US $ 16.5 billion,
including the destruction of up to 10,000 structures damaged or destroyed in California
alone Bayham et al. (2022). PM 2.5 emissions generated by wildfires potentially reverse
the improvements in air quality from the Clean Air Act8 and contribute to approximately
25 % of the overall PM 2.5 emissions in the US (Burke et al., 2021).
In this section, I evaluate the cost of each scenario we have discussed in this paper. The
focus is on damage magnitude rather than event frequency. For each scenario, I create a
counterfactual population density distribution to compute localized wildfire probabilities. Utilizing emissions data from the Missoula Fire Lab Emissions Inventory, I estimate
expected PM 2.5 emissions for each location. To quantify the cost of these emissions
dispersing from the ignition point, I employ the InMAP pollution transport model. Neglecting the role of population density results in an underestimation of wildfire-related
damages. Wildfire damages from smoke exposure are 8 % higher when we account for
the effects of population density(Table 1.6).
Social costs of wildfire PM 2.5 emissions comes from epidemiological concentrationresponse functions which estimate a 6-14 % increase in mortality for 10 µg/m3
increase
in PM 2.5 emissions (Turner et al., 2017; Lepeule et al., 2012) multiplied by a measure of
8Wildfire Smoke Is Erasing Progress on Clean Air - The New York Times, Source: https://
www.nytimes.com/interactive/2022/09/22/climate/wildfire-smoke-pollution.html, Accessed on
10/07/2023.
40

Value of Statistical Life (Dept. of Transportation)9
. This is a lower bound on the social
cost of wildfires since I only measure one aspect of wildfire damages and I only account
for emissions occurring within 16 sq. kilometers of the ignition location and I do not
model how fires spread. These estimates make up 6 % of total property damage in the
state of California in a bad year such as 2020 or 33 % of total property damage in a mild
year such as 2022, as measured by CALFIRE10
.
Table 1.6 outlines the cost implications of various counterfactual scenarios compared to
a baseline where wildfire risk is decoupled from population density. Each counterfactual
incorporates 2060 climate projections (RCP 4.5) and the impact of population mobility.
In the baseline, I examine the effects of higher temperatures on wildfire probabilities
independent of human activity. In Los Angeles County, 98 % of areas show a rise in
wildfire probability. Given the preference of households for moderate climates, there’s a
migration away from the city’s hottest areas, which are also more prone to wildfires. This
shift leads to a reduction in wildfire risk: only 50 % of areas encounter an increased fire
risk, a significant decrease from the 98 % observed in the scenario considering only climate factors. A risk-proportional tax leads to an even greater reduction in fire risk since
households receive a strong price signal regarding living in the firezone. Finally, restricting housing supply in high-risk locations while relaxing zoning regulations in low-risk
locations leads to a marginal reduction in fire risk since the increase in housing supply
also attracts more households to move into the city center.
Demographic changes exacerbate fire risk since retired households are not constrained
to workplace locations and tend to live in the peripheral areas of the city. Restricting
9The Value of Statistical Life used is US $ 9.6 million. See 2016 Revised Value of Statistical Life Guidance, US Department of Transportation.
10CALFIRE Redbooks 2020, 2022. CALFIRE measures property damage in terms of property and content
loss in terms of replacement of like kind and quantity. This measure does not include suppression costs or
indirect losses such as business interruptions.
41

development in high-risk areas and allowing housing supply to respond flexibly in the
city center results in a marginal reduction in wildfire risk, since increased housing supply
also attracts people to move in to the city and does not necessarily induce relocation out
of high-risk locations. Demographic changes exacerbate this effect but including a corrective tax helps reduce overall wildfire risk, even in the presence of increased pressure
from population growth.
Scenario % Locations with ↑ Fire Risk PM 2.5 Exposure
Extensive Margin Intensive Margin
1. 2060 Temp. 98 % $ 325 - $ 731 mil.
(w/o human activity)
2. Climate Migration 50 % $ 107 - $ 242 mil.
3. Fire Tax 40 % $ 104 - $ 234 mil.
4. Building Restrictions 95 % $ 108 - $ 243 mil.
Table 1.6: Cost of Wildfires Through Expected PM 2.5 Exposure and Model Welfare
Notes:
1. Baseline with higher temperatures (projected 2060 values, RCP 4.5) and without human activity.
2. Scenario with higher temperature and household relocation in response to changes in local climate
amenities.
3. Scenario with risk-proportional fire tax.
4. Scenario where housing supply is fixed at baseline values in high-risk locations and housing supply
responds flexibly in low-risk locations (with a relaxation of zoning restrictions).
42

1.8 Conclusion
In this paper, I examined the extent to which population density contributes to wildfire
risk. This is an example of an endogenous disamenity, where the number of people in a
location increases the risk that a wildfire occurs in that location. I estimate an inverted
U-shaped relationship between wildfires and population density such that places with
intermediate density have the highest fire risk. Resilience to future wildfire risk depends
on the spatial configuration of population across the landscape of the US.
I developed a quantitative spatial model that captures how urban economic forces shape
the incidence of local environmental hazards. I capture a unique feature of wildfires as a
natural disaster where the incidence depends largely on human activity. I combine this
model with fine-scale data of Los Angeles county and I quantify the cost of wildfires in
terms of increased exposure to PM 2.5. I look at three scenarios in which households
might change their location choices: (1) households change their location choices in response to changing temperatures induced by climate change (2) households change their
location choices in response to a tax levied by a local government and (3) households
change their location choices in response to building restrictions. A tax that is proportional to baseline wildfire risk reduces the cost of wildfires, even though it does not fully
account for the effect of additional households on realized wildfire risk.
The findings in this paper constitute a lower bound to the overall effect of population
density for several reasons. Due to the limitations of the empirical design, I am only able
to capture the effect of population density on wildfires based on locations which are already populated and become more dense over time. I am not able to measure how fire
risk would increase if currently unpopulated locations became populated. In my analysis, I do not include a “feedback loop” where increased population density increases the
43

probability of wildfires and the increased probability of wildfires then affects how households make location decisions, although this is an open question in the literature and will
likely evolve as the costs of wildfires become more salient over time. I use estimated relationships generated from historical data but the future distribution of wildfires is not
necessarily similar to what we have already experienced. For instance, we expect to see
more simultaneous wildfires, which would put a strain on fire suppression resources,
which might increase overall exposure to wildfire PM 2.5. Finally, in my paper, I focus
on one of the most fireprone metropolitan areas in the US. Since it is already fireprone,
future changes in fire risk specific to Los Angeles may not be as large as those we might
expect for other metropolitan areas which have not experienced as many fires and so have
more fuel to burn, and potentially more severe fires, such as Portland, Oregon or Seattle,
Washington.
44

Chapter 2
Social Cost of Flooding
2.1 Introduction
Storms and flooding account for 74 % of total damages from loss of life attributed to
anthropogenic climate change (Newman and Noy, 2023). Our current global landscape
is one of increased flood frequency and severity and also rapid urbanization. Current
projections indicate that between 190 and 230 million individuals reside in areas anticipated to be below the 2100 flood levels (Kulp and Strauss, 2019). Recent scientific research highlights the importance of responding to high frequency, low magnitude flooding (Michael, 2007; Moftakhari et al., 2018; Wobus et al., 2019). Non-disastrous floods
are often not salient targets for city planners, but they can still lead to high levels of damage, especially when there are few policy or recovery solutions in place (Moftakhari et al.,
2017; Merz et al., 2021; Paulik et al., 2021).
Urban growth trends lead to an increase in non-permeable surfaces, resulting in increased storm runoff (Barnard, 1978). The capacity of municipal drainage systems to
manage surplus water from intense precipitation events is limited, and many are approaching the terminus of their operational lifespan, presenting significant maintenance
or replacement challenges, particularly in the developing world (O’Donnell and Thorne,
45

2020; Merz et al., 2021). Damages from natural disasters occurring in urban areas is estimated to be US$ 314 billion and may increase to US$ 415 billion in the absence of resilient
infrastructure development (Sharifi and Yamagata, 2018). The São Paulo Metropolitan
Region is the fourth largest urban agglomeration in the world, holding 10% of the population of Brazil and accounting for 19 % of national GDP. The city’s major east-west
highway is located next to the Tietê River and rapid urban expansion has not been accompanied by improvements in infrastructure (Haddad and Teixeira, 2015; Roncancio
and Nardocci, 2016).
Historically, flood impact assessments have focused on direct damages such as mortality and infrastructural damage (McGrath et al., 2019) as well as indirect damages in the
form of disruptions to commercial operations (Taberna et al., 2023). Recently, attention
has expanded to include intangible consequences of flooding, such as the psychological
distress associated with proximity to inundated areas (Lekuthai and Vongvissessomjai,
2001; Veldhuis, 2011; Abu et al., 2024). Our research is directed towards understanding the effects of recurrent urban flooding on travel delays. Our approach integrates
real-time travel time data obtained using the Google Directions API and we employ an
instrumental variables framework to ascertain the causal influence of urban flooding on
travel durations. Other studies use simulated origin and destination points (Hanna et al.,
2017; Akbar et al., 2018, 2023b,a) to study travel speeds in cities around the world. In
contrast, we use a representative household survey (Ang et al., 2020) to determine the
set of trips for analysis and complement it with repeated observations in order to assess
typical congestion and flood occurrences encountered by households. Other research has
been conducted combining this origin-destination survey with road network simulations
to construct a measure of local vulnerability to flooding (Tomás et al., 2022). Our paper
differs in using real-time travel data and an economic methodology to assess the impact
of observed flooding on travel times. Some authors employ enumerators for the collection
46

of travel times but this approach limits the number of trips and range of conditions that
can be studied (Lu et al., 2012; He et al., 2021). In our work, each origin-destination pair
is queried up to 75 times under varying traffic and weather conditions.
The concentration of economic activities in low-lying urban areas perpetuates migration
to these flood-prone areas (Rentschler et al., 2023). This underscores the need for infrastructure investments capable of safeguarding these assets (Hettiarachchi et al., 2018;
Kocornik-Mina et al., 2020; Pour et al., 2020). Using our estimation framework, we obtain
trip-specific measures of the travel delays generated by urban flooding. We combine these
travel delays with individual wages to construct a measure of the social cost of flooding.
We can sum these costs at the individual level to estimate the annual flood-related expenses per household (Table 2.7) or at the location level to identify intersections that
would benefit the most from infrastructure upgrading (Figure ??). Our contribution lies
in generating a demand-driven assessment of flood costs, where the benefits of upgrading
infrastructure in a given location depends on both the volume of individuals traversing
the location and the severity of flood-related expenses each person encounters.
2.2 Dataset
São Paulo Origin-Destination Travel Survey We obtained a set of origin and destination coordinates from a travel survey that collected detailed information about 53,305
trips taken by 8,115 households. The survey was designed to be representative of commuting patterns in the Metropolitan Region of São Paulo on a regular weekday. Our sample of representative trips is restricted to the subset of 15,483 trips taken by car, which
are designed to be representative of 12.49 million motorized trips taken on a weekday in
São Paulo.
47

Demographics There are 17,890 individual survey respondents. Of these, 5,985 of
them took car trips which were recorded by the survey. On average, car trip respondents are: older, have higher household income, more likely to be male and have regular
work. We construct a measure of household income per working age adult to use in our
estimation of the value of additional time spent in traffic due to flooding. We use this
measure to account for how students, retirees and those who are unemployed may have
zero reported individual income in the survey.
Trips Table 2.1 shows how many of the survey trips we used in this study are distributed
across different types of trips. In the 2012 travel survey, responses were balanced across
31 survey zones. We define morning peak as 7 am - 10 am and evening peak as 5 pm
- 8 pm, according to the definition used by São Paulo’s vehicle license plate restriction
policy1
. We define “downtown” as survey zones 1, 14, 15, 16 and 23. The survey data
captures heterogeneity in travel times: morning peak hour trips going towards the downtown area take 7 minutes longer on average and evening peak hour trips coming from
downtown take 14 minutes longer. We estimate the cost of encountering a flood by generating the added travel time due to a flood as a proportion of the travel time you are
already incurring on your trip. If my trip takes a long time because I am far away from
my destination or because the road I am traveling on is congested, and I encounter a flood,
then I am bearing a relatively larger cost of flooding.
1Source: https://en.wikipedia.org/wiki/Vehicle_restriction_in_S%C3%A3o_Paulo
48

No. of Obs. Share Av. Travel Time
(mins)
Survey Trips 15,483 1.00 31.18
Morning Peak 4,007 0.2588 32.94
To Downtown 1,248 0.0806 38.41
From Downtown 1,016 0.0656 31.44
Evening Peak 4,441 0.2868 36.03
To Downtown 963 0.0621 38.69
From Downtown 1,220 0.0787 45.03
Off-Peak 7,035 0.4543 27.11
To Downtown 1,962 0.1267 29.49
From Downtown 1,932 0.1247 28.94
Table 2.1: São Paulo Household Survey Descriptive Statistics
Flood Events There are 1,059 flood events recorded by the main traffic agency2
in the
city of São Paulo which overlap with the time of our data collection (July 2016 - September 2017). The agency uses satellite imagery, data from weather monitoring stations and
observer reports to determine where flood events are occurring3
. In the paper, we address
the question of whether some areas in the city are more likely to have a flood reported
than others in our empirical strategy.
Reported floods data include: (1) the date on which the flood occurred (2) the start and
end time of the flood event (3) the road segment on which the flood event occurred and
any identifying landmarks (4) whether or not the road segment is passable or blocked and
(5) which direction of travel is affected by the flood. Using (1) and (2), we generate the
duration of the flood event. We use flood event duration in our main analysis to estimate
the amount of added travel time generated per minute of a flood event. We use (3) to
2Flood events are reported at https://www.cgesp.org/v3/alagamentos.jsp. CGESP is a division of
the main traffic agency in the city of São Paulo, CET.
3Monitoramento - CGE, Source: https://www.cgesp.org/v3/monitoramento.jsp, Accessed on July
16, 2020.
49

generate geographic coordinates for each flood events. We query Google Maps using all
available road segment data (usually two street names, or a street name with a number)
and take the intersection between two streets as the location of the flood. The coordinates
are automatically generated by Google after entering in the road segment names. Of the
1,059 flood events, there are 408 unique locations (unique lat-long pairs). Flooding is
concentrated in the area near the city center of São Paulo. For our analysis, we generate a
100 m buffer zone around the flood coordinates and intersect with the baseline trip route
to see if a given trip encounters a flood (Figure 2.1, Panel D).
Figure 2.1: Dataset Construction
Observing trip durations using Google Transit API From July 2016 to September
2017, we queried Google’s Directions API using origin-destination pairs from a representative travel survey. We collected real-time trip data at 20 minute intervals within 2 hours
of the time that the surveyed trip was originally taken. For each surveyed trip, we obtain
approximately 75 observations of real-time trip durations collected over the course of a
week. The Google API generates trip durations based on optimal routes which account
50

for real-time traffic conditions. This allows us to observe the same trip under flooding
and non-flooding traffic conditions (Figure 2.1, Panel B). We can then estimate the added
travel time facing drivers who optimally adjust to traffic conditions created by flooding
(Table 2.5). The assumption is that if my regular commute is impassable or very congested due to a flood, I will adjust my trip route to arrive at my destination in the shortest
amount of time possible, given the new situation.
Table 2.2 shows the share of trips and average travel times under different categories
in the queried trips dataset. Of the 1.4 million queries made on the Google API, 909 are
designated as “flooded trips." A flooded trip is one where (1) queried trip departure time
takes place after the beginning of a flood event and before the ending of a flood event and
(2) the baseline trip route intersects with the location of the flood event.
No. of Obs. Share Av. Travel Time
(mins)
Queried Trips 1,471,923 1.00 19.99
Flooded Trips 909 0.0006 31.09
Morning Peak 149 0.0001 38.32
To Downtown 93 6.32e-05 28.36
From Downtown 55 3.8e-05 15.858
Evening Peak 353 0.0002 34.54
To Downtown 127 8.63e-05 35.98
From Downtown 123 8.42e-05 34.22
Off-Peak 405 0.0002 25.424
To Downtown 222 0.0001 21.595
From Downtown 235 0.0001 24.896
Table 2.2: Queried Trips Descriptive Statistics
51

Figure 2.2 illustrates the density of trip times between what is reported in the 2012 survey (“Survey Trips”) and what we collected from Google (repeated queries of the time it
takes to get from survey trip origin to survey trip destination, “Queried Trips”). The survey trips tend to be bunched around certain values compared to the queried trips. Survey
respondents may have rounded their responses to the nearest 5 or 10 minutes. We do not
observe the routes taken by drivers, we only observe the time it takes to get from a given
origin to a given destination. The Google Directions API uses an algorithm which generates the lowest travel time between origin and destination based on data from mobile
phones which are in the vicinity, which have location tracking on Google Maps enabled
(Google, 2009). (Hanna et al., 2017) report that the data obtained from the Google API
are reflective of actual times.
Figure 2.2: Comparing Survey and Queried Trips
Figure 2.3 shows the real-time trip durations we collected for a given survey trip. Each
52

line represents a different day on which data was collected for this survey trip. Data was
collected within 4 hours of the trip departure time reported in the travel survey to ensure
that we are capturing a similar kind of trip from that which was reported. For instance,
if a trip takes place at 5 pm on a week day because the respondent is going home from
work, we collect data on trip times between the same origin and destination between 3
pm and 7 pm to capture the kinds of traffic conditions this individual would encounter
on their way home. For this particular trip, a flood event occurred on July 18, 2016, and
we collected data on how long it would have taken for this individual to travel from origin to destination while the roads were flooded, either by sitting in traffic or by taking
a longer route which avoids the flood. In our analysis, we compare the additional time
from the flooded trip query (collected on July 18, 2016) against the average trip time from
all other days on which we collected data for this particular survey trip. We also remove
any addition variation that is due to the particular day of the week, the hour of the day,
the month of the year so that we are able to isolate the effect of the flood alone and not
other conditions which affect travel time (e.g. public holidays).
Figure 2.3: Repeated Observation of One Survey Trip
53

2.3 Measuring the impact of urban flooding on travel times
Table 2.3 shows the results of estimating the coefficients of the following equation:
log(Trip Time)it = β (Flood)it + γ (Rain Bins)it + αi + δt + ϵit
The outcome variable log(Trip Time)it is the log of trip duration (in minutes) of the i-th
survey trip queried at time t. (Flood)it is an indicator variable which takes the value 1 if
the queried trip takes place during a flood event and if the baseline trip route intersects
with the flood location and 0 otherwise. (Rain Bins)it take the value 1 if the average rainfall in the city of São Paulo is light, moderate or heavy4
, 0 otherwise. αi
is a survey trip
fixed effect which removes variation which is specific to survey trip i and δt
is a number
of time fixed effects which remove variation which is specific to particular periods of time
e.g. time of day, day of week, month of year. ϵit is the error term. Including the fixed
effects terms removes all survey-trip-specific and time-specific variation and leaves us
with the average effect of a flood event on travel times in the city.
Table 2.3 shows how the estimated effect changes as we remove more time-specific variation. In (4), we see that flood events are associated with a 2.6 % increase in travel time.
The average effect is small because flood events (and flooded trips) are relatively rare
compared to the total number of trips being taken throughout the city. In our paper, we
are interested in how some individuals / locations bear a large share of the overall cost of
flooding.
4We use definitions provided by the American Meterological Society. Light rain is any rain that is less
than 0.25 cm per hour, moderate rain is rain that is between 0.26 cm per hour and 0.76 cm per hour and
heavy rain is rain that greater than 0.76 cm per hour. Source: Rain - AMS Glossary, http://glossary.
ametsoc.org/wiki/Rain, last modified on April 25 2012, at 19:44. Accessed on July 16, 2020.
54

Trip Duration (log)
(1) (2) (3) (4) (5)
Flood 0.041∗∗∗ 0.038∗∗∗ 0.038∗∗∗ 0.031∗∗∗ 0.038∗∗∗
(0.013) (0.013) (0.013) (0.012) (0.013)
Light Rain 0.009∗∗∗ 0.012∗∗∗ 0.012∗∗∗ 0.020∗∗∗ 0.022∗∗∗
(0.001) (0.001) (0.001) (0.001) (0.001)
Moderate Rain 0.021∗∗∗ 0.026∗∗∗ 0.026∗∗∗ 0.027∗∗∗ 0.028∗∗∗
(0.002) (0.002) (0.002) (0.001) (0.001)
Heavy Rain 0.005∗ 0.010∗∗∗ 0.010∗∗∗ 0.004 0.001
(0.003) (0.003) (0.003) (0.003) (0.003)
Trip FE Y Y Y Y N
Month FE N Y Y Y Y
Day of Week FE N N Y Y N
Hour FE N N N Y N
Trip x Day x Hour FE N N N N Y
Treated Obs. 488 488 488 488 488
Observations 1,471,923 1,471,923 1,471,923 1,471,923 1,471,923
R
2 0.985 0.986 0.986 0.988 0.990
Adjusted R2 0.985 0.986 0.986 0.988 0.990
Residual Std. Error 0.114 0.113 0.113 0.103 0.095
Note: ∗p<0.1; ∗∗p<0.05; ∗∗∗p<0.01
Standard errors clustered at survey trip level.
Table 2.3: Effect of flooding on trip duration
without accounting for non-random flood reporting
Since the floods data is dependent on reports from observers, we want to account for nonrandom selection of floods reporting. For instance, our results may overestimate the effect
of floods on travel times if floods which occur in locations with poor drainage are more
likely to be reported than floods which occur elsewhere. Using instrumental variable
(IV) approach, we look for a variable that explains a large proportion of the variation in
the independent variable (flood duration) and is related to the outcome variable (travel
time) only through the independent variable. We use accumulated rain in the 12 hours
preceding the query trip departure time as our instrumental variable. We exclude data
from the 3 hours immediately preceding the departure time because we want to isolate
55

the effect of accumulated rain on flooding and we do not want to include the effect of
accumulated rain on the trip duration itself. We estimate the following equations:
(Flood Duration)it = λt + β (Flood)it × (Accumulated Rain)it + γ (Rain Bins)it + δt + νit
log(Trip Time)it = αi +δt +β (Flood)it ×(Predicted Flood Duration)it +γ (Rain Bins)it +ϵit
From the first equation, we generate predicted values for floods duration (in minutes) as
predicted by accumulated rain in the hours leading up to the time the queried trip was
taken (Table 2.4). Every additional millimeter per hour of accumulated rain results in
3.1-3.4 fewer minutes of flood duration. We use both multiplicative and additive fixed
effects and find that the predicted effect of accumulated rain remains consistent between
the two specifications. In the second regression, we regress the log of trip duration of the
predicted values generated by the first regression (Table 2.5). Every additional minute of
flooding (as predicted by accumulated rain) causes a 0.0424 % increase in trip duration.
Since the average flood duration in our dataset is 148 minutes, we can expect that, on
average, floods cause a 6.3 % increase in travel time and a 9.3 % increase in travel time if
the flood occurs during the evening peak hours.
56

Floods Duration (mins)
Flood × Acc. Rain 16.246∗
(9.246)
Light Rain −0.979
(1.781)
Moderate Rain 3.316
(2.919)
Heavy Rain 0.494
(3.461)
Month FE Y
Day of Week FE Y
Hour of Day FE Y
Treated Obs. 488
Observations 1,471,923
R
2 0.006
Adjusted R2 0.006
Residual Std. Error 183.503
Note: ∗p<0.1; ∗∗p<0.05; ∗∗∗p<0.01
Standard errors are clustered at the survey trip level.
Table 2.4: Effect of accumulated rain on flood duration
57

Trip Duration (log)
(1) (2) (3) (4)
Av. Effect 0.00025∗∗ 0.00040∗∗∗
(0.00012) (0.00015)
Morning Peak 0.00024 0.00054∗∗
(0.00019) (0.00025)
Evening Peak 0.00069∗∗∗ 0.00077∗∗∗
(0.00027) (0.00027)
Off-Peak −0.00008 −0.00011
(0.00013) (0.00013)
Light Rain 0.02025∗∗∗ 0.02024∗∗∗ 0.02241∗∗∗ 0.02241∗∗∗
(0.00088) (0.00088) (0.00088) (0.00088)
Moderate Rain 0.02705∗∗∗ 0.02707∗∗∗ 0.02793∗∗∗ 0.02795∗∗∗
(0.00143) (0.00143) (0.00140) (0.00140)
Heavy Rain 0.00380 0.00376 0.00099 0.00096
(0.00284) (0.00284) (0.00312) (0.00312)
Trip FE Y Y N N
Month FE Y Y Y Y
Day of Week FE Y Y N N
Hour FE Y Y N N
Trip x Day x Hour FE N N Y Y
Treated Obs. 488 488 488 488
Observations 1,471,923 1,471,923 1,471,923 1,471,923
R
2 0.98817 0.98817 0.99028 0.99028
Adjusted R2 0.98805 0.98805 0.98975 0.98975
Residual Std. Error 0.10250 0.10250 0.09491 0.09491
Note: ∗p<0.1; ∗∗p<0.05; ∗∗∗p<0.01
Standard errors are clustered at the trip level.
Table 2.5: Effect of predicted flood duration on trip duration
58

2.4 Estimating damages
The coefficients in Table 2.5 are the average effect of flooding on travel time per minute of
flooding. We multiply the coefficients from Table 2.5 by the average flood duration in our
dataset to get an average increase in travel time (Table 2.6). We generate the annualized
damages from increased travel time due to flooding using the following equation:
(Annual Damage)i = (Value of Time)i × (Annual Average Added Travel Time)i
× (Survey Expansion Factor)i × (Conversion to 2017 USD)
Annual Damage is how much 2017 USD is generated by the added travel time from flooding in the city of São Paulo per survey trip. Value of Time converts added travel time
borne by the individual taking the trip into a monetized value. We define value of time as
50 % of the individual’s household income per working age adult. Household income in
the survey is reported in 2012 BRL and so we convert it into 2017 USD. Annual Average
Added Travel Time is the percentage values from Table 2.6 multiplied by the number of
days of flooding in 2017 (58 days of flooding) and the survey trip duration.
Lower Bound Estimate Upper Bound
Av. Effect 1.60 % 5.94 % 10.27 %
Morning Peak -2.45 % 13.21 % 28.87 %
Evening Peak 2.73 % 10.55 % 18.37 %
Off-Peak -7.40 % -1.24 % 4.93 %
Table 2.6: Estimated Average Added Travel Time
We exclude floods which occur on weekends because the survey is only representative
of trips which occur on Monday-Friday. The travel survey includes “expansion factors”
59

which we use to extrapolate from the survey to the population of the Metropolitan Region
of São Paulo. The annual damage from flooding accruing to all the people in São Paulo
who are observably similar to the individual who takes survey trip 6784 is 20,382 USD.
Table 2.7 shows the annual damages per person (sum across all trips taken by the same
individual) and the aggregate annual damages (sum across all trips).
Annual Damages Per Person
(2017 USD)
1st Qu. Median 3rd Qu. Max.
30.16 69.86 155.29 3,822.47
% of Income
1st Qu. Median 3rd Qu. Max.
0.50 % 0.99 % 1.98 % 11.88 %
Annual Damages Per 5 sq. km
(2017 USD)
1st Qu. Median 3rd Qu. Max.
27,723.00 61,107.00 236,925.00 2,804,313.00
Aggregate Annual Damages
(2017 USD)
Average Effect Peak Hours
119,065,891.00 134,330,079.00
% of Overall Cost of Congestion
5.26 % 5.92 %
% of Flood Damages from Business Disruptions
(Haddad et. al. (2015)
273.31 % 307.76 %
% of São Paulo GDP
0.05 % 0.06 %
Table 2.7: Estimated annual damages (2017 USD)
In order for the monetized damages we calculate to represent the actual damages from
added travel time due to flooding in 2017, we assume that (1) the 2012 travel survey is
representative of trips taken in 2016-2017 and (2) that the trips taken in 2012 were taken
60

in 2016-2017 (with 100 % probability) and delayed by floods. In other words, we do not
account for how flooding might affect the decision to take a trip or the decision to change
trip characteristics (e.g. trip departure time). We interpret our results as estimating what
would have happened if people did take the trip, encounter floods and spent extra time
on the road as a result. To address assumption (1), we discuss the similarities between
the trips taken in the 2012 survey with the trips taken in another representative travel
survey conducted in 2017. We evaluate assumption (2) below.
2.4.1 Robustness Checks
Effect of Flooding on the Probability of Car Travel
We want to test whether or not people take fewer trips on days when floods occur compared to days when floods do not occur. We aggregate the number of survey trips / floods
to the survey zone level and conduct a Poisson regression of number of trips per survey
zone per day on (1) whether or not it flooded that day and (2) the number of floods that
occurred that day. We include fixed effects for the survey zone and the week of the year
(Table 2.8) Compared to the average number of trips taken in that survey zone and in
that week of the year, days where there is flooding are associated with a slight positive
increase (significant at the 10 % level) in the number of trips taken and increasing the
number of floods occurring in a day is associated with increasing the number of trips
taken (significant at the 1 % level). This means that, if anything, people tend to take more
trips when it floods and that we may be underestimating the total damages caused by
flooding because our data collection does people who decide to take trips on flooded days
(who otherwise might not have).
61

Number of Trips Per Zone
(1) (2)
Flood 0.044∗
(0.023)
No. of Floods 0.024∗∗∗
(0.005)
Light Rain -0.175∗∗∗ -0.177∗∗∗
(0.018) (0.018)
Moderate Rain -0.567∗∗∗ -0.571∗∗∗
(0.047) (0.047)
Heavy Rain -1.139∗∗∗ -1.159∗∗∗
(0.123) (0.123)
Constant 0.186 0.186
(0.707) (0.707)
Zone FE Y Y
Week of Year FE Y Y
Observations 9,025 9,025
Note: ∗p<0.1; ∗∗p<0.05; ∗∗∗p<0.01
Table 2.8: Probability of taking a trip on a day with flooding
Pooled observations from 2012 and 2017 travel survey
Effect of Flooding on Time Adjustments
We want to test whether people choose to take trips at different times on days which
are flooded. We aggregate the number of survey trips / floods to the survey zone level
and conduct a Poisson regression of the number of trips per survey zone per hour on (1)
whether or not it flooded in that hour (2) the number of floods in your survey zone in that
hour and (3) a dummy variable indicating the hours of the day that are not flooded on a day
that is flooded. There is a small decrease in the number of trips associated with flooding in
that hour (not statistically significant), a negligible positive effect of the number of floods
in that hour and a small positive effect during non-flooded hours of a flood day. This is
evidence that people are not changing their departure times in response to flooding. This
62

suggests that our estimates reflect the costs of flooding as borne by people who take trips
at around the same time as reported in the travel survey.
Trips Per Zone Per Hour
(1) (2) (3)
Flood -0.050
(0.085)
No. of Floods 0.001
(0.026)
Not Flooded Hours × Flooded Day 0.050
(0.085)
Light Rain -0.101∗∗ -0.101∗∗ -0.101∗∗
(0.046) (0.046) (0.046)
Moderate Rain 0.030 0.028 0.030
(0.110) (0.110) (0.110)
Heavy Rain -0.080 -0.080 -0.080
(0.143) (0.143) (0.143)
Zone FE Y Y Y
Date FE Y Y Y
Note: ∗p<0.1; ∗∗p<0.05; ∗∗∗p<0.01
Table 2.9: Probability of taking a trip in an hour with flooding
Pooled observations from 2012 and 2017 travel survey
Effect of Flooding on Route Substitution
Using the Google Directions API, we test for route substitution in response to increased
travel time because of flooding. From our main results, the average flood adds 5.94 % to
travel time. Our hypothesis is that flood costs are concentrated around major thoroughfares in the city because there are few alternate routes and even if there are not alternate
routes, the amount of time saved by re-routing is less than the extra time spent on an
alternate route. Thus, even though damages are small at the individual level, they can
63

be large at the aggregate level since we are adding up costs over a large base. Figure
2.4 shows two randomly selected survey routes, one on which the majority of the trip
uses the Marginais and another which does not use the Marginais. For trips which use
the Marginais, we generate alternate routes by selecting the “avoid highways” option in
Google. The alternate route is 12 minutes longer than the original route which uses the
Marginais. If the person traveling encounters a flood, on average, 1.53 minutes (5.94 % of
trip time) will be added to the trip and thus, the person would rather stay on the flooded
route than switch to the alternative route which still takes longer. For routes which do
not use the Marginais, there are more alternative routes and travel times are more similar
between alternatives. For the trip which does not use the Marginais, the different between
the main route and the alternative route is 2 minutes, while the effect of the average flood
on this trip would be 2.01 minutes, and so the person may be indifferent between staying
on their current route or taking an alternate route.
64

(a) 70 % of trip on Marginais
(b) Does not use Marginais
Figure 2.4: Alternative Routes Generated Using Google Directions API
65

Figure 2.5 shows how trips which use the Marginais and trips which do not use the
Marginais differ in terms of the additional travel time added from re-routing (to avoid
a flood) versus the additional travel time added from encountering a flood. Marginais
trips tend to have higher costs of re-routing and thus, people who use the Marginais tend
to bear most of the social cost of flooding in the city of São Paulo.
Figure 2.5: Difference between Re-routing and Flood Effect
66

Chapter 3
Amenities in QSE Models
3.1 Introduction
Urban economics seeks to study the forces which underlie the internal structure of city.
A standard model (Ahlfeldt et al., 2015; Redding and Rossi-Hansberg, 2017; Monte et al.,
2018) includes the following factors which explain the number of households residing in
a given location i and commuting to a work location j: (1) The attractiveness of a given
residential location Xi
(2) the attractiveness of a given workplace location Ej
(3) wages wj
(4) the cost of commuting between i and j, dij and (5) the cost of housing qi
.
The residential amenity X is a model residual which captures all other factors aside from
housing prices, commuting costs and labor market access which explain the number of
households in a given location. A key contribution of (Ahlfeldt et al., 2015) is being able
to distinguish between agglomeration forces (where the attractiveness of a location is augmented by the number of people who live / work in that location) and the other characteristics of a location which are not affected by the number of people living there (location
“fundamentals”). Location fundamentals could include the natural geographic features
of a location e.g. distance to the coastline, natural beauty (Lee and Lin, 2018), local climate (Albouy et al., 2016), etc. The agglomeration forces present in a given location could
67

depend on the ratio of college to non-college workers (Diamond, 2016), tourism (Almagro and Domínguez-Iino, 2022), number of retirees (Komissarova, 2022), crime (Khanna
et al., 2023) and so on.
Prior studies have underscored the significance of specific variables in modeling residential amenity X. Yet, what remains unclear is the relative significance of these variables in
shaping the value of X and what this implies for the use of this class of spatial models.
From a broader perspective, we are asking the questions: What drives household residential location choice? To what extent are these factors captured in this standard class of
spatial models?
In this paper, we seek a comprehensive understanding of the components of X. We construct a spatial model of the Los Angeles metropolitan area (comprised of 5 counties).
We calibrate the model using data from the American Community Survey 2012-2016.
We construct a census tract-level dataset of 11 main types of amenities which include:
topography (ruggedness), traffic, heat exposure (urban heat island effect), localized pollution, seasonal precipitation and solar radiation, racial demographics, zoning, crime,
school quality and a measure of local consumption amenities (e.g. number of pet stores,
etc.). We conduct a series of analyses to determine the relative importance of these amenities in explaining variation in X. Finally, we look at different ways of calibrating X and
see which version best matches the available data.
3.2 Residential Amenity
68

According to our baseline model, households choose residential location i and workplace
location j using the following indirect utility function:
νij =
XiEjwj
q
γ
i
dij
If idiosyncratic preference shocks are distributed using a Fréchet distribution, then the
number of households living in location i and commuting to j can be expressed as:
πij =
ν
θ
ij
P
i
′
j
′ ν
θ
i
′
j
′
We obtain tract-level commuting data from the US Census LEHD Origin-Destination
Employment Statistics (LODES) database. We construct a housing price index using
data from the American Community Survey and we construct a wage index using data
from the Census Transport Planning Products (CTPP) database (Delventhal et al., 2022;
Delventhal and Parkhomenko, 2023). We calibrate values of X and E such that modelgenerated commuting flows are equal to observed commuting flows, up to a tolerance
level. We decompose X into an endogenous and exogenous component as follows:
Xi = x¯i

NRi
Λi
!ρ
where x¯i
refers to the component of X which does not depend on population density, NRi
is the number of residents living in i, Λi
is the land area of location i and ρ is a model
parameter which governs the degree to which X is determined by population density
NRi
/Λi
. We use both X and x¯ in our empirical analysis.
69

Figure 3.1: Calibrated Residential Amenities (Los Angeles Metropolitan Area)
70

3.3 Dataset
Ruggedness. Ruggedness is measured using the Terrain Ruggedness Index (TRI). The
index captures differences in elevation within a given unit of surface area. Grid-level
data on ruggedness is taken from Nunn and Puga (2012) and is publicly available 1
.
Traffic. Traffic is measured as the number of vehicle-kilometers per hour per road length
within 150 meters of the census tract boundary. It is part of the CalEnviroScreen dataset.
Heat exposure. Heat exposure is measured using the urban heat island index produced
by the California Environmental Protection Agency2
. The index measures the difference
in temperatures between an urban census tract and a nearby upwind rural census tract.
Daytime temperatures can be 1-6 deg. F. higher in urban areas and up to 22 deg. F. higher
at night because of heat radiating from buildings and pavement.
Local pollution. Pollution data is taken from CalEnviroScreen3 dataset produced by
the California Office of Environmental and Health Hazard Assessment. Each location is
given a pollution burden score which ranges from 0-10. The score represents an average
of percentiles from a number of different measures of pollution, including ozone, PM
2.5, diesel, drinking water contamination, lead exposure, pesticides, toxic releases from
factory facilities and incineration.
Seasonal precipitation and solar radiation. Seasonal precipitation and solar radiation
is taken from the UOregon PRISM dataset4
.
1Data on Terrain Ruggedness and Other Geographic Characteristics of Countries, Source: https://
diegopuga.org/data/rugged/
2Understanding the Urban Heat Island Index | CalEPA, Source: https://calepa.ca.gov/climate/
urban-heat-island-index-for-california/understanding-the-urban-heat-island-index/
3CalEnviroScreen 4.0 | OEHHA, Source: https://oehha.ca.gov/calenviroscreen/report/
calenviroscreen-40
4PRISM Climate Group, Oregon State University, https://prism.oregonstate.edu
71

Racial demographics. We use self-reported race variables from the American Community Survey. We focus on self-identified White (non-hispanic), African American, Asian,
Latino and Other Race individuals. The Other Race category includes Pacific Islanders,
Native Americans, and individuals of mixed race or two or more races.
Zoning. We observe the share of properties within each census tract that are zoned
for different uses: Single-family residence, multifamily residences, non-residential use,
non-determined. Zoning data is taken from the Southern California Association of Government Area Residential Zoning Dataset5
School quality. School quality data is taken from SchoolDigger.com which publishes
rankings and test scores for schools across the US. We construct a measure of school
quality by taking the average test score of the public elementary, middle and high schools
closest to each tract’s centroid.
Crime Data on the number of crimes per census tract is taken from the city of Los
Angeles’ Open Data Portal6 The dataset includes the geographic coordinates where each
crime took place, the time of crime, date and time of day when the crime took place.
In our analysis, crimes are divided into personal (e.g. assault) and property crime (e.g.
vandalism). We apply a logarithmic transformation to these variables in the empirical
analysis to ascertain the degree to which a change in X correlates with a 1% change in the
number of crimes committed in a given census tract.
5Southern California Association of Government Area Residential Zoning Data | The Othering & Belonging Institute, Source: https://github.com/OtheringBelonging/SCAGZoning/blob/main/README.md
6Crime Data from 2010 to 2019 | Los Angeles - Open Data Portal, Source: https://data.lacity.org/
Public-Safety/Crime-Data-from-2010-to-2019/63jg-8b9z/data
72

Consumption amenities. Data on the number of different types of retail establishments
is taken from the National Neighborhood Data Archive (NaNDA)7 We apply a logarithmic transformation to these variables in the empirical analysis to ascertain the degree to
which a change in X correlates with a 1% change in the number of retail establishments
in a given census tract.
3.4 Empirical Analysis
We run a series of analyses to examine the degree to which variation in X is explained
by observable data on amenities. We calculate the correlation coefficients of each variable and X to determine which variable is most correlated with X (Table 3.1). We also
calculate R squared values of regressing X against each variable individually (Table 3.2)
to see what share of variation in X is explained by each variable. In the aforementioned
approaches, each individual variable could be acting as a proxy for other variables. For
instance, if the share of White residents in a given location is correlated with natural
amenities, high school quality, etc. then the share of Whites could explain a high share of
the variation in X but it is not clear which amenities are really driving high values of X.
To address this, we run a regression of X against all variables at once (Table 3.3). In the
regression results, we can see which variables have a statistically significant relationship
with X, while controlling for all the other variables. We have attempted to ensure that
variables are within a similar range of magnitude and / or have values that can be easily
interpreted. However, it is hard to assess the relative importance between two statistically significant variables. Finally, we conduct a Shapley-Owen R squared decomposition
to assess the relative importance of each variable in the overall R squared value of the
regression (Table 3.4).
7National Neighborhood Data Archive (NaNDA): Retail Establishments by Census Tract, United States,
2003-2017. Source: https://www.openicpsr.org/openicpsr/project/115972/version/V2/view
73

Exog.
Full X Component
Only
Traffic 0.015 -0.076
Pollution Score (0-10) -0.040 -0.149
Ruggedness Index 0.225 0.338
Urban Heat Island Index 0.040 -0.212
White 0.424 0.464
African American -0.147 -0.118
Asian 0.205 0.102
Latino -0.400 -0.433
Other Race 0.324 0.274
Property Crime -0.039 -0.182
Personal Crime -0.065 -0.209
Multifamily Housing 0.041 -0.282
Mixed Development -0.002 -0.031
Zoning Not Determined -0.047 0.022
Non-Residential -0.080 0.148
Single-Family Housing 0.147 0.167
Sch. Quality (Elem.) 0.377 0.394
Sch. Quality (Middle) 0.320 0.319
Sch. Quality (High) 0.185 0.194
Precip. (Jan) 0.133 0.052
Precip. (April) 0.029 0.077
Precip. (July) -0.141 0.174
Precip. (Oct) 0.104 0.175
Solar Radiation (Jan) -0.093 -0.029
Solar Radiation (April) -0.063 0.015
Solar Radiation (July) -0.132 0.089
Solar Radiation (Oct) -0.119 0.126
Eating Places 0.356 0.214
Full Restaurants 0.358 0.210
Fast Food 0.190 0.170
Coffee Shops 0.252 0.187
Bars 0.211 0.106
Furniture Stores 0.223 0.201
Electronics Stores 0.286 0.230
Supplies Stores 0.162 0.205
Clothing Stores 0.332 0.238
Sports / Music Stores 0.309 0.279
Department Stores 0.094 0.107
Pet Stores 0.101 0.090
Used Merchandise Stores 0.161 0.101
Cosmetics Stores 0.164 0.125
Food / Health Stores 0.153 0.160
Table 3.1: Correlation Coefficients
74

Full X Exog. Component
Variable R Sq. Variable R Sq.
White 0.179 White 0.216
Latino 0.160 Latino 0.188
Sch. Quality (Elem.) 0.142 Sch. Quality (Elem.) 0.155
Full Restaurants 0.128 Ruggedness Index 0.114
Eating Places 0.127 Sch. Quality (Middle) 0.102
Clothing Stores 0.110 Multifamily Housing 0.079
Other Race 0.105 Sports / Music Stores 0.078
Sch. Quality (Middle) 0.102 Other Race 0.075
Sports / Music Stores 0.095 Clothing Stores 0.056
Electronics Stores 0.082 Electronics Stores 0.053
Coffee Shops 0.063 Eating Places 0.046
Ruggedness Index 0.051 Urban Heat Island Index 0.045
Furniture Stores 0.050 Full Restaurants 0.044
Bars 0.044 Personal Crime 0.043
Asian 0.042 Supplies Stores 0.042
Fast Food 0.036 Furniture Stores 0.040
Sch. Quality (High) 0.034 Sch. Quality (High) 0.038
Cosmetics Stores 0.027 Coffee Shops 0.035
Supplies Stores 0.026 Property Crime 0.033
Used Merchandise Stores 0.026 Precip. (Oct) 0.030
Food / Health Stores 0.023 Precip. (July) 0.030
Single-Family Housing 0.022 Fast Food 0.029
African American 0.021 Single-Family Housing 0.028
Precip. (July) 0.020 Food / Health Stores 0.025
Precip. (Jan) 0.018 Pollution Score (0-10) 0.022
Solar Radiation (July) 0.017 Non-Residential 0.022
Solar Radiation (Oct) 0.014 Solar Radiation (Oct) 0.016
Precip. (Oct) 0.011 Cosmetics Stores 0.016
Pet Stores 0.010 African American 0.014
Department Stores 0.009 Department Stores 0.012
Solar Radiation (Jan) 0.009 Bars 0.011
Non-Residential 0.006 Asian 0.010
Personal Crime 0.004 Used Merchandise Stores 0.010
Solar Radiation (April) 0.004 Pet Stores 0.008
Zoning Not Determined 0.002 Solar Radiation (July) 0.008
Multifamily Housing 0.002 Precip. (April) 0.006
Pollution Score (0-10) 0.002 Traffic 0.006
Urban Heat Island Index 0.002 Precip. (Jan) 0.003
Property Crime 0.002 Mixed Development 0.001
Precip. (April) 0.001 Solar Radiation (Jan) 0.001
Traffic 0.000 Zoning Not Determined 0.000
Mixed Development 0.000 Solar Radiation (April) 0.000
Table 3.2: R Sq. Values (Ranked in Descending Order)
75

Dependent variable:
Full X Exog. Component Only
Tra
ffic (log) -0.002 -0.008
(0.003) (0.005)
Pollution Score (0-10) -0.003 0.011∗∗
(0.003) (0.005)
Ruggedness Index (log) 0.023∗∗∗ 0.044∗∗∗
(0.003) (0.006)
Urban Heat Island Index (log) 0.009∗∗∗ -0.007∗∗∗
(0.001) (0.002)
Race (shares) White 0.469∗∗∗ 0.419∗∗∗
(0.043) (0.071)
African American 0.280∗∗∗ 0.245∗∗∗
(0.048) (0.078) Asian 0.350∗∗∗ 0.166∗∗
(0.045) (0.073)
Latino 0.268∗∗∗ 0.048
(0.041) (0.068) Other 0.334 -1.570∗∗∗
(0.237) (0.388)
Crime (log)
Property Crime -0.024∗∗∗ -0.033∗∗∗
(0.008) (0.013)
Personal Crime 0.024∗∗∗ 0.016
(0.009) (0.014)
Zoning (shares)
Multifamily Residences 0.073∗∗∗ -0.068∗∗∗
(0.015) (0.025) Mixed Use 0.049∗ 0.058
(0.027) (0.045) Non-determined 0.045 0.163∗∗∗
(0.030) (0.049) Non-residential 0.049∗∗∗ 0.144∗∗∗
(0.015) (0.025)
Single-family Residences 0.089∗∗∗ 0.121∗∗∗
(0.015) (0.025)
School Quality (0-100)
Elementary School 0.0002 0.001∗∗
(0.0002) (0.0003) Middle School 0.0001 0.0003
(0.0001) (0.0002)
High School 0.0002∗∗ 0.001∗∗∗
(0.0001) (0.0002)
Precipitation
January 0.001 -0.004∗∗∗
(0.001) (0.001)
April -0.014∗∗∗ -0.001
(0.002) (0.004)
July -0.053∗∗∗ -0.046∗∗∗
(0.008) (0.013) October 0.007∗∗∗ 0.013∗∗∗
(0.003) (0.004)
Solar Radiation
January -0.021∗∗ 0.029
∗
(0.009) (0.015)
April -0.006 0.011
(0.008) (0.014)
July -0.014∗∗ -0.006
(0.006) (0.009) October 0.007 0.035∗∗∗
(0.008) (0.013)
Consumption Amenities (log)
Eating Places 0.023∗∗ 0.003
(0.010) (0.016)
Full Restaurants 0.004 0.013
(0.009) (0.015)
Fast Food Places -0.014∗∗ 0.004
(0.006) (0.010) Coffee Shops 0.003 -0.008
(0.008) (0.013)
Bars -0.002 -0.030∗∗
(0.008) (0.013)
Furniture Stores 0.0003 0.008
(0.005) (0.008)
Electronic Stores 0.012∗∗ 0.006
(0.005) (0.009)
Supplies Stores 0.006 0.032∗∗∗
(0.006) (0.010)
Clothing Stores 0.013∗∗∗ 0.012
∗
(0.004) (0.006)
Sport / Music Stores 0.009 0.017
∗
(0.006) (0.009)
Department Stores -0.015 0.012
(0.013) (0.021)
Pet Stores -0.020 -0.025
(0.016) (0.026) Used Merchandise Stores 0.001 -0.012
(0.010) (0.016) Cosmetic Stores 0.0003 -0.010
(0.010) (0.017)
Food / Health Stores -0.015 0.014
(0.012) (0.019)
Obs 2,345 2,345 Adjusted R2 0.398 0.403
Note:
∗
p
<0.1; ∗∗
p
<0.05; ∗∗∗
p
<0.01
Table 3.3: Regression Results
76

Full X Exog. Component Only
Pollution Score (0-10) 0.00233 0.00517
Ruggedness Index 0.01968 0.03796
Urban Heat Island Index 0.02071 0.01248
White & Asian 0.07386 0.05359
White 0.06283 0.05777
Asian 0.01172 0.00483
Single-Family Housing 0.00970 0.00693
Multifamily Housing 0.00391 0.04760
Personal Crime 0.00278 0.01832
Property Crime 0.00265 0.01755
Sch. Quality (Elem.) 0.03095 0.03500
Sch. Quality (Middle) 0.02054 0.02009
Sch. Quality (High) 0.00624 0.00854
Precip. (July) 0.02086 0.00648
Precip. (Oct) 0.00453 0.00649
Solar Radiation (July) 0.00908 0.00353
Solar Radiation (Oct) 0.00427 0.01266
Eating Places 0.06261 0.02197
Total 0.3647 0.3724
Table 3.4: R Sq. Decomposition
3.5 Discussion
From the main regression, pollution, ruggedness, urban heat island effect, race shares,
property crime, zoning types, elementary and high school quality, lower precipitation in
January and July, higher solar radiation in January and October, and supplies / sports and
music stores are important for determining X (Table 3.3). From the R sq. decomposition,
the most important variable seems to be the share of White residents (Table 3.4).
Our future steps for research include:
77

1. What is the mechanism by which share of White residents affects X? Is this a question of path dependence? Perhaps White residents were the first to occupy the most
attractive locations within the city and those locations remain attractive to this day.
What is it about a location with many White residents that makes it attractive? In
reference to existing literature on “white flight” and racial tipping points (Boustan, 2007; Boustan et al., 2023; Blair, 2023), are there any thresholds beyond which
neighborhoods decline significantly in attractiveness?
2. Thus far, we have assumed that all variables have a linear relationship with X. However, more could be done to find the right functional form. For instance, using
binned variables or adding polynomial terms. We could also use machine learning
methods to find the functional form that maximizes fit. Which variables should we
include? What is the relationship of each of those variables to the amenity X?
3. Are there any other empirical strategies which could improve fit? For example, the
use of instrumental variables or regression discontinuities.
4. Finally, local amenities might be group-specific. Neighborhoods might be particularly attractive to certain groups e.g. retired households, college-educated workers,
racial enclaves, etc. If so, we can calibrate different versions of X using different
groups in society and examine the extent to which different groups favor different
types of observable amenities.
78

Bibliography
Abu, M., Heath, S. C., Adger, W. N., Codjoe, S. N. A., Butler, C., and Quinn, T. (2024).
Social consequences of planned relocation in response to sea level rise: impacts on
anxiety, well-being, and perceived safety. Scientific Reports, 14(1):3461.
Ahlfeldt, G. M., Redding, S. J., Sturm, D. M., and Wolf, N. (2015). The economics of
density: Evidence from the Berlin Wall. Econometrica, 83(6):2127–2189.
Akbar, P., Couture, V., Duranton, G., and Storeygard, A. (2023a). Mobility and congestion
in urban india. Am. Econ. J. Appl. Econ.
Akbar, P. A., Couture, V., Duranton, G., Ghani, E., and Storeygard, A. (2018). Mobility
and congestion in urban India. The World Bank.
Akbar, P. A., Couture, V., Duranton, G., and Storeygard, A. (2023b). The fast, the slow,
and the congested: Urban transportation in rich and poor countries. NBER Working
Paper.
Albouy, D., Graf, W., Kellogg, R., and Wolff, H. (2016). Climate amenities, climate change,
and american quality of life. Journal of the Association of Environmental and Resource
Economists, 3(1):205–246.
Almagro, M. and Domínguez-Iino, T. (2022). Location sorting and endogenous amenities:
Evidence from amsterdam. Available at SSRN 4279562.
Ang, A., Christensen, P., and Vieira, R. (2020). Should congested cities reduce their speed
limits? evidence from são paulo, brazil. Journal of Public Economics, 184:104155.
Balboni, C. A. (2019). In harm’s way? infrastructure investments and the persistence of coastal
cities. PhD thesis, London School of Economics and Political Science.
Barnard, J. R. (1978). Externalities from urban growth: The case of increased storm runoff
and flooding. Land Econ., 54(3):298–315.
Baum-Snow, N. and Han, L. (2019). The microgeography of housing supply. Work in
progress, University of Toronto.
Bayham, J., Yoder, J. K., Champ, P. A., and Calkin, D. E. (2022). The economics of wildfire
in the united states. Annual Review of Resource Economics, 14:379–401.
79

Baylis, P. and Boomhower, J. (2023). The economic incidence of wildfire suppression in
the united states. American Economic Journal: Applied Economics, 15(1):442–473.
Baylis, P. W. and Boomhower, J. (2022). Mandated vs. voluntary adaptation to natural
disasters: the case of us wildfires. Technical report, National Bureau of Economic Research.
Bento, A. M., Cropper, M. L., Mobarak, A. M., and Vinha, K. (2005). The effects of urban spatial structure on travel demand in the united states. Review of Economics and
Statistics, 87(3):466–478.
Blair, P. Q. (2023). Beyond racial attitudes: The role of outside options in the dynamics
of white flight. Technical report, National Bureau of Economic Research.
Borgschulte, M., Molitor, D., and Zou, E. Y. (2022). Air pollution and the labor market:
Evidence from wildfire smoke. Review of Economics and Statistics, pages 1–46.
Boustan, L., Cai, C., and Tseng, T. (2023). Jue insight: White flight from asian immigration: Evidence from california public schools. Journal of Urban Economics, page 103541.
Boustan, L. P. (2007). Black migration, white flight: the effect of black migration on
northern cities and labor markets. The Journal of Economic History, 67(2):484–488.
Burke, M., Driscoll, A., Heft-Neal, S., Xue, J., Burney, J., and Wara, M. (2021). The changing risk and burden of wildfire in the united states. Proc. Natl. Acad. Sci. U. S. A.,
118(2).
Cruz, J.-L. and Rossi-Hansberg, E. (2021). The economic geography of global warming.
Technical report, National Bureau of Economic Research.
Davis, M. A. and Ortalo-Magné, F. (2011). Household expenditures, wages, rents. Review
of Economic Dynamics, 14(2):248–261.
Delventhal, M. and Parkhomenko, A. (2023). Spatial implications of telecommuting.
Available at SSRN 3746555.
Delventhal, M. J., Kwon, E., and Parkhomenko, A. (2022). JUE insight: How do cities
change when we work from home? Journal of Urban Economics, 127:103331.
Desmet, K., Kopp, R. E., Kulp, S. A., Nagy, D. K., Oppenheimer, M., Rossi-Hansberg, E.,
and Strauss, B. H. (2018). Evaluating the economic cost of coastal flooding. Technical
report, National Bureau of Economic Research.
Diamond, R. (2016). The determinants and welfare implications of US workers’ diverging
location choices by skill: 1980–2000. American Economic Review, 106(3):479–524.
Dillon, G. K., Scott, J. H., Jaffe, M. R., Olszewski, J. H., Vogler, K. C., Finney, M. A., Short,
K. C., Riley, K. L., Grenfell, I. C., Jolly, W. M., and Brittain, S. (2023). Spatial datasets of
probabilistic wildfire risk components for the United States (270m) (3rd edition). Forest
Service Research Data Archive. Source: https://doi.org/10.2737/RDS-2016-0034-3.
80

Glaeser, E. L. and Kahn, M. E. (2010). The greenness of cities: Carbon dioxide emissions
and urban development. Journal of urban economics, 67(3):404–418.
Google (2009). Google official blog. Accessed on 4/18/2018, Source: Link.
Haddad, E. A. and Teixeira, E. (2015). Economic impacts of natural disasters in megacities: The case of floods in são paulo, brazil. Habitat International, 45:106–113.
Hanna, R., Kreindler, G., and Olken, B. A. (2017). Citywide effects of high-occupancy
vehicle restrictions: Evidence from “three-in-one” in jakarta. Science, 357(6346):89–
93.
Hauer, M. E. (2019). Population projections for us counties by age, sex, and race controlled to shared socioeconomic pathway. Scientific data, 6(1):1–15.
He, Y., Thies, S., Avner, P., and Rentschler, J. (2021). Flood impacts on urban transit and
accessibility—a case study of kinshasa. Transportation Research Part D: Trans. Environ.,
96:102889.
Heblich, S., Redding, S. J., and Sturm, D. M. (2020). The making of the modern metropolis: evidence from London. The Quarterly Journal of Economics, 135(4):2059–2133.
Hettiarachchi, Wasko, and others (2018). Increase in flood risk resulting from climate
change in a developed urban watershed–the role of storm temporal patterns. Hydrol.
Earth Syst. Sci.
Hornbeck, R. and Moretti, E. (2018). Who benefits from productivity growth? direct and
indirect effects of local tfp growth on wages, rents, and inequality. Technical report,
National Bureau of Economic Research.
Hsiao, A. (2021). Coordination and commitment in international climate action: Evidence from palm oil. Working Paper. Source: https://allanhsiao.com/files/Hsiao_
palmoil.pdf.
Issler, P., Stanton, R., Vergara-Alert, C., and Wallace, N. (2020). Mortgage markets with
climate-change risk: Evidence from wildfires in California. Available at SSRN 3511843.
Kahn, M. E. (2000). The environmental impact of suburbanization. Journal of policy
analysis and management, 19(4):569–586.
Keeley, J. E., Guzman-Morales, J., Gershunov, A., Syphard, A. D., Cayan, D., Pierce, D. W.,
Flannigan, M., and Brown, T. J. (2021). Ignitions explain more than temperature or
precipitation in driving santa ana wind fires. Science advances, 7(30):eabh2262.
Khanna, G., Medina, C., Nyshadham, A., Ramos, D., Tamayo, J., and Tiew, A. (2020).
Spatial mobility, economic opportunity, and crime. NEUDC presentation.
Khanna, G., Medina, C., Nyshadham, A., Ramos, D., Tamayo, J., and Tiew, A. (2023).
Spatial mobility, economic opportunity, and crime. Working paper.
81

Kocornik-Mina, A., McDermott, T. K. J., Michaels, G., and Rauch, F. (2020). Flooded cities.
Am. Econ. J. Appl. Econ., 12(2):35–66.
Komissarova, K. (2022). Location Choices over the Life Cycle: The Role of Relocation for
Retirement. Job Market Paper.
Kulp, S. A. and Strauss, B. H. (2019). New elevation data triple estimates of global vulnerability to sea-level rise and coastal flooding. Nature communications, 10(1):1–12.
Lee, S. and Lin, J. (2018). Natural amenities, neighbourhood dynamics, and persistence
in the spatial distribution of income. The Review of Economic Studies, 85(1):663–694.
Lekuthai, A. and Vongvissessomjai, S. (2001). Intangible flood damage quantification.
Water Resource Management.
Lepeule, J., Laden, F., Dockery, D., and Schwartz, J. (2012). Chronic exposure to fine
particles and mortality: an extended follow-up of the Harvard Six Cities study from
1974 to 2009. Environmental health perspectives, 120(7):965–970.
Lin, W. and Wooldridge, J. M. (2019). Testing and correcting for endogeneity in nonlinear
unobserved effects models. In Panel data econometrics, pages 21–43. Elsevier.
Lu, Q.-C., Peng, Z.-R., and Du, R. (2012). Economic analysis of impacts of sea level rise
and adaptation strategies in transportation. Transportation Research Record, 2273(1):54–
61.
Mangum, K. (2017). The role of housing in carbon emissions. Andrew Young School of
Policy Studies Research Paper Series, (17-05).
McCoy, S. J. and Walsh, R. P. (2018). Wildfire risk, salience & housing demand. Journal of
Environmental Economics and Management, 91:203–228.
McGrath, H., Kotsollaris, M., Stefanakis, E., and Nastev, M. (2019). Flood damage calculations via a RESTful API. International Journal of Disaster Risk Reduction, 35:101071.
Merz, B., Blöschl, G., Vorogushyn, S., Dottori, F., Aerts, J. C. J. H., Bates, P., Bertola, M.,
Kemter, M., Kreibich, H., Lall, U., and Macdonald, E. (2021). Causes, impacts and
patterns of disastrous river floods. Nature Reviews Earth & Environment, 2(9):592–609.
Michael, J. A. (2007). Episodic flooding and the cost of sea-level rise. Ecol. Econ.,
63(1):149–159.
Moftakhari, H. R., AghaKouchak, A., Sanders, B. F., Allaire, M., and Matthew, R. A.
(2018). What is nuisance flooding? defining and monitoring an emerging challenge.
Water Resour. Res., 54(7):4218–4227.
Moftakhari, H. R., Salvadori, G., AghaKouchak, A., Sanders, B. F., and Matthew, R. A.
(2017). Compounding effects of sea level rise and fluvial flooding. Proc. Natl. Acad. Sci.
U. S. A., 114(37):9785–9790.
82

Monte, F., Redding, S. J., and Rossi-Hansberg, E. (2018). Commuting, Migration, and
Local Employment Elasticities. American Economic Review, 108(12):3855–3890.
Moretti, E. (2017). Wildfires aren’t the only threat to the california dream. New York Times
Op. Ed. Source: https://eml.berkeley.edu/~moretti/nyt3.pdf.
Newman, R. and Noy, I. (2023). The global costs of extreme weather that are attributable
to climate change. Nat. Commun., 14(1):6103.
Nunn, N. and Puga, D. (2012). Ruggedness: The blessing of bad geography in africa.
Review of Economics and Statistics, 94(1):20–36.
O’Donnell, E. C. and Thorne, C. R. (2020). Drivers of future urban flood risk. Philosophical
Transactions of the Royal Society A, 378(2168):20190216.
Ospital, A. (2023). Urban policy and spatial exposure to environmental risk. Job Market
Paper. Source: https://www.aospital.com/uploads/ospital_jmp.pdf.
Parkhomenko, A. and Delventhal, M. J. (2023). Spatial implications of telecommuting in
the united states.
Paulik, R., Stephens, S., Wild, A., Wadhwa, S., and Bell, R. G. (2021). Cumulative building
exposure to extreme sea level flooding in coastal urban areas. International Journal of
Disaster Risk Reduction, 66:102612.
Pour, S. H., Wahab, A. K. A., Shahid, S., Asaduzzaman, and Dewan, A. (2020). Low impact development techniques to mitigate the impacts of climate-change-induced urban
floods: Current trends, issues and challenges. Sustainable Cities and Society, 62:102373.
Radeloff, V. C., Helmers, D. P., Kramer, H. A., Mockrin, M. H., Alexandre, P. M., BarMassada, A., Butsic, V., Hawbaker, T. J., Martinuzzi, S., Syphard, A. D., et al. (2018).
Rapid growth of the US wildland-urban interface raises wildfire risk. Proceedings of the
National Academy of Sciences, 115(13):3314–3319.
Redding, S. J. and Rossi-Hansberg, E. (2017). Quantitative Spatial Economics. Annual
Review of Economics, 9:21–58.
Rentschler, J., Avner, P., Marconcini, M., Su, R., Strano, E., Vousdoukas, M., and Hallegatte, S. (2023). Global evidence of rapid urban growth in flood zones since 1985.
Nature, 622(7981):87–92.
Roncancio, D. J. and Nardocci, A. C. (2016). Social vulnerability to natural hazards in são
paulo, brazil. Natural Hazards, 84(2):1367–1383.
Severen, C. (2021). Commuting, labor, and housing market effects of mass transportation:
Welfare and identification. Review of Economics and Statistics, pages 1–99.
Sharifi, A. and Yamagata, Y. (2018). Resilience-oriented urban planning. In Resilienceoriented urban planning, pages 3–27. Springer.
83

Short, K. C. (2021). Spatial wildfire occurrence data for the United States, 1992-2018
[fpa_fod_20210617]. Source: https://www.fs.usda.gov/rds/archive/Catalog/
RDS-2013-0009.5.
Syphard, A. D., Radeloff, V. C., Keeley, J. E., Hawbaker, T. J., Clayton, M. K., Stewart,
S. I., and Hammer, R. B. (2007). Human influence on california fire regimes. Ecological
applications, 17(5):1388–1402.
Taberna, A., Filatova, T., Hochrainer-Stigler, S., Nikolic, I., and Noll, B. (2023). Economic implications of autonomous adaptation of firms and households in a resourcerich coastal city. Scientific Reports, 13(1):20348.
Tomás, L. R., Soares, G. G., Jorge, A. A. S., Mendes, J. F., Freitas, V. L. S., and Santos, L.
B. L. (2022). Flood risk map from hydrological and mobility data: A case study in são
paulo (brazil). Trans. GIS.
Turner, M. C., Krewski, D., Diver, W. R., Pope III, C. A., Burnett, R. T., Jerrett, M., Marshall, J. D., and Gapstur, S. M. (2017). Ambient air pollution and cancer mortality in
the cancer prevention study ii. Environmental health perspectives, 125(8):087013.
Valentinyi, A. and Herrendorf, B. (2008). Measuring factor income shares at the sectoral
level. Review of Economic Dynamics, 11(4):820–835.
Veldhuis, J. A. E. (2011). How the choice of flood damage metrics influences urban flood
risk assessment. J. Flood Risk Manag., 4(4):281–287.
Wang, D., Guan, D., Zhu, S., Kinnon, M. M., Geng, G., Zhang, Q., Zheng, H., Lei, T., Shao,
S., Gong, P., and Davis, S. J. (2020). Economic footprint of California wildfires in 2018.
Nature Sustainability, 4(3):252–260.
Wibbenmeyer, M. and Robertson, M. (2022). The distributional incidence of wildfire
hazard in the western united states. Environmental Research Letters, 17(6):064031.
Wobus, C., Zheng, P., Stein, J., Lay, C., Mahoney, H., Lorie, M., Mills, D., Spies, R., Szafranski, B., and Martinich, J. (2019). Projecting changes in expected annual damages from
riverine flooding in the united states. Earths Future, 7(5):516–527.
84

Abstract (if available)

Abstract This dissertation comprises three essays examining household location choice and the cost of climate change. The first essay identifies an inverted U-shaped relationship be- tween population density and wildfires in the US. I conduct a case study of Los Ange- les County, and find that risk-proportional taxation is an effective means for addressing higher wildfire risk. The second essay assesses the economic impact of flood-induced travel delays in São Paulo, Brazil, estimating an annual cost of $125 million USD, equivalent to 0.06 % of the city’s GDP. The third essay utilizes a spatial equilibrium model to analyze local residential amenity measures, informing infrastructure investment strategies and managed retreat considerations in areas susceptible to natural disasters.

Linked assets

University of Southern California Dissertations and Theses

Conceptually similar

PDF

Essays in climate change adaptation: role of market power in incentivizing adaptation behavior

PDF

The impact of digital transformation on urban communities, welfare, and distributional outcomes

PDF

Essays on the dual urban-rural system and economic development in China

PDF

Essays in environmental economics

PDF

Sustaining open source software production: an empirical analysis through the lens of microeconomics

PDF

Essays on the economics of climate change adaptation in developing countries

PDF

A spatial analysis of beef production and its environmental and health impacts in Texas

PDF

Endogenous H2O2 signaling positively regulates the release of neuropeptides during a neuron-gut axis mediated oxidative stress response in caenorhabditis elegans

PDF

Income eligible school choice: the effect of parent knowledge on advocacy

PDF

Climate change, air pollution and health: how policy can induce adaptation

PDF

Essays on development economics and adolescent behavior

PDF

Essays on narrative economics, Climate macrofinance and migration

PDF

Essays on environmental economics: education, employment, and experiments

PDF

Three essays on strategic commuters with late arrival penalties and toll lanes

PDF

From first principles to machine intelligence: explaining charge carrier behavior in inorganic materials

PDF

Predicting autism severity classification by machine learning models

PDF

Bright portals: illness & the environment in contemporary poetry

PDF

The metaphysics of social groups: structure, location, and change

PDF

Stolen culture: a discussion on decolonization and repatriation of Indigenous artifacts

PDF

Spatial dimensions of stratification: neighborhood change, urban inequality, and the neighborhood-school link in the U.S.

Asset Metadata

Creator Ang, Qi Qi Amanda (author)

Core Title Location choice and the costs of climate change

School College of Letters, Arts and Sciences

Degree Doctor of Philosophy

Degree Program Economics

Degree Conferral Date 2024-05

Publication Date 04/17/2024

Defense Date 04/15/2024

Publisher Los Angeles, California (original), University of Southern California (original), University of Southern California. Libraries (digital)

Tag climate change,Economics,environmental,location choice,natural disasters,OAI-PMH Harvest,quantitative spatial models,Urban

Format theses (aat)

Language English

Contributor Electronically uploaded by the author (provenance)

Advisor Kahn, Matthew (committee member), Oliva, Paulina (committee member), Parkhomenko, Andrii (committee member)

Creator Email amandaangqiqi@gmail.com,qang@usc.edu

Permanent Link (DOI) https://doi.org/10.25549/usctheses-oUC113877817

Unique identifier UC113877817

Identifier etd-AngQiQiAma-12820.pdf (filename)

Legacy Identifier etd-AngQiQiAma-12820

Document Type Dissertation

Format theses (aat)

Rights Ang, Qi Qi Amanda

Internet Media Type application/pdf

Type texts

Source 20240417-usctheses-batch-1141 (batch), University of Southern California (contributing entity), University of Southern California Dissertations and Theses (collection)

Access Conditions The author retains rights to his/her dissertation, thesis or other graduate work according to U.S. copyright law. Electronic access is being provided by the USC Libraries in agreement with the author, as the original true and official version of the work, but does not grant the reader permission to use the work if the desired use is covered by copyright. It is the author, as rights holder, who must provide use permission if such use is covered by copyright.

Repository Name University of Southern California Digital Library

Repository Location USC Digital Library, University of Southern California, University Park Campus MC 2810, 3434 South Grand Avenue, 2nd Floor, Los Angeles, California 90089-2810, USA

Repository Email cisadmin@lib.usc.edu