Essays in Environmental & Resource Economics by Léopold Temoana

Essays in Environmental & Resource Economics

L´eopold Temoana Marc Biardeau

A dissertation submitted in partial satisfaction of the

requirements for the degree of

Doctor of Philosophy

Agricultural and Resource Economics

in the

Graduate Division

of the

University of California, Berkeley

Committee in charge:

Professor Maximilian Auﬀhammer, Chair Professor Solomon Hsiang, Co-chair Professor Lucas W. Davis Professor Gordon Rausser

Summer 2020

Abstract

Essays in Environmental & Resource Economics

L´eopold Temoana Marc Biardeau

Doctor of Philosophy in Agricultural and Resource Economics

University of California, Berkeley

Professor Maximilian Auﬀhammer, Chair

Professor Solomon Hsiang, Co-chair

At its core, environmental & resource economics seeks to identify and correct market failures, i.e. situations in which markets fail to allocate resources in a way that maximizes society’s economic welfare. Some of the greatest issues of our time such as water and air pollution, the over-exploitation of natural resources, or climate change all result from market failures. While market failures can take many forms, they are often associated with time-inconsistent preferences, negative externalities, or the diﬃculty to provide and manage common goods. In this dissertation, I investigate three key issues relative to each of these types of market failure.

In the first chapter, coauthored with Solomon Hsiang and SébastienAnnan-Phan, we propose that the probability that individuals focus attention on a moment a fixed temporal distance from their present moment is stable. We call Kernel of Attention to Time (KAT) the resulting probability distribution across moments in relative time, which directs human attention across the past, present, and future. We then analyze how populations across the world query Google Search for information related to specific moments in time and provide the first evidence of a coherent KAT for most humans on Earth. We discover consistent structure to the distribution of attention across time, regardless of populations’ language or country, with the present strongly dominating all other moments and capturing roughly 25% of time-related attention on average. Attention to the past and future decays rapidly with increasing temporal distance, much faster than exponentially.

Despite consistency in the form of the KAT around the world, we ﬁnd regional patterns in attention to the past, present, and future. Furthermore, it appears that over the last decade, attention to the present has been increasing at the expense of attention to the past. Together, these ﬁndings suggest that human populations exhibit strong common patterns 2

of thought with respect to time, but some non-biological factors that vary across space and over time can alter these patterns.

While the KAT does not capture time-related economic tradeoﬀs directly (e.g. foregoing present consumption to increase future consumption), the structure of its future-oriented portion could enable better understanding of the origin of time-based preferences and provide new insights on time-inconsistent behavior.

In the second chapter, coauthored with Lucas Davis, Paul Gertler and Catherine Wolfram, we develop new measures of global air conditioning potential using temperature data from more than 14,000 monitoring stations around the world. We combine this information with disaggregated global population estimates to calculate cooling degree days (CDDs) and other measures of air conditioning potential by region, country, and city. Overall, the evidence points to enormous potential growth in air conditioning, particularly in low-income and middle-income countries. India, China, Indonesia, Nigeria, Pakistan, Brazil, Bangladesh, and the Philippines all have greater air conditioning potential than the United States, a country where a staggering 400 terawatt hours of electricity are currently used annually for air conditioning. We find, moreover, that a significant portion of total global air conditioning potential comes from the earth’s largest cities. Mumbai, for example, has by itself the air conditioning potential of 25% of the entire United States. Our estimates imply that, were global air conditioning usage to reach U.S. levels, total electricity consumption worldwide for air conditioning could reach 20,000 terawatt hours annually, which roughly corresponds to the current net global electricity consumption. If unmitigated by improvements in air conditioner energy-efficiency or updated power network infrastructures, this rise in overall electricity demand could generate two negative externalities. First, it could lead to black- outs around the globe, especially in low and middle income countries. What is more, most electricity worldwide continues to rely on fossil fuels. Consequently, growing air conditioner adoption could lead to hundreds of millions of tons of increased carbon dioxide emissions, further aggravating climate change.

In the third and last chapter, coauthored with David Zilberman, we rely on satellite-based data tracking vessel fishing hours to investigate the extent to which Very Large Marine Protected Areas (VLMPAs), – Marine Protected Areas spanning at least 100,000 km2,– prohibiting all types of fishing have been successful at deterring fishing effort. These VLM- PAs have been created in an attempt to protect and replenish fish stocks, 34.5% of which have fallen below biologically sustainable levels,1 partly as a result of overfishing. Indeed, most fisheries have traditionally been open access resources, leading individually-acting fishers to collectively extract more than the efficient level and threatening the viability of the resource over time, a situation known as the tragedy of the commons.2 3

In spite of their large size which may constitute a challenge for enforcement, we find that VLMPAs have on average been able to deter fishing effort. However, a case-by-case analysis reveals varying levels of success, with the most successful VLMPA managed by the Republic of Kiribati and the worst performing one managed by the United States.

To better understand the nature of illegal fishing effort in these VLMPAs, we focus on the characteristics of the vessels infringing the fishing bans in these VLMPAs and find that most of the infractions can be traced back to a few industrialized countries. i

Contents

Contents i

List of Figures ii

List of Tables iii

Acknowledgements iv

1 The Distribution of Human Attention to Moments in Time1 1.1 Introduction...... 1 1.2 Model: The Kernel of Attention to Time (KAT)...... 2 1.3 Estimation: Recovering the KAT using Google Search ...... 6 1.4 Results...... 6 1.5 Discussion...... 11

2 Population, Warming, and Global Air Conditioning 15 2.1 Introduction...... 15 2.2 Methods...... 16 2.3 Results...... 22 2.4 Discussion...... 29

3 Evaluating the Eﬀectiveness of Very Large Marine Protected Areas at Deterring Fishing Eﬀort 30 3.1 Introduction...... 30 3.2 Methods...... 34 3.3 Results...... 38 3.4 Discussion...... 42

References 43

A Full derivation of the Kernel of Attention to time 52 A.1 A Probabilistic Model for Attention to Time...... 52 A.2 Predictions for Internet Search Volume...... 53 ii

A.3 Google Search Data...... 56 A.4 Empirical estimation of the KAT ...... 57

B Supplemental Figures and Tables 61 B.1 Figures...... 61 B.2 Tables...... 73

List of Figures

1.1 Google search volume reﬂects changes in relative attention over time...... 3 1.2 Attention to time expressed through Google search queries...... 5 1.3 Google search volume for time-related queries and KAT model predictions... 7 1.4 The estimated Kernel of Attention to Time (KAT) ...... 9 1.5 Relative attention to past, present, and future around the world and over time . 11 1.6 The relative attention to the past, present and future for individual countries . . 12

2.1 Simulation results...... 18 2.2 Reporting Frequencies for Monitoring Stations...... 19 2.3 Global Cooling Degree Days...... 23 2.4 Comparing Our Estimates to Previous Estimates ...... 25 2.5 Does Population-Weighting Matter?...... 26 2.6 Robustness check...... 28

3.1 List of all the Very Large Marine Protected Areas (VLMPAs) ...... 32 3.2 Evolution of daily hours of fishing effort in each VLMPA per 1,000 km2. . . . . 33 3.3 Displacement of fishing effort ...... 36 3.4 Quartic polynomial curves of best fit and their associated 95% confidence intervals 39

B.1 Google search as convolution of KAT with daily dummies...... 62 B.2 Predicted search volume using estimated KATs ...... 63 B.3 Estimated KAT with alternative model choices ...... 64 B.4 Robustness checks for the shares of attention to different moments in time . . . 65 B.5 Estimated KAT using daily, monthly and yearly targets...... 66 B.6 Country boundaries shapefile ...... 67 B.7 Global CDD maps with alternative CDD thresholds...... 68 B.8 Expansion of vessel tracking and VLMPA coverage over time...... 69 B.9 Evolution of average daily hours of fishing effort per 1,000 km2 for all VLMPAs 70 iii

B.10 Displacement of ﬁshing eﬀort ...... 71 B.11 Evolution of the number of vessels by country...... 72

List of Tables

2.1 Rankings by Total Cooling Degree Day Exposure ...... 22

3.1 Area under the 4th-order polynomial curve estimating the difference in annualized fishing hours...... 40 3.2 Ranking of the countries associated with prohibited fishing inside of the established VLMPAs...... 41

B.1 Country-specific holidays...... 73 B.2 Shares of attention to the past, present and future for alphabetically ranked countries 1 to 50 ...... 74 B.3 Shares of attention to the past, present and future for alphabetically ranked countries 51 to 100...... 75 B.4 Shares of attention to the past, present and future for alphabetically ranked countries 101 to 150 ...... 76 B.5 Shares of attention to the past, present and future for alphabetically ranked countries 151 to 181 ...... 77 B.6 Total CDD exposure for the Top 50 Countries...... 78 B.7 Total CDD exposure for Countries Ranked 51 to 100 ...... 79 B.8 Total CDD exposure for Countries Ranked 101 to 150...... 80 B.9 Total CDD exposure for the Top 50 Cities...... 81 B.10 Total CDD exposure for Cities ranked 51 to 100...... 82 B.11 Total CDD exposure for Cities ranked 101 to 150 ...... 83 B.12 Changes in the Country-level counts of Population-Weighted Cooling Degree Days under the Alternative CDD Thresholds...... 84 B.13 Changes in the Country-level ranking of total CDD exposure under the Alterna- tive CDD Thresholds...... 85 B.14 Breakdown of fishing effort in each of the 8 VLMPAs...... 86 B.15 Areas under three distinct order polynomial curves estimating the difference in annualized fishing hours ...... 87 B.16 Characteristics of the eight VLMPAs under consideration...... 88 iv

Acknowledgments

First and foremost, I must give credit to Professor Gordon Rausser, whose precious support allowed me to join the Agricultural & Resource Economics (ARE) PhD program. He ﬁrst entrusted me with a Graduate Student Instructor (GSI) position for his Introduction to Environmental Economics and Policy class back when I was studying at the Goldman School of Public Policy (GSPP). I learned a lot from teaching the sections for this course and from attending Gordon’s lectures. His intimate familiarity with economic policy-making & policy analysis, stemming from his remarkable career in government service, along with his profound knowledge of Environmental & Resource Economics – a ﬁeld which was undoubt- edly shaped by his own research discoveries, – permeated his lectures and convinced me to apply to the ARE department. Gordon subsequently supported my application and, upon my acceptance in the program, took me under his wing for a year as his Graduate Student Researcher (GSR). I learned a lot from the experience and owe him my very presence in the ARE PhD program. I am forever indebted to him.

I was also extremely lucky to have Professor Maximilian Auﬀhammer’s both as my dissertation chair and as a teacher. Our meetings have always been extremely constructive and reassuring. In the classroom, Max’s almost supernatural ability to convey knowledge in a fun and engaging yet rigorous way made me enjoy econometrics to the point of volunteering to teach the discipline both at the undergraduate level and at the graduate level. I can only hope to become one day as true a teacher as he is. His insistence on the primacy of kindness, which he never failed to display himself, did not go unnoticed.

Early on in my academic journey at UC Berkeley, I personally met Professor Solomon Hsiang, whose research had convinced me to apply to the GSPP while I was studying at Sciences Po. Sol is an eminently brilliant yet approachable scientist, with an insatiable de- sire to understand and tackle the great policy challenges of our time. I am grateful to have worked with him for ﬁve years and honored to have him co-chair my dissertation.

I am also grateful to have worked with Professor Lucas Davis. I was thoroughly impressed by Lucas’ unparalleled work ethic, his great attention to detail and his incredible ability to convey complex ideas in the simplest way. He taught me to focus on one idea at once and to stray away from complexity for its own sake.

I am also indebted to my wonderful co-authors S´ebastienAnnan-Phan, Paul Gertler, Catherine Wolfram and David Zilberman. I am truly honored to have been given a chance to work with people constantly pushing the boundary of knowledge. Working with them was an honor and a pleasure. I hope we will get a chance to collaborate again in the future.

I would also like to thank the brilliant administrative and pedagogical staﬀ, both at GSPP and at the ARE department. In particular, I would like to thank Cecille Cabacun- v gan, Martha Chavez and Jalilah Labrie from GSPP, along with Carmen Karahalios and Diana Lazo from ARE. They helped me get back on my feet whenever I encountered a seemingly insurmountable obstacle on my path. Had it not been for their support, I may have given up along the way.

I would like to think the members of my adoptive cohort. I was accepted into the ARE PhD program after having completed most of the coursework for the ﬁrst year while still studying Public Policy at GSPP. As a result, I started the PhD program in 2017 by taking the second-year coursework with the 2016 cohort. They have all been most welcoming. I also greatly appreciated our cohort seminars and our occasional get together even after our coursework ended. I would like to thank Daniel Kannell in particular for being a terriﬁc housemate and an amazing friend.

Last, this journey would not have been possible without the unfaltering support I got from my ever-loving parents, Marie-Fran¸coiseLefèvreand Antoine Biardeau and my ever- optimistic life partner, Cassidy Nicolette. I also owe my thanks to my dauntless stepfather Maurice Brisard and my mother’s ever-welcoming partner Jean-Jacques Ripoll. Je vous aime et vous dédiema thèsede doctorat. 1

Chapter 1

The Distribution of Human Attention to Moments in Time1

1.1 Introduction

In the course of our individual lives, how much do we think about the past, present, and future? How much attention is focused on events tomorrow, compared to the day after? What about a date three years ago? How we as individuals distribute our attention across moments in time aﬀects numerous aspects of society and is intrinsic to the human experience, yet we know little about it quantitatively. For example, we do not know answers to simple questions, such as whether individuals naturally think about the past more often than they think about the future, or sophisticated questions, such as whether attention to the future decays exponentially as some economic models of “discounting” would suggest.3,4 Laboratory-based work has made progress understanding certain aspects of how humans experience the passage of time,5 such as how the brain estimates the length of intervals lasting a fraction of a second,6–8 but experimentally measuring the natural distribution of attention across many moments in time is challenging because asking a subject whether they pay attention to something (e.g. “tomorrow”) focuses their attention on that thing, thereby contaminating their response.9 Thus, while it is understood that there is a limited “budget” of human attention9, 10 such that individuals cannot pay attention to all moments in time with equal focus, how this budget is allocated remains an open question.

Here we propose a quantitative model for how a population of individuals distribute their attention across moments in time and we provide the first measurement of this distribution under natural conditions. As individuals direct their attention towards different objects, ideas, tasks, and events—which we refer to collectively as targets— some will be unrelated to any particular time, such as a poem or abstract life goal, while others are associated with a specific moment in time, such as a birthday or work appointment. When individuals

1The material in this chapter was coauthored with S´ebastienAnnan-Phan and Solomon Hsiang. 2

think about these time-associated targets, some fraction of these thoughts will lead them to execute an internet search that is related to the target, and among these internet searches, a subset will utilize a query whose text contains information that identiﬁes a time period associated with the target (e.g. ‘2012’ or ‘Memorial Day’). By studying patterns in search volume for such time-associated terms, we are able to recover the fraction of attention a population directs towards targets that occur at ﬁxed moments in time.

In this study, we use publicly available data compiled from Google searches11 to study how populations direct their attention at different targets. This approach allows us to study a large portion of all humankind, since roughly 1.8B individuals execute a Google search each day12 (totaling 63,000 searches per second), and it allows us to observe natural patterns of attention passively without interfering in normal behavior. This strategy requires that short- term changes in search volume over time reflects changes in average attention to the targets of those searches, an assumption which appears highly consistent with many patterns in the data.13–16 For examples, queries for “sunscreen” cycle annually, peaking in the Northern Hemisphere summer, queries for “Donald Trump” increased abruptly in 2015 when global attention focused on the then presidential candidate, and queries for “cat” have remained essentially constant for eleven years (Figure 1.1). Thus, at the population level, changes in search query volume are likely a valid measure for changes in attention to targets. For this reason, across the social sciences, these data has proved useful for studying a variety of outcomes, from racial animus17 and individual wellbeing18 to unemployment19 and sexual preferences,20 although, to our knowledge, ours is the first study to use these data to measure how humans experience time.

1.2 Model: The Kernel of Attention to Time (KAT)

Our model is described using two parallel timelines: targets are ﬁxed at locations in absolute time (θ) while a population passes along in query time, where there are observed at each moment (t) only transiently. As the population proceeds through query time, their attention to targets in absolute time evolves as their focus is constantly re-centered on the moment in time that they inhabit (Figure 1.2a). Attention is then allocated across all possible targets depending on their position in absolute time relative to the population’s current position in query time, a distance we call relative time (θ −t) (Figure 1.2b). The situation is analogous to how passengers on a moving train might allocate their attention to stationary landmarks that they pass. Similar to how the gaze of passengers move across landmarks depending on how near or far they are in physical distance from the train at a given moment, the attention of subject populations moves across targets depending on their distance in time.

Under natural conditions, attention in a subject population is likely to ﬂicker rapidly across numerous targets located at diﬀerent times in a seemingly random pattern. For example, it is theoretically possible that the process governing human attention is uniformly 3

Figure 1.1: Google search volume reﬂects changes in relative attention over time

Notes: Google Trends data indicates the total number of queries globally that contain a speciﬁc target string. (a) Queries for the topic “sunscreen” have a cyclical temporal structure that peaks during summer in the Northern Hemisphere. (b). Queries for the topic “donald trump” increased abruptly during the 2016 presidential campaign, which began in 2015. (c) Queries for the topic “cat” are consistently high with little temporal structure. Raw Google Trends data are rescaled by Google, indicating relative volumes of queries with a maximum value of 100. In 2019, Google reported receiving roughly 1.8B queries per day. 4

random, in the sense that this probability is the same for all targets at each moment — although we do not think this example is realistic. Our hypothesis is that the distribution of this flickering attention is governed by a stable and measurable function defined over moments in relative time, which we call the “kernel of attention to time” (KAT) and denote κ(.) (Figure 1.2c). Specifically, we hypothesize that instances of human attention flicker across time-associated targets as if they are being drawn randomly, such that their position in relative time follows the marginal probability density function κ(θ − t) (see Appendix A). As a population moves forward through query time, the distribution of its attention will shift simultaneously such that the KAT always peaks at the corresponding position in absolute time — where relative time equals zero (i.e. θ = t). In the moving train analogy, our model would suggest that while passenger attention across stationary landmarks may appear random, they actually follow a fixed probability distribution that is always anchored around the current position of the train. 5

Figure 1.2: Attention to time expressed through Google search queries

Notes: Search volume for time-specific queries captures the kernel of attention to time for a population of users. (a) Users at time t conduct searches for specific targets (e.g. holidays) that occur at moments in absolute time θ = θ1, θ2, etc. Queries where θ < t indicate attention to past events, θ = t indicates attention to the present, and θ > t indicates attention to the future. (b) A large number of time-related queries can be organized by their position in relative time (θ − t), i.e. the distance in time between when the target occurs (θ) and the time when the query is executed (t). (c) We hypothesize that the probability distribution of time-related queries in relative time is approximately stable for a fixed population. We call this probability distribution the kernel of attention to time (KAT). (d) Data on search queries is organized around the query targets (e.g. searches for “Diwali”), not around the individuals who execute the query. As individuals move through time, passing a target that is fixed in absolute time, attention to the target will evolve as the target is initially in the future, then present, then past. (e) The quantity of attention to a target at a moment is scaled by the KAT. As a population approaches and passes a target, attention to the target traces out a KAT distribution that moves through the time, centered on the moment in time inhabited by the population. (f) Search volume data is recorded based on the time when the search is executed. Thus if search volume scales with the KAT, then a time series of search volume for a single target at a fixed moment will have the same shape as the KAT, but reversed (reflected) in time (e.g. before the event occurs, search volume for the target will reflect attention to the future). (g) When multiple targets span an interval of time X, it is not possible to identify which specific moment is the target of a given search query (e.g. queries containing “september” or “2014” could come from targets on any date during those periods). (h) Each target identified by a query for X will have a distribution of search volume associated with it. (i) The resulting search volume associated with interval X is the sum of search volume distributions for all targets in interval X. This implies that the distribution of search volume for interval X is the convolution of the KAT, shown in (c), with the time series of targets in interval X, shown in (g). 6

1.3 Estimation: Recovering the KAT using Google Search

If the above KAT hypothesis is correct, it generates predictions that can be tested using Google Search data and which allow us to estimate the shape of the KAT. As a population travels through query time and “passes” a target that is fixed at a moment in absolute time (Figure 1.2d), attention allocated to the target will trace out the KAT as the population approaches and then passes the moment of the target (Figure 1.2e). If these changes in attention are reflected in a changing number of search queries, then a time-series of search query volume will be a reversed mirror-image of the KAT, when plotted in query time (Fig- ure 1.2f). Note that query time is the timeline marking when each search query is executed. In cases where a target occupies multiple moments in time (e.g. an event that lasts for a year, Figure 1.2g) then aggregate search volume for that target will be a summation of KAT functions that are all flipped and shifted, each corresponding to a different moment that the target occupies (Figure 1.2h). Thus, if the KAT hypothesis is correct, total search volume for any target (in query time) should be the convolution of the KAT with a time-series of Dirac or indicator-functions that span the absolute time occupied by the target (see Figure 1.2i and Appendix A), scaled by a constant.

Using this result, we should be able to recover the shape of the KAT through a deconvolution of Google Search volume time-series,21, 22 so long as we know when in absolute time the targets occur (see Appendix A). Knowing the date of the target is important because we must compute the distance in relative time between when the search query is executed (t) and when the target occurs (θ). We cannot assign the target of most search queries to a date, but we can do this for the subset of queries that contain date-speciﬁc strings. Speciﬁcally, we estimate the KAT for queries that contain a year in numerical form (e.g. “2016”), a month (e.g. “August”), or a holiday (e.g. “Cinco de Mayo”) during the period 2009-2018 (see Appendix A).

1.4 Results

The temporal structure of Google search volume for holidays, months, and years is well- predicted by the KAT model for human attention (e.g. Figure 1.3). After adjusting for secular trends in search volume, which generally reﬂect expansion of the internet, the KAT model exhibits high predictive accuracy for holidays (R2 = 0.81), months (R2 = 0.83), and years (R2 = 0.93) in our sample.

Figure 1.4a displays the estimated average distribution of attention to time (the KAT, using year targets) for all humans on Earth using Google Search, based on trillions of queries. We estimate that roughly 25% of attention is directed at the present, dramatically dominating any individual moment in the future or past. For example, we estimate that on average, 7

Figure 1.3: Google search volume for time-related queries and KAT model predictions

Notes: (a) Google search volume for queries containing year targets “2015” (top) and “2016” (bottom). Month and year targets produce search distributions similar to (i ﬁg1). (b) Google search volume for queries containing month targets: “april” (top) and “september” (bottom). Spikes span these target months. (c) Global daily Google search volume for queries containing single-day targets that are holidays: “diwali” (top) and “valentine’s day” (bottom). Spikes occur on the holiday date. Single-day targets produce search distributions similar to (Figure 1.2-f). Predictions from our smoothed KAT are plotted in red. 8

attention is 5.7 CI95% = (5.6, 5.8), 18.7 CI95% = (18.3, 19.1), and 28.5 CI95% = (27.5, 29.4) times more likely to be directed at the current day compared to 10 days, 50 days, and 100 days in the future, respectively; and 6.3 CI95% = (6.1, 6.4), 16.0 CI95% = (15.3, 16.8)] and 22.8 CI95% = (21.6, 23.9) times more likely than days of corresponding distance in the past (Figure B.4b).

As targets move further in relative time, the rate (%∆/day) at which attention declines is not constant. This is true for both the future and the past. Attention to dates in the immediate future or past 7 days decline roughly 25% and 30% per day, respectively (red segments, Figure 1.4a). In the more distant future and past, rates of decline fall to 0.6% per day for both the future and past 56 to 150 days (yellow segments). In between is a transition region (1-8 weeks) where rates of change in attention evolve quickly (green and blue segments). For reference, if the KAT were an exponential function, analogous to having a constant discount factor for attention, this rate of decline would be constant.

Although the present period is the single moment receiving the most attention, the total attention allocated to all future or past moments is larger than attention allocated to the present. We compute the total attention allocated to the “future” by summing attention across all periods in the future, and compute similar sums for all periods in the past. In the global average KAT (Figure 1.4a), we estimate that roughly 39% of attention is allocated to the past, while 36% is allocated to the future (see Figure B.4a). It is important to note that these values reflect attention to specific targets associated with identifiable moments in time, but they do not capture attention allocated to abstract notions of “the distant Future” or “the distant Past” that are not associated with specific dates.

The structure of the KAT — a sharp peak in the present, slowing rates of decline with growing temporal distance, and roughly equal mass in the past and future — is broadly reproducible using search data on annual targets from any region of the world, regardless of language, region, or culture. For example, Figure 1.4b shows separate KAT estimates for a diverse selection of twenty diﬀerent countries. While diﬀerences exist across populations, which we explore below, the overall similarity in how populations everywhere allocate attention in relative time is notable.

We also compare the average KAT we estimate using language-independent numerical annual targets (e.g. “2012”) to those estimated using month targets (defined in the local language) or national holidays (see Table B.1). We find that the structure of the KAT is similar regardless of the class of target used (Figure 1.4c-e). However, holiday targets produce a KAT that is relatively more future-oriented, perhaps a result of anticipation specific to holidays, and we have difficulty detecting attention to targets more than 100 days before or after the present using either holiday or monthly targets, due in part to different statistical properties of these data (see Figure B.5). Nonetheless, all three approaches to estimating the KAT indicate a roughly symmetric structure with consistent mathematical form (described 9

Figure 1.4: The estimated Kernel of Attention to Time (KAT)

Notes: (a) displays the average KAT, i.e. the distribution of attention to different moments in time, for countries representing 97% of the global population. The KAT is best represented by the combination of two rational functions, estimated separately on both side of the present day. Exponential functions are also fitted around local portions of the envelope, showing that the discount rate k decreases monotonically as one moves away from the present. Each one of the rational functions can be expressed in terms of the relative distance tot the present day and five parameters (see Appendix A). (b) shows the estimated KAT for a sample of 20 countries using annual query targets. (c)-(e) display the estimated KAT computed from different target lengths: daily targets (c), monthly targets (d), and annual targets (e) for the 20 countries shown in (b) using a selection of four fitting functional forms (Exponential, Hyberbolic, Polynomial, Rational) on the coefficient estimates resulting from the deconvolution (black dots). 10

below) that peaks sharply on the present moment.

In seeking an analytical mathematical form that describes the structure of the KAT, we find that it is generally well approximated by two functions, one for the future and one for the past, where each is from the family of rational functions (i.e. the ratio of two polynomials, green lines in Figure 1.4c-e; also see see Appendix A for analytical forms). Exponential functions cannot fit the structure of the KAT (red lines) while high-order polynomials can approximate the data but have the undesirable feature of being non-monotonic (yellow lines). Following research in behavioral economics and psychology,23–27 we also at- 1 tempt to fit “hyperbolic discounting-like” functions ( 1+k(θ−t) ) that are within the family of rational functions but restrict the numerator to be constant—note that these are not hyperbolic functions. This restriction on the rational function produces visually comparable, albeit poorer, fit to the KAT (blue lines) but results in substantially biased predictions for total search volume of monthly and yearly targets (Figure B.2). Only the unrestricted rational function approximation reliably generalizes across annual, monthly and daily holiday targets.

Although there is striking consistency in the overall structure of the KAT around the world, we explore how the estimated KAT differs across countries. We first estimate a KAT only using search data originating from a single country, then we compute the fraction of attention allocated to the past, present, and future based on that country-specific KAT (see Figure 1.6). To summarize these results, we map the ratio of attention to the future relative to the past (Fig 1.5a) and the fraction of attention allocated to the present (Fig 1.5b). We observe systematic regional patterns. For example, much of tropical Latin America, coastal Africa, and Asia allocates greater attention to the past than the future, while most of North America, Western Europe, North-Central Africa, and Southern South America allocate more attention to the future than past, as do Australia, Malaysia, Taiwan, and Japan (regional outliers). Geographic patterns of attention to the present are less clear. Future work may seek to understand the origin or consequences of these patterns in attention.

Lastly, we examine whether the distribution of attention to moments in time has been recently changing in a coherent way around the world. To do this, we estimate a KAT for each year separately using each class of target (years, months and holidays) and plot these trends in Figure 1.5c. Regardless of which class of target is used, we ﬁnd that between 2009 and 2018, the fraction of attention to the present is rising roughly 0.012 per year globally (p < 0.01), at the expense of the fraction of attention to the past, which is declining roughly by the same amount (0.012, p < 0.05). The fraction of attention allocated to the future appears to have been somewhat stable over this decade. The causes for and consequences of this trend in the distribution of attention is a potentially important question for future investigation. 11

Figure 1.5: Relative attention to past, present, and future around the world and over time

Notes: (a) The ratio of the shares of attention to the future and the past. A ratio greater than one indicates that a people in the country are relatively more attentive to the future than the past. (b) The share of attention to the present. (c) The evolution of the shares allocated of attention towards the past, present, and future, for the twenty countries in Figure 1.4b using day, month, or year search targets.

1.5 Discussion

Here we have proposed that the distribution of human attention to moments in time maintains a coherent probabilistic structure and we recover estimates for this structure, the KAT, for a population of billions of Google users. We discover that the shape of the KAT is 12

Figure 1.6: The relative attention to the past, present and future for individual countries

Notes: (a) Example KAT for countries that are either past-oriented, present-oriented, and future-oriented. (b) Ternary plot highlighting the temporal attentiveness proﬁle of the 181 countries used in the analysis. A country will an equal attentiveness to the past, present, and future would be located in the center of the triangle. The more focused it is toward a direction in time, the closer it would be to the corresponding edge. The three countries: Peru, the United States and France are highlighted in red, green and blue, respectively. (c) Ternary plots highlighting the ﬁve geographical regions the 181 countries in our sample belong to. 13

highly consistent around the world, regardless of the measurement procedure, suggesting that it describes a general pattern of thought fundamental to human minds everywhere. However, we also document some modest differences around the world and over time in attentiveness to the past, present, and future — suggesting that social, cultural, environmental, economic, or other factors can influence how attention is distributed across moments. We believe that disentangling the influence of these factors is a potentially important area for future research.

These findings may inform understanding of basic questions in social and cognitive sciences. For example, economics research has studied how individuals make tradeoffs over time (e.g. foregoing present consumption to increase future consumption), captured in models using “discounting” and “time preference” parameters.3,4, 28, 29 Attention to time does not capture economic tradeoffs directly, but the structure of the future-oriented portion of the KAT could enable better understanding of the origin of time-based preferences. In another example, political science research has examined how constituents evaluate leaders for their past performance.27, 30 The past-oriented portion of the KAT may inform analysis of such constituent retrospection. In a third example, psychology research has considered many ways that thoughts about the past or future influence individuals, such as altering affect31, 32 or contributing to anxiety.33 The methods we present here, perhaps combined with additional data, might provide insights into these relationships or clinical tools to support mental health.

While this analysis is, to our knowledge, the first to present an estimate of the KAT observed in situ, our reliance on Google Search present two challenges that might be ad- dressed in future research using other data sources. First, our approach cannot capture attention to abstract notions of the “distant Future” or “distant Past” if attention directed towards those topics is not associated with specific moments in time that can be queried. For example, the diffuse concept of future generations living centuries from the present will not be well-captured by our approach. Nevertheless, the structure of the KAT we recover might be consistent with how humans think about these distant periods, since finite attention allocated to multi-century periods would likely indicate that only a very small fraction of attention can be spread across individual moments spanning those long intervals.

Second, the sample of aggregated and anonymized Google Search queries we use is not fully representative of the global population. For example, access to the internet and by extension to Google varies from country to country,34 with a higher penetration rate for higher income countries. Google is also used less frequently by older populations,35, 36 and it is not widely used in countries where access is restricted by governments. Thus, it is possible that some of the diﬀerences we observe in aggregated data across locations or over time is due to diﬀerences in the composition and behavior of Google users. Nonetheless, the overall consistency of the KAT we recover around the world might suggest that unobserved populations could behave relatively similarly to those we do observe.

It is also possible that Google searches may be executed by individuals based on some 14 expectation about the availability of online content, with individuals searching less for future targets if they believe associated content does not yet exist. While this behavior may inﬂuence our results, it cannot account for our main ﬁndings. The increase in attention to the present, relative to attention to the future, is too rapid to be explained by changing content: a similarly abrupt increase in online content between sequential moments would require that total searchable online content grows continuously at a rate of 25% per week, orders of magnitude faster than actual content growth. Additionally, a content-driven bias would suggest past-oriented queries should strongly dominate future-oriented queries, which we do not observe.

We believe that understanding how human attention is distributed across time may inform the design of policies, interventions, or institutions. For example, the timing of public health reminders (e.g. “use condoms to avoid STD transmission”) or natural hazard warn- ings (e.g. “store clean water”) could be designed to maximize their beneﬁts by accounting for populations’ overall attentiveness to moments in time, especially when inattention to such messages is costly. In other examples, policies, designed to encourage retirement savings, depend on attentiveness to the future, while the design of census questionnaires that require accurate recall might depend on attentiveness to the past.

We close noting that in discussions surrounding long-term planning, such as managing the global environment, it is sometimes argued that challenges arise because populations do not think enough about the future.37 Our results suggest that while some factors may cause populations to think slightly more or less about the future, the overall general pattern of thought in humans everywhere is to allocate essentially no attention to moments in time further than two-hundred days from the present. Thus, these results might suggest that eﬀorts to directly focus large populations on the distant future are unlikely to succeed. As an alternative, societies could design institutions that systematically evaluate and address long-term challenges, rather than relying on the attentiveness of individual minds. 15

Chapter 2

Population, Warming, and Global Air Conditioning1

2.1 Introduction

Air conditioner sales are booming worldwide, especially in warm countries with growing economies. Thailand, Indonesia, and Vietnam, for example, increased air conditioner sales by 60%, 129%, and 159%, respectively, over the last five years, according to data from Euromonitor International. As is the case with many durable goods, air conditioner adoption follows an “S”-like pattern.38 At low levels of income adoption rates are near zero but then, as incomes rise, adoption can spike dramatically.39, 40 Many low- and middle-income countries are approaching the steep part of this S-curve, and so are poised to experience rapid air conditioner adoption.41 This is mostly good news. Air conditioning brings relief on hot days, makes people more comfortable, and increases productivity.42 During extreme heat events, air conditioning can make the difference between life and death.43, 44 In the United States, heat-related mortality decreased more than 70% with the spread of air conditioning, saving an estimated 20,000 lives each year.45 At the same time, meeting this increase in air conditioning poses an enormous challenge. A typical air conditioner uses 20 times more electricity than a ceiling fan, so air conditioning growth can significantly increase total electricity consumption. In the United States alone, air conditioning uses 400 terawatt-hours of electricity annually, representing 17% of total residential electricity consumption and 12% of total commercial electricity consumption (see the supplementary information for details). Recent studies have also emphasized the role of air conditioning in driving peak electricity demand.46, 47 Growth in air conditioning increases the intensity and frequency of peak events,

1The material in this chapter was published as Heat exposure and global air conditioning in Nature Sus- tainability 3, pages 25–28 (2020). It was coauthored with Lucas Davis, Paul Gertler and Catherine Wolfram. The published version can be found online at https://www.nature.com/articles/s41893-019-0441-9. 16 requiring large investments in electricity generation and transmission infrastructure.48 En- ergy suppliers need accurate predictions about where and when air conditioning adoption will occur if this increased demand is going to be met efficiently. Moreover, most electricity worldwide continues to be generated using fossil fuels. Thus growing air conditioner adoption could mean hundreds of millions of tons of increased carbon dioxide emissions.49 In addition, the refrigerants used in air conditioning are themselves a potent greenhouse gas. Predicting future demand for air conditioning is crucial for the recent Kigali Agreement, which seeks to significantly reduce the use of hydrofluorocarbons.50 In this paper, we compile recent data on population and temperature, two key determi- nants of potential air conditioning use. We use ten years of daily data from 14,500+ global weather monitoring stations to calculate cooling degree days (CDDs) for 219 countries and 1,692 cities. We combine these weather data with highly disaggregated population measures to calculate a measure of CDDs that reflects the climatological conditions where people live. We then multiply population-weighted CDDs by population to get a measure of the total CDD exposure in each country and city. This is the total number of CDDs experienced annually by a country’s (or city’s) population. For example, a country with 1,000,000 people and 3,000 average annual CDDs would have a total CDD exposure of 3 billion. This measure has been used in the previous literature to quantify air conditioning potential.51, 52 We avoid such an interpretation, however, because this measure ignores a large number of economic, demographic, and technological factors. For example, this measure scales linearly with population, and thus ignores cross-country differences in household size. This imperfect metric nonetheless provides a valuable first step, and jumping off point for more comprehensive analyses.

2.2 Methods

Imputing Temperatures for Global Grid We impute daily average temperatures for our 5km by 5km grid of the world using inverse distance weighting. In particular, we weigh temperature observations from nearby temperature monitoring stations using 1/d2 where d is the distance between the monitoring station and the cell centroid. With inverse distance weighting, the weight given to data points decreases with distance.2 Inverse distance weighting with weights equal to the inverse of squared distance has been widely used in the academic literature.53 In addition to selecting this weighting function, we must decide how many nearby monitoring stations to use. The closest stations have the most information content and as one moves to more distant stations they have less information content and more noise. To de-

2For example, consider the case in which there are two stations, one 2 kilometers away and another 4 kilometers away. In this case, 80% of the weight is put on the nearest station, with only 20% on the farther away station. 17 termine the optimal number of nearby monitoring stations to include, we run a simulation on the 14,538 monitoring stations available in our sample. We use the following steps.

1. We randomly select with replacement 100,000 days from the 3,641 days available over the period 2009 to 2018. 2. For each day, we randomly select a monitoring station, hide its temperature report and extract from GPW the population living in the 5km by 5km cell the station is located in. 3. We interpolate the temperature for that station using a number of neighbors that varies from 1 to 30. For each of these 30 we record the diﬀerence between the prediction and the temperature that was actually recorded by that station. 4. We select the number of neighbors that minimizes the population-weighted root mean squared error (RMSE) over the 100,000 station-days of the simulation.

That is, this procedure allows us to determine how many nearby monitoring stations is optimal from the perspective of minimizing RMSE. The results of the simulation described above can be seen in the Panel (A) of Figure 2.1. The population-weighted RMSE decreases substantially when the ﬁrst few neighbors are added, only to stabilize slightly below 2 ◦C for n = 10 and beyond. RMSE is minimized for n = 17, so we use the nearest 17 monitoring stations for all calculations. Panel (B) of Figure 2.1 shows the number of stations recording on each day over the period 2009-2018. Imputing temperatures for 30,000,000+ cells using the 17 closest monitoring stations for each of 3,641 days is a daunting computational task. In order to signiﬁcantly speed up the nearest-neighbor search, we rely on a k-dimensional tree (k-d tree), a space-partitioning data structure for organizing points in a k-dimensional space3. In the case of a nearest-neighbor search, the algorithm seeks the point in the tree that is closest to the given input location. Doing so allows the search for randomly distributed points to be performed in O(log n) time by using the tree properties to eliminate large irrelevant portions of the search space. Since the available k-d tree packages available in R use Euclidean distances to build the binary tree, we convert the WGS84 coordinates of our stations and the cells of our empty grid of the world to ”Earth-Centered, Earth-Fixed” (ECEF) coordinates. ECEF is a geographic coordinate system that also stands as a Cartesian coordinate system. It represents positions (in meters) as X, Y, and Z coordinates. Using ECEF coordinates allows Euclidean distances in that plane to mimic great circle distances, i.e. accounting for the fact the Earth’s surface is curved. Conversion from WGS84 coordinates to ECEF coordinates is done using standard formulas. First, we use weather data 2009-2018 to construct cooling degree days (CDDs) using 14,500+ temperature monitoring stations around the world. Second, we merge this information with gridded global population estimates at the 5km by 5km level. Third, we combine 3I would like to thank my fellow Bear and friend Thibault Doutre for introducing me to the concept. 18

Figure 2.1: Simulation results

Notes: (A) Optimal number of neighboring stations to use for the inverse distance interpolation is n = 17. (B) Number of monitoring stations recording each day from 2009 to 2018. 10,580 stations report each day on average. this merged dataset with data on the geographic boundaries of the world’s countries, along with UN population estimates for cities around the world to calculate population-weighted measures at the region, country, and city-level.

Data Weather Data We collected weather information from the U.S. National Climatic Data Center (NCDC) “Global Summary of the Day” database. These data describe daily weather data recorded at 14,500+ monitoring stations around the world. Stations typically report daily mean, minimum, and maximum temperature, as well as relative humidity, precipitation, and wind speed. For our analysis we use daily mean temperature only. 19

We base all of our calculations on data from 2009-2018. Our selection of this time period reﬂects several objectives. We wanted to use recent data because global mean temperatures are increasing, so older data are less comparable. At the same time, we also wanted to include an entire decade of data to reduce the impact of idiosyncratic year-to-year variation.

Figure 2.2: Reporting Frequencies for Monitoring Stations

Notes: The map shows the percentage of days from January 1, 2009 to December 31, 2018 each of the 14,538 stations reported temperatures for.

The geographical breakdown of the monitoring stations is shown in Figure 2.2. Overall the dataset oﬀers a rich coverage of the world. While 14,538 stations record at least once over the period, the number of stations recording on any given day is systematically lower. Recording frequency is also shown in Figure 2.2. While the great majority of stations in Brazil report between 20% and 40% of the days in 2009-2018, the stations in India, China, and South-East Asia tend to report more than 80% of the time over the entire period. More generally, half the 14,538 stations report at least 92% of the time while the typical station reports 73% of the time. 20

We take several steps to minimize any impact that missing weather data might have on our estimates. Most importantly, our approach for measuring temperature at individual locations takes advantage of all stations for which there is non-missing data. In particular, when data are not available for a given station, the approach incorporates information from stations farther away. While stations farther away are not a perfect substitute, this approach allows us to impute temperatures for all days in the sample, avoiding introducing compositional changes in the sample. As we discuss in the paper, we measure total CDD exposure using cooling degree days (CDDs), a widely-used measure of cooling demand calculated as the sum of daily mean temperatures above 18.3◦C (65◦F). This particular threshold, 18.3◦C, has been widely used in previous studies,54–57 with at least one other study using a very similar 18.0◦C threshold.58 After calculating CDDs for each day, we sum across all days in the calendar year to calculate annual CDDs. The annual measure thus reﬂects both the number of days with hot weather and the intensity of heat on those days. We exclude ocean-based measures to avoid any possible biases from ships and buoys.59

Population Data We collected population data from the Gridded Population of the World (GPW) project at Columbia University’s Center for International Earth Science Information Network (CIESIN). We are using data from the fourth version (GPWv4) which provides global population estimates at approximately the 1km by 1km level for 2000, 2005, 2010, 2015, and 2020. See http://sedac.ciesin.columbia.edu/data/collection/gpw-v4 for details. Our calculations are based on the estimates for 2015. The GPW estimates are based on population censuses that were performed around the world between 2005 and 2014. The GPW estimates are created using the most detailed geographic units available in each census. For most countries, this means population measures at the city- or municipality- level. The GPW estimates assume a uniform distribution of population within geographic units when mapping population to the 1 km grid. Finally the GPW estimates incorporate population projections from the United Nations’ World Popu- lation Prospects, to project the global distribution of population in 2015. The advantage of the GPW estimates is that they provide high-quality measures of where people live within each country. This doesn’t matter much for small countries, but it matters a great deal for many of the countries with the largest total CDD exposure. CDDs vary massively within India, for example, from the relatively cool Northern mountainous areas to the blistering hot Southern areas. We aggregate these data to a 5km by 5km grid, which reduces the total number of cells from 780 million to a more manageable 30 million. This is still computationally burdensome but we prefer not to further aggregate the data because of within-cell climate variation. 21

Geographic Boundaries To calculate daily cooling degrees for each cell, we take the difference between daily mean temperature and 18.3◦C, or zero, whichever is greater. Thus, a day with average temperature of 15◦C has zero cooling degrees whereas a day with average temperature of 28.3◦C has 10 cooling degrees. This particular threshold, 18.3◦C (65◦F) has been widely used in previous studies including.58 After calculating CDDs for each day, we sum across all days in the calendar year to calculate annual CDDs, and average across the ten years in our data. These calculations yield annual average CDDs for a 5km by 5km grid of the global landmass. These estimates are plotted in Figure 2.3. We next calculate population-weighted average CDDs at the region- and country-level. This requires us to categorize cells by region and country. We collected the 1:10m “Cultural Vectors” shapefile (Figure B.6) from Natural Earth http://www.naturalearthdata.com/, a public domain map dataset supported by the North American Cartographic Information Society. This shapefile tells us the region and country corresponding to each 5km by 5km cell. We then calculate population-weighted CDDs by region and country by taking a weighted average of all cells in each region and country, using GPW populations as weights.4 Finally for cities we use imputed CDDs for the 5km by 5km cell corresponding to the city centroid as reported by.61 Then we multiply CDDs by the population for each “urban agglomeration” from61 for our measures of total CDD exposure by city.

U.S. Data on Air Conditioning Use We report in the paper that air conditioning in the United States uses 400 terawatt-hours of electricity annually, representing 17% of total residential electricity consumption and 12% of total commercial electricity consumption. The 400 terawatt-hours comes from the U.S. Department of Energy and reﬂects total residential and commercial electricity consumption for space cooling.5 The 17% and 12% shares are then calculated by dividing by total delivered electricity in each sector.6 It is also worth noting that in the United States, 87% and 80%

4As a check on our results, we also take a sum of the weights to compare the implied country-level populations to country populations in.60 Overall, the population estimates are extremely close except for a few countries with highly irregular coastlines (e.g. Philippines, Aruba, French Polynesia). To avoid any ambiguity we use population county from60 in Table 1 of the paper and throughout the paper for our country- level calculations of total CDD exposure, though results using populations measured using GPW data are essentially identical. 5,62 Table A4 “Residential Sector Key Indicators and Consumption” and Table A5 “Commercial Sector Key Indicators and Consumption,” reports that U.S. households and ﬁrms in 2015 used 0.80 and 0.55 quadrillion Btu of electricity for space cooling, respectively. The vast majority of this energy use is for air conditioning, though this also includes electric fans, evaporative coolers, and heat pumps. Each quadrillion Btu is equivalent to 293.1 terawatt-hours (293.1 million megawatt-hours). 6Table A4 “Residential Sector Key Indicators and Consumption” and Table A5 “Commercial Sector Key Indicators and Consumption,” report that total delivered electricity in 2015 in these two sectors was 4.78 and 4.64 quadrillion Btu, respectively. 22 of residential and non-residential buildings have air conditioning, respectively.7

2.3 Results

Main Findings

Table 2.1: Rankings by Total Cooling Degree Day Exposure

Country Population Population-Weighted Product of Global (in millions) Annual Cooling Population Share Degree Days and CDDs (CDDs) (in billions)

Panel A. Top Ten Countries

1. India 1,309 2,848 3,728 28% 2. China 1,397 1,009 1,410 10% 3. Indonesia 258 3,284 848 6% 4. Nigeria 181 3,429 621 5% 5. Pakistan 189 2,504 474 4% 6. Brazil 206 2,108 434 3% 7. Bangladesh 161 2,644 426 3% 8. Philippines 102 3,266 332 2% 9. United States 320 867 277 2% 10. Vietnam 94 2,777 260 2%

Panel B. Top Ten Cities

1. Mumbai, India 21.0 3,544 74.6 0.6% 2. Delhi, India 25.7 2,831 72.8 0.5% 3. Dhaka, Bangladesh 17.6 2,955 52.0 0.4% 4. Karachi, Pakistan 16.6 3,108 51.6 0.4% 5. Manila, Philippines 12.9 3,572 46.2 0.3% 6. Kolkata, India 14.9 3,047 45.3 0.3% 7. Lagos, Nigeria 13.1 3,227 42.3 0.3% 8. Tokyo, Japan 38.0 1,040 39.5 0.3% 9. Jakarta, Indonesia 10.3 3,772 38.9 0.3% 10. Bangkok, Thailand 9.3 3,995 37.0 0.3%

Notes: This table ranks the top ten countries and top ten cities worldwide by total cooling degree day exposure i.e. the product of population and annual CDDs. Country and city (“urban agglomeration”) populations are from United Nations data. Annual cooling degree days (CDDs) are calculated as the sum of daily average temperatures above 18.3◦C. We report annual average CDDs for the period 2009-2018. For countries, we calculate population- weighted CDDs using global population estimates at the 5km by 5km level from the Gridded Population of the World project. Temperature data are from 14,500+ land-based monitoring stations tracked by the U.S. National Climatic Data Center. See the supplementary information for a complete list of sources and additional details.

7DOE, Residential Energy Consumption Survey (RECS) 2015, Table HC7.1 “Air Conditioning in U.S. Homes,” Released February 2017. DOE, Commercial Buildings Energy Consumption Survey (CBECS) 2012, Table B30 “Cooling Energy Sources, Number of Buildings and Floorspace,” released May 2016. 23

Figure 2.3 plots our global CDD estimates. Vast areas of the world are orange and yellow, indicating 3,000+ and even 4,000+ CDDs annually. The map highlights Africa, the Middle East, India, and Southeast Asia as the areas with most extreme high temperatures. The highest CDDs on the planet are found in Northern Africa along a horizontal band passing through Mauritania, Mali, Niger, Chad, and Sudan. Large areas of the world are also purple and black, indicating less than 1,000 CDDs annually. This includes many of the highest-income areas of the world including Western Europe, the United States, Canada, South Korea, and Japan. In this paper we focus on CDDs, but for areas above 35.0◦ latitude heating degree days are at least as important as CDDs.47, 57, 63 Figure 2.3: Global Cooling Degree Days

Notes: Average annual cooling degree days for the period 2009-2018. We calculate cooling degree days as the sum of daily mean temperatures above 18.3◦C (65◦F). The underlying temperature data is drawn from 14,500+ land-based monitoring stations tracked by the U.S. National Climatic Data Center. The resolution of the ﬁgure is 5 km by 5km.

Table 2.1 ranks the top ten countries globally by total CDD exposure. India is at the top 24 of the list with 1.3 billion people and 2,848 annual CDDs. Strikingly, India represents 28% of total global CDD exposure, 14-times the CDD exposure of the United States, and more than twice as much as any other country. The top four countries (India, China, Indonesia, and Nigeria) have almost half of total global CDD exposure. The list is dominated by low- and middle-income countries with warm climates. Except for the United States, all of the countries in the top ten have an annual GDP per capita under $10,000. There are eight countries with more total CDD exposure than the United States, many with substantially smaller populations, such as the Philippines with only one- third the population but four-times the CDDs. Except for China, all of the countries in the top ten have at least twice as many CDDs as the United States. Table 2.1 also ranks the top ten cities worldwide. India again takes a prominent role with three cities in the top ten. Mumbai, by itself, has total CDD exposure equal to 25% of total CDD exposure for the United States. The list is dominated by cities in low- and middle-income countries. The only city in the top ten from a high-income country is Tokyo, due primarily to its very large population, and the top U.S. city (Miami) appears at number 39. See the supplementary information for a complete list of countries and cities.

Comparing Our Results to Previous Estimates Figure 2.4 compares our CDD estimates to the estimates from the World Resources Re- port.58 Virtually all countries are above the 45 degree line (in red), reﬂecting the global warming that has occurred during the period between which the two datasets were con- structed. According to NOAA, the ten hottest years in recorded history globally all have occurred since 1991, so incorporating recent temperature data is crucial.

Does Population Weighting Matter? Our CDD estimates are population weighted and thus reﬂect where people live in each country. An alternative would be to weigh all areas equally. Figure 2.5 compares these two measures. Deviations between the two measures are relatively small overall, but illustrate the importance of population weighting. For small countries the two measures are essentially identical and, indeed, countries like Hong Kong, Singapore, and Kuwait are right on the 45◦-line (in red). For larger countries, however, the population-weighting clearly matters. Brazil, Australia, and Algeria, for example, are well below the 45◦-line, showing that people tend to live in relatively cooler areas of these countries. China, Myanmar, and Djibouti, in contrast, are above the 45◦-line, meaning that people tend to live in relatively warmer parts of these countries. Overall, there are many more countries below the line than above the line, potentially reﬂecting a tendency of people to sort into relatively cooler areas. Migration patterns can change over time, however, and, for example, some historians have argued that air conditioning has facilitated migration toward the American South.64 25

Figure 2.4: Comparing Our Estimates to Previous Estimates

Notes: Baumert and Selman (2003) use a lower threshold to compute their CDDs (18◦C in lieu of 18.3◦C. For a similar average temperature during the day, their CDD estimate would be no lower than that obtained with the 18.3◦C. Should our estimates match those from Baumert and Selman (2003), the observations would fall exactly on the 45◦-line. 26

Figure 2.5: Does Population-Weighting Matter?

Notes: This graph compares our population-weighted CDD estimates to area-weighted CDD estimates. Should the two be equivalent, the observations would fall exactly on the 45◦-line. 27

Cross-validation Berkeley Earth oﬀers gridded daily temperature estimates at the 1◦ by 1◦ Latitude- Longitude level8. We use the estimates for the latest decade available to construct an alternative set of CDDs, which we then weight by population. Figure 2.6 shows that for the vast majority of countries the two measures are essentially identical, lying very close to the 45 degree line. And, in the few cases where the measures meaningfully diﬀer, they tend to be small island countries (e.g. Turks and Caicos, Saint Lucia, Bahamas), for which Berkeley Earth’s 1◦ by 1◦ resolution is particularly coarse.

Confidence Intervals A novel feature of our analysis is that we also report 95% confidence intervals for all estimates. Our analysis does not rely on survey information or other data derived from an explicit sampling of the population. Instead, the sense in which our estimates are a sample is that temperature is observed only for a sample of all possible locations. Even with 14,538 total temperature monitoring stations, there are large areas of the planet far from a monitoring station, as well as stations that fail to report temperature data on some days. We use a bootstrap simulation to assess how much this incomplete coverage matters. First, we use our data to construct 50 bootstrap samples. For each bootstrap sample, we draw with replacement a sample of 14,538 monitoring stations. These bootstrap samples are different from the original sample with different numbers of duplicates of various stations as well as omissions of stations.9 Second, we calculate all of our estimates for each bootstrap sample. For example, we calculate population-weighted annual cooling degrees in India for each bootstrap sample. At the end of this, we thus have a different estimate of population-weighted annual cooling degrees in India for each bootstrap sample, in addition to the estimate from the original dataset. Third, we calculate the standard deviation for these 50 estimates. This standard deviation is our standard error, and we use this to calculate 95% confidence intervals. These confidence intervals reflect the sensitivity of our results to including or excluding particular stations. For parts of the world with large numbers of monitoring stations, this is like having a large sample size and the estimates don’t tend to vary much across bootstrap samples. However, for parts of the world with sparse coverage, the estimates can vary significantly depending on which particular stations are included, and this gets reflected in wider confidence intervals.

8See See http://berkeleyearth.org/data/ for details. 9It is worth clarifying one subtle issue. Because we are sampling with replacement, we have duplicate monitoring stations in all bootstrap samples. In using inverse distance weighting to impute average temperatures for the 5km by 5km grid, we do not drop these duplicates. We include these duplicate measures as if they were coming from multiple stations all the same distance from a given cell. 28

Figure 2.6: Robustness check

Notes: Comparing our estimates to estimates derived from Berkeley Earth daily temperature data. Should the two be equivalent, the observations would fall exactly on the 45◦-line. 29

2.4 Discussion

Almost three billion people live in the tropics, 40% of the world’s population, mostly in low- and middle-income countries, and most currently without air conditioning. Our estimates of total CDD exposure point to the enormous potential growth in air conditioning in these countries. India, by itself, has an almost unfathomable amount of potential demand for cooling, both because it is so hot and because so many people live there. But our rankings also feature many middle-income countries such as China, Indonesia, Brazil, and Philippines, all poised to dramatically increase air conditioner use in the near term. Our paper contributes to a small existing literature on global demand for air conditioning. Previous studies have shown that electricity consumption increases on hot days,53, 65 and that there is a positive correlation between income and having an air conditioner.41, 66–68 Several previous studies have focused on India,69–71 with relatively few global analyses.51, 67 Much is left to be done. As we emphasized above, CDD exposure is a highly imperfect measure of air conditioning potential. For example, we explained that this measure scales linearly with population, and thus ignores cross-country differences in household size. There are also cross-country differences in building size, construction methods, materials, urban form, and many additional economic, demographic, and technological factors that are not accounted for with CDD exposure. Another important factor is access to electricity. Air conditioning adoption is being made possible, in part, due to increases in electrification. For example, in Bangladesh, electricity now reaches 80% of the population, up from 20% in 2000. In Indonesia, electricity now reaches nearly 95% of the population, up from 50% in 2000. There are still nearly 1 billion people worldwide without access to electricity but this is expected to decrease significantly over the next decade.72 What air conditioning adoption will mean for electricity consumption depends on technological change. If air conditioners can be made more energy-efficient, due to induced innovation or economies-of-scale,73, 74 this could reduce the energy consumption impacts con- siderably. Similarly, if growth in renewables can reduce the carbon intensity of electricity, this could mitigate the carbon dioxide impacts.49 Air conditioning adoption and usage also depend on prices. Putting a price on carbon dioxide emissions would increase electricity prices and thus slow air conditioning adoption and encourage energy efficiency. Carbon policy would also incentivize less energy-intensive forms of cooling. Evaporative cooling, for example, is a viable alternative in many parts of the world. Making homes better insulated, using natural shade, cool roofs, and passive cooling systems are also lower-energy approaches to cooling.75 30

Chapter 3

Evaluating the Eﬀectiveness of Very Large Marine Protected Areas at Deterring Fishing Eﬀort1

3.1 Introduction

Background In October 2010, the tenth meeting of the Conference of the Parties adopted in Nagoya (Japan) the Strategic Plan for Biodiversity. The plan included the Aichi Biodiversity Tar- gets, a set of goals named after the prefecture that hosted the meeting. Aichi Target 11 of the United Nations’ Convention on Biological Diversity (CBD) stipulates that 10% of the ocean should be protected by some form of Marine Protected Area (MPA) by 2020.76 An MPA refers to a marine area where certain activities are limited, or prohibited, in an attempt to protect part, or all, of the natural resources it contains. They are classiﬁed in various categories by the International Union for Conservation of Nature (IUCN).77 Marine reserves (IUCN category Ia), also known as “no-take” areas, refer to MPAs in which ﬁshing, hunt- ing, or collecting are entirely prohibited in order to protect sensitive habitats or threatened species.

There is scientific evidence that MPAs can be effective at increasing biomass, species rich- ness, and population size within the boundaries of the reserves.78, 79 In particular, marine reserves appear to be the most effective type of MPA when it comes to achieving conservation targets.80–83

However, this body of knowledge has focused on smaller MPAs. This is not surprising given the fact that 87% of the 13,000+ existing MPAs are no larger than 100km2, while

1The material in this chapter was coauthored with David Zilberman. 31

95.7% are under 1,000 km2.84 Recent years however have seen the emergence of Very Large Marine Protected Areas (VLMPAs), MPAs spanning more than 100,000 km2. Establishing large MPAs has become the new trend in an attempt to meet the CBD conservation targets.85–88 As such, 31 of the 35 existing VLMPAs have been created within the last decade alone,84 as shown in Panel B of Figure B.8.

There exists scientiﬁc support for large MPAs. In theory, they may better at protecting diverse ecosystem than their smaller counterparts, since their size makes them more likely to encompass species’ range.89 Yet, VLMPAs are not yet well-understood, an observation shared by researchers90, 91 and policymakers alike. As such, Dr. Jane Lubchenco, the head of the US National Oceanic and Atmospheric Administration declared in 2011: “We don’t have the resources that we need to actually monitor, enforce and understand these areas”.92

A lack of enforcement could lead to a failure in meeting their initial objectives.79, 93, 94 In- deed, the effectiveness of MPAs in providing ecological benefits has been shown to be eroded by illegal fishing effort.95–98

In the meantime, recent years have marked the emergence of new data which have of- fered new means to assess fishing effort across the globe. In particular, the Global Fishing Watch (GFW), a non-profit organization has made available to the public a database which tracks vessels using their Automatic Identification System (AIS) data. This database allows to track individual fishing vessels at an unprecedented spatio-temporal resolution2 and manages to detect whether these vessels are currently fishing, by looking at their speed and trajectory99 (see Panel A of Figure B.8 to see the evolution of the number of fishing hours and vessels tracked by GFW). A growing number of studies have used this database to study fishing effort in a variety of contexts.94, 100–104

In this paper, we use data on fishing effort gathered by GFW to study whether the eight “no-take” VLMPAs implemented between 2012 and 2018 (Figure 3.1 and Table B.16) have been successful at deterring fishing effort. We also investigate the characteristics of the vessels associated with illegal fishing effort in these protected areas.

Our analysis shows that the introduction of VLMPA has significantly reduced overall fishing in these important ecological areas, although the impact varied by region. Since each VLMPA is supporting its own unique marine life, it is important to enforce policies that would reduce fishing throughout all these reserves.

We also ﬁnd that the major violators of the sanctity of the VLMPAs are some of the major powers and members of the United Nations’ Security Councils, countries which are

2The data are available at a 0.1◦ ×0.1◦ resolution, on a daily basis from January 2012 to December 2018. 32

Figure 3.1: List of all the Very Large Marine Protected Areas (VLMPAs)

Notes: Exclusive Economic Zones (EEZ) are shaded in blue. The deepest location inside each of the eight VLMPAs is shown in orange. 1 corresponds to the Ross Sea Protected Area, 2 to the Papah¯anaumoku¯akea Marine National Monument, 3 to the Paciﬁc Remote Islands Marine National Monument, 4 to the Terres Australes Fran¸caises, 5 to the Pitcairn Islands Marine Reserve, 6 to the Phoenix Islands Protected Area, 7 to the Nazca-Desventuradas Marine Park, and 8 to the Revillagigedo National Park. 33 supposed to provide global leadership and enforce arrangements that will sustain the marine environment for future generations.

Stylized facts Seven out of the eight areas under consideration in this paper are managed by six sovereign states, including one Small Island Developing State (SIDS) (see Table B.16. The world’s largest VLMPA, the Ross Sea Protected Area is managed by the Commission for the Conservation of Antarctic Marine Living Resources (CCAMLR), an international commission composed of 26 Members. Table B.16 shows that it takes on average 525 days (17.5 months) for the no-take area to come into eﬀect after the original announcement than an area would eventually become protected is made.

Figure 3.2: Evolution of daily hours of ﬁshing eﬀort in each VLMPA per 1,000 km2.

Notes: Blue dots measure the sum of fishing hours spent by vessels emitting prior to the announcement of the creation of an VLMPA. The date announcing a possible implementation implementation of a fishing ban is shown in green. Red dots show the sum of fishing hours spent by vessels emitting for the first time after the announcement. The solid orange vertical line corresponds to the date when the fishing ban becomes effective in each VLMPA. The length of the vessels tracked prior to the announcement averages 54.95 meters while that of those tracked after it averages 48.58 meters.

Figure 3.2 shows the daily sum of fishing hours per 1,000 km2 in each of the 8 VLMPAs 34 over the period 2012 to 2018. It distinguishes hours fished by vessels tracked by GFW prior to the announcement of the creation of each of the areas from hours fished by vessels first tracked after the announcement. One can see that fishing intensity varies widely between the 8 areas. Prior to the date when the the announcement of the creation of the Phoenix Islands Protected Area became public, fishing effort averaged 0.263 hours per 1,000 km2 as opposed to 0.013 hours per 1,000 km2 for the extended portion of the Papah¯anaumoku¯akea Marine National Monument. For comparison, the Global Fishing Watch has measured on average 83,991 daily fishing hours over the period 2012-2018 across all oceans. Given that oceans span approximately 360 million km2, this would yield 0.223 hours (14 minutes) of fishing effort per 1,000 km2, a conservative estimate of fishing effort intensity,105 given that many active fishing vessels are yet to be tracked by AIS. At a glance, it would appear that some areas experience overall declines in the intensity of fishing effort after the fishing ban comes into effort (Phoenix Islands Protected Area, Revillagigedo National Park and Ross Sea Pro- tected Area) while others do not (Nazca-Desventuradas Marine Park, Pacific Remote Islands Marine National Monument, Papah¯anaumoku¯akea Marine National Monument,Pitcairn Is- lands Marine Reserve, Terres Australes Fran¸caises). In spite of the apparent heterogeneity the intensity of fishing effort over time between these eight VLMPAs, it would appear that on average, fishing effort dropped after the enforcement of the ban relative to the period preceding the original announcement, especially when one restricts their attention to the vessels that were tracked prior to the announcement date (blue dots in Figure 3.2). This is highlighted further in Figure B.9 by averaging the daily sums fishing hours for all eight VLMPAs.

The type of fishing effort practiced in each of these eight areas also varies, as shown in Table B.14. The table nonetheless shows that the most prevalent fishing methods in the VLMPAs are set longlines and drifting longlines, regardless of the period under consideration (pre-announcement, post-announcement & pre-implementation and post-implementation). Trawling, a particularly destructive fishing method where a net is dragged across the ocean sea floor, is observed at some point between 2012 and 2018 in three out of the eight VLMPAs, and even accounts for 1% of all the illegal fishing effort monitored in the Terres Australes Fran¸caisesVLMPA.

3.2 Methods

In order to analyze the deterrence effect of VLMPAs, we investigate the change between annualized fishing effort inside the area prior to the announcement and after the implementation of the fishing ban. For this endeavor, we rely on the Incursion Ratio (IR), which we define as the ratio of the distance between the location of fishing effort and the closest point on the VLMPA border over the distance to the border for the deepest location inside the VLMPA, i.e. distance of fishing effort to border IR = distance of deepest point to border 35

For example, the most isolated point within the Revillagigedo National Park is located 130.7 km away from the nearest boundary point. On January 31, 2018, a vessel was identified to be illegally fishing 23.8 km away from the nearest point on the border. It would consequently correspond to IR = 0.18. We represent the deepest point inside each VLMPA in Figure 3.1. The evolution of the distribution of fishing effort inside each of the eight VLMPAs is shown in Figure 3.3, along with the distribution shift for all areas combined (last panel). We display the counts of fishing hours per incursion ratio to convey the change in magnitude along with the level of displacement. The horizontal axis in Figure 3.3 displays the Incursion Ratio. The two vertical green lines show the limits of the range of values IR could take. Whenever IR = 0, this would imply that a vessel if fishing right on the border of the VLMPA. When- ever IR = 1, the vessel would be fishing at the deepest possible location inside of the VLMPA.

At first glance, it would appear that there is also heterogeneity between the eight VLM- PAs when it comes to the displacement of fishing effort during the period following the implementation of the fishing ban relative to the period preceding the announcement that the area would eventually become a marine reserve. The entire distribution shifted towards the border for Phoenix Islands Protected Area and for Revillagigedo National Park. The spread of the distributions seems to have remained virtually unchanged for the remaining areas. When it comes to the overall magnitude, it would appear that the overall count of fishing hours dropped in seven out of eight areas. Pacific Remote Islands Marine National Monument appears to be the only VLMPA witnessing a slight increase in the count of post- implementation annualized fishing hours relative to the pre-announcement period. As was the case for Figure 3.2, Figure 3.3 would indicate that, at the aggregate level, fishing effort decreased after the ban, and would also indicate a general displacement of fishing effort towards the border of the VLMPA.

In Figure 3.3 and for the rest of the analysis, we exclude the lapse of time between the announcement and the creation of the areas as recent research104 has shown that the announcement can lead to a rise in fishing effort until the enforcement of the ban. Provided that fishing effort during that period does not cause stocks to fall below the minimum viable stock threshold, a permanent ban that successfully deters fishing effort will eventually allow the resource to recover inside the marine reserve. By excluding this period, we therefore see how fishing effort has decreased relative to the time when vessels fished in an area not scheduled to be protected.

When attempting to measure the deterrence effect of the creation of these “no-take” VLMPAs, we focus solely on the evolution of fishing effort inside the areas for vessels tracked by GFW prior to the announcement of the creation of these areas. Indeed, as shown in panel A of Figure B.8, the cumulative number of vessels tracked by the database has skyrocketed over the years 2012-2018 as countries started mandating their fleet to acquire AIS tracking 36

Figure 3.3: Displacement of ﬁshing eﬀort

Notes: Comparison of the annualized count of fishing effort inside the VLMPAs prior to the announcement of their creation and after their implementation. Counts of fishing hours are restricted to the vessels tracked by the GFW database prior to the announcement date that a restriction on fishing effort may eventually take place. A version of the density plot that includes all the vessels is shown in Figure B.10. devices3. Incorrectly assuming that a vessel only started fishing upon being tracked by AIS could lead to biases in our analysis. For instance, we cannot rule out that vessels appearing after the implementation of an MPA were not fishing prior to being monitored. Should such vessels fish in the area yet to be announced respect the ban and get tracked by the GFW after the announcement, our results would underestimate the extent of the deterrence effect.

Data Fishing Effort We have access to daily fishing effort information provided by the Global Fishing Watch (GFW), a non-profit organization that offers near real-time tracking of global commercial 3This can also be inferred from Figure B.11. Indeed, restricting the number of vessels practicing banned fishing to those tracked prior to the announcement of an eventual ban has a heterogeneous impact across countries. China, Taiwan, the United States, Japan and South Korea are the most affected by the restriction while most of the other countries have their fleet virtually unaffected by the restriction 37

fishing activity. The database gathers Automatic Identification System (AIS) data gathered from ships at sea. It allows to track individual vessels by using their individual Maritime Mobile Service Identity (MMSI) and is available at a daily, 0.1◦×0.1◦ resolution from January 2012 to December 2018. Fishing effort detection is based on a machine learning algorithm that tracks the behaviors associated with specific fishing activities (long-lining, purse seining, trawling, etc.), a methodology described in.99

MPA shapeﬁles MPA shapeﬁles were obtained from.106 The land polygons drawn in Figure 3.1 were downloaded from Natural Earth. EEZ boundaries were obtained from.107

Creation Timelines Sea Protected Area Announced by the Commission for the Conservation of Marine Living Resources (CCAMLR) on October 27, 2016, the Ross Sea protected area came into force on December 1, 2017 and has become the largest MPA in the world. The VLMPA’s no-take zone covers three separate areas that cover 1.12 million km2, or 72% of the 1.55 million km2.

Papah¯anaumoku¯akea Marine National monument The Papah¯anaumoku¯akea Ma- rine National monument was originally established on June 15, 2006 by the Bush administration. Eﬀorts to expand the monument started on January 29, 2015 when a group of prominent Native Hawaiians wrote to president Barack Obama asking the federal government to expand its protections around the Northwestern Hawaiian Islands. The monument was expanded from its original size of 362,073 km2 to 1,508,870 km2 on August 26, 2016 by presidential proclamation. It is now the fourth largest MPA in the world. It is a fully “no-take” area.

Pacific Remote Islands Marine National Monument The Pacific Remote Islands Marine National Monument was originally established by presidential proclamation on Jan- uary 6, 2009. On May 20, 2014 of a scientific report promoting the expansion of the existing monument was published. On September 25, 2014 president Barack Obama expanded the Pacific Remote Islands Marine National Monument from 215,000 km2 to 1,270,000 km2, making it the sixth largest MPA to date.

Terres Australes Fran¸caises On October 3, 2006, an inter-ministerial decree established the original natural reserve in the territorial waters of the French Southern Territories. On June 14 and June 15, 2016, the scientiﬁc council in charge of monitoring the reserve discussed the possibility of extending the coverage of the reserve.108 On December 12, 2016, the French 38 government published a decree extending the existing MPA by 672,969 km2, setting up “no- take” zones, covering more than 120,000 km2. Enforcement started on Mars 31, 2017, upon publication of the prefectoral order.

Pitcairn Islands Marine Reserve On March 18, 2015, the BBC reported the British government’s intention to create a marine reserve covering 834,000 km2 around the Pitcairn Islands, once monitoring and enforcement of the reserve could be funded.109 The ordinance implementing the area was then promulgated on September 12, 2016.

Phoenix Islands Protected Area The area was established in 2006. For years however, the government of Kiribati claimed that the Phoenix Islands Protected Area was a “no-take area” despite evidence that it was the most ﬁshed MPA in the world. Eventually, the cabinet of president Anote Tong voted on January 29, 2014 to actually close the reserve by January 1, 2015. News of the vote only broke on May 9, 2014,110 since no public announcement was made at the time of the vote.

Nazca-Desventuradas Marine Park In February 2013, the National Geographic and Oceana launched the “Pristine Seas” expedition in the Desventuradas islands, oﬀ the coast of Chile. The expedition led to the publication on February 23, 2013 of a report calling for the creation of an MPA.111 On August 24, 2016, the Chilean government oﬃcially ordered the creation of the Nazca-Desventuradas Marine Park, a 300,000 km2 “no-take” reserve.

Revillagigedo National Park On July 17, 2016, the United Nations Educational, Scien- tific and Cultural Organization (UNESCO) recognized the Revillagigedo islands as a World Heritage site,112 a first step toward the sanctuarization of the area. On November 24, 2017, Mexican president Enrique PeñaNieto officially created of the Revillagigedo National Park, a “no-take” protected area spanning 150,000 km2.

3.3 Results

Investigating Displacement of Fishing Eﬀort We estimate the following fourth-order polynomial equation:

4 X k (3.1) yi = βk · IRi + i, k=0

where yi corresponds to the difference in the number of annualized fishing hours after the implementation of the fishing ban and the number of annualized fishing hours preceding the 39

announcement that a marine reserve would eventually be created. IRi corresponds to the incursion ratio, rounded to the second decimal, and i corresponds to heteroskedasticity-robust disturbances. The ﬁtted regression curves and their associated 95% conﬁdence intervals are shown in Figure 3.4.

Figure 3.4: Quartic polynomial curves of best ﬁt and their associated 95% conﬁdence intervals

Notes: The scatter points represent the difference between the annualized fishing hours at a given incursion ratio after the implementation of the ban and prior to the announcement of the creation of the VLMPA. Negative values signify that there were fewer fishing hours at a given depth inside the VLMPA after the ban was put in place than during the period preceding the announcement of a potential marine reserve.

After estimating Equation 3.1, we calculate the area between the curve and the horizontal axis. A positive (negative) number would imply that fishing effort has increased (decreased) overall after the implementation of the ban, relative to the period preceding the announcement of the possible creation of a marine reserve. The area calculations are displayed in Table 3.14. When pooling all the eight VLMPAs together, we would conclude that “no-take” VLM- PAs have overall been successful at deterring fishing effort since the area between the fitted fourth-order polynomial and the x-axis would be negative (-86.9) and significant at a 5%

4In Table B.15, we display the area calculations corresponding to third-order polynomial and fith-order polynomial curves, along with their corresponding heteroskedasticity-robust 95% confidence intervals curves. 40 significance level with an associated 95% confidence interval of [−120.1; −53.7]. Should we only value the aggregate effectiveness of VLMPAs to deter fishing effort, then this be good news. Unfortunately, a closer look at each of the VLMPAs projects a more nuanced picture. As shown in Table 3.1, seven out of the eight VLMPAs are associated with an overall reduction in fishing effort after the ban. The Pacific Remote Islands Marine National Monument, an VLMPA managed by the United States, is associated with an overall increase in annualized fishing effort during the period following the creation following the implementation of the ban, relative to the period preceding the announcement that a ban would eventually be implemented. Four of the VLMPAs – the Ross Sea Protected Area, Papah¯anaumoku¯akea Marine National Monument, Revillagigedo National Park and Phoenix Islands Protected Area, – are associated with reductions in fishing effort that are significant at the 5% level. The largest reduction in fishing effort occurs in Phoenix Islands Protected Area, an VLMPA managed by the Republic of Kiribati, a country listed among the Small Island Developing States (SIDS).

Figure 3.3 shows that with the exception of the Papah¯anaumoku¯akea Marine National monument, an VLMPA managed by the United States, the VLMPAs where fishing took place deep within the protected area whose possible creation was yet to be announced witnessed a significant shift of the distribution of fishing effort from the vessels tracked prior to the announcement towards the border (Phoenix Islands Protected Area, Revillagigedo National Park). We fail to observe similar shifts in the spatial distribution of fishing effort for VLMPAs where the bulk of pre-announcement fishing effort took place at in incursion ratio smaller than halfway through the deepest point inside of the VLMPA.

Table 3.1: Area under the 4th-order polynomial curve estimating the diﬀerence in annualized ﬁshing hours

VLMPA name Lower bound Estimate Upper bound Ross Sea Protected Area -10.6 -6.1 -1.6 Papah¯anaumoku¯akea Marine National Monument -4.1 -2.1 -0.1 Pacific Remote Islands Marine National Monument -3.4 9 21.3 Pitcairn Islands Marine Reserve -5.4 -2.5 0.4 Terres Australes Fran¸caises -14.3 -6.2 1.8 Phoenix Islands Protected Area -407.9 -332.7 -257.4 Nazca-Desventuradas Marine Park -3.1 -1.5 0.2 Revillagigedo National Park -12.6 -7.1 -1.6 Pooled VLMPAs -120.1 -86.9 -53.7 Notes: A positive (negative) number means that annualized banned fishing effort exceeds (subceeds) annualized fishing effort during the period preceding the announcement of a possible VLMPA creation. 41

Analyzing the Profile of Infringing Vessels The small number of VLMPAs under consideration restricts our ability to systemically identify the area-centered characteristics responsible for the success or failure to deter fishing effort. However, we have at our disposal a plethora of information concerning the vessels responsible for fishing effort inside these areas, allowing us to investigate illegal fishing effort from a vessel-centered perspective. Indeed, thanks to the MMSI uniquely identifying each fishing vessel, GFW is able to track each vessel’s length, tonnage, engine size, class, and its flag. Table 3.2 ranks the countries of origin of all the vessels associated with banned fishing by the number of infractions committed by vessels carrying their flag.

Table 3.2: Ranking of the countries associated with prohibited ﬁshing inside of the established VLMPAs

All VLMPAs Rank Country of Origin Infractions Vessels Mean Length 1. South Korea 1,527 112 58 2. Taiwan 1,264 122 38 3. China 872 114 50 4. United States 613 84 42 5. France 559 8 58 6. Japan 376 52 50 7. Kiribati 254 12 68 8. Vanuatu 104 23 51 9. Spain 89 5 64 10. Papua New Guinea 70 17 75 11. Micronesia (Federated States) 38 12 73 12. Marshall Islands 33 7 71 13. Honduras 31 1 54 14. Mexico 31 15 72 15. Solomon Islands 26 3 68 16. Slovakia 24 2 29 17. New Zealand 23 4 50 18. Maldives 13 2 55 19. Russia 6 1 56 20. Australia 5 1 68 21. Ukraine 4 1 55 22. Cook Islands 3 1 57 23. El Salvador 3 2 83 24. Kosovo (UNK) 1 1 72 Unidentified 21 6 44 Total 5,990 608 58.4 Notes: We define infractions at the VLMPA-daily-vessel level. If a vessel were to fish illegally two days in a row in a given VLMPA, each daily occurrence would be counted as a distinct infraction. If a vessel were to fish in two VLMPAs during the same day, these would also get counted as two separate infractions.

Table 3.2 shows that the five countries most responsible for banned fishing are associated 42 with 80.7% of all the infractions observed in the eight VLMPAs, while their trespassing fleet represents 72.3% of all infringing vessels. Three of these countries are permanent members of the United Nations Security Council.

It is also worth noting that the average size of the trespassing vessels stands at 58.4 meters long, which corresponds to about 2.3 times the size of the average vessel (25.7 meters long) tracked by GFW over the 2012-2018 period. This might be explained by the remoteness of the VLMPAs under analysis, which would require larger vessels to resist the journey and allow for a catch large enough to make the ﬁshing excursion proﬁtable.

3.4 Discussion

We leverage a database tracking fishing effort at an unprecedented resolution to evaluate the effectiveness of “no-take” very large marine protected areas (areas spanning more than 100,000 km2) to deter fishing effort, measured in hours spent fishing inside the protected areas. We find that VLMPAs have overall been successful at deterring overall fishing effort. This would by itself be quite encouraging. However, a more granular approach reveals some heterogeneity between the eight VLMPAs under consideration, something which might be problematic if each area was designed to preserve distinct ecosystems. We find find that that the area associated with the largest reduction in fishing effort (Phoenix Islands Protected Area) is managed by the Republic of Kiribati, a SIDS, while the least successful one (Pacific Remote Islands Marine National Monument) is managed by the United States.

The wealth of information on the vessels tracked by GFW allows us to identify the characteristics of the infringing vessels. We consequently analyze the profile of fishing vessels practicing banned fishing in the VLMPAs and find that they can be traced back to a mere 24 countries, four of them being permanent members of the United Nations Security Coun- cil (China, United States, France, and Russia). Together, their fishing fleets account for a little over a third of all the daily fishing infractions we observe in these VLMPAs. This is particularly worth mentioning given that the next United Nations’ CBD conference is to be held in Beijing in 2020. These four countries will likely be key stakeholders during a conference which purpose will consist in assessing the progress made in meeting the conservation targets set in 2010. Focusing strictly on meeting the quantitative objectives set by the Aichi Target 11 at the expense of the actual effectiveness of the established protected areas may mislead the general public into believing that progress in ecological conservation has been achieved across the board, an outcome which would be all-the-more disappointing given that the increasing wealth of data concerning fishing effort, along with an increased cooperation between national stakeholders, could help these marine protected areas meet their intended objective to reduce fishing effort. 43

References

[1] The State of World Fisheries and Aquaculture 2020 (FAO, 2020). URL https://doi. org/10.4060/ca9229en. [2] Hardin, G. The tragedy of the commons. Journal of Natural Resources Policy Research 1, 243–253 (2009).

[3] Gollier, C. Pricing the planet’s future: the economics of discounting in an uncertain world (Princeton University Press, 2013).

[4] Millner, A. & Heal, G. Discounting by committee. Journal of Public Economics 167, 91–104 (2018).

[5] McClure, S. M., Laibson, D. I., Loewenstein, G. & Cohen, J. D. Separate neural systems value immediate and delayed monetary rewards. Science 306, 503–507 (2004).

[6] Bueti, D., Bahrami, B. & Walsh, V. Sensory and association cortex in time perception. Journal of Cognitive Neuroscience 20, 1054–1062 (2008).

[7] Olivers, C. N. & Meeter, M. A boost and bounce theory of temporal attention. Psy- chological review 115, 836 (2008).

[8] Lewis, P. A. & Miall, R. C. Brain activation patterns during measurement of sub-and supra-second intervals. Neuropsychologia 41, 1583–1592 (2003).

[9] Bruya, B. & Tang, Y.-Y. Is attention really effort? revisiting daniel kahneman’s influential 1973 book attention and effort. Frontiers in psychology 9 (2018).

[10] Kahneman, D. Attention and eﬀort, vol. 1063 (Citeseer, 1973).

[11] Google. Google trends. Data retrieved from the Google Trends website (accessed October 2019), https://www.google.com/trends. [12] ArdorSeo. How many google searches per day on average in 2020? Data retrieved from the Ardor Seo website (accessed March 12, 2020), https://ardorseo.com/blog/ how-many-google-searches-per-day. 44

[13] Gunn III, J. F. & Lester, D. Using google searches on the internet to monitor suicidal behavior. Journal of aﬀective disorders 148, 411–412 (2013).

[14] Wu, L. & Brynjolfsson, E. The future of prediction: How google searches foreshadow housing prices and sales. In Economic analysis of the digital economy, 89–118 (Uni- versity of Chicago Press, 2015).

[15] Drake, M. S., Roulstone, D. T. & Thornock, J. R. Investor information demand: Evidence from google searches around earnings announcements. Journal of Accounting research 50, 1001–1040 (2012).

[16] D’Amuri, F. & Marcucci, J. The predictive power of google searches in forecasting us unemployment. International Journal of Forecasting 33, 801–816 (2017).

[17] Stephens-Davidowitz, S. The cost of racial animus on a black candidate: Evidence using google search data. Journal of Public Economics 118, 26–40 (2014).

[18] Algan, Y., Murtin, F., Beasley, E., Higa, K. & Senik, C. Well-being through the lens of the internet. PloS one 14 (2019).

[19] Ettredge, M., Gerdes, J. & Karuga, G. Using web-based search data to predict macroe- conomic statistics. Communications of the ACM 48, 87–92 (2005).

[20] Stephens-Davidowitz, S. & Pabon, A. Everybody lies: Big data, new data, and what the internet can tell us about who we really are (HarperCollins New York, 2017).

[21] O’Haver, T. Intro to signal processing-deconvolution. University of Maryland at Col- lege Park, Retrieved (2008).

[22] Tukey, J. W. The extrapolation, interpolation and smoothing of stationary time series with engineering applications. (1952).

[23] Laibson, D. Golden eggs and hyperbolic discounting. The Quarterly Journal of Eco- nomics 112, 443–478 (1997).

[24] Rubinstein, A. “economics and psychology”? the case of hyperbolic discounting. International Economic Review 44, 1207–1216 (2003).

[25] Karp, L. Global warming and hyperbolic discounting. Journal of public economics 89, 261–282 (2005).

[26] Anthoﬀ, D., Tol, R. S. & Yohe, G. Discounting for climate change. Economics: The Open-Access, Open-Assessment E-Journal 3 (2009).

[27] Findley, T. S. Hyperbolic memory discounting and the political business cycle. Euro- pean Journal of Political Economy 40, 345–359 (2015). 45

[28] Krusell, P. & Smith, A. A., Jr. Consumption–savings decisions with quasi–geometric discounting. Econometrica 71, 365–375 (2003). [29] Becker, G. S. & Mulligan, C. B. The endogenous determination of time preference. The Quarterly Journal of Economics 112, 729–758 (1997). [30] Nordhaus, W. D. The political business cycle. The review of economic studies 42, 169–190 (1975). [31] Zimbardo, P. & Boyd, J. Putting time in perspective: A valid, reliable individual- differences metric. Journal of Personality and Social Psychology 77, 1271–1288 (1999). [32] Boyd, J. N. & Zimbardo, P. G. Time perspective, health, and risk taking. In Under- standing behavior in the context of time, 97–119 (Psychology Press, 2006). [33] Wohlford, P. Extension of personal time, affective states, and expectation of personal death. Journal of Personality and Social Psychology 3, 559 (1966). [34] Internet World Stats. Internet users statistics for africa (2019). Data retrieved from the Internet World Stats Indicators (accessed December 26, 2019), https: //www.internetworldstats.com/stats.htm. [35] Weber, I. & Castillo, C. The demographics of web search. In Proceedings of the 33rd international ACM SIGIR conference on Research and development in information retrieval, 523–530 (2010). [36] Weber, I. & Jaimes, A. Who uses web search for what: and how. In Proceedings of the fourth ACM international conference on Web search and data mining, 15–24 (2011). [37] King, M. W. How brain biases prevent climate action (2019). (accessed July, 2020), https://www.bbc.com/future/article/ 20190304-human-evolution-means-we-can-tackle-climate-change. [38] Gertler, P. J., Shelef, O., Wolfram, C. D. & Fuchs, A. The demand for energy-using assets among the world’s rising middle classes. American Economic Review 106, 1366– 1401 (2016). [39] Wolfram, C., Shelef, O. & Gertler, P. How will energy demand develop in the developing world? Journal of Economic Perspectives 26, 119–137 (2012). [40] Auffhammer, M. & Wolfram, C. D. Powering up china: Income distributions and residential electricity consumption. American Economic Review Papers and Proceedings 104, 575–580 (2014). [41] Davis, L. W. & Gertler, P. J. Contribution of air conditioning adoption to future energy use under global warming. Proceedings of the National Academy of Sciences 112, 5962–5967 (2015). 46

[42] Park, J. R., Goodman, J., Hurwitz, M. & Smith, J. Heat and learning. American Economic Journal: Economic Policy (forthcoming).

[43] Burgess, R., Deschenes, O., Donaldson, D. & Greenstone, M. Weather, climate change and death in india. Working Paper (2017).

[44] Carleton, T. et al. Valuing the global mortality consequences of climate change accounting for adaptation costs and beneﬁts. University of Chicago Working Paper (2018).

[45] Barreca, A., Clay, K., Deschenes, O., Greenstone, M. & Shapiro, J. S. Adapting to climate change: The remarkable decline in the u.s. temperature-mortality relationship over the twentieth century. Journal of Political Economy 124, 105–159 (2016).

[46] Waite, M. et al. Global trends in urban electricity demands for cooling and heating. Energy 127, 786–802 (2017).

[47] Wenz, L., Levermann, A. & Auﬀhammer, M. North–south polarization of european electricity consumption under future warming. Proceedings of the National Academy of Sciences 114, E7910–E7918 (2017).

[48] Auﬀhammer, M., Baylis, P. & Hausman, C. H. Climate change is projected to have severe impacts on the frequency and intensity of peak electricity demand across the united states. Proceedings of the National Academy of Sciences 114, 1886–1891 (2017).

[49] International Energy Agency. The future of cooling: Opportunities for energy-eﬃcient air conditioning. OECD/IEA (2018).

[50] Shah, N. et al. Opportunities for simultaneous eﬃciency improvement and refrigerant transition in air conditioning. Lawrence Berkeley National Laboratory Working Paper (2017).

[51] Sivak, M. Potential energy demand for cooling in the 50 largest metropolitan areas of the world: Implications for developing countries. Energy Policy 37, 1382–1384 (2009).

[52] Sivak, M. Will air conditioning put a chill on the global energy supply? American Scientist 101, 330–333 (2013).

[53] Deschˆenes,O. & Greenstone, M. Climate change, mortality, and adaptation: Evidence from annual ﬂuctuations in weather in the u.s. American Economic Journal: Applied Economics 3, 152–185 (2011).

[54] Sailor, D. J. & Mu˜noz,J. R. Sensitivity of electricity and natural gas consumption to climate in the usa—methodology and results for eight states. Energy 22, 987–998 (1997). 47

[55] Shorr, N., Najjar, R. G., Amato, A. & Graham, S. Household heating and cooling energy use in the northeast usa: Comparing the eﬀects of climate change with those of purposive behaviors. Climate Research 39, 19–30 (2009).

[56] Alberini, A. & Filippini, M. Response of residential electricity demand to price: The eﬀect of measurement error. Energy Economics 33, 889–895 (2011).

[57] Petri, Y. & Caldeira, K. Impacts of global warming on residential heating and cooling degree-days in the united states. Scientiﬁc Reports 5 (2015).

[58] Baumert, K. & Selman, M. Data note: Heating and cooling degree days. World Resources Institute (2003).

[59] Karl, T. R. et al. Possible artifacts of data biases in the recent global surface warming hiatus. Science 348, 1469–1472 (2015).

[60] United Nations. World population prospects: The 2017 revision, key ﬁndings and advance tables, ﬁle pop/1-1: Total population for 2015. ESA/P/WP/248 (2017).

[61] United Nations. World urbanization prospects: The 2014 revision, population of urban agglomerations with 300,000 inhabitants or more. ST/ESA/SER.A/352 (2014).

[62] Department of Energy, Energy Information Administration. Annual energy outlook (2016).

[63] Atalla, T., Gualdi, S. & Lanza, A. A global degree days database for energy-related applications. Energy 143, 1048–1055 (2018).

[64] Arsenault, R. The end of the long hot summer: The air conditioner and southern culture. Journal of Southern History 50, 597–628 (1984).

[65] Auﬀhammer, M. & Aroonruengsawat, A. Simulating the impacts of climate change, prices and population on california’s residential electricity consumption. Climatic Change 109, 191–210 (2011).

[66] Biddle, J. Explaining the spread of residential air conditioning, 1955–1980. Explo- rations in Economic History 45, 402–423 (2008).

[67] Isaac, M. & Van Vuuren, D. P. Modeling global residential sector energy demand for heating and air conditioning in the context of climate change. Energy Policy 37, 507–521 (2009).

[68] McNeil, M. A. & Letschert, V. E. Modeling diﬀusion of electrical appliances in the residential sector. Energy and Buildings 42, 783–790 (2010). 48

[69] Akpinar-Ferrand, E. & Singh, A. Modeling increased demand of energy for air conditioners and consequent co2 emissions to minimize health risks due to climate change in india. Environmental Science & Policy 13, 702–712 (2010).

[70] Gupta, E. Global warming and electricity demand in the rapidly growing city of delhi: A semi-parametric variable coeﬃcient approach. Energy Economics 34, 1407–1421 (2012).

[71] Abhyankar, N., Shah, N., Park, W. Y. & Phadke, A. Accelerating energy eﬃciency improvements in room air conditioners in india: Potential, costs-beneﬁts, and policies. Lawrence Berkeley National Laboratory Working Paper, LBNL-1005798 (2017).

[72] International Energy Agency. World energy outlook 2018. OECD/IEA (2018).

[73] Newell, R. G., Jaﬀe, A. B. & Stavins, R. N. The induced innovation hypothesis and energy-saving technological change. Quarterly Journal of Economics 114, 941–975 (1999).

[74] Acemoglu, D. & Linn, J. Market size in innovation: Theory and evidence from the pharmaceutical industry. Quarterly Journal of Economics 119, 1049–1090 (2004).

[75] Georgescu, M., Moreﬁeld, P. E., Bierwagen, B. G. & Weaver, C. P. Urban adaptation can roll back warming of emerging megapolitan regions. Proceedings of the National Academy of Sciences 111, 2909–2914 (2014).

[76] Convention on Biological Diversity (CBD). Strategic plan for biodiversity 2011-2020 and the aichi biodiversity targets. https://www.cbd.int/decision/cop/?id=12268 (2010).

[77] International Union for Conservation of Nature (IUCN). 17th session of the general as- sembly of IUCN AND 17th technical meeting. https://portals.iucn.org/library/ efiles/documents/GA-17th-011.pdf (1988). [78] Lester, S. E. et al. Biological eﬀects within no-take marine reserves: a global synthesis. Marine Ecology Progress Series 384, 33–46 (2009).

[79] Zupan, M. et al. Marine partially protected areas: drivers of ecological eﬀectiveness. Frontiers in Ecology and the Environment 16, 381–387 (2018).

[80] Costello, M. J. & Ballantine, B. Biodiversity conservation should focus on no-take marine reserves: 94% of marine protected areas allow ﬁshing. Trends in ecology & evolution 30, 507–509 (2015).

[81] Sala, E. & Giakoumi, S. No-take marine reserves are the most eﬀective protected areas in the ocean. ICES Journal of Marine Science 75, 1166–1168 (2017). 49

[82] Claudet, J. et al. Marine reserves: size and age do matter. Ecology letters 11, 481–489 (2008).

[83] Edgar, G. J. et al. Global conservation outcomes depend on marine protected areas with ﬁve key features. Nature 506, 216 (2014).

[84] MCI. Mpatlas (2019). URL www.mpatlas.org. [Online; accessed 2019-04-08]. [85] Boonzaier, L. & Pauly, D. Marine protection targets: an updated assessment of global progress. Oryx 50, 27–35 (2016).

[86] Dulvy, N. K. Super-sized mpas and the marginalization of species conservation. Aquatic Conservation: Marine and Freshwater Ecosystems 23, 357–362 (2013).

[87] Magris, R. A. & Pressey, R. L. Marine protected areas: Just for show? Science 360, 723–724 (2018).

[88] Toonen, R. J. et al. One size does not ﬁt all: the emerging frontier in large-scale marine conservation. Marine pollution bulletin 77, 7–10 (2013).

[89] O’leary, B. C. et al. Addressing criticisms of large-scale marine protected areas. Bio- science 68, 359–370 (2018).

[90] Gruby, R. L., Gray, N. J., Campbell, L. M. & Acton, L. Toward a social science research agenda for large marine protected areas. Conservation Letters 9, 153–163 (2016).

[91] Ban, N. C. et al. Social and ecological eﬀectiveness of large marine protected areas. Global Environmental Change 43, 82–91 (2017).

[92] Cressey, D. Ocean conservation: Uncertain sanctuary. https://www.nature.com/ news/ocean-conservation-uncertain-sanctuary-1.9568 (2011). [93] Jone, P., Qiu, W. & De Santo, E. Governing marine protected areas: getting the balance right (United Nations Environment Programme (UNEP), 2011).

[94] Dureuil, M., Boerder, K., Burnett, K. A., Froese, R. & Worm, B. Elevated trawling inside protected areas undermines conservation outcomes in a global ﬁshing hot spot. Science 362, 1403–1407 (2018).

[95] Arias, A., Cinner, J. E., Jones, R. E. & Pressey, R. L. Levels and drivers of ﬁshers’ compliance with marine protected areas. Ecology and Society 20, 1–14 (2015).

[96] Pollnac, R. et al. Marine reserves as linked social–ecological systems. Proceedings of the National Academy of Sciences 107, 18262–18265 (2010). 50

[97] Bergseth, B. J., Russ, G. R. & Cinner, J. E. Measuring and monitoring compliance in no-take marine reserves. Fish and Fisheries 16, 240–258 (2015).

[98] Bergseth, B. J., Gurney, G. G., Barnes, M. L., Arias, A. & Cinner, J. E. Addressing poaching in marine protected areas through voluntary surveillance and enforcement. Nature Sustainability 1, 421 (2018).

[99] Kroodsma, D. A. et al. Tracking the global footprint of ﬁsheries. Science 359, 904–908 (2018).

[100] Guiet, J., Galbraith, E., Kroodsma, D. & Worm, B. Seasonal variability in global industrial ﬁshing eﬀort. PloS one 14, e0216819 (2019).

[101] Queiroz, N. et al. Global spatial risk assessment of sharks under the footprint of ﬁsheries. Nature 572, 461–466 (2019).

[102] Englander, G. Property rights and the protection of global marine resources. Nature Sustainability 2, 981–987 (2019).

[103] Bradley, D. et al. Leveraging satellite technology to create true shark sanctuaries. Conservation Letters 12, e12610 (2019).

[104] McDermott, G. R., Meng, K. C., McDonald, G. G. & Costello, C. J. The blue paradox: Preemptive overﬁshing in marine reserves. Proceedings of the National Academy of Sciences 116, 5319–5325 (2019).

[105] Weimerskirch, H. et al. Ocean sentinel albatrosses locate illegal vessels and provide the ﬁrst estimate of the extent of nondeclared ﬁshing. Proceedings of the National Academy of Sciences 117, 3006–3014 (2020).

[106] UNEP-WCMC and IUCN. Marine protected planet (2019). URL www. protectedplanet.net. [Online; accessed 2019-04-08]. [107] Flanders Marine Institute. Maritime boundaries and exclusive economic zones (200nm), version 10. http://www.marineregions.org/. [108] TAAF. Bilan d’activit´es2016 de la r´eserve naturelle des Terres australes fran¸caises (2017).

[109] BBC News. Budget 2015: Pitcairn Islands get huge marine reserve. https://www.bbc. com/news/science-environment-31943633 (2015). [Online; accessed 2019-04-09].

[110] Science News. Mega-Marine Reserve Going Oﬀ-Limits to Fishers. https://www. sciencemag.org/news/2014/05/mega-marine-reserve-going-limits-fishers (2014). [Online; accessed 2019-04-02]. 51

[111] National Geographic Society & Oceana. Islas desventuradas – Biodiversidad marina y propuesta de conservaci´on.Report (2013).

[112] UNESCO. World heritage list. https://whc.unesco.org/en/list/1510 (2016). [On- line; accessed 2019-04-10].

[113] Li, T. Robust and consistent estimation of nonlinear errors-in-variables models. Jour- nal of Econometrics 110, 1–26 (2002).

[114] Li, T. & Hsiao, C. Robust estimation of generalized linear models with measurement errors. Journal of Econometrics 118, 51–65 (2004).

[115] Delaigle, A., Hall, P., Meister, A. et al. On deconvolution with repeated measurements. The Annals of Statistics 36, 665–685 (2008).

[116] Ramsey, F. P. A mathematical theory of saving. The economic journal 38, 543–559 (1928).

[117] Fisher, I. Theory of interest: as determined by impatience to spend income and oppor- tunity to invest it (Augustusm Kelly Publishers, Clifton, 1930).

[118] Von Mises, L. Human action (Lulu Press, Inc, 2016). 52

Appendix A

Full derivation of the Kernel of Attention to time

A.1 A Probabilistic Model for Attention to Time

We discretize time into “moments” that are ﬁnite and equal in length. In principle, time can be sub-divided into arbitrarily small moments. Although, in our data analysis, we consider discrete dates since that is the resolution of data available to us.

Let φx be a potential target of attention, and let Φ = {φ1, φ2, ···} be the countably infinite set of all possible targets. Let the subset Φtime ⊆ Φ be those targets associated with time specific moments in absolute time, such that for each target in φx ∈ Φ there exists an interval of time Θx associated with the target φx. Specifically, φx is associated with Θx if φx is associated with some (or all) moments contained in the interval Θx. Let the interval Θx be composed of a set of K consecutive moments in time Θx = {θx1 ...θxK } and we consider

φx associated with the interval Θx if it is associated with some or all moments θxK in the interval. Let the notation φx 7→ θxk mean that the target φx are associated with moment time θxk contained in the interval of time Θx (deﬁne φx 7→ ∅ if φx ∈/ Φ ).

For an example, a birthday party on July 15, 2017 during the time between 10:00 and 12:00 could be a target of attention (φx). In this case, we might consider the two hours of the party the moment in time directly associated with the target (θxk ). This party could then also be associated with the date “July 15, 2017”, which describes an interval of time containing that moment (Θx). In our analysis of internet query data, we sometimes study identiﬁable periods described by just the month, in which case Θx=“July”, or by just a year, in which case Θx=“2017”. This framework and notation is important because in search data, we often cannot identify the precise moment to which individuals are attentive when they execute a query — e.g. they do not search for “10:00-12:00 on July 15, 2017” — but we can identify information about a larger interval of time containing that moment (e.g. 53

“July”). Nonetheless, our aim is to understand attention to narrower moments using the search query data identifying only longer time intervals, which forces us to consider how attention is allocated to moments which are themselves nested within intervals.

At each instant in time t experienced by humans (what we call “query time”), a randomly selected individual j in a subject population focuses their attention Aj on a target according to some unobserved process. Let the notation Aj(t) = φx denote the event that at moment t, individual j directs their attention at target φx. Let Pr[Aj(t) = φx] denote the probability of this event occurring. We propose that the probability distribution of attention across targets is such that the total attention allocated to all targets associated with the moment θ depends on the temporal distance between that moment and the time t inhabited by the subject population. We call the temporal distance θ − t relative time (see Figure 1.2a-c). A positive relative time means an event is in the future of the subject population, a negative relative time means the event is in the population’s past, and a zero value means that the event is in the population’s present.

Formally, we hypothesize that attention is allocated across targets such that   X (A.1) C ·  Pr[Aj(t) = φx | φx 7→ θ] = κ(θ − t) time φx∈Φ holds (approximately) at each t. Here, κ(.) is a probability density function that integrates to one, which we call the kernel of attention to time (KAT). The scaling factor 1 (A.2) C = P time Pr[A (t) = φ ] φx∈Φ j x accounts for the fact that at each moment, some attention may be allocated towards the subset of targets that are not associated with any particular time (e.g. a poem). For parsimony, we assume this scaling is constant. The expression inside the brackets on the left-hand-side of Eq. A.1 is the a sum of probabilities that attention is directed at targets associated with the time θ. The summation sums over all possible time-associated targets. The elements of this sum may change as the time of the subject population (t) moves forward. Eq. A.1 allows for many possibilities in how attention across individual targets evolves, however it constrains the sum of probabilities associated with each moment θ to be proportional to the value of the KAT at the corresponding relative time θ − t.

A.2 Predictions for Internet Search Volume

Eq. A.1 generates predictions about the time-series structure of time-associated internet search volume. In each moment, if an individual’s attention is allocated to a target, then 54

there is some probability that it will lead to them executing an internet search query related to the target. Of those searches, a fraction will be associated with a specific moment in time. From this fraction, an even smaller fraction will contain a string that identifies a time interval containing the associated moment in time. Thus, we can compute the probability that an individual j allocates attention at time t toward some time-associated target φx and we observe an internet search query that identifies the time interval Θx associated with that target:

Pr[observe j search for φx at time t and identify Θx] = X time time (A.3) Pr[Aj(t) = φx|φx 7→ θ, φx ∈ Φ ] · δ(φx, θ) · Pr[φx ∈ Φ ] · ω .

θ∈Θx where

time δ(φx, θ) = Pr[φx 7→ θ | φx ∈ Φ ]

and

ω = Pr[j searches for φx at time t | Aj(t) = φx, ...] 0 · Pr[j s search query identiﬁes Θx | j searches for φx, ...])

Eq. A.3 decomposes this probability into five conditional probabilities for each moment in time θ, which are multiplied and summed over all moments in the identifiable interval of time Θx. The first factor is the probability that j allocates attention to the specific target φx at time t, if the target is associated with the specified moment θ, which we denote δ(φx, θ) for clarity below. The second factor is the probability that the target of attention is associated with the moment θ. The third factor is simply a “probability” that the target is associated with any moments in time – this value is either one if the target φx is in the set Φtime and zero otherwise, but it is written as a probability for consistency. The fourth factor is the probability that j executes an internet search for the target at time t, given that they were attentive to the target at t. The fifth factor is the probability that the executed search query contains information that identifies the time interval Θx associated with the target. The product of the last two factors is assumed to be a constant ω during the period of observation. In principle, gradual changes in internet access or slowly changing patterns of behavior might alter ω because it changes the likelihood that attention to a topic generates an internet query, but we find in practice that such trends do not affect our results (see Figure B.3).

We compute the predicted total volume of searches executed at time t that explicitly identify time period Θ, denoted SΘ(t), by summing across N individuals in population P 55

and summing over all possible targets of attention in Φ: X X SΘ(t) = Pr[observe j search for φx at time t and identify Θx]

j∈P φx∈Φ

! Nω X X = δ(φ , θ) κ(θ − t)(A.4) C x |{z} θ∈Θ φx∈Φ α | {z } δ¯(θ) where the second equality follows by substitution of Equations A.1, A.2 and A.3 and re- arranging terms. Eq. A.4 illustrates one way that the function κ(.) behaves like a kernel function, hence KAT, since it weights relative search volume for all possible targets based on Nω the distance in relative time. The scaling factor ( C ) captures the size of the population (N), the likelihood that individuals in that population allocate attention to any time-associated 1 targets ( C ), and the likelihood that allocating attention towards a time-associated target generates an internet search query identifying a period of time (ω). For parsimony, we rewrite this scaling factor as α. Eq. A.4 can be further simpliﬁed by observing that the summation across all targets (φx) of the joint probability distribution δ(φx, θ) over targets and moments in time is simply the marginal distribution over moments in time, denoted δ¯(θ). This marginal distribution describes the fraction of all possible time-associated targets that are associated with each speciﬁc moment in absolute time.

The final expression for predicted search volume at time t identifying time period Θ thus has the simplified form X ¯ (A.5) SΘ(t) = α δ(θ)κ(θ − t). θ∈Θ Eq. A.5 describes how total query volume that identifies a period of time (Θ) across all targets should evolve as a population moves through time (t), approaching and then passing the identified moments (θ) that occupy Θ. This query volume is the product of a scaling factor P ¯ (α) and a time series ( θ∈Θ δ(θ)κ(θ − t)). Behavioral factors, such as how easily attention to a target generates a related search query, as well as structural factors, such as the size of the population, are captured by α. In estimation, this scaling factor is not directly observed because the data are normalized by the data provider, but regardless, this normalization is immaterial because α is a nuisance parameter that can be accounted for by allowing for flexible multiplicative trends in the data. The time series is a super-position of reflected (with respect to t), shifted (by θ), and rescaled (by δ¯) KAT functions, one for each moment θ in the identified period of time Θ. One KAT function is reflected and shifted so that it is centered at each moment θ (Figure 1.2f). These reflected and shifted KAT functions are each then multiplied by the fraction of targets associated with their corresponding θ and summed at each value of t, producing a single time-series observable through Google (Figure 56

1.2i).

The time series of search volume SΘ(t) therefore appears as a super-position of impulse- response functions. The impulses are located at positions in time θ over the interval Θ, with a magnitude δ¯(θ). Each impulse-response has the shape of the reﬂected KAT. Thus, Eq. A.5 is mathematically equivalent to a re-scaled (discrete form) convolution of the KAT with a function fΘ(t): (A.6) SΘ(t) = α fΘ(t) ∗ κ(t) ¯ where fΘ(t) traces out the values of δ(θ = t) for all moments during the interval Θ and is zero everywhere else (see Figure B.1). Thus fΘ(t) describes the number of time-associated targets associated with the period Θ at each time t. If δ¯ is a constant for all moments in time, then fΘ(t) is a square wave with nonzero values during t = θ ∈ Θ. This formulation allows us to recover the form of the KAT impulse-response function using standard signal-processing techniques for deconvolution21, 22 applied to internet search data.

A.3 Google Search Data

We obtain data on queries made to the Google Search engine from the Google Trends API. These data describe changes over time in the number of queries originating from a geographic region during a discrete time period which contain a specific string. For example, we examine queries that contain the string “2016”, which aggregates search volume for queries such as “taxes 2016”, “halloween 2016”, and “2016 presidential election”. Time- series provided by Google have been rescaled by an unknown factor, but this does not affect our analysis because search volume is already predicted to be rescaled by several unknown constants (Eq. A.4) and these are all removed during the deconvolution analysis. We focus on country-level data, which are available for most of the countries in the world and use a sample that begins January 1, 2008 to ensure data quality. The temporal resolution of the available search volume data available depends on the length of the sample, with daily searches only available when the sample period does not exceed 270 days. This length of time is sufficient of analysis of monthly and daily targets (e.g. Θ = “July”), but is insufficient for the study of annual targets. Longer time series are provided using weekly search volumes, which we use in our analysis of annual targets, adjusting for the alternative time-step.

Targets corresponding to years We obtain search volume for queries containing the numerical representation of each year (e.g. “2016”) between 2009 and 2018, inclusive (e.g. Fig 1.3a). Because these numerical values are language-independent, we obtain data from 181 countries representing 96.7% of the global population. This results in a longitudinal sample where an observation corresponds to a unique {country, week-of-query, target-year} triplet. Individual country-weeks appear 57 more than once, because we examine searches for different target years originating on the same date. The final sample in our main analysis contains 246,131 observations. In our main analysis (Fig 1.4), we use data on queries occurring within the year that is the target period as well as 270 days before and after that year. We note that search volume for years exhibit irregular outliers where search volume can surge up to 27-times for just a week. For example, notable cases include queries during the Olympics Games or the Football World Cup. To limit the influence of these outliers, raw query data is rescaled to match medians in each country-by-target-year time series.

Targets corresponding to months We obtain search for queries containing the name of each month (e.g. “january”) for each month. Similar to the annual targets data, we construct samples that include 120 days before and after the endpoints of each target month. We focus this analysis on a more limited sample of 20 countries from ﬁve continents that are broadly representative of the global population—selected for breadth of representation across languages, cultures and religions— and where rates of Google search usage are high (Argentina, Australia, Brazil, Chile, Colom- bia, France, Germany, Hong-Kong, India, Indonesia, Mexico, Netherlands, New Zealand, Norway, Peru, Singapore, Spain, Sweden, the United Kingdom, and the United States) and examine queries where the month is described using the primary language for each country (English, French, German, Spanish, Portuguese, Bahasa Indonesia, Swedish, Norwegian, or Dutch). The resulting sample contains 712,620 observations, each corresponding to a unique {country, day-of-query, target-month } triplet.

Targets corresponding to speciﬁc days We identify searches associated with speciﬁc dates by collecting search volume for searches containing the name of single-day holidays. We use the same sample of 20 countries used for month-related targets and select three holidays for each each country (e.g. “Cinco de Mayo” for Mexico, “Waitangi Day” for New-Zealand, and “National Day” for Singapore, see Table B.1 for complete list). For each holiday, we examine search queries during a window of 270 days, centered on the holiday, resulting in a sample of 178,762 observations, each corresponding to a unique {country, day-of-query, target-holiday} triplet.

A.4 Empirical estimation of the KAT

For each class of target, we estimate the KAT using Google search queries in a two step procedure. First, we non-parametrically deconvolve query volume to recover the KAT. This ﬂexible approach recovers the KAT without requiring that we impose any assumptions about its form, however, because we have limited data, there is noise in these estimates which do 58

not result in a parsimonious analytical expression. To address these two issues, in the second step we ﬁt smoothing functions to the estimates of the KAT obtained in the ﬁrst step.

Deconvolution We obtain an estimate for the KAT by deconvolving Eq A.6, a standard procedure in signal processing21, 22 and econometrics113–115 that is the inverse of the convolution operation. The approach isolates an impulse-response function from time-series data when the timing and magnitude of impulses are known. We separately pool all search data for targets associated with years, months or days to recover an average estimate of the KAT corresponding to each class of target. For targets associated with years, we deconvolve these time series by estimating the distributed-lag regression via ordinary least-squares with a weekly time-step:

35 X (A.7) Syit = βLfyi,t−L + yit L=−35

where observations are indexed by the search target year y (e.g. y = 2016 for queries containing the string “2016”), country i where queries originate from, and week that queries are executed t. In the notation of the theoretical framework above, the interval of interest Θ is equal to the target year y. The variable fyit is equal to one if the date of the search t is contained in the year y, and zero otherwise. This dummy variable corresponds to the function f(t) in Eq. A.6 and appears as a square wave if it is plotted for a single i and y over time (as in Figure B.1). Coefficients βL are estimated for 35 weekly leads and 35 weekly lags of the variable fyit, in addition to its contemporaneous values where L = 0, and they flexibly recover the impulse-response function without imposing any assumptions ˆ on its form. The estimated coefficient βL = κ(−L) corresponds to our estimate for the the ˆ KAT at relative time −L. The βL are displayed as black dots in Figure 1.4c-e. Unexplained variations in search volume correspond to the residuals yit. We also estimate a separate version of Eq A.7 for monthly targets, where year index y is replaced by month index m and fmit is equal to one if t is in month m and zero otherwise, as well as a version for holiday targets, where y is replaced by holiday h and fhit is one if t = h and zero otherwise. Results from these three estimates explain a large amount of variation in search queries for these three classes of target (years: R2 = 0.93; months: R2 = 0.83; holidays: R2 = 0.81) and are displayed in Figure B.2. To examine the robustness of these results, we also estimate Eq A.7 including year and/or country-specific constants (fixed effects) that flexibly adjust for differences in search volume for each target at baseline between countries, for example accounting for differences in internet access or culture, and for nonlinear trends in search volume over time, for example due to changing usages patterns within countries. We find that adjusting for these factors in Eq. A.7 has essentially no effect on the estimated structure of the KAT (see Figure B.3). 59

Fitting an analytical form We seek a smooth analytical function that approximates the KAT estimated in Eq. A.7 by fitting different families of functions, evaluating both their goodness-of-fit to the coefficients ˆ βL and evaluating their ability to accurately predict search volume when convolved with f(t). For each family of functions, we independently fit parameters to the future-oriented and past-oriented portions of the KAT, keeping the present in both samples and forcing the two functions to intersect at the present. We estimate exponential functions of the form eη·t+β, motivated by intuition from economic discounting3, 116 and market interest rates117, 118 where inter-temporal tradeoffs between sequential future decisions should be self-consistent. th P8 ˙p We also estimate 8 order polynomial functions of the form p=0(γpct ) , motivated by their potential parsimony and simplicity. We also estimate functions that resemble models of “hyperbolic discounting” behavior widely documented in psychology and behavioral 23–27 a economics, which take the form 1+b·t . Note that while exponential and hyperbolic- discounting-like forms are motivated by prior research studying tradeoffs over time (i.e. “discounting”), there is not a clear theoretical linkage between the form of the KAT and the tradeoffs we expect populations to make over time. Additionally, note that neither of these existing research literature provide any guidance on the form of the past-oriented portion of a+b·t+c·t2 the KAT. Lastly, we estimate rational functions of the form 1+b·t+c·t2 where the numerator is allowed to be up to the second order and the denominator is restricted to be of order 2. The family of rational functions, broadly defined, technically includes the polynomial forms and hyperbolic-discounting-like forms that we examine, with specific restrictions on the order of the numerator and denominator. Comparisons between the fitted functions within each family are shown as colored lines in Figure 1.4c, and results from their convolution with f(t) are shown in Figure B.2. Fitted values from the rational family of functions minimize RMSE and also predict search volume with lowest errors. Using our sample of annual targets and 181 countries, the average KAT is estimated to have the form:

( −0.022·t2−4.777·t+99.975 −0.796·t+1 if t − θ < 0 (i.e. the past) (A.8)κ ˆ(t) = −0.004·t2+0.520·t+100.040 −0.002·t2+0.510·t+1 if t − θ > 0 (i.e. the future) which is displayed in Figure 1.4a.

Computing total attention to the future and past To compute attention to the past or future, we sum total attention to moments in time ˆ before or after the present. To do this, we sum βL = κ(−L) from L = −35 weeks (245 days) to L = −1 for the past and from L = 1 to +35 for the future. We compare these values to the probability mass at L = 0, which we interpret as attention to the present. We display these three values in Figure 1.6a (Figure 1.4a breaks down attention to the future and past into “distant” and “near” components by introducing an additional cutoﬀ at 100 days in both directions). We use block-bootstrap resampling to compute uncertainty in these measures 60

of attention, estimating Eq. A.7, ﬁtting an analytical function, and integrating the resulting function following each resampling of the data. Figure B.4a shows the distribution of total attention allocated to past, present, and future moments using the years targets, and Figure B.4b details the relative fraction of attention from these re-sampled values.

Country specific estimates To compare the allocation of attention over time across countries, we estimate the KAT only using data from one country at a time. We estimate Eq A.7 for each country separately ˆ and use the country-specific values for βL to fit country-specific rational functions. Figures 1.4b and Figure B.5 show country specific KAT estimates for our sample of twenty representative countries, the later displaying estimates using annual, monthly, and daily targets. We integrate these curves using the approach above to compute total attention to the past, present, and future for each country, using annual targets. Tables B.2-B.5 list these values for each country, which are also displayed in three dimensions in Figure 1.6, broken down by region, and in the maps of Figure 1.5a-b.

Estimating trends in the KAT To compute trends in attention to the past present and future, we pool query data for targets across our sample of twenty countries shown in Figure 1.4b. We then estimate the KAT for each class of target only using data from a single year. We integrate the resulting estimated fraction of attention to past/present/future. These values are plotted in Figure 1.5c using models for each class of target for each of the ten years in our sample. 61

Appendix B

Supplemental Figures and Tables

B.1 Figures 62

Figure B.1: Google search as convolution of KAT with daily dummies

Notes: Google search as convolution of KAT with daily dummies. Illustration of the convolution associated with input functions (black curves) of diﬀerent lengths. These would correspond to the daily, monthly and yearly event lengths (f(t)) considered in our analysis. The response function (orange curves) is the same across all three panels. It corresponds to the Kernel of Attention to Time (KAT) we wish to recover. The output ﬁgures (purple curves) correspond to the convolution, i.e. the raw data one would observe when querying temporal events on Google Trends (as shown in Figure 1.2g-h). 63

Figure B.2: Predicted search volume using estimated KATs

Notes: The three panels show the predicted search volume – using the process described in Figure B.1, – of the four classes of ﬁtted functions used to determine the KAT estimates shown in Figure 1.4c-e against the average raw queries data generated by Google Trends when searching for the holidays shown in Table B.1, the months of the year, and the calendar years from 2009 to 2018 for the 20 countries shown in Figure 1.4b. The black dots correspond to the country-target averaged search scores observed in Google for the daily targets, monthly targets and annual targets. The four KATs derived using the ﬁtting functions shown in Figure 1.4 are convolved to predict the relative search volume for each target length. The RMSE associated with each functional form is shown in parentheses. 64

Figure B.3: Estimated KAT with alternative model choices

Notes: The red dashed line in each panel represents a KAT estimated with a different set of fixed effects that account for possible omitted variable bias. The solid gray line is our main result without fixed effect also plotted in Figure 1.4. The shape of the KAT appears robust to the addition and choice of fixed effect. Four different set of fixed effect models are plotted here: (a) country level to account for country-specific characteristics that remain constant over time, (b) country and year level to account for country-specific characteristics that remain constant over time and for time trends affecting all countries, (c) country, year, and query level to account for constant country-specific unobserved characteristics, time trends affecting all countries and query-specific unobserved characteristics (e.g. years associated with international competitions such as the Olympic games may have a different KAT structure than years without such high-focus events), and (d) country-by-year level to account for unobserved variation at the country level within any given year. 65

Figure B.4: Robustness checks for the shares of attention to diﬀerent moments in time

Notes: (a) Computation of the the total attention allocated to past, present, and future for 3000 iterative block-bootstrap at the query-by-location level using annual targets. (b) shows the ratio of attention allocated to the present relative to every day in the relative past (blue curve) and future (red curve) using bootstrap. It shows that individuals are equally likely to think about the past or future, relative to the present, for the first 20 days. For more distant days (170 days to 200 days) in relative time, individuals are estimated to be more prone to thinking about the past than the future, albeit the difference is not statistically significant. 66

Figure B.5: Estimated KAT using daily, monthly and yearly targets

Notes: Estimated KAT using daily, monthly and yearly targets. Estimated coeﬃcients con- stitutive of the KAT. The panel on the left, center and right show the coeﬃcients capturing interest for the holiday targets for the sample of 20 countries listed in Table B.1, monthly targets (e.g. “July”) and annual targets (e.g. “2016”), respectively. 67

Figure B.6: Country boundaries shapeﬁle

Sources: Natural Earth’s 1:10m Cultural Vectors shapeﬁle. Notes: The map was cropped at 85.0◦N and 60.0◦S by the authors. 68

Figure B.7: Global CDD maps with alternative CDD thresholds

Notes: (A): Global Cooling Degree Days using a 15.6◦C (60◦F) threshold. (B): Global Cooling Degree Days using a 21.1◦C (70◦F) threshold. (C): Global Cooling Degree Days using a 23.9◦C (75◦F) threshold. 69

Figure B.8: Expansion of vessel tracking and VLMPA coverage over time

Notes: (A) Evolution of the number of vessels and ﬁshing hours tracked by GFW over the 2012 to 2018 period. (B) Expansion of all types of Very Large Marine Protected Areas over time (1975-2019). VLMPAs now cover 17,002,904 km2, which includes 7,802,962 km2 of marine reserves, i.e. areas where no ﬁshing is permitted (also known as “no-take” areas). 70

Figure B.9: Evolution of average daily hours of ﬁshing eﬀort per 1,000 km2 for all VLMPAs

Notes: Blue dots measure the sum of fishing hours spent by vessels emitting prior to the announcement of the creation of an VLMPA. Red dots show the sum of fishing hours spent by vessels emitting for the first time after the announcement. The dates have been standardized so that each period – before announcement, between announcement & implementation, and after implementation – would last for 100 days, meaning that the period ranging from January 1, 2012 to December 31, 2018 corresponds to 300 days under this standardized date format. The announcement of the creation of an VLMPA occurs on the 100th day (solid green line). The implementation occurs on the 200th day (solid orange line). 71

Figure B.10: Displacement of ﬁshing eﬀort

Notes: Comparison of the annualized count of fishing effort inside the VLMPAs prior to the announcement of their creation and after their implementation. All vessels are included, regardless of the date of their first appearance in the GFW database. 72

Figure B.11: Evolution of the number of vessels by country

Notes: Number of vessels practicing banned fishing in any of the 8 VLMPAs. Restricted vessels refer to the vessels which were tracked prior to the announcement that a protected area would be created. All refer to all the vessels to ever practice banned fishing, regardless of the date of their first appearance in the database. 73

B.2 Tables

Table B.1: Country-speciﬁc holidays

Country First holiday Second holiday Third holiday Argentina Day of Remembrance for International Workers’ New Year’s Eve Truth and Justice Day Australia Australia Day Easter Monday National New Year’s Eve Brazil Tiradentes Day Independence Day All Souls’ Day Chile Valentine’s Day Christmas Day Day New Year’s Eve Colombia Maundy Thursday Teacher’s Day New Year’s Eve France Victory in Europe Day Bastille Day New Year’s Eve Germany International Women’s German Unity Day New Year’s Eve Day Hong-Kong Valentine’s Day Halloween Christmas Day India Republic Day Independence Day New Year’s Eve Indonesia Valentine’s Day Independence Day The Prophet Muham- mad’s Birthday Mexico Cinco de Mayo Cry of Dolores Columbus Day Netherlands Valentine’s Day Liberation Day Ascension Day New-Zealand Waitangi Day Queen’s Birthday Halloween Norway 17 May Constitution Whit Sunday New Year’s Eve Day Peru Valentine’s Day International Workers’ New Year’s Eve Day Singapore Valentine’s Day Vesak Day National Day Spain Spanish National Holi- Constitution Day New Year’s Eve day Sweden Valentine’s Day Walpurgis Night National Day United Kingdom Good Friday Saint Patrick’s Day Boxing Day United States Valentine’s Day Independence Day Mother’s Day Notes: Country-speciﬁc holidays used to estimate the KAT in Figure 1.4c. 74

Table B.2: Shares of attention to the past, present and future for alphabetically ranked countries 1 to 50

Country Share of attention Share of attention Share of attention to the past to the present to the future Afghanistan 0.4 (0.03) 0.25 (0.03) 0.35 (0.03) Albania 0.5 (0.02) 0.16 (0.02) 0.34 (0.02) Algeria 0.39 (0.02) 0.24 (0.02) 0.37 (0.02) Angola 0.44 (0.02) 0.28 (0.03) 0.28 (0.02) Argentina 0.35 (0.01) 0.23 (0.01) 0.42 (0.01) Armenia 0.65 (0.02) 0.09 (0.03) 0.27 (0.02) Aruba 0.34 (0.02) 0.26 (0.03) 0.4 (0.02) Australia 0.28 (0.01) 0.33 (0.01) 0.39 (0.01) Austria 0.3 (0.01) 0.2 (0.01) 0.5 (0.01) Azerbaijan 0.41 (0.02) 0.21 (0.02) 0.38 (0.02) Bahamas 0.37 (0.02) 0.35 (0.02) 0.28 (0.02) Bahrain 0.41 (0.01) 0.22 (0.02) 0.37 (0.01) Bangladesh 0.46 (0.02) 0.28 (0.03) 0.26 (0.02) Barbados 0.41 (0.02) 0.3 (0.02) 0.29 (0.02) Belarus 0.5 (0.01) 0.15 (0.02) 0.34 (0.01) Belgium 0.31 (0.01) 0.22 (0.01) 0.46 (0.01) Belize 0.42 (0.03) 0.29 (0.03) 0.29 (0.03) Benin 0.37 (0.03) 0.31 (0.03) 0.32 (0.03) Bhutan 0.35 (0.05) 0.36 (0.06) 0.29 (0.05) Bolivia 0.46 (0.01) 0.19 (0.01) 0.35 (0.01) Bosnia and Herz. 0.46 (0.02) 0.17 (0.02) 0.37 (0.02) Botswana 0.49 (0.03) 0.23 (0.03) 0.28 (0.03) Brazil 0.3 (0.01) 0.24 (0.01) 0.46 (0.01) Brunei 0.43 (0.02) 0.27 (0.02) 0.29 (0.02) Bulgaria 0.36 (0.01) 0.27 (0.01) 0.36 (0.01) Burkina Faso 0.38 (0.03) 0.29 (0.03) 0.33 (0.03) Burundi 0.35 (0.05) 0.28 (0.06) 0.37 (0.05) Cabo Verde 0.46 (0.03) 0.18 (0.04) 0.36 (0.03) Cambodia 0.47 (0.03) 0.23 (0.03) 0.3 (0.03) Cameroon 0.47 (0.02) 0.18 (0.03) 0.35 (0.02) Canada 0.32 (0.01) 0.35 (0.01) 0.33 (0.01) Central African Rep. 0.24 (0.09) 0.34 (0.11) 0.42 (0.09) Chad 0.24 (0.04) 0.45 (0.05) 0.31 (0.04) Chile 0.29 (0.01) 0.25 (0.02) 0.46 (0.01) China 0.42 (0.02) 0.19 (0.03) 0.39 (0.02) Colombia 0.45 (0.01) 0.21 (0.01) 0.34 (0.01) Comoros 0.38 (0.09) 0.39 (0.11) 0.23 (0.09) Congo 0.41 (0.04) 0.28 (0.05) 0.31 (0.04) Costa Rica 0.42 (0.01) 0.21 (0.02) 0.37 (0.01) Croatia 0.42 (0.01) 0.17 (0.01) 0.41 (0.01) Cuba 0.45 (0.02) 0.24 (0.02) 0.31 (0.02) Cura¸cao 0.37 (0.03) 0.26 (0.03) 0.38 (0.03) Cyprus 0.48 (0.01) 0.21 (0.02) 0.31 (0.01) Czechia 0.36 (0.01) 0.23 (0.01) 0.4 (0.01) Cˆoted’Ivoire 0.45 (0.02) 0.19 (0.02) 0.36 (0.02) Dem. Rep. Congo 0.34 (0.03) 0.24 (0.04) 0.42 (0.03) Denmark 0.34 (0.01) 0.26 (0.02) 0.4 (0.01) Djibouti 0.43 (0.03) 0.24 (0.04) 0.33 (0.03) Dominican Rep. 0.57 (0.01) 0.2 (0.01) 0.23 (0.01) Ecuador 0.44 (0.01) 0.26 (0.01) 0.29 (0.01)

Notes: This table shows the shares of temporal attention and their related standard errors based on variance estimates given by the inverse of the negative Hessian matrix (in parentheses) for countries alphabetically ranked 1 to 50. 75

Table B.3: Shares of attention to the past, present and future for alphabetically ranked countries 51 to 100

Country Share of attention Share of attention Share of attention to the past to the present to the future Egypt 0.42 (0.02) 0.18 (0.02) 0.4 (0.02) El Salvador 0.44 (0.02) 0.22 (0.02) 0.34 (0.02) Eq. Guinea 0.35 (0.05) 0.29 (0.06) 0.36 (0.05) Estonia 0.41 (0.02) 0.23 (0.02) 0.35 (0.02) Eswatini 0.26 (0.06) 0.44 (0.07) 0.3 (0.06) Ethiopia 0.57 (0.06) 0.25 (0.06) 0.18 (0.06) Fiji 0.47 (0.02) 0.34 (0.02) 0.19 (0.02) Finland 0.28 (0.01) 0.17 (0.01) 0.56 (0.01) Fr. Polynesia 0.38 (0.02) 0.22 (0.02) 0.4 (0.02) France 0.31 (0.01) 0.18 (0.01) 0.51 (0.01) Gabon 0.51 (0.03) 0.2 (0.03) 0.29 (0.03) Gambia 0.42 (0.04) 0.25 (0.05) 0.33 (0.04) Georgia 0.56 (0.02) 0.1 (0.02) 0.34 (0.02) Germany 0.3 (0.01) 0.24 (0.01) 0.47 (0.01) Ghana 0.45 (0.02) 0.26 (0.02) 0.29 (0.02) Greece 0.41 (0.01) 0.18 (0.02) 0.42 (0.01) Grenada 0.46 (0.05) 0.32 (0.05) 0.22 (0.05) Guam 0.39 (0.02) 0.24 (0.02) 0.37 (0.02) Guatemala 0.46 (0.01) 0.27 (0.02) 0.28 (0.01) Guinea 0.31 (0.06) 0.35 (0.07) 0.34 (0.06) Guyana 0.39 (0.03) 0.43 (0.03) 0.18 (0.03) Haiti 0.52 (0.02) 0.14 (0.03) 0.34 (0.03) Honduras 0.5 (0.02) 0.2 (0.02) 0.3 (0.02) Hong Kong 0.41 (0.01) 0.23 (0.01) 0.36 (0.01) Hungary 0.36 (0.01) 0.39 (0.01) 0.26 (0.01) Iceland 0.44 (0.02) 0.27 (0.02) 0.3 (0.02) India 0.37 (0.01) 0.29 (0.02) 0.34 (0.01) Indonesia 0.31 (0.01) 0.42 (0.01) 0.27 (0.01) Iran 0.46 (0.03) 0.15 (0.03) 0.4 (0.03) Iraq 0.27 (0.02) 0.21 (0.03) 0.52 (0.02) Ireland 0.36 (0.01) 0.33 (0.01) 0.31 (0.01) Israel 0.35 (0.01) 0.25 (0.01) 0.4 (0.01) Italy 0.32 (0.01) 0.2 (0.01) 0.47 (0.01) Jamaica 0.44 (0.01) 0.34 (0.02) 0.22 (0.02) Japan 0.33 (0.01) 0.2 (0.01) 0.47 (0.01) Jordan 0.53 (0.02) 0.07 (0.02) 0.4 (0.02) Kazakhstan 0.47 (0.02) 0.23 (0.02) 0.3 (0.02) Kenya 0.5 (0.02) 0.23 (0.02) 0.27 (0.02) Kuwait 0.44 (0.01) 0.21 (0.02) 0.35 (0.02) Kyrgyzstan 0.42 (0.03) 0.28 (0.03) 0.3 (0.03) Laos 0.43 (0.02) 0.23 (0.03) 0.34 (0.02) Latvia 0.43 (0.02) 0.18 (0.02) 0.39 (0.02) Lesotho 0.33 (0.06) 0.38 (0.07) 0.29 (0.06) Liberia 0.49 (0.05) 0.35 (0.06) 0.17 (0.05) Libya 0.33 (0.02) 0.18 (0.03) 0.49 (0.02) Lithuania 0.39 (0.01) 0.16 (0.01) 0.45 (0.01) Luxembourg 0.29 (0.01) 0.23 (0.01) 0.49 (0.01) Macao 0.42 (0.02) 0.23 (0.02) 0.35 (0.02) Macedonia 0.49 (0.02) 0.11 (0.02) 0.41 (0.02) Madagascar 0.41 (0.02) 0.23 (0.02) 0.36 (0.02)

Notes: This table shows the shares of temporal attention and their related standard errors based on variance estimates given by the inverse of the negative Hessian matrix (in parentheses) for countries alphabetically ranked 51 to 100. 76

Table B.4: Shares of attention to the past, present and future for alphabetically ranked countries 101 to 150

Country Share of attention Share of attention Share of attention to the past to the present to the future Malawi 0.46 (0.04) 0.31 (0.04) 0.23 (0.04) Malaysia 0.39 (0.01) 0.18 (0.01) 0.43 (0.01) Maldives 0.5 (0.03) 0.19 (0.03) 0.31 (0.03) Mali 0.44 (0.03) 0.21 (0.03) 0.35 (0.03) Malta 0.38 (0.02) 0.27 (0.02) 0.35 (0.02) Mauritania 0.49 (0.03) 0.23 (0.03) 0.28 (0.03) Mauritius 0.47 (0.02) 0.25 (0.02) 0.28 (0.02) Mexico 0.42 (0.01) 0.26 (0.01) 0.31 (0.01) Micronesia 0.33 (0.09) 0.41 (0.1) 0.26 (0.09) Moldova 0.65 (0.02) 0.07 (0.02) 0.28 (0.02) Mongolia 0.54 (0.03) 0.2 (0.03) 0.26 (0.03) Montenegro 0.39 (0.02) 0.13 (0.02) 0.47 (0.02) Morocco 0.47 (0.02) 0.19 (0.02) 0.34 (0.02) Mozambique 0.4 (0.02) 0.28 (0.03) 0.32 (0.02) Myanmar 0.51 (0.04) 0.29 (0.05) 0.2 (0.04) Nepal 0.51 (0.03) 0.31 (0.03) 0.18 (0.03) Netherlands 0.31 (0.01) 0.34 (0.01) 0.35 (0.01) New Caledonia 0.39 (0.02) 0.22 (0.03) 0.39 (0.02) New Zealand 0.38 (0.01) 0.26 (0.01) 0.36 (0.01) Nicaragua 0.52 (0.02) 0.17 (0.02) 0.31 (0.02) Niger 0.35 (0.04) 0.37 (0.05) 0.28 (0.04) Nigeria 0.4 (0.02) 0.27 (0.02) 0.33 (0.02) Norway 0.31 (0.01) 0.27 (0.01) 0.42 (0.01) Oman 0.46 (0.02) 0.18 (0.02) 0.36 (0.02) Pakistan 0.47 (0.01) 0.34 (0.02) 0.19 (0.02) Palestine 0.46 (0.03) 0.12 (0.03) 0.42 (0.03) Panama 0.42 (0.01) 0.2 (0.02) 0.37 (0.01) Papua New Guinea 0.38 (0.05) 0.33 (0.05) 0.29 (0.05) Paraguay 0.38 (0.02) 0.26 (0.02) 0.36 (0.02) Peru 0.47 (0.01) 0.22 (0.01) 0.31 (0.01) Philippines 0.51 (0.01) 0.23 (0.01) 0.26 (0.01) Poland 0.43 (0.01) 0.24 (0.01) 0.34 (0.01) Portugal 0.32 (0.01) 0.24 (0.01) 0.45 (0.01) Puerto Rico 0.4 (0.01) 0.25 (0.01) 0.35 (0.01) Qatar 0.42 (0.01) 0.24 (0.02) 0.34 (0.01) Romania 0.43 (0.01) 0.3 (0.01) 0.28 (0.01) Rwanda 0.43 (0.03) 0.3 (0.04) 0.27 (0.04) S. Sudan 0.31 (0.04) 0.47 (0.04) 0.22 (0.04) Saint Lucia 0.33 (0.04) 0.39 (0.04) 0.28 (0.04) Samoa 0.42 (0.06) 0.35 (0.07) 0.22 (0.06) Saudi Arabia 0.41 (0.02) 0.19 (0.02) 0.4 (0.02) Senegal 0.5 (0.02) 0.25 (0.02) 0.25 (0.02) Serbia 0.35 (0.01) 0.13 (0.01) 0.52 (0.01) Sierra Leone 0.48 (0.04) 0.26 (0.05) 0.26 (0.05) Singapore 0.38 (0.01) 0.26 (0.01) 0.36 (0.01) Slovakia 0.4 (0.01) 0.26 (0.01) 0.34 (0.01) Slovenia 0.37 (0.01) 0.17 (0.01) 0.47 (0.01) Somalia 0.35 (0.05) 0.37 (0.06) 0.28 (0.05) South Africa 0.45 (0.01) 0.27 (0.02) 0.29 (0.01) South Korea 0.42 (0.02) 0.16 (0.02) 0.42 (0.02)

Notes: This table shows the shares of temporal attention and their related standard errors based on variance estimates given by the inverse of the negative Hessian matrix (in parentheses) for countries alphabetically ranked 101 to 150. 77

Table B.5: Shares of attention to the past, present and future for alphabetically ranked countries 151 to 181

Country Share of attention Share of attention Share of attention to the past to the present to the future Spain 0.44 (0.01) 0.24 (0.01) 0.32 (0.01) Sri Lanka 0.49 (0.02) 0.22 (0.02) 0.29 (0.02) St. Vin. and Gren. 0.47 (0.03) 0.27 (0.04) 0.26 (0.03) Sudan 0.4 (0.02) 0.24 (0.03) 0.36 (0.02) Suriname 0.44 (0.02) 0.28 (0.03) 0.28 (0.02) Sweden 0.32 (0.01) 0.33 (0.01) 0.35 (0.01) Switzerland 0.26 (0.01) 0.27 (0.01) 0.47 (0.01) Syria 0.35 (0.03) 0.08 (0.03) 0.57 (0.02) Taiwan 0.4 (0.01) 0.15 (0.01) 0.45 (0.01) Tajikistan 0.48 (0.03) 0.23 (0.03) 0.29 (0.03) Tanzania 0.43 (0.03) 0.31 (0.03) 0.26 (0.03) Thailand 0.45 (0.02) 0.24 (0.02) 0.32 (0.02) Timor-Leste 0.34 (0.03) 0.44 (0.04) 0.22 (0.03) Togo 0.38 (0.03) 0.3 (0.04) 0.32 (0.03) Trinidad and Tobago 0.44 (0.02) 0.25 (0.02) 0.31 (0.02) Tunisia 0.42 (0.01) 0.22 (0.02) 0.37 (0.01) Turkey 0.37 (0.01) 0.26 (0.02) 0.38 (0.01) Turkmenistan 0.38 (0.03) 0.2 (0.03) 0.42 (0.03) U.S. Virgin Is. 0.27 (0.04) 0.44 (0.05) 0.29 (0.04) Uganda 0.44 (0.02) 0.28 (0.03) 0.28 (0.02) United Arab Emirates 0.42 (0.01) 0.22 (0.01) 0.36 (0.01) United Kingdom 0.33 (0.01) 0.32 (0.01) 0.35 (0.01) United States of America 0.32 (0.01) 0.35 (0.01) 0.34 (0.01) Uruguay 0.39 (0.01) 0.19 (0.01) 0.41 (0.01) Uzbekistan 0.49 (0.03) 0.14 (0.03) 0.37 (0.03) Vanuatu 0.43 (0.06) 0.26 (0.07) 0.31 (0.06) Venezuela 0.48 (0.01) 0.25 (0.01) 0.27 (0.01) Vietnam 0.48 (0.02) 0.15 (0.02) 0.36 (0.02) Yemen 0.35 (0.03) 0.21 (0.03) 0.43 (0.03) Zambia 0.37 (0.03) 0.35 (0.04) 0.28 (0.03) Zimbabwe 0.49 (0.02) 0.3 (0.03) 0.21 (0.02) Notes: This table shows the shares of temporal attention and their related standard errors based on variance estimates given by the inverse of the negative Hessian matrix (in parentheses) for countries alphabetically ranked 151 to 181. 78

Table B.6: Total CDD exposure for the Top 50 Countries

Country Population-Weighted 95% Conﬁdence Total CDD 95% Conﬁdence Annual Cooling Interval Exposure Interval Degree Days (in billions)

1. India 2,848 [2,762, 2,934] 3728.2 [3615.6, 3840.8] 2. China 1,009 [970, 1,048] 1409.6 [1355.1, 1464.1] 3. Indonesia 3,284 [3,096, 3,472] 847.8 [799.3, 896.3] 4. Nigeria 3,429 [3,311, 3,547] 621.3 [599.9, 642.7] 5. Pakistan 2,504 [2,316, 2,692] 474.2 [438.6, 509.8] 6. Brazil 2,108 [2,026, 2,190] 434.2 [417.3, 451.1] 7. Bangladesh 2,644 [2,494, 2,794] 426.2 [402, 450.4] 8. Philippines 3,266 [3,101, 3,431] 332.2 [315.4, 349] 9. United States 867 [857, 877] 277.4 [274.2, 280.6] 10. Vietnam 2,777 [2,654, 2,900] 259.8 [248.3, 271.4] 11. Thailand 3,471 [3,417, 3,525] 238.3 [234.6, 242] 12. Egypt 1,934 [1,778, 2,090] 181.4 [166.7, 196] 13. Myanmar 3,082 [2,915, 3,249] 161.5 [152.8, 170.3] 14. Dem. Rep. Congo 2,115 [1,938, 2,292] 161.2 [147.7, 174.6] 15. Mexico 1,229 [1,097, 1,361] 154.7 [138.1, 171.3] 16. Ethiopia 1,459 [889, 2,029] 145.7 [88.8, 202.6] 17. Sudan 3,746 [3,209, 4,283] 144.8 [124, 165.5] 18. Iran 1,515 [1,410, 1,620] 120.2 [111.9, 128.6] 19. Japan 886 [834, 938] 113.4 [106.7, 120] 20. Saudi Arabia 3,347 [3,037, 3,657] 105.6 [95.8, 115.4] 21. Venezuela 3,376 [3,221, 3,531] 105.2 [100.4, 110] 22. Malaysia 3,375 [3,265, 3,485] 103.7 [100.3, 107.1] 23. Iraq 2,658 [2,520, 2,796] 96.0 [91, 101] 24. Ghana 3,326 [3,238, 3,414] 91.7 [89.3, 94.2] 25. Colombia 1,879 [1,176, 2,582] 90.6 [56.7, 124.5] 26. Niger 4,030 [3,881, 4,179] 80.2 [77.2, 83.1] 27. Cˆoted’Ivoire 3,156 [3,072, 3,240] 72.9 [71, 74.9] 28. Turkey 917 [868, 966] 71.8 [67.9, 75.6] 29. Burkina Faso 3,817 [3,734, 3,900] 69.1 [67.6, 70.6] 30. Cameroon 3,022 [2,664, 3,380] 69.0 [60.8, 77.2] 31. Mali 3,800 [3,678, 3,922] 66.4 [64.2, 68.5] 32. Uganda 1,612 [1,309, 1,915] 64.7 [52.5, 76.9] 33. Angola 2,293 [2,133, 2,453] 63.9 [59.4, 68.3] 34. Sri Lanka 2,967 [2,216, 3,718] 61.5 [45.9, 77] 35. Yemen 2,269 [1,523, 3,015] 61.1 [41, 81.2] 36. Mozambique 2,129 [1,951, 2,307] 59.6 [54.6, 64.6] 37. Cambodia 3,649 [3,389, 3,909] 56.6 [52.6, 60.7] 38. Chad 3,981 [3,750, 4,212] 55.8 [52.5, 59] 39. Senegal 3,443 [3,131, 3,755] 51.6 [46.9, 56.2] 40. Taiwan 2,162 [1,984, 2,340] 50.8 [46.6, 55] 41. Kenya 1,051 [786, 1,316] 49.6 [37.1, 62.2] 42. Algeria 1,217 [1,108, 1,326] 48.5 [44.2, 52.9] 43. Nepal 1,642 [1,162, 2,122] 47.1 [33.3, 60.8] 44. Afghanistan 1,367 [1,194, 1,540] 46.1 [40.3, 52] 45. Guinea 3,537 [3,368, 3,706] 42.8 [40.7, 44.8] 46. Madagascar 1,703 [928, 2,478] 41.3 [22.5, 60.1] 47. Argentina 890 [843, 937] 38.6 [36.6, 40.7] 48. Haiti 3,580 [3,126, 4,034] 38.3 [33.5, 43.2] 49. Somalia 2,678 [2,177, 3,179] 37.2 [30.3, 44.2] 50. Italy 610 [546, 674] 36.3 [32.5, 40.1]

Notes: This table ranks the top 50 countries worldwide by total CDD exposure. Annual cooling degree days (CDD) and CDD exposure are deﬁned as in Table 1. 79

Table B.7: Total CDD exposure for Countries Ranked 51 to 100

Country Population-Weighted 95% Conﬁdence Total CDD 95% Conﬁdence Annual Cooling Interval Exposure Interval Degree Days (in billions)

51. Morocco 1,039 [888, 1,190] 36.2 [30.9, 41.4] 52. Benin 3,388 [3,268, 3,508] 35.8 [34.6, 37.1] 53. South Africa 645 [598, 692] 35.7 [33.1, 38.3] 54. Korea 704 [673, 735] 35.6 [34, 37.2] 55. Russia 245 [237, 253] 35.3 [34.1, 36.4] 56. Spain 754 [697, 811] 35.0 [32.3, 37.6] 57. United Arab Emirates 3,678 [3,512, 3,844] 33.7 [32.1, 35.2] 58. Uzbekistan 1,083 [960, 1,206] 33.5 [29.7, 37.4] 59. Dominican Rep. 3,147 [2,915, 3,379] 33.1 [30.7, 35.6] 60. Malawi 1,829 [1,425, 2,233] 32.1 [25, 39.2] 61. Peru 962 [572, 1,352] 30.2 [17.9, 42.4] 62. Cuba 2,560 [2,424, 2,696] 29.3 [27.8, 30.9] 63. Syria 1,444 [1,350, 1,538] 27.1 [25.3, 28.8] 64. Zambia 1,637 [1,327, 1,947] 26.4 [21.4, 31.3] 65. Togo 3,391 [3,322, 3,460] 25.2 [24.6, 25.7] 66. S. Sudan 2,102 [1,577, 2,627] 25.0 [18.7, 31.2] 67. Sierra Leone 3,358 [3,200, 3,516] 24.3 [23.2, 25.4] 68. Papua New Guinea 3,015 [2,688, 3,342] 23.9 [21.3, 26.5] 69. Guatemala 1,462 [640, 2,284] 23.8 [10.4, 37.1] 70. Honduras 2,501 [1,861, 3,141] 22.4 [16.7, 28.1] 71. Zimbabwe 1,338 [1,060, 1,616] 21.1 [16.7, 25.5] 72. Singapore 3,683 [3,578, 3,788] 20.4 [19.8, 21] 73. Nicaragua 3,222 [2,746, 3,698] 19.6 [16.7, 22.5] 74. Ecuador 1,185 [615, 1,755] 19.1 [9.9, 28.3] 75. Lao PDR 2,810 [2,676, 2,944] 18.7 [17.8, 19.6] 76. El Salvador 2,913 [2,522, 3,304] 18.4 [15.9, 20.9] 77. Burundi 1,784 [1,250, 2,318] 18.2 [12.7, 23.6] 78. France 281 [268, 294] 18.1 [17.3, 19] 79. Ukraine 371 [361, 381] 16.6 [16.1, 17] 80. Australia 688 [666, 710] 16.4 [15.9, 16.9] 81. Mauritania 3,748 [3,358, 4,138] 15.7 [14, 17.3] 82. Tunisia 1,333 [1,245, 1,421] 15.0 [14, 16] 83. Hong Kong 2,030 [2,019, 2,041] 14.7 [14.6, 14.8] 84. Rwanda 1,236 [940, 1,532] 14.4 [10.9, 17.8] 85. Oman 3,315 [2,687, 3,943] 13.9 [11.3, 16.6] 86. Kuwait 3,523 [3,461, 3,585] 13.9 [13.6, 14.1] 87. Liberia 3,055 [2,952, 3,158] 13.7 [13.3, 14.2] 88. Congo 2,707 [2,587, 2,827] 13.5 [12.9, 14.1] 89. Panama 3,357 [2,944, 3,770] 13.3 [11.7, 15] 90. Paraguay 1,931 [1,829, 2,033] 12.8 [12.1, 13.5] 91. Central African Rep. 2,794 [2,528, 3,060] 12.7 [11.5, 13.9] 92. Israel 1,489 [1,253, 1,725] 12.0 [10.1, 13.9] 93. Eritrea 2,440 [1,958, 2,922] 11.8 [9.5, 14.2] 94. Jordan 1,274 [1,079, 1,469] 11.7 [9.9, 13.5] 95. Dem. Rep. Korea 457 [417, 497] 11.5 [10.5, 12.5] 96. Puerto Rico 3,087 [3,000, 3,174] 11.3 [11, 11.7] 97. Germany 134 [123, 145] 10.9 [10.1, 11.8] 98. Bolivia 997 [534, 1,460] 10.7 [5.7, 15.7] 99. Greece 951 [872, 1,030] 10.7 [9.8, 11.6] 100. Costa Rica 2,194 [1,237, 3,151] 10.5 [5.9, 15.2]

Notes: This table shows countries ranked 51 to 100 worldwide by total CDD exposure. Annual cooling degree days (CDD) and CDD exposure are deﬁned as in Table 1. 80

Table B.8: Total CDD exposure for Countries Ranked 101 to 150

Country Population-Weighted 95% Conﬁdence Total CDD 95% Conﬁdence Annual Cooling Interval Exposure Interval Degree Days (in billions)

101. Libya 1,672 [1,548, 1,796] 10.4 [9.7, 11.2] 102. Kazakhstan 546 [510, 582] 9.7 [9.1, 10.3] 103. Jamaica 3,334 [2,895, 3,773] 9.6 [8.3, 10.8] 104. Qatar 3,482 [3,299, 3,665] 8.6 [8.2, 9.1] 105. Turkmenistan 1,494 [1,392, 1,596] 8.3 [7.7, 8.9] 106. Lebanon 1,345 [742, 1,948] 7.9 [4.3, 11.4] 107. Romania 390 [347, 433] 7.8 [6.9, 8.6] 108. Tajikistan 902 [713, 1,091] 7.7 [6.1, 9.3] 109. Azerbaijan 788 [688, 888] 7.6 [6.6, 8.5] 110. Canada 181 [168, 194] 6.5 [6, 7] 111. Gambia 3,288 [2,192, 4,384] 6.5 [4.3, 8.7] 112. Guinea-Bissau 3,572 [3,381, 3,763] 6.3 [6, 6.7] 113. Poland 145 [132, 158] 5.5 [5.1, 6] 114. Gabon 2,857 [2,768, 2,946] 5.5 [5.3, 5.7] 115. Bahrain 3,444 [3,421, 3,467] 4.7 [4.7, 4.8] 116. Chile 256 [186, 326] 4.5 [3.3, 5.8] 117. Trinidad and Tobago 3,264 [3,119, 3,409] 4.4 [4.2, 4.6] 118. Namibia 1,751 [1,512, 1,990] 4.2 [3.7, 4.8] 119. Timor-Leste 3,386 [3,254, 3,518] 4.2 [4, 4.4] 120. Portugal 380 [308, 452] 4.0 [3.2, 4.7] 121. Serbia 447 [414, 480] 4.0 [3.7, 4.2] 122. Kyrgyzstan 653 [582, 724] 3.8 [3.4, 4.2] 123. Djibouti 4,039 [2,404, 5,674] 3.7 [2.2, 5.3] 124. Eq. Guinea 2,963 [2,842, 3,084] 3.5 [3.3, 3.6] 125. Botswana 1,572 [1,482, 1,662] 3.5 [3.3, 3.7] 126. Hungary 352 [320, 384] 3.4 [3.1, 3.8] 127. Uruguay 824 [689, 959] 2.8 [2.4, 3.3] 128. Bulgaria 384 [289, 479] 2.8 [2.1, 3.4] 129. Guyana 3,245 [2,408, 4,082] 2.5 [1.9, 3.1] 130. Comoros 3,129 [2,959, 3,299] 2.4 [2.3, 2.6] 131. Fiji 2,499 [1,815, 3,183] 2.2 [1.6, 2.8] 132. Croatia 497 [443, 551] 2.1 [1.9, 2.3] 133. United Kingdom 32 [27, 37] 2.1 [1.8, 2.4] 134. Georgia 529 [390, 668] 2.1 [1.5, 2.6] 135. Mauritius 1,655 [873, 2,437] 2.1 [1.1, 3.1] 136. Albania 706 [615, 797] 2.1 [1.8, 2.3] 137. Suriname 3,405 [2,904, 3,906] 1.9 [1.6, 2.2] 138. Solomon Is. 3,170 [2,952, 3,388] 1.9 [1.7, 2] 139. Bosnia and Herz. 494 [410, 578] 1.7 [1.4, 2] 140. Cyprus 1,438 [1,386, 1,490] 1.7 [1.6, 1.7] 141. Austria 191 [170, 212] 1.7 [1.5, 1.8] 142. Maldives 3,886 [3,744, 4,028] 1.6 [1.6, 1.7] 143. Armenia 544 [381, 707] 1.6 [1.1, 2.1] 144. Belarus 164 [141, 187] 1.6 [1.3, 1.8] 145. Czech Rep. 140 [122, 158] 1.5 [1.3, 1.7] 146. Brunei 3,416 [3,274, 3,558] 1.4 [1.4, 1.5] 147. Swaziland 1,059 [779, 1,339] 1.4 [1, 1.8] 148. Cape Verde 2,385 [2,211, 2,559] 1.3 [1.2, 1.4] 149. Netherlands 71 [65, 77] 1.2 [1.1, 1.3] 150. Slovakia 201 [172, 230] 1.1 [0.9, 1.3]

Notes: This table shows countries ranked 101 to 150 worldwide by total CDD exposure. Annual cooling degree days (CDD) and CDD exposure are deﬁned as in Table 1. 81

Table B.9: Total CDD exposure for the Top 50 Cities

City Annual Cooling 95% Conﬁdence Total CDD 95% Conﬁdence Degree Days Interval Exposure Interval (CDDs) (in billions)

1. Mumbai, India 3,544 [3,386, 3,702] 74.6 [71.3, 77.9] 2. Delhi, India 2,831 [2,469, 3,193] 72.8 [63.5, 82.1] 3. Dhaka, Bangladesh 2,955 [2,656, 3,254] 52.0 [46.7, 57.3] 4. Karachi, Pakistan 3,108 [2,886, 3,330] 51.6 [48, 55.3] 5. Manila, Philippines 3,572 [3,261, 3,883] 46.2 [42.2, 50.3] 6. Kolkata, India 3,047 [2,728, 3,366] 45.3 [40.6, 50] 7. Lagos, Nigeria 3,227 [3,082, 3,372] 42.3 [40.4, 44.3] 8. Tokyo, Japan 1,040 [927, 1,153] 39.5 [35.2, 43.8] 9. Jakarta, Indonesia 3,772 [3,402, 4,142] 38.9 [35.1, 42.8] 10. Bangkok, Thailand 3,995 [3,746, 4,244] 37.0 [34.7, 39.3] 11. Chennai, India 3,727 [3,594, 3,860] 36.9 [35.5, 38.2] 12. Cairo, Egypt 1,941 [1,457, 2,425] 36.4 [27.4, 45.5] 13. Kinshasa, Dem. Rep. Congo 2,699 [2,615, 2,783] 31.3 [30.3, 32.2] 14. Shanghai, China 1,246 [988, 1,504] 29.6 [23.5, 35.7] 15. Rio de Janeiro, Brazil 2,241 [2,159, 2,323] 28.9 [27.9, 30] 16. Hyderabad, India 3,204 [3,146, 3,262] 28.7 [28.1, 29.2] 17. Guangzhou, China 2,062 [1,853, 2,271] 25.7 [23.1, 28.3] 18. Ahmadabad, India 3,478 [3,365, 3,591] 25.5 [24.7, 26.4] 19. ThànhPho Ho Ch´ıMinh, Vietnam 3,423 [3,284, 3,562] 25.0 [24, 26] 20. SãoPaulo, Brazil 1,171 [942, 1,400] 24.7 [19.8, 29.5] 21. Kuala Lumpur, Malaysia 3,492 [3,192, 3,792] 23.9 [21.8, 25.9] 22. Lahore, Pakistan 2,687 [2,294, 3,080] 23.5 [20.1, 26.9] 23. Al-Khartum, Sudan 4,569 [4,456, 4,682] 23.4 [22.9, 24] 24. Kinki M.M.A., Japan 1,139 [1,037, 1,241] 23.1 [21, 25.1] 25. Ar-Riyadh, Saudi Arabia 3,572 [3,507, 3,637] 22.8 [22.3, 23.2] 26. Bangalore, India 2,171 [1,682, 2,660] 21.9 [17, 26.8] 27. Shenzhen, China 1,996 [1,513, 2,479] 21.5 [16.3, 26.6] 28. Singapore, Singapore 3,720 [3,508, 3,932] 20.9 [19.7, 22.1] 29. Surat, India 3,306 [3,002, 3,610] 18.7 [17, 20.4] 30. Baghdad, Iraq 2,772 [2,643, 2,901] 18.4 [17.6, 19.3] 31. Beijing, China 807 [539, 1,075] 16.4 [11, 21.9] 32. Yangon, Myanmar 3,419 [3,178, 3,660] 16.4 [15.3, 17.6] 33. Chongqing, China 1,223 [990, 1,456] 16.3 [13.2, 19.4] 34. Jiddah, Saudi Arabia 3,982 [3,868, 4,096] 16.2 [15.8, 16.7] 35. Abidjan, Côted’Ivoire 3,190 [3,072, 3,308] 15.5 [14.9, 16.1] 36. Dongguan, China 2,036 [1,920, 2,152] 15.1 [14.3, 16] 37. Dar es Salaam, Tanzania 2,956 [2,904, 3,008] 15.1 [14.9, 15.4] 38. Hong Kong, Hong Kong 2,023 [1,986, 2,060] 14.8 [14.5, 15.1] 39. Miami, USA 2,502 [2,426, 2,578] 14.6 [14.1, 15] 40. Luanda, Angola 2,612 [2,530, 2,694] 14.4 [13.9, 14.8] 41. Foshan, China 2,035 [1,987, 2,083] 14.3 [14, 14.7] 42. Pune, India 2,358 [2,248, 2,468] 13.5 [12.9, 14.1] 43. Chittagong, Bangladesh 2,951 [2,803, 3,099] 13.4 [12.7, 14.1] 44. Tehran, Iran 1,566 [1,278, 1,854] 13.2 [10.8, 15.6] 45. Buenos Aires, Argentina 826 [673, 979] 12.5 [10.2, 14.9] 46. Fortaleza, Brazil 3,168 [3,106, 3,230] 12.3 [12.1, 12.5] 47. Kano, Nigeria 3,419 [3,343, 3,495] 12.3 [12, 12.5] 48. New York-Newark, USA 661 [484, 838] 12.3 [9, 15.6] 49. Istanbul, Turkey 805 [485, 1,125] 11.4 [6.9, 15.9] 50. Tianjin, China 991 [899, 1,083] 11.1 [10.1, 12.1]

Notes: This table ranks the top fifty cities worldwide by total CDD exposure. City populations are from61 and are for “urban agglomerations” as defined by the United Nations. Cooling degree days (CDD) and total CDD exposure are defined as in Table 2. 82

Table B.10: Total CDD exposure for Cities ranked 51 to 100

City Annual Cooling 95% Conﬁdence Total CDD 95% Conﬁdence Degree Days Interval Exposure Interval (CDDs) (in billions)

51. Recife, Brazil 2,952 [2,702, 3,202] 11.0 [10.1, 12] 52. Phoenix-Mesa, USA 2,684 [2,588, 2,780] 10.9 [10.5, 11.3] 53. Ouagadougou, Burkina Faso 3,847 [3,635, 4,059] 10.5 [10, 11.1] 54. Ibadan, Nigeria 3,335 [3,198, 3,472] 10.5 [10.1, 11] 55. Salvador, Brazil 2,906 [2,447, 3,365] 10.4 [8.8, 12.1] 56. Surabaya, Indonesia 3,604 [3,374, 3,834] 10.3 [9.6, 10.9] 57. Belo Horizonte, Brazil 1,789 [1,761, 1,817] 10.2 [10.1, 10.4] 58. Houston, USA 1,818 [1,758, 1,878] 10.2 [9.9, 10.6] 59. Jaipur, India 2,932 [2,900, 2,964] 10.1 [10, 10.3] 60. Al Kuwayt, Kuwait 3,627 [3,543, 3,711] 10.1 [9.8, 10.3] 61. Wuhan, China 1,261 [1,192, 1,330] 10.0 [9.4, 10.5] 62. Chukyo M.M.A. (Nagoya), Japan 1,033 [940, 1,126] 9.7 [8.8, 10.6] 63. Port-au-Prince, Haiti 3,954 [3,932, 3,976] 9.6 [9.6, 9.7] 64. Monterrey, Mexico 2,102 [2,092, 2,112] 9.5 [9.4, 9.5] 65. Caracas, Venezuela 3,247 [3,234, 3,260] 9.5 [9.4, 9.5] 66. Dubayy (Dubai), United Arab Emirates 3,908 [3,899, 3,917] 9.4 [9.4, 9.5] 67. Dallas-Fort Worth, USA 1,630 [1,597, 1,663] 9.3 [9.1, 9.5] 68. Douala, Cameroon 3,143 [3,045, 3,241] 9.2 [9, 9.5] 69. Santo Domingo, Dominican Republic 3,107 [3,003, 3,211] 9.2 [8.8, 9.5] 70. Faisalabad, Pakistan 2,572 [1,893, 3,251] 9.2 [6.8, 11.6] 71. Kozhikode, India 3,598 [2,449, 4,747] 8.9 [6.1, 11.8] 72. Lucknow, India 2,765 [2,003, 3,527] 8.9 [6.5, 11.4] 73. Nagpur, India 3,251 [2,678, 3,824] 8.7 [7.2, 10.2] 74. Bamako, Mali 3,451 [3,250, 3,652] 8.7 [8.2, 9.2] 75. Maracaibo, Venezuela 3,981 [3,436, 4,526] 8.7 [7.5, 9.9] 76. Makkah, Saudi Arabia 4,804 [4,600, 5,008] 8.5 [8.1, 8.9] 77. Kanpur, India 2,775 [2,206, 3,344] 8.4 [6.7, 10.1] 78. Kochi, India 3,473 [3,226, 3,720] 8.4 [7.8, 9] 79. H`aNoi, Vietnam 2,298 [1,951, 2,645] 8.3 [7.1, 9.6] 80. Nanjing, Jiangsu, China 1,107 [463, 1,751] 8.2 [3.4, 12.9] 81. Dakar, Senegal 2,334 [1,975, 2,693] 8.2 [7, 9.5] 82. Kumasi, Ghana 3,120 [2,949, 3,291] 8.1 [7.7, 8.6] 83. Abuja, Nigeria 3,310 [2,935, 3,685] 8.1 [7.2, 9] 84. Hangzhou, China 1,246 [588, 1,904] 8.0 [3.8, 12.2] 85. Coimbatore, India 3,130 [2,994, 3,266] 8.0 [7.6, 8.3] 86. San Juan, Puerto Rico 3,243 [3,001, 3,485] 8.0 [7.4, 8.6] 87. Thrissur, India 3,396 [2,840, 3,952] 7.9 [6.6, 9.2] 88. Xiamen, China 1,757 [1,412, 2,102] 7.8 [6.3, 9.3] 89. Lima, Peru 775 [616, 934] 7.7 [6.1, 9.2] 90. Los Angeles, USA 624 [456, 792] 7.7 [5.6, 9.7] 91. Zhongshan, China 2,060 [1,562, 2,558] 7.6 [5.8, 9.4] 92. Guayaquil, Ecuador 2,791 [2,619, 2,963] 7.6 [7.1, 8] 93. Accra, Ghana 3,319 [3,126, 3,512] 7.6 [7.1, 8] 94. Malappuram, India 3,424 [2,805, 4,043] 7.6 [6.2, 9] 95. Shantou, China 1,905 [1,729, 2,081] 7.5 [6.8, 8.2] 96. Kannur, India 3,477 [2,966, 3,988] 7.5 [6.4, 8.6] 97. Port Harcourt, Nigeria 3,177 [2,920, 3,434] 7.4 [6.8, 8] 98. Seoul, Republic of Korea 755 [283, 1,227] 7.4 [2.8, 12] 99. Medan, Indonesia 3,332 [3,051, 3,613] 7.3 [6.7, 8] 100. Chengdu, China 926 [381, 1,471] 7.0 [2.9, 11.1]

Notes: Cooling degree days (CDD) and total CDD exposure are deﬁned as in Table 2. 83

Table B.11: Total CDD exposure for Cities ranked 101 to 150

City Annual Cooling 95% Conﬁdence Total CDD 95% Conﬁdence Degree Days Interval Exposure Interval (CDDs) (in billions)

101. Barranquilla, Colombia 3,520 [3,135, 3,905] 7.0 [6.2, 7.8] 102. Alexandria, Egypt 1,472 [1,031, 1,913] 7.0 [4.9, 9.1] 103. Samut Prakan, Thailand 3,875 [3,711, 4,039] 7.0 [6.7, 7.3] 104. Belém,Brazil 3,103 [2,971, 3,235] 6.8 [6.5, 7.1] 105. Phnum Pénh,Cambodia 3,946 [3,630, 4,262] 6.8 [6.3, 7.4] 106. Thiruvananthapuram, India 3,457 [2,984, 3,930] 6.8 [5.9, 7.7] 107. Visakhapatnam, India 3,540 [3,238, 3,842] 6.8 [6.3, 7.4] 108. Bandung, Indonesia 2,689 [2,415, 2,963] 6.8 [6.1, 7.5] 109. Patna, India 2,991 [2,453, 3,529] 6.6 [5.4, 7.8] 110. Manaus, Brazil 3,190 [2,798, 3,582] 6.5 [5.7, 7.3] 111. Conakry, Guinea 3,340 [3,235, 3,445] 6.5 [6.3, 6.7] 112. Indore, India 2,624 [2,306, 2,942] 6.4 [5.6, 7.2] 113. Vadodara, India 3,220 [3,049, 3,391] 6.4 [6, 6.7] 114. Suzhou, Jiangsu, China 1,152 [804, 1,500] 6.3 [4.4, 8.2] 115. Vijayawada, India 3,560 [3,469, 3,651] 6.3 [6.1, 6.4] 116. Yaoundé,Cameroon 2,033 [1,739, 2,327] 6.2 [5.3, 7.1] 117. Nanning, China 1,884 [1,590, 2,178] 6.1 [5.1, 7] 118. Madurai, India 3,743 [3,394, 4,092] 6.0 [5.4, 6.5] 119. Multan, Pakistan 3,114 [2,746, 3,482] 6.0 [5.3, 6.7] 120. Semarang, Indonesia 3,640 [3,507, 3,773] 5.9 [5.7, 6.1] 121. Hyderabad, Pakistan 3,335 [3,055, 3,615] 5.9 [5.4, 6.4] 122. Ciudad de Panamá,Panama 3,531 [3,401, 3,661] 5.9 [5.7, 6.1] 123. Davao City, Philippines 3,634 [3,446, 3,822] 5.9 [5.6, 6.2] 124. Kitakyushu-Fukuoka M.M.A., Japan 1,047 [795, 1,299] 5.8 [4.4, 7.2] 125. Muqdisho, Somalia 2,709 [2,557, 2,861] 5.8 [5.5, 6.1] 126. Agra, India 2,875 [2,497, 3,253] 5.7 [4.9, 6.4] 127. Bhopal, India 2,700 [2,457, 2,943] 5.7 [5.2, 6.2] 128. Valencia, Venezuela 3,263 [3,010, 3,516] 5.7 [5.2, 6.1] 129. Bras´ılia,Brazil 1,344 [700, 1,988] 5.6 [2.9, 8.3] 130. Tel Aviv-Yafo, Israel 1,542 [1,270, 1,814] 5.6 [4.6, 6.5] 131. Taipei, China 2,066 [1,856, 2,276] 5.5 [4.9, 6.1] 132. Xi’an, China 903 [682, 1,124] 5.5 [4.1, 6.8] 133. Madrid, Spain 891 [825, 957] 5.5 [5.1, 5.9] 134. Santa Cruz, Bolivia 2,544 [2,356, 2,732] 5.4 [5, 5.8] 135. Rajkot, India 3,330 [3,243, 3,417] 5.3 [5.2, 5.5] 136. Halab, Syria 1,493 [707, 2,279] 5.3 [2.5, 8.1] 137. Atlanta, USA 1,025 [527, 1,523] 5.3 [2.7, 7.8] 138. Tampa-St. Petersburg, USA 2,012 [1,655, 2,369] 5.3 [4.4, 6.3] 139. Cali, Colombia 1,972 [1,648, 2,296] 5.2 [4.4, 6.1] 140. La Habana, Cuba 2,420 [2,126, 2,714] 5.2 [4.5, 5.8] 141. Al-Madinah, Saudi Arabia 4,070 [3,834, 4,306] 5.2 [4.9, 5.5] 142. Changsha, China 1,369 [1,033, 1,705] 5.1 [3.9, 6.4] 143. Medell´ın,Colombia 1,303 [413, 2,193] 5.1 [1.6, 8.6] 144. Brazzaville, Congo 2,699 [2,683, 2,715] 5.1 [5.1, 5.1] 145. Makassar, Indonesia 3,445 [3,414, 3,476] 5.1 [5.1, 5.2] 146. Rawalpindi, Pakistan 2,050 [2,021, 2,079] 5.1 [5.1, 5.2] 147. N’Djaména,Chad 3,944 [3,897, 3,991] 5.0 [4.9, 5] 148. Palembang, Indonesia 3,405 [3,316, 3,494] 5.0 [4.8, 5.1] 149. Benin City, Nigeria 3,313 [3,270, 3,356] 5.0 [4.9, 5] 150. Gujranwala, Pakistan 2,350 [2,299, 2,401] 5.0 [4.9, 5.1]

Notes: Cooling degree days (CDD) and total CDD exposure are deﬁned as in Table 2. 84 C ◦ C threshold) ◦ C threshold) (23.9 ◦ C threshold) (21.1 ◦ (15.6 C threshold Annual CDDs CDDs count Annual CDDs CDDs count Annual CDDs CDDs count ◦ Country Ranking using Population-Weightedthe 18.3 Change in Population-Weighted Change in Population-Weighted Change in 1. India2. China3. Indonesia4. Nigeria5. Pakistan6. Brazil7. Bangladesh8. Philippines9. United10. States Vietnam11. Thailand12. Egypt13. Myanmar14. Dem. Rep.15. Congo Mexico16. Ethiopia17. 3,822 4,294 Sudan 1,54318. Iran19. Japan 4,444 3,266 +1,01020. Saudi 3,653 +974 Arabia21. +534 Venezuela 4,290 1,345 3,055 +1,01522. Malaysia +762 +1,00923. Iraq24. +1,024 3,746 Ghana 3,124 4,49625. +478 +947 Colombia26. Niger 4,12227. 2,273 Cˆoted’Ivoire 2,761 +1,009 +969 +1,025 2,03528. Turkey29. 2,426 +1,040 607 Burkina Faso 1,99830. 2,432 -1,011 1,829 1,827 Cameroon +827 2,272 -813 -1,003 4,733 4,298 1,254 -402 +769 492 +973 -675 -817 2,108 1,381 4,384 1,128 1,892 2,484 -994 +987 +951 4,385 -854 2,154 -375 +1,008 1,271 +593 +495 1,317 -987 +1,010 -885 -987 2,777 1,257 4,167 4,336 3,347 1,440 -928 -2,013 299 681 4,828 1,234 1,071 +1,011 5,096 662 +1,010 -617 -1,591 +898 -1,989 1,274 +689 1,397 2,733 2,556 4,035 +1,011 +1,066 2,369 -1,270 575 -1,573 1,079 -710 -548 230 -797 -1,992 519 2,363 -1,013 1,060 1,511 +1,013 +480 361 -791 -1,007 -1,533 1,254 -436 2,144 2,314 -1,012 -637 -1,717 -1,960 1,073 -1,754 -367 2,085 779 -1,828 2,807 3,085 -1,012 -1,012 -1,155 331 1,812 -806 2,013 232 -573 1,824 560 -1,010 1,368 -945 1,354 -1,934 -1,009 -1,227 683 -898 -1,523 253 -2,008 -357 1,135 1,304 -2,021 -832 1,560 484 1,806 -633 -2,021 -2,022 2,128 1,051 -1,098 -1,395 -2,011 -1,902 283 -1,971 -634 Notes: Changes in the counts of Population-Weighted CDDs are all expressed relative to the counts computed using the 18.3 Table B.12: ChangesAlternative CDD in Thresholds the Country-level counts of Population-Weighted Cooling Degree Days under the threshold. 85

Table B.13: Changes in the Country-level ranking of total CDD exposure under the Alternative CDD Thresholds

Ranking using Rank change Ranking using Rank change Ranking using Rank change the 15.6◦C the 21.1◦C the 23.9◦C threshold threshold threshold

1. India 0 India 0 India 0 2. China 0 China 0 China 0 3. Indonesia 0 Indonesia 0 Indonesia 0 4. Nigeria 0 Nigeria 0 Nigeria 0 5. Brazil +1 Pakistan 0 Pakistan 0 6. Pakistan -1 Bangladesh +1 Bangladesh +1 7. Bangladesh 0 Brazil -1 Philippines +1 8. Philippines 0 Philippines 0 Brazil -2 9. United States 0 Vietnam +1 Thailand +2 10. Vietnam 0 Thailand +1 Vietnam 0 11. Thailand 0 United States -2 United States -2 12. Egypt 0 Egypt 0 Egypt 0 13. Mexico +2 Myanmar 0 Sudan +4 14. Ethiopia +2 Sudan +3 Myanmar -1 15. Dem. Rep. Congo -1 Dem. Rep. Congo -1 Saudi Arabia +5 16. Myanmar -3 Mexico -1 Iraq +7 17. Sudan 0 Iran +1 Iran +1 18. Japan +1 Saudi Arabia +2 Venezuela +3 19. Iran -1 Iraq +4 Niger +7 20. Venezuela +1 Venezuela +1 Mexico -5 21. Saudi Arabia -1 Malaysia +1 Malaysia +1 22. Malaysia 0 Japan -3 Ghana +2 23. Colombia +2 Ethiopia -7 Burkina Faso +6 24. Iraq -1 Ghana 0 Japan -5 25. Ghana -1 Niger +1 Mali +6 26. Turkey +2 Colombia -1 Chad +12 27. Uganda +5 Burkina Faso +2 Dem. Rep. Congo -13 28. Niger -2 Côted’Ivoire -1 Côted’Ivoire -1 29. Côted’Ivoire -2 Mali +2 Cambodia +8 30. Kenya +11 Cameroon 0 Cameroon 0

Notes: Changes in ranking are all expressed relative to the ranking derived from using the 18.3◦C threshold. 86 Table B.14: Breakdown of fishing effort in each of the 8 VLMPAs PrevalenceRank Before announcement Fishing type Between Percentage announcement total and implementation Fishing type After implementation Percentage total Fishing Type Percentage total The three most prevalent fishing methods in each VLMPA during each of the three periods characterizing the life cycle of an MPA: before Notes: MPA name Ross SeaProtectedAreaPapah¯anaumoku¯akea1.Marine NationalMonument 1. DriftingPacific 2. longlines Remote 69.6 2.Islands MarineNational Monument 3.Pitcairn 1. 3. 3. SetIslands Drifting longlines 2. longlines Pole 2.3 andMarine line Reserve DriftingTerres longlines 97.2 Pole Drifting andAustrales 27.7 Tuna longlines line purse Unidentified 91.2 seines0.4 3.Fran¸caises Other fishing 1.Phoenix Islands 2.4 2.7Protected 2. 4.9 Drifting longlinesArea 91.3 1. /Nazca-Desventuradas1. 2. 3. Tuna 1. Drifting purse Set Drifting longlinesMarine seines longlines 100 longlinesPark /Revillagigedo 2. Drifting Drifting longlinesNational 6.8 Trawling 97.2 longlines Passenger Pole Tuna Drifting 99.2 Set purse and longlines Other longlines seines linePark fishing / 8.6 75.1 3. 1. Drifting Tuna 72.6 / 93.2 purse longlines 2. 98.3 seines2.7 0.7 26.6 3. Drifting / 2. Drifting longlines 15.2 Drifting longlines longlines Pole Tuna and purse line seines59.3 9.7 Trawling 0.2 Drifting longlines 99.9 3. 27 0.1 / Set Tuna Pole Set Trolling longlines purse and longlines seines line 92.3 / Tuna 0.8 purse 86.1 99.6 Trawling seines / 40.7 Tuna purse seines / Drifting Pole / longlines and 73 line Trawling 5.9 Research 13.9 1 Tuna 99.9 purse Drifting / seines 100 longlines 82.3 0.9 Squid 0.1 / jigger Pole and line 0.6 99.5 / 62 / Drifting longlines 0.1 Set longlines / / Tuna purse seines 38 0.3 Trawling 17.7 passenger 99.9 96.6 / / / / Trawling 2.1 Pole Squid and jigger / line 0.1 1 0.1 0.5 / / / / / / / / / announcement, after announcement & before implentation and after implementation.Estimation of the type of fishing effort is provided by GFW. 87 -order polynomial th -order polynomial 5 th -order polynomial 4 LB Est. UB LB Est. UB LB Est. UB rd 3 A positive (negative) number means that annualized banned fishing effort exceeds (subceeds) annualized VLMPA name Ross Marine Sea National Protected MonumentPapah¯anaumoku¯akea AreaPacific -4.0 Remote Islands MarinePitcairn National Islands -2.1 Monument MarineTerres Reserve -10.1 Australes Fran¸caisesPhoenix -1.2 -0.2 Islands ProtectedNazca-Desventuradas Area Marine Park -4.1 7.7Revillagigedo National ParkPooled -2.1 VLMPAs -3.4 -0.1 9 -4.1 -2.1 21.3 -10.3 -5.8 -29.8 -6.2 0.0 -2.6 -2.9 -8.7 -400.3 -332.7 -13.4 -265.0 -2.1 -1.6 -6.2 0.6 -407.9 12.3 -332.7 -257.4 -10.6 -0.2 -414.1 0.9 -332.7 -5.4 -6.1 -14.7 -251.3 -3.1 -7.5 -2.5 -14.3 -1.6 -1.5 -6.2 -0.3 0.4 -10.7 -6.1 0.2 1.8 -12.6 -5.6 -7.1 -117.1 -15.2 -1.4 -3.5 -2.5 -86.6 -6.3 -1.6 -1.6 -56.0 0.6 -120.1 -12.0 2.6 -86.9 0.3 -7.2 -53.7 -123.5 -2.4 -87.0 -50.6 Notes: Table B.15: Areas underalized three fishing distinct hours order polynomial curves estimating the difference in annu- fishing effort during theacross period all preceding specifications thewitnesses announcement for a of all reversal a of theeven. possible sign VLMPAs VLMPA for creation. but the main the Results area remain Pacific estimate stable Remote depending Islands on whether Marine the National order Monument, of the which polynomial is odd or 88 ) 2 Country/Entity Date Date of no-take area Table B.16: Characteristics of the eight VLMPAs under consideration The Announcement date corresponds to the earliest evidence we could find that an area would eventually be designated to VLMPA nameRoss Marine Sea National Protected MonumentPapah¯anaumoku¯akea Area UnitedPacific States Remote Islands MarinePitcairn National Islands Monument Marine UnitedTerres Reserve States Australes Fran¸caisesPhoenix Islands ProtectedNazca-Desventuradas Area Marine ParkRevillagigedo National Park January 29, 2015 August 26, 2016 May 20, 2014 September United 25, CCAMLR Kingdom 2014 (International) October 27, Chile 2016 Managing 1.15 December Kiribati France 1, 2017 1.06 March 18, 2015 Mexico September 12, 1.12 2016 Announcement May Implementation 9, February June 0.83 2014 23, 14-15, 2013 2016 August 24, March 2016 Size 31, (millions 2017 km January July 1, 17, 2015 2016 November 24, 0.30 2017 0.12 0.41 0.15 Notes: become a marine protectedcorresponds area to (see the the dayVLMPA. timelines The when leading VLMPAs the to can the fishing also creation ban be of becomes comprised each of effective. VLMPA areas in The where the Size some main column type text). of indicates The fishing the Implementation effort size date is of allowed. the “no-take” area withing each