DISCUSSION PAPER Ma y 2015; revised April 2016 RFF DP 15-19-REV Weather, Traffic Accidents, and Exposure to Climate Change Benjamin Leard and Kevin Roth 1616 P St. NW Washington, DC 20036 202-328-5000 www.rff.org Weather, Traffic Accidents, and Exposure to Climate Change Benjamin Leard and Kevin Roth∗ April 8, 2016 Abstract Quantifying the costs of climate change requires measurement of direct effects as well as behavioral responses. While behavioral responses have been shown to increase costs, we identify responses that reduce costs, which we define as “voluntary exposure.” We quantify the response of the transportation sector in terms of traffic accidents and travel demand to daily variation in weather. We find warmer temperatures and reduced snowfall are associated with a significant increase in fatal accidents. We find, however, that almost all of the estimated effect of temperature on fatalities is due to changes in exposure for pedestrians, bicyclists, and motorcyclists. While the application of these results to middle-of-the-road climate predictions suggests that weather patterns for the end of the century would lead to 397 additional fatalities per year, the associated welfare losses are almost completely offset by voluntary exposure benefits from increased travel by walking, biking, and motorcycling. To facilitate more accurate estimates of climate change impacts, we also introduce the empirical cumulative distribution function method to the economics literature for correcting baseline weather errors from climate simulations. Key Words: Traffic Accidents, Exposure, Climate Change JEL Classification Numbers: Q58, Q52, H23, R41 ∗ Leard: Resources for the Future, 1616 P St. NW, Washington, D.C. 20036, e-mail: leard@rff.org. Roth: University of California, Irvine 3297 Social Science Plaza, Irvine, CA 92697, e-mail: kroth1@uci.edu. The authors thank Edson Severnini and Sol Hsiang for useful suggestions and are grateful to seminar participants at 2016 AEA AERE session, UC-Irvine, and USC for helpful comments. 1 1 Introduction Understanding the channels by which climate change will affect the economy is required to estimate the costs of climate change and to develop adaptation strategies. Extreme heat may result in direct costs such as increased mortality from heat stress and lower productivity. Economists have also made clear that individuals are likely to engage in costly defensive activities to mitigate these outcomes (Harrington and Portney, 1987). For example, where heat aggravates respiratory illness, individuals may use medication and air conditioning to reduce exposure to risk (Deschênes, Greenstone and Shapiro, 2013; Graff Zivin and Neidell, 2014). In such cases, measuring direct costs is a lower bound of the true welfare effects because it omits these defensive expenditures. Less attention has been paid to situations where climate change generates costs that are, at least partially, explained by voluntary exposure to risk. Warmer weather facilitates spending time outdoors. Choosing to spend time outdoors provides many benefits, but it also exposes individuals to potential risks from air pollution, crime, UV radiation, and traffic. Where costs from these risks occur as part of a utility maximizing decision to spend time outside, these costs must be offset with the welfare gain from spending time outside. But the costs of time spent outside in terms of illness, injury, or death are often much easier to measure than the benefits. To understand the welfare consequences of climate change, it is important to examine exposure to risk. Where climate change aggravates an existing risk, individuals will engage in costly behavior to avoid exposure to that risk, but when climate change increases the utility of an activity with an associated risk, individuals will voluntarily increase exposure to that risk. We define the resulting welfare effects of the latter case as voluntary exposure benefits. Under the extreme assumption where an entire observed change in risk is due to voluntary exposure, the net welfare costs of climate change would be zero. Whether or not an observed increase in risk is accompanied by additional costs from defensive expenditure or offsetting benefits can be ascertained by examining exposure. Where individuals avoid risk, there are additional costs; where individuals engage in more risky activity, there are offsetting welfare benefits. Problematically, in many empirical settings there is almost no information on exposure to determine which effect generates the direct cost. In this paper we analyze these issues in the context of the transportation sector by estimating the relationship among weather, traffic accidents, and travel demand. We choose to focus on the transportation sector because even small changes to traffic fatalities are likely to carry large costs. Worldwide, nearly 1.24 million people die in traffic accidents annually (WHO, 2013).1 To quantify the effects of climate change on traffic accidents, we examine the effect of weather on accidents and generate future outcomes with climate change using simulated future weather. By exploiting plausibly random daily variation in temperature, rainfall, and snowfall, we are able to estimate the effect of weather on transportation outcomes. We find a large and statistically significant relationship between weather and traffic fatalities. Unsurprisingly, we find precipitation plays a role in fatal car crashes. But we also 1 In 2006, traffic fatalities were the leading cause of death in the United States for individuals between the ages of 4 and 35 (Subramanian, 2009). 2 find large effects for temperature. As temperatures rise, an increasing number of fatalities involve pedestrians, bicycles, and motorcycles, or ultra-light duty (ULD) accidents. Warmer weather is intuitively linked to more exposure as people will derive utility from spending leisure time outside or from traveling by bicycle or motorcycle for transportation or leisure purposes. We are also able to measure exposure directly by estimating the demand for ULD trips as a function of weather. To estimate these effects, we use detailed data on police-reported accidents, daily travel logs of households, weather, and climate change prediction data. We find that for a day with temperature above 80◦ F there is a 9.5 percent increase in fatality rates compared with a day at 50-60◦ F. Our estimates indicate that the majority, but not all, of this effect is due to fatalities involving individuals traveling by ULD modes. We also find evidence that individuals avoid ULD travel on cold days, while estimates for demand on hot days are positive but imprecise. We cannot rule out the possibility that reduced cognitive ability brought on by heat is responsible for some of the estimated effect particularly for light-duty travel on hot days, and we find some evidence on the hottest days that drivers avoid trips by light-duty vehicles, consistent with defensive behavior. We estimate that the discounted costs of additional traffic fatalities caused by climate change are $40 billion from 2015 to 2099, an amount on the same order of magnitude as others such as profit changes to the agriculture industry (Deschênes and Greenstone, 2007; Fisher et al., 2012) and social welfare effects from criminal activity (Ranson, 2014). But we find that welfare gains from increased ULD travel reduces these costs by at least $34 billion. Under the extreme assumption that the entire observed change in traffic fatalities is due to voluntary ULD exposure, the marginal costs of climate change stemming from the increase in traffic fatalities are less than $9 billion and not statistically distinguishable from zero. Including reduced injuries and property damage only (PDO) accidents suggests that net welfare effects may even be positive. In sum, we find that omitting voluntary exposure behavior from welfare analysis may lead to a significant overestimate of climate change costs. Methodologically, we also introduce a new technique for using predicted weather from climate models when simulating future outcomes derived from an approach used in the climate science literature. It is well known that climate prediction models can forecast baseline weather that is different from the weather actually observed (Auffhammer et al., 2013). Various methods to correct baselines have been used in prior papers, but none are able to align the full distribution of baseline outcomes with observed weather, while also allowing the variance of events to change. This is particularly important for studies that seek to understand how economic outcomes respond to a change in climate variability. We show that prior methods may yield unintuitive results. Furthermore, these methods are often not well suited to predict precipitation changes, which are important in our setting. We make use of a quantile-mapping method inspired by techniques used in the climate science literature. This technique allows us to avoid the baseline climate modeling error that can generate spurious results in areas with complex geography, while also applying the full distributional changes predicted by climate models. The rest of the paper proceeds as follows. Section 2 reviews the economic, transportation, and medical literature relevant to our study. Section 3 develops a framework for estimating how much households are willing to pay to avoid climate change. Section 4 describes the data and Section 5 develops the econometric framework for our outcomes and the results of 3 our estimation. Section 6 details our simulation of future weather outcomes and Section 7 concludes. 2 Literature Review Several papers from the traffic safety literature explore the relationships between weather outcomes and traffic accidents. This literature has primarily estimated the impact of rainfall or snowfall on accident frequency (Eisenberg, 2004; Eisenberg and Warner, 2005), but does not examine the relationship with temperature or the impact of climate change. An exception is Froom et al. (1993), who show that high temperatures are associated with higher rates of helicopter pilot errors. In the medical literature there is extensive research that suggests plausible physiological reasons that accidents and fatalities may be affected by temperature. While prior work has documented effects of temperature on ability to perform a task, the relationship is complex (Hancock and Vasmatzidis, 2003).2 In a laboratory setting, Broadbent (1963) show that heat in excess of 85◦ F did not affect the speed of task completion but did result in a higher error rate. Heat has been found to primarily affect vigilance, tracking, and multitasking but simple perception and reaction time remain unaffected (Grether, 1973). While air conditioning may mitigate these effects, many effects from heat are due to prolonged exposure that increases core temperature (Hancock and Vasmatzidis, 2003). The economics literature has studied the impact of climate change using temperature and occasionally precipitation data in a variety of areas such as agriculture (Deschênes and Greenstone, 2007; Fisher et al., 2012), economic growth (Dell, Jones and Olken, 2014), and health (Neidell, 2004). Graff Zivin and Neidell (2014) examine how weather influences time allocation showing that warmer weather draws individuals outdoors, possibly exposing them to extreme pollution events (Bäck et al., 2013). Graff Zivin, Hsiang and Neidell (2015) find that heat is associated with more errors on cognitive assessment exams but these transitory shocks do not affect long-run human capital formation. Our study is also related to work on crime and conflict (Jacob, Lefgren and Moretti, 2007; Hsiang, Burke and Miguel, 2013; Ranson, 2014) that has found a strong relationship between heat and violent activity. Deschênes and Moretti (2009) use mortality data from the Multiple Causes of Death files and include regressions for motor vehicle deaths. They find some evidence for a decrease in fatalities from cold weather, which they conjecture is due to avoided travel. Finally, this work contributes to a literature that attempts to incorporate behavior into welfare calculations when health and safety risks change. Harrington and Portney (1987) and Graff Zivin and Neidell (2013) emphasize that defensive behavioral adjustments to an environmental risk can be crucial for providing a comprehensive welfare assessment, and prior literature has attempted to empirically estimate these expenditures in environmental settings (e.g., Deschênes and Greenstone (2011), Deschênes, Greenstone and Shapiro (2013) and Barreca et al. (2013)).3 We expand this literature by identifying and quantifying a new 2 This is because many factors influence the effect of heat stress on cognitive performance, such as task type, exposure duration, skill and acclimatization level of the individual. 3 See Auffhammer and Mansur (2014) for a review of the economics literature on how climate change affects energy expenditures. 4 direction of individual response to climate change. Our research contributes to this literature by showing that exposure is an important consideration in the interpretation of these costs and shows that excluding voluntary exposure behavior can overstate welfare costs. 3 Framework for Climate Change Estimating Welfare Effects of In this section, we present a simple analytical framework for characterizing the welfare effects of climate change on traffic accidents building on Harrington and Portney (1987). Subsection 3.1 presents the model assumptions and subsection 3.2 includes a derived welfare formula that we use for estimating the welfare effects of climate change on traffic accidents. 3.1 Assumptions Suppose that a representative household is endowed with a light-duty vehicle (LDV) and decides how to spend income to maximize utility. The household decides how many miles to travel by LDV and by walking, biking, or motorcycling, or any other ultra-light duty (ULD) mode. We examine ULD travel separately from automobile travel because the utility of traveling by this mode is likely to be strongly influenced by weather and is subject to a higher fatality rate per collision, all other crash characteristics equal. The household chooses a number of LDV miles, denoted by m, ULD miles, denoted by b, and consumption of a composite good x. It faces a fixed LDV operating cost per mile denoted by pm , a fixed permile cost of ULD travel denoted by pb , and a price of the composite good that we normalize to one. The household then maximizes utility subject to a budget constraint with endowed income I. Utility is a function of LDV and ULD miles, consumption of the composite good, and the probability that traveling causes a traffic accident. The likelihood of a traffic accident is a function of the household’s LDV miles, ULD miles, and weather conditions, denote by W . We characterize weather conditions with measured temperature, rainfall, and snowfall. We assume that the probability of an accident, a = am (m, W ) + ab (b, W ), is additively separable in the probability of getting into an accident while driving (represented by the function am (·)) and the probability of getting into an accident while walking/bicycling/riding a motorcycle (represented by the function ab (·)). The probability of an accident is assumed to increase in LDV or ULD miles, so that ∂a ∂a > 0 and ∂a > 0. We leave the sign of ∂W for our empirical analysis. ∂m ∂b The household makes choices based on the following problem: V = max {U (m, b, x, a)} subject to (1) a = am (m, W ) + ab (b, W ), (2) pm m + pb b + x = I. (3) m,b,x ∂U The household’s utility function, U (·), is assumed to satisfy ∂m > 0, ∂U > 0, ∂U > 0, ∂U > 0, ∂b ∂x ∂l ∂U < 0, and sufficient second-order derivative properties to ensure an interior solution. ∂a 5 3.2 Expressions for WTP/WTA for a Change in Weather For a change in a weather variable, the household must be compensated with income I for changes in W to hold utility constant. This implies that we can define income as a function of W : I = I ∗ (W ). (4) The function I ∗ (W ) keeps utility constant in response to change in a weather outcome. Holding fixed all other exogenous parameters, we can express indirect utility as V = V (I ∗ (W ), W ). (5) Differentiating Equation 5 with respect to W yields ∂V dI ∗ ∂V dV = ∗ + = 0. dW ∂I dW ∂W (6) Rearranging terms yields a general expression for WTP/WTA for a change in weather conditions: ∂V dI ∗ = − ∂W . (7) ∂V dW ∗ ∂I In Appendix A, we show that Equation (7) can be expressed as ∂U ∂U ∂U dI ∗ dam ∂U dab ∂am ∂ab = − ∂a + ∂a mεm + ∂a bεb . − ∂a dW dW} |λ {z dW} |λ ∂m ∂b } |λ {z {z } |λ {z a W T Pm W T Pba WTPm (8) WTPb The first and second terms, W T Pma and W T Pba , are the welfare effect from a change in the probability of an LDV accident and ULD accident, respectively, resulting from a change in weather. Each term is the product of the monetary value from a change in the probability of an accident and the effect of a change in a weather variable on the probability of an accident.4 The third term, W T P m , is the welfare effect from a change in miles driven induced by a change in weather. The fourth term, W T P b , is the welfare effect from a change in ULD miles traveled induced by a change in weather. Each term is the product of the price of travel, in terms of accidents per mile, and the change in travel induced by weather. The change in ∂m , and the automobile miles from weather is given by the LDV trip semi-elasticity, εm = m1 ∂W 1 ∂b change in ULD miles from weather is given by εb = b ∂W . The third and fourth terms are central to our analysis and measure the welfare effect of households reacting to climate change in potentially two distinct ways. If households respond in a defensive manner by reducing exposure to negative outcomes, such as reduced driving to avoid accidents when conditions are rainy, these terms will represent costly defensive activity (Harrington and Portney, 1987). This behavior is calculated as a welfare cost; studies omitting this component underestimate the welfare costs of climate change. The second potential reaction is that households may expose themselves more to negative outcomes 4 An increase in the fatality rate from more snow, for example, lowers household welfare, implying that m the household is willing to pay to avoid this weather change. This term will be positive when da dW > 0 because ∂U ∂a < 0 and λ > 0. 6 if the private benefits of doing so exceed the costs. We define this behavior as voluntary exposure. Where these terms are positive, voluntary exposure is calculated as a welfare benefit and omitting this component will overestimate the welfare costs of climate change. As an example, assume warmer weather increases fatal motorcycle accidents. If this ∂b < 0 and W T P b increase is accompanied by households avoiding motorcycle travel, then ∂W 5 will be a defensive expenditure cost. If, however, warmer weather encourages voluntary ∂b exposure so that ∂W > 0, then W T P b will be positive. Such an outcome occurs if individuals engage in more motorcycle driving in warmer weather leading to more motorcycle accidents. At an extreme, if the entire change in accidents is due to voluntary exposure, the envelope dI ∗ theorem implies that W T Pba = W T P b and W T Pma = W T P m , and dW in Equation (8) is 6 equal to zero. In our empirical analysis we will mostly focus on fatal accidents but we will also consider PDO accidents and injuries. For fatal crashes we also examine two subcases where all LDV fatalities are due to exposure or where all ULD fatalities are due to exposure by setting the welfare effect of each group to zero to give a sense of the relative contribution of each mode to the final welfare cost. 3.3 Implications for Measuring Welfare Effects of Climate Change Equation (8) is valuable because empirical estimates for all terms are available from our own estimates, the literature, or existing datasets for fatalities. In our empirical model, we m , the marginal effect of the change of a weather variable estimate the following terms: da dW dab , the marginal effect of the change of a weather on the probability of an LDV accident; dW variable on the probability of a ULD accident; εm , the LDV mile semi-elasticity with respect to a change in weather; and εb , the ULD mile semi-elasticity with respect to a change in weather. Estimates for the remaining terms in Equation (8) are available from our data or /λ, the monetary value from the change in the additional sources. These include ∂U ∂a ∂a probability of a accident, which is the value of a statistical life (VSL) for fatal accidents; ∂m and ∂a , the marginal effect of a change in miles on the probability of an LDV fatality or ∂b ULD fatality respectively, which can be approximated by the per mile fatality rate; m, the number of LDV miles; and b, the number of ULD miles.7 It is important to note that a change in the number of miles is not the only behavioral adjustment drivers may make. Warmer weather may allow other channels of voluntary exposure, such as driving faster, or more scenic but dangerous routes. Although we examine some of these voluntary exposure channels in the appendix, such as speed, there will be 5 At the optimum, the household is equating the monetary value of the marginal disutility from the increased likelihood of an accident stemming from driving one more mile to the marginal value of driving that is unrelated to accidents. 6 A similar assumption underlies the calculation of the value of a statistical life where individuals consume risk that is compensated with benefit. Our framework also assumes they understand and properly value these risks when engaging in travel. 7 For PDO accidents and injuries we do not have some elements of this calculation as national level statistics are not collected. Although we believe accidents and injuries are consistently collected within a state-year in our SDS data, they vary across years and states, creating ambiguity on which injury or accident rate is most appropriate. 7 some adjustments we cannot measure.8 Similarly, we do not observe defensive expenditures such as using air conditioning or purchasing safety helmets, implying that the cost of climate change is higher than what we report.9 Equation (8) also does not include long-run margins of adjustment that households or governments may adopt in response to changes in climate.10 We think there are three key types of long-run adjustment that could occur. First, there may be a private adaptation to local conditions. Such adaptation includes purchasing a different vehicle that is safer to drive or making changes in behavior to adapt to local conditions. To some degree we are able to empirically test for this type of adaptation by separately estimating our results across regions, which we do in the appendix.11 Second, migration may occur if individuals, for example, move farther north in response to climate change. This is another type of private adaptation but one that we cannot examine with our data. Both of these forms of private adaptation imply that the true costs of climate change are lower than what we report. The third type is public adaptation through infrastructure or institutional changes that attempt to address these costs. Estimating these changes is beyond the scope of our paper, but our welfare calculation serves as a component in determining the amount of spending warranted in a public response to traffic fatalities from climate change. It is also important to note that if the primary mechanism for our observed effects is voluntary exposure, we may see private or public adaptation but not migration. If warmer weather increases the amount of biking, helmet sales may increase and bike lanes may pass a cost-benefit test whe they would not have earlier, but we would not expect people to flee northward. 4 Data and Summary Statistics 4.1 4.1.1 Data Sources Accident, Injury, and Fatality Data We obtain the population of police-reported accidents for 20 states from the State Data System (SDS) maintained by the National Highway Traffic Safety Administration (NHTSA). These data are collected and used by the NHTSA to provide analysis and policy recommendations for U.S. DOT. The benefit of these data is that they include not only fatalities, as recorded in other sources, but also non-fatal accidents. 8 See Appendix Table H17 for speed regressions. Another possible welfare consideration in our context is the external cost of traffic accidents. Although we do not explicitly model external costs of traffic accidents, our empirical estimate of the direct effect of a climate change on traffic accidents (W T Pm and W T Pba in Equation (8)) incorporates the entire social cost of the change in accidents. 10 This equation also omits welfare effects from the marginal impact of climate change on travel quality. For example, taking a trip to the beach on a warm and sunny day is more valuable than taking the same trip on a cold day, even if the accident probability is the same. To model the value of a trip taken by the household that is independent of the accident probability as a function of weather conditions, we could incorporate comfort into the choice problem by defining quality-adjusted miles traveled functions for LDV and ULD travel, qm = q(m, W ) and qb = q(b, W ), that enter directly in the household’s utility function. Poor weather conditions (e.g., heavy rain or extreme temperature) lowers quality-adjusted miles traveled. Leaving this component out of our empirical evaluation of welfare will understate the welfare effects of climate change. 11 See Appendix Table H10. 9 8 Accident reports, completed by police officers, are administered at the state level. These files are requested annually by the NHTSA from the state agencies that computerize the data, which are then formatted for consistency and compiled into the SDS.12 We obtained permission to use SDS data from Arkansas, California, Florida, Georgia, Iowa, Illinois, Kansas, Michigan, Minnesota, Missouri, Montana, Nebraska, New Mexico, New York, North Carolina, Ohio, Pennsylvania, South Carolina, Washington, and Wyoming.13 While there is considerable variation in what each state records, all states in our sample provide a record of each police reported accident, the day when the accident occurred, the county where the incident occurred, the types of parties involved (e.g., light-duty vehicle, motorcycle, bicycle, pedestrian) and the number of fatalities involved. Others variables such as vehicle information or factors contributing to the accident are subject to considerable state-level variation in the manner and detail with which they are recorded. Our main regressions focus on fatalities, PDO accidents, and injuries. Thus we are able to largely avoid variables that are inconsistently recorded.14 We also make use of fields that record whether or not the accident involved any ULD parties, which is almost universally recorded.15 However, in some robustness checks in the appendix we use a subset of states with intoxicated or young operators, which are subject to more variation among states and years.16 We note that an accident appears in our dataset only if the police file a report. PDO accidents may not always be reported, and reporting rates could be subject to weather changes. For minor accidents, police may have different reporting thresholds by state, and policy changes may affect reporting rates over time, which we can control for with countyyear-month fixed effects. There is, however, some concern that weather may influence the likelihood that a report gets filed. In particular when bad weather results in more accidents, departments may become overwhelmed, resulting in higher threshold for filing. If this is the case then we underestimate the effect of weather on accidents. This concern is less important for fatalities, which form our primary analysis, because they will always be reported (Blincoe et al., 2014). 4.1.2 NHTS Daily Travel Data We construct household vehicle miles traveled (VMT) and trip count from the 1990 National Personal Transportation Survey and the 1995, 2001, and 2009 National Household Travel 12 The agencies that usually collect the data are state police, state highway safety department, or the state’s Department of Transportation. 13 Years of coverage in the SDS data include AR 1998-2010, CA 1995-2008, FL 1995-2008, GA 1995-2008, IA 2001-2005, IL 1995-2009, KS 1994-2008, MI 1995-2009, MN 1995-2007, MO 1995-2008, MT 1995-2008, NE 2002-2008, NM 1991-2010, NY 2002-2010, NC 1999-2008, OH 2000-2010, PA 1991-1999 and 2003-2010, SC 1997-2008, WA 1994-1996 and 2002-2010, and WY 1998-2007. 14 Injuries are often but not always recorded as five levels of severity including fatality, incapacitation, injury, possible injury and property damage only. Incapacitation, injury, and possible injury are included in ‘crashes with an injury.’ There are two state-year combinations where injuries cannot be discerned and are dropped from the analysis. 15 Besides pedestrian, motorcycle, and bicycle, we also include mopeds, motorized scooters, pedalcycles, unicycles, and tricycles. 16 See Appendix Table H11. Because some states do not disaggregate drugs from alcohol use, we consider drivers to be intoxicated if they are tested to be beyond the legal limit for alcohol or if they are reported to have taken any illicit drug. 9 Surveys (NHTS). Administered by the U.S. DOT Federal Highway Administration, these surveys are representative cross sections of randomly selected U.S. households.17 The NHTS has several data files available to researchers, one of which includes data on household daily travel diaries. Travel diaries are trip-by-trip travel logs for a single individual. Each trip reports where the respondent went (name of place), what time the trip started and ended, why the respondent made the trip, how the respondent traveled, and the travel distance of the trip, in miles.18 Staff at the Federal Highway Administration and Oak Ridge National Laboratory helped us acquire the confidential version of the NHTS data files that contain either zip code or county of residence for all households in each sample.19 The restricted files that we acquired include the day, month, and year of the household’s assigned day of travel, which is required to merge the travel data with our daily weather data.20 We take several steps to clean and merge the travel diary data. Since our unit of analysis is the household, we aggregate trip count and VMT to the household level for three categories of trips: light-duty, ULD, and public transit. After minor data cleaning, we arrive at a final sample of 283,857 household by travel day observations.21 4.1.3 Historical Weather Data Daily weather data come from the National Climatic Data Center (NCDC) Global Historical Climatology Network-daily, which provides daily minimum and maximum temperature and total daily rainfall and snowfall for weather stations in the United States. This database collects and performs quality control for weather data from land based weather stations around the globe and is archived by the National Oceanic and Atmospheric Administration. We use data from 2,607 stations located in all 50 states and the District of Columbia.22 Weather stations are used to calculate county-level weather. Prior literature has documented that missing weather station data can account for a substantial portion of the variation in weather measures if naively averaged (Auffhammer et al., 2013). Therefore we impute data using a regression of temperature or precipitation for a detector on its nearest neighbor (Auffhammer and Kellogg, 2011; Schlenker and Roberts, 2009). The coefficients from this regression are then used to predict missing values to correct for systematic 17 Each wave is a survey of the non-institutionalized population of the United States using ComputerAssistant Telephone Interviewing (CATI) Technology. The 2009 survey had an average response rate of 19.8 percent. 18 The NHTS specifies that the beginning of a travel day is 4:00 a.m. An example of a recorded trip taken from the 2001 User Guide is the following: “from 7:14 p.m. to 7:22 p.m., return home, by car, 1 mile.” 19 The 2001 confidential file includes zip codes for most households but has limited county information. We assign households to counties using the 2000 U.S. Census zip code to county cross walk. In a few cases, the zip codes reported in the NHTS data do not match any zip codes in the 2000 U.S. Census cross walk. In these cases, we use the 2010 U.S. Census zip code to county cross walk or the U.S. Department of Housing and Urban Development zip code to county cross walk. 20 The public files include the month and year of the travel day. 21 See Appendix B for more details. 22 For more information see Peterson and Vose (1997). We also perform some additional minor quality control. In some instances outliers cannot be confirmed through other sources and appear to be misplaced decimals. As a rule, we impute all observations with snow or rain greater than 1000 cm. We also drop a detector in Nevada that reported several temperatures above 5000◦ C. 10 differences in levels among stations. For the case of missing rain and snowfall, to ensure positive predicted values, we restrict the regression to have a zero intercept. Where the weather data for the next closest detector is missing the imputation is done using the following detector, up to a maximum distance of 200 km. These stations are then averaged to predict daily, county-level weather using inverse distance weighting to the county centroid. This average uses all detectors within 200 km of the county centroid, and where no detectors are within this distance, the county is dropped.23 Because the creation of county-level weather often predicts extremely small levels of precipitation whenever any station has rain or snow, any value less than 0.01 cm is rounded to 0 cm. This results in a balanced panel of 3,140 counties with weather data from 1991 to 2010. Of these counties, 1,474 are then matched to SDS data and 2,944 are matched to NHTS trip data. 4.1.4 Weather Prediction Data CCSM4 is the model used in the Fifth Assessment of the International Panel on Climate Change to predict future climate and weather under a variety of scenarios. We use scenario RCP6.0, which represents a middle-of-the-road prediction of future warming and changes in precipitation. It represents a future with a balanced development of (fossil fuel and non fossil fuel) energy technology.24 Although it is less carbon development than the CCSM4 RCP8.5 scenario, it still entails 2.3◦ C (4.2◦ F) of warming in the United States. In the appendix, we examine outcomes for other climate change scenarios and an alternative model and find that our broad conclusions are unchanged.25 We obtained the CCSM4 RCP6.0 scenario daily weather predictions from the Centre for Environmental Data archival website, made available through the British Atmospheric Data Centre.26 The data include predictions from January 1, 2006, to December 31, 2100 for 0.94 degree latitude by 1.25 degree longitude grid points throughout the entire world.27 Available weather variables include average, minimum, and maximum daily temperatures and rainfall and snowfall rates.28 23 This drops seven counties from the total sample: Aleutians East, Aleutians West, Bethel Census Area, Dillingham, Nome, Northwest Arctic, and Yakutat City all of which are located in Alaska. 24 Other scenarios represent extreme predictions. For example, the RCP8.5 scenario represents a fossil-fuel intensive future, while RCP4.5 represents a predominantly non fossil fuel future. 25 We examine outcomes from climate scenarios RCP4.5 and RCP8.5 using CCSM4 output. We also report outcomes using prediction data from the A1B middle-of-the-road warming scenario presented in the Fourth IPCC Report using the Hadley 3 climate prediction model, which has been used in prior economics literature. See Appendix Table H20 and Figure H.1 for the results from these models. 26 We use the CCSM4 model exclusively for future weather predictions because CCSM4 is one of the few models that report separate outputs for rainfall and snowfall. In contrast, most models report a single precipitation variable that combines rainfall and snowfall. Since we expect rainfall and snowfall to have different impacts on accidents and travel demand, it is crucial to use climate predictions that report separate values for each precipitation type. 27 CCSM4 predictions provide 365 days per year of weather predictions. 28 We convert temperature, reported in Kelvins, to degrees Fahrenheit and precipitation rates, reported in kilograms per meter squared per second to centimeters of precipitation. To convert the rainfall predictions to centimeters per day, we multiply the reported value by 8,640 (http://www.cpc.ncep.noaa.gov/products/ outreach/research_papers/ncep_cpc_atlas/2/cont_data.html). For snowfall, about 1 centimeter of snowfall represents 1 millimeter of water. To make our prediction data consistent with our observed weather 11 To assign predicted weather outcomes to counties, we use the same method that was used to assign observed weather to counties based on weather station locations. Here, for every county, we locate every CCSM4 grid point that is within 200 kilometers of the county’s centroid. The weather predictions at these grid points are then averaged to predict daily county-level weather using inverse distance weighting to the county centroid. 4.1.5 Other Data Sources To predict the change in fatalities nationwide, we require the average daily fatality rate by county, which the accident files provide for 20 states. The Fatality Analysis Reporting System (FARS) tracks annual automobile fatalities for all states and provides information on the county in which the fatality occurred. This allows us to calculate average daily fatality rates from which to project changes in future fatalities. We sum the fatalities recorded by FARS data from 2000-2009 for each county and divide by the number of days to calculate this baseline fatality rate. We also use the FARS data from 1975 to 2013 in the appendix as a check on our main fatality results.29 Finally, in regressions run on urban, suburban, and rural samples, we use the classification system of the National Center for Health Statistics generated by the CDC. 4.2 Summary Statistics Our analysis uses weather data matched to accidents and travel demand for our primary estimation. For our simulation we use daily observed and predicted weather data to generate county-day measures of future weather. Table 1 presents summary statistics for our two primary datasets used in the estimation. Panel A gives key summary statistics on the matched accident data. These data consist of observations of county-day weather matched to counts of accidents and fatalities. Panel B gives statistics on the NHTS travel demand data. These data consist of daily weather matched to aggregate household VMT and trips in that 24 hour period. Panel A describes some key statistics of the SDS data by Census Region. Of the 46 million accidents in our data, 267,984 record a fatality, the vast majority of which do not involve a ULD mode. Additionally, there are 15 million recorded accidents with injuries, but not all states record injuries in all years. When aggregated by county-day, our unit of observation, we record the count of incidents for 6.69 million county-day observations, and where no incidents occur, the day is assigned a count of zero.30 In the average county there are 7.2 accidents per day per 100, 000 people, and 0.07 fatalities per day per 100, 000. The summary statistics by census region reveal that there is considerable variation in temperature, rainfall, and snowfall among regions. We display the 5th, 50th, and 95th quantiles of temperature and the 75th and 95th quantiles of rainfall and snowfall. The station data, which do not report water equivalent snowfall, we must scale the predicted snowfall values. Therefore, we multiply predicted snowfall by 8,640*10 to convert the snowfall predictions to centimeters per day. 29 See Section C for more detail and Appendix Table H8 for results. 30 We assume that if a county has observations in a given year, all days where no accidents are recorded are assigned a zero. 12 Northeast is the snowiest region with 1.91 cm of snowfall at the 95th quantile. It also has the lowest accident rate, 3.6 per 100, 000, and fatality rate, 0.04 per 100, 000. The Midwest has the largest temperature fluctuations with 61 degrees between the 5th and 95th quantiles. The South is generally warmer and rainier than other regions and has little snowfall, with a high accident rate of 8.8 per 100, 000 and fatality rate of 0.08 per 100, 000. The West is drier with less variation in temperature than other regions. This regional variation without location based fixed effects may be cause for concern if, for example, colder climate is correlated with public transportation in cities like New York, Philadelphia, and Chicago, while low-density car oriented cities like Los Angeles and Atlanta are located in warm regions. Panel B describes the NHTS household travel survey data. Because the trip data are a random sample, they do not capture the total number of trips or miles driven in a county but has the benefit of wider coverage with observations in the majority of counties in all 50 states and the District of Columbia and are collected through the course of a year. These data provide information on 283,857 households in 2,944 counties. The data record 1.3 million trips with a total distance of 11.5 million miles. The average household in our sample drives 52.5 miles per day with a total of 6.2 trips. Households in the South take the fewest trips, 6.0 daily, but for the greatest number of miles, 55.2 miles. The Midwest takes the most trips, 6.5 per household, and the Northeast, with 49.3 miles per household per day, travels the fewest miles. The weather observed for these households is generally similar to the accident data, although the West is hotter and less snowy than in the accident data, due to the inclusion of additional states.31 Table 2 presents weather statistics used in our simulation. The sample includes all countyday measures in all states. Panel A depicts the observed weather data for 2000-2009, Panel B shows CCSM4 RCP6.0 predictions for 2006-2009, and Panel C has CCSM4 RCP6.0 scenario for 2090-2099.32 Comparing Panels A and B we can see that CCSM4 generally predicts 5th quantiles and median temperatures in 2006-2009 that are colder than observed in the actual weather data from 2006-2009, while the 95th quantile is predicted to be slightly hotter than actually observed. This indicates that the baseline CCSM4 data display excessive dispersion and that an error correction method targeting the mean may not adequately correct the extreme events. The table indicates that the baseline CCSM4 predicts rainfall well; however, it over-predicts snowfall, both in frequency and in amount. Comparing Panel B with Panel C we note that CCSM4 predicts warmer temperatures, an increase in rainfall (0.11 cm at the 95th quantile), and much lower snowfall. Table 2 also demonstrates why error correction is important. Without error correction, one might conclude that global warming would change temperatures less than CCSM4 predicts, because CCSM4 2006-2009 is a colder baseline than the observed data. In our application where snowfall is important, without error correction the naive comparison of observed data with CCSM4 2090-2099 would suggest that climate change will increase snowfall for all regions except the Northeast. 31 We show in Appendix Table H1 that all regions have some observations in each of the weather bins used in the specification. 32 In Appendix Table H3 we report disaggregated statistics by state for our observed weather station data for 2000-2009. 13 5 5.1 Estimating the Effect of Weather on Accidents and Travel Demand Estimation Methodology This section details the econometric framework we use to determine the effects of weather on accidents and travel demand. Our main analysis uses a Poisson regression model because the distributions of our dependent variables are nonnegative and skewed. We first describe the estimation of accidents that is the framework used for fatalities, injuries, and PDO accidents. Next we describe the estimation of travel demand in terms of daily trip count and miles per trip for LDVs, ULD modes, and public transit. 5.1.1 Accidents We chose a Poisson model for our initial analysis based on several aspects of our data. Accident counts are all non-negative, integer-valued random variables. For data characterized as a counting process, the Poisson distribution is the benchmark model (Cameron and Trivedi, 2013). Poisson regression will yield consistent estimates provided the conditional mean is correctly specified.33 We assume that the count of accidents on date d in county c given xd,c is Poisson distributed with density f (yd,c |xd,c ) = e−µ µyd,c , yd,c ! yd,c = 0, 1, 2, ... (9) We specify the mean µ using the conventional exponential mean function: E[yd,c |xd,c ] = µ = exp 8 X j αj Td,c + j=1 8 X j=1 j j α−1 Td−1,c + 5 X 5 X j=1 j j β−1 Rd−1,c + j=1 5 X j β j Rd,c + 5 X j γ j Sd,c + j=1 j j γ−1 Sd−1,c + θscym + z0d,c δ (10) j=1 j where Td,c is an indicator for mean daily temperature on date d in county c lying within j j j j j the bounds of bin j, Rd,c is for rain, Sd,c is snow, Td−1,c , Rd−1,c , and Sd−1,c indicate lagged weather, θscym is a state-county-year-month fixed effect, and zd,c includes other possible covariates. The appropriate functional form of the daily weather variables is unknown and we adopt the semi-parametric approach of Deschênes and Greenstone (2011). This concern is particularly relevant here, where even after controlling for precipitation, there may be 33 Cameron and Trivedi (2013) note that for many common negative binomial models, consistency requires not only correct specification of the mean and variance but also that the data have a negative binomial distribution. A violation of the assumed Poisson distribution will allow for valid inference only if the standard errors are appropriately computed, which requires correction particularly when there is over- or under-dispersion. 14 differential effects above and below freezing that could be difficult to capture using a parametric specification.34 We assume that the impact of temperature is constant within 10◦ F intervals, and constant for rain or snow falls between 0.0 cm< x ≤ 0.1 cm, 0.1 cm< x ≤0.5 cm, 0.5 cm< x ≤1.5 cm, 1.5 cm< x ≤ 3.0 cm, and 3.0 cm< x.35 Because drivers who are unaccustomed to snow may face an elevated risk, we also create an indicator variable for snow of more than 0.1 cm following a month without any recorded snow.36 Our preferred empirical estimates include lagged variables of weather. The motivation for including these lags is that unfavorable travel conditions may cause individuals to delay travel. We therefore include lags of temperature and precipitation by one day, or one week, to account for this inter-temporal displacement.37 In our setting we include shorter lags than are typical in this literature because it is unlikely that weather is bringing forward accidents that were bound to happen at a later date, but weather may defer trips increasing rates at a later date, although the time-span for such deferment is unlikely to be longer than a week. Consistent estimation of Equation (10) requires that we control for unmeasured shocks that covary with weather. Both regressions include a set of state-county-year-month fixed effects to capture all unobserved determinants of incidents that vary at the county and monthly levels.38 These fixed effects absorb both temporal and spatial changes related to population, employment, and gasoline prices, as well as policy changes such as drunk driving laws and graduated drivers licenses. Conditioning on these fixed effects, we identify αj , β j , and γ j from weather deviations within a county in a given month. Once controlling for these factors, it seems plausible, due to the random nature of weather, that weather is orthogonal to unobserved determinants of accidents and travel demand. The first two moments of the Poisson distribution, E[Y ] = µ and V [Y ] = µ show the equidispersion property that is often violated. The presence of overdispersion, while still providing consistent estimates, will inflate the t-ratios in a Poisson model. To correct the standard errors we block bootstrap at the annual level. 34 For example, even after several days without snow, melting and refreezing may create slick roads. The primary restriction of bin choice lies in the NHTS household survey data, which are more limited than the SDS data for observations on days with extreme weather conditions. We have run specifications with more weather bins for our non-fatal accident and fatality regressions (see Appendix Table H9) and find nearly identical results to those estimated here. 36 In robustness tests included in Appendix H, we also create a variable for infrequent rainfall after one month of no rain. This variable will also capture the effect of oil or debris that may be dislodged by infrequent rainfall. 37 In many other fatality settings, for example, respiratory illness, there is a concern that inclement weather may harm only those who were likely to die shortly thereafter and this literature has stressed the inclusion of lags sufficiently long to capture the net effect (Deschênes and Moretti, 2009). In the main text we report the sum of coefficients from the contemporaneous and lagged weather. Appendix Table H6 gives the full disaggregated set of coefficients. We also include specifications with longer lag periods and more weather bins. In Appendix Table H14 we also examine aggregation of our data to the monthly level. 38 In Appendix Tables H8, H12, and H13 we report estimation results for models that include state-month, county-year fixed effects, which are often used in studies with more aggregate data, and find results that are similar in sign and magnitude. 35 15 5.1.2 VMT and Trips To model travel demand we fit the following equation: E[yi |xd,c ] = exp 8 X α j j Td,c + j=1 8 X j=1 j j α−1 Td−1,c + 5 X j j β−1 Rd−1,c 5 X β j j Rd,c + j=1 + j=1 5 X j j γ−1 Sd−1,c 5 X j γ j Sd,c + j=1 + θscym + (11) z0d,c δ j=1 We avoid log-linearizing and then estimating the equation using ordinary least squares for several reasons. First, there are some households that have zero daily trips or miles for which log-linearization is infeasible. Second, as shown by Santos Silva and Tenreyro (2006), Jensen’s inequality implies that interpreting the coefficients from such an estimate as an elasticity can be incorrect in the presence of heteroskedasticity. Instead, we estimate Equation (11) using Poisson regression.39 The covariates include county-year-month fixed effects, first snowfall, and controls for household characteristics. These controls include the household size, the number of adults, vehicles, and workers in the household, the NHTS defined life-cycle stratum and income group, race, and the day of the week on which the household was followed.40 The estimation procedure is identical to that of Equation (10) except that we now bootstrap the standard errors at the state-by-survey-year level. 5.2 5.2.1 Results Estimates of the Impact of Weather on Accidents We estimate Equation (10) for three mutually exclusive and collectively exhaustive types of accidents: accidents involving a fatality, (PDO) accidents, and accidents involving an injury. Table 3 presents the estimates of the impact of temperature, rainfall, and snowfall on these three types of accidents. We present the sum of the current and lagged coefficients to account for any inter-temporal offsetting that may occur for a given weather fluctuation. In each regression, the temperature bin of 50–60◦ F is omitted, implying that each estimated coefficient is the percent change in accidents compared to a day at 50–60◦ F. Bins for rainfall of 0 cm and snowfall of 0 cm are also omitted. The initial set of columns, (1) through (5), display point estimates associated with weather variables on fatalities. In column (1), our central specification, we find that temperature has a strong and statistically significant effect on fatalities. Compared with a day at 50◦ F, fatality risk rises from -14.3 percent at <20◦ F to 9.5 percent for a day at >80◦ F. The estimated effects for precipitation are somewhat complex. While snowfall increases fatalities, rainfall decreases total fatalities. We suspect this is because drivers avoid trips or drive cautiously enough to reduce overall fatality risks on rainy days, which we will confirm with our PDO accident and travel demand regressions.41 39 In Appendix Table H15 we examine other functional forms of our trip demand model, including the Inverse Hyperbolic Sine Function. 40 Specifications excluding controls given in Appendix Table H16. 41 Note that the number of observations changes between regressions and is larger than the total count of 16 The following columns examine the source of these effects and test the robustness of our fatality result. In column (2) we estimate the daily count of fatalities, omitting fatalities where any party involved was a ULD mode, leaving only fatalities where all parties were LDVs.42 The estimated coefficients are small and generally insignificant, with one potentially important exception being the >80◦ F bin with a 5.4 percent increase in fatalities. If this increase was due to more drivers on the road, we could attribute this effect to exposure. But, as we will show in the travel demand estimates, the hottest days are associated with less travel, which is consistent with heat aggravating an existing risk. The effects from rainfall are also largely removed but snowfall effects remain positive. Column (3) uses only the sample of fatalities where at least one party traveled by a ULD mode. The effects are extremely large. Compared with a day at 50◦ F, a day below 20◦ F sees a 61 percent decrease in fatalities, while a day above 80◦ F sees an 18 percent increase. These magnitudes are large because ULD accidents are a relatively small share of fatalities. With a small base, a change of a few fatalities will result in a large percentage change. Together columns (2) and (3) suggest that ULD fatalities may be a minority of total fatalities but they are the primary mechanism of the temperature-fatality relationship found in column (1). Although it is possible that heat somehow makes an existing set of ULD trips more at risk, this seems unlikely. If cognition were the primary factor, one would expect the minimum number of fatalities to occur between 60◦ F and 80◦ F when comfort is highest, but instead we find the minimum at days when temperatures are below freezing.43 In column (4) we consider the possibility that weather may affect behavior beyond the one-day lag of our main specification in column (1). In this specification we include additional lags for the entire week and find that the sum of the contemporaneous and all lagged coefficients is nearly identical to that with only a single lag.44 Column (5) examines only the set of counties considered large or medium metro counties by the National Center for Health Statistics 2006 Urban-Rural Classification Scheme. The gradient of point estimates from coldest to hottest temperature bins remains as strong as when estimated from the entire sample, suggesting that our results are not driven by counties with low population.45 fatalities in our dataset. This is because all days within a county-year-month are included, many of which are zero. Whenever a county-year-month group has only zero outcomes, the group is omitted, reducing the sample size for regressions such as that in column (3) because ULD crashes are relatively rare. In Appendix Table H8 under the column titled FARS, we also estimate our fatality model, Equation (10), using the FARS records of fatal accidents between 1975 and 2013, finding results are nearly identical to those that we find with the SDS data. See Appendix C for a detailed description of FARS. These data cover only fatal crashes but in all states. 42 We also include heavy-duty vehicles over 4,000 lbs. 43 In Appendix D we estimate models of compositional changes to hold fixed the number of daily accidents. We find that given an accident the likelihood that it involves a ULD vehicle increases with temperature. We also find that fatalities are no more likely to involve alcohol, young drivers, or males as temperatures increase. 44 This specification also removes the possibility our effects are individuals simply picking the warmest day of the week for discretionary travel or ULD travel as exercise, helping to differentiate between relative and absolute effects. 45 Because the fixed effect specification removes all county-year-month groups without any variation, the central specification will generally contain more urban counties. We also estimate models of adaptation in the appendix to identify whether safety technology or driver behavior has altered the relationship between weather and traffic accidents. We find little evidence that the relationship has changed in the past 20 years. 17 In column (6) we examine the daily PDO accident count as a function of weather. We find that heat does not increase accidents over 50◦ F. If the mechanism for increased fatalities were cognition or aggression, it would be surprising that small crashes follow a different pattern. We mostly find effects of temperatures below freezing, which could be due to persistent ice. We also find that rainfall or snowfall increases PDO accidents. Our largest rainfall coefficient is associated with the ¿3 cm bin, indicating that PDO accidents increase by 18.8 percent over a day without rainfall. The effect of snowfall is more than two times larger, with a day of ¿3 cm snowfall increasing PDO accidents by 43.3 percent. This suggests that precipitation generates more accidents but drivers, through either reduced trips or careful driving, reduce the per accident fatality rate. In the case of rainfall these behavioral changes lower the total fatalities, but for snowfall the increase in accidents is so large that the total fatality rate increases, albeit less than the accident rate. Finally, column (7) reports the effect of weather on accidents with at least one injury but no fatalities. There is a slight positive association between temperature and accidents with injuries. The precipitation effects display a similar pattern as that for accidents but of a smaller magnitude. For example a day with snowfall of > 3 cm is associated with an increase in injuries of only 25.9 percent. Broadly these results are a transition between PDO accidents and fatalities. 5.2.2 Estimates of the Impact of Weather on Travel Demand The estimated effects of weather on daily fatalities are suggestive of an exposure mechanism for ULD accidents but the mechanism for the remaining LDV effects is unclear. In this section we further explore these mechanisms using the NHTS logs of daily household travel. Table 4 gives the point estimates and standard errors for a regression of weather on several aspects of travel demand. Our coefficient estimates are again the sum of contemporaneous and lagged effects.46 Given the limited amount of data, statistical precision is lower than in the prior section, but some broad patterns are found in these results. The estimates in column (1) indicate that mean daily temperatures below 20◦ F see 7.6 percent fewer LDV miles per household. Warmer days do not show evidence of a statistically significant change in demand but the bin above 80◦ F does have a negative point estimate. This provides some evidence that the increase in LDV fatalities above 80◦ F cannot be attributed to additional driving and exposure as a mechanism. Similarly the point estimates for precipitation are marginally significant but consistently negative. For LDVs it appears possible that precipitation and possibly hot weather aggravate an existing risk and avoided trips are an additional defensive expenditure. Because our travel demand estimates are based on a self-reported survey, we believe households record trip count with more precision than miles. While miles are a complete measure of adjustment, distance can be difficult to judge and is often rounded. In column (2), we examine the total trip as a function of weather. We find similar patterns as the total miles regression but the bin above 80◦ F finds a larger effect and greater precision, again showing avoided travel.47 See Appendix E for details. 46 See Appendix Table H7 for the disaggregated results. 47 In the appendix, we combine distance with the duration of the trip to generate speed. Speed can be a 18 In columns (3) and (4) we present results with only ULD travel by households. In column (3) we document a positive correlation between temperature and ULD miles below 50◦ F. Above this temperature the point estimates continue to grow but are not precisely estimated. Comparing these estimates with those in column (1), we can see that the changes in demand are an order of magnitude larger than for LDV travel. On days below 20◦ F demand decreases by 76 percent. As might be expected these exposed modes of transit are also unpopular on days with precipitation. The ULD trip count, estimated in column (4), shows that cold temperatures and rainfall also decrease the trip demand. The general pattern confirms our earlier fatality results that ULD demand is closely linked to temperature. The final column, (5), examines only trips taken with public transit options such as bus and subway. Our strongest results come from cold days when users reduce trip demand. Although not measured with precision, the point estimates are negative for hot days when waiting for buses and subways is more unpleasant. Several general observations can be made from these estimates. In Table 4 we find no evidence that ULD travel is reduced as temperature and fatality risk increase. This suggests an exposure mechanism. For LDV travel a different picture emerges where elevated fatality risk on days above 80◦ F (or accident risks on cold, rainy, or snowy days) reduces demand, suggesting reduced LDV travel is a defensive action against aggravated risk due to weather.48 6 Simulation of Future Outcomes The estimates above present a mixed picture of the effect of climate change on traffic accidents. While our estimates suggest warmer temperatures will result in more fatalities, a transition from snow to rain will reduce fatalities. It is also possible that changes in travel demand will offset or exacerbate these costs. In this section we use predicted future weather data to examine the welfare changes from each of these weather changes. Simulating these changes requires a baseline and a predicted future outcome. It is natural to use observed weather as the baseline and the output from climate forecasting models as the future. A common problem arising with this setup is that the current observed weather can be very different from the current weather predicted from the model. Failure to correct for this discrepancy will lead to biased estimates. Prior correction methods typically adjust either the current observed weather or prediction data by adjusting mean outcomes. This is problematic because, as noted in our summary statistics, global warming involves not only shifting mean values, but also increased variability in weather. In this section we initially describe the framework of how our estimates from above are applied to a new distribution of weather outcomes generated by climate change. Then we examine in depth how that new weather distribution is developed from climate models taking account of the baseline error inherent in these models. We first describe the existing methods function of driver choice but also of congestion, complicating interpretation of the speed estimates. But one might be concerned that deadly crashes were generated by higher intensity crashes. In Appendix Table H17 we do not find any evidence that fatalities are due to speed effects. If anything, drivers reduce speed on days with temperatures over 80◦ F and when there is snowfall, again suggesting defensive behavior. 48 These results are robust to alternative functional form assumptions. We estimate an alternative model that can handle zero values, the Inverse Hyperbolic Sine, in Appendix Table H15 and find that our results are broadly consistent with those reported using the Poisson model. 19 and discuss cases where they may result in counterintuitive results. We also illustrate some of these concerns with an example. Finally, we provide an improved methodology inspired by techniques used by climate scientists. To calculate the change in accidents we sum the daily changes to each county: XX ∆y = ∆yd,c , (12) c d where ∆yd,c is the predicted change in accidents on day d in county c. To calculate the daily changes in accidents, we multiply the baseline accident rate in the county with the percent change in accident rate for each weather measure: " 8 # X j j j ∆yd,c = α bj + α b−1 · Tbd,c − Td,c yd,c + j=1 " 5 X j b j − Rj (βbj + βb−1 )· R d,c d,c # yd,c + (13) j=1 " # 5 X j j j yd,c . (b γj + γ b−1 ) · Sbd,c − Sd,c j=1 j j j The α bj , βbj , γ bj and α b−1 , βb−1 ,γ b−1 terms denote our estimated contemporaneous and lagged coefficients from Equation (10), respectively, which are summed to give the net effect of a j bj and Sbj represent indicators for future day with particular weather conditions. The Tbd,c ,R d,c d,c weather on date d in county c within bin j. Equation (13) generates the percent change in accidents by multiplying the number of bin changes by the marginal effect of a bin change. This percent change is then multiplied by the baseline daily level of accidents in the county yd,c .49 Because our SDS data have fatality rates, yd,c , for only 20 states, we use FARS data to generate average daily fatality rates at the county level.50 For simulations of LDV and ULD fatalities, injuries, and PDO accidents, we project the county baseline using a Poisson model by regressing the outcome variable (e.g., LDV fatalities) on the fatality rate and population by county for the counties in the SDS data and use the estimated coefficients to impute missing counties. For our trip demand simulations we use the observed average per household in each county in our NHTS sample, and where no NHTS data is given in a county, we apply the national average.51 This daily household value is then multiplied by 365 days and the number of households in the county as taken from the census. Evaluating Equation (12) requires obtaining estimates of future weather predictions. We base our weather predictions on the CCSM4 RCP6.0 Scenario, which includes daily (minimum, maximum, and average) temperature and precipitation throughout the United States. Problematically, these models do not predict weather at a county level but rather at 49 As an example, suppose α b7 = 0.05, α b8 = 0.08. If future predicted values indicate that one day will move th th 7 8 from the 7 bin to the 8 bin, then Td,c = 1 and Tbd,c = 1. If the baseline level of accidents yd,c = 2, then ∆yd,c = [0.08 − 0.05] 2 = 0.06, interpreted as a predicted increase of 0.06 accidents in county c on day d. 50 We use FARS data averaged from 2000 to 2010 for these baseline estimates. 51 These averages are generated across all of our NHTS data. 20 equally spaced grid points. This implies that what the model predicts in the baseline years from 2006 to 2009 will not match observed weather outcomes particularly in areas with complex terrain (Wilcke, Mendlik and Goblet, 2013). If these baseline discrepancies are not adjusted, the simulation will generate changes that are the result of this baseline discrepancy as opposed to changes in weather. In the next section we detail the shortcomings of prior methods used to correct this discrepancy and compare them with quantile-based methods, based on methods developed in the climate science literature. 6.1 Prior Calibration Methods Traditionally, economists have corrected biases between the baseline observed data and the baseline predicted data from models like CCSM4 with an additive term or a multiplicative factor to match means (Deschênes and Greenstone, 2011; Ranson, 2014). We characterize these methods as four unique corrections: predicted additive, predicted multiplicative, observed additive, and observed multiplicative. Although in theory any method can be used, additive methods tend to be used on temperature, while multiplicative methods tend to be used on precipitation. A predicted additive correction adds a value to each future prediction (CCSM4 20902099), where the correction is defined by ξτ,c = 1 X (xd,c − xd,c ) . Nτ d∈τ (14) The value of the correction term is the average of the difference between the baseline CCSM4 prediction (2006-2009), denoted by xd,c , and the observed weather station data, denoted by xd,c , for a given weather variable and predefined time period τ . The term Nτ represents the number of days observed in the time period τ . The time period τ can be as short as a single day to as long as an entire decade.52 The correction term is then added to the CCSM4 future prediction, denoted by x ed,c , to obtain a corrected future prediction: x bd,c = x ed,c + ξτ,c . (15) A predicted multiplicative correction takes the product of the CCSM4 future prediction and a multiplicative factor: x bd,c = x ed,c ντ,c . (16) The term ντ,c is defined as ντ,c = 1 Nd 1 Nd P xd,c P xd,c d∈τ d∈τ . (17) The term ντ,c is the ratio of the average observed weather station data and the average CCSM4 baseline data for a given month and county. 52 The choice of the time period varies across studies. For example, Deschênes and Greenstone (2011) calculate errors for each county by day of year by computing the difference between county by day of year specific average temperature from observed weather data and CCSM4 predictions during the baseline period. Ranson (2014) calculates errors for each county by month of year. 21 Another class of correction method uses the current observed distribution as its base and adds or multiplies the change observed in the CCSM4 prediction. For the additive observed correction a projected change is added to each outcome in the observed weather. The change is defined by 1 X (18) (e xd,c − xd,c ) . ψτ,c = Nd d∈τ The correction term is then added to the observed weather data, xd,c , to obtain a corrected future prediction: x bd,c = xd,c + ψτ,c . (19) The multiplicative observed correction scales each outcome in the observed weather, where the scaling factor is defined by P 1 ed,c d∈τ x Nd . (20) ζτ,c = 1 P d∈τ xd,c Nd The correction term can be multiplied by the observed weather data, xd,c , to obtain a corrected future prediction: x bd,c = xd,c ζτ,c . (21) Our concerns with existing methods fall into two categories. First, they can correct biases only in the mean and not other moments of the distribution. Failing to match higher order moments may lead to unrealistic weather changes. One benefit of the observed additive correction, which has been used in recent work (e.g., Schlenker and Roberts (2009); Ranson (2014)), is that corrected future data will not have dispersion or compression of weather values that is purely the result of baseline modeling error. The drawback is that where CCSM4 predicts a change in weather variability, it will be omitted from the simulation.53 Second, existing correction methods often imply unreasonable corrections to daily precipitation variables, which have many zeros. Although additive methods are generally not used for precipitation, if ξτ,c > 0 or ψτ,c > 0, then the corrected rainfall prediction will shift all days with zero precipitation to a positive value. Prior studies have instead corrected precipitation with a multiplicative method. This correction can, nevertheless, be P problematic. If xd,c = 0, which is likely to be the case for snowfall in many places or d∈τ rainfall in dry counties, ντ,c is undefined.54 Thus any correction method must be performed at a highly aggregate level and is sensitive to trace precipitation.55 The multiplicative correction method also has no ability to generate or remove trace precipitation days and can only scale up or down already existing precipitation. In Panels (a) and (b) of Figure 1 we illustrate these concerns using a temperature distribution from a representative county.56 Panel (a) depicts the PDF of observed daily 53 Changes in weather variability are of particular interest for projecting crop yield changes (Schlenker, 2006). P 54 xd,c is slightly greater than zero, ντ,c can be large, which is unrealistic as most In some cases when d∈τ climate models predict modest changes in precipitation. 55 Trace precipitation is defined as daily rainfall or snowfall less than 0.01 inch or 0.0254 cm. 56 The data that we used to generate these distribution functions are taken from Orange County, CA; however, we have shifted and enhanced some features to aid in exposition. 22 mean temperature data, the CCSM4 2006-2009 baseline prediction, and the CCSM4 20902099 future prediction. CCSM4 predicts a baseline mean temperature of 40.0◦ F, which in this particular county is nearly equal to the observed mean temperature of 41.0◦ F. Crucially, the dispersion of the baseline CCSM4 data is much higher and it predicts more extreme temperatures in 2006-2009 than were actually observed. The future CCSM4 prediction of mean temperature is 48◦ F, which is an 8◦ F increase. While CCSM4 predicts that all days increase in temperature, this increase is disproportionately large for days that were already warm. In this example, the CCSM4 predictions suggest that the hottest days are nearly 20◦ F warmer, while the coldest are only 5◦ F warmer. In Panel (b) we demonstrate the additive mean-matching correction methods used in prior papers for temperature. Because the initial means are nearly identical, the predicted additive method does not substantially transform future CCSM4 weather outcomes. This method implies that there will be an increase in the frequency of days below 20◦ F, even though it is obvious from Panel (a) that all days become warmer. The implications for hot days are also somewhat unrealistic. The hottest days increase by more than 40◦ F. No shift of this magnitude is seen between the CCSM4 baseline and future predictions. For the observed additive method, all days are shifted up by 8◦ F. While this method does not spuriously generate colder days, all days warm by an equal amount; this translation omits the change in dispersion that characterized the CCSM4 prediction. In Appendix G we produce a similar figure illustrating the outcomes of the multiplicative methods. The concerns, however, are similar.57 When matching only the mean, broader changes in the distribution are forced and are often unrealistic. 6.2 Calibration Using a Quantile-Mapping Method To correct these biases with more intuitive outcomes, we draw on quantile-based methods used by atmospheric and climate scientists (Wilcke, Mendlik and Goblet, 2013). The major advantage of this method is that it corrects all moments of the distribution, not just the mean. With these methods, it is possible to have positive corrections to some parts of the distribution and negative corrections to others. Our method consists of three steps: defining the Empirical Cumulative Distribution Function, solving for and applying the error correction, and correcting for any wet bias. Quantile-mapping methods generate a corrected prediction by comparing Empirical Cumulative Distribution Functions (ECDFs) for each weather measure X ∈ {T, R, S}. The researcher must pick the relevant geographic area and time window from which to generate these functions. In our simulation, we construct these for the average year at the county level, c, based on 10 years of data. Once the ECDF is generated, we take the inverse. This generates three inverse ECDFs: the observed baseline, Φ−1 X,τ,c (·), the CCSM4 baseline −1 e −1 (·). prediction, ΦX,τ,c (·), and the CCSM4 future (uncorrected) prediction, Φ X,τ,c These three functions allow us to solve for the error correction. For a given observed weather outcome xd,c , we compute the cumulative probability of achieving that outcome in 57 Appendix Figure H.2 uses a different CCSM4 2090-2100 distributional change (an upward translation) and shows how multiplicative methods can generate spurious effects. 23 the observed baseline data, denoted by p: p = ΦX,τ,c (xd,c ). (22) Next, we evaluate the difference between the baseline CCSM4 ECDF and the future prediction CCSM4 ECDF at the probability p: e −1 (p) − Φ−1 (p). ∆xd,c = Φ X,τ,c X,τ,c (23) We then add the difference to the current weather outcome to generate an error-corrected weather prediction:58 x bd,c = xd,c + ∆xd,c . (24) The method is illustrated in Panels (c) and (d) of Figure 1 for temperatures in our representative county. In Panel (c) we show the inverse ECDFs for all three functions and the final corrected outcomes. Take an observed temperature of 33◦ F. The ECDF of the observed temperatures indicates that 20 percent of all days have temperatures below 33◦ F. e −1 (0.2) − Φ−1 (0.2) = 5◦ F. This correction The correction applied to this temperature is Φ T,τ,c T,τ,c increases the temperature to 38◦ F in the future corrected outcomes. This non-parametric correction is applied to each quantile, generating the inverse ECDF denoted by the dotted red line. Panel (d) shows the final corrected PDF, which indicates both increasing temperatures and increasing dispersion, mirroring the changes predicted by CCSM4. For precipitation measures, a few other corrections must be made. First, the correction must be bound from below by zero: x bd,c = max{xd,c + ∆xd,c , 0}. (25) Second, when precipitation in the observed baseline is less frequent than the predicted baseline, and the future prediction indicates more precipitation, the “wet bias” must be corrected.59 This set of circumstances is illustrated in Panels (e) and (f) for a hypothetical county where we exaggerate outcomes to aid in exposition. In this county, 20 percent of all days currently receive no rainfall. For this county CCSM4 predicts an increase in rainfall. Without wet bias correction, any dry day will be mapped to a positive precipitation amount e −1 (0.175)−Φ−1 (0.175) = ∆xd,c = 0.1 cm. To reduce the number of days with rainfall, of Φ X,τ,c X,τ,c we randomly draw precipitation days from the future predicted ECDF below the point where e −1 (p ≤ ΦX,τ,c (0)). the baseline predicted ECDF is 0. Specifically, we randomly draw from Φ X,τ,c This reduces the number of days with trace rainfall to match the CCSM4 model prediction. 6.3 6.3.1 Results Simulation Results using Quantile Mapping Table 5 summarizes the results of our simulation and welfare calculations using our quantile mapping method. Each simulation in Panel A details the changes in accidents or travel 58 See Appendix G for a slightly modified version of the quantile-mapping correction method that uses the CCSM4 baseline weather data to establish the probability p. The two methods generally give nearly identical results. 59 Wet bias is a common phenomenon in climate prediction research and occurs when quantile mapping methods are applied in settings where dry day frequency in prediction models is higher than in the observed weather data. See Wilcke, Mendlik and Goblet (2013) for more details. 24 demand from each weather component as well as the 95 percent confidence intervals.60 Column (1) reports the effect of weather from 2090 on fatalities using our estimates from Table 3 column (1). The reduction in snowfall and increase in rainfall result in reduced fatalities. These effects are more than offset by the increased fatalities due to temperature. The net change in fatalities is an increase of 397. Applying the Department of Transportation’s (DOT) VSL at $9.1 million, these fatalities have a cost of $3.6 billion. When discounted using a 3 percent rate, this has a present value of $360 million.61 Columns (2) and (3) attribute these costs to LDV and ULD crashes using the estimates of columns (2) and (3) from Table 3. These results suggest that LDV fatalities will increase by 63 annually but these estimates are not distinguishable from zero, while ULD fatalities will increase by 298.62 Therefore, we cannot reject the possibility that the entire increase in fatalities will be due to ULD crashes. We estimate welfare changes stemming from changes in demand for LDV and ULD travel using terms W T P m and W T P b in Equation (8).63 Column (4) applies the estimated LDV mile demand effects estimated in column (1) of Table 4 to future weather outcomes. The increase in miles generates a benefit of $48 million although the 95 percent confidence interval overlaps with zero and we cannot reject the possibility there are defensive expenditures that carry additional costs. Column (5) uses the ULD mile demand estimates from column (3) of Table 4. ULD miles will be expected to increase by 0.98 billion by 2090 carrying a benefit of $315 million. This value is statistically different from zero and would offset 88 percent of the total fatality costs.64 Column (6) uses our accident regressions from Table 3, column (6). Because PDO accidents are negatively correlated with temperature, weather from 2090 would decrease the number of PDO accidents by 22,485. We evaluate these costs using a value per PDO accident of $10,633 with a total benefit of $18 million.65 60 These confidence intervals incorporate the uncertainty of our point estimates but not the uncertainty across various climate change models. These are generated by drawing from a normal distribution with the mean and standard deviation of coefficient estimates given in Table 3. These draws are then applied to the predicted weather effects in each decade. From these values the mean and 95 percent confidence interval are calculated. When summing across decades, the mean and 95 percent confidence interval are assigned to the midpoint of the decade and linear interpolation is used between those years. 61 We do not adjust the VSL amount for future changes in income. We use a 3 percent discount rate to be consistent with DOT assumed rates (Blincoe et al., 2014). 62 Because these are taken from the estimates in Table 3 they do not sum to 397 but are not inconsistent with this value. 63 To calculate the change in demand from our estimates, we multiply the estimated percent change in demand times the number of miles taken daily in the county. To generate the daily number of miles, we use the average number of miles across households in the county in the NHTS data, which is multiplied by the number of households in the county. Miles are priced based on the average fatalities per LDV or ULD mile times the VSL. 64 We reiterate that these terms are derived from a stylized framework that leaves out several features that may influence travel demand and thus alter the welfare calculation. For example, due to the lack of appropriate data we cannot measure the welfare gains from improvements in trip quality as temperature increases from cold to warm. These effects are likely positive for both travel modes, implying that we are underestimating the welfare gains from increases in travel demand stemming from climate change. 65 We use the per vehicle value of PDO crashes from Blincoe et al. (2014) and multiply this by the average number of vehicles per crash of 1.75. 25 The net effect on injuries, given in column (7), is an additional 4,909 in 2090. Because the relationship between temperature and injuries is flatter than fatalities, the 95 percent confidence interval overlaps with zero and in many near-term intermediate years there is a decrease of injuries. We evaluate these at $141,677 per injury, which is the frequency weighted average of the five injury levels recognized by the DOT’s Maximum Abbreviated Injury Scale.66 This generates an additional cost of $94 million. Panel B aggregates results from 2015 through 2099. The total net present cost of an additional 18,603 fatalities is $39.9 billion.67 We find that the benefits of ULD travel are $34.1 billion. These benefits would offset more than 85 percent of the fatality damages and we cannot reject full offsetting.68 Taken together, our results suggest that excluding exposure benefits from cost-benefit analysis can dramatically overestimate the expected net costs from climate change. Figure 2 illustrates the effect of end-of-the-century climate change on traffic accidents by plotting impacts for each county in the United States. Panels (a), (b), and (c) map the net effects of climate change on traffic fatalities, injuries, and PDO accidents, respectively. Counties colored red are predicted to see an increase in fatalities, while blue colored counties are predicted to see a decrease in fatalities. The deeper the color, the larger the magnitude of the effect. The ranges for each color are in terms of average annual fatalities per 100,000 people. Panel (a) shows that nearly all counties are predicted to see an increase in the rate of fatalities per person except some interior regions in the North-West. For injuries that appear in Panel (b), much of the country will experience an increase in injuries, except for a swath from Michigan and Ohio through Oklahoma, northern regions bordering Canada, and south Florida. In contrast, we see in Panel (c) that a majority of counties will experience a reduction in PDO accidents, although a significant number of counties in the South and South-West experience an increase. The effect of climate change on PDO accidents is dominated by the transition of days with snowfall to rainfall, as the temperature coefficients are small and the heavy snowfall coefficients are much larger than the corresponding heavy rainfall coefficients. 6.3.2 Comparison of Weather Correction Methodologies Table 6 presents the comparison between our quantile-mapping method and the methods common in the literature that are based on the mean. Column (1) repeats our predicted change in fatalities for the decade 2090-2099 with 95 percent confidence intervals given in brackets. We note that our earlier summary statistics from Table 2 showed that future 66 This standardized scale was developed by the Association for the Advancement of Automotive Medicine. We use the weights of police reported injury levels from Blincoe et al. (2014). These include costs for quality of life, medical, property damage, congestion, insurance, market and household productivity, and legal costs. 67 In Appendix H, we present net present cost estimates under several other warming scenarios and prediction models. Table H20 shows that under the RCP4.5 scenario, there would be 18,115 fatalities from 2015 to 2099 with a total net present cost of $44.7 billion. The RCP8.5 scenario would generate 33,376 fatalities and costs of $73.5 billion during that time span. The Hadley 3 A1B scenario generates 27,493 fatalities for a total cost of $62.6 billion. 68 It is important to note that offsetting can be sensitive to the climate scenario used because there can be substantial variation in the timing and location of weather changes. These additional trips are evaluated by their price, which only incorporates the risk of a fatality. We do not incorporate accident or injury costs because FARS does not track these and the SDS data only track crashes involving LDVs. We also evaluate LDV trips only using fatality costs, excluding other costs such as time and gasoline for consistency. 26 weather was broadly characterized by higher temperatures, increased rainfall, and decreased snowfall. While population is not evenly divided among all counties, intuition suggests that our results should broadly align with these national trends where temperature changes increase fatalities while snow and rainfall effects are negative. First we examine columns (2) and (3) with CCSM4 2090 predicted data as future weather, but correct that prediction based on the baseline discrepancy between observed weather and CCSM4 2006-2009 predicted weather. This is accomplished with either an additive term in column (2) following Equation (15), or a multiplicative term in column (3) following Equation (16). Examining snowfall for both methods we find an increase in fatalities by 406 and 307 annually. This result is counterintuitive given our positive point estimates for snowfall in Table 3. A decrease in snowfall should reduce snow-related fatalities. In columns (4) and (5) we simulate fatalities using the additive observed method given in Equation (19), and multiplicative observed method given in Equation (21), respectively. These methods generate future weather outcomes by adjusting the current distribution of weather such that the mean changes by the amount predicted by the CCSM4 simulation between 2006-2009 and 2090-2099. Here we focus on rainfall. Nationally we expect rainfall to increase, and with negative point estimates for rainfall effects in Table 3, there should be a decrease in fatalities. Therefore, it is unsurprising that fatalities decease. What is surprising about the additive method shown in column (4) is the magnitude, which indicates a decrease of 433 fatalities. This large result is likely due to counties with a predicted increase in rainfall where a common additive term will change all zero events to positive events, resulting in hundreds of days with trace rainfall. By contrast a multiplicative correction, such as that demonstrated in column (5), can only adjust rainfall by changing the quantity on preexisting nonzero events. This implies that frequency will never adjust and that most of the adjustment will occur on extreme events.69 This could explain the small change in fatalities of -12. While the comparison given in this table does not prove our method is correct, we find our results to be intuitive given the summary statistics of weather changes predicted by CCSM4. Many of the discrepancies among methods seem likely due to the mechanics of matching only the mean. Most prior work in this literature has focused on temperature where we find that the confidence intervals generally overlap among all methods, but it is unclear whether this result will hold more generally. Where rainfall and snowfall are important, as in our context, the quantile-mapping method overcomes drawbacks of prior methods. 6.4 Discussion There are several relevant caveats to our predictions of future fatalities. First, there are many aspects of automobile demand that may change given the time scale involved. Car ownership rates in the United States are among the highest in the world and may change substantially. Public attitudes, city structure, congestion levels, and energy prices could change driving behavior in unexpected ways. As an example, in 2008 the number of fatalities was 34,172, but one year later fatalities were 30,862 due to reduced driving from the recession. Second, even if all 397 additional deaths are due to voluntary exposure, this is not an 69 For example, if a location has only 1 rainfall event of 10 cm, a 0.5 adjustment will reduce this amount to 5 cm, which remains in the top bin. 27 argument against public policy to reduce this number. Increased fatalities may change the cost-benefit analysis of a particular policy. There is also a lot of room for policy to reduce fatalities. A gasoline tax of roughly $0.17 would reduce miles sufficiently to remove these fatalities.70 Another limitation of our welfare calculations is that they do not incorporate adaptation. Adaptation may come in the form of private or public adaptation to warmer, rainier weather. Although regressions shown in the appendix do not find evidence that the marginal effect of weather on fatalities is different between cold and hot locations or has changed over the past 20 years,71 it is possible that individuals will invest in adaptation. Driver-assistance technology (e.g., automatic lane changing) and autonomous vehicle technology are two key examples of technology that may help to avoid fatal interactions with ULD vehicles. Leaving these potential technologies out of our welfare calculations imply that our estimates overstate the long-run impacts of climate change. But there is also reason to worry that the role of private adaptation is limited. Any optimal private response to changing fatality rates requires that accident costs be internalized (White, 2004). Most states use “contributory negligence” whereby damages are assigned in proportion to fault. But provided the driver is not negligent, there is no incentive to avoid heavy pedestrian areas, maintain larger distances from bicycles and motorcycles, or avoid automobile travel altogether on warm days when these risks are high. Finally, our estimates do not account for the possibility that individuals could migrate away from affected areas, as a form of private adaptation. If the fatalities we estimate are due to voluntary exposure, migration should not be expected because they would not be associated with lower utility. Because our measure of voluntary exposure does not offset all costs, some of our estimates may be due to aggravated risk, which would encourage migration. Previous estimates of migration in response to climate change suggest that migration from hot to cool regions will be modest (Albouy et al., 2016). Migration and other forms of private adaptation reduce the costs of climate change, so including them in our framework would lower our computed costs. 7 Conclusion This paper estimates the impact of weather on traffic fatalities, injuries, and PDO accidents as well as total trip demand. We exploit plausibly random day-to-day variation in weather to show that fatalities increase with snowfall, decrease with rainfall, and are positively correlated with temperature. The evidence suggests that these effects are primarily due to interactions with pedestrians, bicycles, and motorcycles. We apply these estimates to a climate change scenario RCP6.0, a balanced growth scenario of fossil and non-fossil energy sources resulting in 2.3◦ C (4.2◦ F) of warming by the end of the century across the United States. The climate data are corrected for use in our simulation using a quantile-mapping methodology that allows us to account for baseline modeling error while allowing for changes in the distribution beyond the mean. We find that while the shift from snow to rain will save roughly 72 lives annually, temperature increases will cause 468 additional fatalities. The 70 71 This calculation assumes no change in vehicle characteristics, and a rebound effect of 20 percent. See Appendix Table H10. 28 net annual increase will be 397 fatalities by 2090. But when deaths involving pedestrians, bicyclists, and motorcyclists are removed, the increase is no longer statistically different from zero. It is possible that some fatalities are due to psychological effects or changes in aggression but most evidence points to voluntary exposure as the primary mechanism of increase. We find that these fatalities carry a total net present cost of $40 billion by the end of the century. Increased benefits from travel that capture the benefits of voluntary exposure, however, offset more than 85 percent of these costs. Accounting for these benefits more broadly may help to explain why, despite the numerous documented negative outcomes associated with heat, the predominant migration pattern within the United States in recent decades has been towards warmer climates. One broader implication of this research is that voluntary exposure is an important mechanism for understanding the welfare implications of climate change. Our estimated effects appear to be the result of individuals being drawn outdoors and using forms of transportation that will not protect them in a crash. It is possible that individuals spending time outdoors will also be exposed to street crime or air pollution. In the agriculture sector farmers may choose riskier but higher-profit crops. Maize, with a longer time to maturity, has higher yield but is also riskier to grow.72 It is plausible that warm locations may be susceptible to large profit losses from weather shocks because farmers choose high-risk, highreward crops. Voluntary decisions may also be important for broader economic measures. Warmer climate may allow for outdoor leisure time in activities like biking, for which the welfare benefits may or may not be captured by GDP. It may also increase the opportunity cost of working. A priori it is difficult to ascertain what fraction of GDP loss during warm weather is due to an aggravated cost and what has an offsetting voluntary benefit. In other domains, such as health effects of temperature for infants and elderly, a voluntary exposure mechanism seems less plausible. Finally, it is important to note that the exposure mechanism will vary across countries, particularly for transportation. The United States, with highly developed infrastructure, limited access highways, and urban areas designed to channel high speed traffic away from residential areas, provides significant separation between vehicles and pedestrians, and bicycles. By contrast, developing nations, and even some middle income countries, have larger fatality rates per capita, largely due to vehicles colliding with pedestrians (Kopitis and Cropper, 2005). It should also be noted that our results do not indicate that reliance on walking, biking, and motorcycling imply large fatality rates, as other English-speaking and western European nations have per capita fatality rates that are often less than half that of the United States.73 Some countries like Sweden with extraordinarily low fatality rates have pursued a variety of urban design and legislative changes to reduce fatalities with policies such as replacing intersections with roundabouts to slow vehicles where they are likely to 72 For example, a hybrid with a relative maturity rating of 100 days may produce 210 bu/A, while one with a 101 day rating may produce 212 bu/A. An extra day of maturation requires a warmer climate but also exposes the crop to a higher possibility of an extreme weather event. Generally, perennial crops such as citrus, and wine have high dollar value but are restricted to warm climates as they require a long, uninterrupted growing season and can suffer for many years from short-run shocks, while commodity crops such as wheat are annual, hardy, and can be grown in northern Alberta. 73 Per capita rates in the United States are 11.4 per 100,000. The rate for Australia is 6.1, Canada 6.8, Germany 4.7, Norway 4.3, Sweden 3.0, Switzerland 4.3, and the UK 3.7 (World Health Organization, 2013). 29 encounter pedestrians.74 Relatively simple changes like these may prove to be effective, although unglamorous, adaption strategies to climate change. References Albouy, David, Walter Graf, Ryan Kellogg, and Hendrik Wolff. 2016. “Climate amenities, climate change and american quality of life.” Journal of the Association of Environmental and Resource Economists, 3(1): 205–246. Auffhammer, Maximilian, and Erin T. Mansur. 2014. “Measuring climatic impacts on energy consumption: A review of the empirical literature.” Energy Economics, 46: 522– 530. Auffhammer, Maximilian, and Ryan Kellogg. 2011. “Clearing the air? The effects of gasoline content regulation on air quality.” The American Economic Review, 2687–2722. Auffhammer, Maximilian, Solomon M. Hsiang, Wolfram Schlenker, and Adam Sobel. 2013. “Using weather data and climate model output in economic analyses of climate change.” Review of Environmental Economics and Policy, 7(2): 181–198. Bäck, Danielle, Nicolai V Kuminoff, Eric Van Buren, and Scott Van Buren. 2013. “National evidence on air pollution avoidance behavoir.” Barreca, Alan, Karen Clay, Olivier Deschênes, Michael Greenstone, and Joseph S. Shapiro. 2013. “Adapting to climate change: The remarkable decline in the U.S. temperature-mortality relationship over the 20th century.” NBER Working Paper No. 18692. Blincoe, Lawrence, Ted R. Miller, Eduard Zaloshnja, and Bruce A. Lawrence. 2014. “The economic and societal impact of motor vehicle crashes, 2010.” National Highway Traffic Safety Administration, Washington, DC. Broadbent, DE. 1963. “Differences and interactions between stresses.” Quarterly Journal of Experimental Psychology, 15(3): 205–211. Cameron, A Colin, and Pravin K Trivedi. 2013. Regression analysis of count data. Vol. 53, Cambridge University Press. Dell, Melissa, Benjamin F. Jones, and Benjamin A. Olken. 2014. “What do we learn from the weather? The new climate-economy literature.” Journal of Economic Literature, 52(3): 740–798. Deschênes, Olivier, and Enrico Moretti. 2009. “Extreme weather events, mortality, and migration.” The Review of Economics and Statistics, 91(4): 659–681. 74 The “Vision Zero” policy was adopted by Sweden in 1997 to reduce fatalities to zero by 2020. Similar policy initiatives have now been adopted by other nations and cities including several in the United States. 30 Deschênes, Olivier, and Michael Greenstone. 2007. “The economic impacts of climate change: Evidence from agricultural output and random fluctuations in weather.” American Economic Review, 97(1): 354–385. Deschênes, Olivier, and Michael Greenstone. 2011. “Climate change, mortality, and adaptation: Evidence from annual fluctuations in weather in the US.” American Economic Journal: Applied Economics, 3(4): 152–185. Deschênes, Olivier, Michael Greenstone, and Joseph S. Shapiro. 2013. “Defensive investments and the demand for air quality: Evidence from the NOx Budget Program and Ozone Reductions.” IZA Discussion Paper No., 7557. Eisenberg, Daniel. 2004. “The mixed effects of precipitation on traffic crashes.” Accident Analysis & Prevention, 36: 637–647. Eisenberg, Daniel, and Kenneth E. Warner. 2005. “Effects of snowfalls on motor vehicle collisions, injuries and fatalities.” American Journal of Public Health, 95(1): 120–124. Fisher, Anthony C., Michael Hanemann, Michael J. Roberts, and Wolfram Schlenker. 2012. “The economic impacts of climate change: Evidence from agricultural output and random fluctuations in weather: Comment.” American Economic Review, 102(7): 3749–3760. Froom, Paul, Yeheskial Caine, Igal Shochat, and Joseph Ribak. 1993. “Heat stress and helicopter pilot errors.” Journal of Occupational and Environmental Medicine, 35(7): 720–732. Graff Zivin, Joshua, and Matthew Neidell. 2013. “Environment, health and human capital.” Journal of Economic Literature, 51(3): 689–730. Graff Zivin, Joshua, and Matthew Neidell. 2014. “Temperature and the allocation of time: Implications for climate change.” Journal of Labor Economics, 32: 1–26. Graff Zivin, Joshua, Solomon M Hsiang, and Matthew J Neidell. 2015. “Temperature and human capital in the short-and long-run.” National Bureau of Economic Research (NBER) Working Paper, 21157. Grether, WF. 1973. “Human performance at elevated environmental temperatures.” Aerospace Medicine, 44(7): 747–755. Hancock, PA, and Ioannis Vasmatzidis. 2003. “Effects of heat stress on cognitive performance: The current state of knowledge.” International Journal of Hyperthermia, 19(3): 355–372. Harrington, Winston, and Paul R. Portney. 1987. “Valuing benefits of health and safety regulation.” Journal of Urban Economics, 22: 101–112. Hsiang, Solomon M., Marshall Burke, and Edward Miguel. 2013. “Quantifying the influence of climate on human conflict.” Science, 341(6151). 31 Jacob, Brian, Lars Lefgren, and Enrico Moretti. 2007. “The dynamics of criminal behavior: Evidence from weather shocks.” Journal of Human Resources, 42(3): 489–527. Kopitis, Elizabeth, and Maureen Cropper. 2005. “Traffic fatalities and economic growth.” Accident Analysis & Prevention, 37(1): 169–178. Neidell, Matthew J. 2004. “Air pollution, health and socio-economic status: The effect of air quality on childhood asthma.” Journal of Health Economics, 23(6): 1209–1236. Peterson, Thomas C, and Russell S Vose. 1997. “An overview of the Global Historical Climatology Network temperature database.” Bulletin of the American Meteorological Society, 78(12): 2837–2849. Ranson, Matthew. 2014. “Crime, weather, and climate change.” Journal of Environmental Economics and Management, 67: 274–302. Santos Silva, Joao M.C., and Silvana Tenreyro. 2006. “The log of gravity.” Review of Economics and Statistics, 88(4): 641–658. Schlenker, Wolfram. 2006. “Inter-annual weather variation and crop yields.” New York, NY: Unpublished Working Paper, Dept. of Economics and School of International and Public Affairs, Columbia University. Schlenker, Wolfram, and Michael J Roberts. 2009. “Nonlinear temperature effects indicate severe damages to US crop yields under climate change.” Proceedings of the National Academy of Sciences, 106(37): 15594–15598. Subramanian, Rajesh. 2009. “Motor Vehicle Traffic Crashes as a Leading Cause of Death in the United States, 2006.” National Highway Traffic Safety Administration Traffic Safety Facts Research Note. Report No. DOT HS 811 226, Washington, DC. White, Michelle J. 2004. “The ‘arms race’ on American roads: The effect of sport utility vehicles and pickup trucks on traffic safety.” Journal of Law and Economics, 47(2): 333– 355. Wilcke, Renate A.I., Thomas Mendlik, and Andreas Goblet. 2013. “Multi-variable error correction of regional climate models.” Climatic Change, 120: 871–887. World Health Organization. 2013. WHO global status report on road safety 2013: Supporting a decade of action. World Health Organization. 32 Table 1: Summary Statistics, Accidents and Travel Demand Data Panel A. State Accident Data, 1991-2010 Census Region (States Included) Accidents Fatalities per 100,000 per 100,000 Temperature (◦ F) 5th 50th 95th Quantile Quantile Quantile (3) (4) (5) Rainfall (cm) 75th 95th Quantile Quantile (6) (7) Snowfall (cm) 75th 95th Quantile Quantile (8) (9) (1) (2) Northeast (NY, PA) 3.6 0.04 20.0 50.9 75.3 0.34 1.32 0.01 1.91 Midwest (IA, IL, KS, MI, MN, MO, NE, OH) 7.6 0.06 17.2 53.0 79.0 0.23 1.25 0.00 0.94 South (AR, FL, GA, NC, SC) 8.8 0.08 37.2 65.9 82.6 0.33 1.68 0.00 0.00 West (CA, MT, NM, WA, WY) 5.6 0.08 23.3 53.1 76.6 0.10 0.75 0.00 0.93 All 20 States 7.2 0.07 21.2 56.4 80.4 0.24 1.30 0.00 0.70 Panel B. NHTS Daily Travel Data, 1990-1991, 1995-1996, 2001-2002, and 2008-2009 Temperature (◦ F) Census Region (States Household Household 5th 50th 95th Included) Miles Trip Count Quantile Quantile Quantile (1) (2) (3) (4) (5) Rainfall (cm) 75th 95th Quantile Quantile (6) (7) Snowfall (cm) 75th 95th Quantile Quantile (8) (9) Northeast (CT, MA, ME, NH, NJ, NY, PA, RI, VT) 49.3 6.2 21.4 50.4 74.8 0.30 1.13 0.00 1.75 Midwest (IA, IL, IN, KS, MI, MN, MO, ND, NE, OH, SD, WI) 51.9 6.5 14.4 46.6 75.6 0.24 1.21 0.00 1.39 South (AL, AR, DC, DE, FL, GA, KY, LA, MD, MS, NC, OK, SC, TN, TX, VA, WV) 55.2 6.0 34.8 65.1 83.3 0.22 1.40 0.00 0.00 West (AK, AZ, CA, CO, HI, ID, MT, NM, NV, OR, UT, WA WY) 49.8 6.3 41.1 64.4 81.8 0.02 0.61 0.00 0.00 All States 52.5 6.2 25.5 60.1 82.0 0.20 1.19 0.00 0.36 Notes: Panel A details the state accident data for 20 states grouped by census region. States with available data are listed below each region. Panel A statistics are based on 6,698,935 county-by-day observations of accidents, fatalities, and weather. Columns (1) and (2) give average daily accidents and fatalities per 100,000 residents. Columns (3) through (5) give temperature, (6) and (7) rainfall, and (8) and (9) snowfall for the listed quantiles. Panel B describes the National Household Transportation Survey data grouped by census region with included states listed. The statistics detail 283,857 households and their driving behavior for a 24 hour period. Columns (1) and (2) detail the average household’s daily vehicle miles traveled, and trip count. Columns (3) through (9) detail weather statistics as in Panel A. 33 Table 2: Observed Weather and CCSM4 RCP6.0 Prediction Data Daily Temperature (◦ F) Daily Rainfall (cm) 5th 50th 95th 75th 95th Census Region Quantile Quantile Quantile Quantile Quantile (1) (2) (3) (4) (5) Panel A. Observed Weather Station Data, 2000-2009 Northeast 21.0 54.6 79.5 0.32 1.48 Midwest 14.4 51.8 78.0 0.22 1.20 South 33.4 63.6 82.7 0.30 1.54 West 20.9 51.0 76.1 0.11 0.71 All Regions 21.9 57.3 81.0 0.24 1.32 Daily Snowfall (cm) 75th 95th Quantile Quantile (6) (7) 0.00 0.00 0.00 0.00 0.00 1.43 1.04 0.00 1.12 0.55 Panel B. Baseline Predicted CCSM4, 2006-2009 Northeast 21.0 57.3 80.0 0.34 Midwest 13.7 55.2 84.3 0.23 South 32.2 65.3 84.0 0.38 West 14.6 48.2 77.7 0.17 All Regions 21.4 59.4 83.1 0.30 1.33 1.08 1.32 0.90 1.19 0.00 0.00 0.00 0.02 0.00 0.57 0.67 0.01 2.34 0.46 Panel C. Predicted Future Northeast 25.9 Midwest 20.6 South 35.9 West 22.6 All Regions 26.9 1.45 1.10 1.44 1.11 1.30 0.00 0.00 0.00 0.00 0.00 0.28 0.41 0.00 1.69 0.22 CCSM4, 2090-2099 61.2 84.2 59.2 88.9 69.3 88.1 52.2 81.8 63.3 87.5 0.37 0.23 0.40 0.24 0.31 Notes: The table details the 11,429,977 county-by-date observations used in the simulation. Each panel presents temperature, rainfall, and snowfall for the listed quantile by census region. Panel A describes the observed weather data from the National Climatic Data Center’s Global Historical Climatology Network-daily. Panels B and C describe the data from the CCSM4 RCP6.0 scenario predicting daily weather. 34 Table 3: Poisson Regression of Accidents on Weather Variables Fatalities Current + Lag Mean Temp. <20◦ F 20-30◦ F 30-40◦ F 40-50◦ F 60-70◦ F 70-80◦ F >80 ◦ F Rainfall 0-0.1 cm 0.1-0.5 cm 0.5-1.5 cm 1.5-3 cm >3 cm Snowfall 0-0.1 cm 0.1-0.5 cm 0.5-1.5 cm 1.5-3 cm >3 cm Controls Restrictions LDV Crashesa ULD Crashesb Urban Countiesc Controls Week of Lagsd County-year-monthe First Snowfallf Num. Obs. PDO Accidents Injuries (1) LDV Crashesa (2) ULD Crashesb (3) Week of Lagsd (4) Urban Countiesc (5) (6) (7) -0.143*** (0.029) -0.117*** (0.026) -0.089*** (0.019) -0.063*** (0.014) 0.055*** (0.009) 0.069*** (0.015) 0.095*** (0.015) 0.014 (0.055) 0.007 (0.026) 0.005 (0.015) -0.012 (0.016) 0.021** (0.009) 0.015 (0.014) 0.054*** (0.020) -0.605*** (0.117) -0.632*** (0.074) -0.413*** (0.077) -0.143*** (0.031) 0.128*** (0.018) 0.176*** (0.030) 0.177*** (0.027) -0.098*** (0.047) -0.112*** (0.024) -0.085*** (0.017) -0.058*** (0.013) 0.053*** (0.009) 0.064*** (0.014) 0.090*** (0.016) -0.114** (0.054) -0.128*** (0.033) -0.078** (0.034) -0.033* (0.018) 0.058*** (0.010) 0.069*** (0.018) 0.087*** (0.026) 0.097*** (0.022) 0.028* (0.017) -0.031*** (0.009) -0.023*** (0.005) -0.005 (0.005) -0.011* (0.006) -0.010 (0.008) 0.000 (0.022) -0.041*** (0.015) -0.070*** (0.007) -0.044*** (0.004) 0.020*** (0.004) 0.022*** (0.005) 0.016** (0.007) -0.029*** (0.009) -0.049*** (0.011) -0.059*** (0.015) -0.086*** (0.016) -0.048 (0.032) -0.001 (0.010) -0.006 (0.012) 0.006 (0.015) -0.040** (0.020) 0.024 (0.044) -0.080*** (0.029) -0.176*** (0.023) -0.201*** (0.035) -0.302*** (0.052) -0.227** (0.096) -0.028*** (0.010) -0.054*** (0.011) -0.062*** (0.015) -0.114*** (0.019) -0.052 (0.033) -0.027* (0.015) -0.047*** (0.015) -0.039** (0.018) -0.012 (0.020) -0.018 (0.049) 0.023*** (0.004) 0.071*** (0.005) 0.107*** (0.007) 0.147*** (0.011) 0.188*** (0.015) 0.014*** (0.005) 0.058*** (0.005) 0.088*** (0.007) 0.119*** (0.010) 0.144*** (0.017) 0.027* (0.016) 0.073*** (0.020) 0.130*** (0.028) 0.155*** (0.023) 0.040 (0.044) 0.054*** (0.020) 0.083*** (0.023) 0.158*** (0.029) 0.143*** (0.042) -0.032 (0.047) -0.017 (0.062) -0.170*** (0.060) 0.028 (0.073) -0.032 (0.104) -0.117 (0.094) 0.032* (0.017) 0.069*** (0.020) 0.143*** (0.030) 0.165*** (0.031) -0.002 (0.043) -0.002 (0.023) 0.017 (0.025) 0.070*** (0.026) 0.094** (0.040) -0.085 (0.055) 0.022*** (0.007) 0.098*** (0.010) 0.230*** (0.010) 0.354*** (0.012) 0.433*** (0.014) 0.007 (0.008) 0.078*** (0.007) 0.189*** (0.010) 0.281*** (0.010) 0.259*** (0.022) N N N Y N N N Y N N N N N N Y N N N N N N N Y Y 3,117,797 N Y Y 2,829,908 N Y Y 406,051 Y Y Y 3,117,797 N Y Y 767,735 N Y Y 6,665,499 N Y Y 5,067,561 Notes: The estimates are from a Poisson regression of the daily count of fatalities, property-damage-only (PDO) accidents, or injuries by county on weather and other covariates as indicated. Standard errors, in parentheses, are block bootstrapped by year. Reported coefficients and standard errors are the linear combination of the current and lagged estimates. Disaggregate results presented in Appendix Table H6. * significant at 10% level, ** significant at 5% level, *** significant at 1% level. a Includes only fatality counts where all participant were light-duty vehicles (also includes heavy duty vehicles). b Includes only fatality counts where one participant was an ultra-light duty mode. c Includes only counties classified as large or medium urban and suburban counties as classified by the National Center for Health Statistics 2006 Urban-Rural Classification Scheme. d Includes controls for 6 additional days of lags for each weather bin. Coefficients and standard errors include the sum of all current and lagged weather controls. e Fixed effects for county by year by month. f Indicator for first snowfall after 1 month without snow. 35 Table 4: Poisson Regression of Travel Demand Household Miles (1) Trip Count (2) Household Miles (3) Trip Count (4) Public Transit Trip Count (5) -0.076** (0.031) -0.063** (0.024) -0.011 (0.015) -0.026** (0.011) 0.001 (0.011) 0.002 (0.016) -0.015 (0.030) -0.051** (0.023) -0.037** (0.016) -0.026** (0.010) -0.003 (0.007) -0.007 (0.009) -0.009 (0.008) -0.037** (0.016) -0.760*** (0.154) -0.489*** (0.109) -0.343*** (0.088) -0.129* (0.072) 0.061 (0.051) 0.068 (0.087) 0.129 (0.106) -0.292*** (0.060) -0.144*** (0.039) -0.095*** (0.033) -0.033 (0.022) 0.027 (0.020) 0.006 (0.031) 0.021 (0.042) -0.281** (0.125) -0.199** (0.101) -0.036 (0.051) -0.066 (0.046) 0.048 (0.060) 0.028 (0.063) -0.144 (0.103) -0.004 (0.010) -0.026** (0.010) -0.015 (0.014) 0.010 (0.021) -0.023 (0.037) -0.012* (0.006) -0.005 (0.008) -0.007 (0.007) -0.031** (0.012) -0.034* (0.020) -0.115** (0.052) -0.224*** (0.046) -0.281*** (0.059) -0.571*** (0.114) -0.326 (0.238) -0.007 (0.019) -0.067*** (0.019) -0.091*** (0.018) -0.144*** (0.040) -0.087 (0.059) 0.001 (0.051) 0.027 (0.039) 0.010 (0.060) 0.036 (0.062) 0.214* (0.124) 0.026 (0.024) -0.007 (0.024) -0.014 (0.030) -0.090** (0.045) -0.156*** (0.058) -0.003 (0.015) -0.017 (0.013) -0.012 (0.020) -0.075** (0.030) -0.160*** (0.035) 0.117 (0.087) -0.057 (0.087) -0.100 (0.108) -0.065 (0.160) -0.089 (0.115) 0.025 (0.044) -0.127*** (0.037) -0.151*** (0.044) -0.105 (0.068) -0.104 (0.077) 0.068 (0.055) -0.054 (0.090) -0.155 (0.079) -0.034 (0.114) 0.005 (0.141) Y Y Y 261,667 Y Y Y 261,667 Y Y Y 228,144 Y Y Y 228,144 Y Y Y 177,987 Light-Duty Vehicles Current + Lag Mean Temp. <20◦ F 20-30◦ F 30-40◦ F 40-50◦ F 60-70◦ F 70-80◦ F >80◦ F Rainfall 0-0.1 cm 0.1-0.5 cm 0.5-1.5 cm 1.5-3 cm >3 cm Snowfall 0-0.1 cm 0.1-0.5 cm 0.5-1.5 cm 1.5-3 cm >3 cm Controls County-year-montha First snowfallb Household Controlsc Num. Obs. Ultra-Light Duty Notes: The estimates are from a Poisson regression of the daily count of trips and miles per trip for Light-Duty Vehicles, Ultra-LightDuty Modes, and Public Transit by household on weather and other covariates as indicated. Standard errors, in parentheses, are block bootstrapped by year. * significant at 10% level, ** significant at 5% level, *** significant at 1% level a Fixed effects for county by year by month. b Indicator for first snowfall after 1 month without snow. c Includes controls for count of vehicles in household, household size, number of workers in household, number of adults in household, NHTS life cycle stratum, race, NHTS defined income group, and day of week. Specifications excluding controls given in Appendix Table H16. 36 Table 5: Estimates of the Change in Accidents and the Present Discount Value Costs of Climate Change Panel A. Outcomes in 2090 Fatalities Change in 2090 due to Temperature Change in 2090 due to Rainfall Change in 2090 due to Snowfall Net Change in 2090 Net Present Cost ($2015 Million) (3) LDV Tripsb (in billions) (4) ULD Tripsb (in billions) (5) 98 316 4.33 [387, 547] [49, 469] [276, 257] -27 -2 -24 [-34, -19] [-9, 4] -45 -33 [-57, -33] Full Sample LDVa ULDa 37 Accidentsa Injuriesa (1) (2) (6) (7) 468 1.15 -10,369 7,716 [-5.83, 14.60] [0.72, 1.60] [-16,562, -3,861] [5,426, 10,094] -0.55 -0.20 7,887 2,745 [-28, -20] [-1.07, 0.13] [-0.25, -0.15] [7,431, 8,366] [2,543, 2,956] 7 2.68 0.03 -20,003 -5,552 [-43, -22] [0, 13] [1.34, 3.92] [-0.01, 0.07] [-20,692, -19,350,] [-5,893, -5,226,425] 397 [310, 476] 63 [-25, 144] 298 [259, 341] 6.45 [-3.81, 16.90] 0.98 [0.59, 1.42] -22,485 [-28,949, -15,748] 4,909 [2,399, 7,386] $360c $60c $261c -$48d -$315e -$18f $94g [$276, $438] [-$21, $140] [$225, $300] [-$151, $51] [-$171, -$472] [-$25, -$10] [$58, $130] ULD Tripsb (in billions) (5) Accidentsa Injuriesa Panel B. Outcomes from 2015 to 2099 Fatalities (1) (2) (3) LDV Tripsb (in billions) (4) (6) (7) 18,603 3,290 13,501 247.90 45.1 -1,056,915 220,672 [14,240, 22,802] [-1,149, 7,571] [11,451, 15,631] [-285.31, 787.86] [23.29, 68.19] [-1,385,121, -712,150] [95,091, 348,906] $39.9c $8.3c $27.4c -$4.7d -$34.1e -$3.0f $5.7g [$29.4, $50.1] [-$2.5, $18.8] [$22.4, $32.6] [-$18.0, $8.5] [-$14.0, -$55.1] [-$4.0, -$2.0] [$0.9, $10.5] Full Sample Sum of Net Changes 2015-2099 Sum of Costs ($2015 Billion) LDVa ULDa Notes: Net Present Cost estimates are reported in 2015 dollars. The net change estimates are the sum of county-level changes in weather on the listed outcome. All future weather simulations use quantile mapping to adjust current weather to the changes predicted by the CCSM4 RCP6.0 scenario. Values given in brackets indicate the 95% confidence interval. See text for further details of calculations. All costs assume a discount rate of 3%. a For counties with missing daily average LDV, ULD, injuries, and accidents, rather than applying a national average, we impute using a Poisson regression of the daily count of incidents for states with SDS data on county population and fatalities. b For counties without NHTS data we are missing LDV trips and ULD trips and we apply the national average. c Assumes the value of a statistical life is $9.1 million (Blincoe et al., 2014). d The cost of LDV miles is evaluated with total cost of nationwide LDV fatalities divided by the annual number of LDV miles. This cost is on average $.09 per mile. e The cost of ULD miles is evaluated with total cost of nationwide ULD fatalities divided by the number of ULD miles per household per day (0.59), as taken from the NHTS data, multiplied by 365 days and the number of households in the census. This cost is on average $3.36 per mile. f Assumes the cost per accident is $10,633 (Blincoe et al., 2014). g Assumes the value of a statistical injury of $141,677. Calculated based on the observed frequency of 5 severity levels recognized by the DOT (Blincoe et al., 2014). Table 6: 2090 Fatality Changes, Comparison of Correction Methods (1) Additive Predictedb (2) Multiplicative Predictedc (3) Additive Observedd (4) Multiplicative Observedd (5) 468 464 457 478 -1196 [387, 547] [368, 552] [370, 536] [396, 555] [-1,529, -857] -27 -297 -507 -433 -12 [-34, -19] [-375, -234] [-727, -318] [-615, -271] [-19, -4] ECDFa Change in 2090 due to Temperature Change in 2090 due to Rainfall Change in 2090 due to Snowfall Net Change in 2090 -45 406 307 59 -28 [-57, -33] 397 [310, 476] [122, 687] 573 [255, 862] [-99, 752] 253 [-232, 587] [-24, 150] 103 [-112, 308] [-45, -11] -1235 [-1,567, -891] Notes: All simulations predict the change in fatalities based on weather changes using the 2090 decade of the CCSM4 RCP6.0 scenario. Values given in brackets indicate the 95% confidence interval. See section for further details of various methodologies. a Adjusts current weather to 2090-2099 weather using our nonparametric quantile-mapping detailed in section 6.2. b Adjusts CCSM4 2090-2099 with an additive correction based on the mean difference between observed 2006-2009 data and CCSM4 2006-2009 predictions. c Adjusts CCSM4 2090 with a multiplicative correction based on the mean difference between observed 2006-2009 data and CCSM 2006-2009 predictions. d Adjusts 2006-2009 mean observed weather with an additive factor base on the change between CCSM4’s 2006-2009 and 2090-2099 predictions. e Adjusts 2006-2009 mean observed weather with a multiplicative factor base on the change between CCSM4’s 2006-2009 and 2090-2099 predictions. 38 CCSM 4 2090-2099 Observed Weather Predicted Additive Observed Additive 40 70 Smoothed Density .02 .03 .04 Smoothed Density .02 .03 .04 .05 CCSM 4 2006-2009 .05 Observed Weather 41.0° F .01 .01 40.0° F 0 0 48.0° F 0 10 20 30 40 50 60 70 80 90 100 0 10 20 30 50 60 80 90 100 Temperature Panel (c): Inverse ECDF: Quantile-based Method Panel (d): Quantile-based Correction Methods .06 Temperature 100 90 CCSM 4 2006-2009 CCSM 4 2090-2099 QM Observed Weather Quantile-based Method .05 Observed Weather 80 Smoothed Density .02 .04 .03 70 60 50 38 33 28 23 .01 Temperature Panel (b): Traditional Correction Methods .06 .06 Panel (a): PDF of Hypothetical Temperature Profile 10 0 0 0 .2 .4 .6 .8 1 0 10 20 30 40 50 60 70 80 90 100 Temperature Panel (e): Inverse ECDF: Wet Bias Uncorrected Panel (f): Inverse ECDF: Wet Bias Correction .8 .8 Probability CCSM 4 2090-2099 QM Observed Weather CCSM 4 2006-2009 CCSM 4 2090-2099 QM .2 0 0 .2 Rainfall in CM .4 Rainfall in CM .4 .6 CCSM 4 2006-2009 .6 Observed Weather .1 .15 0.1650.175 .2 .25 .3 .1 Probability .15 0.1650.175 .2 Probability Figure 1: Demonstration of Error Correction Methods 39 .25 .3 (a) Average Annual Predicted Change in Fatalities per 100,000 People (b) Average Annual Predicted Change in Injuries per 100,000 People (c) Average Annual Change in PDO Accidents per 100,000 People Figure 2: County-Level Predictions, 2090-2099 40 Appendix for Weather, Traffic Accidents, and Exposure to Climate Change Benjamin Leard and Kevin Roth∗ April 8, 2016 ∗ Leard: Resources for the Future, 1616 P St. NW, Washington, D.C. 20036, e-mail: leard@rff.org. Roth: University of California, Irvine 3297 Social Science Plaza, Irvine, CA 92697, e-mail: kroth1@uci.edu. The authors are grateful to seminar participants at UC-Irvine for helpful comments. 1 A Proof of Equation (8) In this section we present a proof of the welfare formula (8). Partially differentiating the indirect utility function (1) with respect to I yields ∂V = λ. ∂I (A.1) Partially differentiating the indirect utility function (1) with respect to W yields ∂U ∂am ∂U ∂ab ∂V = + . ∂W ∂a ∂W ∂a ∂W (A.2) Substituting (A.1) and (A.2) into Equation (7) yields ∂U ∂U dI ∗ ∂am ∂ab = − ∂a − ∂a . dW λ ∂W λ ∂W (A.3) Totally differentiating am (·) and ab (·) with respect to W and rearranging terms yields dam ∂am ∂m ∂am = − , ∂W dW ∂m ∂W dab ∂ab ∂b ∂ab = − . ∂W dW ∂b ∂W Substituting (A.4) and (A.5) into (A.3) and rearranging terms gives ∂U ∂U ∂U ∂U dI ∗ dam dab ∂am ∂m ∂ab ∂b = − ∂a − ∂a + ∂a + ∂a . dW λ dW λ dW λ ∂m ∂W λ ∂b ∂W (A.4) (A.5) (A.6) Defining semi-elasticities for miles traveled by automobile and by ULD modes with respect ∂m ∂b to weather as εm = m1 ∂W and εb = 1b ∂W , respectively, and substituting these terms into (A.6) yields Equation (8). B NHTS Data Description Our travel demand data come from the 1990, 1995, 2001, and 2009 waves of the National Household Travel Survey (NHTS). We provide an overview of these data in the main paper and further details on the data here. Although the surveys follow a similar sampling strategy, there are subtle differences. One difference is the sample size. In total, there are 22,317, 42,033, 69,817, and 150,147 usable household observations in the 1990, 1995, 2001, and 2009 surveys, respectively, for a total of 284,314 households.1 There is also some variation in the months that the surveys were administered: the 1990 survey lasted from March 1990 until March 1991; the 1995 survey lasted from May 1995 until July 1996; the 2001 survey lasted 1 Occasionally these samples are supplemented by add-on surveys requested by state and metropolitan transportation planning organizations. In the 2009 survey, the add-on partners with the largest samples included California (18,000 household target), Texas (14,342 household target), New York (14,000 household target), and Florida (14,000 household target). 2 from March 2001 through May 2002; and the 2009 survey lasted from March 2008 through May 2009. In each survey, households were randomly assigned a travel date. The NHTS balanced the variation in travel by day of the week by assigning travel days for one-seventh of the sample telephone numbers to each day of the week.2 Seventy-one percent and seventytwo percent of individuals who completed a survey filled out a travel diary in the 2001 and 2009 NHTS, respectively. The original restricted day travel data files include 149,546, 409,025, 642,292, and 1,167,321 trips in the 1990, 1995, 2001, and 2009 surveys, respectively, for a total of 2,368,184 trips. We took several measures to clean the travel diary and household data. We dropped trips with missing travel distance, start time, or end time. Some trips showed unrealistically high distances. We restricted all distances to be at most the distance implied by reported travel time moving at a speed of 100 mph. To create household measures of daily travel decisions, we aggregated the trip level data to the household level. Because the travel data are reported at the respondent level, a trip taken by two household members is reported twice, once per individual that was involved in the trip. To avoid double-counting these trips, we included only one trip per household with the same transportation mode, vehicle identification number, trip start time and miles traveled.3 This yielded 1,909,488 unique household trips.4 We then summed the number of trips, the total number of miles, and averaged the miles per trip for each household. We separate trips into three groups: light duty (e.g., automobile) trips, ULD trips, and public transit trips. For households that reported no trips of a particular type, we assign them zero total trips and miles traveled. From the initial set of 284,314 households, we drop 447 households with missing county or date information and 10 households that we could not match to weather data, yielding 283,857 household by travel day observations.5 C FARS Data Description Our alternative dataset for traffic fatalities comes from the Fatal Accidents Reporting System (FARS), a nationwide census administered by the National Highway Traffic and Safety Administration (NHTSA). It contains the universe of police reported accidents involving traffic fatalities from 1975 to 2013 and is updated annually.6 The data are collected and recorded by representatives of each state plus the District of Columbia and Puerto Rico. While FARS observations include detail about fatal accidents, 2 When a household was successfully recruited to the survey, the computer-assisted telephone interviewing system assigned the household’s travel date on the selected day of the week 10 to 14 days in the future. 3 About 10 percent of automobiles have a missing identification number. We assigned a new identification number to these vehicles based on observing that the trips taken in vehicles with missing identification numbers appeared to be non-overlapping trips with different miles traveled among all trips taken by a given household. 4 This is 129,708, 324,006, 492,444, and 963,330 unique household trips in the 1990, 1995, 2001, and 2009 surveys, respectively. 5 The households that we could not match all were located in Alaska, where a weather station was not located within 200 km of the households’ county centroid. 6 FARS was created to help NHTSA and other federal agencies evaluate the effectiveness of motor vehicle safety policies. 3 there are no records for non-fatal accidents. Relevant for our requirements, the date and county of each fatal accident are recorded. The FARS data requires little cleaning to convert the raw data to the final sample that we use for estimation. We drop a small number of observations that have missing or unidentifiable day/month/year or county information. Appendix Table H4 demonstrates that we use almost the entire original sample for estimation and that there are no years in particular where we drop a substantial number of fatal accident observations. We match fatal accidents to our weather data using county and day/month/year variables. D Compositional Changes In this section we present a model that allows us to estimate how weather changes the relative composition of the types of accidents that occur. Taking the example of accidents that are fatal, if fatalities increase under particular weather conditions, our count models in section 5 cannot tell if this increase is because total accidents increase while the fatalities per accident remain constant, or if accidents remain constant but become more deadly. We examine the possibility that weather changes the likelihood that an accident will be fatal or that a fatality will involve a particular set of circumstances using a linear probability model. D.1 Estimation Methodology for the Impact of Weather on Accident Composition First we estimate the conditional expectation of a fatality in accident i as a function of weather and the covariates. E[f atalityi |xd,c , accidenti ] = 8 X 5 X j αj Td,c + j=1 8 X j j α−1 Td−1,c + j=1 5 X j β j Rd,c + j=1 j j β−1 Rd−1,c + j=1 5 X 5 X j γ j Sd,c + j=1 j j γ−1 Sd−1,c (D.1) + θscym + z0d,c δ j=1 where f atalityi is an indicator for accident i involving a fatality. We also estimate the conditional expectation that fatality i will involve an individual traveling by a ULD mode, an intoxicated or young driver, or a single vehicle crash as a function of weather and the covariates E[attributei |xd,c , f atalityi ] = 8 X α j j Td,c + j=1 8 X j=1 j j α−1 Td−1,c + 5 X j j Rd−1,c + β−1 5 X j=1 5 X β j j Rd,c + 5 X j γ j Sd,c + j=1 (D.2) j j + θscym + z0d,c δ γ−1 Sd−1,c j=1 j=1 where attributei is an indicator for fatality i involving a particular type of crash. For both models, all covariates, including fixed effects, are identical to that of Equation (10). Because an observation is a particular accident or fatality as opposed to a count of 4 incidents, the observations are no longer aggregated by county and day but weather variables are matched at the county-day level. For these regressions we cluster standard errors at the year level. D.2 Estimation Results for the Impact of Weather on Accident Composition Table H5 presents the estimates of the impact of weather on the composition of accidents and fatalities. Column (1) estimates Equation (D.1), examining the probability that an accident is fatal as a function of temperature and precipitation. This regression helps to distinguish the possibility that fatalities increase because accidents are more severe from the possibility that they are equally severe, but accidents are more frequent. The negative and statistically significant coefficients on cold temperatures and both precipitation measures indicate that if an accident occurs, it is less likely to be deadly. This result is important because it is evidence that drivers behaviorally adjust to changes in accident risk to reduce fatality risk. Given the results in the main paper for fatalities in Table 3, column 1, it appears that when there is snowfall, accidents are less likely to be deadly, but the frequency increases enough to generate more total fatalities. One remaining puzzle is why drivers are successful in reducing the probability of a fatality for precipitation and cold but not for heat. It is possible that accidents from cold and precipitation involve different channels from heat, with cold and precipitation accidents originating from an inability to control the vehicle, while heat involves a cognitive component, but we suspect the most likely explanation is that heat increases LDV-ULD interactions and when an LDV and a pedestrian or ultra-light duty vehicle collide, there may be no behavioral adjustment that can reduce the probability of a fatality.7 Columns (2) through (6) estimate Equation (D.2) using only fatal crashes. These regressions examine the probability that a participant in a fatal accident changes with the weather. The dependent variable in column (2) is an indicator for fatal crashes involving a pedestrian, bicycle, or motorcycle. This regression indicates that these ULD modes are increasingly involved in fatal accidents as temperature increases. The maximum of point estimates across temperature bins indicates that they are 3.6% more likely to be involved in a fatal crash on a day with mean temperature of 70 to 80◦ F than a day of 50 to 60◦ . Column (3) considers the possibility that changes in drunk driving may explain our result, if, for example, people used alcohol as a coping mechanism for hot or cold weather. We find no evidence that this explains any of the temperature related fatalities.8 Column (4) tests for changes in the number of accidents involving a driver less than 21 years old. The concern is that young, inexperienced drivers may change travel patterns around weather, or that our fixed effects are not properly controlling for temporal patterns such as school attendance. We see no evidence that young drivers substantially contribute to our findings. The results in column (5) illustrate that changes in the frequency of male drivers, who are 7 We provide further robustness tests to explain the effect of temperature on fatalities in Appendix Table H11. In this table we report coefficient estimates from models where we omit accidents involving intoxication and young drivers. The results for these specifications are close to our benchmark specification. 8 There is some evidence that accidents involving alcohol decrease with snowfall. A potential explanation is that drivers avoid intoxication when they anticipate that driving will be difficult. 5 known to be at higher risk of accidents and possibly more aggressive drivers, do not appear to explain our result. Column (6) examines the possibility that our results are due to single vehicle crashes, which may indicate that weather affects a driver’s control of a vehicle. The results here indicate that this dynamic may occur around the freezing point, where we see some statistically significant effects between 20 and 40◦ F. At these temperatures melting and refreezing may cause unexpected ice, resulting in vehicles sliding off the road. We do not see any evidence that such explanations extend beyond this narrow range. In summary, Table H5 presents evidence that exposure to pedestrians, bicycles, and motorcycles is one of the key mechanisms for the temperature effect. To the extent that a psychological mechanism exists, it must act equally across age groups, genders, and the number of involved parties. E Examining the Evidence for Adaptation We return to our count model to examine the capacity for adaptation. We present adaptation results in Table H10. First we examine the possibility that technological change has weakened this relationship over time. Next we consider the possibility that there may be private or public adaptation to the local climate. There have been many safety innovations over the 20 year time period encompassed by our data including dual front airbags and anti-lock brakes. To the extent that this technology would reduce the effects of weather on fatalities, we might see evidence of a reduced gradient in 2000-2009 compared with 1990-1999. Columns (1) and (2) show that no such reduction has occurred for temperature. The number of states reporting to the state data system has grown over time and consequently the precision of our estimates is greater in the later time period, but if anything the relationship between temperature and fatalities has grown stronger.9 Increased automobile air conditioning also seems to have minimally affected this relationship. This may not be entirely surprising as ULD exposure is not affected by vehicle air conditioning. For rainfall there may be some support that these technologies have improved safety, although most changes are not statistically different between the two time periods, and the same improvement cannot be found for snowfall. Columns (3) and (4) examine the possibility that drivers or cities may be adapted to local conditions. Drivers in warmer locations may know how to avoid heat or be more attentive to features such as air conditioning in vehicles. Cities might adapt through policing, bike lanes, or insurance rates that mitigate these risks. Alternatively, residents in locations with frequent hot weather may not avoid hot days if they suspect the following day is unlikely to see improved conditions and a warm climate may facilitate choosing a bike or motorcycle over a light duty vehicle. Columns (3) and (4) show the results estimated from the coldest quartile of counties and the hottest quartile of counties. We do not find evidence that these two sets of counties have statistically different responses to temperature. If anything, the hottest quartile of counties has a larger increase in fatalities on hot days than the coldest counties. One potential explanation for this lack of evidence for adaptation is that a significant share 9 We present robustness checks for this result in Appendix Table H11 by estimating models with early time period data (1990-1995) and later time period data (2005-2010). Our results presented in Table H10 are robust to these alternative specifications. 6 of accident costs is external (Parry, Walls and Harrington, 2007), suggesting that agents will not privately choose the optimal level of adaptation. Together these results suggest that adaptation would need to take the form of either migration or dramatically new technology, such as driverless cars, that some hope will remove human error. F Comparison of Quantile-Mapping Method with Multiplicative Methods Figure H.2 demonstrates a change in temperature profile that cannot be captured by multiplicative methods. Unlike the comparable figure in the text (Figure 1), temperatures in this location increase by a uniform 10◦ F.10 Panel (a) shows that the baseline observed weather has a mean of 41◦ F, while the baseline simulated CCSM4 data has a mean of 40◦ F. With global warming the entire distribution would shift up by 10◦ F such that the new mean is 50◦ F. Crucially, the observed data have a much smaller variance than the simulated data. Panel (b) first illustrates the distribution of outcomes generated by applying a multiplicative correction factor to the predicted 2090 data using Equation (16) from the main text. Because the observed and predicted baseline data have very similar means, the correction changes the CCSM4 2090 data little under the predicted multiplicative factor. The second correction is generated by changing the baseline data with a multiplicative factor based on the change in CCSM4 between 2010 and 2090 using Equation (21) from the main text. Applying a multiplicative change to the observed baseline data increases all temperatures but to a greater degree for warm days than cold days. Mechanically, warmer days will always warm by more than 10◦ F, while cold days will warm by less. Panels (c) and (d) show the correction generated by the quantile-based method. In Panel (c) the figure shows that all points in the distribution increase by 10◦ F between the CCSM4 predictions in 2010 and 2090. This correction is then applied in Panel (d) uniformly translating the observed data up by 10◦ F. This illustrates the flexibility of the quantile-mapping method. Because the correction is non-parametric, it can capture both multiplicative and additive changes to the distribution. G Modified Quantile-Mapping Correction Method A slight variant of the quantile mapping correction method involves changing how one defines the probability p. In the original method, we define the probability p by p = ΦW,τ,c (xd,c ), (G.1) which is the cumulative probability of achieving the observed weather outcome xd,c in the observed weather data. Alternatively, we can define p as p = ΦW,τ,c (xd,c ), 10 (G.2) Because the figure in the main text was generated with a multiplicative increase, it could be captured with a multiplicative method; however, as noted in the text, multiplicative methods are generally not applied to temperature data. 7 which is the cumulative probability of achieving the observed weather outcome xd,c in the CCSM4 baseline prediction. This requires creating an ECDF for each weather variable, county, and month in the CCSM4 baseline data. The remaining steps to compute the errorcorrected prediction are identical to those described in the text. Results using this method for fatalities are given in Table H19, column (6), and generally produce similar but slightly larger numbers of lives lost than with the primary method used in the main text. 8 H Further Tables and Figures 9 Table H1: Count of Observations by Bin Panel A. State Data Total Observations Rainfall 0 cm 0-0.1 cm 0.1-0.5 cm 0.5-1.5 cm 1.5-3 cm >3 cm Snowfall 0 cm 0-0.1 cm 0.1-0.5 cm 0.5-1.5 cm 1.5-3 cm >3 cm Temperature <20◦ F 20-30◦ F 30-40◦ F 40-50◦ F 50-60◦ F 60-70◦ F 70-80◦ F >80◦ F Northeast 644,252 Midwest 3,325,021 South 1,664,548 West 1,065,114 All Regions 6,698,935 182,395 165,783 175,670 95,852 21,290 3,262 1,419,930 727,732 681,449 377,398 99,396 19,116 698,950 301,974 340,543 222,392 77,150 23,539 587,227 215,994 175,820 68,579 14,151 3,343 2,888,502 1,411,483 1,373,482 764,221 211,987 49,260 480,996 38,621 47,999 36,781 20,054 19,801 2,793,250 143,488 157,499 110,686 58,802 61,296 1,625,472 15,642 13,489 6,000 2,065 1,880 893,746 33,393 59,819 42,521 19,853 15,782 5,793,464 231,144 278,806 195,988 100,774 98,759 32,055 69,249 105,173 106,243 104,301 128,689 93,824 4,718 223,391 315,508 488,899 490,963 493,730 598,647 586,654 127,229 1,345 19,864 105,654 229,797 284,596 337,611 463,170 222,511 37,907 62,780 142,433 219,773 221,618 215,618 142,024 22,961 294,698 467,401 842,159 1,046,776 1,104,245 1,280,565 1,285,672 377,419 Northeast 180,151 Midwest 22,546 South 48,505 West 31,924 All Regions 283,126 77,081 38,057 37,613 20,911 5,128 1,361 9,039 4,933 4,930 2,873 652 119 20,529 9,747 9,666 5,935 1,907 721 24,240 2,857 2,495 1,769 458 105 130,889 55,594 54,704 31,488 8,145 2,306 153,545 7,164 7,785 5,228 2,955 3,474 17,710 1,371 1,406 1,059 507 493 47,212 601 447 176 40 29 30,342 305 584 391 169 133 248,809 9,441 10,222 6,854 3,671 4,129 6,711 12,479 22,925 26,887 29,714 35,418 32,521 13,496 2,162 2,530 3,699 3,580 3,260 3,672 3,241 402 57 639 2,770 5,475 8,042 10,979 13,514 7,029 119 205 893 4,680 9,428 8,435 6,994 1,170 9,049 15,853 30,287 40,622 50,444 58,504 56,270 22,097 Panel B. NHTS Data Total Observations Rainfall 0 cm 0-0.1 cm 0.1-0.5 cm 0.5-1.5 cm 1.5-3 cm >3 cm Snowfall 0 cm 0-0.1 cm 0.1-0.5 cm 0.5-1.5 cm 1.5-3 cm >3 cm Temperature <20◦ F 20-30◦ F 30-40◦ F 40-50◦ F 50-60◦ F 60-70◦ F 70-80◦ F >80◦ F Notes: Each observation represents a day-month-year in a given region. In our Accidents (upper panel) and NHTS (lower panel) data, we have 6,698,935 and 283,126 day-month-year observations, respectively. In our State Data, states in the Northeast include New York and Pennsylvania; states in the Midwest include Iowa, Illinois, Kansas, Michigan, Minnesota, Missouri, Nebraska, and Ohio; states in the South include Arkansas, Florida, Georgia, North Carolina, and South Carolina; states in the West include California, Montana, New Mexico, Washington, and Wyoming. In our NHTS Data, states in the Northeast include Connecticut, Massachusetts, Maine, New Hampshire, New Jersey, New York, Pennsylvania, Rhode Island, and Vermont; states in the Midwest include Iowa, Illinois, Indiana, Kansas, Michigan, Minnesota, Missouri, North Dakota, Nebraska, Ohio, South Dakota, and Wisconsin; states in the South include Alabama, Arkansas, District of Columbia, Delaware, Florida, Georgia, Kentucky, Louisiana, Maryland, Mississippi, North Carolina, Oklahoma, South Carolina, Tennessee, Texas, Virginia, and West Virginia; states in the West include Alaska, Arizona, California, Colorado, Hawaii, Idaho, Montana, New Mexico, Nevada, Oregon, Utah, Washington, and Wyoming. 10 Table H2: Statistics State Extended Accident Summary CountyDaily Day Accidents Observations per 100,000 Daily Fatalities per 100,000 Northeast New York Pennsylvania 203,794 440,458 4.67 3.05 0.03 0.04 Midwest Iowa Illinois Kansas Michigan Minnesota Missouri Nebraska Ohio 180,774 558,858 575,295 449,278 381,321 588,110 237,801 353,584 4.94 7.47 8.05 12.81 4.19 6.34 7.20 8.18 0.05 0.05 0.08 0.05 0.05 0.07 0.08 0.05 South Arkansas Florida Georgia North Carolina South Carolina 356,100 337,524 404,006 365,300 201,618 4.79 2.87 20.57 6.80 5.77 0.07 0.06 0.12 0.05 0.07 West California Montana New Mexico Washington Wyoming 286,384 286,384 241,065 170,937 80,344 4.22 6.38 6.30 4.71 7.58 0.06 0.11 0.11 0.04 0.09 Notes: States reported in this table are those that are contained in our accidents State Data. In the column titled County-Day Observations, we report the total number of county-day-month-year observations appearing in each state. Note that these observations are not conditional on having at least one accident; an observation is recorded as having no accidents if there were no reported accidents in the county on a given day. We compute Daily Accidents per 100,000 by taking the daily average of the number of accidents in a county across all years in our sample and dividing by the 2010 population size of the county and then averaging these rates over a given state. Our computed accident rate includes PDO accidents, accidents involving an injury, and accidents involving a fatality. 11 Table H3: Extended Weather Summary Statistics State Daily Temperature (◦ F) 5th 50th 95th Quantile Quantile Quantile Daily Rainfall (in cm) 75th 95th 99th Quantile Quantile Quantile Daily Snowfall (in cm) 75th 95th 99th Quantile Quantile Quantile Northeast New York Pennsylvania 17.0 21.7 49.1 51.6 74.4 75.7 0.36 0.34 1.34 1.31 2.53 2.44 0.06 0.00 2.61 1.51 7.43 5.75 Midwest Iowa Illinois Kansas Michigan Minnesota Missouri Nebraska Ohio 13.4 20.2 23.4 14.5 4.8 23.6 16.6 20.4 51.1 54.8 56.4 47.4 46.2 57.2 51.5 53.1 77.7 78.8 82.4 73.8 74.5 80.8 79.2 76.3 0.18 0.29 0.14 0.26 0.18 0.26 0.12 0.34 1.21 1.39 1.18 1.07 1.03 1.54 0.96 1.27 2.50 2.70 2.61 1.99 2.10 3.09 2.17 2.28 0.00 0.00 0.00 0.01 0.00 0.00 0.00 0.00 0.65 0.51 0.30 2.62 1.71 0.20 0.64 1.05 4.86 3.57 3.24 7.17 6.05 2.64 4.80 3.94 South Arkansas Florida Georgia North Carolina South Carolina 33.5 49.1 38.7 34.7 38.8 63.1 73.0 65.0 61.0 64.1 83.8 83.3 81.9 80.4 82.1 0.29 0.39 0.34 0.32 0.31 1.82 1.74 1.65 1.63 1.53 3.66 3.54 3.43 3.14 3.01 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.01 0.00 0.42 0.00 0.08 0.69 0.12 West California Montana New Mexico Washington Wyoming 41.1 12.3 30.1 29.6 15.8 59.5 45.0 56.3 49.5 45.7 79.1 72.6 79.0 70.8 73.6 0.04 0.09 0.07 0.29 0.07 1.10 0.49 0.52 1.14 0.43 2.71 1.17 1.24 2.36 1.02 0.00 0.00 0.00 0.00 0.02 0.31 1.59 0.51 0.86 1.78 2.29 4.79 3.49 3.34 5.14 Notes: The table displays temperature, rainfall, and snowfall for the listed quartile by state in our observed weather station data from 2000 to 2009. These statistics are derived from observed weather data from the National Climatic Data Center’s Global Historical Climatology Network-daily. 12 Table H4: FARS Sample, by Calendar Year Year 1975 1976 1977 1978 1979 1980 1981 1982 1983 1984 1985 1986 1987 1988 1989 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 Total Original Sample Observation Count 44,522 45,522 47,877 50,330 51,090 51,089 49,282 43,945 42,587 44,250 43,823 46,086 46,388 47,087 45,582 44,597 41,507 39,250 40,146 40,711 41,816 42,065 42,010 41,498 41,716 41,944 42,193 43,004 42,881 42,834 43,510 42,708 41,259 37,423 33,883 32,999 32,479 33,782 32,719 1,658,394 Observations Dropped Remaining Observations 40 3 4 3 10 7 185 15 9 15 13 9 12 10 9 13 21 8 12 17 13 22 24 17 5 5 21 12 21 18 14 0 0 1 1 0 3 1 2 595 44,482 45,519 47,873 50,327 51,080 51,082 49,097 43,930 42,578 44,235 43,810 46,077 46,376 47,077 45,573 44,584 41,486 39,242 40,134 40,694 41,803 42,043 41,986 41,481 41,711 41,939 42,172 42,992 42,860 42,816 43,496 42,708 41,259 37,422 33,882 32,999 32,476 33,781 32,717 1,657,799 Percent of Original Sample Retained 99.91 99.99 99.99 99.99 99.98 99.99 99.62 99.97 99.98 99.97 99.97 99.98 99.97 99.98 99.98 99.97 99.95 99.98 99.97 99.96 99.97 99.95 99.94 99.96 99.99 99.99 99.95 99.97 99.95 99.96 99.97 100.00 100.00 100.00 100.00 100.00 99.99 100.00 99.99 99.96 Notes: The table displays counts of fatalities and dropped observations for FARS data (see Appendix C) used in the national-level projections and auxiliary regressions in Table H8. 13 Table H5: Linear Probability Model, Compositional Changes Given an Accident Given a Fatality (1) Ultralight Duty (2) Intoxicated Driver (3) Young Driver (4) Male Driver (5) Single Vehicle (6) -0.0010*** (0.0002) -0.0007*** (0.0001) -0.0004*** (0.0001) -0.0002** (0.0001) 0.0003*** (0.0001) 0.0004*** (0.0001) 0.0006*** (0.0001) -0.119*** (0.015) -0.114*** (0.009) -0.085*** (0.007) -0.045*** (0.008) 0.023*** (0.005) 0.036*** (0.005) 0.024*** (0.008) 0.000 (0.017) 0.017 (0.012) 0.010 (0.011) 0.006 (0.009) 0.007 (0.005) 0.002 (0.008) -0.005 (0.009) 0.037 (0.013) 0.017 (0.016) 0.016 (0.010) 0.015 (0.005) -0.002 (0.005) -0.006 (0.007) -0.015 (0.009) 0.019** (0.007) 0.002 (0.009) 0.004 (0.007) 0.007 (0.005) 0.000 (0.003) 0.002 (0.004) 0.000 (0.008) 0.019 (0.015) 0.034** (0.014) 0.027*** (0.007) 0.009 (0.007) 0.000 (0.007) 0.007 (0.007) 0.005 (0.012) 0.0003*** (0.0000) -0.0006*** (0.0000) -0.0009*** (0.0001) -0.0013*** (0.0001) -0.0013*** (0.0002) -0.023*** (0.005) -0.039*** (0.005) -0.053*** (0.006) -0.074*** (0.011) -0.081*** (0.015) 0.003 (0.006) 0.004 (0.005) -0.002 (0.007) -0.017 (0.010) 0.009 (0.028) 0.003 (0.003) 0.010*** (0.003) 0.011** (0.005) 0.006 (0.007) 0.008 (0.015) -0.001 (0.004) 0.002 (0.004) -0.008 (0.006) -0.004 (0.006) 0.003 (0.018) -0.006 (0.005) -0.007 (0.006) -0.003 (0.007) -0.017* (0.010) 0.015 (0.023) 0.0001 (0.0001) -0.0001 (0.0001) -0.0005*** (0.0001) -0.0008*** (0.0001) -0.0012*** (0.0001) -0.011 (0.012) -0.001 (0.012) 0.003 (0.010) 0.019 (0.016) 0.040*** (0.012) 0.007 (0.014) -0.009 (0.015) -0.029*** (0.016) -0.087*** (0.023) -0.113*** (0.028) 0.003 (0.011) -0.008 (0.013) -0.005 (0.011) 0.000 (0.013) -0.041** (0.018) -0.003 (0.010) -0.008 (0.011) -0.004 (0.012) 0.007 (0.014) 0.033* (0.017) 0.009 (0.015) 0.015 (0.014) -0.004 (0.015) -0.092*** (0.016) -0.085*** (0.023) Fixed Effects County-year-montha Y Y First snowfallb Num. Obs. 46,570,970 Y Y 222,613 Y Y 222,613 Y Y 222,613 Y Y 222,613 Y Y 222,613 Fatality Current + Lag Mean Temp. <20 ◦ F 20-30 ◦ F 30-40 ◦ F 40-50 ◦ F 60-70 ◦ F 70-80 ◦ F >80 ◦ F Rainfall 0-0.1 cm 0.1-0.5 cm 0.5-1.5 cm 1.5-3 cm >3 cm Snowfall 0-0.1 cm 0.1-0.5 cm 0.5-1.5 cm 1.5-3 cm >3 cm Notes: The estimates are from a linear probability regression of the listed indicator on weather and other covariates. Dependent variable is a 1 where any participant had the characteristic listed and a 0 otherwise. Column 1 is the probability that a fatality occurred in any party given that an accident occurred. Columns 2 through 6 limit the sample to fatal collisions. Standard errors, in parentheses, are clustered by year. * significant at 10% level, ** significant at 5% level, *** significant at 1% level a Fixed effects for county by year by month. b Indicator for first snowfall after 1 month without snow. 14 Table H6: Estimates Poisson Regression of Accidents on Weather Variables—Full PDO Accidents Contemporaneous Temperature <20◦ F 20-30◦ F 30-40◦ F 40-50◦ F 60-70◦ F 70-80◦ F >80◦ F Rainfall 0-0.1 cm 0.1-0.5 cm 0.5-1.5 cm 1.5-3 cm >3 cm Snowfall 0-0.1 cm 0.1-0.5 cm 0.5-1.5 cm 1.5-3 cm >3 cm Lagged Temperature <20◦ F 20-30◦ F 30-40◦ F 40-50◦ F 60-70◦ F 70-80◦ F >80◦ F Rainfall 0-0.1 cm 0.1-0.5 cm 0.5-1.5 cm 1.5-3 cm >3 cm Snowfall 0-0.1 cm 0.1-0.5 cm 0.5-1.5 cm 1.5-3 cm >3 cm First Snowfall Fixed Effects State-county-year-month Num. Obs. Injuries Fatalities 0.027 0.014 -0.033*** -0.026*** 0.010** 0.010* 0.017** (0.028) (0.019) (0.009) (0.005) (0.004) (0.006) (0.007) -0.046 -0.041** -0.064*** -0.045*** 0.025*** 0.030*** 0.027*** (0.029) (0.018) (0.008) (0.005) (0.004) (0.004) (0.007) 0.028*** 0.100*** 0.162*** 0.216*** 0.271*** (0.004) (0.006) (0.008) (0.012) (0.017) 0.020*** 0.088*** 0.144*** 0.190*** 0.230*** (0.005) -0.015** (0.007) 0.007 (0.006) (0.006) -0.027*** (0.009) 0.017** (0.007) (0.009) -0.028* (0.015) 0.029*** (0.010) (0.011) -0.043** (0.017) 0.028** (0.012) (0.017) -0.022 (0.029) 0.059** (0.024) 0.023*** 0.111*** 0.265*** 0.402*** 0.483*** (0.004) (0.009) (0.008) (0.010) (0.013) 0.011** 0.086*** 0.210*** 0.316*** 0.325*** (0.005) (0.006) (0.012) (0.013) (0.017) 0.035*** 0.081*** 0.195*** 0.253*** 0.225*** (0.010) (0.011) (0.014) (0.021) (0.027) 0.070*** 0.014 0.002 0.003 -0.015*** -0.021*** -0.028*** (0.017) 0.046*** (0.018) (0.011) 0.000 (0.010) (0.006) -0.006 (0.005) (0.003) 0.002 (0.004) (0.003) -0.005 (0.003) (0.006) -0.008 (0.006) (0.007) -0.011* (0.006) 0.033 0.000 -0.004 -0.006 0.018** 0.008 0.026 (0.031) -0.014 (0.029) (0.023) -0.023 (0.021) (0.015) -0.022 (0.014) (0.014) -0.015 (0.011) (0.007) 0.019*** (0.007) (0.011) 0.005 (0.010) (0.016) 0.006 (0.016) -0.005** -0.029*** -0.056*** -0.069*** -0.083*** (0.002) (0.003) (0.003) (0.006) (0.007) -0.001 -0.012*** -0.035*** -0.048*** -0.050*** (0.005) -0.004 (0.004) -0.007 (0.005) -0.008** (0.003) -0.007 (0.007) -0.021*** (0.005) -0.065*** (0.011) -0.035*** (0.012) -0.099*** (0.013) -0.066*** (0.014) -0.185*** (0.015) -0.003 (0.015) (0.014) -0.020 (0.016) (0.020) -0.055*** (0.020) (0.020) -0.108*** (0.020) (0.030) -0.177*** (0.029) (0.049) (0.052) 0.022 Y 6,665,499 -0.006*** -0.030*** -0.056*** -0.071*** -0.086*** 0.010 Y 5,067,561 -0.175*** -0.117*** -0.085*** -0.058*** 0.037*** 0.061*** 0.069*** (0.032) (0.021) (0.014) (0.009) (0.009) (0.013) (0.024) Fatalities, Omit ULD -0.043 0.011 0.013 0.001 0.007 0.018 0.040* 0.043*** 0.065*** 0.150*** 0.174*** 0.119*** (0.029) (0.019) (0.011) (0.009) (0.008) (0.013) (0.024) (0.009) (0.012) (0.017) (0.017) (0.027) (0.002) -0.014** (0.006) -0.012* (0.007) (0.003) -0.022*** (0.005) -0.024*** (0.005) (0.004) -0.030*** (0.009) -0.032*** (0.010) (0.006) -0.043*** (0.012) -0.055*** (0.014) (0.008) -0.027 (0.026) -0.048* (0.027) (0.038) 0.093* Y 3,117,797 0.077* (0.046) Y 2,860,068 Notes: The estimates are from a Poisson regression of the daily count of PDO accidents, injuries, or fatalities by county on weather and other covariates as indicated. The reported coefficient estimates include both contemporaneous and lagged variables for each weather category. The final column reports estimates of our fatalities model with omitting fatalities involving ultra-light duty fatal accidents. Because countymonth groups with zero incidents are automatically excluded from the regression, regressions for injuries and fatalities have fewer observations. Standard errors, in parentheses, are block bootstrapped by year. * significant at 10% level, ** significant at 5% level, *** significant at 1% level 15 Table H7: Poisson Regression of Trip Counts on Weather Variables—Full Estimates Contemporaneous Temperature <20◦ F 20-30◦ F 30-40◦ F 40-50◦ F 60-70◦ F 70-80◦ F >80◦ F Rainfall 0-0.1 cm 0.1-0.5 cm 0.5-1.5 cm 1.5-3 cm >3 cm Snowfall 0-0.1 cm 0.1-0.5 cm 0.5-1.5 cm 1.5-3 cm >3 cm Lagged Temperature <20◦ F 20-30◦ F 30-40◦ F 40-50◦ F 60-70◦ F 70-80◦ F >80◦ F Rainfall 0-0.1 cm 0.1-0.5 cm 0.5-1.5 cm 1.5-3 cm >3 cm Snowfall 0-0.1 cm 0.1-0.5 cm 0.5-1.5 cm 1.5-3 cm >3 cm First Snowfall Fixed Effects State-county-year-month Num. Obs. LDV Trip Count ULD Trip Count -0.039* -0.015 -0.016 -0.011 -0.007 -0.015 -0.044*** (0.022) (0.015) (0.012) (0.007) (0.008) (0.012) (0.015) -0.333*** -0.206*** -0.116*** -0.055*** 0.041** 0.010 0.033 (0.052) (0.040) (0.031) (0.018) (0.018) (0.029) (0.043) -0.008 -0.012** -0.017*** -0.054*** -0.038 (0.005) (0.006) (0.006) (0.014) (0.024) -0.034** -0.091*** -0.129*** -0.253*** -0.168*** (0.015) (0.019) (0.026) (0.031) (0.041) 0.002 -0.005 -0.016 -0.076*** -0.154*** (0.011) 0.023 (0.035) (0.015) -0.104*** (0.037) (0.018) -0.082** (0.037) (0.024) -0.044 (0.044) (0.017) -0.083 (0.064) 0.020 0.005 -0.005 0.003 -0.003 -0.003 0.004 (0.018) (0.013) (0.013) (0.008) (0.006) (0.012) (0.014) -0.010 0.047 0.017 0.029 -0.007 -0.009 0.002 (0.049) (0.036) (0.036) (0.026) (0.019) (0.022) (0.034) -0.004 0.008 0.011 -0.007 -0.002 (0.006) (0.007) (0.008) (0.014) (0.018) 0.026 0.024 0.029 0.058* -0.024 (0.019) (0.018) (0.021) (0.031) (0.047) -0.002 -0.011 0.025 0.020 0.000 (0.009) (0.015) (0.017) (0.019) (0.026) 0.018 -0.005 -0.022 -0.046 -0.017 (0.033) (0.037) (0.030) (0.053) (0.039) -0.107** (0.048) 0.003 (0.143) Y 270,121 Y 241,577 Notes: The estimates are from a Poisson regression of the daily count of trips taken on weather and other covariates as indicated. The reported coefficient estimates include both contemporaneous and lagged variables for each weather category. The last column reports estimates of our trip count model with ultra-light duty trips only. Standard errors, in parentheses, are block bootstrapped by year. * 10% level, ** 5% level, *** 1% level 16 Table H8: Additional Specifications for Fatalities Model, FARS, Fixed Effects, and NHTS Months FARS Data Temperature quad <20◦ F quad 20-30◦ F quad 30-40◦ F quad 40-50◦ F quad 60-70◦ F quad 70-80◦ F quad >80◦ F Rainfall quad 0-0.1 cm quad 0.1-0.5 cm quad 0.5-1.5 cm quad 1.5-3 cm quad >3 cm Snowfall quad 0-0.1 cm quad 0.1-0.5 cm quad 0.5-1.5 cm quad 1.5-3 cm quad >3 cm Model Fixed Effects State FE State-county-year-month State-month, County-year Other Regressors First Snowfall First Rainfall Num. Obs. Same Months as NHTS Changing Fixed Effects -0.190*** (0.019) -0.121*** (0.013) -0.081*** (0.007) -0.051*** (0.005) 0.047*** (0.005) 0.061*** (0.006) 0.078*** (0.008) -1.217*** (0.077) -0.901*** (0.070) -0.720*** (0.045) -0.383*** (0.037) 0.067*** (0.022) 0.131*** (0.031) 0.348*** (0.042) -0.301*** (0.032) -0.182*** (0.034) -0.177*** (0.020) -0.237*** (0.019) 0.068*** (0.016) 0.162*** (0.016) 0.261*** (0.014) -0.169*** (0.027) -0.149*** (0.025) -0.106*** (0.015) -0.076*** (0.012) 0.066*** (0.009) 0.073*** (0.013) 0.097*** (0.016) -0.113** (0.055) -0.162*** (0.047) -0.161*** (0.037) -0.100*** (0.013) 0.047 (0.037) 0.057 (0.052) 0.094 (0.067) -0.017*** (0.003) -0.028*** (0.006) -0.052*** (0.007) -0.073*** (0.008) -0.021 (0.020) -0.360*** (0.026) -0.360*** (0.030) -0.323*** (0.027) -0.329*** (0.034) -0.141*** (0.050) -0.099*** (0.009) -0.149*** (0.014) -0.178*** (0.016) -0.179*** (0.017) -0.047 (0.029) -0.034*** (0.009) -0.053*** (0.011) -0.070*** (0.014) -0.104*** (0.016) -0.068** (0.030) -0.052** (0.021) -0.064*** (0.012) -0.069*** (0.014) -0.103*** (0.026) -0.095* (0.057) 0.000 (0.010) 0.038*** (0.009) 0.063*** (0.012) 0.104*** (0.015) -0.059*** (0.022) Poisson 0.287*** (0.046) 0.534*** (0.042) 0.710*** (0.064) 0.686*** (0.055) 0.474*** (0.068) Poisson 0.001 (0.022) -0.046* (0.026) -0.009 (0.039) -0.040 (0.039) -0.188*** (0.050) Poisson 0.029** (0.014) 0.069*** (0.019) 0.123*** (0.029) 0.135*** (0.026) 0.006 (0.043) Poisson 0.104*** (0.032) 0.107*** (0.033) 0.087 (0.104) 0.120*** (0.037) 0.050 (0.073) Poisson N Y N N N N Y N N N N Y N Y N Y N 19,200,000 Y N 6,698,935 Y N 6,698,935 Y N 6,321,916 Y N 489,714 Notes: The estimates are from a Poisson regression of the daily count of fatalities by county on weather and other covariates as indicated. Standard errors, in parentheses, are block bootstrapped by year. Reported coefficients and standard errors are the sum of the current and lagged estimates. The first column reports coefficient estimates from our fatalities model using FARS data for all counties in the United States from 1975 to 2013. In the next three columns we report coefficient estimates without fixed effects, with state, and finally with state-month, county-year fixed effects. * 10% level, ** 5% level, *** 1% level 17 Table H9: Additional Specifications for Fatalities Model, More Weather Bins, and Negative Binomial Model Temperature <0◦ F 0-10◦ F 10-20◦ F 20-30◦ F 30-40◦ F 40-50◦ F 60-70◦ F 70-80◦ F 80-90◦ F >90◦ F Rainfall 0-0.1 cm 0.1-0.5 cm 0.5-1.5 cm 1.5-3 cm >3 cm Snowfall 0-0.1 cm 0.1-0.5 cm 0.5-1.5 cm 1.5-3 cm >3 cm Fixed Effects Model State FE State-county-year-month State-month, County-year Other Regressors First Snowfall First Rainfall Num. Obs. More Weather Variables (1) Maximum Daily Temperature (2) -0.187** (0.088) -0.114** (0.051) -0.152*** (0.034) -0.118*** (0.026) -0.090*** (0.019) -0.063*** (0.014) 0.055*** (0.009) 0.069*** (0.015) 0.096*** (0.015) 0.066 (0.077) -0.469* (0.248) -0.051 (0.088) 0.154** (0.069) -0.080** (0.034) -0.042* (0.024) -0.028 (0.018) 0.055*** (0.014) 0.113*** (0.014) 0.140*** (0.014) 0.171*** (0.019) -0.190** (0.081) -0.118** (0.050) -0.155*** (0.034) -0.120*** (0.026) -0.091*** (0.017) -0.064*** (0.013) 0.056*** (0.009) 0.070*** (0.015) 0.096*** (0.016) 0.067 (0.073) -0.029*** (0.009) -0.049*** (0.011) -0.059*** (0.015) -0.086*** (0.016) -0.049 (0.032) -0.019** (0.009) -0.034*** (0.011) -0.038** (0.015) -0.062*** (0.016) -0.021 (0.031) -0.029*** (0.009) -0.050*** (0.012) -0.060*** (0.015) -0.087*** (0.015) -0.050 (0.037) 0.027* (0.016) 0.074*** (0.020) 0.131*** (0.029) 0.155*** (0.023) 0.042 (0.044) 0.028** (0.014) 0.070*** (0.020) 0.123*** (0.029) 0.146*** (0.024) 0.030 (0.044) 0.027* (0.016) 0.074*** (0.021) 0.132*** (0.027) 0.157*** (0.023) 0.046 (0.042) Poisson N Y N Poisson N Y N Neg. Binomial N Y N Y Y 3,117,797 Y N 3,117,797 Y N 3,117,797 Negative Binomial (3) Notes: The estimates are from a Poisson regression of the daily count of fatalities by county on weather and other covariates as indicated. Standard errors, in parentheses, are block bootstrapped by year. Reported coefficients and standard errors are the sum of the current and lagged estimates. In column 1, we disaggregate the endpoints into more temperature bins and include the first rainfall event after 1 month. The final two columns keep these additional temperature bins. Column 2 replaces average temperature with the maximum daily temperature. Column 3 reports estimates from a Negative Binomial model. * 10% level, ** 5% level, *** 1% level 18 Table H10: Poisson Regression of Fatality Count, Adaptation Current + Lag Temperature <20◦ F 20-30◦ F 30-40◦ F 40-50◦ F 60-70◦ F 70-80◦ F >80◦ F Rainfall 0-0.1 cm 0.1-0.5 cm 0.5-1.5 cm 1.5-3 cm >3 cm Snowfall 0-0.1 cm 0.1-0.5 cm 0.5-1.5 cm 1.5-3 cm >3 cm Sample Restriction Controls County-year-montha First Snowfallb Num. Obs. Time Period Changes (1) (2) Regional Sorting (3) (4) -0.089* (0.045) -0.075 (0.053) -0.029 (0.025) -0.063*** (0.018) 0.055*** (0.020) 0.068** (0.027) 0.096*** (0.034) -0.166*** (0.033) -0.136*** (0.025) -0.114*** (0.019) -0.063*** (0.017) 0.056*** (0.010) 0.070*** (0.019) 0.095*** (0.016) -0.210*** (0.046) -0.170*** (0.057) -0.113*** (0.036) -0.085* (0.046) 0.124*** (0.029) 0.130*** (0.033) 0.105 (0.084) -0.048 (0.335) -0.072 (0.078) -0.062** (0.030) -0.039** (0.016) 0.062*** (0.014) 0.088*** (0.020) 0.131*** (0.021) 0.001 (0.018) -0.005 (0.014) -0.027 (0.028) -0.051** (0.024) -0.013 (0.046) -0.042*** (0.007) -0.068*** (0.009) -0.073*** (0.015) -0.101*** (0.021) -0.066 (0.046) -0.050*** (0.022) -0.095*** (0.025) -0.116*** (0.033) -0.295*** (0.058) -0.341 (0.214) -0.036** (0.016) -0.023 (0.016) -0.016 (0.016) -0.040** (0.019) 0.008 (0.043) 0.019 (0.046) 0.048 (0.039) 0.099*** (0.031) 0.088** (0.050) 0.046 (0.071) 0.032** (0.012) 0.084*** (0.022) 0.144*** (0.033) 0.183*** (0.022) 0.036 (0.053) 0.072** (0.031) 0.163*** (0.038) 0.201*** (0.044) 0.319*** (0.054) 0.321*** (0.058) -0.005 (0.071) 0.096 (0.064) 0.076 (0.062) 0.058 (0.118) -0.259 (0.204) 1990-1999 2000-2010 County in Coldest Quartile County in Hottest Quartile Y Y 907,028 Y Y 2,210,769 Y Y 576,079 Y Y 1,040,523 Notes: The estimates are from a Poisson regression of the daily count of fatalities for a county on weather and other covariates. Standard errors, in parentheses, are block bootstrapped by year. * significant at 10% level, ** significant at 5% level, *** significant at 1% level a Fixed effects for county by year by month. b Indicator for first snowfall after 1 month without snow. 19 Table H11: Additional Specifications for Fatalities Model—Explaining Fatalities and Time Period Changes Explaining Fatalities Omit Omit UltraOmit Young light Intoxicated Driver Duty Accidents Accidents Accidents Temperature <20◦ F 20-30◦ F 30-40◦ F 40-50◦ F 60-70◦ F 70-80◦ F >80◦ F Rainfall 0-0.1 cm 0.1-0.5 cm 0.5-1.5 cm 1.5-3 cm >3 cm Snowfall 0-0.1 cm 0.1-0.5 cm 0.5-1.5 cm 1.5-3 cm >3 cm Fixed Effects State-county-yearmonth Num. Obs. Time Period Changes 1990-1999 2000-2010 1990-1995 2005-2010 -0.057** (0.027) -0.012 (0.026) -0.009 (0.016) -0.014 (0.015) 0.027*** (0.008) 0.023* (0.013) 0.046** (0.021) -0.117*** (0.044) -0.126*** (0.024) -0.083*** (0.027) -0.055*** (0.016) 0.059*** (0.011) 0.076*** (0.018) 0.106*** (0.018) -0.131*** (0.039) -0.094** (0.038) -0.109*** (0.037) -0.071*** (0.024) 0.035* (0.021) 0.033 (0.023) 0.120*** (0.043) -0.089* (0.045) -0.075 (0.053) -0.029 (0.025) -0.063*** (0.018) 0.055*** (0.020) 0.068** (0.027) 0.096*** (0.034) -0.166*** (0.033) -0.136*** (0.025) -0.114*** (0.019) -0.063*** (0.017) 0.056*** (0.010) 0.070*** (0.019) 0.095*** (0.016) -0.111 (0.122) -0.077 (0.105) -0.029 (0.077) -0.058 (0.051) 0.022 (0.034) 0.047 (0.050) 0.057 (0.105) -0.181*** (0.061) -0.107*** (0.032) -0.134*** (0.022) -0.058** (0.024) 0.081*** (0.013) 0.111*** (0.023) 0.109*** (0.034) -0.005 (0.010) -0.007 (0.009) -0.003 (0.013) -0.028* (0.016) 0.010 (0.033) -0.023*** (0.008) -0.033*** (0.012) -0.034** (0.016) -0.037 (0.026) 0.005 (0.034) -0.017 (0.023) -0.062*** (0.018) -0.041** (0.021) -0.073* (0.038) -0.102 (0.081) 0.001 (0.018) -0.005 (0.014) -0.027 (0.028) -0.051** (0.024) -0.013 (0.046) -0.042*** (0.007) -0.068*** (0.009) -0.073*** (0.015) -0.101*** (0.021) -0.066 (0.046) -0.060 (0.042) -0.039 (0.031) -0.063 (0.078) 0.011 (0.058) -0.197* (0.114) -0.052*** (0.007) -0.080*** (0.015) -0.103*** (0.024) -0.130*** (0.026) -0.132** (0.058) 0.040** (0.018) 0.045* (0.024) 0.095*** (0.031) 0.066** (0.026) -0.059 (0.043) 0.025 (0.027) 0.083** (0.036) 0.155*** (0.036) 0.246*** (0.038) 0.205*** (0.055) 0.057 (0.041) 0.032 (0.037) 0.105* (0.060) -0.046 (0.061) -0.271*** (0.063) 0.019 (0.046) 0.048 (0.039) 0.099*** (0.031) 0.088** (0.050) 0.046 (0.071) 0.032** (0.012) 0.084*** (0.022) 0.144*** (0.033) 0.183*** (0.022) 0.036 (0.053) 0.035 (0.046) 0.079 (0.046) 0.103 (0.071) 0.145 (0.103) -0.120 (0.086) 0.016 (0.012) 0.079** (0.031) 0.127** (0.055) 0.148*** (0.021) 0.034 (0.065) Y Y Y Y Y Y Y 2,860,068 1,846,940 1,046,954 907,028 2,210,769 254,147 1,093,558 Notes: The estimates are from a Poisson regression of the daily count of fatalities by county on weather and other covariates as indicated. Standard errors, in parentheses, are block bootstrapped by year. Reported coefficients and standard errors are the sum of the current and lagged estimates. Under the first three columns we restrict the sample to different subsets of fatal accidents by omitting those that share particular characteristics. Under the next group of columns, we restrict the sample to different time periods to determine whether our estimated effects of weather on fatalities change over time. The first two columns under Time Period Changes, titled 1990-1999 and 2000-2010, split the sample in half by decade. The next two columns titled 1990-1995 and 2005-2010, further restrict the split subsample into the earliest and latest six years of our full sample. * significant at 10% level, ** significant at 5% level, *** significant at 1% level 20 Table H12: Additional Specifications for PDO Accidents Model Urban Counties Changing Fixed Effects NHTS Sample Temperature <0◦ F 0-10◦ F 10-20◦ F <20◦ F 20-30◦ F 30-40◦ F 40-50◦ F 60-70◦ F 70-80◦ F >80◦ F -0.596*** (0.068) -0.455*** (0.045) -0.467*** (0.032) -0.275*** (0.033) -0.034** (0.016) -0.063*** (0.018) -0.142*** (0.035) -0.005 (0.024) 0.005 (0.022) -0.095*** (0.018) -0.188*** (0.021) -0.021* (0.012) 0.021** (0.010) 0.017 (0.015) 0.119*** (0.020) 0.029* (0.015) -0.038*** (0.008) -0.030*** (0.005) -0.004 (0.005) -0.012** (0.006) -0.021*** (0.008) 0.106*** (0.022) 0.014 (0.018) -0.039*** (0.009) -0.022*** (0.005) 0.001 (0.004) -0.002 (0.006) 0.003 (0.008) 0.063 (0.088) -0.006 (0.040) -0.020 (0.014) -0.022*** (0.006) 0.007 (0.007) 0.015 (0.019) 0.026 (0.023) 80-90◦ F >90◦ F Rainfall 0-0.1 cm 0.1-0.5 cm 0.5-1.5 cm 1.5-3 cm >3 cm Snowfall 0-0.1 cm 0.1-0.5 cm 0.5-1.5 cm 1.5-3 cm >3 cm Model State FE State-month, County-year State-county-year-month First Snowfall First Rainfall Num. Obs. More Temp. Max. Daily Bins Temp. Negative Binomial 0.403*** (0.051) 0.203*** (0.030) 0.056** (0.023) 0.516*** (0.052) 0.396*** (0.049) -0.090 (0.062) 0.208*** (0.034) 0.042 (0.028) -0.077*** (0.024) 0.034** (0.016) -0.031*** (0.009) -0.023*** (0.005) -0.005 (0.005) -0.011* (0.006) 0.077*** (0.020) 0.047*** (0.012) -0.010 (0.007) 0.013*** (0.004) 0.013*** (0.005) -0.075*** (0.016) -0.107*** (0.008) -0.061*** (0.005) 0.025*** (0.003) 0.035*** (0.004) -0.010 (0.008) -0.005 (0.008) 0.012* (0.007) 0.015** (0.007) 0.048*** (0.007) 0.076*** (0.012) -0.144*** (0.024) -0.042* (0.022) 0.039 (0.027) -0.024 (0.037) 0.053 (0.055) -0.017** (0.008) 0.030*** (0.010) 0.066*** (0.012) 0.059** (0.023) 0.194*** (0.031) 0.021*** (0.004) 0.071*** (0.004) 0.108*** (0.007) 0.148*** (0.010) 0.186*** (0.014) 0.024*** (0.005) 0.077*** (0.005) 0.120*** (0.008) 0.170*** (0.013) 0.206*** (0.019) 0.007 (0.006) 0.063*** (0.005) 0.079*** (0.005) 0.112*** (0.012) 0.161*** (0.018) 0.023*** (0.004) 0.071*** (0.005) 0.107*** (0.007) 0.147*** (0.011) 0.188*** (0.015) 0.023*** (0.004) 0.071*** (0.005) 0.107*** (0.007) 0.148*** (0.011) 0.190*** (0.015) 0.012*** (0.003) 0.049*** (0.004) 0.073*** (0.005) 0.102*** (0.008) 0.154*** (0.012) 0.227*** (0.033) 0.454*** (0.034) 0.702*** (0.048) 0.735*** (0.043) 0.687*** (0.059) -0.014 (0.028) 0.046** (0.022) 0.162*** (0.028) 0.168*** (0.035) 0.185*** (0.043) 0.020*** (0.007) 0.097*** (0.010) 0.223*** (0.010) 0.351*** (0.012) 0.446*** (0.014) 0.015* (0.008) 0.070*** (0.012) 0.177*** (0.014) 0.291*** (0.016) 0.372*** (0.014) 0.025*** (0.005) 0.086*** (0.011) 0.222*** (0.009) 0.361*** (0.017) 0.445*** (0.063) 0.022*** (0.007) 0.098*** (0.010) 0.228*** (0.010) 0.352*** (0.011) 0.436*** (0.014) 0.016** (0.007) 0.092*** (0.010) 0.220*** (0.010) 0.343*** (0.012) 0.426*** (0.014) Poisson Poisson Poisson Poisson Poisson Poisson Poisson N N N Y N 6,698,935 Y N N Y N 6,698,935 N N Y Y N 1,014,853 N N Y Y N 1,067,616 N N Y Y Y 6,665,499 N N Y Y N 6,665,499 0.012* (0.006) 0.106*** (0.008) 0.258*** (0.008) 0.396*** (0.010) 0.466*** (0.017) Neg. Binomial N N Y Y N 6,665,499 N Y N Y N 6,694,960 Notes: The estimates are from a Poisson regression of the daily count of PDO accidents by county on weather and other covariates as indicated. Standard errors, in parentheses, are block bootstrapped by year. Reported coefficients and standard errors are the sum of the current and lagged estimates. In the first three columns we report coefficient estimates without fixed effects, with state, and finally with state-month, county-year fixed effects. Next we report coefficient estimates for a model restricted to counties that are defined as urban using the classification system of the National Center for Health Statistics generated by the CDC. In the next column we restrict our sample to the county-year-months that appear in the NHTS travel diary data. Under the header More Temp. Bins, we desegregate the endpoints into more temperature bins and include the first rainfall event after 1 month. The final two columns keep these additional temperature bins. The column titled Max. Daily Temp. replaces average temperature with the daily maximum temperature. The last column of the table reports estimates from a Negative Binomial model. * 10% level, ** 5% level, *** 1% level 21 Table H13: Additional Specifications for Injuries Model Urban Counties Changing Fixed Effects NHTS Sample Temperature <0◦ F 0-10◦ F 10-20◦ F <20◦ F 20-30◦ F 30-40◦ F 40-50◦ F 60-70◦ F 70-80◦ F >80◦ F -1.230*** (0.078) -0.923*** (0.066) -0.748*** (0.047) -0.385*** (0.042) -0.009 (0.013) 0.055* (0.030) 0.213*** (0.042) -0.143*** (0.030) -0.097*** (0.017) -0.151*** (0.017) -0.245*** (0.026) 0.007 (0.013) 0.094*** (0.010) 0.125*** (0.016) -0.001 (0.018) -0.047*** (0.013) -0.070*** (0.007) -0.047*** (0.003) 0.023*** (0.003) 0.024*** (0.004) 0.010* -0.006 0.025 (0.021) -0.051*** (0.017) -0.074*** (0.007) -0.035*** (0.004) 0.020*** (0.004) 0.022*** (0.005) 0.020*** (0.007) -0.032 (0.080) -0.054 (0.038) -0.066*** (0.024) -0.040*** (0.003) 0.023*** (0.007) 0.041*** (0.013) 0.045*** (0.014) 80-90◦ F >90◦ F Rainfall 0-0.1 cm 0.1-0.5 cm 0.5-1.5 cm 1.5-3 cm >3 cm Snowfall 0-0.1 cm 0.1-0.5 cm 0.5-1.5 cm 1.5-3 cm >3 cm Model State FE State-county-year-month State-month, County-year First Snowfall First Rainfall Num. Obs. More Temp. Max. Daily Bins Temp. Negative Binomial 0.209*** (0.048) 0.071** (0.029) -0.029 (0.022) 0.248*** (0.059) 0.235*** (0.047) -0.059 (0.053) 0.098** (0.041) -0.006 (0.033) -0.092*** (0.025) -0.038** (0.015) -0.070*** (0.007) -0.044*** (0.004) 0.020*** (0.004) 0.023*** (0.005) 0.018 (0.019) -0.001 (0.014) -0.023*** (0.006) 0.034*** (0.003) 0.057*** (0.005) -0.089*** (0.015) -0.107*** (0.008) -0.067*** (0.004) 0.036*** (0.003) 0.048*** (0.004) 0.017** (0.007) 0.018 (0.013) 0.065*** (0.007) 0.063*** (0.006) 0.049*** (0.006) 0.067*** (0.015) -0.316*** (0.029) -0.225*** (0.030) -0.092*** (0.034) 0.024 (0.049) 0.383*** (0.062) -0.098*** (0.010) -0.083*** (0.012) -0.033** (0.013) 0.053** (0.024) 0.287*** (0.034) 0.010*** (0.004) 0.050*** (0.004) 0.077*** (0.005) 0.104*** (0.009) 0.129*** (0.015) 0.015*** (0.005) 0.060*** (0.005) 0.101*** (0.008) 0.138*** (0.012) 0.157*** (0.020) -0.001 (0.010) 0.041*** (0.011) 0.063*** (0.009) 0.091*** (0.010) 0.136*** (0.021) 0.014*** (0.005) 0.058*** (0.005) 0.088*** (0.007) 0.119*** (0.010) 0.144*** (0.017) 0.018*** (0.005) 0.063*** (0.005) 0.095*** (0.007) 0.128*** (0.010) 0.156*** (0.017) 0.008** (0.004) 0.047*** (0.004) 0.068*** (0.006) 0.092*** (0.009) 0.120*** (0.015) 0.254*** (0.065) 0.458*** (0.041) 0.655*** (0.046) 0.763*** (0.028) 0.757*** (0.052) -0.037 (0.026) -0.102*** (0.025) -0.082** (0.039) -0.018 (0.044) 0.061 (0.047) -0.001 (0.006) 0.066*** (0.007) 0.160*** (0.010) 0.236*** (0.013) 0.211*** (0.020) -0.002 (0.008) 0.052*** (0.008) 0.140*** (0.013) 0.203*** (0.013) 0.180*** (0.021) 0.025*** (0.005) 0.075*** (0.014) 0.207*** (0.007) 0.294*** (0.023) 0.307*** (0.085) 0.007 (0.008) 0.078*** (0.007) 0.189*** (0.010) 0.281*** (0.010) 0.260*** (0.022) 0.005 (0.007) 0.074*** (0.007) 0.183*** (0.010) 0.273*** (0.011) 0.252*** (0.022) Poisson Poisson Poisson Poisson Poisson Poisson Poisson N N N Y N 5,489,551 Y N N Y N 5,489,551 N N Y Y N 5,247,345 N Y N Y N 810,928 N Y N Y N 807,775 N Y N Y Y 5,067,561 N Y N Y N 5,067,561 0.008 (0.007) 0.086*** (0.007) 0.211*** (0.009) 0.315*** (0.010) 0.302*** (0.022) Neg. Binomial N Y N Y N 5,067,561 Notes: The estimates are from a Poisson regression of the daily count of injuries by county on weather and other covariates as indicated. Standard errors, in parentheses, are block bootstrapped by year. Reported coefficients and standard errors are the sum of the current and lagged estimates. In the first three columns we report coefficient estimates without fixed effects, with state, and finally with state-month, county-year fixed effects. Next we report coefficient estimates for a model restricted to counties that are defined as urban using the classification system of the National Center for Health Statistics generated by the CDC. In the next column we restrict our sample the county-year-months that appear in the NHTS travel diary data. Under the header More Temp. Bins, we desegregate the endpoints into more temperature bins and include the first rainfall event after 1 month. The final two columns keep these additional temperature bins. The column titled Max. Daily Temp. replaces average temperature with the daily maximum temperature. The last column of the table reports estimates from a Negative Binomial model. * 10% level, ** 5% level, *** 1% level 22 Table H14: Aggregation of Data to Monthly Level Temperature <20◦ F 1-day effect 20-30◦ F 1-day effect 30-40◦ F 1-day effect 40-50◦ F 1-day effect 60-70◦ F 1-day effect 70-80◦ F 1-day effect >80◦ F 1-day effect Rainfall 0-0.1 cm 1-day effect 0.1-0.5 cm 1-day effect 0.5-1.5 cm 1-day effect 1.5-3 cm 1-day effect >3 cm 1-day effect Snowfall 0-0.1 cm 1-day effect 0.1-0.5 cm 1-day effect 0.5-1.5 cm 1-day effect 1.5-3 cm 1-day effect >3 cm 1-day effect Fixed Effects State-month, county-year Num. Obs. PDO Accidents Injuries Fatalities 0.004 (0.004) 0.120 -0.006 (0.005) -0.180 -0.002 (0.002) -0.060 -0.005* (0.002) -0.150 -0.001 (0.001) -0.030 0.000 (0.002) 0.000 -0.001 (0.002) -0.030 0.004 (0.005) 0.120 -0.009 (0.005) -0.270 0.001 (0.003) 0.030 -0.004** (0.002) -0.120 -0.001 (0.002) -0.030 -0.002 (0.003) -0.060 -0.002 (0.003) -0.060 -0.010** (0.004) -0.300 -0.012*** (0.003) -0.360 -0.005** (0.003) -0.150 -0.007*** (0.002) -0.210 0.001 (0.001) 0.030 0.003* (0.001) 0.090 0.005*** (0.002) 0.15 0.002 (0.002) 0.060 0.002* (0.001) 0.060 0.004* (0.002) 0.120 0.002 (0.003) 0.060 -0.001 (0.005) -0.030 -0.002 (0.002) -0.060 -0.002 (0.002) -0.060 0.001 (0.002) 0.030 0.000 (0.004) 0.000 0.006 (0.006) 0.180 -0.002 (0.002) -0.060 -0.006** (0.003) -0.180 -0.010*** (0.003) -0.300 -0.008** (0.004) -0.240 -0.017*** (0.006) -0.510 -0.004 (0.003) -0.120 0.004 (0.004) 0.120 0.008*** (0.003) 0.240 0.006 (0.006) 0.180 0.015*** (0.005) 0.450 -0.007* (0.004) -0.210 0.001 (0.003) 0.030 0.006* (0.003) 0.180 0.010** (0.005) 0.300 0.013* (0.007) 0.390 -0.003 (0.003) -0.090 0.001 (0.003) 0.030 0.009** (0.004) 0.270 -0.003 (0.005) -0.090 -0.011 (0.009) -0.330 Y 218,607 Y 179,111 Y 218,524 Notes: The estimates are from a Poisson regression of the month count of fatalities, PDO accidents, and injuries by county on weather and other covariates as indicated. Standard errors, in parentheses, are block bootstrapped by year. Reported coefficients and standard errors are the sum of the current and lagged estimates. * 10% level, ** 5% level, *** 1% level 23 Table H15: Travel Demand Robustness: Inverse Hyperbolic Sine Inverse Hyperbolic Sine LDV Trip LDV ULD Trip ULD Count Miles Count Miles Temperature <20◦ F 20-30◦ F 30-40◦ F 40-50◦ F 60-70◦ F 70-80◦ F >80◦ F Rainfall 0-0.1 cm 0.1-0.5 cm 0.5-1.5 cm 1.5-3 cm >3 cm Snowfall 0-0.1 cm 0.1-0.5 cm 0.5-1.5 cm 1.5-3 cm >3 cm Fixed Effects State-month, county-year State-county-year-month Num. Obs. -0.069** -0.117*** -0.078*** -0.087*** (0.030) (0.042) (0.018) (0.016) -0.042** -0.070** -0.048*** -0.063*** (0.021) (0.035) (0.013) (0.012) -0.031** -0.030 -0.035*** -0.040*** (0.015) (0.025) (0.009) (0.008) -0.008 -0.012 -0.011* -0.016*** (0.010) (0.017) (0.006) (0.006) -0.014 -0.003 0.014** 0.012* (0.012) (0.018) (0.007) (0.007) -0.026** -0.025 0.008 0.006 (0.013) (0.021) (0.011) (0.012) -0.065*** -0.098*** 0.004 0.009 (0.021) (0.036) (0.015) (0.014) -0.010 (0.009) -0.008 (0.012) -0.004 (0.010) -0.045*** (0.013) -0.055** (0.028) -0.010 (0.016) -0.027 (0.021) -0.018 (0.020) -0.059** (0.024) -0.077 (0.047) 0.000 (0.007) -0.020*** (0.007) -0.031*** (0.007) -0.045*** (0.013) -0.041** (0.018) Weighted Inverse Hyperbolic Sinea LDV Trip LDV ULD Trip ULD Count Miles Count Miles -0.089* (0.046) -0.072** (0.034) -0.050** (0.024) -0.004 (0.014) -0.010 (0.019) -0.019 (0.020) -0.043 (0.028) -0.136** (0.062) -0.099* (0.051) -0.052 (0.042) -0.010 (0.024) 0.015 (0.033) 0.006 (0.036) -0.035 (0.053) -0.058*** -0.070*** (0.020) (0.019) -0.034** -0.050*** (0.014) (0.015) -0.019* -0.026*** (0.011) (0.009) -0.008 -0.010 (0.010) (0.007) 0.015 0.015 (0.011) (0.010) 0.004 0.002 (0.011) (0.010) -0.003 -0.006 (0.016) (0.014) -0.005 -0.009 0.003 0.006 0.001 (0.006) (0.013) (0.020) (0.007) (0.007) -0.029*** -0.001 -0.012 -0.011 -0.021*** (0.007) (0.015) (0.026) (0.008) (0.007) -0.041*** 0.014 0.017 -0.021** -0.033*** (0.006) (0.017) (0.032) (0.008) (0.007) -0.065*** -0.090*** -0.121*** -0.053*** -0.066*** (0.009) (0.026) (0.044) (0.013) (0.012) -0.042*** -0.052 -0.103 -0.015 -0.017 (0.016) (0.042) (0.071) (0.036) (0.030) -0.017 -0.018 0.001 -0.009 -0.043 -0.066 -0.004 -0.015 (0.021) (0.034) (0.014) (0.012) (0.028) (0.044) (0.015) (0.013) -0.015 -0.021 -0.038*** -0.025*** -0.027 -0.049 -0.046*** -0.032*** (0.019) (0.035) (0.011) (0.008) (0.037) (0.060) (0.014) (0.011) -0.025 -0.063** -0.040*** -0.038*** -0.001 -0.035 -0.033** -0.041*** (0.026) (0.030) (0.012) (0.011) (0.037) (0.048) (0.014) (0.012) -0.105** -0.168** -0.031* -0.034* -0.109** -0.182** -0.017 -0.017 (0.043) (0.067) (0.018) (0.019) (0.045) (0.076) (0.014) (0.020) -0.223*** -0.377*** -0.030 -0.022 -0.245*** -0.423*** -0.025 -0.019 (0.045) (0.074) (0.025) (0.019) (0.062) (0.104) (0.016) (0.013) N Y 283,657 N Y 283,657 N Y 283,657 N Y 283,657 N Y 283,657 N Y 283,657 N Y 283,657 N Y 283,657 Notes: The estimates are from Inverse Hyperbolic Sine of the daily count of trips taken, total miles, ultra-light duty trip count, and ultra-light duty miles on weather. Other covariates include household size, number of vehicles, number of workers, number of adults, family structure, race, income, day of week, and fixed effects as indicated. Inverse Hyperbolic Sine has been suggested as a transformation for a skewed dependent variable that is defined for zero but approaches the log transformation for larger values (Burbidge, Magee and Robb, 1988). The reported coefficient estimates include both contemporaneous and lagged variables for each weather category. Standard errors, in parentheses, are block bootstrapped by year. Reported coefficients and standard errors are the sum of the current and lagged estimates. In the first four columns we report coefficient estimates of unweighted Inverse Hyperbolic Sine models. In the last four columns we report coefficient estimates of weighted Inverse Hyperbolic Sine models, where observations are weighted by the inverse frequency of households in each survey wave. Since the NHTS significantly increases in size over time–from about 20,000 households in 1990 to about 150,000 households in 2009–the weights for households in the earlier surveys are much larger than the weights for households in the later surveys. * significant at 10% level, ** significant at 5% level, *** significant at 1% level. a Weighted based on survey size to account for the fact that earlier surveys are smaller. 24 Table H16: Poisson Regression of Travel Demand without Household Controls Trip Count (1) -0.047* (0.028) -0.029* (0.017) -0.025* (0.013) -0.006 (0.008) -0.009 (0.010) -0.017 (0.011) -0.035* (0.020) Miles per Trip (2) -0.017 (0.070) -0.039 (0.051) 0.011 (0.027) -0.051*** (0.017) 0.033* (0.020) 0.043 (0.034) 0.022 (0.054) Trip Count (3) -0.254*** (0.065) -0.112** (0.043) -0.079** (0.032) -0.029 (0.025) 0.025 (0.021) -0.009 (0.033) 0.017 (0.047) Miles per Trip (4) -0.730*** (0.159) -0.456*** (0.119) -0.333*** (0.096) -0.127 (0.088) 0.074 (0.053) 0.100 (0.082) 0.121 (0.110) Public Transit Trip Count (5) -0.207 (0.134) -0.126 (0.092) 0.010 (0.046) 0.002 (0.042) 0.061 (0.053) -0.015 (0.076) -0.072 (0.134) -0.013 (0.010) -0.005 (0.009) -0.002 (0.011) -0.058*** (0.019) -0.039 (0.033) 0.002 (0.017) -0.051*** (0.017) -0.012 (0.029) 0.092** (0.041) 0.061 (0.100) -0.012 (0.020) -0.064*** (0.018) -0.089*** (0.021) -0.169*** (0.037) -0.134** (0.060) -0.090* (0.048) -0.226*** (0.048) -0.291*** (0.068) -0.538*** (0.112) -0.363 (0.239) -0.054 (0.053) -0.008 (0.056) 0.001 (0.057) -0.214*** (0.078) 0.000 (0.168) -0.009 (0.015) -0.028 (0.022) -0.008 (0.024) -0.071** (0.032) -0.169*** (0.038) 0.030 (0.047) 0.017 (0.043) -0.030 (0.062) 0.047 (0.104) 0.061 (0.088) 0.034 (0.044) -0.126*** (0.036) -0.144*** (0.042) -0.095 (0.060) -0.099 (0.064) 0.078 (0.078) -0.078 (0.089) -0.159 (0.104) -0.090 (0.165) -0.049 (0.110) 0.121* (0.071) -0.007 (0.103) -0.011 (0.088) -0.007 (0.114) -0.017 (0.107) Y Y N 261,748 Y Y N 223,766 Y Y N 228,712 Y Y N 228,210 Y Y N 177,659 Light-Duty Vehicles Current + Lag <20 ◦ F 20-30 ◦ F 30-40 ◦ F 40-50 ◦ F 60-70 ◦ F 70-80 ◦ F >80 ◦ F Rainfall 0-0.1 cm 0.1-0.5 cm 0.5-1.5 cm 1.5-3 cm >3 cm Snowfall 0-0.1 cm 0.1-0.5 cm 0.5-1.5 cm 1.5-3 cm >3 cm Controls County-year-montha First Snowfallb Household Controlsc Num. Obs. Ultra-Light Duty Notes: The estimates are from a Poisson regression of the daily count of trips and miles per trip for Light-Duty Vehicles, Ultra-Light Duty Modes, and Public Transit by household on weather and other covariates as indicated. Standard errors, in parentheses, are block bootstrapped by year. * significant at 10% level, ** significant at 5% level, *** significant at 1% level a Fixed effects for county by year by month. b Indicator for first snowfall after 1 month without snow. c Excludes controls for count of vehicles in household, household size, number of workers in household, number of adults in household, NHTS life-cycle stratum, race, NHTS defined income group, and day of week. 25 Table H17: Extended Poisson Regression of Travel Demand with Household Controls Current + Lag <20 ◦ F 20-30 ◦ F 30-40 ◦ F 40-50 ◦ F 60-70 ◦ F 70-80 ◦ F >80 ◦ F Rainfall 0-0.1 cm 0.1-0.5 cm 0.5-1.5 cm 1.5-3 cm >3 cm Snowfall 0-0.1 cm 0.1-0.5 cm 0.5-1.5 cm 1.5-3 cm >3 cm Controls County-year-montha First Snowfallb Household Controlsc Num. Obs. LDV Total Miles (1) -0.076** (0.031) -0.063** (0.024) -0.011 (0.015) -0.026** (0.011) 0.001 (0.011) 0.002 (0.016) -0.015 (0.030) (2) 0.008 (0.018) 0.017 (0.011) 0.021*** (0.006) 0.005 (0.006) 0.002 (0.005) 0.007 (0.006) -0.001 (0.012) LDV Travel Time (3) -0.053** (0.021) -0.055*** (0.019) -0.026** (0.012) -0.018** (0.008) 0.001 (0.008) -0.005 (0.011) -0.027 (0.022) ULD Total Miles (4) -0.760*** (0.154) -0.489*** (0.109) -0.343*** (0.088) -0.129* (0.072) 0.061 (0.051) 0.068 (0.087) 0.129 (0.106) Public Transit Miles (5) -0.122 (0.224) -0.173 (0.170) 0.059 (0.106) -0.018 (0.064) 0.101 (0.075) 0.118 (0.097) -0.073 (0.246) -0.004 (0.010) -0.026** (0.010) -0.015 (0.014) 0.010 (0.021) -0.023 (0.037) 0.000 (0.005) 0.000 (0.004) -0.003 (0.007) 0.029*** (0.008) -0.008 (0.018) -0.003 (0.006) -0.018 (0.009) -0.006 (0.008) -0.016 (0.016) 0.000 (0.027) -0.115** (0.052) -0.224*** (0.046) -0.281*** (0.059) -0.571*** (0.114) -0.326 (0.238) -0.024 (0.099) 0.018 (0.075) -0.030 (0.094) 0.208* (0.115) -0.246 (0.197) 0.026 (0.024) -0.007 (0.024) -0.014 (0.030) -0.090** (0.045) -0.156*** (0.058) 0.000 (0.009) -0.005 (0.014) -0.006 (0.013) -0.030* (0.017) -0.067*** (0.018) 0.015 (0.017) -0.001 (0.019) -0.014 (0.019) -0.058 (0.040) -0.105*** (0.040) 0.117 (0.087) -0.057 (0.087) -0.100 (0.108) -0.065 (0.160) -0.089 (0.115) 0.170 (0.114) -0.079 (0.142) -0.464** (0.215) 0.004 (0.210) 0.101 (0.287) Y Y Y 261,667 Y Y Y 223,676 Y Y Y 261,627 Y Y Y 228,144 Y Y Y 177,987 LDV Speed Notes: The estimates are from a Poisson regression of the daily count of trips and miles per trip for Light-Duty Vehicles, Ultra-Light Duty Modes, and Public Transit by household on weather and other covariates as indicated. Standard errors, in parentheses, are block bootstrapped by year. * significant at 10% level, ** significant at 5% level, *** significant at 1% level a Fixed effects for county by year by month. b Indicator for first snowfall after 1 month without snow. c Includes controls for count of vehicles in household, household size, number of workers in household, number of adults in household, NHTS lifecycle stratum, race, NHTS defined income group, and day of week. 26 Table H18: Change in Fatalities with Climate Change: Intermediate Dates Using CCSM4 RCP6.0 Fatalities 2045 (2) 2025 (1) Annual Change due to Temperature Annual Change due to Rainfall Annual Change due to Snowfall Annual Net Change Net Present Cost ($2015 Million) Annual Change due to Temperature Annual Change due to Rainfall Annual Change due to Snowfall Annual Net Change Net Present Cost ($2015 Million) Annual Change due to Temperature Annual Change due to Rainfall Annual Change due to Snowfall Annual Net Change Net Present Cost ($2015 Million) 2065 (3) 73 220 383 [54, 91] 10 [4, 17] -8 [-11, -64] 76 [56, 95] [173, 263] -3 [-7, 3] -24 [-31, -18] 192 [147, 238] [319, 442] 0 [-4, 5] -34 [-41, -25] 349 [285, 410] $511.37 $718.36 $724.43 2025 (1) LDV Fatalities 2045 (2) 2065 (3) 17 51 76 [-4, 36] -1 [-6, 6] -3 [-6, 1] 14 [-8, 34] [6, 96] 1 [-3, 5] -18 [-24, -12] 34 [-15, 80] [17, 135] -1 [-5, 4] -26 [-33, -19] 49 [-13, 109] $92.20 $127.56 $101.78 2025 (1) ULD Fatalities 2045 (2) 2065 (3) 48 149 261 [38, 59] -3 [-2, 7] 1 [-1, 3] 51 [40, 63] [117, 171] -6 [-9, -3] 3 [-1, 7] 146 [124, 168] [230, 291] 0 [-3, 3] 4 [0, 6] 265 [235, 298] $347.42 $545 $550 Notes: The estimates are the sum of county-level changes in weather on the listed outcome. All future weather simulations use quantile-mapping to adjust current weather to the changes predicted by the CCSM4 RCP6.0 scenario. Values given in brackets indicate the 95% confidence interval. 27 Table H19: Fatality Change Using Prior Correction Methods 2090-2099 Temperature Rainfall Snowfall Additive Predicted (1) 464 [368, 552] -297 [-375, -234] 406 [122, 687] Multiplicative Predicted (2) 457 [370, 536] -507 [-727, -318] 307 [-99, 752] Additive Observed (3) 478 [396, 555] -433 [-615, -271] 59 [-24, 150] Multiplicative Observed (4) -1196 [-1,529, -857] -12 [-19, -4] -28 [-45, -11] ECDF 1 ECDF 2 (5) 468 [387, 547] -27 [-34, -19] -45 [-57, -33] (6) 485 [416, 581] -102 [-142, -55] -55 [-81, -30] 573 253 103 -1235 397 340 [255, 862] [-232, 587] [-112, 308] [-1,567, -891] [310, 476] [265, 439] Annual Net Change Notes: The estimates are the sum of county level changes in weather on the listed outcome. All future weather simulations use quantile-mapping to adjust current weather to the changes predicted by CCSM4 RCP6.0 scenario. Values given in brackets indicate the 95% confidence interval. The Additive Predicted model adjusts the 2090 CCSM4 prediction data using an additive correction of the baseline discrepancy. The Multiplicative Predicted model adjusts the 2090 CCSM4 prediction data using a multiplicative correction of the baseline discrepancy. The Additive Observed method adds the mean shift in weather between the baseline and future CCSM4 to the current observed weather. The Multiplicative Observed method adjusts the current observed weather through multiplication of a common term that matches the baseline CCSM4 to the future CCSM4 data. ECDF 1 repeats Table 6 column 1, which is the Empirical Cumulative Distribution Function outlined in the main text. ECDF 2 uses the Empirical Cumulative Distribution Function outlined in Appendix G. See text for further details of calculations. Table H20: Predicted Changes in Accidents and Net Present Costs under Various Future Weather Scenarios CCSM 4.5 (1) 353 [282, 420] CCSM 6 (2) 397 [310, 476] CCSM 8.5 (3) 701 [541, 857] Hadley 3 A1B (4) 603 [402, 796] Present Cost ($2015 Mil.) $302 [$241, $359] $339 [$265, $408] $695 [$537, $850] $511.49 [$344, $681] Sum of Changes 2015-2099 18,115 [13,910, 22,342] 18,603 [14,240, 22,802] 33,376 [25,419, 41,225] 27,493 [17,964, 37,179] Sum of Costs ($2015 Mil.) $44,743 [$33,982, $55,484] $39,882 [$29,409, $50,134] $73,482 [$55,569, $91,262] $62,603 [$38,688, $86,883] Net Change in 2090 Notes: This table shows net present cost estimates for fatal accidents under the various levels of warming and for various climate model scenarios and models. For each traffic accident category, we report predicted changes in annual accident frequency stemming from changes in temperature, rain and snow caused by climate change in 2090 and summed between 2015 and 2099. The Sum of Costs estimates in the bottom row are the discounted value of costs from 2010 to 2099, using a discount rate of 3 percent and a VSL of $9,100,000 (Blincoe et al., 2014). 28 (a) Average Annual Predicted Change in Fatalities per 100,000 People (b) Average Annual Predicted Change in Injuries per 100,000 People (c) Average Annual Change in PDO Accidents per 100,000 People Figure H.1: County-Level Predictions, 2090-2099, under the RCP8.0 Future Weather Scenario 29 .06 Panel (b): Traditional Correction Methods .06 Panel (a): PDF of Hypothetical Temperature Profile CCSM 4 2090-2099 Observed Weather Observed Multiplicative Smoothed Density .03 .04 .02 Smoothed Density .02 .03 .04 41.0° F 0 .01 50.0° F 0 .01 40.0° F 10 20 30 40 50 60 70 80 90 100 0 10 20 30 40 50 60 70 80 90 Temperature Temperature Panel (c): Inverse ECDF: Quantile-based Method Panel (d): Quantile-based Correction Methods 100 .06 0 100 90 CCSM 4 2006-2009 CCSM 4 2090-2099 QM Observed Weather Quantile-based Method .05 Observed Weather 80 Smoothed Density .02 .03 .04 70 60 50 40 30 .01 20 10 0 0 Temperature Predicted Multiplicative .05 CCSM 4 2006-2009 .05 Observed Weather 0 .2 .4 .6 .8 1 0 Probability 10 20 30 40 50 60 Temperature Figure H.2: Demonstration of Error Correction Methods 30 70 80 90 100 Appendix References Blincoe, Lawrence, Ted R. Miller, Eduard Zaloshnja, and Bruce A. Lawrence. 2014. “The economic and societal impact of motor vehicle crashes, 2010.” National Highway Traffic Safety Administration, Washington, DC. Burbidge, John B., Lonnie Magee, and A. Leslie Robb. 1988. “Alternative transformations to handle extreme values of the dependent variable.” Journal of the American Statistical Association, 83(401): 123–127. Parry, Ian, Margaret Walls, and Winston Harrington. 2007. “Automobile externalities and policies.” Journal of Economic Literature, 373–399. 31