Access the full text.
Sign up today, get DeepDyve free for 14 days.
Background: Autonomous vehicles (AVs) will radically re-shape the health and well-being of people in the United States in good ways and bad. We set out to estimate a reasonable time-to-adoption using cost-effectivenessmodels to estimate the point at which AVs become reasonably safe and affordable for widespread adoption. Methods: We used Waymo data (previously, Google Self-Driving Car Project) and a microsimulation model to explore projected costs and safety issues today and five years from today to get a sense of the speed of consumer adoption were AVs brought to the market. Results: The adoption of AVs for private use was associated with an ICER of 1,396,110/QALY gained today, a figure that would decline to 173,890/QALY gained 5-years in the future. However, AV taxis are both less expensive and potentially already safer than human-piloted taxis. Conclusions: While AVs are not unlikely to be used a family vehicles any time soon, it would make economic sense to adopt them as taxis today. Legislation enhancing the benefits while mitigating the potential harmful health impacts of AV taxis is needed with some urgency. Background Without policies to foster their benefits and mitigate Autonomous Vehicles (AVs) are presently on the road their harms, it is at least conceivable that society could with no human behind the wheel in some localities end up with a public health threat equal in magnitude to (Energy and Commerce Committee 2017; U.S. Depart- the one posed by the introduction of private cars over a ment of Transportation 2016; State of California Depart- century ago. At that time, policymakers did not foresee ment of Motor Vehicles 2018). These “level 4” vehicles are the automobile as a leading cause of death globally. Rather, confined to certain conditions and roads, but could plaus- they focused on licensure and ownership, overlooking its ibly be introduced more broadly in the coming year. The potential impact on injuries, fatalities, a sedentary lifestyle, Self Drive Act (HR3388) is designed to override and unify obesity, and air pollution (McGinnis and Foege 1993). state regulations, with the ultimate goal of expediting the While some level 4 AVs have been mixed with HPVs, transition to a future without human piloted vehicles they are not presently on the market for sale (Kang (HPVs). Federal lawmakers so far are nearly exclusively fo- 2017). We undertook this analysis because cohesive, cused on short-term issues, such as safety, licensing, and in- well-informed legislation is needed that considers both surance issues (Energy and Commerce Committee 2017). the short-term public health benefits and the longer- However, the potential benefits and harms of AVs term social consequences of AVs (Business Insider extend beyond their immediate safety on the road or Insider Reports 2015; KPMG 2013). The longer-term reduced insurance costs. AVs have the potential to trans- consequences will require time to study and regulate. form our lives in ways unseen since the introduction of We do not evaluate these potential long-term challenges. automobiles themselves well over 100 years ago. Rather, in this study, we simply ask whether it is plausible that AVs are affordable and safe in their current state today and five years from today. If AVs are * Correspondence: firstname.lastname@example.org; email@example.com presently unaffordable or are dangerous, they are Yale School of Medicine, New Haven, USA Global Research Analytics for Population Health, Columbia University unlikely to be widely adopted by consumers irrespective Mailman School of Public Health, New York City, USA of whether laws permit their use. Under these Full list of author information is available at the end of the article © The Author(s). 2018 Open Access This article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. Freedman et al. Injury Epidemiology (2018) 5:24 Page 2 of 8 circumstances, governments may debate and study the vehicles have logged over 2 million miles, with billions unintended consequences of introducing driverless cars of simulated miles (Waymo 2017). These data allow for on our roads. As the technology improves and becomes microsimulation modeling of future crashes, costs, and cheaper, laws can be gradually put into place that health effects. We further consider multiple alternative mitigate the harms of AVs while nurturing their benefits. scenarios of the unintended consequences of AVs. However, if AVs are currently both affordable and We conducted a model-based analysis that closely reasonably safe relative to conventional HPVs, the modeled the publicly-reported results from Waymo government should rapidly make large investments in since December 2016, the last point at which publicly- understanding the benefits and threats posed by AVs. available data were posted online (Waymo 2016b). The The urgency arises from two competing problems that dataset included a list of 14 crashes that took place be- arise when this new, affordable, and safer technology is tween 1 October 2012 and 22 August 2016 while com- available. Either: 1) society will fail to deploy AV tech- pleting 2,102,047 miles of fully autonomous travel. nology with the potential to save lives in the short-term, While the car was on fully-autonomous mode for all leading to needless suffering and loss of life; or, 2) AVs crashes used in our analysis, an engineer was in the car will be successfully introduced on the road over broader at all times. The average speed of the AV during colli- geographic areas and will continue to grow in use with sions was 0.23 miles per hour and the average speed of little regulation, leading to potentially serious long-term the other vehicle in collisions was 7.54 miles per hour, health and social consequences. with the majority of crashes (9 of 14, 64%) being rear- In this study, we simulate the economic costs and end collisions with the AV stationary. We used a Markov health benefits of AVs today given their high costs as chain microsimulation model that compares AVs and well as the considerable uncertainty surrounding their HPVs over the average lifespan of a car. All future costs safety. We simply ask whether AVs are affordable and and QALYs were discounted at a rate of 3%, and we ad- plausibly safe given the best available data. We also hered to the reference case scenario of the new Panel on make projections of future costs. Cost-Effectiveness Analysis for Health and Medicine (Neumann et al. 2016). Our chief assumptions are listed Methods in Table 1. Analytic overview Using a representative cohort of drivers with ages While AVs frequently hand control over to a human reflecting the ages of US drivers, we calculated the incre- driver and perform poorly in bad weather (known as dis- mental cost-effectiveness ratio (ICER) of AVs as they engagement) (State of California 2015; Favaro et al. exist today relative to HPVs, while also projecting vari- 2018), there is now sufficient data to estimate crash rates ous future expected and possibly unintended conse- were no driver present when software is optimized for a quences of adopting AVs. Specifically, we estimated the given city. incremental change in cost divided by the incremental Waymo (previously, Google Self-Driving Car Project) change in QALYs of AVs relative to that of HPVs, ac- uses a wider array of expensive technology to improve counting for the differential in QALYs accrued when in- its safety than other commercial AVs, and Waymo’s jured or dead, and various costs. Table 1 Major assumptions and their justifications used to build our Markov microsimulation models Assumption Justification Driverless autonomous vehicles (AVs) will have a higher Minor crash rates of human-piloted vehicles are very difficult to estimate as most go unreported. minor crash rate than human piloted vehicles Simulations from Waymo show lower crash rates, but the company has a financial stake in such outcomes. Reported minor crash rates for Waymo showed higher rates in AVs as compared to HPVs. The Waymo dataset is a key data source for this study. 2 million miles of driving is adequate for inference We used established formulas to extrapolate probabilities of injury and death regarding the safety of AVs using observed mean crash speeds for AVs. If adopted today, AVs would eliminate most AVs can be rented as taxis by private car owners when not being used. parking spaces The cost of autonomous vehicles will decline The efficiency and cost of many technologies follows Moore’s law for central following Moore’s law processing unit speeds. Used cars will have a similar market, whether While technologies in AV have few moving parts, they tend to decline in value as fast human piloted or autonomous as automobiles do because the technology becomes dated very quickly. Autonomous vehicles without a human driver When the driver’s seat can hold a paying passenger, the additional passenger will will increase productivity sometimes perform work on a device (e.g., emailing colleagues). AVs and human-piloted vehicles are of similar The build of a car can also influence the probability of injury and death. In current use, quality of build AV equipment seems to be used across a span of makes and models. Freedman et al. Injury Epidemiology (2018) 5:24 Page 3 of 8 Costs vehicle, as well as salaries for drivers. We assumed We calculated the total costs of property damage only, that increased parking availability, as a result of taxi minor, and severe injury crashes as classified by the usage, would reduce fuel costs and offset the Maximum Abbreviated Injury Scale (MAIS) score, and additional fuel costs from using vehicles as taxis. See death using data obtained from the National Highway Additional file 1 for more notes on cost of taxis. All Traffic Safety Administration (NHTSA) report on the parameters used in estimation were included in sensi- economic and societal impact of motor vehicle crashes tivity analyses. All costs were inflated to 2016 US (Blincoe et al. 2015). We considered a MAIS score of dollars and are shown in Table 2. less than 3 to be a minor injury crash and a MAIS score of 3 or above to be a severe injury crash. These HPV Probability of crash, injury, and death crash costs were then scaled to more accurately reflect Expected annual probabilities of injuries and fatalities in AV crashes since AV equipment is more expensive HPVs were calculated via NHTSA reports (Blincoe et al. (Greenblatt and Saxena 2015; Vallet 2016). The scaling 2015; National Highway Traffic Safety Administration of costs was achieved by estimating the proportion of 2016). We use data from 2,102,047 autonomous vehicle costs attributable to destruction of a vehicle (car dam- miles travelled (VMT) collected between 2010 and 2016 age) and adding the expected marginal cost of damage from Waymo – a period over which there were 14 to an AV relative to a HPV in a similar crash type (see crashes (Waymo 2016b). In all cases, the driver of the Additional file 1: Equation 2). We estimated the average non-AV was at fault. While all Waymo crashes are costs of a HPV (New-car transaction prices up 2 percent reported, many HPV crashes are not. When only in march 2016, along with increases in incentive spend, reported crashes are considered, AVs are 1.30 times according to Kelley blue book [press release] 2016), an more likely to be involved in a minor crash (Table 2). AV (Greenblatt and Saxena 2015), an AV projected five Waymo’s AVs are tested with a human behind the years into the future under the assumption that the cost wheel who can take over if needed. Waymo’s simulations of technology follows Moore’s law—i.e., periodic predict that there would have been an additional 7 halving—with a period of 2 years (Greenblatt and Saxena crashes between 2010 and 2016 were the human not 2015; Strawn and Strawn 2015), the annual cost of able to take control of the car, and that all 7, or 1/3, owning a car (American Automobile Association 2014), would have been caused by the AV. Given the small n of average cost of a funeral (National Funeral Directors 21 crashes, we chose not to use actual crashes to esti- Association 2017), productivity loss (i.e. opportunity mate injuries and fatalities. Rather, we used well- cost) from driving (United States Office of Personnel established formulas that outline the risk of crashes, Management n.d.; American Automobile Association injuries, and fatalities based on mean change in vehicle 2015; Proctor et al. 2016), costs of spots (The Econo- speeds at the time of the crash (Mohit et al. 2017; mist 2017; Chester et al. 2010), and taxi-related costs Flannagan 2013), whether caused by the AV or the (Greenblatt and Saxena 2015; New-car transaction human driver of the other car. Using the variance of the prices up 2 percent in march 2016, along with in- actual distribution of speeds at the time of crash, we creases in incentive spend, according to Kelley blue created a distribution of speeds at the time of the crash, book [press release] 2016; Bureau of Labor Statistics and linked each sample of this distribution to its risk of 2015; New York City Taxi and Limousine Commission injury or death. 2014) using various other sources and assuming na- The mean change in speed (ΔV) for Waymo crashes tional averages. Productivity loss from driving a HPV was 3.91 miles per hour. This injury model classified was calculated by multiplying the mean wage (United injuries using the KABCO scale in which an individual is States Office of Personnel Management n.d.;Proctor either killed (K), has an incapacitating injury (A), a non- et al. 2016), mean commuting time (American Auto- incapacitating injury (B), a possible injury (C), or no mobile Association 2015), and a productivity param- injury (O). In order to use data reported in both MAIS eter between 0 and 1, with a base case productivity and KABCO format, we considered the two scales parameter of 0.3 Cost of parking spots for non-taxis equivalent, with K corresponding to MAIS 6, A corre- were calculated by multiplying the average cost of one sponding to MAIS 3, 4, or 5, B corresponding to MAIS parking spot (including costs to build, buy the land, 1 or 2, and C corresponding to MAIS 0 or no injury. maintain the spot, etc.) (The Economist 2017)bythe ratio of parking spots to registered cars (base case ratio of 3.4) (Chester et al. 2010). The cost of parking Morbidity spots for taxis were set equal to just the average cost Changes in health-related quality of life due to MVCs one parking spot (The Economist 2017). Taxi-related were derived using the EuroQol 5D5L with the assist- costs included costs for buying and maintaining the ance of two pediatric orthopedic surgeons with extensive Freedman et al. Injury Epidemiology (2018) 5:24 Page 4 of 8 Table 2 Selected costs and probabilities used in the Markov microsimulation model Low Value Base Value High Value Selected Costs ($) Property Damage Only Crash (Blincoe et al. 2015) 2976 4251 5526 Minor Injury Crash (Blincoe et al. 2015) 3129 12,282 61,352 Severe Injury Crash (Blincoe et al. 2015) 200,241 274,553 1,101,865 Fatal Crash (Blincoe et al. 2015) 1,072,412 1,532,018 1,991,623 Property Damage Only Crash AV (Blincoe et al. 2015; Greenblatt and Saxena 2015; Vallet 2016) 3123 4461 5799 Minor Injury Crash AV (Blincoe et al. 2015; Greenblatt and Saxena 2015; Vallet 2016) 3864 13,332 62,717 Severe Injury Crash AV (Blincoe et al. 2015; Greenblatt and Saxena 2015; Vallet 2016) 214,206 294,503 1,127,800 Fatal Crash AV (Blincoe et al. 2015; Greenblatt and Saxena 2015; Vallet 2016) 1,086,965 1,552,808 2,018,650 Autonomous Vehicle (Greenblatt and Saxena 2015) 122,218 183,666 265,648 Autonomous Vehicle in 5 Years (Greenblatt and Saxena 2015; Strawn and Strawn 2015) 33,229 56,539 100,383 Human Piloted Vehicle (New-car transaction prices up 2 percent in march 2016, along with increases 17,218 33,666 70,648 in incentive spend, according to Kelley blue book [press release] 2016) Annual Cost of Owning Car (American Automobile Association 2014) 8536 Funeral (National Funeral Directors Association 2017) 6206 7332 8687 Productivity (United States Office of Personnel Management n.d.; American Automobile Association 2015; 552 1657 2762 Proctor et al. 2016) Parking spot for non-taxi (The Economist 2017; Chester et al. 2010) 47,600 102,000 176,800 Parking spot for taxi (The Economist 2017; Chester et al. 2010) 0 30,000 80,000 Taxi Salary (Bureau of Labor Statistics 2015) 19,432 27,760 36,088 Cost of Taxi (New-car transaction prices up 2 percent in march 2016, along with increases in incentive spend, 92,699 132,427 172,155 according to Kelley blue book [press release] 2016; New York City Taxi and Limousine Commission 2014) Cost of AV Taxi (Greenblatt and Saxena 2015; New York City Taxi and Limousine Commission 2014) 197,699 282,427 367,155 Selected Probabilities Human Piloted Vehicle (Blincoe et al. 2015; National Highway Traffic Safety Administration 2016) Crash – 0.0646 – Property Damage Only Crash – 0.6091 – Minor Injury from Crash – 0.3766 – Severe Injury from Crash – 0.0090 – Death from Crash – 0.0054 – Autonomous Vehicle (Waymo 2016b) Crash 0.06729 0.08385 0.09319 Property Damage Only Crash 0.9551 0.9654 0.9777 Minor Injury from Crash 0.0220 0.0315 0.0409 Severe Injury from Crash 2.09E-04 2.99E-03 3.88E-03 Death from Crash 6.48E-05 9.26E-05 1.20E-04 All costs have been rounded to the nearest 2016 dollar clinical experience working with crash victims (Muennig (TreeAge Software, Willamstown, Mass.). The model ex- et al. 2014; Gold et al. 1996). amined the use of a mode of transportation for the life- span of a HPV compared to the use of an alternative Decision-analysis models mode of transportation for the same time period. We simulated the average lifespan of a motor vehicle Following is a description of the baseline HPV versus with an average of 14,133 annual VMT per licensed AV model. For each iteration (N = 10,000) of the deci- driver (National Highway Traffic Safety Administration sion analysis model, two drivers were simulated, each 2016). A decision-analysis model was constructed using with the same age randomly selected from a representa- TreeAgePro 2016 software for the Macintosh computer tive distribution of licensed drivers’ ages in the United Freedman et al. Injury Epidemiology (2018) 5:24 Page 5 of 8 States. The driver was randomized to either pilot a HPV influence on the results of the model using the range of or AV. At each time step for either vehicle, the subjects values in Table 2. Values for several variables of interest could be in one of several states: driving, driving severely were simultaneously varied over their plausible range injured, dead, out of the analytic horizon and healthy, or using Monte Carlo microsimulations (N = 10,000 out of the analytic horizon and severely injured. The microsimulations), with values drawn from probabilis- analytic horizon was defined as the lifespan of a repre- tically weighted triangular distributions or from sentative vehicle, taken from a vehicle lifetable, and an reported heterogeneous distributions, with linear AV was assumed to have an equivalent lifespan to that interpolation. Because NHTSA does not publish a of an HPV (Table 1). It was important to continue to standard willingness-to-pay (WTP) threshold, we used run the model after the analytic horizon was reached, so the EPA’s WTP threshold of $140,000/QALY (Muennig as to accurately calculate the impact on total QALYs of and Bounthavong 2016). The results of this threshold each strategy past the lifespan of the car. While driving analysis are reported below. or driving severely injured, a subject accrued the costs of purchasing a vehicle (only on the first timestep), the cost Results of owning and operating a vehicle, the cost of parking, Simulation of human-piloted car cohort and the productivity loss from piloting a vehicle ($0 for The model replicated the U.S. cohort of conventional AV) (Table 2). For models involving taxis, the salary of car drivers between 2013 and 2016 and the U.S. cohort the taxi driver and other costs were included. Both the of conventional taxis. Over the lifetime of the average expiration of the car (i.e., the end of the analytic car, HPVs cost about $286,000 including the cost of the horizon) and the probability of a crash were drawn from vehicle, maintenance, parking, and other costs. The uniform random variables, as in a Monte Carlo model, average number of QALYs that the driver lives is 16.43. and compared to a lifetable for each timestep. The cost of vehicle maintenance and upkeep for an AV was equal Projected outcomes for the adoption of autonomous to the cost of maintenance for a HPV plus the relative vehicles marginal cost of an AV relative to an HPV—i.e., the cost Quality-adjusted life years was scaled to account for the relative difference in Data from Waymo indicate the relative risk (RR) of vehicle cost. For each timestep, a subject could be crashes for AVs is 1.30, while the RR for fatalities based involved in a crash with a fixed probability. The crash upon the vehicle’s speed at the time of the crash is 0.02 type was determined using national statistics of the fre- (Table 1) due to a much lower speed of impact with any quency of crashes of each MAIS type, where an of MAIS collisions. 0, 1, and 2 was considered a minor injury crash, an AVs were projected to be safer than HPVs, with an MAIS of 3, 4, or 5 was considered a severe injury crash, incremental increase in QALYs of 0.08 (0.05% change), and a fatal crash and property damage only crash were or approximately one month of perfect health, over the considered in their own crash categories. If an individual time that the car is on the road. was involved in a fatal crash, they would be sent to the “Dead” category for subsequent time steps, while if an Costs individual was involved in a severe injury crash, they The vehicle lifetime costs of purchasing, maintaining, and would be sent to the “Driving Severely Injured” category. operating new AVs ($425,757, 95% CI: $288,479–$594,010) Individuals could also be sent to the “Dead” category were projected to exceed costs for HPVs over the same based on a lifetable checked each year given the subject’s period by 49%, with most of the costs attributed to the age. At the conclusion of the model (T = 100 timesteps), initial cost of an AV. The lifetime costs of purchasing, all vehicles were expired and all individuals were dead; maintaining, and operating a new AV with costs reduced the QALYs and costs for each iteration (N = 10,000) by 5-year Moore’s Law projection ($303,535, 95% CI: were calculated and compared within iterations and $173,959–$613,812) still exceed the costs for HPVs by 6%. across the entire simulation. The projected lifetime cost of using AV taxis ($447,667, The decision trees for each model can be found in the 95% CI: $306,002–$634,870) in place of HPVs was found to Additional file 1. be 56% more expensive than owning and operating a pri- vateHPV for thesame period. Thedifferenceincosts is primarily due to profits for the taxi owner. The projected Sensitivity and scenario analyses lifetime societal cost of using a human-piloted taxi We first conducted an analysis for the base-case sce- ($570,032, 95% CI: $222,787–$1,205,646) was 34% higher nario, defined by the input values in Table 2. Sensitivity than those for using an AV taxi over the same period. Since analyses were then conducted using plausible ranges of both vehicles have similar parking requirements, the differ- high and low values for each variable to test their ence was primarily due to the salary of the human driver. Freedman et al. Injury Epidemiology (2018) 5:24 Page 6 of 8 Table 3 The cost, quality-adjusted life years gained, and 95% confidence intervals (CIs), and incremental cost-effectiveness ratios (ICER) from our microsimulation models (based on 10,000 microsumulations) for human piloted vehicles (HPVs) versus: privately-owned autonomous vehicles (AVs), AVs 5 years in the future, and AV taxis. The final simulation compares human piloted taxis versus AV taxiss Cost 95% CI QALYs 95% CI ICER 2.50% 97.50% 2.50% 97.50% Human-Piloted Vehicles Versus AVs HPVs 286,146 155,949 653,505 16.41 0.99 28.18 AVs 425,757 288,479 594,010 16.51 1.47 28.34 1,396,110 AVs in 5 Years 303,535 173,959 613,182 16.51 1.47 28.34 173,890 AV Taxis 447,667 306,002 634,870 16.51 1.47 28.34 1,615,210 Human-Piloted Taxis Versus AV Taxis HPT 570,032 222,787 1,205,646 16.41 0.99 28.18 AV Taxis 447,667 306,002 634,870 16.51 1.47 28.34 Saves money and QALYs Incremental cost-effectiveness ratios in this analysis are only useful insofar as they inform consumers of the value of an AV relative to a conventional car on the grounds of the best available health data. These data are subject to considerable uncertainty AVs were assumed to fall in price according to Moore’s Law, but were not assigned increased effectiveness Cost-effectiveness adoption of this technology, dramatically reduce the cost The adoption of AVs for private use was associated with of ride sharing, and remove jobs. For this reason, legisla- an ICER of $1,396,110/QALY gained (Table 3). Likewise, tors should consider studying the regulation of AVs with the projected adoption of AVs in 5-years was $173,890/ more urgency. If a commercially viable product ends up QALY gained. The scenario comparing current HPVs to on the road, the widespread adoption of AVs could come AV taxis was associated with an ICER of $1,615,210/ sooner than anticipated even if its market price is too QALY gained. The strategy of using human-piloted taxis high for most families to afford as a personal vehicle. as an alternative to AVs was found to save both money For this reason, policies that mitigate the potential and QALYs. harms of AVs—and particularly the widespread and Cost-effectiveness estimates were remarkably stable for rapid use of AVs as taxis, should be studied with more all scenarios and variables, with two notable exceptions. urgency. The model comparing HPVs to a 5-year projection of Many have speculated about the potential harms and AVs was sensitive both to variation in the cost of an AV benefits of AVs on the road. If adopted today, AV taxis or AV taxi and to variation in the probability of crash in would likely make it easier to transport people with an AV or AV taxi. minor disabilities and would also lower costs for taxi trips. This way, AV taxis produce large benefits for the Discussion small segment of the population that regularly relies on While the data are poor, and there is considerable uncer- taxis or disabled transit services. However, it could also tainty, we find that the widespread adoption of AVs expand the number of people who opt to ride share could plausibly save lives, reduce suffering, and produce rather than own a personal vehicle. improvements in productivity (due to fewer injuries), in When used as taxis, AVs could, in the long-term, also the short-term. However, it is unlikely that such vehicles plausibly produce more land in cities for real estate, will find much of a market for personal use in private sidewalks, bike lanes, emergency vehicle lanes, and sector in the near-term due to their high cost of owner- parks. This is because they continuously drive, shuttling ship and the impracticality of limiting the vehicles to people around, and rarely needing to park. By decreasing roads on which they have been programmed to operate. cars owned for personal use, they could eventually This is true even when additional cost savings are con- eliminate office and shopping mall parking spaces, as sidered, such as the economic benefits associated with well as parking spaces provided by cities. By reducing reductions in parking spaces. congestion and increasing drop off space, AVs could However, we find that AVs would save money if used reduce transport times of goods, increasing economic effi- as a taxi, even in their current state of development. ciency at the macro level. Driver costs consume upwards Both companies and families might be incentivized to of 30% of trucking costs (European Parlament 2015). invest in an AV were one available on the market. Such While this brings huge economic efficiencies for a vehicle could be rented out for additional income as a corporations, it can produce large shifts in the labor ride share, rather than depreciating in value in the gar- force. These include the rapid displacement of taxi and age. This could effectively incentivize fairly widespread delivery drivers who might not be able to find jobs that Freedman et al. Injury Epidemiology (2018) 5:24 Page 7 of 8 pay as well. By providing taxi services at a lower price, result, it is difficult to know whether AVs actually cause AV taxis could also move middle-income riders out of more or fewer minor crashes. urban public transportation systems, which too often Waymo’s simulations based on billions of miles of function on the edge of financial viability in the US. simulated driving predict that there would be only 68% (Many countries tend to have cheaper, more stable tran- as many minor crashes if AVs completely replaced HPVs sit systems, however.) today (and many fewer serious ones). We chose not to Because they deliver passengers directly from one rely on Waymo’s simulation data to estimate minor destination to the next and because they displace public crash rates because the data are not available for public transit, AV taxis could lead to less walking and higher scrutiny. But if we did use this estimate, the projected rates of obesity among some segments of the population, safety of AVs would be greatly increased, along with while incentivizing others to take to the safer roads for their cost-effectiveness. exercise. Second, the non-simulated data from Waymo that we Any of the above scenarios are mere speculation. do use not only depends on reporting by the company Without investments in research to study these potential itself, but are also limited to just over 2 million miles of impacts of AVs, including regulatory policy experiments, actual road driving. there will be no public debate or preparation for these Finally, AV crashes may garner more attention than possibilities. HPV crashes (Boudette 2016). If adopted too early, the If the time horizon for a shift to AVs is a year to a few crashes that they do cause may stir backlash, derailing years, as our study suggests it might be, action is needed their use. Our models simply show that AVs could plaus- sooner rather than later. It is certainly difficult to study ibly be used on the road today under ideal conditions. whether AV taxis would reduce congestion and pollution Oneofthe more obviousthreats isthat, withAV taxison (by driving very close to one another and by optimizing the road, society could lose public transit systems. If so, routes) or would increase it (by increasing demand for road they pose threats to mobility for low-income Americans, vehicles). However, legislation is best made on research ra- and those in major urban areas that are dependent on ther than guesswork, and complex systems dynamics public transit systems. models could inform such policymaking (Sterman 2006). We also studied the cost-effectiveness of AV technology Conclusions in as a health investment because it is possible that some In the early 1900s, horse manure and corpses littered the families would be attracted to the safety of AVs relative to streets of major cities, posing a public health threat. When conventional vehicles. At present, AV adoption in passen- the automobile was first introduced on US streets, it was ger vehicles is significantly more expensive per QALY seen as a way of mitigating these public health threats, but gained than medical practices that are deemed unafford- little attention was paid to the public health threats that able (Boulware et al. 2003). It would be more than 5 years automobiles produced as unintended consequences. We before privately owned AVs reached the $140,000/QALY are now at the next frontier of transportation. However, gained threshold, which is considered a high valuation for once again, Federal legislation is focusing on mitigating the willingness-to-pay for a QALY (Muennig and the public health threats associated with human-powered Bounthavong 2016). As such, they are not a good value vehicles without attending to the threats that they pose. even for households that can afford them for personal use, Our study shows that autonomous taxis appear to be on nor are they likely to be in the near future. the threshold of viability from an economic and injury prevention standpoint and may be introduced more Limitations rapidly than some experts believe. There is an urgent need Our study has several limitations ranging from data un- for legislators to begin to focus on the unintended conse- certainty in present terms to wide-ranging assumptions quences of AVs, including their impact on public trans- about what the future might look like. We tested this portation and the built environment. uncertainty and these assumptions using multiple one- way sensitivity analyses and a Monte Carlo analysis, Additional file which examines all sources of uncertainty together. One of the biggest challenges in building our model was Additional file 1: Supplemental Appendix including notes on calculations; complete lists of variables, life tables, and decision trees; and estimating crash rates. Roughly, 47% of crashes among example results. (DOCX 12099 kb) HPVs go unreported (M. Davis and Co. 2015). This is most likely because it is often less expensive for owners to Funding directly pay for minor automobile damages than to make This work was funded by Global Research and Analytics in Population Health an insurance claim. On the other hand, our AV data con- (GRAPH), Mailman School of Public Health, Columbia University. Peter A tain every incident, no matter how minor the crash. As a Muennig serves as Director of GRAPH. Freedman et al. Injury Epidemiology (2018) 5:24 Page 8 of 8 Availability of data and materials Kang, C. Where self-driving cars go to learn. New York Times. 2017. https://www. All data were made publicly available by Google Self-Driving Car Project nytimes.com/2017/11/11/technology/arizona-tech-industry-favorite-self- (now Waymo). The data was removed from online sources on or around driving-hub.html?hp&action=click&pgtype=Homepage&clickSource=story- November 2016. heading&module=second-column-region®ion=top-news&WT.nav=top- news. Accessed 11 Nov 2017. KPMG. (2013). Self-driving cars. Are we ready? https://home.kpmg.com/ Authors’ contributions content/dam/kpmg/pdf/2013/10/self-driving-cars-are-we-ready.pdf. IGF assisted in the design of the study and study concepts, data acquisition, Accessed 31 July 2017. quality control of data and algorithms, data analysis and interpretation, and M. Davis & Co. National telephone survey of reported and unreported motor statistical analysis; EK assisted in the literature review and participated in data vehicle crashes. Washington, DC: National Highway Traffic Safety acquisition. PAM assisted in the design of study concepts, study design, and Administration; 2015. quality control of data and algorithms. All authors assisted in manuscript McGinnis JM, Foege WH. Actual causes of death in the United States. JAMA. preparation, manuscript editing, and manuscript review. All authors read and 1993;270(18):2207–12. approved the final manuscript. Mohit B, Rosen Z, Muennig P. The impact of urban speed reduction programmes on health system cost and utilities. Inj Prev. 2017; https://doi.org/10.1136/ Ethics approval and consent to participate injuryprev-2017-042340. Not applicable. Muennig P, Bounthavong M. Cost-effectiveness analysis in health: a practical approach: John Wiley & Sons; 2016. Competing interests Muennig PA, Epstein M, Li G, DiMaggio C. The cost-effectiveness of new York The authors declare that they have no competing interests. City's safe routes to school program. Am J Public Health. 2014;104(7):1294–9. National Funeral Directors Association. Statistics: National Funeral Directors Author details Association. 2017. http://www.nfda.org/news/statistics.Accessed 17 1 2 Yale School of Medicine, New Haven, USA. Global Research Analytics for May 2017. Population Health, Columbia University Mailman School of Public Health, National Highway Traffic Safety Administration. Traffic safety facts 2014: a New York City, USA. University of Michigan School of Public Health, Ann compilation of motor vehicle crash data from the fatality analysis reporting Arbor, USA. system and the general estimates system. Washington, DC: U.S. Department of Transportation; 2016. Received: 22 December 2017 Accepted: 5 April 2018 Neumann PJ, Sanders GD, Russell LB, Siegel JE, Ganiats TG. Cost-effectiveness in health and medicine: Oxford University Press; 2016. New York City Taxi & Limousine Commission. Taxicab Factbook. New York: New References York City Taxi & Limousine Commission. 2014. http://www.nyc.gov/html/tlc/ American Automobile Association. Owning and operating your vehicle just got a downloads/pdf/2014_taxicab_fact_book.pdf. Accessed 17 May 2017. little cheaper according to AAA’s 2014 ‘your driving costs’ study: American New-car transaction prices up 2 percent in March 2016, along with increases in Automobile Association. 2014. http://newsroom.aaa.com/tag/driving-cost- incentive spend, according to Kelley blue book [press release]. Irvine; 2016. per-mile/. Accessed 17 May 2017. Proctor BD, Semega JL, Kollar MA. Income and poverty in the United States: American Automobile Association. American driving survey: methodology and 2015. Washington, DC: United States Census Bureau; 2016. year 1 results, May 2013–May 2014. Washington, DC: American Automobile State of California. Department of Motor Vehicles. Disenagement Reports, Association; 2015. https://newsroom.aaa.com/tag/american-driving-survey/. 2015. https://www.dmv.ca.gov/portal/dmv/detail/vr/autonomous/ Accessed 17 May 2017. disengagement_report_2016. Accessed 17 May 2017. Blincoe L, Miller TR, Zaloshnja E, Lawrence BA. The economic and societal impact State of California Department of Motor Vehicles. (2018). Deployment of of motor vehicle crashes, 2010 (revised). Washington, DC: National Highway Autonomous Vehicles for Public Operation. https://www.dmv.ca.gov/portal/ Traffic Safety Administration; 2015. dmv/detail/vr/autonomous/auto. Accessed 7 Sept 2017. Boudette NE. Autopilot cited in death of Chinese tesla driver. The New York Sterman JD. Learning from evidence in a complex world. Am J Public Health. Times. 2016. https://www.nytimes.com/2016/09/15/business/fatal-tesla-crash- 2006;96(3):505–14. in-china-involved-autopilot-government-tv-says.html. Accessed 17 May 2017. Strawn G, Strawn C. Moore’s law at fifty. IT Prof. 2015;17(6):69–72. Boulware LE, Jaar BG, Tarver-Carr ME, Brancati FL, Powe NR. Screening for How not to create traffic jams, pollution and urban sprawl. The Economist. proteinuria in US adults: a cost-effectiveness analysis. JAMA. 2003; 2017. https://www.economist.com/news/briefing/21720269-dont-let- 290(23):3101–14. people-park-free-how-not-create-traffic-jams-pollution-and-urban-sprawl. Bureau of Labor Statistics. Taxi drivers and Chauffers: U.S. Department of Labor. Accessed 20 May 2017. 2015. https://www.bls.gov/ooh/transportation-and-material-moving/taxi- U.S. Department of Transportation. Federal Automated Vehicles Policy: U.S. drivers-and-chauffeurs.htm. Accessed 17 May 2017. Department of Transportation. 2016. https://www.transportation.gov/AV. Business Insider Insider Reports. (2015). Report on Self-Driving Cars. http://www. Accessed 7 Sept 2017. businessinsider.com/report-10-million-self-driving-cars-will-be-on-the-road-by- United States Office of Personnel Management. (n.d.). Pay & leave: pay 2020-2015-5-6. Accessed 31 July 2017. administration United States Office of Personnel Management. https://www. Chester M, Horvath A, Madanat S. Parking infrastructure: energy, emissions, opm.gov/policy-data-oversight/pay-leave/pay-administration/fact-sheets/ and automobile life-cycle environmental accounting. Environ Res Lett. computing-hourly-rates-of-pay-using-the-2087-hour-divisor/. Accessed 17 2010;5(3):1–8. May 2017. Energy & Commerce Committee. (2017). The SELFDRIVE Act. https:// Vallet M. Total loss thresholds by state 2016. http://www.carinsurance.com/ energycommerce.house.gov/selfdrive/. Accessed 7 Sept 2017. Articles/total-loss-thresholds.aspx. Accessed 17 May 2017. European Parlament. (2015). Employment conditions in the international road Waymo. On the road: Waymo; 2017a. https://waymo.com/ontheroad/. haulage sector. http://www.europarl.europa.eu/RegData/etudes/STUD/2015/ Waymo. (2016b). Google Self-Driving Car Project Monthly Report. 2010–2016. 542205/IPOL_STU(2015)542205_EN.pdf. Accessed 5 Feb 2017. Favaro F, Eurich S, Nader N. Autonomous vehicles’ disengagements: trends, triggers, and regulatory limitations. Accid Anal Prev. 2018;110:136–48. Flannagan C. A method for estimating delta-v distributions from injury outcomes in crashes: University of Michigan Transportation Research Institute; 2013. https://deepblue.lib.umich.edu/bitstream/handle/2027.42/ 117575/103241.pdf?sequence=1&isAllowed=y. Gold MR, Siegel JE, Russell LB, Weinstein MC. Cost-effectiveness in health and medicine. New York: Oxford University Press; 1996. Greenblatt JB, Saxena S. Autonomous taxis could greatly reduce greenhouse-gas emissions of US light-duty vehicles. Nat Clim Chang. 2015;5:860–5.
Injury Epidemiology – Springer Journals
Published: Jun 4, 2018
Access the full text.
Sign up today, get DeepDyve free for 14 days.