Perceived risk and modal choice: Risk compensation in transportation systems

Accid. Anal. and Prev., Vol. 27, No. 4, pp. 503-521, 1995 Copyright 0 1995 Elsevier Science Ltd Printed in the USA. All tights reserved 0001-4575/95 $9.50 + .OO

Pergamon ooo1&75(94)ooog7-5

PERCEIVED RISK AND MODAL CHOICE: RISK COMPENSATION IN TRANSPORTATION SYSTEMS ROBERT

B.

NOLAND

Department of Economics, School of Social Sciences, University of California, Irvine, Irvine, CA 92717, U.S.A. (Accepted

6 September 1994)

Abstract-A transportation mode choice analysis is performed that examines behavioral responses to perceived risk in the choice of mode for daily commute trips. This methodology provides a technique for examining, by means of disaggregate individual level data, risk-compensating effects in transportation systems.

Various measures of perceived risk are examined for explaining modal choice. Other studies have described how safety regulations have resulted in increases in “driving intensity.” This study defines one component of driving intensity to be the increased probability of commuting by automobile. The results show that modal shifts occur when risk perceptions for a given mode are reduced. To demonstrate potential risk-compensating effects within the transportation system, an estimate of changes in accident fatalities due to commuting is derived using rough estimates of fatalities per person-mile travelled. It is shown that a given change in the perceived risk of commuting by automobile results in a less than proportionate change in net commuting fatalities. The relative magnitude is dependent on how objective reductions in risk translate into perceived reductions in risk. This study also shows that perceived safety improvements in bicycle transportation have an aggregate elasticity value that is greater than one. This means that bicycle safety improvements attract proportionately more people to bicycle commuting (i.e. a 10% increase in safety results in a greater than 10% increase in the share of people bicycle commuting). Keywords-Risk

compensation,

Perceived risk, Mode choice, Transportation

INTRODUCTION

safety, Bicycle transportation

other attributes are held equal; obviously other factors such as travel time and cost are traded off against individual perceptions of modal risk when making a commute mode choice. As will be shown, people perceive bicycling to be the riskiest commute mode, followed by automobiles and walking, with transit being considered the safest. One advantage of this methodology is that it is not necessary to consider the timing of safety improvements in automobiles. In other words, it avoids the problem of estimating turnover of the automobile stock and the consequent change in the amount of vehicles subject to safety regulations. It is also not necessary to track specific improvements in highway safety design or increased usage of limited access freeways (which are generally considered safer). The reliance on individual perceptions of transportation risk provides a mechanism for translating objective criteria, which are difficult to measure, into a perceptive scale utilized by individuals in their choice process. It is also impossible to measure the objective risk that each individual com-

Peltzman (1975) hypothesized the theory of risk compensation as negating or offsetting various automobile safety regulations. His hypothesis was that automobile drivers would compensate for the increased safety of their vehicles by increasing their exposure to risk, or driving more “intensively.” This theory has been debated in the literature with much disagreement over whether it exists and the magnitude of the effect if it does. This paper presents a method for identifying the impact of driver perceptions of risk on the transportation system. A multinomial logit choice methodology was used to examine perceptions of risk as an attribute in modal choice. A model was constructed to test whether perceptions of risk associated with alternative modes of commuting affect people’s choice. If people are less likely to use riskier modes, then the result would be a reduced share of commuting by automobile and bicycle and increased patronage of transit, and in some cases where it is a feasible choice, increased walking. This is the case if all 503

504

R.B.NOLAND

muter faces. Blomquist (1991) shows evidence that motorists are competent in making judgments about risk-taking behavior. In other words, they are competent at relating objective criteria to individual perceptions. Peltzman (1975) defined driving intensity as an increase in reckless behavior, such as speeding. Orr (1982) expanded the notion of driving intensity to include other factors as well. Orr considered driving intensity to be an “unfortunate choice of words” that “creates images of the ordinary citizen turned aggressive by the application of a seat belt” (Orr 1982). He included other changes in behavior that could result from increases in safety regulations. For example, allowing small children to stand in the front seat of automobiles, an increased willingness to allow teenagers to drive, more driving during hazardous weather, and more tired or drunken driving. The analysis presented here will examine increased levels of driving to work as another measure of driving intensity (as well as increased levels of bicycle commuting) and will demonstrate how this could have an impact on net societal and transportation system risk. This study measures risk perceptions using various different measures. These measures are based on scale parameters for perceived accident probabilities and severities collected using survey data. As will be seen, some of these measures are significant in explaining modal choice. The predicted changes in modal shares are used to calculate the net impact on total commuting trip fatalities. Objective risk measures are specified using aggregate measures to calculate changes in fatalities as perceptions of risk are changed. The objective risk of a mode is defined as the actual risk facing a user of the transportation system for the specific mode and route chosen (these can be expressed, for example, as fatalities or injuries per vehicle-km or per trip). It will be shown that when a change in the objective risk of an automobile (or other modes) triggers a change in perceived risk, there is a less than proportional decrease in fatalities. This implies that risk compensation or partial offsetting does occur (i.e. there is a less than expected decline in fatalities based on pure engineering calculations). When perceptions of bicycling risk are reduced without any change in objective risk, increases in fatalities can be one possible outcome. This paper is not meant to analyze specific driver reactions that have generally been associated with risk-compensating behavior. For example, increased speeding and reckless driving (which could apply to bicyclists as well). The point of this paper is to show how relative changes in perceptions of risk

(which may or may not be associated with changes in objective risk) can result in unexpected impacts on total fatalities within the transportation system. That is, behavioral responses to relative changes in perceived risk levels can result in some marginal shifts in modal choice with offsetting increases in expected fatality reductions. This can be seen as a form of risk compensation within the transportation system, i.e. a behavioral response reflected through modal choice decisions resulting in less than expected risk reductions. Obviously, risk compensating behavior as traditionally defined may also play a large role. The following section will discuss previous work on risk compensation, starting with Peltzman’s basic theory and methodology. These will be discussed with regard to the methodology utilized in the current study and its potential advantages. This will be followed by a discussion of the data used in the current study and a comparison of risk perceptions for different transportation modes. An outline of the methodology used and the results of the analysis will then be presented. Conclusions and relevant policy implications are summarized in the final section. PREVIOUS WORK ON RISK COMPENSATION The effectiveness of safety regulations to protect the public welfare has been extensively debated in the literature, especially regarding automobile safety regulations. The question posed is whether regulatory measures taken to reduce societal risks actually achieve this goal. Evans (1985) classifies research on the impact of risk reduction into three general approaches. These include the engineering, economic, and homeostatic approaches. The engineering approach assumes that straightforward engineering calculations on accident reduction will result in a direct reduction in accident rates. The primary drawback is that this approach assumes that the risk taker does not react in any way to reductions in risk. Evans (1985) calls these approaches’ non-interacting. The economic approach attempts to redress this fault by treating risk (or safety) as a commodity. People seek to maximize their utility, which may involve trading off increased levels of safety for other goods. The general result of a reduction in risk will be some substitution of that reduction for other goods. Peltzman (1975) examined this issue with regard to automobile safety regulations and developed the hypothesis of risk compensation. This hypothesis stated that people who utilize safety measures take

Perceived risk and modal choice

greater risks. In the case of automobile safety, this was embodied by an increase in driving intensity. This would include speedier and more reckless driving. This type of response to less risk is a compensating behavioral change that offsets or negates the desired risk reduction. It could potentially even increase the total societal risk. Part of the increase in fatalities measured by Peltzman was that for pedestrians (defined in his sample to include bicyclists also). The methodology used by Peltzman was to project expected accident rates based on pre-1965 (or preregulation) accident rates to the present. Safety measures introduced over the time frame of Peltzman’s study include lap seat belts, energy-absorbing steering columns, penetration-resistant windshields, dual-braking systems, and padded instrument panels. Variables used in his time-series analysis included income, alcohol consumption, driving speed, driver age, a secular trend, and a cost component for insured accidents. The projected accident rate, assuming an automobile stock with no regulations, is compared to actual accident rates, to determine whether there was any significant reduction in accident rates due to safety regulations. Peltzman’s controversial conclusion was that safety regulations did not reduce the accident rate. In other words, the reduction in risk was fully offset by compensating behavioral changes that increased driving intensity. Part of this was due to projected increases in pedestrian accident rates. This is an important conclusion, especially since his methodology did not take into account the reaction of pedestrians and bicyclists to a riskier environment. In other words, they may also engage in compensating behavior and switch to safer modes or increase their levels of caution. Pedestrians, for example, may switch to transit, and bicycle commuters may switch to using automobiles (assuming these options are available for the given individual and the desired trip). The current research accounts for perceptions of risk across modes, and so will take these effects into account. Evans (1985) lists risk homeostasis, as hypothesized by Wilde (1982), as the third theoretical formulation of the impact of risk-reduction measures. The basic premise of this theory is that people maintain the same target level of risk over time. Wilde (1982) developed this line of reasoning with his risk homeostasis theory. He viewed driver behavior as a homeostatic control process similar to a temperature regulation mechanism in animals. Changes in the level of safety would induce a behavioral response mechanism to return to a desired target level of risk per unit of time. This implies that the only way to

505

achieve reductions in accident rates is to develop policies that motivate individuals to reduce their target level of risk. Wilde (1984) acknowledges that technological advances in road and automobile design have reduced risk per kilometer traveled. But, one result, according to Wilde, has been increased automobile usage. Risk reductions are not the only factor (and not even the primary factor) resulting in increased automobile usage. Increased incomes and land use changes are obviously much more important. One consequence is a reduction in transit usage, which is much safer per passenger-kilometer traveled. * Both Wilde’s (1982) and Peltzman’s (1975) studies have been subject to criticism. Evans (1986) concludes that the data reject the risk homeostasis theory. The statistical analyses employed by Peltzman have been criticized in several studies. Graham and Garber (1984) pointed out the Peltzman’s model is very sensitive to plausible changes in the functional form of his model. Using the same data, but without using logarithms in their regressions, they found that safety regulations actually averted fatalities. They question the plausibility of Peltzman’s model based on the sensitivity of the results to changes in the functional form of the model. Joksch (1976a, b) determined that Peltzman neglected several variables known to influence automobile safety-for example, changes in the size and weight of cars and improved door latches introduced in 1956-thereby leading to bias in his regressions. Robertson (1977a) critized Peltzman’s study for using variables that were collinear, resulting in distorted projections. He presents a modified model that avoids multicollinearity problems and that refutes the risk compensation hypotheses. Another study by Robertson (1977b) noted reduced fatalities per number of registered cars for later model years (which are subject to increased safety regulations). He also found similar reductions in pedestrian deaths as opposed to Peltzman’s finding of an increase in pedestrian deaths. Robertson (1981), in a study similar to Peltzman’s, found no evidence of risk compensation. He found death reductions associated with federal safety regulations for pedestrians, bicyclists, and motorcyclists as well as a significant reduction in fatalities for car occupants. These studies based their criticism of Peltzman’s results on the functional form of the model, i.e. his use of logarithms and the omission of key variables. *Transit may not be safer perjoumey if one includes walking and running to access transit. One study has estimated that in Norway the per-journey injury risk of using transit may be 1.5 to 2 times higher than that of driving one’s own car (Lie 1983).

506

R.B. NOLAND

Zlatoper (1984), in a study similar to Peltzman’s, concluded that safety regulations reduced vehicle-occupant fatality rates but that there was an increase in pedestrian fatalities. Zlatoper performed a regression using pedestrian deaths as a dependent variable and found that the proportion of cars with regulated safety features was a significant factor in increasing pedestrian deaths. This may, however, merely represent a time trend for other changes in the pedestrian environment that have increased levels of risk. A study by McCarthy (1986) adds support to Peltzman’s risk compensation theory. Using a binomial logit model, he shows that the decision to wear an occupant restraint device is dependent on the level of “risk price” (defined as the probability of a fatality) experienced by the driver. An increase in the amount of risk price will result in an increased demand for more safety, which can be achieved by using seat belts. The study found that variables associated with a larger risk price, such as speed limit, vehicle weight, and population density were significant in predicting seat belt usage. A study by Conybeare (1980) replicated Peltzman’s results using Australian data. He found a significant increase in nonoccupant fatalities due to mandatory seat belt legislation in Australia and a less than expected decline in occupant fatalities. Graham (1982) interprets the available evidence on risk compensation as showing that it is more likely to occur for policies designed to reduce accident frequency rather than those designed to reduce accident severity. For example, motorcycle helmets would reduce the severity of accidents, but not necessarily the frequency, while road safety improvements may reduce the frequency of accidents and hence be subject to more compensating behavior. Chirinko and Harper (1993) define two effects in a decomposition of offsetting behavior. One is a protective effect from any safety regulations that reduces the level of vulnerability to an accident. The other is a substitution effect that results in an increase in accident frequencies. They define separate vulnerability and accident rates to test whether regulations affected them. They found that vulnerability was significantly reduced by safety regulations, while the frequency of accidents was not affected (although its significance varied with the functional form of the model). Traynor (1993) evaluates individual accident reports to determine the impact of varying safety conditions on driver behavior. Specifically, Traynor defines a variable that measures driver aggression (i.e. intensity) and finds that another variable representing varying unsafe driving conditions affects the like-

lihood that drivers will be more aggressive. From this he concludes that there is a driver response to changes in the safety environment. Singh and Thayer (1992) use individual survey data to determine that compensating behavior is more likely in people who are not strongly risk averse. They conclude that the use of aggregate data in previous studies may have resulted in suspect conclusions since individual characteristics could not be measured. All of these studies fail to consider the changing environment for pedestrian and bicycle transportation. Many trips, which at one time were feasible by these modes, are no longer possible, due primarily to changing land-use patterns. In addition, transit access may be inadequate or nonexistent for many trips. Therefore, decreased usage of many of these modes, perhaps by changes in their relative risk, may confound analysis of aggregate statistics. In this paper, by measuring risk perceptions for alternative transportation modes, this question is analyzed from a new perspective using individual survey data, without some of the methodological problems for which Peltzman was criticized. For example, actual auto safety regulations and highway safety improvements imply a change in the objective levels of risk facing the driver. However, these changes in risk levels may be difficult, if not impossible, to measure, especially for specific individuals with different driving and commuting patterns. Variation in the objective risk facing each individual is avoided by using individual perceptions of the risk of the mode. This allows the model to use the risk information each individual has interpreted about his or her specific situation to determine any behavioral responses, in this case the choice of commuting mode. In addition, the model provides a framework that takes into account the responses of pedestrians and bicyclists. As their perceptions of risk change in response to increases in automobile usage or increases in driving intensity, they may respond by choosing a different mode or adopting other measures to reduce their levels of risk. The hypothesis that decreases in the perceived risk of a commute mode increase the likelihood of the mode being chosen will be tested. If the null hypothesis of no behavioral effect can be rejected, this model can then be used to estimate how modal shifts due to changes in risk perceptions may change the total level of traffic accidents. The increased risk exposure associated with increased usage of any mode may offset the risk reductions from safety improvements. This is a behavioral response similar to what Peltzman hypothesized. Peltzman hypothesized that people compensate for risk reductions

Perceived risk and modal

by being more careless. If automobile or bicycle commuting is made safer, people may increase their exposure to what is still a risk-taking activity (and in particular, if they are attracted from relatively safer modes, the net impact could be detrimental by increasing the level of fatalities within the transportation system). Therefore, while Peltzman showed an increase in driving intensity, defined as speedy and reckless driving, the present analysis expands this definition to include increases in automobile (and bicycle) commuting. DATA

USED

FOR ANALYSIS

A mail survey of the Pennsylvania counties of the Philadelphia metropolitan area was performed to gather data for the analysis. * Among the questions asked was one about identification of the primary transportation mode used for commuting to and from work, and two questions to elicit the perceived risk of various transportation modes (see Fig. 1). The overall response rate was over 55% using Dillman’s total design method without a registered mail followup (Dillman 1978). The question in Fig. 1 measures a score for the perceived probability that an accident will occur and a score for the perceived severity of an accident if it were to occur. The questions in Fig. 1 were asked for each mode of travel available to the survey respondent for his or her daily commute trip. The probability of an accident is measured by eliciting the respondent’s perceived likelihood that an accident will occur. Severity levels are elicited by describing different possible accident outcomes, such as “Injuries requiring medical attention but no hospital stay” and “Permanently disabling injuries (paralysis or coma).” Respondents were asked to consider their risk over a five-year time frame. This was arbitrarily selected as a large enough time frame to allow people to think adequately about potential future risks involved with each transportation mode. Slavic, Fischhoff, and Lichtenstein (1978) have shown that people respond to lifetime estimates of the risk of driving a car more readily than to the risk involved with a single trip. Since risk is defined as the expectation of an undesirable outcome, both scores were used to*The research was originally designed to evaluate the impact of perceptions of risk on the decision to bicycle to and from work. For this reason, the sample consisted of bicycle clubs and a general random sample (Noland 1992). The former yielded a response rate of 65% and the latter a response rate of 36%. The bicycle club sample was used to increase the number of bicycle commuters and does not affect the conclusions reached by this paper, since the logit analysis included adjustments for an exogenous sampling strategy.

choice

501

gether to obtain a measure of individual perceptions of risk. The advantage of using risk perceptions is that it is the relevant behavioral response variable to which people react. This allows the survey respondents to consider their own judgments about risky situations, rather than, for example, relying only on reported accidents. It also would be impossible to measure individual objective levels of risk for each surveyed individual and his or her specific commute route and potential commute modes. Obviously the method of scaling and measuring individual risk perceptions may not be perfect. However, as will be shown, the relative rankings of the four modes studied generally match public perceptions, thereby serving as partial validation that the scaling used is appropriate. Various other information was elicited by the survey instrument. These included measurements of perceived travel time, cost, and comfort of alternative modes and demographic information (see Noland [1992] for more details on the survey instrument). Perceived

risk measures

The data collected using the question in Fig. 1 allowed several measures of perceived risk to be analyzed. Risk can be defined as the product of the probability of an event’s occurring and its outcome when it occurs. Therefore, one simple measure would be to multiply the probability of accident score by the severity of accident score. This defines a simple risk perception (SRP) score (this score is resealed to be between one and seven by taking the square root of the product). Another perceived risk measure may better reflect how people scale the severity of accidents. For example, people place an increasingly greater negative utility on increases in accident severity such that the cardinal distance between the “no injuries” and “minor scrapes and bruises” scores is probably less than that between “injuries requiring a lengthy hospital stay of several months will full recovery” and “permanently disabling injuries.” The “fatal injuries” score would obviously have a very high negative utility. In other words, a linear scaling of the severity measures may not be adequate. Therefore, an enhanced severity risk perception (ESRP) score is devised and analyzed as an alternative measure. The ESRP is defined as: ESRP = (Probability ofaccident) x

l($Severily

The ESRP score weights each severity

ofaccident)

(1)

score as 10

R. B. NOLAND

508

The following ratings were given for the likelihood of being in an accident during the next five years if one were to commute to and from work. This was solicited for the following four transportation modes: automobile, bicycle, transit, and walking. The question was phrased as follows: Please rate how likely yoU think it is for you to be in an accident, during the next five years, ifyou used each of the following forms of transportation for commuting to and from work or school? 1 = Almost certain not to have an accident 2 = Very unlikely 3 = Somewhat unlikely 4 = 50% chance of having an accident 5 = Somewhat likely 6 = Very likely 7 = Almost certain to have an accident

The following ratings were given for the severity of an accident if it were to occur. The question was phrased as follows: Now consider the severity of an accident if it were to occur. Please rate how seriously injured yoU think you would be ifyou were to be in an accident with each of the following forms of transportation? 1 = No injuries whatsoever 2 = Minor scrapes and bruises 3 = Injuries requiring medical attention but no hospital stay 4 = Injuries requiring short hospital stay of a few days 5 = Injuries requiring lengthy hospital stay of several months with full recovery 6 = Permanently disabling injuries (paralysis or coma) 7 = Fatal injuries

Fig. 1. Survey questions used to elicit risk perceptions.

times worse than less severe outcomes. This is a base 10 logarithmic scale. As will be shown, this alternative measure does not give radically different results from those of the SRP measure. The main purpose of presenting the ESRP measure is to validate the robustness of the survey question presented in Fig. 1. The probability score may also not be adequate as a linear scale. The SRP score implies that those perceiving that they are “almost certain to have an accident” feel that the likelihood is seven times the probability of “almost certain not to have an accident.” An alternative measure changes these weights by transforming the scale. This is done according to the scale shown in Table 1. An enhanced probability risk perception score (EPRP) is then derived by multiplying this score by the linear severity

score. This score will also be used to validate the survey question, and as will be seen, does not produce results substantively different from the SRP score. All three measures tend to have some correlation with travel distance (as will be shown below). In other words, the longer the commute distance,

Table

1. Resealing

of perceived questions

Almost certain not to have an accident Very unlikely Somewhat unlikely 50% chance of having an accident Somewhat likely Very likely Almost certain to have an accident

probability 0.01 0.20 0.35 0.50 0.65 0.80 0.99

Perceived risk and modal choice

the greater the exposure to a risky situation and, hence, the greater the probability of an accident and the greater the perception of risk. These measures can be resealed to account for commute travel distance, thereby giving a time and distance independent measure of the perceived risk of alternative modes.* These alternative measures are compared and discussed in the following section. Comparison

of perceived

risk for various modes

Mean values for the perceived risk measures (based on the survey questions in Fig. l), for each mode, are presented in Table 2 by mode chosen for commuting. As can be seen, regardless of the mode actually used, the bicycle is almost always perceived as the riskiest mode for commuting. In only one case is bicycling not perceived as the riskiest mode. Automobile commuting is perceived as the riskiest mode by automobile and transit commuters when the SRP score is scaled by travel time (see Table 3). This is especially interesting since automobile and transit commuters generally travel longer distances. Generally, automobile commuting is considered the second most risky mode by the respondents, when measured using the SRP score. Transit is perceived as the safest mode; however, walking is considered safest when resealed by commute time. The differences in SRP scores between modes are all significant as shown by the F-test derived using analysis of variance. The ESRP scores give somewhat different results. First, bicycling is once again considered the riskiest mode by all subgroups. Walking is rated with a high ESRP, but this disappears when scaled for travel time. This is not surprising since many people would face very long walking times, which would imply a potentially high level of risk. Transit is again perceived to be the safest mode. F-tests for the ESRP by mode chosen shows no statistical difference between modes, except in the case of bicycle risk perceptions. The mean EPRP scores generally show the same pattern as the SRP scores. The only difference is that transit risk is not significantly different by mode chosen. When scaled by travel time, EPRP again shows results similar to the SRP score. *If someone commuting 50 miles a day were to walk to work he/she would obviously be exposed to a very high level of risk, partly due to exhaustion. The logit analysis that follows considers only feasible choice sets. Therefore walking was considered feasible for commute distances of only two miles or less (for bicycling this was restricted to six miles or less). In addition, most people with infeasible commutes using a specified mode would generally leave the survey questions unanswered for those modes. Allowing for variable choice sets allows one not to drop these individuals from the logit analysis. Those individuals were included in calculating the mean risk perception scores discussed below.

509

The relative perceptions between the modes is not surprising. Bicycling is often cited as having greater levels of risk than other modes (U.S. EPA 1979; Mitchell 1978; Holdgate 1981), and transit is perceived to be relatively safe with regard to accidents. People do not necessarily choose the mode they think is safest, implying that other factors are also important in their actual choice of mode, and that more than rationalization is at work here.? METHODOLOGY A multinomial logit analysis was applied to the data to determine which attributes, including the risk perception measure, were significant in predicting the choice of transportation mode. The methodology utilized is outlined in detail in BenAkiva and Lerman (1985) and is briefly presented here. Deterministic choice implies that the choice with maximum utility will always be chosen and that transitive preferences will always be observed. Empirical work has found this is not always the case. Individuals sometimes select different choices when the same attributes are presented, they may violate transitivity when presented with different choice sets, and seemingly identical consumers may not consistently select the same alternative. Error may be introduced from inadequate specification of utility functicns or random behavior on the part of consumers. Random utility models attempt to account for these disturbances. This model states that choice probabilities are equal to the probability that the utility of an alternative is equal to or greater than the utilities of all other alternatives in the given choice set. This can be formally written as: P(iIC,) = Pr[Ui, 2

Uj,,

allj E C,]

(2)

V, is random utility for alternative i and individual n, and C, is the choice set (in this case, the choice set will consist of transportation modes available to individual n). The utilities are assumed to be random variables. Some of the underlying sources of randomness includes attributes that are unobserved by the researcher, unobserved taste variations amongst tPeople may tend to rationalize their survey answers to correspond with their mode choice. That is, bicyclists may respond that bicycling is not the riskiest mode, while automobile commuters would find all other modes riskier. Table 2 shows that this is not the case. Bicyclists consider bicycling to be the riskiest mode and automobile commuters do not consider automobiles to be the safest mode.

R. B. NOLAND

510

Table 2. Mean of perceived risk measures by mode chosen Mode chosen Bicycle

Auto

Walking

Transit

F-test

Prob

Entire sample

73

329

35

69

-

-

506

3.65 (1.03) 2.90 (0.86) 2.40 (1.02) 2.21 (0.84)

4.32 (1.28) 2.84 (0.91)

4.08 (1.17) 3.26 (0.85) 2.69 (1.34) 2.21 (0.79)

6.877

0.000

4.548

0.004

6.913

0.000

(Z) 2.42 (0.87)

3.87 (1.26) 3.12 (1.21) 2.33 (1.05) 2.15 (0.78)

2.640

0.049

4.16 (1.25) 2.92 (0.93) 2.85 (1.38) 2.34 (0.85)

1467 (6854) 438 (3509) 1086 (6038)

5770 (15279) 2027 (11828) 1070 (3736) 3.9 (11.5)

3115 (11319) 611 (4812) 2328 (9346)

3.037

0.029

1.559

0.198

1.895

0.129

0.946

0.418

(3%)

7193 (18954) 329 (2867) 3837 (12234) 202 (2019)

1.81 (1.21) 1.08 (1.73) 0.62 (0.86) 0.56 (0.61)

2.62 (1.66) 1.04 (0.78) 1.28 (1.47) 0.71 (0.67)

2.07 (1.44) 1.41 (1.38) 0.55 (0.87) 0.55 (0.59)

2.30 (1.38) 1.40 (0.77)

6.370

0.000

1.398

0.001

7.408

0.000

2.462

0.062

Number of observations SRP Bicycling Auto Walking Transit ESRP Bicycling Auto Walking Transit EPRP Bicycling Auto Walking Transit

(&

(%) 0.54 (0.53)

5712 (16672) (4z) 3043 (10790) 216 (2147) 2.42 (1.58) 1.12 (0.83) 1.10 (1.36) 0.66 (0.64)

Notes: Standard deviations are in parenthesis. ESRP is in thousands.

individuals, measurement errors, and the substitution of instrumental variables to represent actual attributes. The utility (V,) can be expressed as the sum of components of the utility that are observable (V,) and a random error term (ei,J representing the sources of randomness mentioned above. This can be written as: Ui~ = Vi, + ein

Multinomial logit (MNL) models are generally used to operationalize random utility models. MNL models can be expressed as: e “in

P,(i) = Z.

J’=,

e%

P,,(i) is the probability that individual IZ will choose alternative i. If utility is restricted to linearin-parameters’ functions, then if B represents a vector of utility coefficients, and X, is a vector of attributes for alternative i, and Xj, represents the

vectors have:

for all alternatives

P,(i) =

in the choice

set, we

eB’xin EjEc, eB’%

The procedure used to estimate these models is maxestimation imum-likelihood (Newton-Raphson method). This procedure estimates the vector of attribute coefficients, B. Logit models determine the probability that a mode will be chosen based on comparisons between individual utilities for each mode. The differences between modal attributes for each individual are used to determine the choice. For example, a given individual may have longer travel times for all modes compared to another individual. It is only the difference between a given individual’s travel times that determines their choice. The same is true for the perceived risk variable. Therefore, if some individual uniformly has a higher level of perceived risk

Perceived risk and modal choice

511

Table 3. Mean of perceived risk measures, divided by travel time, by mode chosen Mode chosen

Number of observations SRP/time Bicycling Auto Walking Transit ESRP/time Bicycling Auto Walking

Bicycle

Auto

Walking

Transit

13

329

35

69

0.20 (0.12) 0.17 (0.10) 0.048 (0.043) 0.087 (0.071)

0.12 (0.10) 0.17 (0.14) 0.033 (0.032) 0.070 (0.082)

SO.38 (220.10) 28.65 (233.%) 16.59

156.97 (549.78) 11.64 (104.80) 30.43 (116.94) 2.81 (20.37)

Transit EPRPltime Bicycling

0.068 (0.070) 0.059 (0.074) 0.012 (0.020) 0.020 (0.033)

0.095 (0.079)

Auto Walking Transit

(Z) 0.013 (0.023) 0.020 (0.030)

0.45 (0.29) 0.36 (0.21) 0.14 (0.074) 0.20 (0.17) 854.86 (2669.8) 202.71 (1182.8) 51.09 (176.95) 0.366 (1.20) 0.24 (0.241) 0.15 (0.16) 0.024 (0.031) (KY2)

0.10 (0.076) 0.12 (0.088) 0.030 (0.042) 0.069 (0.056) 13.50 (298.05) 20.13 (160.43) 22.89 (79.59) 0.55 (1.58) 0.050 (0.038) 0.049 (0.039) (~:~) 0.016 (0.020)

F-test

Prob

Entire sample 506

81.946

0.000

25.159

0.000

84.751

0.000

22.770

0.000

8.640

0.000

3.415

0.017

0.790

0.500

2.060

0.105

39.438

0.000

15.533

0.000

4.117

0.007

6.094

0.000

0.15 (0.15) 0.17 (0.15) 0.042 (0.048) 0.081 (0.092) 178.48 (855.W) 28.47 (338.96) 28.83 (115.28) 3.81 (35.56) 0.081 (0.10) (GY) 0.013 (0.021) 0.021 (0.034)

Notes: Standard deviations are in parenthesis. ESRP is in thousands.

for all modes, it does not matter, model evaluates only differences. RESULTS

since the logit

OF ANALYSIS

Modal choice model results Tables 4, 5, and 6 present results for various formulations of a multinomial logit choice model. The results are interpreted based on the statistical significance of the attribute parameters.* The T-statistic for each attribute is shown. For testing the hypothesis that risk has a behavioral impact, a negative value on the risk measure is expected, therefore the T-statistic is one-tailed with levels above 1.65 at the 95% level of significance. The F2 statistic is equal to 1 - (L(p) - W/L(O). L(p) is the log-likelihood function evaluated at the maximum and L(0) is the log-likelihood evaluated at zero. K *While maw of the modal attributes are sienificant in determining choice, &is paper focuses only on &e perceived risk attributes. See Noland (1992) for a full discussion of the other policy variables.

is the number of parameters in the model. This statistic, while similar, is not equivalent to the adjusted R2 used in regression models for explaining the amount of total variation in a model. Instead it is used only as a means of comparing alternative models that use the same data (Ben-Akiva and Lerman 1985). Hence, the values shown, which are all about 0.4, do not indicate that only 40% of the variation is explained. Model A in Table 4 includes the simple risk perception (SRP) score and indicates that perceptions of risk are significant in determining modal choice at above the 95% confidence level. This means that individuals are more likely to choose a given commute mode the safer they perceive it to be. Generally this would tend to favor increased usage of transit and walking as opposed to bicycling and automobile use, all else equal. Recall that modal choice models based on a random utility model are compensatory; other factors, such as travel cost and travel time are also important in people’s choice of mode and are also significant in the model. Model B shows that the enhanced probability risk perception

R. B. NOLAND

512

Table 4. Results of multinomial logit mode choice analysis Coefficient estimates Independent

variables

Simple risk perception

(A) (SRP)

Enhanced probability risk perception Enhanced severity risk perception

(EPRP)

Transit constant Walk constant Perceived cost (US$ per month) Travel time (minutes) Perceived comfort Number of motor vehicles owned (specific to automobiles) Bicycle parking available (specific to bicycling) (specific to bicycling)

Sex (M = 1, specific to bicycling & transit) Income (in 1000 US$, specific to automobile) Hours of exercise per week (specific to bicycling and walking)

$d”) P

6.346 (3.619) 4.788 (2.800) 5.323 (3.155) -0.0103 (-4.954) -0.0274 (-2.660) 0.5508 (7.398) 0.404 (1.487) 1.617 (3.070) 0.496 (1.963) 1.295 (3.612) -0.963 E-5 (-1.304) 0.0738 (1.927) -171.55 0.3965

(C) -

-0.218 (-1.706)

(ESRP)

Automobile constant

Bicycling competency

-0.255 (-1.971) -

(B)

6.434 (3.675) 4.952 (2.905) 5.404 (3.197) -0.0102 (-4.985) -0.0273 (-2.666) 0.555 (7.457) 0.399 (1.473) 1.579 (3.007) 0.509 (2.012) 1.265 (3.546) -0.977 E-5 (-1.329) 0.0749 (1.952) - 172.01 0.3950

-0.317 E-4 (-1.554) 6.613 (3.808) 5.293 (3.142) 5.709 (3.406) -0.0105 (-5.105) -0.0306 (-2.966) 0.556 (7.519) 0.396 (1.470) 1.545 (2.923) 0.530 (2.097) 1.223 (3.418) -0.992 E-5 (-1.353) 0.0744 (1.946) -171.96 0.3952

Notes: T-stats are in parentheses. Dependent Variable: Choice of Mode. Number of records = 354. Number of cases = 510. L(0) = -305.80.

(EPRP) score is also significant above the 95% confidence level. The results, using this score, are basically similar to the results for the SRP measure. Model C includes the enhanced severity risk perception (ESRP) score. The coefficient on this variable is significant in predicting modal choice, but only above the 90% level. Note that the coefficient on travel time has a higher level of significance in this model than in model A. Both the SRP and ESRP score have some correlation with travel times due to the probability of an accident increasing with longer commute times. Additional analyses (not shown here) showed that when the travel time variable is removed, SRP would have a higher level of significance and ESRP a lower level. However, no major changes in the coefficient value occurred, indicating that multicollinearity is not having a major effect. This would seem to imply that the perceived

probability of an accident may be a good predictor of modal choice. Graham (1982) stated that it appeared that those safety measures that reduce the severity of an accident do not exhibit compensating behavior, while those that reduce the frequency of an accident do. Chirinko and Harper (1993) found no offsetting behavior when they tested for accident frequencies. Models D and E in Table 5 examine this hypothesis. It can be tested by using the two components of the perceived risk measure: the expected probability that an accident will occur and its expected severity. The perceived probability that an accident will occur is not significant, while the perceived severity is above the 90% confidence level. This result seems to contradict the hypothesis proposed by Graham. Model F also shows that using the enhanced severity measure ( 10(seueri@)) is not significant in modal choice. Models G, H, and I, in Table 6, scale all three risk measures by travel time to give risk per minute

Perceived risk and modal choice

513

Table 5. Results of multinomial logit mode choice analysis Coefficient estimates Indeoendent

variables

(D) -0.0987 (-0.958) -

Probability of accident Severity of accident Enhanced severity of accident Automobile constant Transit constant Walk constant Perceived cost (US$ per month) Travel time (minutes) Perceived comfort Number of motor vehicles owned (specific to automobiles) Bicycle parking available (specific to bicycling) Bicycling competency

(specific to bicycling)

Sex (M = 1, specific to bicycling & transit) Income (in 1000 US$, specific to automobile) Hours of exercise per week (specific to bicycling and walking)

6.559 (3.767) 5.110 (2.995) 5.518 (3.255) -0.0102 (-5.028) -0.0276 (-2.700) 0.560 (7.513) 0.410 (1.520) 1s79 (3.015) 0.514 (2.043) 1.281 (3.606) -0.976 E-5 (-1.331) 0.0738 (1.929) - 173.09 0.3915

(E)

(F)

-

-

-0.137 (-1.439) 6.519 (3.717) 5.097 (2.995) 5.715 (3.397) -0.0104 (-5.023) -0.0290 (-2.817) 0.552 (7.436) 0.410 (1.514) 1.590 (3.024) 0.518 (2.046) 1.304 (3.638) -0.101 E-4 (-1.370) 0.0715 (1.873) - 172.51 0.3934

-0.613 E-7 (-0.845) 6.651 (3.825) 5.334 (3.169) 5.782 (3.455) -0.0105 (-5.129) -0.0299 (-2.893) 0.558 (7.532) 0.409 (1.521) 1.550 (2.942) 0.531 (2.109) 1.251 (3.496) -0.994 E-5 (-1.358) 0.0732 (1.919) -173.19 0.3911

Notes: T-stats are in parentheses. Dependent variable: Choice of mode. Number of records = 354. Number of cases = 510. L(0)= -305.80.

of travel. This adjusts for any relationship to travel times and is an attempt to measure the inherent risk characteristics perceived for each mode. The SRP measure is significant at just above the 90% level while the ESRP measure is not significant at the 90% level but is just below this level. The EPRP measure has the highest statistical significance, at just below the 95% level. One could also rescale by travel distances, but since different modes have different travel times, the latter seems like a better measure. Wilde (1982) also considers risk per unit time as the relevant variable for measuring homeostatic behavior. Based on these results, it seems likely that the null hypothesis of no behavioral response (i.e. changes in modal choice) due to risk perceptions can be rejected. While the risk coefficient is above the 95% level in model A, the others are generally significant above the 90% level of confidence. The

implication is that when a mode is made safer (or perceived to be safer), more people will use it for commuting. The question is whether this can lead to a net increase in total accidents, or some offsetting of the expected decline in accidents, as Peltzman (1975) implied.

Elasticity of perceived risk coejjicients Table 7 displays the elasticities of perceived risk coefficients calculated from Model A (the SRP measure). These are aggregate elasticities that measure the percentage change in total modal shares for a given percentage change in the specified attribute (see Ben-Akiva and Lerman (1985) for a discussion of how these are calculated). Table 7 shows both direct elasticities (modal shifts from or to a specified mode for a given change in its own perceived risk coefficient) and cross elasticities (modal shifts in a

R.B.

514

NOLAND

Table 6. Results of multinomial logit mode choice analysis Coefficient estimates Independent

variables

Simple risk perception

(G) (SRP)/Commute

Enhanced probability risk perception Enhanced severity risk perception

time (EPRP)/Commute

(ESRP)/Commute

time

time

Automobile constant Transit constant Walk constant Perceived cost (US$ per month) Travel time (minutes) Perceived comfort Number of motor vehicles owned (specific to automobiles) Bicycle parking available (specific to bicycling) Bicycling competency

(specific to bicycling)

Sex (M = 1, specific to bicycling & transit) Income (in 1000 US.& specific to automobile) Hours of exercise per week (specific to bicycling and walking)

-1.049 (-1.295) 6.809 (3.888) 5.417 (3.196) 5.615 (3.319) -0.0106 (-5.155) -0.0312 (-2.967) 0.569 (7.636) 0.411 (1.519) 1.539 (2.918) 0.549 (2.164) 1.293 (3.628) -0.988 E-5 (-1.340) 0.0710 (1.869) - 172.66 0.3929

J-(P) ir2

(H)

(I)

-

-

-2.308 (-1.619) -

-

7.611 (3.825) 6.711 (3.129) 5.549 (3.281) -0.0104 (-5.087) -0.0304 (-2.946) 0.564 (7.604) 0.402 (1.482) 1.531 (2.883) 0.548 (2.155) 1.261 (3.531) -0.100 E-4 (-1.361) 0.070 (1.843) -171.75 0.3958

-0.284 E-3 (-1.246) 6.603 (3.813) 5.294 (3.153) 5.660 (3.385) -0.0104 (-5.094) -0.0297 (-2.893) 0.562 (7.564) 0.409 (1.520) 1.524 (2.872) 0.526 (2.084) 1.236 (3.465) -0.102 E-4 (- 1.394) 0.0743 (1.944) -171.91 0.3953

Notes: T-stats are in parentheses. Dependent variable: Choice of mode. Number of records = 354. Number of cases = 510. L(0)= -305.80.

specified mode for a given change in the perceived risk coefficient of another mode). The most interesting result is that the direct elasticity of bicycling is equal to - 1.19, or less than negative one. This implies that for a given percentage reduction (increase) in bicycling risk there will be an increase (decrease) in the percent share of Table 7. Aggregate direct and cross elasticities from Model A Elasticities Change in modal shares of: Change in perceived risk (SRP) of: Automobile Bicycle Transit Walking

Automobile -0.0843 0.0196 0.0422 0.00936

Bicycle

Transit

Walking

0.177 0.403 0.0469 -1.186 0.0516 0.354 0.0294 -0.379 0.0295 0.325 0.0578 -0.197

Note: Values along diagonal are direct elasticities, all others are cross elasticities.

bicycle commuting that is greater than the percent reduction (increase) in risk. As will be shown in the next section, this could have adverse consequences on net transportation system risk, if bicycling is riskier than other modes and the perceived risk reductions do not correspond with reductions in objective risk. For example, if people perceive bicycle paths (separated from streets) to be safer, but they actually are not, this could increase the total accident rate by increasing the level of bicycle transportation. G&rder, Leden, and ThedCen (1994) show that while “experts” may rate on-street bicycle lanes as less safe for bicyclists, most cyclists perceive them to be safer. Note that increases in bicycle safety tend to have a relatively high cross elasticity with respect to walking (cross elasticity = 0.35). That is, many of the people shifting to bicycles will come from those who are currently walking to work. The great-

Perceived risk and modal choice

515

Table 8. Regression results with SRP as dependent SRP-Bicycle

Dependent

Constant Travel time (alternative specific) Sample dummy variable (= 1 for general sample) Age Sex (Male = 1) Income Home-work population density product Adiusted R-Sauared:

0.000 0.191

0.096

-0.003 -0.147 -0.015 -0.127

Std coefficient

T-test

16.500 4.139 2.182 -0.062 -3.397 -0.335 -2.778

0.000 0.168 0.003 -0.180 -0.118 -0.063 0.007

17.8% 3.875 0.066 -4.010 -2.712 -1.401 0.148

0.095

est shift from increases in automobile safety is a decrease in transit share (cross elasticity = 0.40). The direct and cross elasticities for automobile shares are all relatively small, indicating that changes in the perceived risks of other modes will have a relatively minor impact on changes in the total share of commuters using automobiles. For a more detailed discussion of short-run and long-run elasticities see Noland and Kunreuther (forthcoming). The impact of these effects on net transportation system risk will be modelled using sample enumeration techniques below. Determinants

of risk perception

SRP-Automobile

T-test

Std coefficient

variable:

variable

measures

This section presents results of ordinary least squares regressions that attempt to explain some of the determinants of the perceived risk measures discussed above. The regressions relate both demographic and commute travel pattern variables to the perceived risk measures. Examining the risk measures for alternative modes gives some explanation for differences in how the risks are perceived. The risk scores show some relationship to travel times. The regressions using the SRP score as a dependent variable in Table 8 show this is true for some of the modes. Travel time has a high level of significance in predicting the SRP score. For bicycle and automobile risks it is highly significant and for walking it is above the 90% confidence level. Perceived transit risk does not show any relationship to travel time. Travel time represents a proxy for exposure to a risky situation. The longer the commute, the greater the exposure to risk (except in the case of transit). A study by Barnard (1987) using household survey data that measured perceived bicycle safety, found that bicycle safety is negatively related to the amount of time spent travelling on major roads (but not on smaller side roads). The ESRP score regressions have a significant

0.073

SRP-Transit Std coefficient 0.000

0.035 0.088 -0.090 -0.065 -0.083 -0.081 0.026

T-test 15.187 0.732 1.949 -1.966 -1.461 -1.797 - 1.672

SRP-Walkinn Std coefficient 0.000

0.087 0.118 0.048 -0.129 -0.056 -0.226

T-test 10.863 1.915 2.713 1.088 -2.980 -1.264 -4.929

0.094

travel time coefficient in all cases except for perceptions of automobile risk (see Table 9). The increase in significance was surprising since this variable was expected to show less relationship to travel times, since it weights the severity index more heavily than the probability index. The sample dummy variable indicates that there is a difference between the two samples. For the SRP measure, the general sample has a statistically higher risk perception than the bicyclist sample. However, for the ESRP measure, the coefficient is significant only for bicycling risk. Various demographic variables were also tested. These included age, sex, and income. Income does not show any significance, except at the 90% level for the SRP measure for transit. Generally the sign on the coefficient is negative, indicating that higher income people have a lower perception of the risk of the modes (perhaps because they purchase more safety devices). Age is significant only for the SRP measure for automobiles and transit. The greater the age the lower the risk perception. This is an interesting result, especially for automobile risk perceptions, since younger people generally are riskier drivers. Therefore, this implies that they may recognize this fact. Sex is significant in a few cases also. For both the bicycle risk scores and for automobiles and transit using the SRP measure, the coefficient is negative indicating that males have a lower perception of the risk of the transportation modes. This is also an interesting result, since males generally are riskier drivers. Therefore, this may imply they do not perceive that they face greater risks than female drivers. An alternative interpretation is that female drivers are safer because they do have a better understanding of the risks. The last variable included in the models is a proxy measure for home and work location choice. This is the product of the home population density and the workplace population density. For example,

516

R. B. NOLAND

Table 9. Regression results with ESRP as dependent ESRP-Bicycle

Dependent

variable:

Constant Travel time (alternative specific) Sample dummy variable (= 1 for general sample) Age Sex (Male = 1) Income Home-work population density product Adjusted R-Squared:

Std coefficient 0.000 0.117 0.123 0.066 -0.130 -0.049 -0.079

ESRP-Automobile

T-test 1.199 2.493 2.725 1.462 -2.961 -1.064 - 1.708

0.062

someone commuting from an urbanized area to another urbanized area (or within the same urbanized area) would have a higher value for this variable than someone commuting within or between suburbs. City to suburb or suburb to city commutes would have intermediate values. The risk faced by these alternative commute patterns may differ. For example, bicyclists and pedestrians may face greater risks in urbanized areas with larger amounts of traffic. This seems to be the case as the coefficients on this variable show. It is only highly significant (with a negative sign) for bicycle and walking risk perceptions, using both measures, although less so for the ESRP measure. It is not at all important in determining automobile risk perceptions, implying that they have similar risk perceptions independent of their commute patterns. Suburban automobile commuters may face different risks than urban commuters, such as higher traffic speeds on large arterial roads as opposed to high traffic volumes and interactions on city streets. That is, they may perceive similar risks, but the source of the risk may vary. Transit shows a smaller (90%) level of significance for the SRP measure. This analysis has highlighted some of the underlying factors that may explain the risk perception scores measured. The importance of travel time indicates that people evaluate the impact of exposure to risk in forming their perceptions; i.e. increased travel times result in an increased level of exposure to risk. The importance of commute patterns for the nonmotorized modes implies that large traffic volumes on city street networks is a source of perceived risk. It should be noted that the adjusted R* for all the models is relatively low, implying that much of the variance in the risk perception measures has not been explained. In any case, these regressions show some of the sources of these risk mea-

variable ESRP-Transit

ESRP-Walking

Std coefficient

T-test

Std coefficient

T-test

Std coefficient

T-test

0.000 0.003 0.030 -0.074 -0.050 0.009 0.024

1.845 0.071 0.654 -1.586 - 1.099 0.199 0.523

0.000 0.157 -0.061 0.057 -0.060 -0.005 -0.007

-1.037 3.324 -1.350 1.239 -1.343 -0.109 -0.154

0.000 0.126 0.070 0.016 -0.003 -0.047 -0.103

1.173 2.701 1.560 0.360 -0.073 - 1.013 -2.181

0.000

0.021

0.031

sures and gives some confidence in the metrics used for measuring perceptions of risk.

Some examples of the impacts on transportation fatalities from reductions in objective and perceiued risk

A key question in risk compensation studies is whether the impact of a risk reduction measure can actually increase the net risk to society. Peltzman (1975) indicated that this could occur. Most other studies, however, have not found the effect to be so extreme that net accidents actually increase. The models presented here can be used to demonstrate how changes in modal shares can result from a given change in risk perceptions and the potential consequences on total fatalities. The estimations that follow are intended to show how modal shifts due to increases in safety can have the effect of offsetting some of the expected safety improvements. These results are presented only to demonstrate how this potential effect could occur and should not be interpreted as definitive predictions. These estimates are based on Model A in Table 4. Changes in modal shares are calculated using the sample enumeration procedure described by Ben-Akiva and Lerman (1985), with the data weighted by income categories. Predictions were forecast for l%, 5%, lo%, and 50% decreases in the perceived risk of commuting by automobile and also for equivalent reductions in the perceived risk of commuting by bicycle. A combined decrease in the perceived risk of commuting by car and an increase of the same percentage in the perceived risk of travelling by bicycle was also modelled. This will take into account any impacts from bicyclists responding to increased automobile volumes (or driving intensity) that may make the bicycling environment more risky.

Perceived risk and modal choice

Table 10. Predictions of modal shares using sample enumeration Percent reduction in perceived risk of commuting by car

Car Bicycle Transit Walk

Base case

1%

5%

10%

50%

67.478% 6.502% 9.324% 16.696%

67.535% 6.490% 9.286% 16.689%

67.760% 6.445% 9.138% 16.657%

68.037% 6.389% 8.956% 16.618%

70.072% 5.989% 7.626% 16.313%

Percent reduction in perceived risk of commuting by car and proportionate increase in perceived risk of commuting by bicycle Car Bicycle Transit Walk

67.478% 6.502% 9.324% 16.696%

67.548% 6.414% 9.291% 16.757%

67.824% 6.069% 9.161% 16.946%

68.159% 5.658% 9.001% 17.182%

70.497% 3.108% 7.801% 18.594%

Percent reduction in perceived risk of commuting by bicycle Car Bicycle Transit Walk

67.478% 6.502% 9.324% 16.696%

67.465% 6.579% 9.3 19% 16.637%

67.411% 6.895% 9.299% 16.395%

67.341% 7.305% 9.274% 16.080%

66.694% 11.064% 9.037% 13.205%

The predicted modal shares (and the base case prediction) are all shown in Table 10. In the cases where the perceived risk of automobile commuting decreases, the level of automobile commuting increases. Transit usage and bicycle commuting both decrease, the latter more so when the perceived risk of bicycling increases due to the increased safety of automobile commuting (the percentage increase in bicycie risk perceptions is assumed to equal the percentage decrease in automobile risk perceptions). Note that the share of commuters walking to work increases when the perceived risk of bicycling increases; this indicates a shift from using a bicycle to walking. When the perceived risk of bicycling is decreased, much of the total increase in bicycle usage comes from people who previously walked (as was previously discussed in the section on aggregate elasticities). In the calculations that follow, it is assumed that the fatal accidents from one million commute trips a day for one year are being estimated. Total person-km travelled for each mode are calculated based on the average commute distances, by mode, in the sample. These are shown in Table 11. Table Car Bicycle Transit Walk

11. Average commute distances mode for sample 14.04 9.29 20.06 1.84

km km km km

8.72 5.77 12.46 1.14

by miles miles miles miles

517

Table 12. Fatality rates for each mode per 100 million personkilometers (and miles) travelled Car Bicycle Transit Walk

0.65/person-10s 6.25/person-108 O.O3/person-108 6.25/person-108

km km km km

l.O5/person-108 lO.OS/person-108 O.OS/person-108 lO.OS/person-108

miles miles miles miles

Applying estimates of fatal accident rates then gives an estimate of the total fatalities per mode, which is then used to calculate the net total fatalities for each change in risk perceptions. The accident rates used are shown in Table 12. The automobile accident rate used is the 1991 aggregate fatality rate for automobile passengers reported by the National Safety Council (1992). The transit fatality rate is also taken from this source but is based on the highest level reported during the last 15 years. This is done to prevent overestimation of the safety of transit. Bicycle and pedestrian accident rates are more difficult to estimate since they are not caIculated by miles travelled. For bicycle fatality rates an estimate provided by Mitchell (1978) of five times the total motor-vehicle fatality rate is used (this value is 1.25 per 10’ vehicle-km travelled). It is also assumed that the pedestrian rate is the same as the bicycle accident rate. These values are intended to provide a rough estimate of the objective risk facing commuters of each mode. The bicycling and walking risks are probably very speculative, while the automobile and transit risks are aggregates that include trips other than commute trips.* One problem with this analysis is that it uses perceptions of the risk of modes. Since there is no data available to correlate these perceptions with actual objective measures of risk, two results are presented for each predicted fatality rate. One assumes that the objective risk level correlates precisely with people’s perceptions. That is, when the perceived risk is reduced (or increased) by a given percent, the corresponding fatality rate is adjusted by the same percent. This may, however, be an *As noted earlier, transit injury risk may be higher when accessed as a pedestrian. The aggregate data are used here without this adjustment for three reasons. First, many automobile users in Philadelphia use on-street parking and must also access this mode as a pedestrian. Since the logit model does not include those without access to transit (i.e. mainly suburb-to-suburb commuters) those suburban residents with garages are much less likely to be included as having transit as a feasible option. Second, the analysis focuses on fatalities and not injury rates. It may be that minor injuries are more likely on transit than in a personal vehicle, but that is not the focus here. Third, and last, this analysis is merely meant to demonstrate how offsetting behavior could work. If transit usage were more risky than automobile usage, it could also show how decreases in transit risk are partially offset by modal shifts.

R. B. NOLAND

518

Table 13. Net imnact on fatalities from decreases in risk uerceotions Decrease in automobile risk perceptions

0% (Base case)

1% 5% 10% 50% Slope*

(1) Equal changes in objective and perceived risk 59.60 59.28 57.98 56.36 42.99 -0.324

(2)

with and without eauivalent decreases

Decrease in automobile risk perceptions with increase in bicycle risk perception

No change in objective risk

(3) Equal changes in objective and perceived risk

59.60 59.58 59.53 59.47 59.03 -0.013

59.60 58.28 57.97 56.25 40.55 -0.335

in obiective risk

Decrease in bicycle risk perceptions

(4) No change in objective risk

(5) Equal changes in objective and perceived risk

No change in objective risk

59.60 59.40 58.64 57.73 52.18 -0.187

59.60 59.59 59.53 59.39 54.42 -0.021

59.60 59.78 60.53 61.51 70.46 0.191

(6)

*Slope is calculated based only on values between 0% and lo%, due to increased error in forecasting for larger ranges.

unrealistic assumption for some policy measures that reduce risk. In some cases, perceptions of risk might change while the actual objective level of risk does not (or perceptions may change more than the underlying objective changes in risk). Results are also calculated assuming no change in the objective level of risk. The results for both these cases are shown in Table 13. As can be seen in Table 13, columns (1) and (2), in the cases where the perceived risk of automobile commuting is decreased, the net level of transportation fatalities decreases. Even when the perceived risk of automobile commuting is induced without any corresponding change in objective risk levels there is a slight reduction in total fatalities. For this latter case the effect is trivial, amounting to 0.572 fatalities per one million commuters for a 50% reduction in perceived risk. These results show that the modal shifts to automobile commuting from riskier modes, triggered by the decrease in perceived risk, result in net reductions in fatalities. Net decreases in the objective risk of commuting by automobile offset the risk due to increased levels of commuting by automobile. Column (3) shows the case in which the perceptions of bicycling risk increase (as well as the objective level of bicycling risk) in combination with decreases in automobile risk. This results in even greater reductions in total fatalities due to bicycle commuters’ shifting to safer modes. Even when the objective level of bicycling risk is not increased, one still gets a decrease in net fatalities, as shown in column (4). In the case where the perceived risk of bicycling is decreased, due to an equivalent decrease in objective risk levels (column 5), there is a small net reduction in fatalities. When perceptions of the risk of bicycling are reduced with no equivalent reduction

in objective risk levels (column 6), then there is an increase in fatalities. Most of the increase in bicycle commuters comes from people who previously walked. If the relatively high fatality rate assumed for walking is reduced, then the decrease in bicycling risk could result in an increase in net fatalities, even when objective risk levels decrease. For example, the level of walking risk at which no change in fatalities occurs for a 1% change in bicycling risk is about 7.85 fatalities per hundred million miles traveled. This is certainly a plausible level of risk. When this risk level is set equal to zero, i.e. no walking fatalities occur, the increase in net fatalities from reducing the risk of bicycling (with changes in objective risk levels equal to changes in perceived risk levels) is only 0.026 fatalities per year for a 1% reduction in bicycling risk for one million commuters, and an increase of 3.18 for a 50% reduction in bicycling risk.

Table 14. Net impact on bicycle fatalities from decreases risk perceptions with and without equivalent decreases objective risk

in in

Decrease in bicycle risk perceptions

0% (base case)

1% 5% 10% 50% Slope*

(1)

(2)

Equal changes in objective and perceived risk

No change in objective risk

18.85 18.89 18.99 19.06 16.04 0.021

18.85 19.08 19.99 21.18 32.08 0.233

*Slope is calculated based only on values between 0% and lo%, due to increased error in forecasting for larger ranges.

519

Perceived risk and modal choice

Table 14 provides additional detail on how the number of bicycle fatalities may change with reductions in perceived and objective risk levels. When equal changes in perceived and objective risk levels occurs, bicycle fatalities increase; this is due to the fact that reductions in perceived risk attract a proportionately greater number of people to this mode. Note that fatalities are shown to have decreased when risk levels are reduced by 50%, but this may be due to inaccuracies in the model introduced by forecasting beyond smaller percentage changes. In the case with no equivalent reductions in objective risk, bicycle fatalities increase. If one assumes that the 50% risk reduction estimates are accurate, then this implies that there could be between a 15% reduction in bicycle fatalities and as much as a 70% increase in fatalities, depending on the relative response between actual objective risk level changes and perceptions of risk. The slopes shown in Table 13 are from a plot of the fatalities versus the percentage reduction in perceived risk (between 0% and 10%). These can be interpreted as representing the percentage change in fatalities for a given percentage change in risk perceptions. For example, in column (l), a one-percent reduction in perceived risks of automobile commuting results in a -0.324% reduction in total commuting fatalities. This indicates that the modal shift results in a risk compensating mechanism for the transportation system that gives the result of offsetting some of the reductions in risk. When perceived risks change without an underlying change in objective risk (column 2), the slope is still negative, although in this case the slope is almost zero, indicating virtually no change in fatalities. For reductions in bicycling risk perceptions only (column 6), the slope is 0.191, indicating an increase in fatalities. A value of - 1 for the slope would indicate that a given percentage reduction in objective risk results in the same percentage reduction in fatalities. As can be seen, none of the slopes come close to this value. Another possibility is that a change in objective risk triggers a less than equal change in perceptions of risk. Table 15 shows this result when a change in the objective risk of automobile commuting (column 1) triggers only half the equivalent reduction in perceptions of risk. As can be seen, the slope is -0.6357, indicating that a 1% reduction in perceived risk (or in this case, a 2% reduction in objective risk) still does not result in a 1% reduction in fatalities. The slope is steeper than the value shown in Table 13, column (l), implying less offsetting behavior due to less switching to the automobile mode from safer modes (as would be expected if people perceived the objective risk reductions correctly). Compare

Table 15. Net impact on fatalities when decrease in objective risk is double the decrease in perceived risk

0% (base case) 1% 5% 10% 50%* Slopet

(1) Decrease in automobile objective risk double the decrease in perceived risk

Decrease in bicycle objective risk double the decrease in perceived risk

59.60 58.97 56.43 53.24 26.95 -0.636

59.60 59.40 58.54 57.27 38.38 -0.233

(2)

*At the 50% level, column (1) contains zero automobile fatalities and column (2) contains zero bicycle fatalities. tSlope is calculated based only on values between 0% and lo%, due to increased error in forecasting for larger ranges.

the fatalities for the 10% case in Table 13, column (l), with the 5% reduction in Table 15, column (l), which are for the same level of reduction in objective risk. When perceived risk is only half as large (in Table 15) the number of fatalities is greater. For the bicycle case (column 2), the effect is not as great. Therefore, it seems if people can be made to believe that a risk reduction measure is not as effective as it actually is, the actual reductions in transportation system fatalities may be greater. The assumed relationships between objective risk reductions and perceptions of risk may occur for specific policy measures. The introduction of air bags into new automobiles can serve as an example. If people are aware that they have an air bag in their car, they will reassess their perceived risk accordingly. If they do so accurately, this represents the case where objective risk reductions lead to an equal reduction in perceived risks. If people either do not assess the change in perceived risks accurately or are unaware that they have an air bag in their car, then the aggregate objective risk levels may decline more than perceptions of risk. The case where people overestimate the risk reductions could also occur, if their assessment of the risk reductions is greater than the actual reduction in risk from having an air bag (e.g. this could occur if automobile companies aggressively market the safety of air bags by implying that people can survive a head-on collision, as was done by one automobile company after this actually happened). In the case of bicycles, similar effects can occur. For example, if a policy of constructing on-street bicycle lanes is implemented, it is possible that net fatalities could increase. This will of course depend on the relationship between perceptions of risk and the actual objective reductions in risk from con-

520

R.B.

strutting bicycle lanes. Another example is the case of off-street bicycle paths. Some studies have suggested that these are more dangerous than using streets with traffic (Kaplan 1975; Johnson et al. 1978; G&rder et al. 1994). These facilities tend to attract people who assume that automobiles are not present, even though in many cases they are; for example, at intersections with streets. Based on this faulty assumption, they perceive that the risk of an accident is lower than it may actually be. It is likely that offstreet bicycle paths will decrease the average severity of accidents, but actual objective risk levels may not correspond with perceptions. CONCLUSIONS The results of this research give evidence that increases in the perceptions of the risk of using a given transportation mode may reduce the probability of commuting by that mode. If risk is an important determinant of modal choice, it would lower the probabilities of commuting by bicycle and automobile and increase the relative shares of transit and walking (which are perceived as relatively safer) for those people with a choice of modes. In particular, the choice of the bicycle mode for commuting is very sensitive to perceptions of risk, with an aggregate elasticity of less than --I; i.e. the proportional increase in the percentage of people commuting by bicycle is greater than the percentage decrease in perceived risk. It is important to note that people must have a choice of modes for this effect to occur. In many cases people are captive to a specific mode for their work trip (and in the United States this is generally the automobile). The analysis of the net impact of assumed changes in objective risk measures and their translation into changes in risk perceptions shows how modal shifts within the transportation system can result in offsetting compensations for the risk reductions. That is, reductions in commute travel fatalities do not fall proportionately with reductions in risk. This is also dependent on the relationship between objective risk and how risk is perceived. While this methodology offers some improvements to previous studies that have examined risk compensation, it also has some drawbacks. The first is defining an adequate measure of risk perceptions. While the measures used gave results that indicate a compensating effect within the transportation system, it may be more useful to allow people to have more concrete definitions of risky situations when they evaluate their perceptions. For example, people may be able to evaluate their perceptions more accurately based on specific risky situations, such as

NOLAND

evaluating the risks of a large proportion of speeding drivers or of bad road designs. For example, with regard to bicycle risk, it has been shown that the lack of adequate road space along the shoulder influences people’s decision to commute by bicycle (Noland and Kunreuther forthcoming). This is in spite of the fact that the probability of an accident, but not the severity, is greater at intersections than between them (Cross and Fisher 1977). Another possibility is that people make choices about their commute route to minimize risk. For example, this may lead people to avoid a congested freeway and use an automobile on surface arterial or local residential streets instead. The modeling methodology did not account for these decisions, which may precede the choice of mode. The actual driving (and bicycling) behavior of individuals after they have selected their mode is also a factor, as Peltzman (1975) originally hypothesized. This research has shown another methodology for analyzing the hypothesis that reductions in transportation risk can have less than the expected impact on total transportation system fatalities, due to partial compensation for the risk reductions embodied by shifts between modes. Individual perceptions of risk were related to modal choice behavior, avoiding some of the criticisms of earlier work. The results are supportive of at least partial risk compensation within the transportation system. Further research should focus on what specific factors determine perceptions of risk and how they relate to objective safety measures and which of these may influence not only modal choice but other elements of driving intensity. Acknowledgements-1 would like to thank Howard Kunreuther, Pete Fielding, Amihai Glazer, and an anonymous referee for valuable comments and suggestions. I am also indebted to Kenneth Small for suggesting the analysis in the final section of this paper. This research was begun with the support of the Center for Energy and the Environment at the University of Pennsylvania and completed with funding support from the University of California Transportation Center. Any errors or omissions are the responsibility of the author.

REFERENCES Barnard, P. O., Perceived bicycle safety and commuting use: Results from the ATDATAS study. In: Federal Office of Road Safety, editor. Bikesafe 86. Newcastle, Australia: Australian Government Printing Service; 1987:605-606. Ben-Akiva, M.; Lerman, S. R. Discrete choice analysis: Theory and application to travel demand. Cambridge, MA: The MIT Press: 1985. Blomquist, G. C. Motorist use of safety equipment: Expected benefits or risk incompetence? Journal of Risk and Uncertainty 4: 135-152; 1991. Chirinko, R. S.; Harper, E. P. Jr. Buckle up or slow

Perceived risk and modal choice down? New estimates of offsetting behavior and their implications for automobile safety regulation. Journal of Policy Analysis and Management 12:270-2%; 1993. Cross, K. D.; Fisher, G. A study of bicycle motor-vehicle accidents: Identification of problem types and countermeasure approaches. Volume 1. Washington DC: US Department of Transportation. National Highway Traffic Safety Administration; 1977. Conybeare, J. A. C. Evaluation of Automobile Safety Regulations: The case of compulsory seat belt legislation in Australia. Policy Sciences 12:27-39; 1980. Dillman, D. A. Mail and telephone surveys: The total design method. New York: John Wiley & Sons: 1978. Evans, L. Human behavior feedback and traffic safety. Human Factors 27555-576; 1985. Evans, L. Risk homeostasis theory and traffic accident data. Risk Analysis 6:81-94; 1986. Garder, P.; Leden, L.; Thedeen, T. Safety implications of bicycle paths at signalized intersections. Accident Analysis & Prevention. 26:429-439; 1994. Graham, J. D. On Wilde’s theory of risk homeostasis. Risk analysis 2:235-237; 1982. Graham, J. D.; Garber, S. Evaluating the effects of automobile safety regulation. Journal of Policy Analysis and Management 3:206-224; 1984. Holdgate, M. W. Risks in the built environment. Proceedings of the Royal Society, London, A 376:151-165; 1981. Joksch, H. C. Critique of Sam Peltzman’s study: The effects of automobile safety regulation. Accident Analysis & Prevention 8: 129-137; 1976a. Joksch, H. C. The effects of automobile safety regulation: Comments on Peltzman’s reply. Accident Analysis & Prevention 8:213-214; 1976b. Johnson, M. S.; Jurik, N. C.; Kreb, A. R.; Rose, T. L. The Wheels of misfortune: A time series analysis of bicycle accidents on a college campus. Evaluation Quarterly 2608-619; 1978. Kaplan, J. A. Characteristics of the regular adult bicycle user. San Francisco: Federal Highway Administration; 1975. Lie, T. Personskader ved bussreiser. En samlet oversikt over skader og risiko for forare og passasjerer pa bussene, for gange till og fra stoppested og fra andre trafikanter. Oslo: Tranportokonomisk Institutt. TOI-notat nr 648; 1983. McCarthy, P. S. Seat belt usage rates: A test of Peltzman’s hypothesis. Accident Analysis and Prevention 18: 425-438; 1986. Mitchell, C. G. B. Cycle use in Britain. In: Cycling as a

521

mode of transport, the proceedings of a symposium held at the Transport and Road Research Laboratory. Supplementary report 540. Crowthorne, England: Transport and Road Research. Department of the Environment. Department of Transport; 1978. National Safety Council. Accident Facts: 1992. Chicago, IL: Author; 1992. Noland, R. B. The Role of risk in policies to promote bicycle transportation. Dissertation in energy management and policy. Philadelphia: University of Pennsylvania; 1992. Noland, R. B.; Kunreuther, H. Short-run and long-run policies for increasing bicycle transportation for daily commute trips. Transport Policy: (forthcoming). Orr, L. D. Incentives and efficiency in automobile safety regulation. Quarterly Review of Economics and Business 22143-63; 1982. Peltzman, S. The effects of automobile safety regulation. Journal of Political Economy 83:677-725; 1975. Robertson, L. S. State and federal new car safety regulation: Effects on fatality rates. Accident Analysis and Prevention 9:151-156; 1977a. Robertson, L. S. A critical analysis of Peltzman’s “The Effects of Automobile Safety Regulation.” Journal of Economic Issues 11:587-600; 1977b. Robertson, L. S. Automobile safety regulations and death reductions in the United States. American Journal of Public Health 71:818-822; 1981. Singh, H.; Thayer, M. Impact of seat belt use on driving behavior. Economic Inquiry 30649-658; 1992. Slavic, P.; Fischhoff, B.; Lichtenstein, S. Accident probabilities and seat belt usage: A psychological perspective. Accident Analysis and Prevention 10:281-285; 1978. Traynor, T. L. The Peltzman hypothesis revisited: An isolated evaluation of offsetting driver behavior. Journal of Risk and Uncertainty 7:237-247; 1993. U.S. EPA. Bicycling and air quality information document. Washington DC: U.S. Environmental Protection Agency. Office of Transportation and Land Use Policy; 1979. Wilde, G. J. S. The theory of risk homeostasis: Implications for safety and health. Risk Analysis 2:209-225; 1982. Wilde, G. J. S. Evidence refuting the theory of risk homeostasis? A rejoinder to Frank P. McKenna. Ergonomics 27:297-304; 1984. Zlatoper, T. J. Regression analysis of time series data on motor vehicle deaths in the United States. Journal of Transport Economics and Policy 18:263-274; 1984.