Feasibility Study for Highway Hazardous Materials Bulk Package Accident Performance Data Collection (2013)

Below is the uncorrected machine-read text of this chapter, intended to provide our own search engines and external engines with highly rich, chapter-representative searchable text of each book. Because it is UNCORRECTED material, please consider the following text as a useful but insufficient proxy for the authoritative book pages.

85 S e c t i o n 4 The third objective of HMCRP Project 07, to identify methodologies for analyzing the collected data, was achieved through a literature review and the development of a model that incorporates measures of accident performance. The lit- erature review focuses on studies and reports that evaluate crash risk and the risk of transporting hazardous materials. Methodologies for analysis of accident performance data are explored and a hazardous materials transportation risk model is developed to incorporate measures of accident per- formance. This transportation risk model is presented, and input values for most of the variables can be derived from existing studies and data sets, with the exception of the con- ditional probability of release and the estimated quantity of release. These two variables relate directly to bulk pack- age accident performance. The conditional probability of release is a function of the damage incurred in an accident and the bulk package’s ability to withstand that damage. Similarly, the expected quantity of release is a function of the size of the breach (related to the package’s ability to withstand damage), the physical properties of the hazardous material, and the time elapsed before the breach can be plugged or the hazardous material can be off-loaded. The methodologies for developing equations that esti- mate the conditional probability of release and the estimated quantity of release are discussed. Furthermore, variables that should be considered in these equations are identified and information pertaining to these variables, collected as part of the pilot study, is summarized to identify possible cor- relations. However, it must be stressed that the pilot study data are over-weighted in accidents that had a release and under-weighted in non-release accidents. For this reason, the analyses and evaluations of the pilot study data pre- sented in this report are included here only to illustrate the process. The resultant statistics should not be considered valid estimates of the accident performance of highway bulk packages. Literature Review The purpose of the accident reporting system being inves- tigated in this project is to enable the analysis of bulk pack- age performance in the event of an accident. The database would be designed to accomplish the goal of determining container-design-specific risk factors for hazardous material cargo tanks and intermodal tanks. In addition, this project will identify possible methods to analyze such an accident database for transportation risk analysis. The sources listed below were reviewed for their approaches to support trans- portation risk analysis activities. Comparative Risks of Hazardous Materials and Non- Hazardous Materials Truck Shipment Accidents/Incidents (Battelle 2001) presents a comparative risk assessment of hazardous material and non-hazardous material shipments. Battelle’s (2001) analysis started by examining Class 3 (flam- mable and combustible liquid) accidents/incidents within a 1-year period using the HMIRS database. Subsequently, the analysis was expanded to include Division 2.1 (flammable gases) and Class 8 (corrosives) over a period of 3 years. In the final phase of the project, the analysis was expanded to all types of hazardous material cargo tank truck accidents. The HMIRS database was supplemented with non-spill accidents from the MCMIS database, and the degree of underreporting was estimated using factors developed by comparing HMIRS with other databases for eight states. The degree of under- reporting determined for eight states was used to estimate the degree of underreporting for the nation. In the next step, Battelle (2001) developed event trees for each type of hazard- ous material and the probability of an event occurring was calculated using information, adjusted for underreporting, from the HMIRS and MCMIS databases. Where a hazard- ous material class/division did not have enough accidents to make a statistical evaluation, information from the HMIRS and MCMIS was supplemented with either probabilities from a similar class/division or theoretical accidents. Battelle (2001) Identifying a Methodology for Data Analysis

86 then determined the likelihood that an en-route accident would occur in a year and developed an impact cost associated with each hazardous material class/division. Accident risk and cost per mile were calculated for each hazardous material class/ division. Finally, accident risk and cost per mile were calcu- lated for non-hazardous materials and compared to risk and cost per mile of hazardous materials. A National Risk Assessment for Selected Hazardous Materi- als in Transportation (Brown et al. 2000) presents a quanti- tative risk assessment that considered in-transit releases for three types of hazardous materials: toxic inhalation hazards (TIHs), flammable materials, and explosives. The purpose was to assess and define the nature of the risk of transport- ing hazardous materials. The data used in this analysis were generated from PHMSA HMIRS reported accidents and, for non-gasoline accidents, national commodity flow surveys; a detailed consequence model; routing models; and National Weather Service meteorological observations. The Chemical Accident Stochastic Risk Assessment Model (CASRAM) used this data to generate samples of possible accidents. The model begins by using accident rates to determine whether an acci- dent occurs in a particular run. The occurrence of a release is predicted based on conditional probabilities of release devel- oped by Harwood and Russell (1990). Once a release occurs within a sample run, the size and shape of the affected zone is modeled based on PHMSA’s HMIRS percentage of capacity released. With regard to highway bulk packages, release prob- abilities for MC 306 (using gasoline and fuel oil), MC 312 (using sulfuric acid), and MC 331 (using ammonia and liquid petroleum gas) cargo tanks were developed by dividing the number of HMIRS reported releases by the estimated num- ber of accidents. All probabilities were adjusted by a factor of 1.5 to account for underreporting. The conditional prob- ability of release was estimated to be 6.5% for MC 306 con- tainers, 4.0% for MC 312 steel containers, 1.5% for MC 312 stainless steel containers, 2.5% for MC 331 steel containers with a 0.187-inch thick shell, and 1.05 for MC 331 contain- ers used to transport chlorine and sulfur dioxide (where the lower probability is due primarily to the increased robust- ness of these containers). However, the release probabilities vary depending on the type of chemical transported in the container. The national risk assessment used percentage of capacity released because it offered the most robust statistic. Since the packaged amount was not recorded at the time of the analysis, there is an implicit assumption that all compart- ments within the bulk package are full. The analysis indicated that the amount of material released varies depending on the accident severity and the container specification. The 25th, 50th, and 75th discharge percentiles for MC 306, MC 307, and MC 312 cargo tanks are provided in Table 53. In The Dimensions of Motor Vehicle Crash Risk (Wang, Knipling, and Blincoe 1999), General Estimates System (GES) data were analyzed in terms of both the specific role of a vehi- cle in a specific type of crash and the body type of the vehicle having a critical participating role in the accident. Cargo tank trucks were grouped with other combination or single unit trucks. The analysis developed several statistics by considering several metrics (numerators), including the number of crashes, the number of vehicles having a critical participating role, the number of vehicles involved, the number of people involved, the crash severity, economic cost, and comprehensive societal value costs. Of particular note, Wang, Knipling, and Blincoe (1999) addressed the injury severity level by assigning numer- ical values to injury severity categories defined by the National Safety Council (NSC 2007) and combining serious and fatal injuries because of the unacceptably large sampling errors associated with the small fatality estimates for specific crash/ vehicle types. These metrics were standardized in a number of ways using denominator data such as the U.S. annual total number of reported crashes (by type), the total number of miles traveled per year in the United States, the annual num- ber of U.S. registered vehicles, the expected operational life of a vehicle (by vehicle type), and the average number of crashes per driver over his/her driving career. After percent- age estimates and derived statistics were calculated, the crash and injury statistics were rounded to the appropriate levels of significance. For example, crash and injury counts over 2,000 were rounded to the nearest 1,000, and counts less than 2,000 were rounded to the nearest 100. The paper also provides a discussion concerning the most relevant referents from which Wang, Knipling, and Blincoe (1999) assert that “identifica- tion of the most relevant dimensions of motor vehicle crash risk is even more fundamental to developing a framework for enlightened safety benefits assessment and decision-making.” Cargo Tank Specification Percent of Capacity Released 25th percentile Median 75th percentile MC 306 2.6% 22% 60% MC 307 0.16% 1.9% 18% MC 312 1.3% 11% 50% Table 53. Discharge percentiles for cargo tanks.

87 Large Truck Crash Causation Study (LTCCS) Analysis Series: Using LTCCS Data for Statistical Analysis of Crash Risk (Hedlund and Blower 2006) discusses the strengths and weaknesses of statistical analyses performed using the LTCCS database, a nationally representative sample of 963 injury and fatal crashes involving large trucks. Hedlund and Blower (2006) investigate crash causes in terms of risk-increasing factors because they do not wish to determine fault, crashes typically have more than one cause, and risk-increasing fac- tors are based on facts (not inferences). Limitations of the LTCCS database, identified in this paper, include difficulties in the following: • Applying statistics on a national scale because national ref- erents (denominators) are not very accurate. • Determining confidence errors of statistics due to the multi- stage sampling approach. Hedlund and Blower (2006) esti- mate that first order statistics will have an error of 3% while comparisons of two types of crashes will have an error of approximately 10%. • Reducing bias due to incomplete data or second-hand data. Uncertainty in the Estimation of Risks for the Transport of Hazardous Materials (Saccomanno, Stewart, and Shortreed 1993) discusses uncertainties associated with risks of hazard- ous material transportation. Saccomanno, Stewart, and Short- reed (1993) assert that hazardous material transportation risk analysis requires an estimation of several elements of the risk analysis process, including the probability of release given that an accident has occurred. The paper identifies the follow- ing three problems: • Control—using average exposure (referent) or using expo- sure resulting from controlled conditions as opposed to actual exposure. Saccomanno, Stewart, and Shortreed (1993) emphasized the need to document several environmental factors in an accident record so that variations from mean values can be explained. • Omission—excluding accidents because of specific param- eters (i.e., excluding an accident because it occurred at an intersection). • Bias—inputting values conservatively can result in an over- estimation of risk, especially when analyses are later per- formed that incorporate biases cumulatively. In terms of cargo tank motor vehicle releases, Saccomanno, Stewart, and Shortreed (1993) indicated that the size of the opening as well as the energy absorption characteristics should be recorded in addition to critical exposure values required to sustain damages. Approaches to Analyzing Data Background—How Bulk Package Performance Relates to Hazardous Materials Transportation Risk Traditionally risk (R) is defined as the product of the probability of an event (P) and the consequences of that event (C): R P C= × ( )Eq. 1 In the context of hazardous material transportation, risk refers to the probability and consequence of a release. Hazard- ous material releases can be classified as accident-caused or non-accident-caused, and the circumstances that lead to them are different. This section and this report in general focus on accident-caused release risk. Early attempts to quantify the risk of transporting hazard- ous materials estimated the probability of a release based on the historical number of releases that occur per million miles, and the number of miles traveled. The consequences of these releases were estimated using the number of people residing within the hazardous materials’ evacuation perimeter (Erkut and Verter 1998). Such methods may be sufficient for route selection models, but risk estimates may have large sources of variance depending on the commodity and the container in which it is shipped (Brown, Dunn, and Policastro 2000). Therefore, attempts to better estimate the risk of transport- ing hazardous materials have resulted in the use of conditional probabilities of release. Additionally, the accident itself affects risk in that, at a minimum, the risk associated with transport- ing hazardous materials is equivalent to the risk of transporting other goods (e.g., the risk of injury and/or property damage if the package becomes involved with a collision). Thus the risk of transporting hazardous materials (R) can be described by: R N P P C N P CA R A R A A= × × × + × × ( )Eq. 2 where N = exposure estimate. In this case it represents the number of miles traveled. PA = the probability of an accident. PRA = the probability of a release given that an accident has occurred or the conditional probability of release. Therefore, 1 - PRA corresponds to the events in which the vehicle is involved in an accident but does not result in a release. CR = the consequences of the release. CA = the consequences due to the physical aspects of the accident (e.g., the risk of injury and/or property dam- age if the package becomes involved in a collision).

88 The variables in the risk equation above can be estimated from existing databases (FMCSA’s MCMIS, PHMSA’s HMIRS, NASS GES, or the LTCCS). Exposure estimates can be deter- mined based on the national commodity flow surveys, or, in a route-specific analysis, by the expected number of miles traveled by the bulk container(s). Probabilities of accidents can be determined using FMCSA’s MCMIS and refined using the LTCCS. A basic conditional probability of release can be estimated by comparing the total number of releases (numerator) to the total number of accidents involving haz- ardous materials (denominator). Both variables are recorded in FMCSA’s MCMIS. These estimates can be further refined by using variables available in PHMSA’s HMIRS to consider specification- or component-specific conditional probabil- ity of release. Underreporting in existing databases, especially if bulk packages did not release their contents when involved in an accident, introduce a considerable amount of uncertainty into the estimates of conditional probability of release. To address this, conditional probabilities of release can be fur- ther refined using the accident damage database presented in this report. The consequences of a release can be estimated using information from PHMSA’s HMIRS, NHTSA’s FARS, or using sophisticated hazardous materials dispersion mod- els. In general, the consequences of a release are a function of the amount released, the potential adverse effects of the haz- ardous material, and the population density in the property potentially affected by the material. Finally, the consequences of the accident pertaining to the casualties and property dam- age sustained as a direct result of the accident, excluding any effects of the hazardous material, can be found in FMCSA’s MCMIS for non-release accidents and PHMSA’s HMIRS for release accidents. Two variables in Equation 2 relate directly to bulk pack- age performance: the conditional probability of release and the amount released. The conditional probability of release is a function of the damage incurred during an accident and the bulk package’s ability to withstand that damage. Similarly, the amount released is a function of the size of the opening (related to the package’s ability to withstand damage), the physical properties of the hazardous material, and the time elapsed before the opening can be plugged or the hazardous material can be off-loaded. In general, these two variables are independent of the probability of an accident occurring and the consequences of a particular release quantity. Therefore, independent regression equations can be developed using an accident damage database such as the options presented in Section 3. Indeed, a principal objective of developing such an accident damage database is to provide a robust source of empirical accident information on bulk package performance in a wide range of accident scenarios. As data are accumulated over time, the statistical power of these regression equations will increase, and more sophisticated analyses can be con- ducted. These include developing the conditional probabili- ties of release for various components, in different accident scenarios. Assuming the individual accident scenarios and the individual component-specific conditional probabilities of release are independent, the combined conditional probabil- ity of release can be determined using Equation 3: P PR A jk k mj j n = − − − −[ ]      == ∏∏1 1 1 111 (Eq. 3) where PRA = the probability of a release given that an accident has occurred, or the conditional probability of release. Pjk = the conditional probability of a release from the kth source due to the jth accident scenario. n = the number of accident scenarios. mj = the number of components considered for each acci- dent scenario. Logistic Regression Models Previous experience involving analysis of railroad tank car safety performance using the RSI-AAR TCAD illus- trates the use of statistical techniques to estimate package accident performance (Treichel et al. 2006). Several regression methods were employed to develop the conditional probabili- ties of release as part of the RSI-AAR Railroad Tank Car Safety Research and Test Project, including the following: • Logistic regression in which a binary response variable such as release or no release is selected. Logistic regression makes an exponential transformation that allows prob- ability estimate errors to conform to the required assump- tion of normality. • Ordinal logistic regression in which response variables are binned. For example, binned data regarding the percentage lost would be an appropriate response variable for ordinal logistic regression. • Multiple binary logistic regression in which there are sev- eral response variables representing, for example, release or no release of hazardous materials in several categories. Wen and Simpson, in Multivariate Regression Analysis of Tank Car Lading Loss (Wen and Simpson, 2000), presented the following equations and some other details of the general logistic regression model (see Appendix G): P Y P e e jk i jk i L i L i =( ) = ( ) = + ( ) ( )1 1 4X X X ( )Eq.

89 where Yi = a binary dependent variable associated with the ith record where 0 represents a non-release event and 1 represents a release event. Xi = a vector denoting the values of l independent variables for the ith record: Xi =       1 5 1 2 1 x x x i i i  ( )Eq. L = the logarithmic odds ratio, = Xi′a, and a = the following vector denoting the values of l coefficients of the logarithmic odds ratio, L: a =       β β β β 0 1 2 6  l ( )Eq. The incorporation of a linear equation is beneficial because it “allows consideration of a large number of independent variables and has well-established theoretical basis for fur- ther statistical inference” (Wen and Simpson, 2000), while the incorporation of a quadratic or higher-order equation allows interactions between variables (occurs when there is a covariance between two or more variables in the equation). Since the estimation of the conditional probability of release for bulk packages transported by highway is similar to the estimation of the conditional probability of release for tank cars, similar approaches for developing the response function may be used. Accident scenarios that may initially be considered include accidents in which another vehicle is involved and incidents in which the vehicle and bulk package overturned. Similarly, the components that may be considered include the following: • Valves. • Loading/unloading lines, piping or fittings. • Manway or dome covers. • Tank head. • Tank shell. • Welds or seams. The combination above results in the need to calculate 12 separate regression equations representing the conditional probability of release from the jth source given the kth acci- dent scenario. These regression equations could consider the effects of the following factors: • Material type. • Damage location on bulk package. • Thickness of the bulk package at the damage location. • Bulk package operating pressure. • Bulk package capacity. • Packaged amount. • Damage type. • Speed. • The speed of the other vehicle involved in the accident (if another vehicle was involved). • The type of object striking or struck by the bulk package. Not all of these factors may have a significant effect on a particular component’s loss probability given a particular type of accident. Therefore, regression model building techniques comparing every subset of the full model should be conducted. In Treichel et al. (2006), separate regression models for each of the four major components (head, shell, top fittings, and bottom fittings) were developed by removing factors that had no significant effect on a particular component’s loss prob- ability from that component’s model. Similarly, a regres- sion analysis of highway accident data may show that, for example, the capacity of the bulk package does not have a significant effect on the probability of a release from loading/ unloading lines. Similar to the methodology used in Treichel et al. (2006) to determine the coefficients for each significant factor, acci- dent report selection criteria for the population to be ana- lyzed should be constructed with the intention of eliminating the possibility of undesirable heterogeneities. Without these inclusion criteria, biased conclusions may be drawn, particu- larly if loading characteristics differ (bulk package is empty) or damage is sustained to the bulk package after the initial impact (such as prolonged exposure to fire). Thus, the inclu- sion criteria for accident records are particularly impor- tant when accident data include accidents involving cargo tanks whether or not they carried hazardous materials; were loaded at the time they were damaged; sustained sufficient damage to result in a lading loss; or were exposed to a fire for a prolonged period of time. Population-Wide Accident and Release Rates The amount of time necessary to yield accident perfor- mance measures is important when considering the imple- mentation of such a system. The following analysis estimates the rate at which reports would be generated given current highway bulk package accident rates observed in the pilot

90 study. This was achieved by compiling a data set consisting of PHMSA HMIRS reports, FMCSA MCMIS reports, and news articles over a 7-month period, estimating the total number of accidents per month, and calculating the average rate at which accidents occur. Accident Rate Data Set To estimate the rate at which reports would be generated given current highway bulk package accident rates, a data set consisting of PHMSA HMIRS reports, FMCSA MCMIS reports and news articles published between March and September in 2011 was compiled. The combined data set consisted of 924 accidents (see Figure 15) of which 236 resulted in a release of hazardous materials (see Figure 16). The majority of these accidents were reported by only one source and only five records were captured in all three data sets. News articles and FMCSA records shared a total of 11 records. PHMSA HMIRS reports used in this analysis consisted of in-transit highway accidents involving cargo tank motor vehicles or portable tanks in which a crash occurred. These records correspond to the reportable incidents as defined by 49 CFR 171.15 and 49 CFR 171.16. For highway transporta- tion, these include any incident in which one or more of the following apply: • “As a direct result of a hazardous material: a person is killed, a person receives an injury requiring admittance to a hos- pital, the general public is evacuated for one hour or more, a major transportation artery or facility is closed or shut down for one hour or more” (49 CFR 171.15 7.b.1). • “A situation exists of such a nature that, in the judgment of the person in possession of the hazardous material, it should be reported to the National Response Center (NRC) even though it does not meet other requirements” (49 CFR 171.15 7.b.5). • “There is an unintentional release of a hazardous material or the discharge of any quantity of hazardous waste” (49 CFR 171.16 a.2). • “A specification cargo tank with a capacity of 1,000 gallons or greater containing any hazardous material suffers struc- tural damage to the lading retention system or damage that requires repair to a system intended to protect the lading retention systems, even if there is no release of hazardous material” (49 CFR 171.16 a.3). In reality, accidents in which a bulk package is damaged but the damage does not result in a release are underrepre- sented in this database. Therefore, we can expect that the rate of accidents will be greater than the rate derived solely from considering this data set. Regardless, with these param- eters, 123 accidents were reported to PHMSA HMIRS within the 7-month period. Of these 123 accidents, 98 resulted in a release. FMCSA MCMIS reports consist of accidents involving a cargo tank or intermodal truck that has been placarded for hazardous materials transportation. These parameters yielded 754 accidents between March and September 2011. Some of the discrepancy in the number of accidents from PHMSA’s HMIRS results from the inclusion of accidents in which the bulk package was not damaged (e.g., an accident in which a vehicle was rear-ended by the hazardous materials vehicle); however, the discrepancy cannot be entirely discounted because the MCMIS data set also contains 70 releases that are not included in the HMIRS (see Figure 16). The third data set used in this analysis was derived from news articles found during the same period. These news arti- cles described accidents that resulted in damage to the bulk package. During the 7-month period, 127 accidents were recorded, of which 103 resulted in the release of hazardous materials. Although the number of records in the news article data set was of a similar magnitude to the records in HMIRS, the two data sets only share approximately 30% (37 records were found in both data sets). Similarly, there are only six records in the news article data set that are also in FMCSA MCMIS. Like the PHMSA HMIRS data set, a data set com- posed of news articles tends to underestimate accidents in which the bulk package was damaged but no release occurred. Therefore, an accident rate derived solely from considering the number of accidents reported in news articles is expected to be lower than the actual accident rate. Accident and Release Rates The accidents in the combined data set were grouped by month to estimate the rate at which accidents occur (see Figure 25). If all the FMCSA MCMIS, PHMSA HMIRS, and news report accidents are considered, 132 ± 20 acci- dents can be expected per month (with 95% confidence). Furthermore, from these accidents, approximately 34 ± 7 releases will occur per month (with 95% confidence) (see Figure 26). Therefore, approximately 26% of accidents involving hazardous materials bulk packages result in a release. In contrast, Brown, Dunn, and Policastro (2000) show that the release rates from Harwood and Russell (1990) for highway bulk packages transporting gases and liquids are 8% and 19%, respectively. As mentioned above, the pilot study data are over-weighted in accidents that had a release and under-weighted in non- release accidents. Furthermore, only a subset of data relat- ing to MC 306 and DOT 406 cargo tanks has been used in the sample size analyses that follow. For these reasons, the analyses and evaluations of the pilot study data presented in

91 N um be r o f H ig hw ay H az ar do us M at er ia l A cc id en ts Figure 25. Highway hazardous materials accidents from March to September 2011. N um be r o f H ig hw ay H az ar do us M at er ia l A cc id en ts Figure 26. Hazardous materials releases resulting from highway accidents from March to September 2011.

92 records that corresponded to 77 component-specific damage cases in which a component of the bulk package was damaged. On average, there were approximately 2.3 component-specific damage cases per accident. The data set was modified to con- tain the following variables: • Thickness of the head or shell at the location where the bulk package was damaged. • Total capacity of the bulk package. • Packaged amount. • Speed of the bulk package prior to impact. • Speed of the other vehicle involved in the accident. • Damage location. • Component damaged within that location. • Damage type. • Whether a personal vehicle was involved. • Whether a heavy vehicle was involved. • Whether the bulk package crossed the centerline or median. • Whether it ran off the road. • Whether it rolled over. • Whether it exploded or caught fire. • Whether the units separated in the crash (if the tractor- trailer was also towing a pup trailer). • Whether the bulk package struck the roadway. • Whether it struck the ground. • Whether it struck a concrete barrier. • Whether it struck a guardrail. Since the general logistic regression program used for the analysis did not allow categorical data, the damaged compo- nent, its location, and the type of damage were converted to a series of binomial variables. Probability of a Release Of the 77 component-specific damage cases included in the MC 306/DOT 406 data subset, 74 indicated whether a release occurred as the result of an accident. Of these, 48 damage cases contributed to a release of hazardous materials. Since the records generated during the pilot study correspond to accidents in which a release occurred, the con- ditional probability that damage to a particular component- location combination (damage case) contributed to a release, given that a release occurred, was 65 percent (48 of 74 records). Using the probability that an accident results in a release, the probability that a particular component-specific damage contributed to a release is found by: P damaged component release accident P damag ( ) = ed component release accident release P ac ( ) × cident release accident Eq.( ) ( )7 this report are for illustrative purposes only and should not be considered reliable estimates of the performance of high- way bulk packages. Such analyses will require a much more extensive data set, the development of which is the subject of this report. Minimum Number of Records The requisite sample sizes were estimated in order to deter- mine when sufficient data would be available to conduct sta- tistical analyses with acceptable confidence intervals. These sample sizes must satisfy two conditions: 1. There must be a sufficient number of accident records to minimize Type I errors (where insignificant variables appear to have a significant effect) and Type II errors (where significant variables appear not to have an effect on the probability of a release) when testing hypotheses with the model. While determining acceptable levels of Type I and Type II errors is not necessary for developing a model, sample size estimate approaches used in hypoth- esis testing may provide a rough estimate of the required sample size for developing a regression equation. 2. There must be at least 10 events per variable included in the model (Peduzzi et al. 1996). This is typically checked once the accident data have been collected. For each vari- able considered, there should be accident records per- taining to at least 10 release events and 10 non-release events. Minimum Sample Size to Minimize Statistical Errors To satisfy the first condition, the general multiple logis- tic regression form, given by Equation 4, with a single inde- pendent variable, was considered. The requisite sample size for each variable was determined by using a subset of data corresponding to hazardous materials typically shipped in MC 306 or DOT 406 containers. The probability of a dam- age case release was estimated using this subset of data and the percentage of accidents resulting in a release (26%). The parameters describing the distribution of each vari- able were calculated, and the odds ratio of each variable was determined. Using this information, the number of accident records needed to achieve specified significance and power levels was determined. Sample Size Data Set The pilot study data set was refined to contain records corre- sponding to hazardous materials typically shipped in MC 306 or DOT 406 containers. This subset consisted of 35 accident

93 ability of a damage case release is determined by dividing the mean by the number of records considered in the ini- tial odds ratio estimates (see Table 55). Logistic regression analyses were performed using the original variables to obtain the odds ratios used in the subsequent sample size analysis. Analysis Parameters In statistical analysis, there are generally two types of errors. In the case of estimating release probabilities, a Type I error, denoted by a, occurs when the variable is found to have a significant effect on the probability of release when it actu- ally does not. Type I errors are controlled in experimental analysis by specifying the significance level, or the amount of Type I error allowed in the experiment. In most experiments where Type I errors are the primary concern, the significance level corresponds to a = 0.05. On the other hand, a Type II error, denoted by q, occurs when a variable is not found to have a significant effect when it actually does. Type II errors are controlled by designing experiments that have large val- ues of power (and small values of q). This is usually accom- plished by increasing the number of samples considered in the experiment. Since the minimum sample size required to determine whether a variable has a significant effect on the probability of release is to be identified, the Type I and Type II errors should be balanced. For example, more Type I error may be accepted in the experiment until sufficient sample sizes have been generated to control Type II error. For this rea- son, sample sizes were determined for two significance levels (0.05 and 0.1) and six levels of power (0.70, 0.75, 0.80, 0.85, 0.90, and 0.95). Sample Size Estimates The required number of accident records was estimated using the POWER procedure in SAS 9.2. Once the number of cases required to obtain a significant result with adequate probability had been determined, the number of accidents Assuming an accident rate of 26 percent (34 releases per 134 accidents per month, as developed in Section 3), the probability of release related to component-specific dam- age is 16.7 percent. This means that, in an accident, the bulk package components are able to withstand the incurred dam- age 83.3 percent of the time. Odds Ratios In order to estimate the required sample size for a logistic regression analysis, an initial estimate of the odds ratio is required (see Equation 4). As a measure of the ratio of prob- ability of an event per unit change of a variable, the odds ratio depends upon an estimate of the mean and standard deviation of the variables used in the logistic regression equation. The following odds ratios estimates are based on data collected during the pilot study. Scalar Variables. A normal distribution was assumed in order to estimate the mean and standard deviation of the scalar variables. This assumption yields favorable estimates of the minimum number of records required. It is likely that, given non-normal distributions, larger sample sizes will be required. Therefore, once sufficient data have been collected, the assumption of normality should be re-examined. The following variables were standardized by dividing each value by its standard deviation (see Table 54): • Thickness of the head or shell at the location where the bulk package was damaged. • Total capacity of the bulk package. • Packaged amount. • Speed of the bulk package prior to impact. • Speed of the other vehicle involved in the accident. To get a reasonable value for the odds ratio, logistic regres- sion analyses were performed using the standardized variables. Binomial Variables. The distribution for binomial vari- ables is described by the probability of a damage case release and the number of damage cases considered. The prob- Variable Mean Deviation Adjusted Mean Standard Deviation Odds Ratio Thickness 0.19 0.01 13.04 1 0.653 Capacity 4,160 4,525 0.92 1 1.389 Packaged Amount 6,506 2,621 2.48 1 1.166 Speed 44 20 2.20 1 0.682 Other Vehicle Speed 14 22 0.60 1 1.509 Table 54. Distribution parameters of scalar variables and corresponding odds ratios.

94 Variable n Mean p Odds Ratio Damaged Valve 74 0.068 0.09% 2.273 Damaged Lines, Pipes, and/or Fittings 74 0.108 0.15% 4.266 Damaged Manway 74 0.108 0.15% >999 Damaged Head 74 0.054 0.07% 0.163 Damaged Shell 74 0.622 0.84% 0.259 Damaged Weld and/or Seam 74 0.041 0.05% >999 Abraded 74 0.027 0.04% 0 Bent 74 0.095 0.13% 0.183 Burst or Ruptured 74 0.041 0.05% 1.087 Crushed 74 0.392 0.53% 0.176 Cracked 74 0.068 0.09% >999 Gouged or Cut 74 0.068 0.09% 2.273 Leaked 74 0.081 0.11% >999 Punctured 74 0.054 0.07% >999 Ripped or Torn 74 0.108 0.15% >999 Torn Off or Damaged 74 0.068 0.09% >999 Front Head Damage on Centerline 75 0.027 0.04% 1.000 Front Head Damage Above Centerline 75 0.053 0.07% 0.163 Rear Head Damage Below Centerline 75 0.040 0.05% 0.255 Rear Head Damage Above Centerline 75 0.013 0.02% >999 Bottom Front Driver-Side Damage 75 0.053 0.07% 1.666 Bottom Middle Driver-Side Damage 75 0.040 0.05% 1.087 Bottom Rear Driver-Side Damage 75 0.067 0.09% >999 Top Front Driver-Side Damage 75 0.053 0.07% 1.087 Top Middle Driver-Side Damage 75 0.027 0.04% 0.532 Top Rear Driver-Side Damage 75 0.080 0.11% 0.511 Bottom Front Passenger-Side Damage 75 0.040 0.05% 0.255 Bottom Middle Passenger-Side Damage 75 0.053 0.07% 0.163 Top Front Passenger-Side Damage 75 0.187 0.25% 0.969 Top Middle Driver-Side Damage 75 0.053 0.07% 1.666 Top Rear Passenger-Side Damage 75 0.080 0.11% 1.091 Damage to Piping and/or Undercarriage Below the Tank 75 0.133 0.18% >999 Passenger Vehicle Involved 77 0.273 0.35% 1.729 Heavy Vehicle Involved 77 0.065 0.08% 2.273 Crossed Centerline 77 0.221 0.29% 0.384 Ran-Off-Road 77 0.675 0.88% 0.500 Rolled Over 77 0.818 1.06% 0.439 Units Separated 77 0.195 0.25% 0.384 Struck Roadway 77 0.468 0.61% 0.778 Struck Ground 77 0.688 0.89% 0.397 Struck Concrete Barrier 77 0.065 0.08% 0.334 Struck Guardrail 77 0.130 0.17% 0.891 Struck Tree 77 0.013 0.02% 1.000 Involved Explosion or Fire 77 0.182 0.24% 8.333 Table 55. Distribution parameters of binomial variables and corresponding odds ratios.

95 Significance Level Nominal Power Number of Cases Number of Accidents 0.05 0.70 274 120 0.05 0.75 308 134 0.05 0.80 348 152 0.05 0.85 398 174 0.05 0.90 466 203 0.05 0.95 576 251 0.10 0.70 209 91 0.10 0.75 239 104 0.10 0.80 275 120 0.10 0.85 319 139 0.10 0.90 380 166 0.10 0.95 480 209 Table 56. Sample size required for the variable “Thickness.” Figure 27. Sample size requirements given ` = 0.05. their effect was calculated in a similar manner. These sample sizes indicated that there are three tiers of variables. Tier I vari- ables required sample sizes of less than 800 accident records (see Figure 27 and Figure 28). Of these, damage to the shell or crushing damage required the smallest number of records. The sample sizes required for Tier II variables at a sig- nificance level of 0.1 and 90% power ranged between 858 and 1,286 accident records (see Table 57). In comparison was calculated. Table 56 shows the sample sizes needed to obtain the corresponding power for testing, at the specified significance level, the effect of the thickness of the head or shell at the location where the bulk package was damaged. For example, using a significance level of 0.1, in order to achieve 90% power, 166 accident records are needed. For those variables containing sufficient information and variance, the number of accident records needed to determine

96 Figure 28. Sample size requirements given ` = 0.10. Significance Level Nominal Power Packaged Amount Driver-Side Bottom Front Driver-Side Top Rear Passenger- Side Top Middle Struck Roadway 0.05 0.70 927 661 618 661 737 0.05 0.75 1,042 744 695 744 829 0.05 0.80 1,178 841 786 841 937 0.05 0.85 1,348 962 899 962 1,072 0.05 0.90 1,577 1,125 1,052 1,125 1,255 0.05 0.95 1,950 1,392 1,301 1,392 1,552 0.10 0.70 707 504 472 504 562 0.10 0.75 808 577 539 577 643 0.10 0.80 928 663 620 663 739 0.10 0.85 1,079 770 720 770 859 0.10 0.90 1,286 917 858 917 1,023 0.10 0.95 1,624 1,159 1,084 1,159 1,293 Table 57. Required number of accident records for variables.

97 standard (non-varying) thickness jacket adjusts the curve towards lower probability of release values by the same amount. Note that in Figure 29 the slopes of the lines differ- entiating the presence of a jacket are equal. The minimum sample sizes for models in which variables are not correlated is approximately equal to the largest sample size requirement of the variables included in the model. For example, using the MC 306/DOT 406 pilot study data subset, 277 accident records are needed for modeling the probability of a release as a function of thickness, capacity, and speed at a significance level of 0.1 with 90% power. However, if the variables are correlated, as is typically the case for empirical data, the probability of a damage case release will be different from the sum of the probabilities of release from each variable. Referring to the RSI-AAR TCAD example, if the presence of a jacket and the thickness of the shell material were, in fact, correlated, the slopes of the lines would be different (see Figure 30). Here, given a curve that represents the relationship between the probability of release and tank shell thickness, the addition of a standard (non- varying) thickness jacket adjusts the curve towards lower probability of release values by increasingly greater amounts. In Figure 30, the hypothetical decrease in probability of a release due to the interaction effects of a jacket and shell thickness is represented by the grey area. The required minimum sample size for models in which the variables are correlated will be larger than that required for models with uncorrelated variables. This is because addi- tional records are needed to evaluate the interaction effects of the variables included in the model. The pilot study data set suggests correlations exist between the following variables: • Damage to the shell, crushing type damage, rollover acci- dents, running off the road, “struck roadway,” and “struck ground.” • Which component was damaged and the damage location (if not nested). to Tier I variables, Tier II variables required twice as many accident records. Those variables requiring sample sizes greater than 5,000 records are classified as Tier III variables: • Bulk package design pressure. • Burst or ruptured damage. • Damage to the rear head below the centerline. • Damage to the bottom middle driver-side. • Damage to the top front driver-side. • Damage to the top middle driver-side. • Damage to the bottom front passenger-side. • Damage to the top front passenger-side. • Damage to the top rear passenger-side. • Striking the guardrail. These large sample size requirements may be attributed to a low correlation between these variables and the probabil- ity of a damage case release or limitations in the estimating procedure. Therefore, the required sample sizes should be re-evaluated after the collection of accident data has begun. Multivariate Logistic Regression Sample Sizes In multivariate logistic regression, the minimum sample size must also meet the conditions of a minimum number of records (Condition 1) and a minimum number of events per variable (Condition 2). To determine the minimum number of records, the interaction of the variables included in the logistic regression model must be considered. An example can be drawn from tank car performance analyses using the RSI-AAR TCAD (Treichel et al. 2005). In these analyses, the characteristics in the model were assumed to be indepen- dent of each other; therefore, for example, given a curve that represents the relationship between the probability of release and tank shell thickness (see Figure 29), the addition of a Figure 29. No correlation effects. Pr ob ab ili ty o f a R el ea se Tank Shell Thickness Jacket No Jacket Figure 30. Effect of correlation between variables.

98 ing the required sample sizes to the rate of data acquisition (see Table 59). If the implemented data set were able to obtain records for all accidents that occur (132 per month), the data set would yield significant results for some variables within 1 month of implementation. Within 1 year, there would be sufficient accident records to analyze each of the Tier I and Tier II variables. Since reporting rates may be significantly lower than the expected accident rate, the number of months needed to generate sufficient sample sizes was determined for a range of reporting rates (recall that Battelle Memorial Institute [2009] estimated an HMIRS reporting rate of 26.9% and the underreporting analysis conducted in Section 3 estimated a reporting rate between 13% and 44%). This assumes that the ratio of non-release accidents to accidents in which a release occurs is maintained. As shown in Table 59, if com- plete records for only 20% of accidents are obtained, the data set would need 19 months of data accumulation before each of the Tier I variables could be tested. By the fourth year of data collection, the significance of each of the Tier II variables could be tested. Statistical Summary of the Pilot Study The pilot study resulted in a data set consisting of 50 records. A summary of data in Table 60 illustrates the range of responses obtained in the pilot test, grouped by hazard- ous material type. This data set has varying degrees of com- pleteness, particularly regarding bulk package design, the extent of the damage, and the dimensions of the breach. As • Damage to fittings and accidents involving passenger vehi- cles (i.e., damage to Location 3: front head above centerline and damage to the head itself). • Speed of the bulk package and speed of the other vehi- cle involved (this most likely resulted from the method employed to estimate the other vehicle’s speed in the context of the pilot study, but correlation is nevertheless expected). • Capacity and packaged amount. Since the correlation between variables cannot be esti- mated due to the limited size of the pilot study data set, the increase in sample size cannot currently be determined. Minimum Number of Events Peduzzi et al. (1996) define the number of outcome events as “the smaller number of binary outcomes (e.g., alive versus dead)” and provide the example that “a particular study may have many subjects, but too few deaths for a valid analysis.” In applying the results of Peduzzi et al. (1996) to the question of bulk package performance, each record represents a par- ticular set of independent variables and an outcome (release or no release). Records in the data set are therefore equivalent to the subjects to which Peduzzi et al. (1996) refer. Similarly, since the percentage of accidents resulting in a release is less than the percentage of non-release accidents, release events are equivalent to Peduzzi et al.’s (1996) events. It is also possible that, since the probability of an accident resulting in a release is approximately 26%, the accident data set may have many records but too few release events for a valid analysis. According to Peduzzi et al. (1996), at least 10 events per variable included in the model are desirable to maintain the validity of the model: N l p = ×10 8( )Eq. where l = the number of independent variables in the regression model. p = the lesser of the percentage of release events or non- release events. Given that 26% of accidents result in a release, if a model consisted of 10 variables, the minimum number of accident records required is 385 (see Table 58). Expected Implementation Time Determining the amount of time required for the bulk package accident data collection system to yield statistically significant accident performance measures involved compar- Number of Variables Included in the Model Number of Records 1 39 2 77 3 116 4 154 5 193 6 231 7 270 8 308 9 347 10 385 11 424 12 462 13 500 14 539 15 577 Table 58. Minimum number of records needed to satisfy Condition 2.

99 theless, the following discussion details variable response completeness, summarizes the collected data, and provides a basic interpretation. In the pilot study, the design characteristics, commodity information, and accident information are stored so that one record pertains to one accident. Since one accident may result in damage to more than one part of the bulk package and more than one component, damage information is stored so that one record pertains to one location-component com- bination. For example, a rollover accident that resulted in damage to the shell along the entire length of the tank would result in a minimum of three location-component combina- tions: one for damage to the shell in each location (i.e., given a result, while there are 50 records presented in Table 60, the completeness of the data set is not portrayed. For many variables, the actual number of records for which informa- tion was obtained sum to less than 50. Due to the limited data size and varying degrees of completeness, the statistical summary of this data is not representative of the entire pop- ulation of accidents involving bulk packages. As mentioned above, the pilot study data are over-weighted in accidents that had a release and under-weighted in non-release acci- dents. For this reason, the analyses and evaluations of the pilot study data presented in this report are for illustrative purposes only and should not be considered reliable esti- mates of the performance of highway bulk packages. Never- Variables Required Sample Size Reporting Rate (%) 100 90 80 70 60 50 40 30 20 10 Number of Months Tier I Min 45 1 1 1 1 1 1 1 1 2 3 Max 625 4 5 5 6 7 8 10 13 19 37 Tier II Min 1,084 7 8 9 10 11 13 17 22 33 65 Max 1,624 10 11 12 14 16 20 24 32 48 96 Tier III Min 11,030 83 92 103 118 138 165 206 275 412 824 Table 59. Number of months required to obtain sufficient sample sizes for testing significance at ` = 0.10 and power = 0.90. Table 60. Commodities and containers in the pilot study.

100 als were involved, the bulk package was assumed to have at least two compartments. As a result, 40 accident records (80%) indicate a bulk package with one compartment and 4 records indicate a bulk package with two compartments. Five records correspond to bulk packages with four com- partments and one record corresponds to a bulk package with five compartments. Materials of Construction The materials of construction were determined for 34 records (68%). Aluminum was the material of construc- tion for 28 (56%) of the bulk packages. Two (4%) were con- structed of stainless steel, three (6%) were constructed of carbon steel, and one (2%) was constructed of composite materials. Capacity Since PHMSA only records the total capacity of a bulk package (by type of hazardous material), capacity for indi- vidual compartments was incompletely recorded in the pilot study. Should this data collection system be implemented, it is anticipated that this information will be more readily available. To account for bulk package capacity in the analy- sis of pilot study data, the total capacity of the bulk package was recorded when available. Capacity information is avail- able for 32 of the 50 records (64%). The total bulk package capacities range between 2,500 gallons and 12,500 gallons (see Figure 31). a rollover onto the passenger side of the bulk package, Loca- tion 18—top front passenger-side, Location 19—top middle passenger-side, and Location 20—top rear passenger-side are likely to be the three locations incurring damage). For the purposes of the following discussion, the terms “case” or “damage case” refer to one of these location-component records. A total of 115 damage cases were identified from 46 accident records (4 of the 50 accidents did not have suf- ficient damage information). Container Types The pilot study considered a total of 50 records in which 49 (98% of the pilot study accident records) correspond to cargo tanks and 1 (2%) corresponds to an ISO tank. While specification information was unavailable for 13 of 16 truck- mounted cargo tanks and 8 of 33 trailer-mounted cargo tanks, container specifications are matched to all but two commodi- ties in this pilot study (see Table 60). Number of Compartments In general, it was difficult to ascertain the number of com- partments a bulk package contained from the information provided by PHMSA and/or photos of the bulk package. Where this information was not available, the bulk pack- age was assumed to have at least one compartment with a capacity corresponding to “Cont1 Package Capacity” listed in PHMSA’s HMIRS. Similarly, for those PHMSA-reported accidents in which two separate kinds of hazardous materi- Figure 31. Cumulative proportion of pilot study bulk package capacities.

101 • Petroleum crude oil (UN1267). • Hydrochloric acid (UN1789). • Acrylic acid, stabilized (UN2218). • Propane (UN1978). • Liquefied petroleum gas (LPG) (UN1075). • Flammable liquids, n.o.s. (UN1993). • Isopropenylbenzene (UN2303). • Hydrogen, refrigerated liquid (UN1966). Hazardous Materials Packaged Amount Similar to the capacity for individual compartments, the packaged amount for individual compartments was poorly recorded in the records found for the pilot study. Should this data collection system be implemented, it is anticipated that this information will be readily available. To account for the packaged amount in the pilot study analysis, the total pack- aged amount was recorded when available. Packaged amount information was recorded for 29 (58%) of the 50 records. These range from 500 gallons to 9,501 gallons (see Figure 32). Vehicle Speed The speed of the bulk package vehicle prior to incurring damage was estimated for 48 (96%) of the records. The pilot study grouped speeds in 5 mph bins. These speeds ranged from 0 to 65 miles per hour (see Figure 33). Damage Location The location of damage to the bulk package was determined from photographs accompanying newspaper articles and damage descriptions included in PHMSA’s HMIRS “Descrip- tion of Events.” Since photographs of rollovers typically depict the bulk package’s final resting position, damage to the side in contact with the ground could only be approximated. Damage was estimated for 47 records (94%). Based on this method of approximation, the locations most likely to be damaged are the top front passenger-side and the piping and/or under- carriage below the tank (see Table 61). These damage loca- tions most likely correspond to different types of accidents: damage to the top front passenger-side results from rollover accidents while damage to the piping and/or undercarriage below the tank results from accidents involving other vehi- cles. Note that damage corresponding to the ISO container was converted to the location-naming scheme for cargo tanks for the purposes of this summary. The following observations regarding damage location can be derived from the pilot study data (see Table 62): • The passenger side of the bulk package is more likely to be damaged and result in a release if involved in an accident. Head/Shell Thicknesses Head and shell thickness were recorded for 22 records (44% of the pilot study accident records). These thicknesses were obtained for 11 DOT 406/MC 306 containers (22%), 2 DOT 407/MC 307 containers (4%), 1 DOT 412 container (2%), and 3 MC 331 containers (6%). Four records (8%) did not list a corresponding specification. Additionally, one record (2%) listed only shell thickness information. Examin- ing the data for the largest homogeneous group, the DOT 406/MC 306 containers, the head thicknesses range from 0.175 inches to 0.25 inches while shell thicknesses range from 0.15 to 0.204 inches. Working Pressure Tank pressure ratings were recorded for 27 accidents (54%). The pressure ratings range from 1 psig to 5 psig for the DOT 406/MC 306. The majority (17 records or 34% of the pilot study accident records) indicate a pressure rating of 3 psig. Pressure ratings of up to 29 psig were recorded for the DOT 407/MC 307 containers. The DOT 412 container indicated a pressure rating of 35 psig, and the MC 331 containers indi- cated pressure ratings from 200 psig to 300 psig. Hazard Class The class of the hazardous material was determined for all 50 records included in the pilot study. Six of the records (12%) indicate Hazard Class 2, 41 (82%) indicate Hazard Class 3, 1 (2%) indicates Hazard Class 5, and 2 (4%) indicate Hazard Class 8. Packing Group The packing group of the hazardous material was reported for 36 records (72%). Of these records, 28 (56%) correspond to Packing Group II while 8 (16%) correspond to Packing Group III. Hazardous Materials Identification Number It was possible to determine the hazardous materials identi- fication number for 46 (92%) of the 50 records (see Table 60). Gasoline/Gasohol (UN1203) was the commodity listed for most of the records in the pilot study data set (24 of 50 records). Other commodities included the following: • Ammonium nitrate, liquid (UN2426). • Diesel fuel/heating oil (UN1202). • Diesel fuel/fuel oil/cleaning compounds (NA1993). • Alcohols, n.o.s. (UN1987).

102 Damaged Components Similar to the damage locations, the bulk package com- ponents damaged in an accident were identified using the photographs accompanying newspaper articles and damage descriptions included in PHMSA’s HMIRS “Description of Events” and “What Failed Description.” Detailed information regarding the identification of the type of valve damaged in the accident is limited; therefore, for the purposes of this analysis, the following valve components were grouped together: • 106—Bottom Outlet Valve. • 107—Check Valve. • 115—Discharge Valve or Coupling. • 116—Excess Flow Valve. However, the driver’s side of the bulk package may be more prone to release if it is damaged during an accident. • The top of the bulk package is equally likely to incur dam- age and result in a release as the bottom of the bulk package; however, if piping and/or the bulk package’s undercarriage are excluded, the top of the bulk package is twice as likely to be damaged in an accident. • Although the top front passenger-side is most likely to be damaged, on average the front of the bulk package is equally likely to be damaged and involved in a release as is the rear. This is likely due to the additional protection afforded to the bottom of the bulk package by the tractor and the trailer wheel set. • Both the front and rear of the bulk package are more likely to sustain damage that results in a release than the middle. Figure 32. Cumulative proportion of pilot study bulk packaged amounts. 0.00 0.25 0.50 0.75 1.00 0 10 20 30 40 50 60 70 C um ul at iv e P ro po rt io n of A cc id en ts Estimated Speed (mph) Figure 33. Estimated speed of bulk package prior to impact (mph).

103 Because damage to various components was identified using PHMSA’s “What Failed Description,” the pilot study estimates of component performance are expected to indicate a higher failure rate given that the component has sustained damage. A possible exception is the perfor- mance of the tank shell because the pilot study generally assumed shell damage on the ground-side of rolled bulk packages. Damaged components were identified for 46 accident records (92%). The component most likely to be damaged • 127—Inlet (Loading) Valve. • 134—Liquid Valve. • 144—Pressure Relief Valve or Device. • 154—Valve Body. • 156—Valve Spring. • 157—Valve Stem. • 158—Vapor Valve. Similarly, “Loading or Unloading Lines (135)” was grouped with “Piping or Fittings (141).” Total Damaged Proportion of Releases Per Location Releases Number Damaged Total Number of Releases 1 - Front Head Damage Below Centerline 1 1 1.00 0.02 2 - Front Head Damage on Centerline 3 1 0.33 0.02 3 - Front Head Damage Above Centerline 6 2 0.33 0.05 4 - Front Head Destroyed 0 0 0.00 0.00 5 - Rear Head Damage Below Centerline 3 3 1.00 0.07 6 - Rear Head Damage on Centerline 0 0 0.00 0.00 7 - Rear Head Damage Above Centerline 4 3 0.75 0.07 8 - Rear Head Destroyed 0 0 0.00 0.00 9 - Bottom Front Driver-Side Damage 3 3 1.00 0.07 10 - Bottom Middle Driver-Side Damage 3 2 0.67 0.05 11 - Bottom Rear Driver-Side Damage 5 4 0.80 0.10 12 - Top Front Driver-Side Damage 5 3 0.60 0.07 13 - Top Middle Driver-Side Damage 3 3 1.00 0.07 14 - Top Rear Driver-Side Damage 7 4 0.57 0.10 15 - Bottom Front Passenger-Side Damage 4 1 0.25 0.02 16 - Bottom Middle Passenger-Side Damage 3 1 0.33 0.02 17 - Bottom Rear Passenger-Side Damage 4 1 0.25 0.02 18 - Top Front Passenger-Side Damage 17 10 0.59 0.24 19 - Top Middle Passenger-Side Damage 7 4 0.57 0.10 20 - Top Rear Passenger-Side Damage 8 5 0.63 0.12 21 - Damage to Piping and/or Undercarriage Below the Tank 12 10 0.83 0.24 Note: These locations are identified in Figures 21 through 23. Table 61. Number of accidents resulting in damage and releases by location. Proportion of Releases Per Location Total Damaged Releases Number Damaged Total Number of Releases Driver-Side 14 11 0.79 0.26 Passenger-Side 26 17 0.65 0.40 Top 30 22 0.73 0.52 Bottom (including piping) 26 21 0.81 0.50 Bottom (excluding piping) 16 11 0.69 0.26 Front 27 19 0.70 0.45 Middle 15 10 0.67 0.24 Rear 26 19 0.73 0.45 Table 62. Comparison of damage and release locations.

104 Damage Type Damage type was identified for all 115 damage cases (see Table 65). The pilot study data suggest that there are differences in the probabilities of release depending on the type of damage received. For example, the most prevalent type of damage result- ing in a release is the ripping or tearing of the tank head, shell, or appurtenances. A total of 15 releases can be attributed to rip- ping or tearing although the frequency of this type of damage is low (only 16 cases). This differs from the most prevalent type of damage, crushing damage; there are 39 crushing damage cases of which 12 resulted in the release of hazardous materials. Damage Dimensions Damage dimensions were difficult to ascertain from many of the photographs and damage descriptions. Therefore, only 11 cases listed damage dimensions. However, with the is the tank shell (see Table 63). This is not surprising as it is the largest component of the bulk package. Nevertheless, the pilot study also indicated that the tank shell was the least likely to result in a release if damaged. The pilot study demonstrated the collection of damage information to determine whether component performance varied by damage location. Due to the number of damage locations (see Figures 21, 22, and 23), determining whether component performance varied by location requires a much greater number of records; therefore, only the tank shell is explored in further detail (see Table 64). This analysis indi- cates that tank shell damage probably does vary by location. Additionally, by considering damage cases, analysis of the pilot study data indicates that, in general, additional protection for the top front passenger-side (19% of all releases) and bottom rear driver’s side of the tank (10% of all releases) might be a good idea. Proportion of Releases Per Component Type Total Damaged Releases Number Damaged Total Number of Releases Valves 7 7 1.00 0.17 Loading/Unloading Lines, Piping, or Fittings 12 9 0.75 0.21 Manway/Dome Cover 8 8 1.00 0.19 Tank Head 11 7 0.64 0.17 Tank Shell 28 17 0.61 0.40 Valve Seat 1 1 1.00 0.02 Weld or Seam 5 5 1.00 0.12 Table 63. Number of accidents resulting in damage and releases by component type. Proportion of Releases Per Location Total Damaged Releases Number Damaged Total Number of Releases 9 - Bottom Front Driver-Side Damage 3 3 1.00 0.07 10 - Bottom Middle Driver-Side Damage 3 2 0.67 0.05 11 - Bottom Rear Driver-Side Damage 4 4 1.00 0.10 12 - Top Front Driver-Side Damage 4 2 0.50 0.05 13 - Top Middle Driver-Side Damage 3 1 0.33 0.02 14 - Top Rear Driver-Side Damage 7 2 0.29 0.05 15 - Bottom Front Passenger-Side Damage 4 0 0.00 0.00 16 - Bottom Middle Passenger-Side Damage 3 1 0.33 0.02 17 - Bottom Rear Passenger-Side Damage 4 0 0.00 0.00 18 - Top Front Passenger-Side Damage 16 8 0.50 0.19 19 - Top Middle Passenger-Side Damage 5 2 0.40 0.05 20 - Top Rear Passenger-Side Damage 6 3 0.50 0.07 Table 64. Number of accidents resulting in tank shell damage and releases by location.

105 a release of lading. This translates to a 58% probability of release per instance of damage. Amount Released The amount released was recorded for 40 (80%) of the 50 accidents and 62 (54%) of the 115 damage cases. This quantity reflects an estimate of the difference between the amount packaged and the amount recovered. Therefore, if the accident involved the combustion of hazardous material following its initial release, there was no distinction between hazardous materials spilled versus hazardous materials con- sumed in the fire/explosion. Additionally, since the amount released was obtained from PHMSA HMIRS, if leaks occurred from two different locations on the bulk package or as a result of the failure of two different components in the same loca- tion, the total amount released was assigned to both cases. The amount released ranged between a residual amount and 9,500 LGA (see Figure 34) with a mean of 2,470 gallons. Breach Dimensions The dimensions of the breach were very difficult to ascer- tain from the photos and were therefore recorded for only four cases. As with damage dimension information, with the full implementation of a data collection program, it is antici- full implementation of such a program, it is anticipated that sufficient data would be collected to evaluate the extent and severity of the damage. Release Indicator The pilot study data included an indication as to whether a release occurred due to damage to a particular component in the specified cargo tank location. This variable is used as the dependent variable when evaluating and modeling con- ditional probability of release. In the pilot study, this vari- able was coded “0” if no release occurred and “1” if a release occurred. Of the 46 tanks with damaged components in the pilot study, 42 (91%) tanks were damaged to the extent that a release occurred. Recall that the data are heavily weighted with accident records for which a release occurred and are under-weighted in records for non-release accidents. Each of the 46 cargo tanks with damaged components had at least one, and usually multiple, location-component combina- tions that sustained damage during the accident. In total, the data set included 115 location-component damage records for the 46 cargo tanks. While there were often multiple location-component combinations on a single tank that were damaged in a single accident, not all of the location- component combinations contributed to a release. Of the 115 location-component records, 67 (58%) resulted in Valves Lines, Piping, or Fittings Manways/ Dome Covers Tank Head Tank Shell Valve Seat Weld or Seam Damage Type D R D R D R D R D R D R D R Abraded 5 0 Bent 1 0 2 0 1 1 2 0 2 1 Burst or Ruptured 1 1 1 1 1 1 2 1 Cracked 2 2 5 4 Crushed 1 1 6 2 32 9 Failed to Operate 1 1 Gouged or Cut 1 1 1 1 4 Leaked 2 2 1 1 5 5 Punctured 1 1 9 6 Ripped or Torn 1 1 7 6 3 3 2 2 1 1 1 1 1 1 Structural 1 1 Torn Off or Damaged 1 1 6 4 Vented Note: D represents the number of cases in which the component was damaged. R represents the number of cases in which damage to the component resulted in a release. Table 65. Number of cases corresponding to each damage type by component category.

106 vehicle struck the bulk package and two records (4%) in which the bulk package struck another heavy vehicle. Speed of Other Vehicle Involved in Collision The speed of the other vehicle involved in a bulk package accident was recorded in 5-mph bins. Since vehicle speed was not provided in PHMSA’s HMIRS, it was estimated based upon the accident description and the speed of the bulk package vehicle. These speeds ranged from 0 to 65 mph (see Figure 35). Jackknife Indicator A jackknife occurring as part of the accident was recorded as a “1”; otherwise, a “0” was recorded. Of the 50 records in the pilot study, only one jackknife accident was recorded. Cross Median/Centerline Indicator Of the 50 records in the pilot study data set, 9 (18%) involve the bulk package traveling across the median or centerline of the roadway. These types of accidents correspond to an aver- age release size of approximately 3,360 gallons of hazardous materials, while accidents in which the bulk package did not cross the median or centerline of the roadway resulted in an average release amount of approximately 1,550 gallons. Ran-Off-Road Indicator Of the 50 records in the pilot study data set, 32 (64%) involve the bulk package vehicle being driven out of the lane(s) of travel. Of these 32 records, 7 indicate collision with another vehicle. Of the 18 records where the bulk pated that sufficient data would be collected. The breach dimensions along with the amount released can be used to estimate the rate of release and/or the amount of time until the release was mitigated. For example, large releases com- bined with small breach dimensions may indicate longer response times. Roadway Collision Indicator An incident involving the collision of the bulk package vehicle with another vehicle was coded as “1”; otherwise, it was coded as “0.” The data set generated by the pilot study contains 21 records (42% of the pilot study accident records) in which the bulk package was involved in a collision with another vehicle. Passenger Vehicle Collision Indicator There are two ways in which a passenger vehicle may be involved in a collision with a bulk package. The first, coded as “-1” in the pilot study data set, corresponds to a passenger vehicle striking the bulk package. There were 15 instances of a passenger vehicle colliding with the bulk package (30% of the pilot study accident records). The second, coded as “1,” corresponds to the bulk package vehicle striking a passenger vehicle. There were two instances of this type of collision (4% of the pilot study accident records). Heavy Vehicle Collision Indicator Collisions involving a second heavy vehicle were coded in a similar manner to the passenger vehicle collisions. The pilot study data set includes two records (4%) in which a heavy 0 0.25 0.5 0.75 1 0 2000 4000 6000 8000 10000 C um ul at iv e P ro po rt io n Amount Released (Gallons) Figure 34. Cumulative proportion of pilot study release amounts.

107 package vehicle was not driven out of the lane(s) of travel, 14 indicate collision with another vehicle. An average of 2,350 gallons was released for accidents in which the bulk package was driven off the road compared to approxi- mately 1,140 gallons when the bulk package vehicle was kept on the roadway. Rollover Indicator Thirty-five accidents (70% of the pilot study accident records) involved the bulk package rolling over. Nine (18%) correspond to a roadway collision, eight (16%) correspond to accidents in which the bulk package was driven across the median or centerline of the roadway, and thirty (60%) cor- respond to “ran-off-road” accidents. Of the records that did not indicate a rollover, 12 (24%) correspond to a roadway collision, 1 (2%) involved crossing the median or centerline of the roadway and 2 (4%) involved running off the road. Rollovers resulted in an average release of 2,378 gallons compared to an average of 806 gallons when the package remained upright. Explosion or Fire Indicator A fire or explosion alters how much hazardous material is recovered. Material that is consumed in a fire or explosion is included in estimates of how much material was released in an accident. The accident data in the pilot study included a variable to indicate whether that record corresponded to a fire or explosion. Overall, 15 of the 50 records (30%) indicate that an explosion or fire accompanied the accident. These accidents resulted in an average release of 3,800 gallons com- pared to the mean release of 1,131 gallons for those accidents in which no fire or explosion occurred. Listing of Objects Struck by the Bulk Package In addition to indicating the type of accident a bulk pack- age vehicle was involved in (collision, jackknife, cross median/ centerline, ran-off-road, rollover, and fire/explosion), the pilot study indicated whether or not objects on or near the roadway were struck by the bulk package vehicle. The involvement of these objects was recorded with a “1”; otherwise, a “0” was recorded. Overall, the greatest number of releases involved the bulk package striking both the roadway and the ground (see Table 66). This is largely attributed to rollover accidents. Conclusion This section has summarized and interpreted the data collected in the pilot study. Additionally, minimum sample sizes required to test the effect of the variables on the prob- ability of a release were generated using those pilot study records pertaining to hazardous materials typically trans- ported in MC 306 and/or DOT 406 containers. These mini- mum sample sizes were compared to the expected number of accidents and releases involving hazardous materials to provide an estimate of when the bulk package data col- lection system, once implemented, could be expected to yield results. With a 20% reporting rate, the system can be expected to contain a sufficient number of records to test the significance of most variables within 4 years of its implementation. 0 0.25 0.5 0.75 1 0 10 20 30 40 50 60 70 C um ul at iv e P ro po rt io n of A cc id en ts In vo lv in g A no th er Ve hi cl e Estimated Speed (mph) Figure 35. Estimated speed of other vehicle involved in the accident.

108 Roadway Ground Concrete Barrier Guard Rail Lighting/ Power Line Pole Bride Column Number Average Release Volume (LGA)* 2 32 9 1,444 12 2,730 2 1,240 3 700 5 1,510 1 452 2 9,200 1 736 Not identified Not identified Not identified Not identified Not identified Not identified 13 1,637 2,303 1,240 2,674 736 9,200 Note: * denotes an average of the available release volumes. 2,315 Average Release Volume (LGA)* Table 66. Objects struck by the bulk package and corresponding release sizes.