Identifying and Quantifying Rates of State Motor Fuel Tax Evasion (2008)

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.

51 3.1 Introduction This chapter examines the motor fuel excise tax evasion studies completed in the past 20 years in order to assess the strategies, methods, and tools they employed. Historically, these studies have attempted to determine the percentage of total tax liability captured by current state or federal tax col- lections, and document program characteristics that exacerbate or deter evasion. The focus of this chapter is to examine the approaches these studies used to estimate motor fuel excise tax noncompliance and document their relative strengths and weaknesses for estimating various types of evasion. Some meth- ods seek to estimate total evasion (e.g., econometric analysis), while others seek to estimate specific types of evasion (e.g., bor- der interdiction or use of dyed fuel). Numerous approaches have been employed to study eva- sion at both the state and federal level. These approaches are outlined in the next section. From a conceptual standpoint, the literature review carried out for this study sought to find con- sensus among the evasion studies completed to date to deter- mine the most promising model or accepted practice. No such consensus or preferred approach was found. Rather, methods used in previous studies varied widely from a simple review of previous literature to complex econometric models. In a small number of studies, more than one method was employed and findings were compared to construct ranges of evasion esti- mates. These strategies, methods, and tools are reviewed in the next section of this report, which offers brief analysis of other alternative methods that hold promise but have not yet been used to study motor fuel excise tax evasion. The most successful approaches were designed with flexi- bility in mind, capturing the unique characteristics of the state being examined (e.g., variance in fuel tax rates in the state relative to its neighboring states, or relative enforcement efforts). There are a number of state explanatory variables (e.g., point of taxation, proximity to international borders) that could also have been incorporated into the methodology, regardless of the approach taken in measuring evasion. It is the uniqueness of these characteristics that pose challenges to the modeler, thus requiring a comprehensive approach mindful of the state-by-state variability in tax code, enforcement pro- grams, and geographic location that largely determine levels of evasion. Chapter 5 outlines the methodology for identifying and quantifying state-level fuel tax evasion by the authors of this report. This chapter is intended to document the numer- ous approaches used to date. 3.2 Strategies, Methods, and Tools for Examining Evasion During the past 20 years, states and the federal government have devised a multitude of tools, strategies, and methods for estimating motor fuel tax evasion. These studies have employed a broad spectrum of approaches. Generally, these studies have used one or several of the following methods to estimate eva- sion: (a) literature review, (b) audit review, (c) analysis of bor- der interdictions, (d) survey of tax administrators, (e) compar- ison of fuel supply with taxed gallons, and (f) econometric analysis. Studies employing these methods, including study findings and authors, were identified in Table 1-1 of Chapter 1. This section examines these methods. 3.3 Literature Review Method The literature review method, as applied to estimate motor fuel excise tax evasion, has historically relied on the work of previous studies of evasion, testimony, anecdotes, and inter- views with motor fuel tax administrators. FHWA used the literature review method in 1992 to estimate all federal fuel tax evasion (1.3 billion) and nationwide state fuel tax evasion (1.2 billion) annually (FHWA 1992). The literature review method has also been used to estimate total state fuel tax eva- sion at as high as $1.5 billion when summed to the national level (CSG & CGPA, 1996). WSLTC used the literature C H A P T E R 3 Strategies, Methods, and Tools To Measure and Evaluate State Fuel Tax Evasion

52 method to estimate evasion in Washington State in 1996. The information gathered by the WSLTC estimated an expected range of fuel tax evasion between $15 and $30 million annu- ally in Washington State (WSLTC, 1996). The literature review method is limited to generating rough estimates of fuel tax evasion based on qualitative information and assumptions. One significant drawback to this method is that certain types of information sources used are not based on rigorous analysis and are often anecdotal and unconfirmed. For instance, the FHWA study (1992) estimates were largely based on the unsubstantiated testimony of state and federal officials, industry representatives, and perpetrators of tax evasion. An additional weakness of relying on previous studies is that assumptions used in those studies rarely share common ground (e.g., data sources, time period covered, methods used, etc.). The WSLTC (1996) study, for example, based its lower bound estimate on several state studies that analyzed the impacts of changing the point in taxation. However, the legis- lation analyzed in these studies differed by fuel type, time period, and the place in the distribution system where the point of taxation was moved. The variation in points of taxa- tion, motor fuel type, and other factors can be sorted out through meta regression analysis, which is an application of quantitative methods to the procedure of comparing and combining results from separate yet similar analytic studies. Stanley (2001) briefly defined meta-analysis as “quantitative research synthesis.” This approach, however, was not used in the studies previously conducted in this field. 3.4 Audit Review Method The audit review method estimates evasion by examining audits of motor fuel excise taxpayers. These evasion estimates are aggregated assessments based on the percentage and degree of fraudulent activity found through random audits. Two studies identified here have employed this method. In both cases, audit reviews were combined with other methods to gauge fuel tax evasion and only were used to supplement other forms of evasion analysis. One study combined this method with the literature review and border interdictions methods to examine Washington State motor fuel excise tax evasion (WSLTC, 1996). At the national level, FHWA employed this approach along with a literature review and testimonies from fuel tax administrators and perpetrators of fuel tax evasion to estimate the level of nationwide state and federal motor fuel excise tax evasion at $1.2 billion and $1.3 billion, respectively (FHWA, 1992). Information collected from the audit reviews could be applied to improve the understanding of motor fuel excise tax evasion. At a basic level, audit analysis could be used to gain a grasp of the potential monetary range that individual schemes cost in terms of evasion. Taking this type of analysis further, evasion estimates could be created by taking the percentage of all audits in which illegitimate activities occurred and aggre- gating that percentage out to the population of taxpayers. When using this method, analysts must control for bias asso- ciated with selecting the sample of companies for audits. Neither identified studies that used this method went beyond the basic level just described, nor did they take mea- sures to control for bias. Though analyzing information from audits may enable analysts to investigate the possible magni- tude of specific evasion techniques, there are a number of draw- backs that need to be addressed to make using the audit method useful in estimating base jurisdictional fuel tax eva- sion estimates. First, the number of firms audited is gener- ally a very small percentage of the total universe of taxpaying companies. The characteristics of such a small sample of tax- payers cannot be universally applied to all taxpayers. Efforts must be taken to understand the sample characteristics and understand how representative the sample is compared with the overall population. Second, audits are not always ran- domly selected; sometimes they are the product of a tip, sus- picious return or previous issue with a higher risk taxpayer. Sample results cannot be aggregated with reasonable confi- dence unless they are either random or steps are taken to con- trol for bias. Third, due to the complexity of business opera- tions, audits may fail to uncover all of the fraudulent activity occurring within the investigated entities. Finally, audit data fail to capture evasion occurring outside the legitimate fuel supply system through techniques such as daisy chains and fuel smuggling. 3.5 Border Interdictions Method Border interdictions involve the examination of petroleum import records and inspection of vehicles and vessels cross- ing international and state borders. Operation activities can involve dipping tanks and inspecting fuel, comparing ship- ment loads with shipping documents and checking IFTA documentation. This method can be used to gauge the mis- use of dyed fuel, illegal importation or exportation, the move- ment of chemicals used for fuel cocktails and certain types of IFTA abuse. The WSLTC (1996) study employed this approach to analyze the activities at the border between Washington State and Canada. This method has several weaknesses. First, the border inter- diction method is limited due to time and cost constraints. Second, it does not provide the basis for a jurisdictionalwide estimate of evasion. Rather, it can only identify a few forms of tax evasion while other forms (e.g., underreporting or nonfil- ing, refund and credit schemes) go undetected. Further, for the evasion techniques it can identify, it will only identify them at the location where the operation is set up; it will not iden- tify them at other points in the jurisdiction and other border

53 points. Lastly, border interdictions may not even be able to assess average evasion occurring at borders. Even when U.S. Customs or state police patrol borders, there is evidence to suggest that tanker operators effectively commu- nicate with each other to avoid such stings or checkpoint oper- ations. One study designed to detect cross-border smuggling examined the operations of petroleum tankers crossing from Canada into Washington State through two international bor- der crossings (WSLTC, 1996). As illustrated in Chapter 2, dur- ing the three-day inspection, there was a marked decline in the number of petroleum tankers passing through these inter- national border crossings, thus demonstrating the ability of tanker operators to communicate with each other in order to detect and bypass inspection operations. 3.6 Survey of Tax Administrators Method Surveying state and federal fuel tax administrators is a method that has been used to investigate fuel tax evasion. The CSG & CGPA (1996) study surveyed tax administrators to dis- cern perceptions of the nature and magnitude of fuel tax eva- sion in each state. On average, fuel tax administrators believed that motor fuel tax revenue would increase by 6.53 percent if fuel tax evasion was completely eliminated for both gasoline and diesel (CSG & CGPA, 1996). When aggregated for all states, this would mean a perceived revenue loss of $1.2 billion annu- ally. Denison and Hackbart (1996) also surveyed state fuel tax administrators to explore the affect of enforcement efforts on state tax collections and applied the results to the State of Ken- tucky. However, the information gathered for this research was not opinion based. Rather, it elicited fact-based information on enforcement programs such as the number of auditors and total assessments. This information was used to support a statistical analysis of enforcement activities on fuel tax collections. One weakness associated with estimating fuel tax evasion based on tax administrator surveys is that many resulting fuel tax evasion estimates are based on unsubstantiated perceptions. In the case that survey responses are based on quantitative analysis, there remains the problem that there will most likely be significant differences in the methods and assumptions used in each analysis. General problems that arise in the survey of tax administrators relate to survey design and bias. Survey bias will be reflected in the way information is presented, the order of the questions, question format, and the survey response rate. 3.7 Comparison of Fuel Supply with Taxed Volumes Method The comparison of supply with taxed gallons is another method for estimating evasion. This method was employed by Addanki et al. (1987) for estimating federal gasoline tax evasion. It also was used to estimate evasion of diesel fuel taxes due to the blending of aviation fuels in vehicles for on-road uses (KPMG, 2001). This method of estimating evasion involves comparing taxed gallons to volumes supplied in the distribution system. The primary problem with this approach is the inherent dif- ferences in how data are collected and treated. For instance, methods used by the EIA to develop fuel supply estimates dif- fer from how FHWA develops estimates of fuel consumption and how IRS-taxed gallons data are generated. In fact, an EIA study showed that EIA estimates of gasoline supplied to the sys- tem are actually less than FHWA estimates of taxed gallonage (Hallquist, 1999). The differences in data collection techniques and discrepancies in data collection lead to various factors (e.g., treatment of blending fuels data, varying data sources, breaks in the time series, double counting of shipments, and incom- plete data) that can undermine evasion estimates. Thus, even though this approach has theoretical appeal, it must be applied with care and researchers must account for a number of fac- tors, including data collection techniques, treatment of blend- ing fuels, and allocation of fuels between taxable and non- taxable uses, for this technique to be valid. Relying on data from the Federal Aviation Administration (FAA), the EIA, and the IRS SOI, KPMG (2001) estimated fed- eral tax leakage due to the diversion of jet fuel to highway use. Based on this approach, KPMG (2001) estimated the cost of evasion associated with the diversion of jet fuel to range from $1.7 billion to $9.2 billion over a 10-year time horizon. The lower bound estimate is based on the presumption that only the 4.4-cent commercial jet fuel tax is being evaded, while the upper bound is based on the assumption that the full 24.4- cent diesel tax is the target of the evasion scheme. This study, however, did not properly account for the reclassification of jet fuel between the terminal and final sale to end users and was criticized by industry (API, 2002). In this case, FAA avia- tion fuel consumption data as reported by U.S. carriers is compared with EIA production of aviation fuel data as reported by refineries. However, data presented in the EIA’s Petroleum Marketing Annual (PMA) report suggest that, in 2002, the amount of fuel ultimately sold by prime suppliers of jet fuel was over 3 billion gallons less than that reported in EIA’s Petroleum Supply Annual (USDOE/EIA, 2002). In the event that studies are not capable of accounting for variations in data collection techniques and discrepancies in the datasets, it is difficult to evaluate whether or not the divergence between fuel consumption and supply is due to tax evasion. Addanki et al. (1987) had begun research with the intent of using this method to estimate federal gasoline tax evasion. However, the study deemed it implausible to identify a rea- sonably precise yearly magnitude of evasion due to the limi- tations and bias in available data sources. The authors decided to use an alternative econometric approach examining trends

54 in supply and consumption data rather than comparing data to estimate evasion in any given year. 3.8 Econometric Analysis The econometric approach can be used to develop a com- prehensive fuels model to forecast fuel excise tax collections for each state based on economic activity and demand for fuel use in each sector. Highway travel, freight transportation, res- idential, and industry all consume fuel in the course of busi- ness and the different fuels (gasoline, gasohol, and distillates) can be used somewhat interchangeably between sectors. For example, domestic freight can be hauled by truck, rail, water (ship or barge), or by air. People can commute to business and take pleasure travel by air, rail, bus, or private vehicle (auto- mobiles, light trucks, and sport utility vehicles [SUVs]). Sim- ilarly, industry can use either gasoline, gasohol, or distillate fuels in machinery and vehicles to conduct business not tax- able for state or federal highway trust fund purposes. In addi- tion, home heating oil used in the residential sector can be used interchangeably with the diesel used in freight trucks. Thus, consumption of gasoline, gasohol, diesel, and aviation fuel (gasoline and kerosene types) could be estimated econo- metrically for passenger vehicles, light trucks and SUVs, heavy trucks, residential, industrial sectors, rail, air, and waterborne traffic sectors. When applied in this manner, the econometric method measures motor fuel excise tax evasion by examining historical structural relationships between economic indica- tors [e.g., nonfarm employment, income, and gross state prod- uct (GSP)] and motor gasoline consumption to predict the escalation and decline in total excise tax liability. Estimated tax liability is, in turn, compared to tax collections to estimate evasion rates. This method, along with the literature review, has been the one applied most regularly when estimating fuel excise tax evasion. The econometric method has been applied at the state and federal levels. The econometric method also has been used to examine the relationship between enforcement activities and returns to state agencies as higher motor fuel excise tax collections. Econometric analysis also has been applied to examine the relationship between apparent motor fuel excise tax evasion rates and state characteristics (e.g., geographic proximity to international borders, motor fuel excise tax rate differential relative to border states) that exacerbate or curb evasion. There are shortcomings to the econometric method (e.g., availability and reliability of data, changes in structural rela- tionships between variables over time, and the inability of econometrics to predict future extraordinary events and un- expected trends); however, Battelle in Weimar et al. (2002) viewed this technique as conceptually promising and used it to examine evasion at the federal level for the IRS. Weimar et al. controlled for shortcomings by independently verify- ing components of fuel use wherever possible and using the most accurate available data sources where multiple sources existed. The econometric and statistical models used in previous studies were based on survey data as well as time-series data. Previous state studies have indicated that differences among states in terms of program structure, geographic location, and tax rate have largely determined estimated rates of tax evasion (Addanki et al., 1987; Mingo et al., 1996). Earlier research in this field, however, has largely under-examined the dynamic nature of motor fuel excise tax evasion over time and other significant effects such as the impact of inflation on evasion penalties. 3.8.1 CSG and GPA 1996 Study of Motor Fuel Tax Evasion The CSG & CGPA (1996) study utilized a combined approach to examine motor fuel tax evasion employing the literature review, survey of state tax officials and the econo- metric approaches (CSG & CGPA 1996). The findings of the study are shown in Table 3-1. The variables used in the statis- tical model were (1) income/wealth, (2) demographic charac- teristics of the population, (3) price variables, (4) geographic dispersion variables, and (5) other variables. 3.8.2 Mingo et al. 1996 Study for Proposed Diesel Tax in Oregon Mingo et al. (1996) employed a linear regression analysis to study the impact of tax evasion on a proposed diesel tax for trucks weighing in excess of 26,000 pounds in the State of Oregon (Mingo et al., 1996). Presently, Oregon is the only state in the nation that does not impose a diesel tax on trucks weighing in excess of 26,000 pounds, instead relying on its weight-mile tax. The weight-mile tax is based on a graduated fee schedule with rates that grow in relation to the declared weight of a heavy truck configuration. The amount of the weight mile tax is based on the declared weight of the vehicle and the miles it travels in Oregon. Method Value of State Fuel Tax Avoided in Billion ($) 1- Literature-Based Estimates 1.5 2- Survey-Based Estimates 1.2 3- Statistical Model 0.952 Source: CSG & CGPA, 1996. Table 3-1. Findings of CSG & CGPA 1996 study of fuel tax evasion.

55 Mingo attempted to measure the extent of evasion associated with a proposed tax in Oregon by examining the relationship between various factors and perceived motor fuel excise tax eva- sion and then applying the resulting model in Oregon. Using calculated noncompliance as the dependent variable, the study used the following regressors: (1) whether the state is coastal or bordering another state, (2) the diesel tax rate of nearby states and proximities of their population centers, (3) the intensity of truck ownership and usage within the state, and (4) the relative rates of other truck taxes within a state. The model was success- ful in explaining only three-fourths of variation in compliance rates among states. 3.8.3 Eger and Hackbart 2001 Study of 50 States Eger and Hackbart (2001) reviewed road fund assessment, collection, audit, and enforcement processes using survey data for several states and developed recommendations to improve the efficiency of the road fund collection process in Kentucky (Eger and Hackbart, 2001). To collect data for the statistical model employed in this analysis, an electronic survey was sent to road fund tax administration officials in the 50 states and the District of Columbia. Survey respondents based their answers on information for fiscal years 1997 and 1998. The survey was designed to collect information needed for the statistical model’s dependent and independent (explanatory) variables represented in Table 3-2. The econometric model was devel- oped to explore the affect that enforcement, auditing, and assessments had on highway fund revenue compliance. The strength of the Kentucky study is that the authors relied on cross-sectional data and survey data as opposed to time- series data. By using cross-sectional data, the authors avoided problems typically associated with time-series data, such as statistical aggregation errors, seasonality issues and the problem of using nominal versus real values. Eger and Hackbart also Dependent Variables: Variable Name Variable Definition 1 Assessment per million truck VMT for FY 1997 (for Model 1) Total amount of assessment (defined as the total tax due per audit less the amount reported by the taxpayer with original return) due to audits, of all taxpayers combined, of highway revenue fund (i.e., motor fuel taxes, motor carrier fees, etc) audits in fiscal year 1997 per million VMT. In other words, this variable reflects the absolute assessment value per million VMT without influencing the value by differentials in penalty and interest used by each individual state. 2 Assessment for FY 1997 in real dollars (for Model 2) Total amount of assessment in real dollars (defined as the total tax due per audit less the amount reported by the taxpayer with original return) due to audits, of all taxpayers combined, of highway revenue fund (i.e., motor fuel taxes, motor carrier fees, etc) audits in fiscal year 1997. In other words, this variable reflects the absolute assessment value without influencing the value by differentials in penalty and interest used by each individual state. Independent Variables: Variable Name Variable Definition 1 Border State A dummy variable that includes the states bordering Kentucky: Illinois, Indiana, Missouri, Ohio, Tennessee, Virginia and West Virginia and including Kentucky 2 Number of Field Auditors The number of field or desk auditors as reported on the survey 3 Diesel Tax Excise tax in cents per gallon of diesel for 1997 4 Per Capita Income Per capita income measured in 1997 dollars 5 Urban Road Miles Miles of road in urban areas owned by state highway agencies 6 Rural Road Miles Miles of road in rural areas owned by state highway agencies 7 Federal Tax Contribution Amount of federal tax revenue awarded to the state for FY 1997 8 Location A dummy variable that takes 1 if collection agency/department is revenue, 0 otherwise Adopted from Eger and Hackbart (2001) Table 3-2. Eger and Hackbart (2001) statistical model dependent and independent variables.

56 applied a number of regression analysis functional forms: (1) linear model with assessment per million VMT as a dependent variable; (2) linear model with total assessments in real dol- lars as a dependent variable; and (3) a log-log functional form of Model 2. Reliance on the most representative data set and selection of the correct functional form are basic econometric practices often encouraged before deciding on the most appro- priate model for estimation. A significant problem with this study is its reliance on VMT data, which is usually criticized due to its large margin of error, particularly on lower-order road systems (e.g., county roads and city streets). At the national level, such aggregated data usually hide significance variances between the states. Eger and Hackbart (2001) reported that, of the aforemen- tioned eight explanatory variables, only three were found to be statistically significant at the 0.05 level of significance (95 per- cent level of significance). Based on these equations, total nationwide state motor fuel excise tax evasion was estimated to be $952 million. A shortcoming in the model developed in Eger and Hackbart is that a large proportion of insignificant variables typically indicates that a model is not statistically sound, even in the case where the model’s R-square, the mea- sure of the overall model fitness, may be relatively high (Eger and Hackbart, 2001). 3.8.4 Hackbart and Ramsey 2001 Hackbart and Ramsey (2001) developed a stepwise regres- sion model to estimate revenue loss due to motor fuel excise tax evasion. Three equations outlined below were used by Hackbart and Ramsey (2001) to determine which independ- ent variables provided the best fit in terms of explaining the demand for motor fuels. The first equation in the model deployed by Hackbart and Ramsey (2001) estimated the demand for motor fuel on a gallon-per-resident basis for each state as follows: The second equation estimated the gallons of fuel per driver: The third equation estimated gallons of fuel per vehicle: Equation 3: Gallons of fuel per vehicle = f (per capita personal income, land area per 1,000 residents, and interstate highway miles per 1,000 residents) Equation 2: Gallons of fuel per driver = f (per capita personal income, land area per 1,000 residents, and interstate highway miles per 1,000 residents) Equation 1: Gallons of fuel per resident = f (per capita personal income, land area per 1,000 residents, and interstate highway miles per 1,000 residents) To calculate total motor fuel consumption, Hackbart and Ramsey multiplied the estimated gallons of fuel per resident by each state’s population (Equation 1), gallons of fuel per driver by the number of registered drivers in each state (Equation 2), and gallons of fuel per vehicle by the number of registered vehicles in each state (Equation 3). These esti- mates of motor fuel consumption were combined with tax rates to produce estimated state motor fuel tax liability. These three estimates of liability were compared with actual tax collections in each state and the differences were calcu- lated. In turn, these three evasion estimates were averaged. Finally, the evasion estimates for all states were summed to produce a nationwide estimate of state motor fuel excise tax evasion. Hackbart and Ramsey (2001) employed a statistical model that appears to be more efficient in econometric terms relative to that employed by Eger and Hackbart (2001) because the authors dropped the statistically insignificant variables. Each explanatory variable included in Hackbart and Ramsey (2001) was shown to be statistically significant. In turn, the model’s overall performance is shown to be more statistically sound when compared to the Eger and Hackbart (2001) model. The model showed high F-statistics (overall significance of an overall multiple regression) as well as a high R-square test of a step-wise regression. Though the model deployed by Hackbart and Ramsey (2001) is statistically sound, it doesn’t include a number of economic variables that are theoretically likely to drive motor fuel con- sumption (e.g., gross state product, nonfarm employment and motor fuel prices). Statistical models should be sound from both a theoretical and statistical standpoint. Hackbart and Ramsey (2001) relied on cross-sectional data collected in 1992 as part of the GSG & CGPA (1996). Data used to support the model were collected prior to several advance- ments in motor fuel excise tax collection. For example, the data were collected prior to: • Federal government collecting diesel tax at the terminal rack, • Dyed fuel requirements, • Taxation of kerosene, • Development of ExSTARS and many state fuel tracking systems, and • Movement of many state points of taxation up the distri- bution chain. 3.8.5 Addanki et al. 1987 Gasoline Tax Evasion in New York Addanki et al. (1987) employed the econometric method using time-series data for the period 1974–1982. The study fitted IRS data on taxed gallons to FHWA and EIA gasoline

57 consumption estimates over the 1974–1982 time period using a conventional regression technique. The authors then used the estimated trend line to estimate what IRS taxed gallons should have been for the 1984–1986 time period based on FHWA and EIA estimates of consumption. The study also examined the issue of gasoline tax evasion in New York State, using regression analysis to estimate gasoline tax evasion at $168.4 to $254.5 million annually. This estimate would appear to be extremely high given that estimates of state gasoline tax evasion at the national level have historically hovered in the $600 million to $1.2 billion range. The model employed in Addanki et al. (1987) included taxed gallons as the dependent variable and the FHWA-estimated consumption estimates and the EIA-estimated consumption estimates as the independent variables. The authors in the Addanki study regressed the dependent variable (taxed gallons in billions of gallons) on two explanatory variables (FHWA and EIA consumption statistics) as indicated in the following equation. 3.8.6 Eger et al. 2003 Agricultural Consumption of Tax-Exempt Fuel in Midwestern States Eger et al. (2003) presented a statistical analysis of trends in the consumption of tax-exempt fuel in the agriculture sec- tor in Midwestern states. Eger et al. (2003) used Ordinary Least Square (OLS) and the Autoregressive Analysis to correct for autoregressive error terms over time. The model employed by Egger et al. regressed seasonal monthly gasoline refunds in agri- culture (dependent variable) against a number of explanatory variables: monthly fuel tax rates (cents), number of farms in state, average acreage of farms, and dummy variables for each state included in the analysis. This study used this model and a time-series cross-sectional data set (panel data) to conclude that Wisconsin’s annual consumption of tax-exempt fuel for agri- cultural uses exceeds the average for other Midwestern states by nearly $4 million annually. This higher-than-anticipated level of refunds in Wisconsin suggests relatively higher levels of refund fraud. 3.9 Alternative Econometric Methods In addition to models previously used to estimate motor fuel excise tax evasion, the research team also reviewed two other econometric modeling approaches that hold promise but have not yet been used in this field. Other models of note Equation 4: Taxed gallons in billions of gallons = (FHWA consumption estimate from H f ighway Statistical reports, EIA consumption estimate from Petroleum Supply Monthly reports) mentioned here are the simultaneous fuel supply and demand model and the Tobit model. The Tobit model has been used by economists to assess income and other forms of tax evasion. 3.9.1 Simultaneous Fuel Supply and Demand Model In contrast to single-equation models, a simultaneous sup- ply and demand model uses more than one dependent variable and necessitates as many equations as the number of depend- ent variables. This model uses a system of equations rather than one equation and fits the fuel tax evasion case better because it examines an entire motor fuel system. Simultaneous models take advantage of the fact that supply and demand should be in equilibrium and, to the extent the model identifies large un- defined discrepancies between supply and demand, tax evasion may be present. Both prices and quantities of fuel supplied and demanded are considered in the model structure. A unique feature of the simultaneous model is that a dependent vari- able in one equation may appear as an explanatory variable in another equation in the system. The econometric method is used to examine the relation- ship between taxable motor fuel consumption and various economic indicators (e.g., nonfarm employment, income, and GSP), tax program elements, geographic characteristics, and other factors that exacerbate or curb evasion at the state level. The strength of econometric analysis is that it can use historical data for a dependent variable and several explana- tory variables to measure the relationships among variables to predict future values of the dependent variable or exam- ine trends in the escalation and decline in the dependent variable. However, when applying econometric analysis to tax eva- sion, the analyst is unable to measure the dependent variable (e.g., evasion rates) directly and must rely on indirect measures or assumptions concerning taxable fuel consumption based on estimates of supply or demand, with estimated demand generally being calculated by dividing estimated VMT by estimated motor fuel economy by vehicle class. Data required to perform such a computation, however, often have too many shortcomings to adequately support this form of analysis. For example, there are shortcomings with FHWA VMT data (e.g., sampling techniques are used to estimate VMT on the lower-order road systems, traffic counting devices malfunction, and more emphasis is commonly placed on mature growth areas). Furthermore, estimates of motor fuel economy reported to FHWA as presented in highway statis- tics are largely based on state fuel tax records. To the extent that fuel tax evasion occurs at the state level and errors are present in the survey data collected to estimate fleet MPG, there will be shortcomings in FHWA-estimated vehicle motor fuel economy.

58 3.9.2 Tobit Model Another econometric method that improves upon the sim- ple statistical sampling analysis is the Tobit model, an extension of the Probit model. The Probit model is a binary-choice- based regression model (the dependent variable in the regres- sion analysis is binary in nature taking 0 or 1 values only). The Tobit model was first introduced in 1958 by James Tobin. Tobit and Probit models substitute the normal cumulative distribu- tion function (CDF) in place of the logistic CDF of the Logit model. This expansion has been found to be useful in solving issues related to using OLS regression analysis. Both Probit and Logit solve the problems of the binary choice variable. Tobit is a variation on the Probit. With missing information on the dependent variable, for example the data is censored (having a value of 0 or 100 when the actual value would be less than 0 percent or greater than 100 percent) the coefficients esti- mated by OLS are likely to be biased. Tobit corrects for this bias. Each model is based on the idea that regular forms of regres- sion analysis fail in the case of a binary dependent variable, such as evasion or no evasion. These models consider the appropriate mathematical adjustment needed to ensure that basic regression analysis assumptions do not fail (e.g., depend- ent variable values are not bound, explanatory variable values are fixed in repeated sampling, zero mean value of the error term, homoscedasticity or equal variance of the error term, and no autocorrelation between the error terms). The Tobit model also deals with the issue of censored data. The censored data issue arises when there are two groups of respondents. For the first group, we have information about both the dependent variable (what we are measuring) and the explanatory variables. In the second group, we have infor- mation concerning the explanatory variables but no infor- mation relating to the dependent variable. A sample in which information on the dependent variable is available from some but not all observations is known as a censored sample (Gujarati, 1995). The Tobit model fits the tax evasion prob- lem because the dependent variables (i.e., tax evasion levels) are not completely known. Economists like Jonathan Feinstein (2003) have used the Tobit model to analyze and estimate nationwide income tax evasion based on IRS data. The advantage of the Tobit model over the Logit and Probit models is that it solves the prob- lem associated with the binary variable but also handles the lack of information with respect to some dependent variable observations.