Measuring International Trade on U.S. Highways (2005)

In response to congressional legislation, the Bureau of Transportation Statistics (BTS) submitted a report to Congress that presented the methods and findings of a study that (1) estimated the ton-miles and value-miles of international trade traffic carried by highway in each state; (2) evaluated the accuracy and reliability of such estimates for use in formulas for highway apportionments; (3) evaluated the accuracy and reliability of the use of diesel fuel as a measure of international trade traffic by state; and (4) identified needed improvements in long-term data collection programs to help provide accurate and reliable measures of international traffic for use in formulas for highway apportionments (Bureau of Transportation Statistics, 2003).

BTS estimated that there were nearly 227 billion ton-miles and $488 trillion value-miles of international trade traffic carried by highways in 1997. Although these are impressive amounts, these estimates, particularly at the state level, are not very reliable. There are no direct measures of ton-miles and value-miles of international trade traffic carried by highway in each state. In fact, methods for measuring total volumes of ton-miles and value-miles passing over state highways are still in development. Thus, ton-miles and value-miles must be estimated with data from a multitude of public and private sources, none of which was specifically designed for the task.

The quantity of ton-miles is the product of the tonnage of a shipment times the miles the shipment traveled to its destination. The quantity of value-miles is the product of the declared value of a shipment times the miles the shipment traveled. Both ton-miles and value-miles are derived from separate estimates of either weight or value and miles traveled.

Ton-mile and value-mile figures are combined from separate estimates of export and import trade traffic for a variety of categories. Despite growing attention to measures of tonnage, miles, and value, both ton-mile and value-mile estimates have many substantial sources of error. Because the data sources are not designed to directly yield ton-mile and value-mile measures, the estimates require application of model-based estimation techniques that rely on unvalidated assumptions. The models that were used in the BTS study included the Oak Ridge National Laboratory’s (ORNL) highway network model to derive estimates of the distances traveled by imports and exports. The accuracy of the distance estimates from this model is unknown. Moreover, data on imports were available only at the state level. The accuracy of the assumptions used to distribute state-level imports to counties within a state is unknown because of a lack of data at the substate level, and substate level allocations are needed to determine the likely route taken, which provides a better assessment of the miles traveled within a state.

Moreover, the various data sources utilized to produce estimates of international ton-miles and value-miles transported by highway have known and unknown errors and biases that affect the accuracy and reliability of the estimates. The errors and biases are due not only to typical sources, including sampling variance, nonresponse variance, and measurement error of various types, but also to various data deficiencies, even though the models mentioned above are used to overcome these deficiencies.

The BTS study identified a number of sources of error for a variety of data sets used to estimate volumes and values for freight arriving here as imports or departing as exports, which we summarize below. In addition, there are shipments involving Canada and Mexico that are not straightforward exports or imports: truck shipments between Canada and Mexico and shipments to or from the United States that are transshipped via Canada or Mexico are not fully captured in current trade data for all modes of transportation, which is also a source of error. Also, some truck transportation was not included in the estimates of the amount of international trade because no reliable data were available, especially including drayage, or short-haul shipments.

The 1997 Commodity Flow Survey (CFS), resulting from a partnership between the U.S. Census Bureau and BTS, is a stratified random sample of commodity flows. The CFS is conducted every 5 years as part of the Economic Census Program, and it is mandatory. The establishments that are surveyed are selected from a frame (a current list) constructed from the Census Bureau’s Business Register. For a sample of shipments for a 1-week reporting period each quarter, respondents report the total number of outbound shipments and information on value, weight, primary commodity, domestic destination or port, airport, or border crossing of exit, foreign destination, and mode of transportation. In terms of shipments, the coverage is extensive: it is estimated that the 2002 CFS gathered information on more than 2.5 million shipments. It is a valuable source of information on the flow of goods by truck and other modes in the United States, and it is the primary source of BTS estimates of ton-miles and value-miles for exports. Unfortunately, the CFS does not provide useful information for imports since goods are not covered in the CFS until they reach the first domestic shipper covered by the CFS. A substantial amount of travel can be missed before the shipment reaches the first domestic shipper.¹

While the CFS includes information on shipments of manufacturing, mining, wholesale trade, and selected retail and service industries in the 50 states and the District of Columbia, it excludes shipments of farm-based, forestry, fisheries, transportation, and oil and gas extraction companies. The contribution of these industries to export trade is relatively small, and therefore, while this omission results in the CFS estimates being biased low (excluding other sources of bias), the BTS report says that this bias is a fairly minor source of error. However, CFS also excludes construction, which may represent a more substantial contribution to bias. In terms of all shipments, including imports, only about 60-70 percent of shipments are covered by the CFS.

In addition to sampling variance, the CFS, like any survey, is subject to undercoverage, misresponse, and nonresponse. Not much is known about undercoverage other than that mentioned above. There is no information about the degree of misresponse, which is typically unmeasured. However,

the sampling variance may have increased of late, since the sample size for the CFS has been reduced from 200,000 in 1992 to 50,000 in 2002. Although this still appears to be a substantial sample size, this reduction could reduce the quality of estimates of international trade traffic since export activity in the CFS is a relatively small percentage of total shipments. Specifically, the BTS report states that exports using all modalities accounted for about 8 percent of estimated total value and 8.2 percent of ton-miles in 1997, and the domestic part of exports shipped by truck accounted for 7 percent of the total value shipped by truck and 5 percent of the ton-miles. The coefficients of variation² for the estimates of exports shipped by truck were 5.5 percent for value and 12.4 percent for ton-miles. An analysis of the 1997 response rates indicated an overall unit nonresponse rate of 15 percent, with differences by trade area, size of establishment, and state. Also, the rate of item nonresponse for shipments was 2.7 percent for entries for value and 4.0 percent for entries for weight. Although this amount of unit and item nonresponse likely increases the sampling variance (and therefore the coefficients of variation) of resulting estimates, a more important concern is that this degree of nonresponse could also substantially increase the estimates’ bias.

Imports can be primarily separated into maritime, by land, and by air. With respect to maritime imports, the percentage of imports that arrive at seaports that are subsequently transported by truck is unknown. Also, for those maritime shipments that are subsequently transported by truck, there are only limited data on the state of destination. This lack of direct information necessitates the use of models to estimate the quantities of ton-miles and value-miles for imports transported by highway at the state level.

BTS used the U.S. Foreign Waterborne Transportation statistics to determine the tonnage and value of imports arriving at all U.S. seaports. These data are derived from U.S. foreign trade filings collected by the Customs and Border Protection Agency of the U.S. Department of Commerce. The value of an import is estimated from the weight and the commodity code, which introduces some error. To estimate the amount of maritime imports that are subsequently transported by truck from ports, by com-

modity, BTS used data from the Public Use Carload Waybill Sample, collected annually by the Association of American Railroads, to first estimate those quantities for imports transported by rail. (For this purpose, international freights can be identified by an intermodal service code [ISC] and by type of move code.) When rail freight shipments are subtracted from total international freights transported from seaports (ignoring petroleum products), the remainder is assumed to be transported by truck. In some areas, this assumption introduces a substantial bias since it will result in inland and coastwise shipments by water being counted as being transported by truck.

To allocate total imports transported by truck to the state level, BTS used data from the Port Import/Export Reporting Service (PIERS), collected from vessel manifests. Unfortunately, state information is often missing, and the state information may be the address of the importer and not the actual destination. To address this deficiency, BTS sought state-level destination information from the U.S. Foreign Waterborne Transportation System from the Census Bureau, but this request was denied due to confidentiality requirements under Title 13 of the U.S. Code.

For imports arriving by land, BTS used the Transborder Surface Freight Database to collect data on the tonnage, value, origins, and destinations of truck trade with Canada and Mexico. This database contains monthly freight flow data by commodity type and by surface modes for U.S. exports to and imports from Canada and Mexico. Finally, with respect to imports that arrive by air, subsequent transportation by truck is considered negligible, since the distances traveled by truck from airports are typically short.

The BTS report found that, although international trade data are extensive and detailed, the administrative data collected by the Customs and Border Protection Agency and processed by the Census Bureau are limited in their usefulness in measuring international trade traffic on state highways. Critical data elements necessary to identify destinations and transportation modes are not available. Data from the U.S. Foreign Waterborne Transportation Statistics, PIERS, and the Transborder Surface Freight Database provide important pieces of information, but each suffers from issues of data quality and compatibility. Furthermore, the BTS study found that many of the sources of data required to prepare and improve the estimates of international trade traffic by highway by state are collected under pledges of confidentiality and are protected from other use by legal mandate. Also, some of the potential data sources are developed in the private

sector and are of unknown quality. As a result of these data problems, BTS used inferior sources of information to produce the requested estimates.

Specifically, the BTS report listed five deficiencies: (1) data on shipment weight for imports transported by truck (which is now required by the Census Bureau as of October 2003); (2) information on the state and county of destination for imports and route taken for exports; (3) information on the port of entry for imports rather than the administrative port; (4) data on all modes of transportation used in shipping to a destination, in addition to the mode used when the cargo arrives in or departs from a U.S. port of entry; and (5) data on truck shipments between Canada and Mexico that transit the United States and shipments to or from the United States that are transshipped via Canada or Mexico. Since BTS does not have access to detailed customs data, it must use extracts published as port-of-entry data in the U.S. Foreign Waterborne Transportation Statistics and trade data with Canada and Mexico in the Transborder Surface Freight Database. Access to customs data would alleviate many, but not remove all, of the five above deficiencies in current data.

Since knowing only the state of destination provides limited information as to the specific highways used, it is important to “carry down” information on import freight movements from the state level to the county level. To do this, BTS initially intended to use four “gravity” models. The four models were of flows from border crossing to counties separately by weight and by value, and flows from seaports to counties, again separately by weight and by value. However, due to computational constraints, the gravity models could not be implemented at the county level and so were only used to re-estimate the ton-miles and value-miles for imports at the state level. Since the estimates already available existed at the state level, the main reason for this application of the gravity model would be to reduce the variance in the available estimates through the replacement of the observations with the fitted values from the model.

To carry these modified estimates of state-level transported freight down to the county level (specifically, the centroid of each county), BTS used a model that allocated each state’s quantities (ton-miles and value-miles) to counties in proportion to each county’s share of total state payroll, as reported in the 1997 County Business Patterns data. (These data include statistics on county-level establishments, employment, and payroll by

employment size class for major industry groups and is collected by the Census Bureau.) Given the limited relationship of payroll to international freight, one would expect, at best, that this method would provide only a rough approximation to the true distribution. As a result, there is additional error due to the use of this apportionment method.

Knowing that freight moved from a port or border crossing to a county centroid is not sufficient for understanding what roads were used in that shipment, and more fundamentally, what mode of transportation was used. For the identification of specific highways used in transporting freight from one county centroid to another, BTS used the ORNL national highway network model, which is part of ORNL’s intermodal network model. The model is a geographically based network of major roadways, currently representing 420,000 miles, which simulates shipment routes by roads from two points. The benefit of having this highway model embedded in the ORNL intermodal network model allows one to model, based on transportation costs and travel time, whether freight that is moving from one point to another is likely to be transported using trucks, trains, waterways, or by air. Some of the models that select a mode of transportation or a road use a “least impedance approach” that deterministically selects a transportation option. The remaining models use a logit³ approach that weights various choices based on their estimated probabilities. The performance of these models is hindered by incomplete data, missing and misaligned roadways, and reliance on a number of questionable assumptions. For instance, the assumption that routes are chosen to minimize costs is violated when truck drivers choose routes based on the need to make multiple pick-ups and deliveries and the ability to mix freight shipments. Also, state specific truck weight limits can affect routes selected. Furthermore, sometimes time of day is factored into a decision in order to avoid congestion. Finally, transportation terminals open and close fairly often, affecting the currency of the ORNL road and modal network. All of these factors contribute to a concern that many of the model’s assumptions are invalid. Along the same lines, quantitative measures of the accuracy of the assignments made by the network model do not exist.

The third charge to BTS in TEA-21 was to evaluate the accuracy and reliability of the use of diesel fuel data as a surrogate for international trade traffic by state. The focus on this particular approach to modeling international trade traffic is difficult to understand given that no breakdown of diesel fuel data pertains specifically to international trade. Therefore, there is no direct way to use only that portion of diesel fuel data to model international trade traffic by state.

BTS carried out a modification of the required analysis, using diesel fuel data at the state level as the independent variable in a simple linear regression to model the dependent variable ton-miles of highway traffic—not ton-miles of international highway traffic. It was hoped that if this model was successful, one could make use of another model that would effectively estimate the portion of diesel fuel used by international trade traffic, possibly by modeling the ratio of international freight ton-miles per state to the total freight ton-miles per state. However, in that case, one would need to rely on an assumption concerning the stability of these ratios, which might not be sensible. Presumably, similar approaches might also be applicable to estimating international value-miles at the state level.

In carrying out the indicated simple linear regression, the percentage of variation explained by diesel fuel sales, R², was 88 percent, which represents a correlation between these variables of 0.94. This result was certainly promising, but there were two problems, one of which was pointed out by BTS.

First, while the correlation is quite high, 32 states had differences between the observed and modeled ton-miles that were greater than 15 percent of the observed values, and 11 states had differences of greater than 50 percent. These are important discrepancies that argue against use of these estimates for fund allocation. Yet the discrepancies are not surprising because (1) the fuel sold in a state is not equivalent to the fuel used in a state, (2) there are differences in states’ reporting of motor-fuel sales, and (3) the statistics R² and the correlation coefficient can be dominated by a minority of larger values in a dataset (e.g., a few large states with high quantities of diesel sales and trade traffic could result in an apparently good fit without providing small relative residuals for many of the smaller states).

Second, this model is intended to be used predictively—that is the regression coefficients would be fixed for several time periods, and R² is not a measure of the predictive performance of a regression model. The causal

justification for this regression model was very limited. In particular, the regression model failed to account for changes over time in the fuels used to drive trucks and the changes in demand for diesel fuel from other uses.

As requested in TEA-21, BTS was asked not only to provide estimates of ton-miles and value-miles of international highway trade traffic, but also to investigate the reliability of these estimates. In response to this charge, the BTS report asserted that there are important improvements needed in many of the components of their estimation methods for reliably estimating ton-miles and value-miles. Generally speaking, the improvements are needed because of deficiencies, either in current data coverage and quality or with respect to access to data of higher quality that is collected but is currently unavailable.