information, decisions, and productivity: on-board computers and … · 2004-07-26 · information,...

Information, Decisions, and Productivity:On-Board Computers and Capacity Utilization in Trucking

By THOMAS N. HUBBARD*

Productivity reflects not only how efficiently inputs are transformed into outputs, butalso how well information is applied to resource allocation decisions. This paperexamines how information technology has affected capacity utilization in the truck-ing industry. Estimates for 1997 indicate that advanced on-board computers(OBCs) have increased capacity utilization among adopting trucks by 13 percent.These increases are higher than for 1992, suggesting lags in the returns to adoption,and are highly skewed across hauls. The 1997 estimates imply that OBCs haveenabled 3-percent higher capacity utilization in the industry, which translates tobillions of dollars of annual benefits. (JEL D24, L92, O33, O47)

Theoretical links between economic perfor-mance and the use of information, such as thosein F. A. Hayek’s (1945) famous analysis of eco-nomic organization, are at the core of a recurringtheme in the productivity literature: the premisethat information technology (IT) offers oppor-tunities for large productivity gains. Empiricalevidence showing links between IT diffusionand productivity has been scarce untilrecently,however.1 Researchers in the field refer tothis as“the productivity paradox.” Thedifficulty offinding relationships between IT use and produc-tivity using aggregate data is well-summarized byRobert Solow’s oft-cited observation: “You cansee the computer age everywhere except in theproductivity statistics.”

This paper examines micro-level empiricalrelationships between IT use and productivity inthe trucking industry in the 1990’s. Productivity

in this industry, as elsewhere in the economy,depends critically on how well information isbrought to bear on resource allocation deci-sions.2 Supply and demand conditions changeconstantly; forecasting exactly when and wheretrucks will be available and exactly when andwhere shippers will demand service is difficultmore than a few hours in advance. Informationabout trucks’ availability and value in differentuses is highly dispersed, and communicationcosts create situations where the individuals de-ciding how individual trucks should be used—usually, dispatchers—do not have goodinformation about trucks’ availability. Trucksare not always allocated to their most valuableuse as a consequence. Poor matches betweencapacity and demands lead to underutilizationin the form of idle trucks and partially full orempty trailers.

Using truck-level data collected by the U.S.Bureau of the Census, I examine how on-boardcomputer (OBC) use has affected capacity uti-lization. OBCs help managers at trucking firmsor divisions monitor trucks and drivers. Low-end devices—trip recorders—make truck driv-ers’ activities more contractible and helpmechanics diagnose engine problems. High-end

* Graduate School of Business, University of Chicago,1101 East 58th Street, Chicago, IL 60637, and NBER (e-mail: [email protected]). I would like tothank Ann Merchant for excellent research assistance andmany individuals at trucking firms for useful discussions. Ialso thank Marianne Bertrand, Dietmar Harhoff, Amil Pet-rin, Scott Stern, Jack Triplett, two anonymous referees, andseminar participants for comments. I gratefully acknowl-edge support from NSF Grant No. SES-9975413.

1 Frank R. Lichtenberg (1995); Erik Brynjolffson andLorin Hitt (1996); William Lehr and Lichtenberg (1998);Stephen D. Oliner and Daniel E. Sichel (2000); Dale W.Jorgenson (2001); Susan Athey and Scott Stern (2002). SeeBrynjolffson and Shinku Yang (1996) and Brynjolffson andHitt (2000) for surveys of the evidence.

2 That firms pay thousands of dollars for supply chainmanagement software that provides managers up-to-dateinformation about the status of production processes andinventories testifies that information about capacity is valu-able and costly to obtain in other contexts.

1328

devices—electronic vehicle management sys-tems (EVMS)—also provide dispatchers real-time information about trucks’ location and anefficient means of communicating with distantdrivers. These additional capabilities let dis-patchers make and implement better resourceallocation decisions: they can allocate trucksacross existing orders and market excess capac-ity better than they otherwise could. This, inturn, can lead to better matches between truckcapacity and demands within and across firms.Better matches boost capacity utilization andproductivity in the industry.

I find evidence that OBC use has increasedcapacity utilization significantly in the industry.Estimates using 1997 data indicate that loadedmiles per period in use are 13 percent higheramong trucks for which advanced OBCs havebeen adopted than those without OBCs. Otherevidence suggests that this reflects that OBCshave caused capacity utilization increases byimproving dispatchers’ ability to make and im-plement resource allocation decisions. There islittle evidence of truck utilization increases dueto incentive improvements. The average bene-fits to adopters are higher in 1997 than 1992,suggesting lags in the returns to adoption, andare highly skewed across hauls. About three-quarters of the capacity utilization benefits areon trucks that haul goods long distances innonspecialized trailers. The 1997 estimates implythat OBC-enabled improvements in decision-making have led to 3.3 percent higher capacityutilization in this nearly $500 billion sector ofthe economy, which translates to about $16billion in annual benefits. These benefits arelikely to increase as complementary economicinstitutions such as centralized markets developin the industry and as diffusion becomes morewidespread.

This study stands at the intersection of theproductivity, economics of technology, andeconomics of organizations literatures, and isimportant for several reasons. First, it providesstrong evidence of productivity gains from ITadoption. There is no “productivity paradox” intrucking. This study adds to a growing set ofstudies that document relationships betweenproductivity and IT use, some of which are citedabove. Relative to most other studies, the dataand context studied here provide for an unusu-ally good environment for measuring IT-related

productivity gains. Second, as the Hayek citeindicates, understanding relationships betweeninformational and resource allocation improve-ments is central for understanding the perfor-mance of economic organizations and howdecreases in information costs lead to increasesin welfare. This is one of the first empiricalstudies to examine these relationships in detail.An advantage of this paper’s micro-level indus-try study approach [shared by Athey and Stern(2002)] is that one can understand exactly howand why IT use leads to productivity gains.Third, truck-tracking is one of the first commer-cially important wireless networking applications.Wireless networking applications are expectedto diffuse more broadly in the economy in thenear future; this study helps researchers under-stand their economic implications. The conclu-sion that OBCs have generated large benefits intrucking suggests that new networking applica-tions have the potential to generate large wel-fare gains elsewhere.3 Last, few individualapplications have the potential for as significanta macroeconomic effect as OBC-enabled truck-tracking. OBCs fundamentally improved re-source allocation decisions in an industry thatinteracts with most sectors of the economy andamounts to about 6 percent of GDP (includingthe value added produced by private fleets).OBC diffusion and related logistical improve-ments were nontrivial contributors to economicgrowth in the United States during the 1990’s.

An outline of the rest of the paper follows.The next section describes the institutional set-ting and depicts how OBCs improve incentivesand resource allocation decisions in trucking.Section II presents the data and the basic em-pirical patterns. Section III outlines the empiri-cal framework. Section IV discusses theestimation results. Section V concludes.

I. Information and Capacity Utilization inTrucking

The physical part of the production process intrucking is simple. Cargo is loaded onto a truck,or a truck’s trailer. An individual—a driver—drives the truck to its destination, where the

3 See Robert J. Gordon (2000) for a skeptic’s view.

1329VOL. 93 NO. 4 HUBBARD: INFORMATION, DECISIONS, AND PRODUCTIVITY

cargo is unloaded. The output of the productionprocess is the movement of cargo.

All else equal, costs per unit of output fallwith capacity utilization. The per-unit cost ofmoving cargo on a truck increases less thanproportionately with the weight of the cargo,and firms bear opportunity costs when trucksare idle, especially when idle trucks imply idledrivers. Truck capacity is lumpy and location-and time-specific. Capacity utilization is highwhen trucks haul a series of full loads, each ofwhich starts close to and soon after the previousone finished.

Achieving high levels of capacity utilizationis easy in some circumstances, but hard in oth-ers. When shippers have consistent demands totransport full loads of cargo back and forthbetween two points, high utilization rates can beachieved by dedicating trucks and drivers to ashipper and route. Most situations are not likethis, however. Individual shippers usually donot have demands for both legs of a round tripand shipments often do not fill trailers. In suchsituations, high capacity utilization requirestrucks to haul different shippers’ cargo on thesame run.

Capacity utilization thus depends largely onhow well individuals can identify and agglom-erate complementary demands onto individualtrucks. Higher quality matches increase capac-ity utilization by keeping trucks on the road andloaded more, and therefore raise truck drivers’productivity.4

It follows that understanding the link be-tween information and capacity utilization re-quires some understanding of the institutionsthat facilitate matching, individuals’ role withinthese institutions, and how informational im-provements lead to better matches both directlyand through organizational changes. This is thetopic of the next subsection.

A. Institutions and Market Clearing

Market clearing in trucking is unlike that intextbook economics models. It does not takeplace in centralized markets in which partici-

pants simply observe prices and decide howmuch capacity to sell to or buy from the market.Centralized markets have traditionally been un-important in trucking, in large part because ca-pacity and demands are highly differentiated interms of time, location, and equipment charac-teristics. Organizing centralized markets thatare so narrowly defined is costly relative to thebenefits such markets would generate.5 Instead,capacity and demand are matched in a highlydecentralized manner in which buyers, sellers,and intermediaries engage in costly search.These parties identify trading opportunities bycontacting each other directly rather thanthrough markets.

One way complementary demands are identi-fied is that shippers themselves search for othershippers with complementary demands. For ex-ample, a shipper with one-way demands betweenChicago and St. Louis will search for a shipperwith one-way demands between St. Louis andChicago. However, much of the time complemen-tary demands are identified by intermediaries,who add value by lowering search costs.

There are two main classes of intermediariesin trucking: for-hire carriers and brokers. Theydiffer in whether they own trucks; for-hire car-riers control truck fleets but brokers do not. Asexplained by George F. Baker and Hubbard(2003), truck ownership enhances intermediar-ies’ incentives to find complementary hauls be-cause it allows them to appropriate a greatershare of the surplus. Most intermediaries in theindustry are for-hire carriers. Shippers tend touse for-hire carriers when identifying comple-mentary demands is important, such as for longor less-than-truckload hauls, and private fleetswhen it is not.

Shippers and carriers sometimes contractahead for service. These contracts usually covera series of recurring hauls. Arrangements of thissort reduce search costs by eliminating the needto search for trading partners recurrently, buttend to lower the short-term efficiency of thematch between trucks and hauls.6 Hubbard

4 Links between productivity and the efficiency of themarket-clearing process exist in many markets, particularlythose like trucking in which supply and demand are highlydifferentiated. Labor markets are good examples.

5 Narrowly defined markets tend to be illiquid, andmatches in such markets may not improve much upon thoseachieved through decentralized matching.

6 They may also serve to lower hold-up risks, by pro-tecting relationship-specific informational investments. SeeHubbard (2001).

1330 THE AMERICAN ECONOMIC REVIEW SEPTEMBER 2003

(2001) shows that contracting becomes moreprevalent relative to simple spot arrangementsas local markets become thinner, particularlyfor long hauls. Shippers and carriers tend to relyon short-term arrangements when they use non-specialized equipment for hauls on thick ship-ping lanes, but longer-term arrangements whenthey use specialized equipment or operate onthin shipping lanes. Capacity and demands tendto be matched over longer horizons for haulsinvolving specialized equipment than nonspe-cialized equipment.

Both the presence of intermediaries and thefact that most intermediaries own trucks thuscan be interpreted as institutional responses tothe matching problem. The presence of interme-diaries lowers search costs; truck ownershipprovides intermediaries strong incentives to findgood matches. These institutional features in-crease capacity utilization and thus raise truckdrivers’ productivity.

B. Dispatch and Information

Operationally, the people most directly in-volved in matching capacity to demand are dis-patchers. Dispatchers assign trucks and driversto hauls. Dispatchers who manage shippers’ pri-vate fleets primarily assign trucks to their inter-nal customer’s hauls. Those who manage for-hire carriers’ fl eets assign trucks to externalcustomers’ (shippers’ ) hauls. Dispatchers some-times actively search for additional hauls whendoing so would increase capacity utilization,contacting shippers either directly or throughbrokers.7 For example, they try to find good“backhauls” (return trips).8 Such activities aremore common for dispatchers managing for-hire than private fleets. But they are not unusual

within private fleets, particularly in cases whereshippers use private fleets for long hauls.

Dispatchers work in a highly dynamic envi-ronment. Assignments and schedules are not setfar in advance, particularly when it is hard toforecast exactly when individual shippers willdemand service and exactly when particulartrucks will come free. In practice, dispatchersassign trucks and drivers to a series of hauls atthe beginning of the day or a shift. This is oftena provisional schedule. They then update sched-ules throughout the day as situations warrant,rearranging assignments in response to unex-pected delays and new service orders (some ofwhich they may have actively solicited to fillcapacity). Dispatchers who do this well increasethe productivity of the trucks and drivers theymanage.

Information is a critical input to dispatchers’decisions. In particular, knowing where trucksare and how full their trailers are lets dispatch-ers forecast better the time and location capacitywill become available. Better forecasts, in turn,allow them to allocate trucks across existingorders and market spare capacity more effi-ciently. They also can provide customers betterinformation about arrival times.

Information processing and communicationcapabilities are important as well, because theyhelp dispatchers make good decisions and redi-rect drivers. Most dispatchers use route-planningsoftware packages to help develop schedules.Many of these packages are relatively inexpen-sive and PC-based. Dispatchers commonly usethe software to draft schedules, which they thenrevise to account for factors not accounted forby the software.

Communicating with drivers has traditionallybeen difficult when trucks operate outside radiorange (about 25 miles). Dispatchers and driversrelied on a “check and call” system in whichdrivers stopped and called in every three to fourhours. During the 1990’s, declines in the priceof long-distance cellular communication ledmany dispatchers and drivers to abandon thissystem and communicate with cellular phones.This has significant advantages over the previ-ous system because it allows dispatchers to ini-tiate contact with distant drivers just like theydo with those close by. Dispatchers no longerhave to wait until drivers call in to give theminstructions, and drivers do not have to find a

7 At larger firms, different individuals assign trucks tohauls and solicit business. I will abstract from the fact thatindividuals specialize, assuming that they work closelyenough together so that they can be considered one decision-making unit.

8 In principle dispatchers could also identify other haulsalong the same route that would fill less-than-full trucks. Inpractice, trucks rarely pick up additional loads en routeunless such loads are arranged well in advance. Manyclasses of cargo (especially bulk, liquid, or refrigeratedcargo) cannot be mixed, and extra stops can increase theprobability of late arrivals, especially when they are notplanned in advance.


pay phone just to provide status reports and askif there are schedule changes. Using cell phonesalone has drawbacks, however. In particular,there remain significant coverage gaps, and in-formation about trucks’ location takes time tocollect and is neither verifiable nor in electron-ically processable form.

Thus, information costs have traditionallylowered capacity utilization in the industry be-cause difficulties in monitoring trucks’ locationand communicating schedule changes to drivershave made it hard for dispatchers to matchtrucks to hauls efficiently while trucks are onthe road. This has been particularly the casewhen trucks operate far from home and de-mands are not regular. Finding complementary“backhauls” is particularly important and com-municating with drivers sometimes difficultwhen hauls take trucks far from their base andirregularity makes it hard to arrange for back-hauls in advance.

C. On-Board Computers

Two classes of OBCs began to diffuse in thetrucking industry in the late 1980’s: trip record-ers and electronic vehicle management systems(EVMS).

Trip recorders monitor how drivers operatetrucks. They record when trucks were turned onand off, trucks’ speed over time, and incidentsof hard braking. Trip recorders collect data ontoa storage device. Dispatchers upload these dataonce drivers return to their base. The data triprecorders collect provide dispatchers verifiable in-formation regarding drivers’ activities, includingwhether they were speeding or took unauthorizedbreaks. Trip recorders also track how trucks’ en-gines perform; for example, they track fault codesthat result when engines work improperly. Thisinformation is useful to mechanics because ithelps them diagnose engine problems better.

Trip recorders are thus useful for improvingdrivers’ incentives and mechanics’ maintenancedecisions. They are not particularly useful forimproving dispatchers’ resource allocation de-cisions because they do not provide dispatchersinformation in a timely enough fashion.

EVMS are more advanced than trip record-ers. They contain all trip recorders’ capabilities.In addition, they record trucks’ geographic lo-cation (for example, using satellite tracking)

and provide a close-to-real-time data connec-tion between dispatchers and trucks. These ad-ditional capabilities help dispatchers makebetter scheduling decisions and communicatethem quickly to drivers. Knowing exactly wheretrucks are helps dispatchers allocate trucksacross existing service orders and market excesscapacity better. The communication link helpsthem notify drivers of schedule changes quicklyand effectively. From above, one would expectthese capabilities to be particularly importantwhen trucks haul goods long distances on irreg-ular schedules, since monitoring and communi-cation costs traditionally have had a largeimpact on dispatchers’ ability to match trucks tohauls efficiently in such situations.

The ways in which OBCs affect supply intrucking guide the empirical strategy. Concep-tually, capacity utilization reflects both loadedmiles during the periods that trucks are “ in use,”(i.e., away from their base) and the number ofperiods trucks are in use. From the discussionabove, improvements in drivers’ incentives anddispatchers’ resource allocation decisions pri-marily affect supply by increasing loaded milesduring the periods trucks are in use, for exampleby reducing the time during a run that trucks areidle or run empty. Because this is the marginwhere truck-level relationships between OBCuse and capacity utilization are most likely toreflect their effect on supply, this paper seeks toestimate how much OBCs affect loaded milesper period in use.

In contrast, truck-level relationships betweenOBC use and periods in use are unlikely toreflect OBCs’ effect on supply: monitoring im-provements generally do not affect how manyperiods trucks can potentially be “ in use.”9 Suchrelationships instead are likely to reflect the

9 There may be exceptions to this, though I do notbelieve these exceptions to be significant empirically. Con-sider a situation when scheduling a truck for an out-and-back run would lead it to be idle the following period whenit is at its base. If EVMS newly leads dispatchers to findhauls that would bring it back in a “ triangle,” thus avoidingan idle period at home, this would increase periods in usebut not loaded miles per period in use. I believe such aneffect to be minor; whether trucks are scheduled on “out-and-back” or more complicated routes depends far more onthe size of demands in different shipping lanes than dis-patchers’ ability to match trucks more precisely to thesedemands.


allocation of demand across trucks—theywould exist if shippers shift demands towardtrucking firms whose trucks have OBCs andaway from those whose trucks do not, or ifdispatchers utilize their best-equipped trucksmore than other trucks when capacity exceedsdemand—and might appear even if OBCs hadno supply-side effect on capacity utilization.Alternatively, such relationships may reflect re-verse causation: truck owners may adopt OBCsmore when they expect trucks to be used moreperiods. An important goal of the empiricalframework will be to disentangle relationshipsbetween OBC use and loaded miles per periodin use from those between OBC use and periodsin use. However, this will involve controllingfor rather than interpreting relationships be-tween OBC use and periods in use.10

There is an important economic distinctionbetween trip recorders and EVMS. Both classesof devices are useful for improving incentivesand maintenance decisions. EVMS, however,are also useful for improving resource alloca-tion decisions (“coordination” ).

This paper focuses primarily on the impact ofOBCs’ coordination-improving capabilities oncapacity utilization.11 There are two reasons forthis.

First, evidence from the trade press and plantvisits indicates that OBCs primarily affect supplythrough better dispatch, not through improve-ments in drivers’ incentives or maintenance deci-sions. One exception to this is when drivers’ jobsinvolve cargo handling as well as driving; somefirms attribute productivity gains to the ability totrack how long drivers spend at stops. Trucks canbe utilized more intensively when drivers load andunload cargo faster (see Baker and Hubbard,2003). OBC adoption also may have led somefirms to provide drivers stronger fuel economy-based incentives, and this may have led to pro-ductivity gains, but there is little indication thatthese increases are substantial.

Second, it is difficult to isolate the impact ofOBCs’ incentive-improving capabilities, becauseall OBCs have both incentive- and maintenance-improving capabilities.

II. Data

The data are from the U.S. Bureau of theCensus’ 1992 and 1997 Truck Inventory andUse Surveys (TIUS).12 The TIUS is a mail-outsurvey taken every five years as part of theCensus of Transportation. The Census takes arandom sample of trucks from vehicle registra-tion records, and sends their owners a question-naire that asks them about the characteristicsand use of their trucks.13 For example, ques-tions ask respondents their trucks’ make andmodel. Importantly for this study, the Surveyasks whether trucks have trip recorders orEVMS installed. It also asks many questionsabout how trucks were used during the previousyear, including such things as whether theywere owned by their driver, whether they oper-ated within a private or for-hire fleet, how farfrom home they generally operated, what kindof trailer was attached, what classes of productsthey carried, and the state in which they werebased. Although the TIUS contains observa-tions of a wide variety of truck types, all of theanalysis in this paper uses only observations oftruck-tractors, the front halves of tractor-trailercombinations.

The Survey also asks several questions thatelicit information regarding how intensively in-dividual trucks were utilized. Answers to thesequestions provide the variables used to evaluateproductivity. One question asks how manymiles the truck was driven during the previousyear. Other questions ask what fraction of milesthe truck was driven without a trailer, and whatfraction of miles it was driven empty. Combined

10 To the extent that relationships between OBC use andperiods in use do reflect that OBCs increase periods in use,focusing only on how OBCs affect loaded miles per periodin use would understate how much OBCs affect capacityutilization.

11 Other papers [Baker and Hubbard (2000, 2003)] haveexamined the organizational implications of OBCs’ incentive-improving capabilities.

12 The 1997 Survey is actually called the Vehicle Inven-tory and Use Survey. See U.S. Bureau of the Census (1995,2000) and Hubbard (2000) for more details about theseSurveys.

13 Since draws are taken from vehicle identificationnumbers, sampling is randomized across trucks, not firms orindustry sectors. The trucks in my 1992 sample make upabout 3 percent of truck-tractors registered in the UnitedStates; sampling rates were about one-third lower in 1997than 1992 for budgetary reasons.


with the number of miles the truck was driven,answers to these questions indicate the numberof miles the truck was driven with cargo(“ loaded miles” ). The Survey also asks theweight of the truck when empty and the averageweight of the truck plus cargo during a typicalhaul in the previous year. The difference be-tween these figures is the average weight of thecargo the truck hauled (“cargo weight” ). Multi-plying loaded miles by cargo weight and divid-ing by 2,000 gives an estimate of the truck’soutput during the previous year in ton-miles.Finally, the Survey asks owners how manyweeks out of the year trucks were in use, de-fined as the number of weeks in which a truck isever used to haul cargo. As I discuss below, thisis an important variable in the analysis. Its ab-sence from previous Surveys is the reason I useonly the 1992 and 1997 Surveys.

Responses to these questions likely over-state trucks’ output and capacity utilizationsomewhat, although probably in a similarfashion from year to year. Cargo weight isprobably overstated because respondentslikely report cargo weight when trucks leaveterminals, which is not the average amount ofcargo in trucks’ trailers while loaded whentrucks deliver to multiple points.14 Respon-dents likely understate empty miles, particu-larly when trucks haul trailers for whichbackhauls are unlikely such as auto trailers.This is because respondents who do not try tofind backhauls may not include backhaul ca-pacity in the denominator of this fraction. Butthis bias works against finding relationshipsbetween OBC adoption and capacity utiliza-tion increases if adoption leads firms to re-consider what they think of as unusedcapacity: for example, if it leads them tonewly consider empty backhauls as emptymiles.

The Survey therefore provides detailed infor-mation about production at the individual trucklevel. This level of disaggregation is rare, andprovides a significant advantage in studying

technology adoption, organizational structure,and productivity issues.15 The Survey does not,however, allow one to identify trucks’ owners.It is therefore impossible to determine the for-hire or private fleet in which individual trucksoperated. Although one can aggregate up to theindustry or industry-segment level, the data can-not be used to investigate productivity at thefirm level.

Finally, it is important to recognize that theTIUS does not collect panel data; rather, it is aseries of repeated cross sections. One does notobserve exactly the same trucks or hauls fromyear to year. This limits the extent to which Ican exploit the data’s time dimension. I haveexplored doing so in a way analogous to myother work (Baker and Hubbard, 2000, 2003):aggregating the data up to narrowly definedmarket segments (for example, state-productclass-trailer type-distance combinations) ineach year, and relating segment-level changesin OBC use to segment-level changes in averageloaded miles per period. But cross-sectional pat-terns in the data indicated that this method wasvery likely to produce biased estimates of thetrue relationships between OBC use and capac-ity utilization. I estimated cross-sectional rela-tionships between OBC use and loaded milesper period at the segment and truck level andfound that the segment-level relationships weremuch stronger.16 Because the segment-level re-lationships do not track the micro-level relation-ships in the cross section, I concluded that theywere unlikely to do so in the time series, and infact were likely to bias estimates of OBCs’effect on capacity utilization upward.

The following subsection introduces the dataand shows some broad patterns that indicaterelationships between changes in capacity utili-zation measures and changes in OBC use be-tween 1992 and 1997. However, the main

14 Aggregate mileage estimates for the entire U.S. truck-ing fleet from the TIUS are consistent with those from othersources, but ton-mile estimates are not. This indicates thatthe cargo weight data in the TIUS are not very reliable. Itherefore use loaded miles rather than ton-miles as my mainoutput measure in the analysis below.

15 The manufacturing equivalent perhaps would be tohave data at the level of the production line rather than theestablishment or firm.

16 The specifications are analogous to those reported inthe first column of Table 3, and include controls for dis-tance, trailer type, and other haul characteristics. The segment-level point estimates using the 1997 data suggest that triprecorder adoption raises loaded miles per period by 20–25percent, depending on how narrowly segments are defined;in contrast, the truck-level estimates presented below sug-gest that it raises loaded miles per period by 2 percent.


empirical evidence in this paper will exploitcross-sectional rather than time variation in thedata.

A. Simple Patterns

Table 1 presents simple trends. The top panelindicates that capacity utilization increased be-tween 1992 and 1997. On average, miles pertruck increased by 7.5 percent and loaded milesincreased by 10.1 percent. Although the cargoweight data in the TIUS are not very reliable,there is no indication that average cargo weightdecreased during this time. Reports from thesedata indicate that it increased by 2.5 percent,leading to a 12.5 percent increase in ton-milesper truck. OBC use increased during this periodas well. The fraction of trucks with a trip re-corder installed increased slightly from 7.8 per-cent to 8.4 percent, while the fraction with anEVMS installed more than doubled from 11.1percent to 24.9 percent.

The bottom panel reports similar figures, aver-aging only over trucks that were in use at least 48weeks out of the year. Comparing trends in thesefigures to those in the top panel provides someevidence regarding the extent to which increasesin capacity utilization were due to increases in thenumber of periods in use rather than increases inhow intensively trucks were used conditional onperiods in use. Loaded miles increased by 8.3percent within this subsample—somewhat lessthan the 10.1 percent increase within the full sam-ple, but still a large increase. These figures do not

suggest that increases in capacity utilization dur-ing this period were entirely due to the fact thattrucks were used more weeks out of the year in1997 than 1992. Capacity utilization increasedduring this time even among the most intensivelyused trucks. OBC use was high for these trucks aswell.

Figure 1 provides further evidence. This plotsaverage weeks in use, by truck age, for the 1992and 1997 samples. If increases in loaded milesreflect increases in the utilization of infre-quently used trucks, older trucks should be usedmore weeks in 1997 than 1992. Figure 1 indi-cates that while weeks in use declines steadilywith truck age in both years, the plots track eachother very closely.17 There is no evidence thatolder trucks were used more weeks per year in1997 than 1992.

Figure 2 relates loaded miles per week in useto net EVMS adoption. The lines plot loadedmiles per week in use as a function of age; thebars report the share of n-year-old trucks withEVMS in 1997, less the share of n-year-oldtrucks with EVMS in 1992. There are threeimportant facts. First, old trucks are used lessintensively than new ones, even conditional onweeks in use. Second, the gap between 1997and 1992 trucks is greater when comparing newtrucks than old trucks. Once again the greatest

17 The low figure for brand-new trucks reflects that manywere put into service in the middle of the survey year.

TABLE 1—TRUCK UTILIZATION AND OBC USE—1992, 1997

MilesLoadedmiles

Fractionw/load

Cargoweight Ton-miles

Triprecorder EVMS N

Panel A: All Trucks1992 65,451 58,559 0.882 38,190 1,178 0.078 0.111 36,0821997 70,351 64,500 0.904 39,223 1,325 0.084 0.249 23,183

Change (percent) 7.49 10.15 2.49 2.70 12.48 7.69 124.32Panel B: Trucks in Use � 48 Weeks

1992 77,764 69,993 0.893 37,890 1,399 0.100 0.152 18,6831997 82,488 75,836 0.915 39,602 1,592 0.093 0.301 11,376

Change (percent) 6.07 8.35 2.46 4.52 13.80 �7.00 98.03

Notes: Miles is the average number of miles trucks were operated. Loaded miles is the average number of miles truckswere operated and loaded. Fraction with load is loaded miles/miles, averaged across trucks. Cargo weight is the averageweight of the cargo trucks hauled when loaded. Ton-miles is cargo weight multiplied by loaded miles, averaged acrosstrucks. Trip recorder is the share of trucks with a trip recorder installed. EVMS is the share of trucks with an EVMSinstalled.


increase in capacity utilization is for the trucksthat are already utilized intensively. Third, thegap between the 1997 and 1992 trucks is widestwhere net adoption is highest—for one- to five-year-old trucks. 1992–1996 model year truckshad much higher EVMS use rates in 1997 than1987–1991 model year trucks did in 1992. Ca-pacity utilization rates also appear to increasemore for trucks in this range than younger orolder trucks.

Combined, these tables provide evidenceconsistent with the hypothesis that EVMS adop-tion contributed to increases in capacity utiliza-tion. Capacity utilization increased the most foralready intensively used trucks, and trucks forwhich EVMS tended to be adopted most had thegreatest increases in capacity utilization.

Furthermore, additional evidence indicatesthat capacity utilization increases during thistime also represent increases in labor produc-tivity. Increases in loaded miles per truck wouldnot reflect increases in labor productivity if theratio between drivers and trucks changed, aswould be the case if firms were using trucks (butnot drivers) for double shifts more in 1997 than1992. However, data from the October CPSindicates that the number of truck drivers in-creased by 26.8 percent between 1992 and1997; the 1997 VIUS indicates that the numberof heavy duty trucks increased by 25.7 percent.While this evidence is not necessarily conclu-sive, since the CPS does not distinguish be-tween drivers of heavy- and lighter-duty trucks,these figures do not indicate that there were anyimportant changes in the driver-truck ratio dur-ing this time.

B. Periods and Weeks

As noted above, the goal of this paper is toestimate how much OBCs have increasedloaded miles per period in use. An empiricalproblem arises because while one would likeinformation on the number of periods (e.g.,hours or shifts) trucks are in use, the data in-stead contain information on the number ofweeks trucks are in use. If the relationship be-tween periods in use and weeks in use wereone-to-one, the difference between weeks andperiods in use would just be a difference in unitsand normalizing loaded miles by weeks in usewould amount to the same thing as normalizingby periods in use. But this need not be the casebecause trucks are counted as “ in use” during aweek regardless of whether they are used forone or many shifts. In fact, an increase in thenumber of periods could result in no change inthe number of weeks, for example if it wereaccomplished by utilizing trucks for more shiftsduring the weeks they were already used.

Although a zero elasticity between periods inuse and weeks in use would be an extreme case,the example illuminates a general point: therelationship between periods in use and weeksin use is unknown, and one must estimate it inorder to utilize information on weeks in use tocontrol for periods in use. If part of what hap-pens as periods in use increase is that trucks areused for more periods during weeks they arealready in use, differences in weeks in usewould understate differences in periods in use,and simply normalizing loaded miles by weeks

FIGURE 1. AVERAGE WEEKS IN USE FIGURE 2. LOADED MILES PER WEEK, NET EVMSADOPTION


in use would not completely correct for differ-ences in periods in use.

One of the patterns in Figure 2 manifests this.All else equal, loaded miles per period in useshould not vary across vintages: conditional onbeing in use during a period, old trucks can beused about as intensively as new ones. Thus, thefact that loaded miles per week is considerablylower for old trucks than new ones implies thatsimply normalizing loaded miles by how manyweeks trucks are in use does not completelycontrol for differences in how many periodstrucks are in use.18 Much of the next sectionfocuses on developing a more sophisticated wayto utilize data on weeks in use to control fordifferences in periods in use.

III. Empirical Framework

Let yit equal loaded miles for truck i in periodt, where period corresponds to a day or shift.Let kit be a dummy variable that equals one iftruck i is in use in period t and zero otherwise.Since yit � 0 during periods truck i is not inuse, one can write yi

1, truck i’s loaded milesover the course of T periods, as:

(1) yi1 � �

t � 1

T

yit kit .

Assume for simplicity that loaded miles perperiod in use for truck i is constant acrossperiods: trucks are used in similar ways fromperiod to period, conditional on being in use.19

Let si equal the share of periods that truck i is inuse. Then one can rewrite yi

1 as:

(2) yi1 � yi si T

where yi is loaded miles for truck i per period in use.yi is influenced by many factors, including

the characteristics of the hauls for which thetruck is used, the characteristics of the firm findinghauls for the truck, and the informational envi-ronment. Haul characteristics matter becausethey affect how much time trucks spend at stopsbeing loaded and unloaded and how fast theytravel when moving; for example, yi tends to behigher for trucks used for long than short haulsbecause such trucks spend less time at loadingdocks or on congested city streets. Firm char-acteristics matter if some firms have better in-formation about demand than others and thislets them find better backhauls for the truck.Whether trucks have OBCs affects the infor-mational environment, and can affect yi by im-proving drivers’ incentives or by improvingdispatchers’ knowledge and communicationcapabilities. The latter may facilitate bettermatches between trucks and hauls. I specifyln yi as:

(3) ln yi � Xi� � �1Di � �1i

where Xi includes observable haul and firmcharacteristics that affect loaded miles per pe-riod in use and Di is a vector of dummies thatreflect whether and what kind of OBCs areinstalled on the truck. �1i captures the effect ofunobserved haul and firm characteristics. Tosimplify exposition, assume for now that �1does not vary.

I next discuss si. I assume that si is related todemand, truck, and firm characteristics by thefollowing reduced-form equation.

(4) ln si � Zi� � �2Di � �2i .

Zi includes observable variables that are corre-lated with the share of periods trucks are in use.These may include variables that are also in Xi.One variable that I will assume to be part of Zibut not Xi is truck age: trucks’ age may becorrelated with the share of periods they areused (perhaps because dispatchers put theirnewest trucks in use when capacity exceedsdemand) but does not affect how intensivelythey can be used, given that they are in useduring a period.20 Di is as above. �2 captures

18 Although Figure 2 shows unconditional differences,these differences remain economically and statistically sig-nificant when including controls for how trucks are used.

19 There is some evidence on this in the data. For exam-ple, the Survey asks owners to report individual trucks’share of miles by haul length, product class, and governanceform. Though the quality of the share data may not be good,these data strongly suggest that most trucks are used inconsistent ways from period to period.

20 This restriction produces conservative estimates ofOBCs’ effect on loaded miles per period in use; see below.


correlations between OBC use and the share ofperiods truck i is in use. As described above, suchcorrelations could arise for several reasons.OBC use may lead the share of periods to behigher because shippers may reallocate demandtoward trucking firms with OBC-equippedtrucks, or dispatchers may put OBC-equippedtrucks in service and idle others when capacityexceeds demand. Alternatively, correlationsmay arise because of reverse causation: OBCsare more valuable when trucks are in use more.�2i is a residual, and represents relationshipsbetween si and unobserved factors that are or-thogonal to both Zi and Di. Since this is areduced form, by construction, E(�2i�Zi, Di) � 0.

Taking logs of equation (2) and substitutingin equations (3) and (4), I obtain:

(5)

ln yi1 � Xi� � Zi� � ��1 � �2 �Di � �1i � �2i .

This equation relates loaded miles to OBCuse.21 The empirical goal is to estimate OBCs’effect on loaded miles per period in use, �1.However, as this equation shows, even if theorthogonality condition E(�1i�Di) � 0 holds,least-squares estimates of loaded miles on OBCuse reflect both OBCs’ effect on capacity utili-zation and correlations between OBC use andthe share of periods trucks are in use. I nextdiscuss a method to estimate �1 separately from�2 that exploits the fact that the data containinformation on the share of weeks trucks are inuse. A key step in this method is identifying � lnyi

2/� ln si , the elasticity between weeks in useand periods in use. Thereafter I discuss theorthogonality condition, and interpretations ofthe estimates when OBCs’ effect on yi differsacross hauls.

Let yi2 equal the share of weeks truck i is in

use over the course of T periods, and specify:

(6) yi2 � si

�hi .

� is the elasticity between the share of weeks inuse and the share of periods in use. If 0 � � �

1, the relationship between yi2 and si is concave;

trucks that are used a higher fraction of periodsare used more weeks per year, but at a decreas-ing rate.22 hi includes factors that affect thenumber of weeks in use, conditional on thenumber of periods in use. hi would be higherwhen demands for the truck are more cyclical:for example, trucks that primarily haul agricul-tural goods tend to be used a low number ofweeks relative to periods because demandcomes in spurts. Assuming that ln hi � Wi� ��3i , I therefore have the following:

(7)

ln yi1 � Xi� � Zi� � ��1 � �2 �Di � �1i � �2i

ln yi2 � Wi� � �Zi� � ��2Di � �3i � ��2i .

�1 and �2 are now separately identified. Thelogic is that if trucks with OBCs are used moreweeks than those without them, this should re-flect differences in the number of periods truckswith and without OBCs are used. One can thususe the relationship between weeks in use andOBC use to back out how much relationshipsbetween loaded miles and OBC use reflect dif-ferences in loaded miles per period in use. Do-ing so is simple if � � 1: subtracting the secondequation from the first differences out �2Di. Butas the discussion above emphasizes, unitaryelasticity between periods and weeks in use isunlikely; one must instead estimate �. This re-quires having at least one variable that is relatedto the share of periods trucks operate but doesnot affect loaded miles per period. I assume thisto be the case for truck vintage, and estimate �from the ratio of the relationships between vin-tage and the two dependent variables.23

My identification strategy implies the follow-

21 To simplify the exposition, I have dropped the termln T from the right-hand side of this equation. This iswithout loss of generality, since ln T is not separatelyidentified from �0 , the constant term in the vector �.

22 Concavity would be an important property of morestructurally derived expressions of the relationship betweenperiods and weeks in use. The reduced-form specificationused here captures this feature in a parsimonious way, andproduces a straightforward set of estimating equations fromwhich it is clear how each of the parameters is identified bythe data. I discuss identification further below.

23 Although the estimation procedure below allows theerror terms in the two equations to covary, this covariancedoes not help me identify � unless I were to put furtherrestrictions on the variance of �1 , �2 , or �3.


ing. Suppose trucks differ only in their vintageand whether they have OBCs. Suppose youngtrucks are used 10 percent more weeks, but have20 percent more loaded miles, than old ones.Suppose trucks with OBCs are used 10 percentmore weeks than those without them, but have25 percent more loaded miles. Then �� 0.1,� � 0.2, ��2 � 0.1, and �1 � �2 � 0.25.Solving for �1 , the estimates would indicatethat trucks with OBCs have 5 percent higherloaded miles per period in use than those with-out them.

An important identifying assumption is thatcorrelations between truck vintage and loadedmiles reflect only differences in the number ofperiods in use, not differences in loaded milesper period in use. This assumption tends toproduce conservative estimates of the relation-ship between OBC use and loaded miles perperiod in use: if new trucks can be used moremiles per period than old trucks, my estimate of�1 would be downward biased. To see this,consider the example above, but suppose newtrucks can be used 5 percent more miles perperiod than old trucks. If they have 20 percentmore loaded miles, this implies that they areused 15 percent, not 20 percent, more periodsthan old trucks: the true value of � is 0.15, not0.20. The equations �� 0.1, ��2 � 0.1, and�1 � �2 � 0.25 would then imply that the truevalues of the rest of parameters are � � 0.67,�2 � 0.15, and �1 � 0.10. My estimate of �1would indicate that loaded miles per period inuse was 5 percent higher for trucks with OBCsthan those without them when it was really 10percent higher.

A. Causality

Interpreting �1 as OBCs’ impact on loadedmiles per period in use requires the orthogonal-ity condition E(�1i�Di) � 0 to hold: OBC use isindependent of unobserved haul and firm char-acteristics that affect loaded miles per period inuse. Note that the relevant issue does not con-cern whether adoption is higher when trucks areused more periods—this is a reason why nor-malizing loaded miles by periods in use is im-portant. If within firms, OBCs are installed ontrucks that are expected to be used heavily, orif firms that are able to keep trucks out on

the road more of the time also adopt OBCsmore, this is picked up by �2.24 The relevantissue is narrower. It concerns whether biasesarise because adoption is greater on trucks that,absent OBCs, would accumulate more loadedmiles during the periods they spend out on theroad.

Unobserved haul characteristics in �1i in-clude factors that affect how much time trucksspend at loading docks and their speed while onthe move, conditional on Xi.

One potential violation of the orthogonalitycondition arises because drivers’ jobs differacross hauls in unobserved ways, and this couldboth drive differences in the time trucks spendat loading docks and the extent to which differ-ent classes of OBCs are used. Baker and Hub-bard (2003) report that hauls differ in whetherdrivers have nondriving service responsibilitiessuch as sorting and shelving cargo upon deliv-ery. Loaded miles per period in use tend to belower when drivers have such responsibilitiesbecause stops take longer: drivers do not merelydrop off cargo and leave. Furthermore, thereturns to adoption may differ with this un-observed haul characteristic. OBCs’ incentive-improving capabilities would be more valuableif monitoring’s benefits are greater in multitask-ing environments; their coordination-improvingcapabilities would be less valuable if givingdrivers service responsibilities interferes withdispatchers’ ability to identify and implementgood matches by making it more difficult fordispatchers to forecast when trucks will comefree, even when they know where trucks are. Ifso, one would expect trip recorder adoption tobe high, and EVMS adoption to be low relativeto trip recorder adoption, in circumstanceswhere loaded miles per period is low because ofunobserved service. This would bias estimatesof trip recorders’ effect downward and EVMS’effect relative to trip recorders’ upward.

24 However, it turns out that firm effects such as thisprobably are not empirically important once one controlsfor truck age and haul characteristics. The estimates of �2

below will not show that trucks with advanced OBCs areused more periods than trucks without OBCs. This sug-gests that unobserved firm characteristics that broadlyaffect truck utilization are not strongly correlated withEVMS adoption, and is a reason I focus more below onproblems associated with unobserved haul than firmcharacteristics.


I investigate the extent to which unobservedservice responsibilities are biasing the results inthe following way. Drivers’ responsibilities dif-fer systematically between private and for-hirecarriage; on average, they have much greaterservice responsibilities within private fleets.25 Ifunobserved differences in drivers’ responsibili-ties bias estimates of OBCs’ impact on loadedmiles per period in use, one would expect thatomitting the private fleet dummy would do so aswell. Finding instead that �1 does not changewhen excluding the private fleet dummy is ev-idence that the bias due to unobserved servicedifferences is likely small.

Another possibility also concerns unobserveddifferences in time spent picking up and deliver-ing cargo. Shippers and receivers differ in thesophistication with which they handle goods, andthis is not directly observed in the data. Supposesome receivers of goods have better logistics prac-tices than others, and sophisticated receivers areboth able to unload trucks faster because of betterhandling methods and value using carriers withOBC-equipped trucks.26 This would induce a spu-rious correlation between loaded miles per periodin use and OBC use, even if OBCs did not causeloaded miles per period in use to increase. I ex-amine this possibility in the following manner.Receivers’ organizational sophistication varieswith the products they receive—it tends to behigher for product classes that are delivered tomanufacturers or warehouse facilities than thosedelivered to raw input processors or retail outlets.I therefore examine whether the coefficientschange when including a set of dummy variablesthat control for the products trucks generally haul;finding that they do not suggests that relation-ships between OBC use and loaded miles perperiod in use do not reflect spurious correlationsdue to unobserved differences in logisticalsophistication.

Unobserved haul characteristics also in-clude factors that affect trucks’ speed whileon the road. Thus, another reason that theorthogonality condition might not hold is thatboth loaded miles per period in use and OBCadoption may be greater when trucks oper-ate in less congested areas. Loaded milesper period in use might be greater becauseof fewer traffic problems; EVMS adoptionmight be greater because satellite-basedcommunication links are more valuable whencell phone coverage is spottier and payphones scarcer. Once again, I investigate thisthrough a sort of robustness check. Conges-tion varies substantially geographically:greater in the East than the West, for example.If there are spurious correlation problemsrelated to cross-sectional differences in con-gestion, estimates of �1 should change whenI include additional controls for where trucksare based. I explore this by comparing thecoefficients when including and excludingdummies that indicate the state in whichtrucks are based from Xi. Finding that �1is robust to whether state dummies are in-cluded is evidence that any biases inducedby such spurious correlations are probablysmall.

�1i also includes unobserved firm charac-teristics. These reflect, for example, how wellthe truck’ s owner (or an intermediary theowner uses) can find backhauls for the truckabsent OBCs, holding constant the owner’ sability to keep trucks “ in use.” Such factorswould cause E(�1i�Di) � 0 to fail undercircumstances such as the following. Con-sider firms that are similar in their ability tokeep trucks “ in use”— demands for outbound“ fronthauls” are the same— but differ in theirknowledge of backhaul demands. If firms thatcan match trucks to hauls better absent OBCsare also more likely to adopt OBCs, then �1will overstate OBCs’ effect on loaded milesper period in use.

This alternative hypothesis is difficult to ex-amine using methods such as above because thedata contain little information about the firmthat owns the truck. While I cannot completelyrule out this alternative hypothesis, the resultsbelow shed some light on its empirical impor-tance. In particular, I will find no evidence of arelationship between OBC use and loaded miles

25 Industry publications commonly remark on this; forexample Standard and Poor’s (1995) states that using pri-vate fleets is valuable because of “overall superior service tocustomers.” Baker and Hubbard (2003) propose that ship-pers’ make-or-buy decision is complementary to decisionsregarding whether drivers have service responsibilities, andfind evidence in favor of this proposition.

26 Hubbard (2000) provides evidence that OBC use isgreater on trucks that haul products with high sales-inventoryratios than low ones, suggesting that logistical sophisticationand OBC use are related.


per period in use in 1992. There is thus noevidence that early adopters were firms thatwere particularly effective in finding backhauls,conditional on the number of periods theirtrucks were in use; if they were, one wouldexpect to observe a positive relationship be-tween OBC use and loaded miles per period inuse in 1992 even if OBCs had no effect oncapacity utilization. While I will find a posi-tive relationship between OBC use and loadedmiles per period in use in 1997, it is unlikelyto reflect that firms that find backhauls effi-ciently absent OBCs are systematically morelikely to adopt OBCs because if it did, onewould expect such a relationship to show up in1992 as well.

In the results section, I will therefore interpretthe estimates under the assumption that OBCuse is independent of firms’ unobserved abilityto match trucks to hauls absent OBCs, with thecaveat that I cannot rule out interpretationswhere this assumption holds in 1992 but not1997.27

B. Heterogeneity in OBCs’ Effect

As noted above, equation (3) assumes awayunobserved heterogeneity in OBCs’ impact oncapacity utilization. In fact, OBCs are likely toaffect yi differently across hauls and be used themost where they have the greatest impact.28 Amore general specification is:

(8) ln yi � Xi� � �1iDi � �1i

� Xi� � ��1 � �i �Di � �1i .

Here the marginal impact of OBCs on capacityutilization varies with omitted factors. Standardselection issues arise. If E(�1i�Di) � 0, least-squares estimates of the coefficient on Di pro-duce the following quantity:

(9) �̂1,ls � �1 � E��i�Di � 1�.

This coefficient captures the average effect ofOBCs among adopters—the average effect oftreatment on the treated. Least-squares esti-mates thus provide the quantities of interest inthis paper: OBCs’ realized impact on capacityutilization among the trucks for which they havebeen adopted.

Along with results from basic specifica-tions, below I will report results from speci-fications that interact the OBC dummies withvariables that I observe in the data; theseprovide estimates of the average returnsamong adopters within haul characteristic-governance form segments. From equation(9), the average returns among adopterswithin a segment does not just reflect themean return to adoption within the segment,but also other moments of the distribution ofreturns. The average returns among adopterswithin a segment could be high even if themean return to adoption is low if there is alarge upper tail. I therefore cannot use theseestimates to test propositions about cross-segment differences in the average returns toadoption; finding that the estimates are higherin long- than short-haul segments would notnecessarily imply that the average returns toadoption increase with haul length. Rather, Iwill combine these estimates with data onadoption and the distribution of trucks acrosssegments to produce estimates of how theoverall returns from OBC adoption are dis-tributed across segments of the industry.

Although it is not the focus of this paper,the results from the interaction specificationswill shed some light on the question: howmuch would OBC use increase capacity uti-lization for the average truck? I will find noevidence that the average capacity utilizationbenefits among adopters are positive withinsome segments. The fact that these benefitsappear small or nonexistent among manyadopters suggests that they were probably verysmall among nonadopters as well, especially

27 Such interpretations would involve a nonmonotonicrelationship between firms’ unobserved ability to find back-hauls absent OBCs and their speed of adoption, since theywould require firms adopting by 1992, between 1993 and1997, and after 1997 to be average, above average, andbelow average, respectively. There is no indication fromtrade press accounts that early adopters were worse on thisdimension than later ones. Indeed, adoption between 1993and 1997 sometimes involved exactly the same firms as inthe earlier period; these firms adopted OBCs for part of theirfleet during the early period, then more of their fleet in thelater period.

28 See Hubbard (2000) for a detailed analysis of adoptionpatterns.


inframarginal nonadopters. Only about 35 per-cent of trucks had OBCs as of 1997; hence,OBCs’ capacity utilization benefits were prob-ably close to zero for the average truck at thistime.

IV. Results

A. Simple Cross-Sectional Regressions

Table 2 presents results from univariatecross-sectional regressions that take the form ofequation (5).29 I present these as preliminary tothe main results below. The dependent variableis loaded miles. The vector Xi contains a set ofdummy variables that indicate how far fromhome the truck operated, a set of dummies thatindicate what class of trailer was commonlyattached to the truck, and dummies that indicatewhether trucks were part of private fleets, usedfor contract carriage, were driven by owner-operators (and if so whether they were operatingunder long-term arrangements with largertrucking firms), and whether trucks were used tohaul “ less-than-truckload” shipments. The vec-tor Zi consists of a vector of dummy variablesthat characterize the truck’s vintage. The coef-ficients of interest are those on OBC andEVMS, which correspond to (�1 � �2). OBC isthe coefficient on a dummy that equals one ifthe truck had either a trip recorder or EVMSinstalled and zero otherwise; EVMS is that on adummy that equals one if the truck had anEVMS installed and zero otherwise. OBC re-flects the correlation between trip recorder useand loaded miles; EVMS reflects the differencein loaded miles for trucks with EVMS andtrucks with trip recorders. Thus, OBC picks uprelationships between loaded miles and OBCs’incentive- and maintenance-improving capa-bilities and EVMS picks up those betweenloaded miles and OBCs’ coordination-improvingcapabilities.

The upper panel contains results using the1992 data. The specification in the first columnrestricts all coefficients other than OBC andEVMS to zero, the second estimates the Xi

coefficients but not the Zi coefficients, and thethird estimates all of the coefficients. From thefirst column, trucks with trip recorders had 45percent more loaded miles than those withoutany IT. Trucks with EVMS had about 29 per-cent more than those with trip recorders. Theseestimates decrease sharply when including thecontrols, and the R2 increases from 0.04 to 0.48.OBC remains positive and significant, and indi-cates that controlling for trucks’ age and haulcharacteristics, trucks with trip recorders had13.3 percent more loaded miles than those with-out them. Trucks with EVMS had 7.8 percentfewer loaded miles than those with triprecorders.

The lower panel reports analogous estimatesusing the 1997 data. The general patterns aresimilar to the 1992 data. The estimates in thethird column imply that trucks with trip record-ers had 7.6 percent more loaded miles thanthose without them, and that there is no signif-icant difference in loaded miles between truckswith trip recorders and trucks with EVMS.

Estimates from these simple specificationsindicate relationships between OBC use andloaded miles, but do not distinguish between

29 The sample size is lower here than in the previoustables because some observations have missing values forweeks in use.

TABLE 2—OBCS AND LOADED MILES: 1992 AND 1997CROSS-SECTIONAL REGRESSIONS

Dependent variable: ln(loaded miles)

Panel A: 1992 SampleOBC 0.450* 0.203* 0.133*

(0.025) (0.022) (0.021)EVMS 0.291* �0.072* �0.078*

(0.030) (0.028) (0.026)Controls? None X vector X, Z vectorsR2 0.044 0.408 0.476N � 35,766

Panel B: 1997 SampleOBC 0.643* 0.207* 0.076*

(0.028) (0.024) (0.024)EVMS 0.189* 0.098* 0.024

(0.028) (0.025) (0.025)Controls? None X vector X, Z vectorsR2 0.102 0.440 0.495N � 22,206

Notes: X vector includes distance dummies, trailer dum-mies, private carriage, contract carriage, independent own-er-operator, subcontracted owner-operator, LTL, and LTL �short-haul dummies. Z vector includes truck vintage dum-mies. Eicker-White standard errors are in parentheses.

* Significantly different from 0 at the 5-percent level.


differences in loaded miles per period in use anddifferences in the number of periods trucks areused. The next subsection reports estimatesfrom multivariate regressions that do so.

B. Multivariate Regressions

Table 3 presents GLS estimates of (7) usingthe 1997 data.30 Xi is the same as above. Ziincludes all of the variables in Xi , plus a full setof truck vintage dummies: if newer trucks areused more weeks than older trucks, this reflectsdispatchers’ (or the market’s) choice of whichtrucks to use when demand is low.31 Wi in-cludes other variables that correlate with thecyclicality of individual trucks’ use: dummies

that indicate whether the truck was primarilyused to haul fresh farm products and live ani-mals. Trucks used to haul these goods are usedfar fewer weeks than other goods.32

The first column contains results from thisbase specification. OBC1 and EVMS1 are esti-mates of �1 , and reflect relationships betweenOBC use and yi , loaded miles per period in use.OBC1 is small and not statistically significantlydifferent from zero; this estimate provides noevidence that OBCs’ incentive-improving capa-bilities affect loaded miles per period in use.EVMS1 is positive and significant, suggestingthat OBCs’ coordination-improving capabilitiesdo so. The point estimate indicates that, control-ling for differences in the number of periodsdifferently equipped trucks are used, trucks withEVMS have 10.4 percent more loaded milesthan those with trip recorders. Assuming for now30 These utilize information from least-squares residuals

to produce an estimate of the variance-covariance matrix ofthe errors in the two equations. While using this as aweighting matrix for systems estimation increases the effi-ciency of the estimates, in this case doing so has little effecton either the estimates or the standard errors.

31 Estimates of �1 are robust to excluding variables in Xi

from Zi , in large part doing so does not change whichvariables are included as controls in the loaded milesequation. See Table A1 for the full set of coefficients fromthe specifications reported in the first column of Tables 3and 4.

32 Preliminary regressions indicated that these variableswere correlated with number of weeks in use. The fact thatthese variables have explanatory power at all is interesting,considering that the unit of observation is a truck-tractor,and truck-tractors are highly mobile and are not specific tofirms, trailers, or products outside of the short run. That haulcharacteristics are significant is evidence of frictions inshifting trucks across uses when demand is low for whatthey generally haul.

TABLE 3—OBCS AND LOADED MILES PER PERIOD IN USE:1997 COEFFICIENT ESTIMATES OF EQUATION (7)—MULTIVARIATE REGRESSIONS

Dependent variables: ln(loaded miles), ln(weeks in use)

OBC1 0.023 0.019 0.027 0.024 0.032(0.029) (0.030) (0.029) (0.029) (0.029)

EVMS1 0.104* 0.105* 0.094* 0.102* 0.092*(0.029) (0.030) (0.029) (0.029) (0.029)

OBC2 0.056 0.059 0.062 0.047 0.046(0.029) (0.031) (0.029) (0.029) (0.028)

EVMS2 �0.078* �0.079* �0.078* �0.071* �0.071*(0.029) (0.031) (0.029) (0.029) (0.029)

Lambda (�) 0.406* 0.387* 0.409* 0.410* 0.412*(0.016) (0.015) (0.016) (0.016) (0.016)

R2

Loaded miles equation 0.497 0.497 0.501 0.501 0.505Weeks in use equation 0.203 0.203 0.203 0.204 0.204

Omits private carriage dummy from X? N Y N N NIncludes state dummies in X? N N Y N YIncludes product dummies in X? N N N Y YN � 22,206

Notes: OBC1 and EVMS1 measure relationships between OBC use and trucks’ loaded miles per period in use. OBC2 andEVMS2 measure relationships between OBC use and the number of periods trucks are in use. Lambda is the estimatedelasticity between number of periods in use and number of weeks in use. Eicker-White standard errors are in parentheses.



that the orthogonality condition E(�1i�Di) � 0holds, this is an estimate of the average impactof EVMS’ coordination-improving capabilitieson loaded miles per period in use among adopt-ers as of 1997. The sum of OBC1 and EVMS1is 0.127 with a standard error of 0.018. Thisgives a point estimate of EVMS’ total impact onloaded miles per period in use, averaged acrossadopters: 12.7 percent.

Moving down the table, OBC2 and EVMS2are estimates of �2. These reflect relationshipsbetween OBC use and periods in use.33 OBC2 ispositive and EVMS2 is negative. The former isnot statistically significantly different from zerousing a t-test of size 0.05, but is using one ofsize 0.15. The latter is significant using one ofsize 0.05. The point estimates indicate that,holding constant truck vintage and other con-trols, trucks with trip recorders are used 5.6percent more periods than trucks without OBCsand 7.8 percent more periods than trucks withEVMS. One interpretation of this is that triprecorders tend to be used for hauls with regularschedules, and these hauls tend not to be cyclical.The sum of OBC2 and EVMS2 is not signifi-cantly different from zero, implying that truckswith EVMS are used almost exactly the samenumber of weeks on the average as trucks with-out OBCs. While periods in use appears high fortrucks with trip recorders, it is not for trucks withEVMS. Controlling for periods in use thereforemostly adjusts for differences between truckswith trip recorders and the other categories, notbetween trucks without OBCs and with EVMS.

The estimate of � indicates that doubling theshare of periods a truck is in use increases theshare of weeks it is in use by about 40 percent.34

One can strongly reject the hypothesis that thiselasticity equals one. As expected, trucks that arein use twice as many weeks are used much morethan twice as many periods, and simply normal-izing loaded miles by number of weeks in use

would not have fully corrected for differences inthe number of periods trucks are used.

Comparing the estimates of OBC and EVMSin the right column of Table 2 to those of OBC1and EVMS1 in Table 3 allows one to observethe effect of the controlling for differences innumber of periods in use. Whereas the coeffi-cient on OBC in Table 2 is positive and signif-icant, that on OBC1 in Table 3 is much lowerand is not statistically significantly differentfrom zero. In contrast, whereas the coefficienton EVMS in Table 2 is small and not statisti-cally significantly different from zero, that onEVMS1 in Table 3 is positive and significant.Ignoring the fact that the trucks with trip record-ers are used more periods than other trucksleads one to overstate OBCs’ incentive benefitsand understate their coordination benefits.

The rest of the columns report results fromspecifications that provide evidence regardingwhether the potential biases from reverse cau-sation and spurious correlation discussed aboveare economically significant. These thus exam-ine the assumption E(�1i�Di) � 0. The secondcolumn omits the private carriage dummy fromXi; if the estimate of EVMS1 in the first columnreflects that EVMS adoption is high whereloaded miles per period is high because of dif-ferences in drivers’ service responsibilities,omitting the private carriage dummy should ex-acerbate this bias and cause the coefficient toincrease. However, the estimate of EVMS1 isalmost exactly the same as in the first column,increasing by a very small and statistically in-significant amount. The rest of the columnsinclude in Xi a full set of state, product, andstate and product dummies, respectively. Thecoefficients on these dummies themselves, notreported here, are jointly significant; this pro-vides evidence that loaded miles per period inuse varies with the products trucks haul and thestate in which they are based. However, theestimates of OBC1 and EVMS1 are almost ex-actly the same as in the first column, particu-larly those in the last column that contain thefull set of controls. This provides evidence thatthe estimates of OBC1 and EVMS1 in the firstcolumn do not reflect the effect of spuriouscorrelation related to unobserved congestion orlogistical sophistication. If they did, one wouldexpect OBC1 and EVMS1 to decrease when

33 Multiplying these by the estimate of � provides esti-mates of relationships between OBC use and weeks in use.

34 In specifications not shown here, I have estimated themodel holding � constant at values between 0.3 and 0.5—arange 20 times the standard error—and find that the esti-mates of OBC1 and EVMS1 are stable within this range.Also, the estimates change little when allowing � to be afunction of Xi.


including these additional controls. In sum, therobustness of the estimates to the inclusion orexclusion of these controls provides evidencethat biases related to unobserved haul charac-teristics discussed above are likely quite small.35

The results in Table 3 thus provide evidencethat OBC adoption has increased capacity utili-zation in trucking through better resource allo-cation decisions. Taking the coefficients aspoint estimates of the benefits to adopters,EVMS increased loaded miles per period in useamong trucks for which they were adopted byan average of 12.7 percent. Using the means inTable 1, this translates to about 8,200 moreloaded miles per truck per year: about one moremedium-distance haul per week. Alternatively,one can think of this as about five fewer hoursper 40-hour week of empty or idle time. Most ofthis increase was due to EVMS’ coordination-improving capabilities; point estimates indicatethat they increased capacity utilization among

adopters by an average of 10 percent. In con-trast, Table 3 provides no evidence that OBCs’incentive- and maintenance-improving capabil-ities increased loaded miles per period in use.Trucks with trip recorders do have higherloaded miles than those without them, but thisappears to be due mainly to differences in thenumber of periods they are used—possibly dueto the regularity of the hauls—rather than theeffects of technology.

Table 4 contains analogous estimates usingthe 1992 data. The first column contains esti-mates from the base specification. Strikingly,the estimates of OBC1, EVMS1, and (OBC1 �EVMS1) are all small and not statistically sig-nificant. In contrast to the 1997 estimates, theseestimates provide no evidence that OBCs in-creased loaded miles per period in use amongadopters as of 1992. The estimates of OBC2 andEVMS2 show similar patterns to 1997, but aregreater in absolute value. They indicate thattrucks with trip recorders were used 14.4 per-cent more periods than those without OBCs and10.0 percent more than those with EVMS. Com-paring these estimates to those in the right col-umn of Table 2 indicates that, as in 1997,ignoring differences in periods in use leads one

35 More generally, it indicates that any important failureof the orthogonality condition E(�1i�Di) � 0 would have tobe due to unobserved haul or firm characteristics that are notstrongly correlated with geographic regions, product differ-ences, or drivers’ service responsibilities.

TABLE 4—OBCS AND LOADED MILES PER PERIOD IN USE:1992 COEFFICIENT ESTIMATES OF EQUATION (7)—MULTIVARIATE REGRESSIONS


OBC1 �0.011 �0.027 �0.010 �0.009 �0.008(0.027) (0.028) (0.027) (0.026) (0.027)

EVMS1 0.022 0.048 0.013 0.027 0.017(0.032) (0.033) (0.032) (0.031) (0.032)

OBC2 0.144* 0.159* 0.143* 0.137* 0.137*(0.027) (0.028) (0.027) (0.026) (0.026)

EVMS2 �0.100* �0.123* �0.100* �0.096* �0.096*(0.029) (0.031) (0.027) (0.029) (0.028)

Lambda (�) 0.431* 0.409* 0.432* 0.436* 0.438*(0.011) (0.010) (0.012) (0.011) (0.012)

R2

Loaded miles equation 0.476 0.476 0.480 0.484 0.488Weeks in use equation 0.191 0.190 0.191 0.192 0.192

Omits private carriage dummy from X? N Y N N NIncludes state dummies in X? N N Y N YIncludes product dummies in X? N N N Y YN � 35,766

Notes: OBC1 and EVMS1 measure relationships between OBC use and trucks’ loaded miles per period in use. OBC2 andEVMS2 measure relationships between OBC use and the number of periods trucks are in use. Lambda is the estimatedelasticity between number of periods in use and number of weeks in use. Eicker-White standard errors are in parentheses.



to overstate OBCs’ incentive effect and under-state their coordination effect on capacityutilization.

The rest of the columns report results fromspecifications that omit and exclude variables asbefore. In the second column, I omit the privatecarriage dummy. Unlike in the 1997 data, theEVMS1 estimate increases substantially and theOBC1 estimate decreases somewhat, as onewould expect under the reverse causation storyaddressed earlier. This provides evidence that theestimates in the first column may reflect the effectof unobserved differences in drivers’ jobs as wellas any causal effects. But since OBC1 andEVMS1 were not statistically significantly differ-ent from zero in the first place, this does notchange the conclusion that there is no evidencethat OBCs increased loaded miles per periodamong adopters during 1992. The rest of the col-umns include the state, product, and state andproduct dummies, respectively. Like in the 1997data, the estimates in the first column are robust tothe inclusion of these dummies, indicating thatonce again, it is unlikely that they reflect spuriouscorrelations related to unobserved differences incongestion or recipients’ logistical sophistication.

Thus, Table 4 provides no evidence of OBC-related increases in loaded miles per period inuse as of 1992, roughly four to five years afterthe first OBCs appeared on the market. Con-trasting this with the 1997 results, the fact thatthe average returns among adopters increaseover time is inconsistent with a simple “movingdown the demand curve” diffusion story wherethe highest return adopters adopt first and ap-propriate the benefits instantaneously, but isconsistent with interpretations where the bene-fits of adoption come with a lag.

Lags in the returns to technology adoption arebelieved to be common by some economists,even for some very important innovations.36

Though not the focus of this paper, interviewswith dispatchers and other industry participantsprovide some candidate explanations for suchlags in this context. One is that improvements indispatching software throughout the 1990’s en-abled dispatchers to utilize the information

OBCs collect better. For example, software pre-sented truck location information in graphical(i.e., on a map) rather than text format, and thismade it easier for dispatchers to use this infor-mation to forecast trucks’ availability andmatch them to hauls. Another is that, softwareimprovements aside, it took time for dispatchersand firms to learn how to use the new informa-tion OBCs provided effectively. Evidence fromthe trade press provides further support for thispoint. For example, Jim Mele (1993) reportsthat while advanced OBCs were initially “ac-cepted as alternatives to telephones that alloweddrivers to make check-in calls without leavingtheir trucks ... some fleets are beginning to ex-ploit the real potential that comes from ... takinginformation from vehicles and feeding it di-rectly into management systems to make thebest possible decisions on dispatching and loadmatching.” Given this observation, made inearly 1993, it is unsurprising to find far moreevidence of OBC-related capacity utilization in-creases in 1997 than 1992.

C. Heterogeneity in the Returns to Adoption

Table 5 reports 1997 estimates from analo-gous specifications that allow the OBC andEVMS coefficients to vary across 12 cells.These cells are distance/trailer/governance per-mutations; each coefficient therefore reflects athree-way interaction. Short-haul trucks includethose that generally operate less than 50 milesfrom their base; long-haul trucks are those thatgenerally operate more than 50 miles from theirbase.37 These estimates provide evidence re-garding whether the returns to adopters vary inthe sample according to variables I observe. Theleft panel reports a specification where I esti-mate all of the model’s coefficients; the rightpanel reports results when I restrict all of theOBC1 coefficients to zero.

The table shows two general patterns. First,with the exception of the common/van/short cell,any evidence that OBCs’ incentive-improvingcapabilities lead to increases in capacity utiliza-tion is weak. None of the other OBC1 coeffi-cients are statistically significantly different

36 See Paul A. David (1990) and Timothy F. Bresnahanand Shane Greenstein (1996) for discussions of lags in thereturns to adoption in the context of electrification andcomputers, respectively.

37 I have estimated the models dividing the long-haulcells more finely. The results are similar to those below.


from zero. Furthermore, one can reject the nullthat the OBC1 coefficients are jointly equal tozero using a likelihood ratio test of size 0.05.

Second, the estimates indicate that the av-erage returns among adopters from OBCs’coordination-improving capabilities differacross segments. Moving to the right panel, theEVMS1 coefficients are statistically signifi-cantly different across cells. There are two no-table patterns when comparing the estimatesacross governance forms. One is that the coef-ficients in the private carriage cells are similarto their counterparts in the common carriagecells; in fact, one cannot reject the null hypoth-esis they are the same. This is interesting be-cause many private fleet dispatchers areconstrained with respect to the extent that theycan match trucks to the demands of external

customers; such constraints would lower thereturns to adoption, averaged across the entiresegment. If so, the fact that average returnsamong private carriage and common carriageadopters are similar suggests heterogeneity inthe returns among private fleets. One interpre-tation of the results is that some private fleetdispatchers are relatively unconstrained, andEVMS helps them improve capacity utilizationin the same way it helps for-hire fleet dispatch-ers. The other pattern is that there is less evi-dence of OBC-related capacity utilizationincreases in the contract carriage cells thanthe other cells. The coefficient on EVMS1 ispositive and significant only in the contract/van/long cell, and the coefficient in thiscell is statistically significantly lower than itscounterpart in the common/van/long cell. This

TABLE 5—OBCS AND LOADED MILES PER PERIOD IN USE:1997 COEFFICIENT ESTIMATES OF EQUATION (7)—MULTIVARIATE REGRESSIONS ESTIMATES OF

OBC1 AND EVMS1 FOR TRAILER-DISTANCE-GOVERNANCE CELLS

Unrestricted specification OBC1 � 0

Short haul Long haul Short haul Long haul

OBC1Private, van �0.062 0.055

(0.132) (0.065)Private, not van 0.124 �0.054

(0.262) (0.089)Contract, van �0.149 0.001

(0.620) (0.057)Contract, not van �0.019 0.119

(0.485) (0.059)Common, van 0.600* �0.064

(0.156) (0.065)Common, not van �0.415 0.152

(0.218) (0.087)EVMS1

Private, van 0.462* 0.116 0.404* 0.161*(0.172) (0.067) (0.148) (0.047)

Private, not van �0.081 0.147 0.028 0.094*(0.281) (0.095) (0.127) (0.055)

Contract, van 0.375 0.097 0.225 0.097*(0.631) (0.52) (0.189) (0.036)

Contract, not van 0.422 �0.103 (0.398) �0.001(0.469) (0.055) (0.236) (0.052)

Common, van �0.223 0.280* 0.364* 0.228*(0.156) (0.066) (0.103) (0.042)

Common, not van 0.556 �0.020 0.152 0.116*(0.323) (0.093) (0.308) (0.054)

Log of likelihood function �40,009 �40,015

Notes: OBC1 and EVMS1 measure relationships between OBC use and trucks’ loaded milesper period in use. Specifications are analogous to those in Table 3. Eicker-White standarderrors are in parentheses.



indicates that the distribution of the returns toadoption differs between these cells, and is con-sistent with the interpretation that schedule reg-ularity tends to make the capacity utilization-related returns uniformly low within contractcarriage. OBCs are sometimes installed ontrucks used for hauls governed by long-termarrangements, but more of the benefits probablycome in ways other than truck utilization; forexample, it may enable shippers’ customers toallocate resources better by helping them trackand anticipate deliveries.

Table 6 explores the distribution of EVMS-related capacity utilization increases. The firstrow reports the estimate of (OBC1 � EVMS1)from the first column in Table 3 (0.127), fol-lowed by several calculations. Reading across,the “all trucks” cell is 100 percent of the indus-try, EVMS adoption in this cell is 25.6 percent,and adopters in this cell make up 25.6 percent ofthe industry. Taking 12.7 percent as the averagecapacity utilization increase among adopters inthe industry, these imply that EVMS use byadopters in this (universal) cell increased capac-ity utilization by 3.3 percent. This is an estimateof advanced OBCs’ effect on capacity utiliza-tion in the industry as of 1997.

The rest of the rows use the estimates fromthe right panel of Table 5 to investigate how the3.3 percent capacity utilization increase splits

across trailer/distance/contractual form cells.For example, the EVMS1 coefficient in the pri-vate/van/short cell is 0.404. This cell made up2.7 percent of the industry and adoption was15.1 percent in this cell. Thus adopters in thiscell made up 0.4 percent of the industry and onthe average increased capacity utilization by40.4 percent. Adoption within this cell in-creased capacity utilization in the industry by0.17 percent (0.404 � 0.004), which is 4.8percent of the industry total. Although the av-erage returns among adopters are high withinthis cell, there are so few adopters in this cellthat it contributes a small amount to the overallcapacity utilization increase.

The main result from this table is that thedistribution of IT-related productivity increasesappears highly skewed across segments. Only5.5 percent of the trucks in the industry—adopt-ers in the common/van/long cell—account forabout 36 percent of the capacity utilization in-crease.38 Approximately another 37 percentcomes from the other two long-haul van cells.

38 For all rows save the first, column (6) equals column(5) divided by 3.48 percent, which is the sum of the column(5) entries from the cells. This differs from 3.25 percent, theestimate of industry capacity utilization gains from Table 3,because the coefficient estimates in column (1) are from adifferent specification.

TABLE 6—DISTRIBUTION OF EVMS-RELATED CAPACITY UTILIZATION INCREASES, 1997

Column (1) (2) (3) (4) (5) (6)

Label Coefficientestimate

Share ofindustry

EVMSadoption

Share � EVMSadoption

Industry CU gainsfrom cell

Share of CUgains

Formula (2) � (3) (1) � (2) � (3)

All trucks 0.127 1.000 0.256 0.256 0.033 1.000

Private, van, short 0.404 0.027 0.151 0.004 0.002 0.048Private, not van, short 0.028 0.118 0.070 0.008 0.000 0.007Contract, van, short 0.225 0.009 0.100 0.001 0.000 0.006Contract, not van, short 0.398 0.007 0.158 0.001 0.000 0.013Common, van, short 0.364 0.019 0.146 0.003 0.001 0.029Common, not van, short 0.152 0.017 0.094 0.002 0.000 0.007Private, van, long 0.161 0.146 0.310 0.045 0.007 0.209Private, not van, long 0.094 0.182 0.166 0.030 0.003 0.081Contract, van, long 0.097 0.135 0.444 0.060 0.006 0.166Contract, not van, long �0.001 0.086 0.294 0.025 �0.000 �0.001Common, van, long 0.228 0.161 0.343 0.055 0.013 0.361Common, not van, long 0.116 0.094 0.237 0.022 0.003 0.074

Notes: “All trucks” coefficient estimate is (OBC1 � EVMS1) from the first column in Table 3. Cell coefficient estimates are fromthe right panel of Table 5. Share of industry is the cell’s share of trucks. EVMS adoption is the share of trucks in the cell that haveEVMS installed. Share of CU gains is the cell’s share of industrywide OBC-related increases in loaded miles per period in use.


Thus, about 15 percent of the U.S. fleet ac-counts for about 73 percent of the benefit. Morethan half of the rest comes from adopters in thelong-haul nonvan cells.

D. How Much of the Increase in CapacityUtilization Between 1992 and 1997 Was

EVMS-Related?

The estimates in Table 3 imply that EVMSenabled increases in capacity utilization of theU.S. tractor-trailer fleet of 3.3 percent in 1997.In contrast, there is no evidence from Table4 that they led to significant increases in capac-ity utilization as of 1992. Table 1 reported thatloaded miles per truck increased by 10.1 percentbetween 1992 and 1997. The point estimates inthis paper suggest that about 33 percent of thisincrease (0.033/0.101) was related to the grow-ing use of on-board computers to achieve bettermatches between trucks and hauls. A substantialpart of the rest is likely due to the expansion ofthe economy during this time.

This estimate of 33 percent should probablybe considered an upper bound, because EVMSuse may have led to capacity utilization in-creases within certain segments as of 1992. Ta-ble A2 in the Appendix shows 1992 results fromspecifications analogous to Table 5. In the rightpanel, the estimates of EVMS1 are positive andsignificant for the private/not van/long andcommon/van/long cells. These point estimatesindicate that adoption within these cells in-creased capacity utilization fleetwide by 0.4percent.39 If one assumes that capacity utiliza-tion increases are zero in the rest of the cells,this would imply that about 29 percent [(0.033–0.004)/0.101] of the capacity utilization increasebetween 1992 and 1997 was due to EVMS-related improvements in resource allocation.

E. What Are the IT-Enabled Increases inCapacity Utilization Worth?

Trucking makes up a significant part of theeconomy; thus, even small proportional in-creases in productivity imply large benefits in

absolute terms. The American Trucking Asso-ciations estimates that trucking (including pri-vate fleets) was a $486 billion industry in 1998,or 6.1 percent of GDP.40 Operating margins aresmall in trucking; therefore, this is a roughapproximation of costs. Multiplying $486 bil-lion by 3.3 percent gives a back-of-the-envelopeestimate of the value of OBC-related increasesin capacity utilization: $16 billion per year. Thisestimate does not account for productivity ben-efits other than in truck utilization, such as anybenefits that accrue to shippers and receiversfrom being better able to anticipate trucks’ ar-rivals. Sixteen billion dollars in annual benefitstherefore may well be a conservative estimatefor the general productivity gains associatedwith OBC use as of 1997.

These increases in capacity utilization haveinvolved costs, but the costs are probably verysmall relative to the benefits. Although there aredepreciation and labor costs from using trucksmore intensively, these are probably quite smallin many cases. For example, running trucksloaded rather than empty causes little extra de-preciation, and does not require drivers to workmore hours. Furthermore, the OBCs themselvesare very inexpensive; the most popular EVMScosts users only $100 per month per truck tolease, including messaging costs. While mypoint estimate of the average capacity utiliza-tion increase among adopters is 13 percent,EVMS hardware and messaging costs increaseoperating costs by less than 1 percent.41 Finally,while using OBCs effectively usually requiressome complementary investments in humancapital and back-office IT, it generally does notinvolve changes in dispatchers’ or drivers’ jobsthat require significant amounts of training, andbackoffice hardware and software is usually PCbased and supplied by competitive firms. Thenet benefits would be very high even if the

39 Adopters within these two cells made up 0.88 percentand 4.01 percent of the fleet, respectively; 0.004 �(0.0088 � 0.160 � 0.0401 � 0.097).

40 American Trucking Associations (2000). I quote theestimate for 1998 because methodological changes and newdata led this and other publications to substantially increasetheir estimate of the size of the industry, starting first withestimates for 1998. These methodological changes accountfor the fact, for example, that much of “ rail” and “air”freight travels by truck for all or part of the way.

41 Assuming operating costs of $2/mile (AmericanTrucking Associations, 2000) and 6,000 miles per truck permonth, average monthly operating costs are on the order of$12,000 per month.


amortized cost per truck of these complemen-tary investments were five times hardware andmessaging costs, and there is no indication frominterviews and the trade press that the costsassociated with such investments are nearly thislarge.

V. Conclusion

Technologies that collect and disseminate in-formation play a unique role in the economy. AsHayek stated more than 50 years ago, suchtechnologies increase productivity by improv-ing decisions, in particular resource allocationdecisions. This paper examines the impact ofone such technology—on-board computers—on capacity utilization in the trucking industry.The evidence in this paper indicates that on-board computer use has increased capacity uti-lization significantly: in 1997, EVMS increasedcapacity utilization by 13 percent on adoptingtrucks. This increase appears to be mostly due

to advanced capabilities that let dispatchers de-termine trucks’ position in real time, and allowdispatchers and drivers to communicate whiledrivers are in their trucks. These capabilitiesenable dispatchers and drivers to keep trucks onthe road and loaded more.

On-board computers in trucking are amongthe first commercially important applications ofwireless networking technologies. Many othersuch applications are likely to follow in the nearfuture, as companies are currently attempting todevelop and commercialize wireless applica-tions that work off a diverse set of hardwareplatforms, including phones and handheld com-puters. The economic value of these applica-tions is based on the same principle as OBCs:information improves decisions; communicationenables decisions to be executed. This allows dis-persed individuals to identify and avail themselvesof economic opportunities. The estimates in thispaper indicate that the productivity gains fromsuch applications can be quite large.


TABLE A1—OBCS AND LOADED MILES PER PERIOD IN USE:1992 AND 1997 COEFFICIENT ESTIMATES OF EQUATION (7)—MULTIVARIATE REGRESSIONS


1997 1992

EstimateStandard

error EstimateStandard

error

� vector OBC1 0.023 0.029 �0.011 0.027EVMS1 0.104* 0.029 0.022 0.033OBC2 0.056 0.029 0.144* 0.027EVMS2 �0.078* 0.029 �0.100* 0.029

� vector C 10.037* 0.041 10.008* 0.033Area: 50–100 miles 0.178* 0.046 0.264* 0.031Area: 100–200 miles 0.454* 0.046 0.527* 0.033Area: 200–500 miles 0.753* 0.046 0.778* 0.031Area: �500 miles 1.009* 0.043 0.993* 0.031Private carriage �0.118* 0.024 �0.134* 0.022Contract carriage 0.086* 0.022 0.091* 0.019Owner-operator: Independent 0.216* 0.061 0.117* 0.033Owner-operator: Subcontractor 0.264* 0.038 0.239* 0.032Trailer: Lowboy 0.023 0.066 0.029 0.052Trailer: Platform 0.031 0.033 0.023 0.026Trailer: Refrigerated van �0.021 0.027 0.002 0.024Trailer: Logging 0.474* 0.070 0.277* 0.044Trailer: Grain body 0.533* 0.085 0.475* 0.063Trailer: Dump 0.363* 0.046 0.359* 0.037Trailer: Tank 0.034 0.037 0.011 0.029Trailer: Other �0.113* 0.033 �0.208* 0.026LTL �0.091* 0.029 �0.082* 0.028LTL � (area � 50) �0.161* 0.080 �0.405* 0.056

� vector C – – – –Area: 50–100 miles 0.466* 0.055 0.316* 0.035Area: 100–200 miles 0.486* 0.054 0.338* 0.036Area: 200–500 miles 0.444* 0.052 0.300* 0.034Area: �500 miles 0.305* 0.051 0.196* 0.034Private carriage �0.176* 0.027 �0.230* 0.023Contract carriage �0.027 0.025 �0.063* 0.020Owner-operator: Independent �0.298* 0.084 �0.114* 0.039Owner-operator: Subcontractor �0.231* 0.045 �0.163* 0.036Trailer: Lowboy �0.512* 0.083 �0.656* 0.060Trailer: Platform �0.147* 0.039 �0.159* 0.029Trailer: Refrigerated van 0.097* 0.034 0.076* 0.026Trailer: Logging �0.264* 0.088 �0.090* 0.047Trailer: Grain body �0.958* 0.100 �0.789* 0.077Trailer: Dump �0.138* 0.056 �0.213* 0.043Trailer: Tank 0.013 0.039 0.011 0.032Trailer: Other �0.051 0.036 0.006 0.025LTL 0.024 0.033 �0.015 0.026LTL � (area � 50) 0.599* 0.086 0.463* 0.052Model year 1996 (1991 for 1992 specification) 0.229* 0.028 0.375* 0.025Model year 1995 (1990 for 1992 specification) 0.202* 0.027 0.415* 0.023Model year 1994 (1989 for 1992 specification) 0.154* 0.029 0.339* 0.023Model year 1993 (1988 for 1992 specification) 0.118* 0.031 0.288* 0.025Model year 1992 (1987 for 1992 specification) 0.114* 0.035 0.241* 0.026Model year 1991 (1986 for 1992 specification) 0.029 0.036 0.146* 0.027Model year 1990 (1985 for 1992 specification) �0.070* 0.039 0.094* 0.028Model year 1989 (1984 for 1992 specification) �0.076* 0.037 0.044 0.029Model year 1988 (1983 for 1992 specification) �0.190* 0.041 0.006 0.038Model year 1987 or before (1982 for 1992 specification) �0.643* 0.033 �0.529* 0.027

� vector Farm products �0.174* 0.018 �0.185* 0.016Live animals �0.190* 0.038 �0.179* 0.022

� 0.406* 0.016 0.431* 0.011

Note: Standard errors are Eicker-White.* Significantly different from 0 at the 5-percent level.


REFERENCES

American Trucking Associations. American truck-ing trends 2000. Alexandria, VA: TransportTopics Press, 2000.

Athey, Susan and Stern, Scott. “The Impact ofInformation Technology on EmergencyHealth Outcomes.” RAND Journal ofEconomics, Autumn 2002, 33(3), pp. 399 –432.

Baker, George F. and Hubbard, Thomas N. “Con-tractibility and Asset Ownership: On-BoardComputers and Governance in U.S. Truck-ing.” National Bureau of Economic Research(Cambridge, MA) Working Paper No. 7634,2000.

. “Make Versus Buy in Trucking: AssetOwnership, Job Design, and Information.”

American Economic Review, June 2003,93(3), pp. 551–72.

Bresnahan, Timothy F. and Greenstein, Shane.“Technical Progress and Co-invention inComputing and the Use of Computers.”Brookings Papers on Economic Activity: Mi-croeconomics, 1996, pp. 1–77.

Brynjolfsson, Erik and Hitt, Lorin. “ParadoxLost? Firm-Level Evidence on the Returns toInformation Systems Spending.” Manage-ment Science, April 1996, 42(4), pp. 541–58.

. “Beyond Computation: InformationTechnology, Organizational Transformationand Business Performance.” Journal of Eco-nomic Perspectives, Fall 2000, 14(4), pp. 23–48.

Brynjolfsson, Erik and Yang, Shinku. “ Informa-tion Technology and Productivity: A Review

TABLE A2—OBCS AND LOADED MILES PER PERIOD IN USE:1992 COEFFICIENT ESTIMATES OF EQUATION (7)—MULTIVARIATE REGRESSIONS ESTIMATES OF

OBC1 AND EVMS1 FOR TRAILER-DISTANCE-GOVERNANCE CELLS

Length of haul Length of haul

Short Long Short Long

OBC1Private, van 0.674* �0.145*

(0.327) (0.040)Private, not van �0.004 0.007

(0.202) (0.061)Contract, van 0.015 �0.085

(0.236) (0.043)Contract, not van �0.356 �0.112

(0.391) (0.066)Common, van 0.426 �0.100

(0.297) (0.075)Common, not van �0.266 0.323*

(0.172) (0.075)EVMS1

Private, van �0.673 0.158* �0.048 0.043(0.405) (0.065) (0.259) (0.060)

Private, not van �0.385 0.153 �0.390* 0.160(0.263) (0.103) (0.139) (0.089)

Contract, van �0.271 0.108 �0.256 0.031(0.487) (0.054) (0.447) (0.040)

Contract, not van 0.443 �0.057 0.104 �0.158*(0.412) (0.114) (0.169) (0.098)

Common, van �0.120 0.189* 0.296 0.097*(0.410) (0.077) (0.290) (0.035)

Common, not van 0.247 �0.499* �0.002 �0.201*(0.194) (0.115) (0.139) (0.092)

Log of likelihood function �68,183 �68,219

Notes: OBC1 and EVMS1 measure relationships between OBC use and trucks’ loaded milesper period in use. Specifications are analogous to those in Table 3. Eicker-White standarderrors are in parentheses.



of the Literature,” in Marvin Zelkowitz, ed.,Advances in computers, volume 43. NewYork: Academic Press, 1996.

David, Paul A. “The Dynamo and the Computer:An Historical Perspective on the ModernProductivity Paradox.” American EconomicReview, May 1990 (Papers and Proceed-ings), 80(2), pp. 355–61.

Gordon, Robert J. “Does the ‘New Economy’Measure Up to the Great Inventions of thePast?” Journal of Economic Perspectives,Fall 2000, 14(4), pp. 49–74.

Hayek, F. A. “The Use of Knowledge in Soci-ety.” American Economic Review, September1945, 35(4), pp. 519–30.

Hubbard, Thomas N. “The Demand for Moni-toring Technologies: The Case of Trucking.”Quarterly Journal of Economics, May 2000,115(2), pp. 533–60.

. “Contractual Form and Market Thick-ness in Trucking.” RAND Journal of Eco-nomics, Summer 2001, 32(2), pp. 369–86.

Jorgenson, Dale W. “ Information Technologyand the U.S. Economy.” American EconomicReview, March 2001, 91(1), pp. 1–32.

Lehr, William and Lichtenberg, Frank R. “Com-

puter Use and Productivity Growth in U.S.Federal Government Agencies.” Journal ofIndustrial Economics, June 1998, 46(2), pp.257–79.

Lichtenberg, Frank R. “The Output Contribu-tion of Computer Equipment and Person-nel: A Firm-Level Analysis.” Economics ofInnovation and Technology, 1995, 3, pp.201–17.

Mele, Jim. “Forget the Crystal Ball.” FleetOwner, January 1993, 88(1), p. 58.

Oliner, Stephen D. and Sichel, Daniel E. “TheResurgence of Growth in the Late 1990s: IsInformation Technology the Story?” Journalof Economic Perspectives, Fall 2000, 14(4),pp. 3–22.

Standard and Poor’s. Standard and Poor’s Indus-try Surveys. Trucking Industry Overview,November 30, 1995.

U.S. Bureau of the Census. 1992 census of trans-portation: Truck inventory and use survey.Washington, DC: U.S. Government PrintingOffice, 1995.

. 1997 census of transportation: Vehicleinventory and use survey. Washington, DC:U.S. Government Printing Office, 2000.


information, decisions, and productivity: on-board computers and … · 2004-07-26 · information,...

Documents