Incremental Nonlinear Model for Predicting Pavement

Serviceability
Jorge A. Prozzi1 and Samer M. Madanat2

Abstract: A recursive nonlinear model was developed for the prediction of pavement performance as a function of traffic characteristics,
pavement structural properties, and environmental conditions. The model highlights some of the advantages of relaxing the linear
restriction that is usually placed on the specification form of pavement performance models. First, a functional form that better represents
Downloaded from ascelibrary.org by Osaka Daigaku on 01/22/24. Copyright ASCE. For personal use only; all rights reserved.

the physical deterioration process can be used. Second, the estimated parameters are unbiased, owing to a proper specification and the use
of sound statistical techniques. Finally, the standard error of the prediction is reduced by half that of the equivalent existing linear model.
This improved accuracy has important economic implications in the context of pavement management. The model developed as part of
this research enables the determination of an unbiased exponent of the so-called power law and of the equivalent loads for different axle
configurations. The estimated exponent confirms the value of 4.2 traditionally used. However, it should be noted that this exponent is only
to be used for determining damage in terms of serviceability. On the other hand, equivalent loads estimated for different axle configura-
tions tend to differ from traditionally used values.
DOI: 10.1061/共ASCE兲0733-947X共2003兲129:6共635兲
CE Database subject headings: Pavements; Serviceability; Models; Cost allocations.

Introduction experiment gave origin to the Present Serviceability Rating

共PSR兲. The PSR was found to correlate highly with surface
The objective of this research was the development of sound flex- roughness and, to a lesser extent, with rutting, cracking, and
ible pavement performance models to be used primarily for the patching.
management of the road infrastructure because accurate predic- Models for the prediction of riding quality are very important
tion of pavement performance is important for effective manage- to highway agencies for the purpose of managing their road net-
ment of the infrastructure. work. The prediction of riding quality is also important for road
Owing to the great complexity of the road deterioration pro- pricing and regulation studies. Both the rate of riding quality de-
cess, performance models are, at best, only approximate predic- terioration over time and the contribution of the various factors to
tors of expected conditions. Models that only produce a determin- such deterioration are important inputs to these studies. Useful
istic prediction of performance without any quantification of the models are those that establish the contribution of pavement
accuracy of the prediction are unrealistic; hence, an estimation of structure, traffic, environment and any other factors that are rel-
the prediction error is essential. The absence of such an estimate evant for cost allocation.
imposes important limitations on the applicability of the perfor-
mance prediction model.
The Present Serviceability Index 共PSI兲 was developed in the Data Considerations
early 1960s and constituted the first comprehensive effort to es-
tablish performance standards based upon considerations of riding There are a number of possible data sources that have been used
quality 关Carey and Irick 1960; Highway Research Board 共HRB兲 for the development of pavement performance models. Some of
1962兴. A panel of highway users from different backgrounds these sources are: 共1兲 randomly selected in-service pavement sec-
evaluated seventy-four flexible pavement sections and rated them tions; 共2兲 in-service pavement sections selected according to an
on a discrete five-point scale 共0 for poor, 5 for excellent兲. This experimental design; 共3兲 purposely built pavement test sections
subjected to the action of actual traffic and the environment; and
1
Assistant Professor, Dept. of Civil Engineering, The Univ. of Texas, 共4兲 purposely built pavement test sections subjected to the accel-
ECJ 6.8, Austin, TX 78712. 共corresponding author兲. E-mail: erated action of traffic and environmental conditions.
[email protected] Data originating from in-service pavement sections subjected
2
Professor, Dept. of Civil and Environmental Engineering, Univ. of to the combined actions of highway traffic and environmental
California, 114 McLaughlin, Berkeley, CA 94720. E-mail: conditions are those that most closely represent the actual dete-
[email protected] rioration process of pavements in the field. All other data sources
Note. Discussion open until April 1, 2004. Separate discussions must would produce models that would suffer from some kind of bias
be submitted for individual papers. To extend the closing date by one
or restrictions unless special considerations are taken into account
month, a written request must be filed with the ASCE Managing Editor.
The manuscript for this paper was submitted for review and possible during the parameter estimation process. Some of these consider-
publication on November 26, 2001; approved on December 20, 2002. ations are briefly described in the following paragraphs.
This paper is part of the Journal of Transportation Engineering, Vol. The most common problems encountered in models developed
129, No. 6, November 1, 2003. ©ASCE, ISSN 0733-947X/2003/6- from randomly selected in-service pavement sections are caused
635– 641/$18.00. by unobserved events typical of such data sets, and by the prob-

JOURNAL OF TRANSPORTATION ENGINEERING © ASCE / NOVEMBER/DECEMBER 2003 / 635

J. Transp. Eng., 2003, 129(6): 635-641

 lem of endogeneity bias generated by the use of endogenous vari- oped based purely on statistical considerations without any
ables as explanatory variables 共Paterson 1987; Prozzi and attempt to represent the actual physical phenomenon underlying
Madanat 2000兲. These two issues are discussed in detail in the the performance process. Different researchers have approached
following paragraphs. the problem in different ways, especially in the way in which the
Data-gathering surveys during experimental tests are usually form of the model specification is developed.
of limited duration. This may cause problems when such data are In the majority of empirical work found in the literature, ex-
used for developing predictive models of pavement failure, or planatory variables are used and discarded solely on the basis of
other discrete events 共such as cracking initiation兲. If only the consideration of the statistics of their parameters. Often, relevant
events observed during the survey are included in the statistical variables are discarded, owing to low statistical significance 共as
analysis 共ignoring the information of the after and before events兲 measured by t statistics兲. On the other hand, irrelevant variables
the resulting models would suffer from truncation bias. If censor- are often incorporated into the model, based on the same consid-
ing of the events is not properly accounted for, the model may erations. Any models developed following such an approach will
suffer from censoring bias 共Paterson 1987; Prozzi and Madanat undoubtedly suffer from specification biases. Most of the specifi-
2000兲. cations are a linear combination of the available regressors, and
Another common problem is that of endogeneity bias. Pave- the criterion for the selection of the best specification among sev-
Downloaded from ascelibrary.org by Osaka Daigaku on 01/22/24. Copyright ASCE. For personal use only; all rights reserved.

ments that are expected to carry higher volumes of traffic during eral alternatives is to obtain the best possible fit to the data 共usu-
their design life are designed to higher standards. The bearing ally measured by the coefficient of determination, R 2 ).
capacity of these pavements is higher than those of pavements A few researchers have used specification forms that simulate
designed to carry lower volumes of traffic. Thus, any variable the actual physical process of deterioration. In their work, the
which is an indicator of a higher bearing capacity, such as the form of the specification, even though relatively simple 共by com-
structural number, will indeed be an endogenous variable which is parison with the actual physical phenomenon兲, is not constrained
determined within the system and cannot be used as an indepen- to linear equations. This approach is often referred to as
dent explanatory variable. If such a variable were incorporated empirical–mechanistic approach.
into the model, the resulting models would suffer from endoge-
neity bias 共Madanat et al. 1995兲. Mechanistic Models
These problems can be accounted for through the use of sta-
tistical techniques that take into account the presence of trunca- These models are based on the use of material behavior and pave-
tion or endogeneity or, alternatively, by developing models that ment response functions, which are believed to represent the ac-
are based on data originating from in-service pavement sections tual behavior of the pavement structure under the combined ac-
whose selection is based on an experimental design. tions of traffic and the environment. Although there are currently
Purposely built pavement sections subjected to the action of various attempts in this direction, a comprehensive and reliable
actual traffic and the environment are often desirable sources of mechanistic pavement model has yet to be developed. Material
data possible. However, time and budget limitations constrain this models presently used are simplifications and only represent ma-
type of experiment to a very limited number. Constructing pave- terial behavior under restricted conditions. They lack empirical
ment test sections and subjecting them to the accelerated action of validation under a wide range of traffic and environmental condi-
traffic and the environmental solves some of the budgetary and tions. Owing to the complexity of the road deterioration process,
time constraints. However, this produces models that are depen- this approach is, at present, unfeasible.
dent on the testing conditions. One way of overcoming some of
these limitations is through the use of data from multiple sources. Mechanistic – Empirical Models
Archilla and Madanat 共2001兲 have successfully developed models These models make use of material characterization 共laboratory
for the prediction of pavement rutting by combining two different or in situ testing兲 and pavement response models 共usually
experimental data sources. multilayer linear elastic or finite element type models兲 to deter-
mine pavement response. This response is, in turn, correlated to
pavement performance and finally calibrated to an actual pave-
Modeling Approaches ment structure. Both pavement test sections and in-service pave-
ment sections are used for this purpose. The models are usually
Pavement performance models can be classified into two groups: calibrated by applying a bias correction factor 共usually referred to
empirical models and mechanistic models, depending on the ap- as the shift factor兲. The determination of this factor is performed
proach followed for the development of the performance function. following ad-hoc procedures that are not supported by rigorous
A third group comprises the so-called mechanistic–empirical statistical procedures. It is often found that these shift factors are
models that make use of both mechanistic concepts and empirical not supported by actual performance data but by engineering
modelling. Some of the main characteristics of each type are de- judgment.
scribed in the following sections. In the past 20 years there has been a tendency to direct their
efforts toward mechanistic–empirical models because they are
more appealing from the engineering point of view. It should be
Empirical Models
noted that mechanistic–empirical approaches have still to prove
In these models, the dependent variable is some indicator of pave- that they constitute an improvement over more traditional empiri-
ment performance. Both subjective indicators 共riding quality, ser- cal models. The main advantage that mechanistic models claim is
viceability, condition index, etc.兲 and objective indicators 共rough- their ability to extrapolate predictions out of the original data
ness, rutting, cracking, etc.兲 are used as dependent variables. range. This advantage constitutes, in turn, their main disadvan-
These performance indicators are related to one or more explana- tage, since it is impossible to assess the reliability of their predic-
tory variables, such as pavement structural strength, traffic load- tions when they are used out of the original data range for which
ing and environmental conditions. These models are often devel- these models were calibrated.

636 / JOURNAL OF TRANSPORTATION ENGINEERING © ASCE / NOVEMBER/DECEMBER 2003

J. Transp. Eng., 2003, 129(6): 635-641

 Table 1. Axle Configuration and Loads During American Associa- Most of the sections on the flexible pavement tangents were
tion of State Highway Officials Road Test part of a complete experimental design, the design factors of
Weight 共kN兲 which were surface thickness, base thickness, and subbase thick-
Axle
Loop Lane configurationa Front axle Rear axle Gross weight
ness. In each loop, the dimensions of the main factorial designs
were 3⫻3⫻3, that is, three levels of surface thickness combined
2 1 1-1 8.9 8.6 17.8 with three different base thicknesses and three subbase thick-
2 1-1 8.9 26.7 35.6 nesses. Typical layer thicknesses were selected. For the pavement
3 1 1-1-1 17.8 53.4 124.6 structural sections comprising the main experimental design the
2 1-2-2 26.7 106.8 240.3 surface thickness varied from 0 to 150 mm, in increments of 25
mm. The base layer varied in thickness from 0 共no base layer兲 to
4 1 1-1-1 26.7 80.1 186.9 225 mm, in increments of 75 mm. The subbase layer thickness
2 1-2-2 40.1 142.4 324.9 varied from 0 共no subbase layer兲 to 400 mm, in increments of 100
5 1 1-1-1 26.7 99.7 227.0 mm. The materials used for the construction of the AC surface,
2 1-2-2 40.1 178.0 396.1 base and subbase layers were the same in all sections. Hence, the
effect of material properties on pavement performance cannot be
Downloaded from ascelibrary.org by Osaka Daigaku on 01/22/24. Copyright ASCE. For personal use only; all rights reserved.

6 1 1-1-1 40.1 133.5 307.1 assessed from the data of the main experimental design. Other
2 1-2-2 53.4 213.6 480.6 experiments aimed at assessing different surface and base mate-
a
1-1-1 indicates single front axle and two single rear axles, while 1-2-2 rials were also conducted during the AASHO Road Test but were
indicates single front axle and two tandem rear axles. not part of the main experimental design and were not therefore
considered in the development of the models presented in this
Data Source: American Association of State research.
Highway Officials Road Test
Existing Models Based on American Association
The American Association of State Highway Officials 共AASHO兲 of State Highway Officials Road Test Data
Road Test was sponsored by AASHO and took place in the late The first pavement performance model was developed based on
1950s near Ottawa, Ill. 共HRB 1962兲. The site was chosen because the data provided by the AASHO Road Test. The AASHO equa-
the soil in the area was uniform and representative of soils in tion estimates deterioration based on a dimensionless parameter g
large areas of the country. The climate was also considered to be referred to as damage. The damage parameter was defined as the
representative of many states in the northern United States. The loss in serviceability at any given time
average annual precipitation in the region of the test was 864 mm
共34 in.兲. The precipitation occurred throughout the year without
any significant differentiation between dry and wet seasons. The
g t⫽
p 0 ⫺p t
p 0 ⫺p f冉 冊
⫽
Nt b
␳
(1)

soil remained mostly frozen during the winter months, with the where g t ⫽dimensionless damage parameter; p 0 ⫽initial service-
depth of freezing depending on the length and severity of the cold ability at time t⫽0; p t ⫽serviceability index at time t;
season. p f ⫽terminal serviceability index, i.e., serviceability at time of
Only one subgrade material was evaluated during the experi- failure; N t ⫽cumulative number of equivalent 80 kN single axle
ment, as well as only one climatic region. Even though both con- loads applied until time t; and ␳,b⫽regression parameters, which
ditions are typical of large areas of the United States, the use of were found to be functions of axle configuration and load and
the results outside these conditions should be subjected to detailed pavement strength.
assessment of their applicability. Estimation of the effects of other The original serviceability equation as developed from the
subgrade material and/or environmental conditions cannot be at- AASHO Road Test was the following:
tained with this data set. This limitation could be overcome by
p t ⫽5.03⫺1.9 log共 1⫹SV t 兲 ⫺0.01冑C t ⫹ P t ⫺1.38RDt2 (2)
incorporating new data sets and by applying joint estimation
共Archilla and Madanat 2001兲. where SVt ⫽slope variance at time t; C t ⫽crack length in feet per
The test tracks consisted of four large loops, numbered 3– 6, 1,000 ft2; P t ⫽patching in ft2 per 1,000 ft2; and RDt ⫽average rut
and two small loops, numbered 1 and 2. Each loop was a segment depth in inches.
of a four-lane divided highway whose north tangents were sur- A number of studies have indirectly incorporated mechanistic
faced with asphalt concrete 共AC兲 and the south tangents with principles for serviceability prediction. They were not based on a
Portland cement concrete. Only loops 2– 6 were subjected to traf- mechanistically developed model for serviceability but they in-
fic and all vehicles assigned to any one traffic lane had the same corporated a mechanistic subsystem for the prediction of cracking
axle arrangement–axle load configuration. Table 1 shows a sum- and rutting, which, in turn, is used for the prediction of service-
mary of the traffic-loading configuration applied to each loop and ability. This indirect approach has been used by Ullidtz 共1979兲
lane. A total of approximately 1,114,000 axle load repetitions and by Uzan and Lytton 共1982兲. Uzan and Lytton developed the
were applied from November 1958 until December 1960. The following model that is an updated version of the serviceability
number of axle repetitions is equal to the number of truck passes model developed during the AASHO Road Test
for loop 2, while, for the other loops, the number of truck passes
is equal to one half of the number of axle repetitions. This is p t ⫽4.4⫺1.7 log关 1⫹350Var共RDt 兲 ]⫺0.9RDt2.5⫺0.03冑C t ⫹ P t
because of the axle configuration of the truck given in Table 1. (3)
A total of 142 flexible pavement sections were built into the These equations are examples of statistical methods in which
various loops. Each section covered the two lanes and each lane the specification was derived mainly on the basis of the best sta-
was subjected to different traffic loading, so the total number of tistical fit to the data, without any physical considerations. Even
test sections was 284. Out of these, there were 252 original de- though these models fit the data well, they suffer from important
signs and 32 duplicates. specification biases.

JOURNAL OF TRANSPORTATION ENGINEERING © ASCE / NOVEMBER/DECEMBER 2003 / 637

J. Transp. Eng., 2003, 129(6): 635-641

 Model Specification power law is restricted to the conditions under which it was de-
rived, a fact that is often forgotten by pavement engineers. The
Two basic considerations were taken into account for the selec- present serviceability index was the failure criteria, and the for-
tion of the form of the specification: 共1兲 for a given pavement mula converts dual wheeled single axles of different loads into
structure, riding quality decreases as traffic increases, and 共2兲 for their equivalent number of standard axles. A dual wheel single
a given traffic level, the riding quality of weaker pavements de- axle of 80 kN was used as a standard axle. In addition, due to the
creases more rapidly than that of stronger pavements. Therefore, use of inadequate statistics techniques the original estimate of the
the following general and flexible form was selected: parameter ␩ is biased. Hence, it was decided to define different
y⫽ f 共 x 兲 ⫽a⫹bx c (4) power laws for the different axle configurations that are present in
the experimental data set.
where y⫽variable representing a measure of riding quality; Different standard loads 共denominator of the power law兲 are
x⫽variable representing a measure of traffic; a⫽parameter or necessary to transform different axle configurations into ESALs.
function that represents the initial condition; b⫽parameter or It should be noted that three axle configurations were used during
function that represents the rate at which quality deteriorates with the AASHO Road Test: single axles with single wheels, single
traffic; and c⫽parameter that represents the curvature of the func-
Downloaded from ascelibrary.org by Osaka Daigaku on 01/22/24. Copyright ASCE. For personal use only; all rights reserved.

axles with dual wheels and tandem axles with dual wheels. These
tion. three configurations currently cover the vast majority of highway
The sign of the rate parameter b indicates whether the riding traffic in the U.S.
quality is increasing or decreasing with traffic. In addition, the
A number of studies have shown the close dependence of the
magnitude of b gives an indication of how much quality varies
exponent of the power law on the type of distress being consid-
with traffic. Finally, the parameter c represents the curvature of
ered and on the pavement structure 关Christison 1986; Committee
the function, i.e., for c⬍1, c⫽1, and c⬎1, the curve is concave,
of State Road Authorities 共CSRA兲 1986; Prozzi and de Beer
linear or convex respectively.
1997; Archilla 2000兴. The decision was therefore made to esti-
The form of Eq. 共4兲 is suitable for predicting some perfor-
mate the exponent from the data instead of using a precalibrated
mance indicator at any time and, therefore, suitable for design and
exponent. It was also decided to keep the exponent of the power
analysis purposes. However, from the pavement management per-
spective, an incremental form is more beneficial, since condition law constant for the various axle configurations.
data are usually available on a regular basis and predictions are Based on the above considerations, the concept of the equiva-
usually only desired for the next one or two time periods. By lent damage factor 共EDF兲 is introduced. The equivalent damage
using a first order series approximation, the same specification factor is a dimensionless factor, which depends on the character-
can also be used in its incremental form istics of the truck and, which, when multiplied by the number of
trucks, yields up the equivalent number of standard axles. The
y t ⫽y t⫺1 ⫹ f ⬘ 共 x t 兲共 x t ⫺x t⫺1 兲 (5)
following equation applies for the traffic configurations used dur-
In the present research, serviceability 共measured in PSI兲 was ing the AASHO Road Test:

冉 冊 冉 冊 冉 冊
selected as the riding quality indicator to be predicted. Hence, the ␭2 ␭2 ␭2
FA SA TA
recursive model for predicting serviceability has the following EDF⫽ ⫹n 1 ⫹n 2 (9)
form: ␭ 1 80 80 ␭ 3 80
␦
p t ⫽p t⫺1 ⫹␣N t⫺1 ⌬N t (6) where EDF⫽equivalent damage factor; FA⫽load in kN of the
front axle 共single axle with single wheels兲; SA⫽load in kN of the
where p t ⫽pavement serviceability at time t; N t ⫽cumulative traf-
single axle with dual wheels; TA⫽load in kN of the tandem axles
fic up to time t; ⌬N t ⫽traffic increment from time t⫺1 to time t;
with dual wheels; ␭ i ⫽parameters to be estimated; and n 1 ,
and ␣,␦⫽parameters or functions to be estimated.
n 2 ⫽number of single axles and tandem axles per truck, respec-
If Eq. 共6兲 is applied from the beginning of the experiment, the
tively.
following expression is obtained:
s⫽t⫺1

p t ⫽p 0 ⫹␣ 兺
s⫽0
N ␦s ⌬N s⫹1 (7) Specification for Pavement Strength

The function ␣ in Eq. 共7兲 is a decreasing function of the strength

Specification for Aggregated Traffic of the pavement. That is, the serviceability of stronger pavements
decreases more slowly than that of weaker pavements. The speci-
An extension of the traditional approach of aggregating all traffic fication of the function ␣ is based on the concept of structural
into its equivalent number of 80 kN 共18,000 lb兲 single standard number 共HRB 1962兲; however, an alternative designation is pro-
axle loads is used in this paper. This number is usually referred to posed in order to differentiate the present specification from the
as equivalent single axle load 共ESALs兲 or E80s. All axle load concept developed by the AASHO researchers. Thus, ␣ is speci-
configurations are converted into their equivalent number of fied as a decreasing function of the equivalent thickness 共ET兲
ESALs by means of a load equivalence factor 共LEF兲 共AASHO according to the following expression:
1981兲. The most widely used form for the determination of the
LEF is the so-called power law ␣⫽␣ 0 ET ␣ 4 ⫽␣ 0 共 1⫹␣ 1 T a ⫹␣ 2 T b ⫹␣ 3 T s 兲 ␣ 4 (10)

LEF⫽ 冉 冊
P ␩
80
(8) where T a ,T b ,T s ⫽thickness of the surface, base, and subbase lay-
ers, respectively 共in cm兲; ␣ 0 – ␣ 4 ⫽set of parameters to be esti-
where LEF⫽load equivalence factor; P⫽axle load in kN; and mated; and ET⫽equivalent thickness 共in cm兲.
␩⫽parameter. The parameters ␣ 1 , ␣ 2 , and ␣ 3 represent the contribution of
The LEF multiplied by the actual number of axles of the given each layer, relative to the subgrade’s contribution, to the resis-
load, P, represents the number of ESALs. The validity of the tance of the pavement to serviceability loss.

638 / JOURNAL OF TRANSPORTATION ENGINEERING © ASCE / NOVEMBER/DECEMBER 2003

J. Transp. Eng., 2003, 129(6): 635-641

 Environmental Considerations r⫽t⫺1

Environmental conditions are of significant importance to pave-

p it ⫽p i0 ⫺␣ i 兺
r⫽0
␦
N ir ⌬N i,r⫹1 (11)
ment performance; in the hypothetical case that a pavement sec-
where the first subscript, i, indicates the pavement test section and
tion is not subjected to traffic, deterioration will still take place.
the second subscript, t, indicates the time period. For the final
There are two main environmental effects to be taken into ac-
formulation all the parameters are renamed as follows:
count: 共1兲 the effect of high temperatures, which reduce the stiff-
ness of the asphalt layer, and 共2兲 the effect of moisture, which p it ⫽␤ 1 ⫹␤ 2 e ⫺T 1 ⫹ 共 1⫹␤ 3 T ai ⫹␤ 4 T bi
reduces the stiffness of the untreated granular layers. r⫽t⫺1
The viscosity of the asphalt binder decreases as the tempera-
ture increases. Thus, the stiffness of the asphalt concrete also
⫹␤ 5 T si 兲 ␤ 6 兺
r⫽0
␤
exp共 ␤ 7 G r 兲 N ir8 ⌬N i,r⫹1 (12a)
varies with temperature. At low temperatures the asphalt concrete where G r ⫽frost gradient in period r 共in cm/day兲; ⌬N ir , repre-
layer becomes very stiff and brittle, thus, prone to cracking. As sents the cumulative traffic in ESALs for section i in period r,
the temperature increases, the asphalt concrete softens, losing obtained by multiplying the equivalent damage factor of each
stiffness very rapidly and becoming more susceptible to rutting. truck configuration by the actual number of truck passes over the
Downloaded from ascelibrary.org by Osaka Daigaku on 01/22/24. Copyright ASCE. For personal use only; all rights reserved.

The presence of moisture reduces the inter-particle friction of pavement test section as follows:

冉冉 冊 冉 冊 冉 冊 冊
the untreated materials, resulting in a significant loss of material
FA i ␤ 11 SA i ␤ 11 TA i ␤ 11
shear strength and stiffness. This, in turn, results in loss of support ⌬N it ⫽n it ⫹D i ⫹D i
of the asphalt concrete, producing increased stress and strain lev- ␤ 9 80 80 ␤ 1080
els for the same applied load. As stresses and strains increase, so (12b)
does the rate of deterioration of the pavement structure. In the where ␤ i ⫽parameters to be estimated; n it ⫽actual number of
context of the present data set, the presence of water in the pave- truck passes for section i at time period t; and D i ⫽dummy vari-
ment layers has a significant effect during the spring months be- able (D⫽1 for one rear axle, D⫽2 for two rear axles兲.
cause of thawing of the water that was captured and then froze The parameter ␤ 2 was incorporated into the specification 关Eq.
during the winter months. 共12a兲兴 to take into account the dependence of the initial service-
The effect of environmental conditions can be taken into ac- ability value on the thickness of the asphalt layer.
count following either of two approaches: 共1兲 by modifying the
pavement strength accordingly, or 共2兲 by weighing the traffic load Estimation Using Ordinary Least-Squares
repetitions with factors that account for the relative damage of
each repetition. The latter approach was used during the initial The data consist of a panel data set: time series data as well as
analysis of the AASHO Road Test data 共HRB 1962兲 by introduc- cross-sectional data. Several approaches can be followed to ad-
ing weighting factors. The calculation of the weighting factors dress the estimation, the simplest being the estimation of one time
was based on the effect of environmental conditions on surface series regression for each section or, alternatively, the estimation
deflections. In the present study, however, the former approach of one cross-sectional regression at each point in time. These
will be followed because it is believed that it is a more accurate techniques are commonly used, however, at a high cost. If the
representation of the physical deterioration process. parameters in the model are believed to be constant across section
From observation of the data, it was apparent that three phases and along time, more efficient parameters can be estimated by
could be distinguished in the serviceability trend of the pavement combining all the data into a single regression.
sections: The most common and general technique used to estimate Eq.
共12兲 is by combining all time series data and cross sectional data
1. a normal phase, characteristic of the summer and fall peri- and carrying out ordinary least-squares 共OLS兲 estimation. In this
ods, when pavement serviceability of the sections decreased case, the intercept is assumed to be the same for all sections. This
slowly; assumption is reasonable, as it considers that the serviceability of
2. a stable phase, characteristic of the winter period, when ser- all pavements is the result of the same process for all sections,
viceability remained stable; and and it only depends on the variables that are observed. However,
3. an accelerated phase, during which the rate of deterioration unobserved heterogeneity is often present as a result of unob-
increased significantly; this phase corresponds to the spring served section-specific variables. Some of the unobserved vari-
months.
ables that may affect the performance of the various sections dif-
Furthermore, it was observed that the three phases described
ferently are: materials density, drainage conditions, foundation
above were almost identical to the periods of 共1兲 zero frost pen-
type, etc. This unobserved heterogeneity could be dealt with in a
etration, 共2兲 increasing frost penetration, and 共3兲 decreasing frost
number of ways. In this study, both the fixed effect approach and
penetration, respectively. Consequently the frost gradient was the
the random effects approach were used. Some of the results of
variable chosen to capture the effect of the environmental condi-
the estimation using random effects are reported in the following
tions on road serviceability. The frost gradient represents the daily
section.
change in the depth of frost penetration and is expressed in cen-
Eq. 共11兲 represents the conditional expectation of the level of
timeters 共cm兲 of frost penetration per day.
serviceability at a given time t for a given section i. This expec-
tation is dependent on a vector of parameters ␤⫽(␤ 1 , . . . ,␤ 11)
Final Specification and on a vector of explanatory variables X⫽(X 1 . . . X 11) which
take on a specific value for a given section i and at a given time
In the preceding sections the form of the specification was given t. These variables describe the combination of pavement proper-
as a function of the relevant variables for each pavement test ties, traffic characteristics and environmental conditions at a given
section. In this section, the full specification is given taking into time for a given section. By applying the specification to all ob-
account that the data set consists of a panel data set, i.e. time servations 共all sections i⫽1, . . . ,S and all times periods t
series data and cross sectional data are available. ⫽1, . . . ,T i ) we obtain the following equation:

JOURNAL OF TRANSPORTATION ENGINEERING © ASCE / NOVEMBER/DECEMBER 2003 / 639

J. Transp. Eng., 2003, 129(6): 635-641

 p it ⫽E 共 p it /Xi ,␤兲 ⫹␧ it (13) Table 2. Estimated Parameters and Corresponding Statistics

Eq. 共13兲 is not linear in the parameters so the estimation does Parameter OLS estimates t value RE estimates t value
not have a closed-form solution; hence, a nonlinear minimization ␤1 4.13 267.0 4.33 159.4
routine was applied to estimate the parameters. Two assumptions ␤2 ⫺1.7 15.9 ⫺2.2 15.4
are necessary to proceed with the estimation: ␤3 0.59 18.2 0.58 18.3
1. the random error term, ␧ it , is assumed to have mean zero ␤4 0.20 15.1 0.21 14.7
and constant variance: E(␧ it )⫽0, and ␴ 2␧ is constant; and ␤5 0.15 15.0 0.13 14.9
2. the covariance of the error terms is zero across sections and ␤6 ⫺3.11 ⫺42.9 ⫺2.66 ⫺39.6
along time: Cov(␧ it ,␧ js )⫽0, for i⫽ j, or t⫽s. ␤7 ⫺0.075 ⫺48.0 ⫺0.071 ⫺57.0
Under the assumption of normality, the values of the parameters ␤8 ⫺0.47 ⫺39.0 ⫺0.55 ⫺58.3
that minimize the sum of squares deviations will be the maximum ␤9 0.57 26.2 0.55 21.7
likelihood estimators as well as the nonlinear least squares esti- ␤ 10 1.8 104.0 1.9 67.9
mators 共Green 2000兲. The sum of square deviations is given by ␤ 11 3.8 48.6 4.2 44.9
S Ti

兺 兺 关 p it ⫺E 共 p it /Xit ,␤兲兴 2
Downloaded from ascelibrary.org by Osaka Daigaku on 01/22/24. Copyright ASCE. For personal use only; all rights reserved.

F 共 ␤兲 ⫽ (14)
i⫽1 t⫽1 based on multistep linear regressions. The variance of the esti-
where T i represents the number of observation periods for section mated serviceability of the original linear model was approxi-
i. Since the panel data set is unbalanced, in general, T i ⫽T j for mately 0.51, while the variance of the estimate of the nonlinear
i⫽ j. A consistent estimator of ␴ 2␧ is also based on the residuals model is 0.26. Most importantly, this better fit-to-the-data was
共note that the degree of freedom correction is not necessary here, achieved using the same data source and the same number of
since all the results are asymptotic兲 explanatory variables as a result of the flexibility of the nonlinear
S Ti formulation.
1
␴ˆ 2⑀ ⫽ 兺 t⫽1
兺 关 p it ⫺E 共 p it /Xit ,␤ˆ 兲兴 2
n i⫽1
(15)
Statistical testing was carried out using Lagrange multipliers
共LMs兲 to test the hypothesis that ␴ 2u ⫽0, in other words, that the
OLS estimation was appropriate. The resulting LM was 105.07,
where ␤ˆ ⫽OLS estimate of ␤; and n⫽兺 i T i .
which is much larger that the corresponding value for a 5% sig-
nificance level 共3.84兲. Therefore, the null hypothesis was rejected
and it was concluded that the ordinary least squares approach was
Estimation Using Random Effects inappropriate: the unobserved heterogeneity cannot be ignored.
As indicated earlier, there are two commonly used methods to
assess the effect of the unobserved heterogeneity: fixed effects Conclusions
and random effects. Both methods assume that the differences
across sections can be captured by differences in the intercept In both models, the parameter of the asphalt concrete layer 共layer
term. The fixed effects method is based on the estimation of sec- coefficient兲 is approximately three times the parameter of the un-
tion specific intercepts, while the random effects method assumes treated granular base layer. This indicates that, per unit of layer
that the intercept term is randomly distributed in the population. It thickness, the contribution of asphalt to the resistance against loss
should be borne in mind that the OLS approach assumes that the of serviceability is three times that of the granular base. The layer
intercept is constant for all sections. parameters of the base and subbase layers are of the same order of
One of the disadvantages of the fixed effects approach is that magnitude. However, the difference is between 33 and 62% for
the estimated intercepts are specific to the given sample and are the OLS and RE approaches, respectively. This is an indication of
not necessary applicable to the population. In addition, there is a the importance of using the correct approach when carrying out
high cost 共reduction of degrees of freedom兲 associated with its the parameter estimation. These differences are even more impor-
estimation, especially when the number of sections is large. For tant in the case of the parameters corresponding to the load vari-
these reasons, the random effects approach was used in this study. ables.
The condition that the unobserved effects are uncorrelated with The parameter corresponding to the exponent of the power law
the explanatory variables is satisfied because the data used for the is 3.8 for the OLS approach and 4.2 for the RE approach. This is
estimation is originated from an experimentally designed test. significant, not only from the statistical point of view, but also for
Hence, using the random effects 共RE兲 approach, the specification its impact on the allocation of the cost of deterioration produced
becomes by the various vehicle classes. This effect is twofold. On one hand
p it ⫽␤ 1 ⫹u i ⫹ f 共 Xit ,␤ 2 ,...,␤ 11兲 ⫹␧ it (16) the damaging effect of heavy axles 共⬎80 kN兲 is underestimated
by the OLS approach and therefore the damage usually attributed
to heavy vehicles is greater than that which they actually cause.
Estimation Results On the other hand, the damaging effect of lighter vehicles 共those

The parameters of the proposed model were estimated using OLS

and random effects with the data corresponding to the 252 test Table 3. Estimates of Variance Components of Two Approaches
sections with original designs. The parameter estimates, as well as Estimate Ordinary least squares Random effects
their asymptotic t values, are given in Table 2. The high values of
the asymptotic t statistics that were obtained with the sample ␴ 2␧ 0.263 0.150
should be noted. The variance estimates are given in Table 3. ␴ 2u N/Aa 0.106
The nonlinear equation developed in this research fits the data ␴ w2 ⫽␴ 2␧ ⫹␴ 2u N/Aa 0.256
a
of the AASHO Road test much better that the original model Not available.

640 / JOURNAL OF TRANSPORTATION ENGINEERING © ASCE / NOVEMBER/DECEMBER 2003

J. Transp. Eng., 2003, 129(6): 635-641

 with axle loads less than 80 kN兲 is overestimated with the OLS LEF ⫽ load equivalency factor;
approach, thus overpredicting the contribution of lighter traffic in N ⫽ cumulative traffic in equivalent single axle
the estimation of the design traffic. loads;
The formulation of the equivalent load in terms of the EDFs n ⫽ number of axle loads;
enabled equivalent loads for various axle/wheel configurations to p ⫽ serviceability index;
be determined. The equivalent load of a single axle with single R 2 ⫽ coefficient of determination;
wheels is estimated to be 46 and 44 kN for the OLS and the RE RD ⫽ rutting depth;
approach, respectively. These values are smaller than those tradi- SA ⫽ single axle load;
tionally used. If it is borne in mind that the RE estimates are SV ⫽ slope variance;
unbiased estimates, this value indicates that a single axle with T ⫽ layer thickness;
single wheels with a 44 kN load has the same effect on service- TA ⫽ tandem axle load;
ability as an 80 kN single axle with dual wheels. This finding has u, ␧ ⫽ error terms;
important implications on the allocation of the amount of damage X ⫽ vector of explanatory variables;
that can be attributed to the front axles and it is recommended that ␣, ␤, ␦, ␩, ␳, ␭ ⫽ regression parameters;
further research be carried out in this area. ␤ ⫽ vector of regression parameters; and
Downloaded from ascelibrary.org by Osaka Daigaku on 01/22/24. Copyright ASCE. For personal use only; all rights reserved.

The equivalent load for a tandem axle with dual wheels is ⌬N ⫽ traffic increment in equivalent single axle
estimated to be 144 and 152 kN for the OLS and RE approach, loads.
respectively. The OLS estimate is close to the equivalent load Subscripts
determined during the original analysis of the AASHO Road Test i ⫽ positive integer indicating section number; and
data. However, the RE 共unbiased兲 estimate is significantly differ- t ⫽ positive integer indicating time period.
ent. The result of the RE estimation highlights the benefit of using
tandem axles from the point of view of road damage in terms of
serviceability. References
The above discussion emphasizes the three most important
American Association of State Highways and Transportation Officials
aspects that need to be taken into account when pavement perfor-
共AASHTO兲. 共1981兲. AASHO Interim Guide for the Design of Pave-
mance models are being developed: ment Structures, AASHTO, Washington, D.C.
1. a physically realistic model specification; Archilla, A. R. 共2000兲. ‘‘Developing of rutting progression models by
2. an adequate data source; and combining data from multiple sources.’’ PhD dissertation, Univ. of
3. statistically sound estimation techniques. California, Berkeley, Calif.
The model specification should be supported by engineering Archilla, A. R., and Madanat, S. M. 共2001兲. ‘‘Development of pavement
knowledge of the materials behavior under load and environmen- rutting model by combining data from different experimental
tal conditions. The data should be obtained from a well-conceived sources.’’ J. Transp. Eng., 126共4兲, 291–299.
experimentally designed test aimed at addressing all the important Carey, W. N., and Irick, P. E. 共1960兲. ‘‘The pavement serviceability per-
variables that have been identified during the development of the formance concept.’’ Highway Research Bulletin No. 250, Highway
theory. Unfortunately, as this is seldom the case, it is up to the Research Board, Washington, D.C., 40–58.
Christison, J. T. 共1986兲. Pavement response to heavy vehicle test pro-
modeler to take into account these limitations in order to develop
gram: Part 2—Load equivalency factors, Vol. 9, Technical Steering
models that are statistically defensible. Committee of the Vehicle Weights and Dimensions Study, Canroad
This research proved the importance of these three steps: a Transportation Research Corporation, Ottawa.
model was formulated using the same data set and the same vari- Committee of State Road Authorities 共CSRA兲. 共1996兲. Structural design
ables, however, the prediction error of the new model was re- of flexible pavements for interurban and rural roads, Draft No. TRH4,
duced by half. By halving the prediction error agencies can affect Department of Transport, Pretoria, South Africa.
significant budget savings by timely intervention and accurate Greene, W. H. 共2000兲. Econometric analysis, 4th Ed., Prentice–Hall,
planning. Englewood Cliffs, N.J.
Highway Research Board 共HRB兲. 共1962兲. ‘‘The AASHO road test.’’ Spe-
cial Rep. Nos. 61A– 61G and 73, Washington, D.C.
Madanat, S., Bulusu, S., and Mahmoud, A. 共1995兲. ‘‘Estimation of infra-
Acknowledgments
structure distress initiation and progression models.’’ J. Infrastruct.
Syst., 1共3兲, 146 –150.
The writers wish to express their gratitude to the University of Paterson, W. D. O. 共1987兲. Road deterioration and maintenance effects:
California Transportation Center 共UCTC兲 for providing the fund- models for planning and management, The Highway Design and
ing necessary to carry out this research. Transportek 共CSIR兲 is Maintenance Series, The John Hopkins Univ. Press, Baltimore.
also acknowledged for its support. Prozzi, J. A., and de Beer, M. 共1997兲. ‘‘Mechanistic determination of
equivalent damage factors for multiple load and axle configurations.’’
Proc., 8th Int. Conf. on Asphalt Pavements, Vol. 1, Univ. of Washing-
Notation ton, Seattle, 161–178.
Prozzi, J. A., and Madanat, S. M. 共2000兲. ‘‘Analysis of experimental
The following symbols are used in this paper: pavement failure data using stochastic duration models.’’ Transporta-
D ⫽ dummy variable; tion Research Record 1699, Transportation Research Board, National
Research Council, Washington, D.C.
E( ) ⫽ expectation operator;
Ullidtz, P. 共1979兲. ‘‘A fundamental method for prediction of roughness,
EDF ⫽ equivalent damage factor; rutting and cracking of pavements.’’ Proc., Association of Asphalt
ET ⫽ equivalent thickness; Pavement Technologists, Minnesota Univ., Minneapolis.
FA ⫽ front 共steering兲 axle load; Uzan, J., and Lytton, R. L. 共1982兲. ‘‘Structural design of flexible pave-
G ⫽ frost gradient; ments: a simple predictive system.’’ Transportation Research Record,
g ⫽ damage; 888, Transportation Research Board, Washington, D.C.

JOURNAL OF TRANSPORTATION ENGINEERING © ASCE / NOVEMBER/DECEMBER 2003 / 641

J. Transp. Eng., 2003, 129(6): 635-641