J. Infrastructure Plan. and Man., JSCE, No.

744/IV-61, 61-67, 2003 October

[Special Topic]

DEVELOPMENT OF A STOCHASTIC MODEL OF

PAVEMENT DISTRESS INITIATION

Hee Cheol SHIN1 and Samer MADANAT2

1Post.doctoral Researcher, Dept. of Civil and Environmental Engineering, University of California

Berkeley,CA 94720, U.S.A.

E-mail: hcshin@ce.berkeley.edu
2Professor,Dept. of Civil and Environmental Engineering,University of California,

Berkeley,CA 94720, U.S.A.

E-mail: madanat@ce.berkeley.edu

Pavement deterioration models provide useful predictions of distress initiation, for purposes of pavement

design and management. A common problem in modeling the initiation is the inappropriate treatment of

data censoring. If the censoring is not accounted for properly, the model may suffer from statistical biases.

In this paper, an analysis of pavement crack initiation data based on the duration modeling techniques is

presented. Duration models enable the stochastic nature of pavement crack initiation to be represented as
well as censored data to be incorporated in the statistical estimation of the model parameters. The results

show that the model predictions are more accurate than those obtained with the original AASHO model.

Key Words: crack initiation, censoring, duration model

1. INTRODUCTION 2. LITERATURE REVIEW

The planning of maintenance and rehabilitation Pavement deterioration models relate indicators of
(M&R) for highway agencies involves the allocation pavement condition to explanatory variables such as
of limited resources to a number of pavements traffic loads, age, and environmental factors. The
within the highway network. This planning process most common indicators of pavement condition are
uses as inputs both measured and predicted highway surface distresses such as cracking, rutting, potholes,
pavement performance indicators. Therefore, accu- etc. These surface distresses are caused by load,
rate predictions of future pavement performance moisture, temperature, construction defects or a
indicators are necessary for efficient planning of combination of the above.
maintenance and rehabilitation activities. Cracking performance depends on many factors,
The focus of this paper is the development of a including
stochastic duration model for crack initiation. Sec- ●The thickness of various pavement layers;

tion 2 provides a review of previous research con- ● The quality of the constmction materials and

ducted on cracking of asphalt concrete, including practices;

the causes of cracking and the modeling approaches ● Environmental considerations, such as tempera-

that have been used. Section 3 formulates the model. ture and moisture;and

Section 4 presents the specification and results of ● The axle loads and axle configurations to which

the estimation of the model parameters using a data the pavement is subjected.
set from the American Association of State High- The focus of this research is on the statistical es-
way Officials (AASHO) Road Test 1). Section 5 timation of empirical models that relate field data on
concludes this paper. crack initiation to explanatory variables representing

61
the pavement structure, traffic loading, and climate. coefficients that represent the resistance of the mate-
The AASHO Road Test was an accelerated loading rial being used in each layer of the pavement struc-
test experiment. A crack initiation model was de- ture. The following general equation for structural
veloped as part of the test. The crack initiation number reflects the relative impact of the layer coef-
model uses traffic repetitions as the dependent vari- ficients (al) and thickness (Di):
able and pavement thickness and load type as ex-
SN = aD (2)
planatory variables. Though the AASHO cracking
model's functional form was relatively arbitrary, the
model has been widely accepted. It forms the basis The estimated values of the coefficients, al, a2,
for most current pavement design procedures in the and as, were: 0. 33, 0. 10, and 0.08 respectively.
world today. The AASHO Road Test Report 5 pro- The Queiroz-GEIPOT models 3),4) have separate
posed the following crack initiation equation. regression equations that predict crack initiation and
the rate of crack progression. The crack initiation
model used the number of equivalent single axles to
(1) initiation as the dependent variable and the struc-
where tural number as the explanatory variable. The equa-
We = number of weighted axle applications sus- tion for the crack initiation model is as follows:
tained by the pavement before appearance of Class 2
1og10N= a + J31og10SN (3)
Cracking;
D1, D2, D3 = thickness of surfacing, base and sub- where
base respectively, in inches; N = the number of Equivalent Single Axle Loads
L1= nominal axle load, in kips; (ESAL) needed to initiate cracking;
L2 =1 for single axle configuration and 2 for tandem SN = structural number; and
axle configuration; a, /3= regression coefficients.
al, a2, a3, a4 = coefficients that were assigned ear- The World Bank's Highway Design and Mainte-
lier; and nance (HDM) models5) predict the initiation and
Al, M, A3, A4 = regression coefficients. progression of various pavement distresses such as
The AASHO model suffered from severe prob- cracking, rutting, raveling and roughness. Each dis-
lems. These are discussed below. First, the analysis tress model includes a number of explanatory vari-
did not account for censoring. The data are consid- ables such as age, traffic, design parameters, envi-
ered censored when cracking is not actually ob- ronmental factors and other distresses. A probabil-
served. In the case in which the section had cracked istic parametric duration model represented crack
before the first inspection, the observation is left initiation, where the dependent variable is the prob-
censored, or if it had yet to crack at the last inspec- ability distribution of the time to cracking. The ba-
tion, it is considered right censored. In the AASHO sic concepts of probabilistic duration models will be
Road Test, there were several sections that had not described in section 3 of this paper. The HDM-III
cracked by the time the experiment ended, and these crack initiation model used a hazard function, h(t) of
constitute right-censored data. If censoring is not the following form:
accounted for correctly in the statistical estimation
h(t)= yexP(-YN)tr-1 (4)
of model parameters, the estimates can be expected
to be biased 2). Second, the model form was arbi- When 'y < 1, the hazard function is decreasing
trary. One specific problem is the variable (L1 + L2), through time; when y =1, it is a constant; and when
which consists of the sum of two quantities with y > 1, the hazard function is increasing. In the
different units. Finally, the coefficients, a1 to a4, analysis of crack initiation, the parameter t is re-
were determined a-priori instead of being estimated placed by a linear vector function of explanatory
simultaneously with the other parameters. The pre- variables x, =
determined parameters were used to compute the The resulting model for prediction of expected
Structural Number (SN) of the pavement. cumulative traffic loading to crack initiation is:
The Structural Number is related to the thickness
TECR
2 = J3,SN PZe SY (5)
of flexible pavement layers through the use of layer

62
where probability that a pavement cracks in the next small
TECR2= mean cumulative traffic loading at initiation interval, At, given it lasts at least until time t, is
of narrow cracking (in millions of ESAL); given by
SN = structural number;
g(t) = Prob(t T < t + At IT t (8)
SY= SN4I(1;000YE4),whereYE4is the annual The hazard function, h(t), which is the instantane-
traffic loading (in millionsof ESAL/lane/year);and ous rate of change of g(t), is defined as:
f3, 132,f33= regression coefficients.
Several deterioration models reviewed above in- (9)
clude separate equations for distress initiation and
progression. Most crack initiation models were de- The hazard function quantifies the instantaneous
veloped without accounting for censoring, which risk that the pavement sections crack at time t. The
may introducebias in the parameterestimates. cumulative (or integrated) hazard function is ex-
Madanat et al 6) applied a structured econometric pressed as
method for developing deterioration models of
H(t)=Jh(ukiu (10)
pavement crack initiation and progression. A model
system consisting of a discrete model for distress The density function, the survival function, and
initiation and a regression model for distress pro- the hazard function are all related;

gression was developed. The estimation sample for

the progression model is self-selected,as it contains (11)
a disproportionatelylarge fraction of weaker pave-
ments, because they are more likely to have already The following relationships between these func-
started cracking (they have lower initiation times). tions also hold:
This selectivity bias was corrected by using Heck- H(t)-logs(t) (12)
man's sequential procedure. Madanat and Shin 7)
extended this research to account for unobserved and
heterogeneity in the panel data set, using random-
5(t) = et) (13)
effects specifications in both the discrete and con-
tinuous models. Right censoring of the data occurs when the test
ends at time C, before all pavement sections have
failed. For each section i, we either know T if T, <
3. MODEL FORMULATION C, or that T1> C, in which case the time to cracking
for pavement i is censored. The time recorded is
(1) Stochastic duration models min (T1, C1) together with the censoring indicator
Let T denote the time to cracking of a pavementin variable b, which is a dummy variable that takes
a test experiment. T is a random variable that takes
values in (0, oo). Its continuousdistributionis speci- the value 0 if the observation (pavement section) is
fied by a cumulative function F(t) with a density censored and 1 otherwise.
functionf(t). The cumulative distribution function The full likelihood function is obtained by multi-
is plying the respective contributions of values of den-
sity function f for uncensored observations and val-
F(t) = 1f(s)ds=Prob(T <t) (6)
ues of survival function S for censored observations.
The probability that the pavement cracks after t is In the presence of right censoring the likelihood for
given by the survival function, all observations in a sample of size n is 8) :

S(t) = 1- F(t) = Prob(T >t) (7) l =11f (t)l lS (t) =11Lf (t)i -, [S(t1)j (14)
Since we collect data at specific times, and we
know the condition of pavements at these times, the To estimate the parameters of the distributions, we
hazard rate is a more useful function than the cumu- maximize the log likelihood function,
lative density function or the survival function. The

63
 mated by maximum likelihood. In the Weibull
L=1ogl={S iog[f(t )]+(i-s)lag[s(t)p (15) model, the expected time to crack initiation is given
by (9):
Upon choosing a particular distribution, we can
substitute the appropriate expressions for the density
function and the survival function.

(2) The Weibull hazard model

Though the hazard function h(t) could be constant (24)
over time, there are many situations in which it is where the gamma function, I'(z), is defined as
more realistic to suppose that h(t) either increases or FZ= Wwz-e-wdw
decreases over time. A flexible form for such a haz- (25)
ard function is given by for z>0.

h(t) = ae1 t>0 Because the model is not a linear regression model,
(16) there is no obvious equivalent to the conventionally

where a and y are positive constants. The hazard reported standard error.

function given by Equation (16) is called the

Weibull hazard function with parameters a and y.
The parametric model that follows the Weibull haz- 4. MODEL SPECIFICATION AND ESTI-

ard function is called the Weibull model. Note that NATION RESULTS

h(t) increases when y> 1; decreases when y < 1; and

is constant when y =1. (1)Model specification

The Weibull distribution function obtained from The AASHO Road Test Data were used for devel-

Equation (1,6) is: opment of the pavement cracking initiation model.

F(t)=1-exp(-Jth(s)
ds) t>0 We

The
have

number
a total

of
of

sections
2520bservations(test

that had cracked by

sections).

the end

=1-exp(-w) (17) of the test was 185. The remaining 670bservations

were censored(i. e., cracking had not occurred yet

Its density function is: by the time of the test). The model shown in this

section used right wheel path cracking as the de-

f (t)=aytY-'exp(-atr) t>0 (18) pendent variable. The variables included are:

The survival function is thus: ● avt: number of accumulated load repetitions in

the traffic lane before crack initiation(the de-

S(t)=1- F(t) = exp(- aty (19) pendent variable);

If a vector of explanatory variables x is observed ●D1: surface thickness in inches(1 to 6 inches);

with the duration data, the Weibull hazard function ・D2: base thickness in inches(0, 3, 6, and g

is written as: inches;

●D3: sub-base thickness in inches(0, 4, 8, 12, and

h(txj3)= e-YY-1
16inches);

● LOAD: nolninal axle load(in kips);and

= e YX, r-i
(20) ●TYPE: single dummy variable, 1for single axle

where = x J. Then, the distribution function, den- and O for tandem axle.

sity function,and survival function are as follows: We estimated a Weibull model using two different

F(tx13)=1-exp-e-Y--tr model specifications. Our first specification(Model

(21) 1)was a modified version of the original AASHO

model, which was presented earlier(equation 1).

f (tx)= e-YxriY-eXP-e-tr (22) The function u (of the hazard model in equation 20),

using this first specification, is shown below:

S(t)=exp- e-yx-t)
r (23) = J1 + /32D1+F33D2+F4D3+ISL2+/36(LI+L2)
(26)
The parameters y and J3of the model can be esti- where L1 and L2 are defined after equation (1). Es-

64
 sentially, this is similar to the AASHO specification, Table 1 Parameter Estimates for Model 1
with the predetermined coefficients of the structural
number in equation (1) replaced by parameters to be
estimated.
Our second model specification improved on the
original AASHO specification, which is problematic
because it assumes that the effects of L1 and L2 are
separable and additive. This is physically unrealis-
tic for two reasons. First, it assumes a constant rate
of substitution between axle load and axle type,
which is inconsistent with pavement engineering
knowledge. Pavement engineers typically use coefficients have the correct signs, which confirm
nonlinear equations to convert between single and our a priori hypotheses. The ratio of the estimated
tandem axle loads. Second, the specification in- resistance of the asphalt concrete to that of the base
volves an addition of two variables with different is less than 3, which is lower than what was ob-
units (L1 is in units of kips while L2 is in units of tained in the original AASHO model. On the other
axles). hand, the ratio of the estimated resistance of the
Our improved specification accounts for the dif- base to that of the sub-base is about 1. 5, which is
ference in the effects of a single axle and a tandem higher than what was obtained in the original
axle of equal load through the use of interaction AASHO model. The estimated value of the parame-
terms. The function µ in the hazard model with the ter y is close to 1. 0, which seems to indicate a rela-
improved specification (Model 2) is: tively constant hazard rate in terms of load repeti-
= YI+flD1+l3D2+14D3+J35TYPE*LOAD+6(1- tions.
TYPE)*LOAD (27) Table 2 presents the results of the Weibull model
Note that we use two interactive terms for LOAD using our second specification. Again, the t-
and TYPE: the first interactive term is statistics show that each variable is a significant ex-
J35TYPE*LOAD and the second interactive term is planatory variable of crack initiation at one percent
J36(1-TYPE) *LOAD. In this specification, J35is the significance level. Again, it can be seen that the
effect of 1 unit of a single axle load, while 136is the coefficients have the expected signs.
effect of 1 unit of a tandem axle load. The ratios of the estimated resistances of the three
layers are close to those obtained in the original

(2) Estimation results AASHO model. The results indicate that the asphalt
It is expected that an increase in pavement layer concrete layer is about 3. 1 times more effective in
thickness increases the time to cracking of the reducing the rate of crack initiation than the base

pavements, and that an increase in axle load, of ei- layer, and the base layer is about 1. 3 times more
ther type, will decrease the time to pavement crack effective than the sub-base layer. The AASHO re-
initiation. The effect of the surface layer should be sults indicated that the asphalt concrete layer is

greater than the effects of the two unbound layers, about 3. 3 times more effective in reducing crack
and the effect of the sub-baseshould be the smallest. initiation than the base layer, and the base layer is
Moreover, the effect of single loads should be about 1. 3 times more effective than the sub-base

greater than that of tandem loads. The dependent layer 1).

variable in the model is hazard rate, not the time to The coefficients of the two interactive load terms
cracking of the pavements,so the parameters should indicate that a tandem axle load of 1. 85 kip has the
be interpretedaccordingly. same effect on crack initiation as single axle load of

Table 1 presents the results of the Weibull model 1 kip. These relative magnitudes are consistent with
using the first specification. The t-statistics show pavement engineering knowledge. Finally, note that
that each variable is a significant explanatory vari- the estimated value of the parameter y is greater than
able of crack initiation at one percent significance one, indicating an increasing hazard rate with load
level. Furthermore,it can be seen that the repetitions.
The Survival function and the cumulative hazard

65
 Table 2 Parameter Estimates for Model 2 Table 3 RMSE of the Mean Predicted Axle Load Repetitions to

crack initiation

model. The results are shown in Table 3. The en-

tries in the table are the RMSE of the predicted
mean time to cracking in units of axle load repeti-
tions. It can be seen that Model 1 achieves about
20% improvement in the prediction accuracy over
the original AASHO model, while Model 2 is 50%
more accurate than the AASHO model. In other
words, for this data set, about 20% improvement in
prediction accuracy can be attributed to the contri-
bution of rigorous statistical method (stochastic du-
ration techniques) while about 30% improvement
can be attributed to the contribution of an improved
model specification.

5. CONCLUSION

In this study, an analysis of the pavement cracking

initiation data collected during the AASHO Road
Figure 1 Survival Function
Test was conducted. This analysis is based on the
use of probabilistic duration modeling techniques.
Duration techniques enable the stochastic nature of
pavement failure time to be evaluated as well as
censored data to be incorporated in the statistical
estimation of the model parameters. Due to the na-
ture of pavement crack initiation, the presence of
censored data is almost unavoidable and not ac-
counting for such data would produce biased model
parameters.
The main advantages that distinguish this stochas-
tic duration model from the original AASHO model
are as follows. First, the duration model explicitly
recognizes the stochastic variations in the pavement
Figure 2 Cumulative Hazard Function
cracking initiation process. Second, the stochastic
duration model accounts for the fact that some of
function computed at the means of the explanatory
the data are censored. Third, all the parameters of
variables are shown in Figure 1 and Figure 2. . The
the model were estimated simultaneously. There-
unit of duration at the x-axis is number of accumu-
fore, this model is statistically more efficient than
lated load repetitions.
the AASHO model. Fourth, our specification was
The prediction accuracies of the three models
more realistic than that used in the AASHO model,
(AASHO, Model 1 and Model 2), for the sample of in that it did not assume that the effects of axle type
observations used in estimating the parameters,
and load were separate and additive. Finally, the
were compared by computing the Root Mean
predictions obtained with our hazard rate model,
Squared Error (RMSE) of the mean axle load repeti-
using our improved specification, were about 50%
tions until crack initiation, as predicted with each
more accurate than those obtained using the original

66
AASHO model. (PCIR), Final Report, 12 volumes, Empresa Brasileira de
Planejamento de Transportes (GEIPOT), Ministry of Trans-
port, Brasilia, 1982.
ACKNOWLEDGMENT: The authors would like 5) Paterson, W.D.O.; Road Deterioration and Maintenance
to acknowledge the assistance of Dr. Marshall Effects, The Highway and Maintenance Standard Series,
Thompson in obtaining the original AASHO road Johns Hopkins University Press, 1987.
test data. 6) Madanat, S., Bulusu, S. and Mahmoud, A.; Estimation of
infrastructure distress initiation and progression models,
ASCE Journal of Infrastructure Systems,Vol. 1, No 3, 1995.
REFERENCE 7) Madanat, S. and Shin, H.C.; Development of distress pro-
1) HRB: The AASHO Road Test - Report 5- Pavement Re- gression models using panel data sets of in-service pave-
search, Highway Research Board Special Report 61E, Pub- ments, Transportation Research Record 1643, TRB, Na-
lication No. 954, National Research Council, 1962. tional Research Council, 1998.
2) Greene, W. H.; Econometric Analysis, Third Edition, Pren- 8) Kalbfleisch, J. D. and Prentice, R. L.; The Statistical Analy-
tice-Hall, 1997. sis of Failure TimeData, John Wiley and Sons, 1980.
3) Queiroz, C.A. V.; Performance Prediction Models for Pave- 9) Meeker, W. Q. and Escobar, L.A.; Statistical Methodsfor
ment Management in Brazil, Ph. D. dissertation, University ReliabilityData, John Wiley and Sons, 1998.
of Texas, Austin, TX, 1981.
4) GEIPOT: Research on the Interrelationships between Costs (Received November 13, 2002)
of Highway Construction, Maintenance and Utilization

67