You are on page 1of 21

The Estimation of an Origin-Destination Matrix from Traffic Counts

Author(s): M. G. H. BELL
Source: Transportation Science, Vol. 17, No. 2 (May 1983), pp. 198-217
Published by: INFORMS
Stable URL: http://www.jstor.org/stable/25768090
Accessed: 15-07-2015 05:54 UTC

Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at http://www.jstor.org/page/
info/about/policies/terms.jsp
JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of content
in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms of scholarship.
For more information about JSTOR, please contact support@jstor.org.

INFORMS is collaborating with JSTOR to digitize, preserve and extend access to Transportation Science.

http://www.jstor.org

This content downloaded from 129.78.139.29 on Wed, 15 Jul 2015 05:54:34 UTC
All use subject to JSTOR Terms and Conditions

The Estimation of an Origin


Destination Matrix fromTraffic Counts
M. G. H. BELL
University

College

London,

London

that will under certain


is described
this paper, a model
matrix
circumstances yield the most likely origin-destination
link
volumes.
which is consistent with measurements
of
traffic
Given knowledge of theproportionate
usage of each link by the
scale
the model parameters
traffic between each zone pair,
each element of a prior estimate of the origin-destination
matrix up or down so that the measurements
of link traffic
volumes are reproduced. The fitted values are shown by ex
In

of uniform scaling to
ample to be invariant to the application
is out
the prior estimates. A Newton model fitting procedure
are
volumes
lined. In reality, measurements
link
of
traffic
random variables and the proportionate
links
usage of
by the
zone
not
each
is
known
with
certainty. An
traffic for
pair
the
variances
and
covariances
is
derived
of the
expression
for
terms
the
values
in
the
variances
and
logarithms of
fitted
of
link
the
covariances
the
measured
volumes,
propor
of
taking
tionate usage of links as given. The expression permits
the
calculation
of asymmetric confidence intervals for the elements
of the fitted origin-destination matrix.

INTRODUCTION

here has recently been considerable interest inmethods for estimat


(O-D) matrix (sometimes referred
ing the most likely origin-destination
to as a trip table) from measurements
of traffic volumes on the links in
a network. Although it is possible to obtain O-D information directly by
surveying tripmakers
(using, for example, household, workplace or road
X.

interviews), this tends to be both costly and labor intensive. By


of traffic volumes on links in a network can be
contrast, measurements
made relatively inexpensively and, when automatic traffic counters are
side

198

Science
Transportation
Vol. 17,No. 2,May 1983

0041-1655/83/1702-0198 $01.25
1983 Operations Research Society ofAmerica

This content downloaded from 129.78.139.29 on Wed, 15 Jul 2015 05:54:34 UTC
All use subject to JSTOR Terms and Conditions

0-d matrix

estimation

from

traffic

counts

199

it is not necessary
(as it is,
used, very little labor is required. Moreover,
for example, in the case of a roadside interview) to stop vehicles, so the
associated risk of causing a disruption to traffic (possibly also resulting in
a modification of its distribution) is avoided.
to the inference of an O-D matrix from meas
Some early approaches
urements of link traffic volumes have been reviewed by Willumsen.[1]
and Nguyen[2] have described an approach based on equilib
Jornsten
rium traffic assignment which subsequently formed the basis of a program

of Transport.[3]
Van Zuylen and
developed by the U.S. Department
have developed closely related models for the estimation
Willumsen[4]
of an O-D matrix, Van Zuylen using information minimization
and
The associated problem of inferring
Willumsen
entropy maximization.
of
turning movements within an isolated junction given measurements
traffic volumes on the approach roads has also received attention (Beil,[5]

et al.[7]). However,
this problem is
Hauer
Keller,[6]
an
is it generally
since
for
isolated
different
junction
only
qualitatively
true that flows entering and leaving the network can be measured
by
traffic counts.
A further model for the inference of an O-D matrix from link traffic
in order to overcome
volumes, originally suggested by Van Zuylen[8]
anomalies encountered with earlier models/1'4'9] is given a probabilistic

Cremer

and

derivation. This new model, unlike the earlier models,


generates an
of
estimated O-D matrix that is always invariant to the application
uniform scaling to the prior estimates.
In common with the earlier models, a matrix of route choice proportions
is required to be known. In practice, such a matrix may be hard to obtain.
In the absence of direct observations, an assignment algorithm will have

to be used. This problem has been discussed elsewhere.[M]


Making use of the matrix of route choice proportions, some properties
of flow within a complex network, namely linear dependency and corre
are used to reduce the number
lation, are examined. Linear dependencies
of model parameters to be fitted. A Newton model fitting procedure is
of the
described, and some results concerned with the convergence
and
variances
are
for
the
An
approximate
expression
given.
procedure
covariances of the logarithms of the fitted values (the estimated origin
is derived assuming that route choice propor
destination movements)
tions are fixed but allowing link volumes to be random variables. The
expression takes correlations between link volumes into account.

Following Potts and Oliver,[10] the transport network is characterized


or undirected
links.
by a graph of nodes connected by either directed
are
or
attract
areas
traffic
which
represented by
generate
Geographical
zone centroids, which are connected to the nodes of the graph by centroid

This content downloaded from 129.78.139.29 on Wed, 15 Jul 2015 05:54:34 UTC
All use subject to JSTOR Terms and Conditions

200

/ m. g. h. bell

connectors. Nodes which are not connected in this way to zone centroids
are referred to as intermediate nodes. The
level of resolution of the
representation of the transport network can be adjusted to fit the avail
at
able data. For example, when information about turning movements
can each be represented by a link
junctions is available, these movements

in the graph. It is therefore possible to associate with each movement


within the transport network about which there is some information a
particular link in the corresponding graph.

2. LINEAR DEPENDENCY

IN COMPLEX NETWORKS

In this

paper,
steady state traffic flows are considered,
effectively
the
time
dimension from the problem. By observing the prop
removing
agation over the network of fluctuations in the levels of link flows, it is
possible tomake additional inferences about the nature of the distribution
of traffic. Such inferences can be used to further constrain the set of
feasible O-D matrices.
For example, Cremer and Keller[6] suggest a
method for estimating turning movements
at an intersection by correlat
ing variations in the flow of traffic on the links into and out of it,making
allowance for the time required for vehicles to pass between entrance and
exit points ofmeasurement.
It is assumed here that Kirchoff's law applies within the network. In
this context, Kirchoff's law asserts that traffic does not appear or disap
law
pear at nodes or in links. Under steady state conditions Kirchoff's
implies that the total flow into a link or node is equal to the total flow
out. These conservation relationships reduce the degrees of freedom for
link volumes, since some are linearly dependent on others.
Kirchoff s law has implications at two levels (see Willumsen[11]).
At
the level of the node, conservation requires that there be one linearly
link volume for each intermediate node. At a higher level,
dependent
conservation of flows on routes may lead to additional linear dependen
cies. For example, if two directed links lie on one and only one route,
then Kirchoff's law requires that the flow on each be equal, even though
the two links may not be connected to a common intermediate node.
Define
=
tj the quantity of traffic between a pair of zones designated by j
(there being J zone pairs)
Vi= the volume of traffic on measured
link i (there being Jmeasured
links); it is not assumed that all links in the network are measured
= the
Pij
proportion of the traffic between zone pair j that uses meas
ured link i.

This content downloaded from 129.78.139.29 on Wed, 15 Jul 2015 05:54:34 UTC
All use subject to JSTOR Terms and Conditions

o-d matrix
Using

from

estimation

traffic

counts

201

the vector and matrix notation


tT=[tu

vT=

...,tj]

[vu
Pn,

?the available

network

*,P\j

:
_Piu

of simultaneous

Vi]

:
--,Pu

information can be expressed

as the following set

equations

v = Et (1)
The number of linearly independent link volumes is identically equal to
the rank ofmatrix E
It is assumed throughout this paper that the matrix of route assignment
proportions, ?, is known. In practice, this may have to be obtained by
making some assumption about route choice behavior. It should be noted,
however, that it is not in general sufficient to assume that the network is

user optimized (i.e. that Wardrop


equilibrium obtains) since the equilib
rium assignment model will not usually yield unique estimates for route
choice proportions (see Willumsen[11]).
Given a set of observed link volumes, y*, and the matrix of route
assignment proportions, the set of feasible solutions for t is defined by
relationship (1). There are a number of possibilities. First, there may be
no feasible solution. When this is the case, at least one row ofmatrix
g is
linearly dependent and the equations are said to be inconsistent (see
Hadley[12]).
Although the inconsistency may be due to the violation of
the
in the network, this is not necessarily
Kirchoffs
law somewhere
an
route
incorrect hypothesized
it may be the result of
explanation;
assignment. Second, feasible solutions for tmay exist but none of them
may be strictly nonnegative
(negative values for the elements of t have
no meaning). This situation may arise if there is a divergence between
route assignment. Third, while strictly nonne
actual and hypothesized
gative feasible solutions may exist, none of them may be strictly positive.
This may reflect an underlying truth (namely that tj does in fact equal
zero for some j), but is more likely to be the result of a divergence
between hypothesized and actual route assignment.
If any feasible solutions exist, then the rank ofmatrix P is equal to the

rank of the augmented matrix of the system of equations


(the matrix [P,
y*] formed by appending vector y*, as an additional column, to matrix
all the rows of matrix P are linearly
P)
(see Willumsen[11]). When

This content downloaded from 129.78.139.29 on Wed, 15 Jul 2015 05:54:34 UTC
All use subject to JSTOR Terms and Conditions

202

m. g. h. bell

independent, the rank of P is equal to the rank of the augmented matrix


and therefore at least one feasible solution exists.
The procedures described in subsequent sections require the identifi
cation of a set of linearly independent link volumes. This is equivalent to
determining a set of rows of matrix P which may be regarded as being
linearly dependent on the other rows. There are a number of suitable

ortho
for doing this, of which two (Gram-Schmidt
numerical methods
elimination) are described here.
gonalization and Gaussian
for the sequential search
The use of Gram-Schmidt
orthogonalization
for linear dependency among the rows ofmatrix P has been independently
and Branston.[13]
suggested by the author and Van Zuylen
Let pi be the ith row of P. On the rth step of the procedure
gr=Pr~

??l

r = 2, (2)

(giTPr)gi/(giTgi),

where gi is equal to pi. Vectors gr and gs (r^ s) are orthogonal (i.e. g?gs
=
2 can be written as
be equal to {giTpr)/\giTgi). Equation
0). Let
gr=Pr~

JjZl

airgi

Pr

?*=1

T =

a'irPi,

2,

,U

(3)

where alr are new coefficients formed from ajr (j' < r). Ifgr = 0, then the
rth row of matrix P is a linear combination of the preceding rows (with
coefficients

, ar-i,r)

a[r,

and

vr = SPi1 a'irVi.
In fact, gr

0 is both a necessary

and

(4)

sufficient condition

for linear

dependency.
In the case of Gaussian
elimination, matrix P is reduced by the
successive elimination of the variables in (1) to a new matrix with zeros
below the principal diagonal. If the rth row of this new matrix contains
only zeros then the corresponding row ofmatrix P is linearly dependent
on the preceding rows. As before, the reduction of a row to zero elements
only is both a necessary and sufficient condition for linear dependency.

Further

information about

and Gaussian

the Gram-Schmidt
process
orthogonalization
be found in Hadley[12] or Isaacson
and

elimination may

Keller.[14]

follows, and which constitutes a key component of


that the matrix PPT could equally well be analyzed
for linear dependency. Typically,
the number of zone pairs, J, greatly
exceeds the number of measured
links, J, so the matrix PPT, which is /
x J,may be significantly smaller than matrix P, which is J x J.
The result which
later proofs, asserts

1. The rth row of the matrix PPT


is linearly dependent on the
rows
row
and
the
rth
preceding
if
only if
of matrix P is linearly
Result

This content downloaded from 129.78.139.29 on Wed, 15 Jul 2015 05:54:34 UTC
All use subject to JSTOR Terms and Conditions

O-D MATRIX ESTIMATION


on its preceding
dependent
involved are the same.

FROM TRAFFIC COUNTS

rows. Moreover,

Proof. Let P have rows equal to pi,


,zi. Note that zT = pfPT.
?i,
If
(i) Necessity.

,pi and PPT

pr =^l\iPj
where Ai,

,Ar-i are not all equal

the linear

203

combinations

have rows equal

to

(5)

to zero, then
(6)

Pr"PT=%=l\ipTPT

(bypost-multiplying
byPT). Thus
Zr=

hSi (7)

(ii) Sufficiency.If
Sr= ??l\i3i
where, as before, Ai,

(8)

,Ar-i are not all zero, then

P?PT

(9)

Z=l\ipTPT

or
u'P'^fr

where

(10)
(11)

u=pr-^Zl\iPi.

Vector u is both contained in the space spanned by pi,


,pr (see (11))
is
and orthogonal topi,
u
(see
(10)).
->pr
Therefore,
equal to 0 and

pr = 2d1 ^Pi-

(12)

An immediate implication of this result is that the rank of PPT is equal


to the rank of P.
The matrix PPT has an interesting interpretation. Its (i, y)th element,
equal to pTPj, has the value zero if there is no route which carries some
traffic and uses both links i and y, and is positive otherwise. From (1), we
have

that

Y(y)

PV(t)PT (13)

matrices for the elements of


where V(y) and V(t) are variance-covariance
t
of
If
elements
the
and
? (the trip table) are independent
y
respectively.

This content downloaded from 129.78.139.29 on Wed, 15 Jul 2015 05:54:34 UTC
All use subject to JSTOR Terms and Conditions

204

M. G. H. BELL

random variables with variance

a2, then

V(v) = a2PPT. (14)


the covariance

Thus

Vi and vj is equal

between

to o2pTpj, or alternatively

theircorrelationis equal to (piTpj)/( \pi\\


Pj \)where \pk\=^(PkTPk).
3. model

Define

additional

formulation

notation

=
tj a prior estimate of tj
= the
qj
probability that a trip is between

zone pair j given that a trip

occurs.

Also

let
...,?,]

F=[fi,

fi, ...,ln^]

,?/]-On

?T=

,ln?/]

,&]-Pnfi,

=[L

ki,

A strictly positive solution for t in Equation


1, if it exists, is unlikely to
be uniquely defined because
in practice the number of unknowns, J,
usually greatly exceeds the number of linearly independent equations. In
this situation, a strictly convex, single-valued objective function, repre
sented by /(?), can be introduced in order to obtain a unique solution.
The use of different forms of objective
function in the constrained

minimization

problem

Mmtf(t)
yield a family ofmodels,
Willumsen,[4] Willumsen

=
subject to v Pt

(15)

including those considered by Van Zuylen and


and Van Vliet,[9] Beil,[5] and the model derived

here.

From entropy maximizing considerations,


obtained the following objective function

Van Zuylen and Willumsen[4]

m-ZjtAMtj/tj)-!).
The

(16)

resulting model

tj^tjUAxW,

7 = 1,

(17)

has one parameter, x?> for each link volume which scales the prior
estimate up or down so that the measured
link volumes are reproduced.
Willumsen
and Van VlietE9] suggest that (16) constitutes an appropriate
measure
of the difference between t and t. If a constant term,
?y ij, is

This content downloaded from 129.78.139.29 on Wed, 15 Jul 2015 05:54:34 UTC
All use subject to JSTOR Terms and Conditions

estimation

o-d matrix

from

counts

traffic

205

added to (16), the resulting function is the sum of J strictly convex


functions of one variable, tj, each of which takes a value zero if tj= tj and
is positive otherwise. This interpretation is particularly attractive when
the model is used to update a previous estimate of t (perhaps obtained by
a roadside, household or workplace
interview) in the light ofmore recent

measurements

Fitted values

of traffic volumes.
t formodel (17) may be obtained by solving the equations

i= l,...,/

Vi= 2> PijijUk(Xk)Pk\

(18)

adjustment algorithm, a special case of


perhaps by a multiproportional
the more general balancing method studied by Bregman
(Lamond and
or by a Newton method. The multiproportional
Stewart[15]),
adjustment
as
a
to
ME2
in
has
been
referred
embodied
(Maximum
program
algorithm

It can be shown that these


(Willumsen[16]).
Entropy Matrix Estimation)
fitted values are not in general invariant to the application of uniform
scaling to the prior estimates. This is a disadvantage when there has been
a significant growth (or decline) in the total quantity of traffic between
the time that the prior estimates were obtained and the time that the
traffic counts were made.
the joint probability of observ
Here a model is derived that maximizes
to
the
t
constraints
(1), given the
imposed by relationship
ing subject
are
distrib
that
and
the
trips
multinomially
assumption
probabilities q

uted. If f(t) is this jointprobability,then


/a)

((2y^)Vn>(/!)n>(^ (19)

and

F(t) = In f(t) = ln(?, tj)l 2> In tjl+


Taking

d In xl/dx to be equal
(dF(t)/dtk)

The

maximization

following Lagrangian

ln(?,

of F(t)

tj)

tj Inqj.

(20)

to In x, we obtain
-

with

In tk+

In qky

respect

k=

1,

to relationship

. .,J.

(21)

(1) yields

the

conditions

for

equation

=
L(t, X) F(t) +XT(v- Pt) (22)
where X is a vector of Lagrange multipliers.
a maximum are

The

necessary

=
= (23)
dL/dt (dF(t)/dt) XTP 0T
,dL/dtj] and dF/dt is similarly
dL/dt is the row vector [dL/dti,
is achieved when
defined. The following result confirms that a maximum
(23) is satisfied.

where

This content downloaded from 129.78.139.29 on Wed, 15 Jul 2015 05:54:34 UTC
All use subject to JSTOR Terms and Conditions

206

m. g. h. bell

2. The Hessian
(subject to the constraint

the matrix
for F(t), namely
[d2F/dtidtj],
v = Pt) negative definite for strictly positive

Result

The Hessian

Proof

of the objective

function has the form

if i=j
otherwis
& otherwise.

[
The

form

quadratic

Jjj dtddF/dtidtj)dtj

=
=
=

Let

dtj
2? dU 1j (d2F/dtidtj)
tk) (dti/U))
L dft((2y

(24)

?*02/S*fc-L-(*i2A)

=
=
<j>i dti/lk dtk and ft

definiteness, we require

4. For negative

(25)

/(?=2i*d-(Wft))<0.
To

is
t

investigate /(#) formaxima,

consider the Lagrange

equation

(26)
where
mum

co is a Lagrange

The

multiplier.

for a maxi

conditions

necessary

are

dL/ddi

(<t>i2/62)-w

= 0

for all i

=
=
implying that <f>i \/a>ft,and therefore that w

Furthermore,

since ft >

1, since

0 and

if/=j

df/ddi^-,
the Hessian
a maximum

(27)

otherwise

for /*(#)is negative definite. Thus, the sufficient condition for


=
is satisfied. At this maximum,
ft, implying that

=
dti/lkdtk

tk

forall /. (28)

the maximum
value for f(6) is zero. By differentiating the
Moreover,
constraint equations, we obtain the relationship dv = Pdt = 0. Thus, at
the maximum of f(6), t = m dt (where 7r 0 is a constant of proportion
= 0. This violates the
0.
ality) and therefore Pt
constraints, since v
Hence f(0) < 0 and consequently the Hessian forF(t)
is negative definite.

From (21) and (23)we obtain


]ntk

ln(ZJtj) +

]nqk-li\iPik,

k=

1,

...

,J

(29)

This content downloaded from 129.78.139.29 on Wed, 15 Jul 2015 05:54:34 UTC
All use subject to JSTOR Terms and Conditions

from

estimation

o-d matrix

traffic

counts

207

and thus
tk=

k=

2< hp*),

(2y tj)qkexp(-

1,

. ,J.

(30)

In practice, the probabilities q are unknown. It is assumed


they can be derived from the prior estimates as follows
qk

? =

ik/ljij,

here

that

(31)

1,...,J.

gives the following model

This

tk= rtkexp(-

k=

Xipik),

,J

1,

(32)

where

T= Zjtj/Zjtj.
If

(33)
becomes

to In x* for all i, then the model

Xi is equal

k=

tk= rtkUi(xi)Pik,

,J.

1,

(34)

differs from the model


given by (17) only in so far as a scale
t
is included. It is shown by example in Section 6 that the
parameter
fitted values yielded by the new model are invariant to the application of

This

uniform scaling to the prior estimates.


4. MODEL
is equal
If *//

to In t, the model
tk=

There

are now / +

FITTING

as expressed

tkexp(\p 2* XiPik),
1 parameters

in (32) may be rewritten as


k=

(\p,Ai,

1,

,J.

(35)

, A/) to be determined

by J

equations
Vi= ZjPijtj

lLjPijije*Vty

=
2* hkPkj), i

1,

,/.

(36)

is equal to the
the number of linearly independent equations
However,
rank of matrix P which is in turn equal to the number of linearly
link volumes. Let L be the number of linearly
independent measured
link volumes.
independent measured
link
loss of generality, let us suppose that the firstL measured
Without
volumes are linearly independent. Hence
Vi= IjPijtj,

/=1,

...,L

(37)

are linearly independent equations. The set of parameters can be corre


,Xl) to be determined. There are
spondingly reduced, leaving (^, Ai,
now L + 1 parameters. The following additional relationship

Zj ij

=
=
= r-1
Ij tj exp(-^) Ij tj lj tjexv(- It XiPij)

is therefore required

in order to obtain a unique

solution.

This content downloaded from 129.78.139.29 on Wed, 15 Jul 2015 05:54:34 UTC
All use subject to JSTOR Terms and Conditions

(38)

208

M. G. H. BELL

Define

further notation
...,

vu

yT=[%jij,

(39)

vL]

and

Pll,

,PlJ

R =

(40)
Pli,

The L +

1 constraint equations

,Plj

can now be written as


y

= Rt.

(41)

If the rank ofmatrix R is equal to that formatrix P, then these equations


will not be linearly independent. Additionally define
MT

Ai,

...

(42)

AL]

and

i,
2

(43)
Pli,

thus enabling

the model

,Plj

to be written as

t= D exp(2TM)
where D

,?/) and exp(-) applies


diag(?i,
Alternatively, since the model is log-linear

(44)
to each element

of STy.

(45)
There are a number of numerical solution procedures that may be used
to derive values for the parameters,
so that the model estimates for y
agree with predetermined values, y*. One such is the Newton procedure
tf

<#?+!)
=

_
(y<">)-l(y?>
y*)
tf(n)

(46)

where
y<?>

Rt^

exp^V*"')

(47)

This content downloaded from 129.78.139.29 on Wed, 15 Jul 2015 05:54:34 UTC
All use subject to JSTOR Terms and Conditions

o-d matrix

estimation

from traffic

counts

209

and
,
^,<nl/^Ki

ldyi<n,/aMi<n),

(48)

(superscript n denotes

[dy^/d^,

,dytU/diitl^

the iteration number). Moreover,

ifi = j = 1 (seeEquation 38)

2*(5yi(",/a^<"))-(^*(n)/W">)

rexp^^)^/^,^"'
= ^

if j =
if i>
if i >

exp(-/ii<n))i;^,1
=
vt\
2*Pi-i,ktkM
. Pi-i,kPj~i,ktkM

1 and; >
l andy =
i and7 >

1
1
1

Thus
0

| exp(-/il(n))l;i,n),

vi

iLkPiktk

, expf-ju^W"1

>?fcPlkPLktk

(n)

(49)
VLin)\lkPLkPlktkin\
0

...

,lkPlktk{n)

jexp(-/x1(n))i;(n)T
;(n)|pD(n)pT

The following two results prove that matrix J(n) is nonsingular and that
In order for the Newton procedure to
therefore J{n)~l is determinate.
a
of it, it is sufficient that
in
to
the
solution
converge
neighborhood
fx*
and Rheinboldt[17]).
J(n) be continuous and nonsingular (see Ortega
Result

3. The matrix PD(n)PT


t.

is positive

definite for strictly positive

prior estimates
Proof
positive

It can be verified from (44) that given positive t the model yields
t.Thus D(n\ which is equal to diag(?(n), ..., tjn)),

fitted values

This content downloaded from 129.78.139.29 on Wed, 15 Jul 2015 05:54:34 UTC
All use subject to JSTOR Terms and Conditions

210

m. g. h. bell

is nonsingular. Matrix PD^PT


may be reformulated as QQT where Q is
. .~~V?7^).
to
and
Since
7z)(/1)~is
equal
equal to diag(V^,
D{n) is nonsingular, VZ)(/l) is also nonsingular and therefore the rank of
Q is equal to the rankof P. However, Result 1 implies that the rank of
is equal to the rank of Q, so the rank of PD(n)PT
is equal to the
QQT
rank of P. Thus, since the rank of P is equal to the number of its rows,
which is L X L, must be nonsingular. Since it is
L, the matrix PD(n)PT,
in
the
form
expressible
QQT, itmust also be positive definite.

Result

4. The Jacobian

J{n\ given

by (49), is nonsingular.

Proof. Matrix PD{n)PT, which is L X L, has rank L (Result 3). Hence


columns 2 to L 4- 1 of J(n) are linearly independent. Therefore, ifJ{n) is
singular, then
0

where

the elements

PD(n)pT

of f are not all zero. Thus


=
exp(-iu1(n))i;(n)rf

since exp(?

[iin)) >

(5i)

=?(4

PD(n)PTi;
Hence,

(50)

exp(-/ii(n))i;(n)T

?(n)

?TPDin)PT?=

?Tvin)

which contradicts the result that PD{n)PT


Therefore J{n) must be nonsingular.

= 0

(52)

is positive definite

(Result 3).

Results
3 and 4 imply that the number of zone pairs, J, should be
greater than or equal to L, the number of rows of matrix P. If not, the
number of columns ofP must be less than L implying that the rank ofP,
T
and therefore PD {n)PT, is less than L. Since PD {n)P is no longer positive
definite, J(n) is no longer necessarily nonsingular.
5. APPROXIMATE

VARIANCES

AND COVARIANCES
VALUES

FOR THE FITTED

important
consideration for the practitioner is the robustness of the
fitted values, t (" refers to estimated quantities), given that the measure
ments for y are random variables and that the set of assignment propor
to the
tions, P, are not known with certainty. As a first approximation
is
it
assumed
here
that
the
are
known
problem,
assignment proportions
with certainty. An expression is derived for the variances and covariances
of the logs of the fitted values, |, in terms of the variances and covariances
An

This content downloaded from 129.78.139.29 on Wed, 15 Jul 2015 05:54:34 UTC
All use subject to JSTOR Terms and Conditions

estimation

o-d matrix
for the mean measurments

from

traffic
v

of link traffic volumes,

counts

211

(* refers to observed

values).

Variances and covariances are obtained for ?, rather than


is conventional to assume that the elements of the former,
the latter, are approximately normally distributed. Standard
permit the construction of asymmetric confidence intervals

it
t,because
rather than
errors for ?
for t which

exclude the possibility of negative (or zero) values. Additional advantages


of this approach are that the standard errors for ?may be interpreted as
approximate coefficients of variation for t, that no specific error structure
hypothesis is assumed, and that correlations between traffic volumes are
taken into account.
The expression is based
For x close to 17

on a truncated Taylor

fix) ? f(V) + [df/dx]^(x


=

Let E{x)

7],where E{x)

denotes

series approximation.

3). (53)

the expectation

of x, and letmatrix

Z=[df/dx]x^ (54)
<=*
and the
be the Jacobian of the transformation. Hence E{f(x)}
f(jj)
matrix for f(x) is given by the following approxima
variance-covariance
tion

Y(f(x))

E{(f(x) -f(v))(f(x)
?

E{Z(x

f(v))T)
-

rj)(x

rj)TZT)

(55)
ZV(x)ZT.

matrix
Two stages are required in order to obtain a variance-covariance
matrix fory*. First a variance-covariance
for I from a variance-covariance
matrix is
then from this a variance-covariance
matrix is obtained for
derived for |. Since the first element of vector y*, equal to
tj, is
matrix for
invariant, the first row and column of the variance-covariance
*
y have zero elements only.
approximation

Using

(55)

V{y*)*JV(MT
where
equal

(56)

the Jacobian matrix J is as defined in Equation


to jlLSince J is nonsingular (Result 4)

Y(&
Following

Equation

J~lViy*(dT)~l

45, the fitted model

52 when

J-'Ky*)^-1)7can be written as

This content downloaded from 129.78.139.29 on Wed, 15 Jul 2015 05:54:34 UTC
All use subject to JSTOR Terms and Conditions

/x(n)is

(57)

212

m. g. h. bell

Thus

(59)
Hence

Y(l)

STV((i)S

(60)

^j-'viynid'Ts.

Standard errors for ? can be obtained by taking the square root of the
of matrix V(?). Since maximum
elements on the principal diagonal
likelihood estimators are normally distributed for large samples (see, for
confidence intervals for ?may be obtained from
example, Edwards[18]),

the unit normal distribution.


In practice, a variance-covariance
matrix for link traffic volumes may
on each link. For this
be obtained by making repeated measurements
purpose, the survey period can be divided into intervals. In each interval,
the traffic can be counted on a subset of links with linearly independent

traffic volumes. The traffic count variances for each link, and the covar
iances for each pair of links, can then be calculated.
The extent of the correlation between link volumes will depend on the
size of the chosen interval ofmeasurement.
Since trips are not instanta
neous, the correlations between link traffic volumes are in reality lagged.
It is therefore desirable to select as long an interval of measurement
as
so
as
to
minimize
the
of
the
possible
significance
lags.
It is assumed here that correlations between link volumes are instan
taneous. Let Xij be the count for the jth interval on link i expressed as a
deviation from the mean count for link i.
Define
0,
, Xm

X =

(61)
Xlu

Xln

where L is the number of linearly independent measured


the number ofmeasurements made for each link. Thus

V(y*)

Substituting into approximation


an equality, we obtain

Y(o

(l/N(N-l))(XXT).

links and N

(62)

(60), and treating the approximation

srj-'a/NiN-

D)(xxT)(d~ys

is

as

(63)

This content downloaded from 129.78.139.29 on Wed, 15 Jul 2015 05:54:34 UTC
All use subject to JSTOR Terms and Conditions

o-d matrix

estimation

from

traffic

zone

1. Example

Fig.

Link -

1
2
3
4
5

0
1110
0.7
0.3
0.3

213

of Assignment

AC AB

network.

TABLE
Matrix

counts

Proportions
Zone Pair

BC

10

0
1
0

0
1
0

CB

0
0
1

BACA

11
0
0
0
1

0
0
0
0

6. AN EXAMPLE
A computer
the methods

program has been developed by the author to implement


described
in the preceding sections. In this section, the
results yielded by a simple network, depicted in Figure 1, are considered.
There are six zone pairs, five links and one intermediate node. The
assumed matrix of assignment proportions, which is not based on an all
or-nothing assignment (as evidenced by the existence of some noninteger
values), is shown in Table I.
counts and their mean values are shown in Table II, and
Hypothetical
III. Taken
in Table
and covariances
the corresponding variances
by
law. Since the
II conform to Kirchoffs
columns, the values in Table
also conform to Kirchoffs
law, the system of constraint
is
consistent.
equations
In this case, matrix P has rank 4, the linearly dependent row being due
to flow conservation at the intermediate node. Either v2, v3 or v4may be
elimination method
treated as being linearly dependent. The Gaussian

mean

values

This content downloaded from 129.78.139.29 on Wed, 15 Jul 2015 05:54:34 UTC
All use subject to JSTOR Terms and Conditions

214

M. G. H. BELL

II

TABLE
Hypothetical

and Their Mean

14
13

13 126
21 227
13
8
135 12

23
17
10
7
14

13
20
11
9
11

TABLE
Variances
Link

Values

Measurement

Link
1

Counts

and Covariances

Mean

21
19.2
19
20.8
10.8 6
13
13.0
15

Value

10.0

III
for Hypothetical
Link

Counts

1 35.0
2
7.0

14.0

3 -1.2

7.9

4 8.3
3.3

9.7

6.3
5

-3.0

-1.8

8.0

-3.5

5.0

2.5

is preferred in practice to the Gram


for detecting linear dependency
Schmidt orthogonalization process because, by applying the procedure to
the augmented matrix of the system, it is simple to identify linearly
dependent equations and simultaneously to check their consistency. The
Gaussian
elimination subroutine of the computer program suggests that
link 4 is linearly dependent.
Taking uniform prior estimates equal to one for all zone pairs, the
parameter estimates are

?i

=
=
ju2 Xi
=
Jl'3

fa
?5
The fitted values
Origin

X'2

1.89
0.48
?1.17

=
A3

=
A4 -0.73.

3.19

and their 95% confidence


Destination

Estimate
AB
15.43

intervals are:
95% confidence

region

11.98 to 19.87

AC
2.06

1.13 to

BC
3.32

1.94 to

5.67

CB
3.20

2.24 to

4.59

CA
5.17

3.93 to

6.79

BA
10.72

3.75

7.37 to 15.58.

This content downloaded from 129.78.139.29 on Wed, 15 Jul 2015 05:54:34 UTC
All use subject to JSTOR Terms and Conditions

O-D MATRIX ESTIMATION


If the prior estimates
estimates are

are multiplied

FROM TRAFFIC COUNTS

215

by a factor of 10, then the parameter

= - 0.41
|Si
=
0.48
/x2
= - 1-17
jia
fi4
ju5

=
=

3.19
-

0.73.

to allow for the change in the scaling applied to the


?1 adjusts
prior estimates. The fitted values and their confidence intervals remain
unchanged.
Five of the fitted values are sensitive to a change to nonuniform prior
from zone JBto
the prior estimate for the movement
estimates. When
zone A is twice the prior estimates for all other zone pairs (suggesting
that a trip from B to A is twice as likely as a trip between any other pair

Parameter

of zones, given that a


confidence intervals are

Origin

Destination

trip occurs)

the fitted values

Estimate
15.43
AB

C
A
BC
CB
A
C

2.64
2.73
4.12
4.25
12.22
BA

and

their 95%

95% confidence

region

11.98 to 19.87

1.49 to
1.59 to
2.99 to
3.21 to
8.76

4.69
4.70
5.68
5.64

to 17.03

from A to B is determined by
Since (as Table I shows) the movement
the movement on link 3, the corresponding fitted value and its confidence
interval is insensitive to changes in the pattern of the prior estimates.
matrix for the logarithms of the fitted values,
The variance-covariance
obtained when the prior estimates are uniformly equal to one, is given in
Table IV. It is clear from (63) that this matrix is symmetric, so only the
lower triangular portion is given.
Since the movement fromA to B is equal to 1.43 times the flow on link
3 (see Table I), the variance of the corresponding fitted value should be
on link 3,
equal to 1.43 squared times the variance of the mean flow
an
for
variance
the
3.96.
However,
program supplies
approximate
namely
the logarithm of the fitted value
IV). Since the
(see Table
fitted
is
value
of
the
deviation of the logarithm
approximately
the coefficient of variation for the fitted value itself, the program
a variance approximately
equal to 4.05, which is of the correct

standard
equal to
indicates

order of

magnitude.

This content downloaded from 129.78.139.29 on Wed, 15 Jul 2015 05:54:34 UTC
All use subject to JSTOR Terms and Conditions

216

m. g. h. bell
iv

table
and Covariances

Variances

for the Logarithms

of the Fitted

Values

Zone Pair

Zone Pair

AC AB

AB

BC

CB

BACA

0.017

AC
BC
CB

-0.025

0.094

-0.018

0.076

0.075

-0.021

0.035

0.019

CA
-0.014

0.016

0.018

BA

0.010

-0.016

0.034
0.018

0.008

-0.021

0.019
0.003

0.036

7. DISCUSSION
is developed
in this paper that yields the most probable set of
A model
O-D movements
that are consistent with a set of link flows if the
probabilities obtained from the prior estimates are correct. In practice,
the prior estimates might be based on an old survey that may not
If this is so, or if there is no basis
correctly reflect current probabilities.
on which to obtain prior estimates (in which case a uniform value has to

be adopted), it seems unlikely that the model will yield the most probable
set of O-D movements
consistent with link flows. As demonstrated
in
Section 6, at least some of the fitted values may be sensitive to changes
in the prior estimates.
for a set of links
In practice, traffic counts are likely to be available
which include some whose volumes are linearly dependent on the volumes
of the others within the set. When
inconsistency arises, it is possible to
remove it by adjusting the observations using a maximum
likelihood
criteria (see Van Zuylen and Willumsen[4]
and Van Zuylen and Bran
itwould
ston[13]), thus incorporating the additional information. However,
in principle be preferable to fit the model directly to the traffic counts.
ACKNOWLEDGMENTS
The

author

is grateful to Dirck Van Vliet, Richard Allsop, Stephen


Holden, Henk Van Zuylen, and especially to Luis Wil
lumsen, for helpful discussions at various stages during the development
of the work, and wishes to express appreciation for the financial support
provided by the Science and Engineering Research Council.

Gallivan,

David

REFERENCES
1. L. G. Willumsen,

"Simplified

Transport

Models

Based

on Traffic

Counts,"

Transportation 10, 257-278 (1981).


2. K. 0. Jornsten and S. Nguyen, "On the Estimation of a Trip Matrix from
Network Data," Centre de Recherche sur les Transports Publication 153,
University ofMontreal, 1979.

This content downloaded from 129.78.139.29 on Wed, 15 Jul 2015 05:54:34 UTC
All use subject to JSTOR Terms and Conditions

O-D MATRIX ESTIMATION


3. U.S.

of Transport,

Department

FROM TRAFFIC COUNTS


of an Origin-Destination

"Estimation

217

Trip

Table Based on Observed Link Volumes and Turning Movements," Federal


Highway Administration, Offices of Research and Development, Traffic
Systems Division, 1980.
4. H. J. Van Zuylen and L. G. Willumsen,
"The Most Likely Trip Matrix
Estimated fromTraffic Counts," Trans. Res. 14B, 281-293 (1980).
5. D.

von Teilstromen
zur Ermittlung
aus
for Dr.-Ing.,
Fakultat
fur Bauingenieur
der Universitat
1979.
Karlsruhe,

"Matrixmodelle

Beil,

thesis

messungen,"
sungswesen

Querschnitts
und Vermes

6. M. Cremer and H. Keller,


"Dynamic Identification of Flows fromTraffic
Counts at Complex Intersections," Proceedings 8th International Sympo
sium on Transportation and Traffic Theory, Toronto University, June
1981.
7. E.

Automatic

and

Pagitsas

E.

Hauer,

T.

B.

Trans.

Counts,"

Res.

Shin,
Rec.

"Estimating
759,1-7

from

Flows

Turning

(1981).

8. H. J.Van Zuylen, "Some Improvements in theEstimation of an O-D Matrix


fromTraffic Counts," Proceedings 8th International Symposium on Trans
portation and Traffic Theory, Toronto University, June 1981.
and D. Van Vliet, "Validation of the ME2 Model for
9. L. G. Willumsen
Matrices
from Traffic Counts," Proceedings 8th Inter
Estimating Trip
national Symposium on Transportation and Traffic Theory, Toronto Uni
versity, June 1981.
10. R.

B.

and

Potts

demic

Press,

R. M.

New

Oliver,
1972.
York,

Flows

in Transportation

Aca

Networks,

"An Entropy Maximising Model for Estimating Trip


11. L. G. Willumsen,
Matrices fromTraffic Counts," thesis forPh.D., University ofLeeds, 1981.
12. G. Hadley,

Linear

from Counts,"

Addison-Wesley,

Algebra,

and D. M. Branston,

13. H. J.Van Zuylen

Transportation

14. E. Isaacson and H. B. Keller,


& Sons, New York, 1966.
15. B.

Lamond

and

N.

F.

Stewart,

Reading,

1979.

Mass.,

"Consistent Link Flow Estimation

Research

(forthcoming).

Analysis ofNumerical Methods,


"Bregman's

Balancing

Method,"

JohnWiley
Trans.

Res. 15B, 239-248 (1981).


16. L. G. Willumsen,
"Estimating a Trip Matrix fromTraffic Counts, Empirical
13th
Annual
Conference of theUniversities Transport Study Group,
Tests,"
Leeds University, January 1981 (unpublished).
17.

J. M.

Ortega

and

W.

C.

Rheinboldt,

Iterative

Solution

of Non-linear

in Several Variables, Academic Press, New York, 1970.


Equations
18. A. W. F. Edwards, Likelihood, Cambridge University Press, New Rochelle,
N.Y.,

1972.

This content downloaded from 129.78.139.29 on Wed, 15 Jul 2015 05:54:34 UTC
All use subject to JSTOR Terms and Conditions