Introduction

The Average Growth Factor Model

We turn now to a second model of trip distribution, the Average Growth
Factor model. The interest of this model lies primarily in two related ideas
Philip A. Viton that will gure again in our discussion of the Gravity Model:

1. The idea of approximate satisfaction of conservation conditions

2. The idea of improving our results through iteration
October 23, 2014

Philip A. Viton CRP/CE 5700() Avg Growth October 23, 2014 1 / 27 Philip A. Viton CRP/CE 5700() Avg Growth October 23, 2014 2 / 27

The Average Growth Factor Model (I) The Average Growth Factor Model (II)

Dene the individual zonal growth rates for any two zones i and j as:

Data requirements: Fi0 = Oi /Oi0

T 0 (observed present-day trip matrix) Fj0 = Oj /Oj0
estimates O = (O1 , O2 , . . . , OZ ) of all future originations by zone.
Then we assume that the growth rate in travel between the two zones
Note that this is more data than was needed for the uniform growth is equal to the average of the two zonal growth rates:
model.
The Average Growth Factor model is based on the intuitive idea that Fi0 + Fj0
Fij0 =
we expect that the growth in interzonal travel between any two zones 2
to be related to the growth in travel in each zone individually.
And we take as an estimate of the future interzonal travel between
zones i and j:
Tij1 = Tij0 Fij0

Philip A. Viton CRP/CE 5700() Avg Growth October 23, 2014 3 / 27 Philip A. Viton CRP/CE 5700() Avg Growth October 23, 2014 4 / 27
Conservation of Origins Error Ratios (I)

Since we are interested in checking T 1 for conservation of origins, this

Suppose we generate a new trip matrix T1 according to this scheme. suggests that we look at the error ratios
How can we assess our answer?
Since we dont know anything about the true future zone-to-zone Oi
Ei1 = i = 1, 2, . . . , Z
Tij s, the only way to do this seems to be via the originations: we Oi1
compare the originations implied by T 1 with the known (ie assumed)
How do we assess these? Clearly, the best possible answer would be
future originations O .
to nd Ei1 = 1 for all i; but this is probably too much to hope for in
We are asking if our estimate T 1 satises conservation of origins. practice.
That is, do the originations implied by T 1 add up to those we know
So we should probably try to say when a particular set of originations
(ie to the vector O )?
(here, the vector Oi1 ) is good enough; equivalently, when the error
ratios are close enough to 1.

Philip A. Viton CRP/CE 5700() Avg Growth October 23, 2014 5 / 27 Philip A. Viton CRP/CE 5700() Avg Growth October 23, 2014 6 / 27

Error Ratios (II) Error Ratios (III)

But where do these convergence criteria come from?

Lets try to formalize this.
The short answer is that thats up to you: you need to decide just
We will dene two convergence criteria L , H . We will say that an
how good you want your answers to be.
error ratio like Ei1 is good enough if
It is important to understand that good refers only to whether the
L Ei1 H originations implied by our candidate trip matrix (T 1 ) agree with the
known/assumed future originations(O ): there is no question of
and that our T 1 is an acceptable answer if comparing the individual elements of a candidate trip matrix with
anything, since theres nothing to compare them to.
L Ei1 H for all i = 1, 2, . . . , Z
In practice, many analysts use one of two sets of convergence criteria:
Note that this is a form of approximate satisfaction of conservation of
(L , H ) = (0.95, 1.05) called a 5% criterion
origins.
(L , H ) = (0.99, 1.01) called a 1% criterion

Philip A. Viton CRP/CE 5700() Avg Growth October 23, 2014 7 / 27 Philip A. Viton CRP/CE 5700() Avg Growth October 23, 2014 8 / 27
Iteration (I) Iteration (II)

What can we do about it? The crucial insight is to note that having
generated T 1 and observing that we do not completely satisfy
Suppose weve done all this: computed the average growth factors conservation of origins, we realize that T 1 can be considered as a new
and generated T 1 according to the Average Growth Factor models starting point, just as T 0 was.
formula. And typically T 1 will be better (in the sense of smaller error ratios)
And suppose we nd that the condition than T 0 was.
So the suggestion is that we simply try again, using T 1 as our new
L Ei1 H for all i = 1, 2, . . . , Z
starting point.
is not satised for all zones i (it may be satised for some of them). Note that if we tried this in the Uniform Growth Factor model, it
wouldnt help, since theres no room for improvement: in that model
T = T 1 already satised the only conservation condition
(conservation of total trips) on oer.

Philip A. Viton CRP/CE 5700() Avg Growth October 23, 2014 9 / 27 Philip A. Viton CRP/CE 5700() Avg Growth October 23, 2014 10 / 27

Iteration (III) Iteration Summary

So the proposal goes like this: Suppose L Eik H doesnt hold for at least on zone i. So we want to
try again, ie compute the next trip matrix T k +1 . This will be based on
1. Based on T 0 and O , compute the zonal growth factors
T k , and the model can be written as:
Fi0 = Oi /Oi0 (for each i), and then the average growth factors
matrix Fij0 = (Fi0 + Fj0 )/2. Fik = Eik = Oi /Oik (the zonal growth factors)
2. Generate a new candidate future trip matrix Tij1 = Tij0 Fij0 . k
Fi + Fjk
Fijk = (the average growth rates matrix)
3. Compute the error ratios Ei1 = Oi /Oi1 . Check to see whether they 2
satisfy L Ei1 H , for all i = 1, 2, . . . , Z . Tijk +1 = Tijk Fijk
4. If they do, then we are done: we take T = T 1 .
We sometimes write the iteration step as T k +1 = T k F k , where is
5. If not, then repeat the computation with T 1 as our base.
term-by-term matrix multiplication (not the matrix product), also known
6. Continue until we have converged. as the Hadamard product.

Philip A. Viton CRP/CE 5700() Avg Growth October 23, 2014 11 / 27 Philip A. Viton CRP/CE 5700() Avg Growth October 23, 2014 12 / 27
Example Iteration 1 Growth Factors
We turn now to an illustration of the calibration procedure. The data Zonal growth factors:
needed is:
Fi0 = Oi /Oi0
A base (observed) trip matrix (just as for the uniform growth factor
model). For illustration we assume: = [75, 45, 80, 95] [65, 60, 61, 68]
= [1.15385, 0.75000, 1.31148, 1.39706]
25 12 10 18 65
10 30 14 6 60 Average growth factors matrix:
Tij0 = 8 12 27 14 61 2 3
6 13 17 32 68 1.15385 0.95192 1.23266 1.27545
6 0.95192 0.75000 1.03074 1.07353 7
49 67 68 70 254 Fij0 = 6
4 1.23266 1.03074
7
1.31148 1.35427 5
Predicted originations by zone: we assume: 1.27545 1.07353 1.35427 1.39706

Oi = [75, 45, 80, 95]

Example: F20 = 0.75000, F30 = 1.31148, so
0 = (0.75000 + 1.31148) /2 = 1. 030 74.
F23
Convergence: we shall assume (L , H ) = (0.95, 1.05).
Philip A. Viton CRP/CE 5700() Avg Growth October 23, 2014 13 / 27 Philip A. Viton CRP/CE 5700() Avg Growth October 23, 2014 14 / 27

Iteration 1 New Trip Matrix Iteration 1 Convergence Check

Tij1 = Tij0 Fij0 Target, Oi 75 45 80 95

28.8462 11.4231 12.3266 22.9581 75.5540 Actual, Oi1 75.554 52.8907 76.5997 89.337
9.5192 22.5000 14.4303 6.4412 52.8907 Error ratio Ei1 0.99267 0.85081 1.04439 1.06339
= 9.86129 12.3689 35.4098 18.9597 76.5997 Since these are not all within our convergence bounds (see the 2nd and
7.65271 13.9559 23.0225 44.7059 89.3370 4th entries), we have not converged.
55.8794 60.2478 85.1893 93.0649 294.3810
So we want to try again (iteration 2) using T 1 as our new starting point.
1 = F0
Example: T23 0 = 1.03074
T23 14 = 14. 430 3.
23

Philip A. Viton CRP/CE 5700() Avg Growth October 23, 2014 15 / 27 Philip A. Viton CRP/CE 5700() Avg Growth October 23, 2014 16 / 27
Iteration 2 Growth Rates Iteration 2 New Trip Matrix

The new zonal growth factors are the error ratios from the convergence
check:
Fi1 = Ei1 = [0.992668, 0.850811, 1.04439, 1.06339] Tij2 = Tij1 Fij1
New average growth factor matrix: 28.6346 10.5291 12.5550 23.6016 75.3204
2 3 8.7743 19.1432 13.6742 6.16485 47.7565
0.99267 0.92174 1.01853 1.02803 = 10.0440 11.7207 36.9817 19.9815 78.7279
6 0.92174 0.85081 0.94760 0.95710 7 7.86721 13.3572 24.2632 47.5397 93.0273
Fij1 = 6
4 1.01853 0.94760
7
1.04439 1.05389 5 55.3201 54.7502 87.4741 97.2877 294.8320
1.02803 0.95710 1.05389 1.06339

Philip A. Viton CRP/CE 5700() Avg Growth October 23, 2014 17 / 27 Philip A. Viton CRP/CE 5700() Avg Growth October 23, 2014 18 / 27

Iteration 2 Convergence Check Iteration 3 Growth Factors

Zonal growth factor (ie the error ratios just computed):

Fi2 = Ei2 = [0.99575, 0.94228, 1.01616, 1.02121]

Target, Oi 75 45 80 95
Actual, Oi2 75.3204 47.7565 78.7279 93.0273 Average growth factors matrix:
Error ratio, Ei2 0.99575 0.94228 1.01616 1.02121
2 3
Since E22 is (just) outside our bound (0.95) we have not yet converged. 0.99575 0.96901 1.00595 1.00848
6 0.96901 0.94228 0.97922 0.98174 7
Fij2 = 6
4 1.00595
7
0.97922 1.01616 1.01868 5
1.00848 0.98174 1.01868 1.02121

Philip A. Viton CRP/CE 5700() Avg Growth October 23, 2014 19 / 27 Philip A. Viton CRP/CE 5700() Avg Growth October 23, 2014 20 / 27
Iteration 3 New Trip Matrix Iteration 3 Convergence Check

Tij3 = Tij2 Fij2 Target, Oi 75 45 80 95

28.5128 10.2028 12.6297 23.8017 75.1471 Actual, Oi3 75.1471 45.983 79.515 94.3115
8.50236 18.0383 13.3900 6.0523 45.9830 Error ratio, Ei3 0.998043 0.978623 1.00610 1.00730
= 10.1038 11.4772 37.5792 20.3548 79.5150 Since for all of these we have 0.95 = L Ei3 H = 1.05, we conclude
7.9339 13.1133 24.7165 48.5478 94.3115 that we have converged.
55.0529 52.8316 88.3155 98.7566 294.9570

Philip A. Viton CRP/CE 5700() Avg Growth October 23, 2014 21 / 27 Philip A. Viton CRP/CE 5700() Avg Growth October 23, 2014 22 / 27

Final Result Other Growth-Factor Models

Our nal prediction for the distributed trips is therefore:

One problem with using the average growth factor model is that in
28.5128 10.2028 12.6297 23.8017 75.1471 practice it tends to converge rather slowly, especially in more realistic
8.50236 18.0383 13.3900 6.05229 45.9830 (large number of zones) settings.
Tij = Tij3 = 10.1038 11.4772 37.5792 20.3548 79.5150
To try and speed up convergence, researchers developed several
7.9339 13.1133 24.7165 48.5478 94.3115
variants on it.
55.0529 52.8316 88.3155 98.7566 294.9570
We will not consider them here the point of presenting the growth
Note that this satises conservation of total trips and conservation of factor models was simply to serve as a lead-in to the gravity model
origins only approximately. Of course we could continue to iterate, and but, for reference, here are two of them.
this would provide a better approximation to the conservation conditions.

Philip A. Viton CRP/CE 5700() Avg Growth October 23, 2014 23 / 27 Philip A. Viton CRP/CE 5700() Avg Growth October 23, 2014 24 / 27
Detroit Model Fratar Model
This model was developed by the civil engineer, T.J. Fratar. It requires the
The Detroit model was initially developed for trip distribution in Detroit. same inputs as the average growth factor model and the Detroit model. In
It requires the same data as the average growth factor model: T 0 and some ways, it is an example of the lengths to which people were prepared
zone-by-zone predictions of future originations (O ). Its structure (at the to go to try and make their models predict better. Its structure (at
completion of iteration k) is: iteration k) is:
Lki Lkj
Fik Fjk Tijk +1 = Tijk Fik Fjk
Tijk +1 = Tijk 2
Fk where:
where:
Fik is the i-th zonal growth factor implied by T k , ie Fik = O /Oik
F k is the regional growth factor implied by T k , ie And
k
F k = S (T )/S (T k ). m Tim
Lki =
Fik is the i-th zonal growth factor implied by T k , ie Fik = Oi /Oik m Fmk Tim
k

k ; but
This term is not easy to interpret. In the numerator we have Om
the denominator isnt so easily understood.
Philip A. Viton CRP/CE 5700() Avg Growth October 23, 2014 25 / 27 Philip A. Viton CRP/CE 5700() Avg Growth October 23, 2014 26 / 27

Final Comments

Youll note that all these iterative models have used only future data
on originations (the O ).
Of course, they would work equally well in principle with predictions
of the attractions.
But its not clear how they would incorporate predictions of both
originations and attractions.
And surely we believe that both may matter in the determination of
future interzonal trips.
One advantage of the Gravity Model is that it makes the distribution
of the trips depend on predictions of both originations and attractions.

Philip A. Viton CRP/CE 5700() Avg Growth October 23, 2014 27 / 27