Geographical and Planning Models for Data on Spatial Flows

Author(s): Mike Baxter

Source: Journal of the Royal Statistical Society. Series D (The Statistician), Vol. 35, No. 2,
Special Issue: Statistical Modelling (1986), pp. 191-198
Published by: Wiley for the Royal Statistical Society
Stable URL: http://www.jstor.org/stable/2987523 .
Accessed: 03/01/2015 13:28

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.

Wiley and Royal Statistical Society are collaborating with JSTOR to digitize, preserve and extend access to
Journal of the Royal Statistical Society. Series D (The Statistician).

http://www.jstor.org

This content downloaded from 128.235.251.160 on Sat, 3 Jan 2015 13:28:42 PM

All use subject to JSTOR Terms and Conditions
The Statistician(1986) 35, pp. 191-198

Geographical and planningmodels fordata

on spatial flows

MIKE BAXTER

DepartmentofMathematics,Statisticsand OperationalResearch,TrentPolytechnic,
NG] 4BU, U.K.
BurtonStreet,Nottingham

Abstract.Modelsforflowsbetweensetsof originsand destinations have been developedand widely

appliedin thegeographical and planningliteratures.
Manyofthesemodelsare essentially mathemati-
cal models.Manymethodsofmodelfitting and assessment havebeenproposed,oftenwithout explicit
considerationofthestochastic natureofthedataused.Thesemethodsarereviewed. Froma statistical
viewpointit is arguedthatmostof the modelsand methodsof fitting similarin the
are essentially
sensethattheyare generalised linearmodelsfittedby maximumlikelihood(or approximations to it)
undersimilardistributional assumptions.This formatmakesexplicittestableassumptions aboutthe
natureofvariationin thedata and leads naturallyintotechniquesforassessingand improving model
performance. These techniques, whichincludetheuse of quasi-likelihoodmodelsforover-dispersed
data, analysisof residualsand exploratorymodelfitting are illustrated,
usingdata drawnfromthe
literature.

1 Introduction

Models forflowsbetweenspatiallylocated originsand destinationsare widelyused in

geographyand urbanand regionalplanning.Phenomenastudiedincludefreight flows
(Chisholm & O'Sullivan, 1973); recreationaltravel (Edwards & Dennis, 1976); air-
line travel (Fotheringham,1984); shopping behaviour (Batty & Mackie, 1972)
and migration(Masser & Gould, 1975). Models have been used fortheirtheoretical
interest,descriptivepurposes,for forecasting,and as componentsof economic de-
mand models.
FollowingWilson (1970), much of the literatureadopts a mathematicalmodelling
perspective.Wilson (1974, p. 4) favourablycontrastssuch a deductiveapproachwith
an inductivestatisticalapproach. The view adopted here is thatthe two approaches
are complementary, but that the benefitsof a statisticalmodellingapproach are
relativelyneglected.Given a model, explicitstochasticassumptionsabout the data
guide the stages of model estimationand assessment.These assumptionsmay be
testedand can lead to model reformulation. Commonlyused models and methodsof
estimationare introducedin Section2, and a generalformulation is givenin Section3
where it is shown that many models can be viewed as generalisedlinear models
estimatedby maximum-or quasi-likelihoodmethods,or approximationsto such
methods. In Section 4, some consequences of such an approach are explored and
illustratedin Section 5.
The discussion is appropriatefor small and moderatedata matrices.Large data
matrices,estimatedfromsample data or onlypartiallyobserved,can presentpractical
(Knott, 1979), that are not dealt with here. The modifiableareal unit
difficulties,
problem(Openshaw, 1984), whichariseswhenoriginsand destinationsare arbitrarily
definedzones, is also not considered.

This content downloaded from 128.235.251.160 on Sat, 3 Jan 2015 13:28:42 PM

All use subject to JSTOR Terms and Conditions
192 Baxter

2 Models

Let T, be the observedflowbetweenorigini and destinationj. A classical 'gravity'

model is
T,,=aP,PjD,7 (1)
whereP,,P, are thepopulationsof i andj representingtheir'masses',D,,is thedistance
between i and j, and a, A, y,fiare unknownparameters.Alternativedefinitionsof
'mass' and the separationbetweenoriginand destinationcan be used; alternative
distancefunctionsare possible;and additionalvariablesmaybe added. An exampleis
given in Section 5. Estimationis usuallyby ordinaryleast squares aftera logarithmic
transformation, with the assumptionthat the transformedmodel has an additive
normallydistributederrorterm. Equation (1) is an example of an unconstrained
model. It is oftenrequiredthatconstraintsof the form
T,,= ET, (2)
are satisfied,whereT,,is the predictedvalue of T,. If summationin (2) is over one of
the indices,i orj, the model is singlyconstrained;if summationis satisfiedover each
index separately,the model is doublyconstrained.
A typicalformof doubly-constrained model (Bennett& Haining, 1985) is

T,0=A,BO,Djf(C,0) (3)
where
=
A,=[ B,D,f (C)] - (4)

B,= [ XA,O1f(C11)]'- (5)

and
XT,,=D, ET,,=O, (6)

ACI) is a functionof C,,the

The 'balancingfactors'Al and B, ensurethat(2) is satisfied;
cost of movingbetweeni and j. Singlyconstrainedmodels can be similarlydefined
(Batty& Mackie, 1972).
Many methodshave been proposedforestimatingconstrainedmodels (Openshaw,
1979; Baxter, 1983). The most common can be viewed as maximum likelihood
estimationassumingthat T, are independentPoisson random variables. This ap-
proach includes the entropymaximizingmethod of Wilson (1970); the Bayesian
approach of Hyman (1969); and the product-multinomial samplingmodel of Batty&
Mackie (1972). Several otherapproachescan be regardedas approximationsto thefull
maximumlikelihoodapproach,includingthe logitregressionmethodof Baxter(1979)
and weightedleast-squaresor minimumchi-squaredmethods that take the form.
Minimise:
S-S (T,,-T,)2/Wl, (7)
where
W,,=T, orT,,.
In the next section,a formulationis givenbased on the theoryof generalisedlinear
models (McCullagh & Nelder, 1983). From a statisticalviewpoint many of the
distinctionsbetweenthe different models then disappear. This generalformulation

This content downloaded from 128.235.251.160 on Sat, 3 Jan 2015 13:28:42 PM

All use subject to JSTOR Terms and Conditions
 Modelsfordata on spatialflows 193

paves the way for a discussion of modellingimplicationsin the remainderof the

paper.

3 A generalformulation

The most common models have the followingstructure(Kirby,1974). Functionsof

the flows,g(T), are independentrandomvariables,withmean h(,u,),takenfromsome
specifiedmemberof the exponentialfamilyof distributions,
with
,u,,=a,/3,
f,, (8)
In equation (8), a, and fl,are originand destinationspecificfunctionsor factors;andfL,
is a functionof variablesmeasuringthe separationof i and j. Usually a,, f3,andf] are
definedso that
,a,,= exp(X'l,,l) (9)
origins,destinationsand theirseparation
Where X,, is a (pX 1) vectorcharacterising
and fiis a parametervector.Thus, forthe usual unconstrainedmodel,g(T,,)=log (T7')
is normallydistributedabout log (C) where,u,,is definedby the righthand side of(1).
For a doublyconstrainedmodel g(T,,)= T,,;h(u,)=,, and T,,has the Poisson distribu-
tion withmean ,u where
exp (#C,,)
,u,= aJf3 (10)
for explicitness,and a, and f3jare parametersto be estimated.Dealing only with
constrainedmodels and assumingthat T,,are independentPoisson randomvariables,
it can be shown(Pregibon,1981; Jorgensen,1984; Green, 1984) that,usingiteratively
reweighted least squares (IRLS), parameterestimatesare givenby
/3=[X'VX]-'X'VZ (11)
whereX is the (nXp) matrixwithtypicalrowX',, V is a diagonalmatrixwithtypical
vectorof the form
element,,; Z is a pseudo-observation
Z=X/3+ v-IS (12)
withS the (n X 1) vectorwithtypicalelement(T,,-,i,).
If, instead of the Poisson assumption,we assume only that T,,has mean ,u,and
variance 2a,,u equation (11) also definesa quasi-likelihoodsolution. For maximum
likelihood, standard errorsof the elements of /3are given by the square-rooted
diagonal elementsof fX'VX]-'; for quasi-likelihoodthese must be multipliedby a
suitableestimateof a. One such maybe obtained(McCullagh& Nelder, 1983) from
a =XA2(n-p) (13)
where
AX7= J(T_
lj
ji)21/i (14
~~~~~~~~~~~~~~~~~(
In the foregoingdevelopment,it is assumed that the requiredinversesexist. The
truthof thisassumptioncan usuallybe ensuredby a suitabledefinitionofthetermsin
X,, and /3,though care then needs to be exercised in the interpretationof the
parameters(Constantine& Gower, 1982; Baxter,1985).
The approach outlinedwill give a doublyconstrainedmodel if X,,containsdummy
variables specificto each originand destinationwithcorresponding parametersin,8
(Baxter, 1985). Singlyconstrainedmodels are obtained by incorporatingoriginor
destinationspecificvariablesand parameters.

This content downloaded from 128.235.251.160 on Sat, 3 Jan 2015 13:28:42 PM

All use subject to JSTOR Terms and Conditions
194 Baxter

4 Some aspectsof modellingspatialflows

In treatingmodels for flows as generalisedlinear models, a statementabout the

models' stochasticpropertiesis needed. This statementexposes similaritiesbetween
particularmodels,oftentreateddistinctly, and can guide the choice of an appropriate
model in particularcases.
The usual approach to estimatingparametersin (1) by log-linearleast squares
impliesthatthe model containsa multiplicative log-normalerrorterm.Flowerdew&
Aitkin(1982) have criticisedthis approach to estimation.Analysisis on the logarith-
mic scale ratherthan directlyin termsof the observedvalues, and resultsin biased
estimatesof the ,u,and, in particular,the underprediction of large flowsand of the
totalflow.This bias is a seriousdrawbackiftheparticularapplicationrequiresthatthe
largerflowsbe accuratelymodelled.Flows are oftenmeasuredas non-negative integers
and, whenmanyofthemare small,thereis littlereasonto supposethatthelog (T,) are
normallydistributedabout theirmean.
Theiralternativeapproachrecognisestheintegernatureofthedata and treatsthe T,,
as Poisson randomvariables.This gives an IRLS procedurewhichis thatsummarised
in Section3. For thisapproach,theonlydifference betweenunconstrained, singly-and
doubly-constrained modelsis in.the particulartermsincludedin XY,and fGin equation
(9). In thissense,the distinctionbetweenthe differenttypesof model disappears.
Estimationof parametersin modelssuch as (1) or (3) usingnon-linearleast squares
assumingan additive,homoscedasticerrortermhas sometimesbeen proposed(Ces-
ario, 1974; Giles & Hampton, 1981). This approach ignoresthe integernatureof the
data, importantwhenflowsare small,and almostcertainlymisrepresents thevariation
in the data.
Thus, many of the more suitable methodscan be viewed as proceedingfromthe
assumptionthatthe T, are Poisson randomvariables,an assumptionhowever,which
often understatesthe observed variation in the data (Flowerdew & Aitkin, 1982;
Baxter,1985). Global measuresof fitoftenused are the likelihoodratiostatistic.

G2=2XT, log (T,l,,) (15)

or the chi-squaredgoodnessof fitstatistics,X-,givenby (14). Theirsize, relativeto the

degrees of freedom,is oftenlarge enough to show that the Poisson assumptionis
incorrect.
Several approachesto modellingthis extra-Poissonvariationhave been tried.For
migrationdata flowsFlowerdew& Aitkin(1982) arguethatthe numberof household
heads migratingmay followa Poisson distribution,but the numberof individuals
movingwill followa compound-Poissondistribution, havinggreatervariancethanthe
Poisson,because of variationin the size of the households.Methodsof estimationare
discussedin Hinde (1982).
A more 'black-box'approach is to make second momentassumptions-specifically
thatthe varianceof T, is proportionalto its mean-and use quasi-likelihood(Baxter,
1985). Anotherapproach is to distinguishbetween specificationerror in ,u and
inherentrandomerrorin T,, each of whichenterthe model in a different way. This
approach has rarelybeen triedwith models of flowdata (Stetzer,1976), thoughthe
methodsof Breslow(1984) should be possible.An analysisof residualsand influence
can show thata poor fitis exaggeratedby a sub-setof badly-fitting points.Identifica-
tion and treatmentof these points may improvethe fitand understandingof the
phenomenonmodelled.
For example,Flowerdew& Aitkin's(1982) analysisnotes that some of the worst
model predictionsseem to be associated with flowsbetweennaval bases, probably
accountedforby troopmovements.Geographicalfeaturesmay also influenceflowsto

This content downloaded from 128.235.251.160 on Sat, 3 Jan 2015 13:28:42 PM

All use subject to JSTOR Terms and Conditions
 Modelsfordata on spatialflows 195

an extentunaccountedforby the basic model (Masser & Gould, 1975). Some specific
examplesand unresolvedproblemsare discussedin the finalsection.

5 Some examples

5.1 The use ofdiagnosticstatistics

Detailed analysis of the componentsof a model's fitis ofteninformative.Recent

developmentsin the statisticalliterature(e.g. Cook & Weisberg,1982) have begunto
penetratethegeographicalliterature(Wrigley,1983, 1984) but have had littleapplica-
tion to data on spatial flows.Mappingof residualsfromflowdata is possible(Masser
& Gould, 1975) but can be difficultto interpret.
Baxter(1985) reanalysesdata on migrationin theUnited States,usingdata givenin
Dorigo & Tobler (1983). Using a doublyconstrainedmodel and statisticsforidentify-
ing outliersand influentialpointsgivenin Pregibon(1982), threeunusual pointswere
identified.Deletion of these singlyand in pairs, and recalculationof the statistics,
suggestedthatall threewerepoorlyfittedby a model suitableforthe restof the data;
no otherunusual pointswere highlighted. The deletionof all threereducedthe value
of the x2 goodness of fitstatisticby 40% and substantiallyaffectedthe parameters
associated withtwo of the deleted flowsand the estimateof fi.The two flowswere
reciprocalflowsbetweentwo regions,one centredon Salt Lake Cityand the otheron
San Francisco.The model fittedwiththe deletedpointsover-predicted flowsby 300%
or more.
Possible causes of mis-specification
suggestedin Baxter(1985) includetheuse of an
inappropriatemeasureof separationor theexistenceof 'sociological'factorsthatdeter
movementbetweenthe two regions.Whateverthe cause, it is clearthatanyinferences
to be drawnfromthemodel maybe heavilydependenton thesetwoobservations.The
data set has been used several times to contrastalternativemodellingapproaches
(Tobler, 1983; Dorigo & Tobler, 1983; Ledent, 1985) withoutallowingforthe effect
on the comparisonsof influential data points.

5.2 Exploratorymodelfitting

It has been observed that the models fittedto the data typicallyhave a simple
structure,and it is not surprisingthattheyfailto capturethecomplexitiesof observed
flowdata (Knott, 1979). In some cases, featuresof the geographyof a regionthatmay
act as actual or perceivedbarriersto travelcan be identified(Masser & Gould, 1975).
In other cases, afteranalysis,characteristicsof the structureof a region may be
identifiedas influencing flowsto an extentnot accountedforin themodel (Flowerdew
& Aitkin,1982; Ewing& Baxter,1981).
Such featurescan be incorporatedinto an exploratorymodel buildingapproach,
thoughthis is rarelyattempted,Using data on migrationwithinUganda given in

]
Masser & Gould (1975), a model of the form
4 4
t,u,=exp a+ E, bX,+ Yt.Xk+lD,k
E1 (16)
k= =
L
was fittedwherethe X, measureoriginpopulation,income,urbanisationand educa-
tion forK= 1, 2, 3, 4 and theXk,measurethe same variablesfordestinations.
Two dummyvariables were created; B,= 1 for flowsfor which the lake system
appeared to be a 'barrier'betweenoriginand destinationon themap and 0 otherwise;

This content downloaded from 128.235.251.160 on Sat, 3 Jan 2015 13:28:42 PM

All use subject to JSTOR Terms and Conditions
196 Baxter

and N,,= 1 for flowsbetweenregionshavinga borderin common and 0 otherwise.

Resultsare presentedin Table 1 formodel (16); thatwithboth B,,and N,,added; and
thatwithN, onlyadded. One possibleoutlierwas identified,but its removalhad little
effecton the resultsobtained.

Table 1. Parameter
estimates
forthreemodels

Estimates
Parameter Model 1 Model 2 Model 3
a* -14 65 -18 47 -17 49
J5* 1 36 1 36 1 36
C52* -1 30 -1 12 -1 19
(53 0-10 0-09 0-10
(54* 067 070 073
Yi* 084 083 082
Y2 -0 16 003 -005
Y3* 1149 1 49 1 50
Y4 0009 0 13 0.15
/3* 0.0104 0 0075 -0 0079
3 - -0 21
,B* - 091 093
RSS 1742 153 1 1548
DF 200 198 199
R2 0075 0 78 0 78
Notes
E4 exp
(1) The modelis T,,=a P,' J2U53E64 ]J' 1)2 Ub?[EiD,+/11/1,,+/12N,1]
WhereP, I, U, E measureregionalpopulation,income,urbanisation and
education;D, is distance;B,,= 1 if the lake is a barrierbetweeni andj;
N,,= 1 ifi andj havea borderin common.
(2) An asterisk* indicatesthattheparameter is significant;
usuallythe
estimateis at least4 timesitsstandarderror.Exceptfor/, in model2 with
a standarderrorof 0 14, non significant estimatesare smallerthantheir
standarderror.
(3) RSS is theresidualsumof squaresand DF thedegreesoffreedom.
For a nullfittheseare 696.7 and 209.

Qualitatively,the models produce similarresultsin the sense thatthe same para-

metersare significant or not in each case. The barriereffectas modelledhere is not
significant, thoughhas the expectedsign; the 'adjacency' variable is significantand
leaves most of the parameterestimatesrelativelyunaffectedapart fromthe distance
parameter.
Interpretation is not clear-cut.There is a tendencyto under-predict flowsbetween
adjacentregionswhenusinga simpledistancedeterrencefunction.The significance of
the adjacency factor,N, may reflectthe fact that a simple measure of distance
betweenregionalcentroidsoverstatesthe effective distancebetweenadjacentregions.
Alternatively, the factof adjacencymay promotemigrationto an extentgreaterthan
thatpredictedusinga simpledeterrencefunction.This resultcould itselfbe function
of thevariableinformation available at an originabout different
destinations.
The foregoinginterpretation is speculative.Difficultiesof interpretation,
as with
difficultiesof definingnew variables,may explainthelack of exploratory applications.
Anotherreasonmaybe resistanceto using'ad-hoc' modelshavingno theoreticalbasis.
Two points can be made here. The firstis that particulardata sets are likelyto
contain unusual featuresnot accounted forby the model initiallyused. Exploratory
modellingof identifiablefeaturesmay removetheirinfluencefromthe results,allow-
ing morereliableestimationof othercomponentsof themodel. Second,some features
of spatial structure,such as adjacency or nearness,can be definedfordifferent data

This content downloaded from 128.235.251.160 on Sat, 3 Jan 2015 13:28:42 PM

All use subject to JSTOR Terms and Conditions
 Modelsfordata on spatialflows 197

sets. Regular applicationof an exploratoryapproach could serve to identifyconsis-

tentlyimportantfactorsthatany theoryof spatial flowsshouldaccountfor.

assumption
5.3 The independence

Finally,a problemthathas receivedlimitedattentionis noted.The modelsdiscussed

assume thatthe T, can be treatedas independentrandomvariables.Ripley(1985) has
arguedthatthiswill rarelybe the case and suggeststhatthespatialdependenceshould
be directlymodelled.It is not clear how thiscan be done (Bennett& Haining, 1985).
As a firstapproximation,the independenceassumptionis reasonableas one of the
objectivesof modellingis to explainthespatialdependencebetweenthe T, in termsof
observableinfluences. The validityoftheindependenceassumptionis, however,rarely
examined. Methods exist for regressionmodels with a singlevalue at each location
that has been modelled (Cliff& Ord, 1981). For flowdata, whereeach location is
associated with a set of values, such methodsare not directlyapplicable. Residuals
froma fitcan be examinedin variousways.They maybe mapped foreach originand
destinationseparatelyor arrangedin a matrixformwhere the rows and columns
correspondto some 'sensible'ordering,reflecting the locationsof originsand destina-
tions.This mappingcan revealpatternsthatsuggestmodelmisspecification, including
spatial autocorrelationin the residuals.
It may be possible to identifyvariableswhose inclusionin the model removesthe
dependence among the residuals;if not, the problemthen arises of modellingthe
dependence. Apart fromBrandsma & Ketellapper(1979), therehave been few at-
temptsto do this. For constrainedmodels,one approachis to workwithan approxi-
mate model given by the firstiteration in the IRLS procedure summarisedin
equations (11) and (12). This methodgives rise to weightedlinear least squares for
whichexistingmethodscould be modified.These methodsare,however,complexand,
in certainrespects,unsatisfactory. In particular,the structureof any spatial depen-
dence needs to be specifiedin a detail thatmanyconsiderunrealistic(Arora& Brown,
1977). More workneeds to be done beforethisproblemis resolved.

References

ARORA,S.S. & BROWN,M. (1977) Alternativeapproaches to spatial autocorrelationan improvement

over currentpractice,InternationalRegional Science Review, 2, pp. 67-78.
BATTY, M. & MACKIE, S. (1972) The calibration of gravity,entropy and related models of spatial
and PlanningA, 4, pp. 205-233.
Environment
interaction,
BAXTER, M.J. (1979) The application of logit regressionanalysis to production constrained gravity
models, Journalof Regional Science, 19, pp. 17 1-177.
BAXTER, M.J. (1983) Estimation and inferencein spatial interaction models, Progress in Human
7, pp. 40-59.
Geography,
BAXTER, M.J. (1985) Quasi-likelihood estimation and diagnostic statistics for spatial interaction
and PlanningA, 17,pp. 1627-1635.
models,Environment
modelling
and spatialinteraction:
BENNETT, R.J.& HAINING, R.P. (1985) Spatialstructure approaches
to the statisticalanalysis of geographicaldata, Journalof the Royal StatisticalSociety,Series A, 148,
pp. 1-36.
BRANDSMA, A.S. & KETELLAPPER, R.H. (1979) A biparametricapproachto spatialautocorrelation,
Environment and PlanningA, 11,pp. 51-58.
33, pp. 38-44.
models,AppliedStatistics,
BRESLOW, N.E. (1984) ExtraPoissonvariationin log-linear
CESARIO, F.J. (1974) More on the generalized tripdistributionmodel, JournalofRegional Science, 14,
pp.389-397.
CHISHOLM, M. & O'SULLIVAN, P. (1973) FreightFlows and SpatialAspectsof theBritishEconomny
(Cambridge UniversityPress).

This content downloaded from 128.235.251.160 on Sat, 3 Jan 2015 13:28:42 PM

All use subject to JSTOR Terms and Conditions
198 Baxter

CHISHOLM,M. & O'SULLIVAN,P. (1973) FreightFlows and SpatialAspectsof theBritishEconomy

(CambridgeUniversity
Press).
CLIFF,A.D. & ORD,J.K.(1981) SpatialProcesses
(Pion,London).
CONSTANTINE, migration,
A.G. & GOWER,J.C. (1982) Models for the analysisof inter-regional
and PlanningA, 14,pp. 477-497.
Environment
COOK, R.D. & WEISBERG,S. (1982) Residualsand Influence (London, Chapman & Hall).
inRegression
DORIGO, G. & TOBLER,W. (1983) Push-pullmigration laws,Annalsof theAssociationofAnmerican
73, pp. 1-17.
Geographers,
in GreatBritain,
EDWARDS,S.L. & DENNIS,S.J.(1976) Longdistanceday-tripping ofTransport
Joulrnal
Economiesand Policy,10,pp. 237-256.
Scotland,ScottishGeo-
EWING,G.O. & BAXTER,M.J. (1981) Recreationalday tripsin east-central
graphicalMagazine,97, pp. 147-157.
R. & AITKIN,M. (1982) A methodof fitting
FLOWERDEW, modelbased on the Poisson
the gravity
JournalofRegionalScience,22, pp. 191-202.
distribution,
A.S. (1984) Spatialflowsand spatialpatterns,
FOTHERINGHAM, and PlanningA, 16,pp.
Environment
529-543.
4, pp. 26-28.
GILCHRIST,R. (1981) Calculationofresidualsforall GLIM models,GLIM Newi'sletter,
GILES,D.E.A. & HAMPTON,P. (1981) Interval estimationin thecalibration
ofcertaintripdistribution
ResearchB, 15,pp. 203-219.
models,Transportation
reweightedleast squares formaximum likelihood
GREEN,P.J.(1984) Iteratively estimation
and some
robustand resistant
alternatives, SocietyB, 46, pp. 149-192.
JournaloftheRoyalStatistical
(Ed.) GLIM82:Proceedings
HINDE,J. (1982) Compound Poisson regressionmodels, in: R. Gilchrist of
Conference
the International on GeneralisedLinear Models,pp. 109-121 (New York,Springer-
Verlag).
HYMAN,G. (1979) The calibration models,Environment
of tripdistribution and PlanningA, 1, pp.
105-112.
JORGENSEN, B. (1984) The delta algorithmand GLIM, International StatisticalReview,52, pp.
283-300.
KIRBY,H. (1974) Theoretical
requirements models,Transportation
gravity
forcalibrating Research,8,
pp. 97-104.
aspectsoftheRHTM, in Forecasting
KNOTT,M. (1979) Statistical on TrunkRoads:A Reporton
Traffic
theRegionalHighwayTrafJficModelPr-oject(London,HMSO).
LEDENT,J. (1985) The doubly constrainedmodel of spatial interaction:a more general formulation,
and PlanningA, 17,pp. 253-262.
Environment
of British
Migrationin TropicalAfrica(Institute
MASSER,I. & GOULD,W.T.S. (1975) Inter-regional
Geographers).
LinearModels,(London, Chapman & Hall).
MCCULLAGH,P. & NELDER,J.A. (1983) Generalized
methodsofestimating
S. (1979) Alternative
OPENSHAW, modelsand theirperform-
spatialinteraction
ance in short-term in: C. P. A. Bartels& R. H. Ketellapper(Eds) Exploratory
forecasting, and
Explanatory AnalysisofSpatialData, pp. 201-225 (Leiden,Nighoff).
Statistical
OPENSHAW, S. (1984) TheModifiable ArealUnitProblem(Norwich,Geo Books).
PREGIBON,D. (1981) Logisticregression AnnalsofStatistics,
diagnostics, 9, pp. 705-724.
PREGIBON,D. (1982) Scoretestsin GLIM withapplications, in: R. Gilchrist(Ed.) GLIM82: Proceed-
ingsoftheInternational on Generalized
Conference LinearModels,pp. 87-97 (New York,Springer-
Verlag).
RIPLEY,B. (1985) Discussion of paper by Bennett& Haining (1985).
STETZER, F. (1976) Parameter estimation for the gravity model: a comparison of six methods,
and PlanningA, 8, pp. 673-683.
Environment
TOBLER,W. (1983) An alternative forspatial-interaction
formulation modelling, and
Environment
PlanningA,15,pp. 693-703.
WILSON,A.G. (1970) Entropyin Urbanand RegionalModelling(Pion,London).
WILSON,A.G. (1974) Urbanand RegionalModelsin Geogr-aphy and Planning(London,Wiley).
inHumanGeography,
WRIGLEY,N. (1983) Quantitative methods: on data and diagnostics,Progress 7,
pp. 567-577.
methods:diagnostics
WRIGLEY,N. (1984) Quantitative inHumanGeography,
Progress
revisited, 8, pp.
525-535.

This content downloaded from 128.235.251.160 on Sat, 3 Jan 2015 13:28:42 PM

All use subject to JSTOR Terms and Conditions