Fitting Segmented Regression Models by Grid Search

Author(s): P. M. Lerman
Source: Journal of the Royal Statistical Society. Series C (Applied Statistics), Vol. 29, No. 1 (1980)
, pp. 77-84
Published by: Wiley for the Royal Statistical Society
Stable URL: http://www.jstor.org/stable/2346413
Accessed: 20-06-2015 03:40 UTC

REFERENCES
Linked references are available on JSTOR for this article:
http://www.jstor.org/stable/2346413?seq=1&cid=pdf-reference#references_tab_contents

You may need to log in to JSTOR to access the linked references.

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 [email protected].

Wiley and Royal Statistical Society are collaborating with JSTOR to digitize, preserve and extend access to Journal of the
Royal Statistical Society. Series C (Applied Statistics).

http://www.jstor.org

This content downloaded from 192.122.237.41 on Sat, 20 Jun 2015 03:40:33 UTC
All use subject to JSTOR Terms and Conditions
Appl.Statist.(1980),
29, No. 1,pp. 77-84

FittingSegmented
RegressionModelsby Grid
Search
By P. M. LERMAN
Department ofCambridge,CambridgeCB2 3DX, England
ofAppliedBiology,University
[ReceivedMarch1977.Finalrevision
October1979]

SUMMARY
A grid-searchmethodoffitting
segmented curveswithunknown
regression transition
andcompared
pointsis described witha standard
method. It is shownto be suitablefor
a widerclassofmodelsthanthestandard
fitting method andtoprovide as a by-producta
wayofmakingreliableinferencesabouttheabscissaeofthetransitions.
Keywords: MULTIPHASE; SEGMENTED REGRESSION; TRANSITION POINT

1. INTRODUCTION
WE examinean alternativeto the standardmethodof fittingsegmented(sometimescalled
multiphase)regressionmodels of the type
E(yjx)=-f1(x,P1) forX0,x,X1
=f2(x, P2) forX1<x < X2

f=r(x,Pr) for Xr.<x<Xr (1)

withE(y Ix) continuousin x throughoutthe interval[X0,Xr], and wherer, X0, Xr and the
functional formof each]i are assumedknown,but the parameters P ..., Pr and the vectorof
transition points X = (X1,..., X,. -) must be estimated from the data. Continuity of E(yI x)
impliesthatthe submodelsfi are subjectto constraints
Ji(Xi, Pd)=fi+ 1(Xi, [i+ 1) (2)
for i1, ..., r-1. Various practical applications for such models have been found in recent
years;see,forexample,Sprent(1961),Hudson(1966),Dunicz (1969),Owen,Davis and Ridgman
(1969),Williams(1970),Bacon and Watts(1971) and Kirby(1974).
Characteristically thedata are a set ofn points,(y, x n),...,(y,x), whereyj has a variance
a2/wj,wjis knownbutc2 is usuallyunknowna priori.Wheninferences are to be made,theyjare
assumed to be independentlyand normallydistributed.The least-squaresestimatorsof
PX,x (denotedby P1, P, X) are thevalues whichminimizethe residualsum ofsquares
function

S- E L w i{y-t(xj, pi)}2 (3)

i=lxi < x <x
subjectto constraints
(2). The minimumvalue ofS is denotedbyS. Determining theseestimates
is non-trivial
becauseS possessespotentially manylocal minima,and is non-differentiableat a
numberof points(specifically if any Xi coincideswiththe abscissa of an observation)which
impliesthat some minimaof S need not be stationarypoints.Beforeconsideringhow to
overcomethese difficulties, we note that the fittingof segmentedmodels which are not

This content downloaded from 192.122.237.41 on Sat, 20 Jun 2015 03:40:33 UTC
All use subject to JSTOR Terms and Conditions
78 APPLIED STATISTICS

constrained pointsis a distinctand easierproblem:thereader

to be continuousat thetransition
to the paper by Hawkins(1976) fora discussionof this.
is referred
2. FITTING THE MODEL
Thestandardtechnique
The standardmethodis thatdescribedby Hudson (1966),and laterrefinedforthe "two
intersecting straightlines"case by Hinkley(1969, 1971) and Singpurwalla(1974). It is a two-
stage technique,the firststage being to fitthe model to everyfeasiblepartitionof the data
withoutimposingcontinuityacross the transitions.. For each partitionthe points where
adjacentfJ intersect are computedfromthefittedvalues of PI, --,Pr. Iftheseintersectionpoints
dividethedata intothesame partition, thenthefitis admissibleand itsresidualsumofsquares
noted.The second stage of themethodinvolvesone in returning to thosefeasiblepartitions
whichdid notgiveadmissiblefitsin stageone and doingthefitagain,thistimewithconstraints
(2) imposedexplicitly.The least-squaressolutionis thatwithsmallestS overall.
Hudson's methodis appropriatewhenthemodel(1) is made up entirelyof linearfiwhose
intersection points may be derivedalgebraicallyfromPj..., Pr, forin such cases it can be
programmed to be fullyautomaticand to be relatively of
quick.It is notwellsuitedto thefitting
othertypesof model because it requiresthe use of time-and-effort consumingsuccessive
approximationtechniquesin thefitting stage(ifanyfiis non-linear),orwhencomputingX. Nor
is itan idealapproach,whatever themodel,iftheluser wishesto makeinferences usinglikelihood
methods(theapproach generallyrecommendedespeciallyfortransitionpoints Beale, 1960;
Draper and Smith,1966; Hudson, 1966; Hinkley,1969, 1971; Feder, 1975a,b), because the
necessarysurfacesmustbe generatedas a separateexercise.
Thealternativefittingtechnique("gridsearch")
We observethata subsetof the unknowns,namelythe vectorof transitionpointsX, is
distinguishedbecause S is usuallymoreeasilyminimizedwhenthevalue ofX is fixed.It would
appear thereforethatan alternativeto Hudson's methodis to adapt an approach commonly
used in non-linearregressionand carryout a gridsearchoverX to map Sm(X),theminimum
valueofS at X,to be followedbyiterativerefinementoftheapproximateminimumdisclosedby
themap.
Thisapproachhas twofeatures whichcommendit.Firstly,itcopes easilywitha widerclass
ofmodelsthandoes Hudson'stechnique.We notethatifa modelis linearwhenX is fixed,Sm(X)
maybe obtainedalgebraicallybystandardmultiplelinearregression methods.The mappingof
Sm(X),and thefinalrefinement forsuchmodels(a farwiderclass thanthatforwhichHudson's
techniqueis suitable),is therefore Secondly,themethodprovidesas a by-
verystraightforward.
productthe surfacesneeded formakinglikelihoodinferences about the transitionabscissae.
This second pointis shownin the nextsection.

3. INFERENCE IN SEGMENTED REGRESSION

methodis influenced
Because our choiceoffitting by theway inferences are to be made,I
brieflyreviewcurrentpractice.In the discussionwhich follows,S? denotes the minimum
attainablevalue ofS underhypothesisHo, p denotesthedimensionofthemodel'sparameter
space ( = numberof unknownsminus numberof constraints), S2 ( = S/(n-p)) denotes the
residualmean square and F, denotestheuppera percentvalue oftheF distribution withv
and w degreesof freedom.
The LR (likelihoodratio)statisticfortestinghypothesisHo is
R -2 ln (A) = nln(S?/S) (4)
whichis oftenapproximatedby
R2= (S3)/S2 (5)

This content downloaded from 192.122.237.41 on Sat, 20 Jun 2015 03:40:33 UTC
All use subject to JSTOR Terms and Conditions
 SEGMENTED REGRESSION 79

7he identified case

It can be shown (Feder, 1975b) that if the model is identified(thatis, if all r segments
postulatedin themodelare knownto be presentand adjacentsegmentsare distinct), and ifHo
is true,the R1 and R2 are asymptotically %2 (whereq is the dimensionalreductionin the
model's parameterspace imposed by Ho) and p,...,Pr,X are asymptotically normal with
knownvariance-covariancestructure.
ofR1/qand R2/qto be Fq,n-p (Feder,
In finitesamplesitis betterto assumethedistributions
1975a).Hence,an approximate(100-oc) percentconfidenceregionforthetransitionpoints,X,
derivedfromR1 or R2 comprisesvalues whichsatisfy
Sm(X)< C", (6)
F _/n) ifR1is used,and Cx = C' = 3 + s2(r- 1)F 1 np if
whereCx = C' = Sexp ((r- 1)F_1)
R2 is used. MonteCarlo studiescarriedout byme have shownthatC' givesreliableresultsfor
quitesmallsampleswhenallJiare linear,and thisis thevalueusedintheexampleswhichfollow.
Theinformation neededto fixtheseregionsis preciselythemap of Sm(X)producedbythegrid-
searchtechnique.
standarderrorsfor ,
Approximate ..., 1971)fromthe
X maybe obtained(Hinkley,
informationmatrixconditionalon thedata partitionimpliedbyX, butwithouttheimposition
ofcontinuity (2). The partitionis ambiguousifanyXicoincideswiththeabscissaof
constraints
observationswhencomputing
in whichcase it is reasonableto omitoffending
an observation,
the requiredinformation matrix.Standard errorsquoted in examples are derivedby this
method.
Thenon-identified case
Whenfewerthanrsegmentsmayactuallybe present(thenon-identified case) thesituationis
moreproblematic.Feder(1975a) showsthatifr segmentsare fitted whenthereare fewerthanr
presentthen X1, normal,noris theLR statisticasymptotically
..., j,X are notasymptotically X2.
The approximatemethodsof inferencerecommendedforuse in the identifiedcase cannot
thereforebe justifiedwithnon-identified models.I do not pursuethe subjectfurther in this
paper,and onlyidentified modelsare treatedin the exampleswhichfollow.

4. EXAMPLES
A fewcase studieswillillustratethegrid-search techniqueand bringout variouspractical
straightlines"model
points.In thefirststudywe look at a specialcase ofthe"twointersecting
forwhichHudson's fittingtechniqueis unsuitable.The secondstudyconsidersa modelwitha
non-linearsub-model.Finallywe look at a three-phasemodel,and see how to proceedwith
morethan two segments.
Example1. Shoot-apexdevelopment in cereals
Fig. l(a) presentsdata froman experiment on wheatperformed byDr E. J.M. Kirbyofthe
PlantBreedingInstitute, Cambridge(Kirby,1974).The independentvariable,x, is numberof
days since sowing,and y is the naturallogarithmof the numberof primordia,a measureof
shoot-apexdevelopment.Two intersecting straightlinesrepresenttherelationshipbetweenx
and y quite well;thatis
E(ylx)=ocl+flx forx<Xl
= oc2 + 2X forx>X1 (7)
subjectto cc1+ f1X1 = a2 + f2 X1. If no otherconstraintsare placed upon themodel,itcan be
fitted
straightforwardly byHudson'stechnique.However,thereare good biologicalgroundsfor
believingthatthetransition stageintheplant'sdevelopment,
occursat an identifiable namelyat
theend ofspikeletinitiation.Furthermore, an accurateestimate oftheordinateofthispointis

This content downloaded from 192.122.237.41 on Sat, 20 Jun 2015 03:40:33 UTC
All use subject to JSTOR Terms and Conditions
80 APPLIED STATISTICS

(a)

y 3 -
V

E /

~~ ~ ~ ~~0
E5-
0c 2 -

0 10 20 30 40 50 60 70
Dayssincesowing.x b

0 12

S,,,=1114 /

010 _

44 46 4 50 52 54

FIG. 1.(a) Data pointsand best-fitting

line; (b) relationbetweenSm(Xl)and X1(break ordinateconstrainedto 3 3358
shown-, break ordinateunconstrainedshown---) forExample 1.

availablefrommatureplants: forthesedata it was 3-3358.It is desirableto be able to use this

information whenfitting themodelin orderto testwhetherthetrueordinateofthebreak-point
differsfromthe specifiedvalue and, ifit does not,to improvethe precisionwithwhichthe
remainingunknownparametersare estimated.
Thereis no simplewayin whichHudson's techniquecan be appliedto thismorerestricted
model.The grid-search approach,on theotherhand,copes quiteeasilybecause,forfixedX1,the
modelis linear.The resultsare shownin Fig. l(a), wherethefittedcurveis superimposedon the
data points,and in Fig. l(b) whereSm(Xi)is plottedagainstX1 in the range44 < X1< 54. The
function of0-25alongtheX1-axis,and individualpointsjoined
Sm(Xi) was evaluatedat intervals
up,to givea continuouscurve.The minimalvalueS ( = 008828) was locatedbylinearsearchat
I= 48&4.
From (6) an approximate95 per centconfidenceregionforX1 is givenby
Sm(Xl)<OO- 114.
A horizontalline at thisheightin Fig. l(b) indicatesthisregionto be 47 2 X1 4 49X8.
Fig. l(b) also showstheplotofSm(Xl)whenthemodelis notconstrainedto have a specified
ordinateatX1.The minimalvalueS inthiscase was 0X08645, and was locatedatX1= 48X0which
is theabscissaofan observation, and nottherefore a stationarypointofS. The LR testusingR2
showed no significantevidence that the break ordinate differsfromthe specifiedvalue
(R2 = 0 34 with1 and 16degreesoffreedom). The 95 percentconfidence regionforX1 whenthe
modelis not constrainedto have a specifiedordinateat thebreakis givenby Sm(Xl) 041107,

This content downloaded from 192.122.237.41 on Sat, 20 Jun 2015 03:40:33 UTC
All use subject to JSTOR Terms and Conditions
 SEGMENTED REGRESSION 81

thatis theinterval45-6<X1 < 52-0;thisintervalis almosttwo and a halftimesas wideas that

obtainedwhenthebreakordinateis specifiedand showstheusefulness ofbeingable to impose
thevalue of thatordinate.
Example'2.Twophases,one of themnon-linear
chosen model
We consideran arbitrarily
E(ylx) = c+#cos(x-y) forx<X1
=xa+f,cos(X1-y) forx>X1, (8)
and observethattherequirement ofcontinuityat thetransitionis incorporatedintothemodel
(see Section5 below fordiscussionofconstrainthandlingwithnon-linearfi). Twentyx-values
weresampledfroma uniform in therange0 < x < 1,and thecorresponding
distribution y-values
formedfromtheadditionto (8) (withparametervalues a = 0, /3= 1, y = 0-7 and X1 = 0-5) of
errortermssampledfroma normaldistribution withstandarddeviation0-02.The resulting data
pointsare shownin Fig. 2(a).
(a)

10

VS * 0
/z

09 _
Y
y~~~~~~ //

Sm(Xl)

08

I III
0 025 05 075 10
x

0 02 - (b)

0015 _

0010 -S. 000907

I I I.
0 025 05 075 10
X,

FIG. between
line;(b) relation
2. (a) Data pointsand best-fitting S.(Xl) andXl, forExample2.

This content downloaded from 192.122.237.41 on Sat, 20 Jun 2015 03:40:33 UTC
All use subject to JSTOR Terms and Conditions
82 APPLIED STATISTICS

The function Sm(Xi)was evaluatedat intervals of 0-02in therange0-07<X1 0-99(the

entirerangeofx-valuesin thedata)usinga Marquardttechnique (Draperand Smith,1966;
Adbyand Dempster, 1974).Final valuesofoc,,Band vfromone minimization wereusedas
startingvaluesinthenext;a strategywhichmakesminimization rapidandreliable.Theresults
areshownin Fig.2(b).Theminimal value3 ( = 0-00710)was locatedatXS1= 0-63.The fitted
curveis shownsuperimposed on thedatain Fig.2(a).Anapproximate 95 percentconfidence
regionforX1 usingcriterion Sm(Xi)< 0 00907which
(6) satisfies Fig.2(b)showstobetheinterval
0-34 X1 0-90.
Theshapeoftherelationship inFig.2(b)deservessomecomment. Ofparticular are
interest
theturning pointsat X1= 0-14and X1= 0-26,bothofwhichareinterior pointsofthesame
intervalbetween adjacentobservations. Suchbehaviour possible,butrarein
is theoretically
practice.
Secondly, oftwominima
theexistence withalmostequalvaluesofSm,oneatX1= 0-63
andtheotheratX1= 0-48,demonstrates howusefula plotofSm(X)canbe.Although X1= 0-63
isthebestestimate, X1= 0-48is almostas good.A fairdescription
oftheresults
wouldrequire
thisfactto be mentioned.Without a plotofSm(Xi)it mighthaveescapednotice.Finally, the
curve,withitsthreeminima, showsthata confidenceregionneednotbea singleconnected set:
it couldbe a collection
ofdisjointsets.
feed intakein ruminants
Example3. Voluntary
One of the modelsproposedforrelatingvoluntary feedintakein ruminants, y, to
metabolizable
energy inthediet,z = 1/x,is givenbyOwenetal.(1969)as
(ME) concentration
E(yI x)= ax forx X1
forX1<x X2
=yx/(x-1) forX2<x (9)
subjectto continuity
requirements
oxX1 = ,B= yX2/(X2-1). So farit has notbeenpossibleto
produceexperimentaldatainwhichallthreesegments
ofthemodelareadequately represented,
(a)

036

1 4 %
Feed

1 2s

i oL
/

11 16 20 25 30
(metabohzbl. energy) 'x

(b)

s.K) .t S.,,0-227

0 3~~~~~~~~~~~~~~0

01

11 15 2 0 25 30

between
line;(b) relations
FIG. 3. (a) Data pointsandbest-fitting Sm(X2) and
Sm(Xi)andX1, shown-, andbetween
X2, shown----, forExample 3.

This content downloaded from 192.122.237.41 on Sat, 20 Jun 2015 03:40:33 UTC
All use subject to JSTOR Terms and Conditions
 SEGMENTED REGRESSION 83

andforthepurposeofthisexampleI useartificial tothosegivenbyOwen

data(Fig.3(a))similar
etal.(1969).TwelveME valueswerechosen,andthecorresponding yvalueswerederivedfrom
theadditionto (9) (withparametervalues a = 1, ,B= 1V5,
y = 0 75,X1 = 1P5,X2 = 2) of error
terms sampledfroma normaldistribution withstandarddeviation 0-1.
ThefunctionSm(X1, of0-05alongboththeX1andX2axesin
X2)wascalculatedat intervals
theregion1 1<X1 < X2< 2-95,theentirerangeofx inthedata.Theresulting contourmapis
shownin Fig. 4. Thereis evidently onlyone minimum (S = 04145) whichis locatedat
(X1,X2)= (1-46,1-95).Anapproximate95percentconfidence regionfor(X1,X2)usingcriterion
(6)isboundedbytheisobarSm(X)= 0X282 andisindicated 2.Thebestfitting
inFig.4 as contour
30 - - - -- - __
_,
- - \

2 --

X, 20

3
-y-

10 .

15 20 25 30

FIG.4. ContoursofSm(X1, X2) forExample3. Values ofSm defining thecontoursare (1) 0 208,(2) 0 282,(3) 0 403,(4)
0 673,(5) 1 5,(6) 2-5. Contours1,2, 3 and 4 are theboundariesofan 80,95, 99 and 99 9 percentconfidenceregion,
respectively, for(X1,X2).

curveis showninFig.3(a).Fig.3(b)displays
plotsofapproximations
to Sm(Xi)andSm(X2),
the
minimum valuesofS forfixedX1and X2 respectively.
Theseapproximations wereobtained
from themapofSm(Xi,X2),and from regionsforX1andX2
Fig.3(b)the95percentconfidence
separatelyare obtained as those values satisfyingSm(Xi)0227 (i = 1,2), namely
1-25<X1 < 1 68 and 1 72< X2< 2X35
5. DiscussioN
CalculatingSm(X)
Ifallf inthemodel(1)arelinear,
anefficient
wayofperforming calculations
thenecessary is
tofitthemodelwithout
first imposing (2),andthento useLagrangian
constraints methods to
obtainthe adjustmentsto the residualsum of squares,and to the values of P,..., PSr
when
continuity X isimposed.
ata specific Themethod becausethesameunconstrained
isefficient fit
willapplytoa number ofX-valuesonthegrid.Thus,itneedonlybecarriedoutonceforthoseX,
andseveralvaluesofSm(X)maybe obtainedbytherelatively process.The
quickconstraining
necessarycalculations Iftheleast-squares
areas follows. parameter andresidual
estimates sum
ofsquaresfora linearmodely = Xp+ e aredenotedby andS,thentheadjustedvalueswhen
constraintsHp = c areimposedare(Plackett,1960,p. 53)
=
p p-(X'WX)' H'[H(X'WX)- H']- 1(H-c),
(10)
S = S + (H-c)'[H(X'WX) H']-(Hp-c),
whereWis thediagonalmatrixofweights
(Wjj= wj).
Ifanyofthefti evenwhenX isfixed,
arenon-linear, thevalueofSm(X)willusuallyneedtobe
byiterative
determined minimization
methods.Insuchcasestheusualproblems with
associated

This content downloaded from 192.122.237.41 on Sat, 20 Jun 2015 03:40:33 UTC
All use subject to JSTOR Terms and Conditions
84 APPLIED STATISTICS

regression
non-linear Ross(1970)givesa gooddiscussion
willbeencountered. use
oftheefficient
methods
ofiterative leastsquares.
in non-linear
Theimposition ofcontinuity withnon-linearJi
constraints of
can be handledin a variety
ways(seeAdbyandDempster, 1974,Chapter5).A simplemethodistoreducetheproblem toan
minimization
unconstrained byincorporating intothemodel(seeExample2
directly
continuity
above),butthisis notalwayspossible.
DisplayingSm(X)
TheshapeofSm(X)iseasilyapprehended ifitisdisplayed graphicallyratherthanintabular
form.Thispresents ifthenumber
nodifficulties oftransitionpointsis oneortwo;Sm(Xi)can be
plotted X1ifthereis onlyonetransition
against (Figs.l(b),2(b)),andiftherearetwo,a contour
mapofSm(Xi,X2)can be produced(Fig.4). For threeor morebreaksitis simplest to display
lower-dimensional ofSm(X).A sensible
extracts choiceis Sm(V),theminimum valueofS forthe
V,oftransition
fixedvalueofa subvector, required
points(Fig.3(b)),sinceitis thefunction for
derivingconfidenceregionsforV. The valueofSm(V)can be approximated fromthemapof
Sm(X)provided thegridis suitablyfineand spansthe-required minimum.
ofX
Final refinement
SinceS maynotbe a stationarypointofS,itslocationfromtheapproximationrevealedby
themapisbestachieved bya techniquewhichdoesnotemploy Linearsearch(Adby
derivatives.
andDempster, point,andpattern
ifthereis onlyonetransition
1974)is suitable search(Powell,
1964;Adbyand Dempster, 1974)fortwoor morebreaks.
REFERENCES
ADBY,P. R. and DEMPSTER, to Optimization
M. A. H. (1974). Introduction Methods.London: Chapman and Hall.
BACON, D. W. and WATTS, betweentwointersecting
D. G. (1971).Estimatingthetransition straight 58,
lines.Biometrika,
525-534.
BEALE,E. M. L. (1960). Confidenceregionsin non-linearestimation.J. R. Statist.Soc. B, 22, 41-81.
DRAPER,N. R. and SMITH,H. (1966). AppliedRegressionAnalysis.London: Wiley.
DUNICZ,B. L. (1969). Discontinuitiesin the surfacestructureof alcohol-watermixtures.Sond. aus der Kolloid-
Zeitschrift undZeitschriftfuir Polymere,230, 346-357.
FEDER,P. I. (1975a). The log likelihoodratio in segmentedregression.Ann.Statist.,3, 84-97.
(1975b). On asymptoticdistribution theoryin segmentedregressionproblems-identified case. Ann.Statist.,3,
49-83.
HAWKINS, D. M. (1976). Point estimationof the parametersof piecewiseregressionmodels.Appl.Statist.,25, 51-57.
HINKLEY,D. V. (1969). Inferenceabout the intersection in two-phaseregression.Biometrika, 56, 495-504.
(1971). Inferencein two-phaseregression.J. Amer.Statist.Ass.,66, 736-743.
HUDSON,D. J. (1966). Fittingsegmentedcurveswhosejoin pointshave to be estimated.J. Amer.Statist.Ass.,61,
1097-1129.
KIRBY,E. J. M. (1974). Ear developmentin springwheat.J. agric.Sci., Camb.,82, 437-447.
OWEN,J.B., DAVIES,D. A. R. and RIDGMAN,W. J.(1969).The controlofvoluntaryfoodintakein ruminants. Animal
Production,11, 511-520.
PLACKETT,R. L. (1960). PrinciplesofRegressionAnalysis.London: OxfordUniversity Press.
POWELL, M. J. D. (1964). An efficient methodforfindingthe minimumof a functionof severalvariableswithout
calculatingderivatives.ComputerJ.,7, 155-162.
Ross, G. J.S. (1970). The efficientuse of functionminimizationin non-linearmaximumlikelihoodestimation.Appl.
Statist.,19, 205-221.
SINGPURWALLA,N. D. (1974).Estimationofthejoin pointin a heteroscedastic regressionmodelarisingin accelerated
lifetests.Commun.Statist.,3 (9), 853-863.
SPRENT, P. (1961). Some hypothesesconcerningtwo-phaseregressionlines.Biometrics, 17, 634-645.
WILLIAMS,D. A. (1970). Discriminationbetweenregressionmodels to determinethe patternof enzymesynthesis in
synchronouscell cultures.Biometrics, 26, 23-32.

This content downloaded from 192.122.237.41 on Sat, 20 Jun 2015 03:40:33 UTC
All use subject to JSTOR Terms and Conditions