Stability

Coefficientsmaychangeovertime
Evolutionoftheeconomy
Policychanges

TimeVaryingParameters
yt = t + xt t + et
Coefficientsdependonthetimeperiod
Ifthecoefficientsvaryrandomlyandare
unpredictable,thentheycannotbeestimated
Astherewouldbeonlyoneobservationforeach
setofcoefficients
Wecannotestimatecoefficientsfromjustone
observation!

SmoothlyTimeVaryingParameters
yt = t + xt t + et
Ifthecoefficientschangegraduallyovertime,
thenthecoefficientsaresimilarinadjacenttime
periods.
Wecouldtrytoestimatethecoefficientsfortime
periodt byestimatingtheregressionusing
observations[t w/2,, t+ w/2]wherew is
calledthewindowwidth.
w isthenumberofobservationsusedforlocal
estimation

RollingEstimation
Thisiscalledrolling estimation
Foragivenwindowwidthw,yourollthrough
thesample,usingw observationsfor
estimation.
Youadvanceoneobservationatatimeand
repeat
Thenyoucanplottheestimatedcoefficients
againsttime

Whattoexpect
Rollingestimateswillbeacombinationoftrue
coefficientsandsamplingerror
Thesamplingerrorcanbelarge
Fluctuationsintheestimatescanbejusterror

Ifthetruecoefficientsaretrending
Expecttheestimatedcoefficientstodisplaytrendplus
noise

Ifthetruecoefficientsareconstant
Expecttheestimatedcoefficientstodisplayrandom
fluctuationandnoise

Example:GDPGrowth

STATArolling command
STATAhasacommandforrollingestimation:
.rolling,window(100)clear:regressgdp L(1/3).gdp
Inthiscommand:
window(100) setsthewindowwidth
w=100
Thenumberofobservationsforestimationwillbe100

clear
Clearsoutthedatainmemory
Thedatawillbereplacedbytherollingestimates
Itisnecessary

rolling command
.rolling,window(100)clear:regressgdp L(1/3).gdp
Thepartafterthe:
regressgdp L(1/3).gdp
ThisisthecommandthatSTATAwillimplementusing
therollingmethod
AnAR(3)willbefitusing100observations,rolling
throughthesample

Example
GDP,quarterly,1947Q1through2009Q4
251observations

Usingw=100
Thefirstestimationwindowis1947Q21972Q1
Thesecondis1947Q31972Q2
Thereare152estimationwindows
Thefinalis1985Q12009Q4

 STATAExecution:

AfterRollingExecution
Theoriginaldatahavebeenclearedfrommemory
STATAshowsnewvariables

start
end
_stat_1
_stat_2
_stat_3
_b_cons

start andend arestarting/endingdatesforeachwindow

start runsfrom1947Q2to1985Q1
end runsfromto1972Q12009Q4

Theothersaretherollingestimates,ARandintercept

Timereset
Astheoriginaldatahavebeencleared,sohasyourtime
index.
Sothetsline commanddoesnotworkuntilyouresetthe
time
Youcansetthetimetobestartorend
.tsset start
.tsset end

Or,moreelegantly,youcansetthetimetobethemid
pointofthewindow

.gent=round((start+end)/2)
.formatt%tq
.tsset t
Thistimeindexrunsfrom1959Q4through1997Q3

Timeresetexample
Example

PlotRollingCoefficients
Nowyoucanplottheestimatedcoefficients
againsttime
Youcanuseseparateorjointplots
.tsline _b_cons
.tsline _stat_1_stat_2_stat_3

RollingIntercept

RollingARcoefficients

Analysis
Theestimatedinterceptis
decreasinggradually
TheAR(1)coef isquitestable
TheAR(2)coef startsincreasing
around1990
TheAR(3)coef is0mostofthe
period,butisnegativefrom
19601973andafter1995
Allofthegraphsgoabitcrazy
over19901997

Sequential(Recursive)Estimation
Asanalternativetorollingestimation,
sequential orrecursive estimationusesallthe
datauptothewindowwidth
Firstwindow:[1,w]
Secondwindow:[1,w+1]
Finalwindow:[1,T]

Withsequentialestimation,window isthe
lengthofthefirstestimationwindow

RecursiveEstimation
STATAcommandissimilar,butaddsrecursive aftercomma
.rolling,recursivewindow(100)clear:regressgdp
L(1/3).gdp
STATAclearsdataset,replaceswithstart,end,and
recursivecoefficientestimates_b_cons,_stat_1,etc.
Useend fortimevariable
.tsset end
Thissetsthetimeindextotheendperiodusedforestimation

RecursiveIntercept

RecursiveARcoefficients

Analysis
Therecursiveintercept
fluctuates,butdecreases
Dropsaround1984,and1990

TherecursiveAR(1)and
AR(2)coefs areverystable
TherecursiveAR(3)coef
increases,andthenbecomes
stableafter1984.

Summary
Userollingandrecursiveestimationto
investigatestabilityofestimatedcoefficients
Lookforpatternsandevidenceofchange
Trytoidentifypotentialbreakdates
InGDPexample,possibledates:
1970,1984,1990

TestingforBreaks
Didthecoefficientschangeatsomebreakdate
t*?
Wecantestifthecoefficientsbeforeandafter
t* arethesame,oriftheychanged
SimpletoimplementasanF testusing
dummyvariables
KnownasaChowtest

GregoryChow
ProfessorGregoryChowof
PrincetonUniversity
(emeritus)
ProposedtheChowTest
forstructuralchangeina
famouspaperin1960

DummyVariable
Foragivenbreakdate t*
Defineadummyvariabled
d=1ift>t*

Included andinteractionsd*x totestfor

changes

ModelwithBreaks
OriginalModel

yt = + xt + et

Modelwithbreak

yt = + xt + d t + d t xt + et
Interpretingthecoefficients
=changeinintercept
=changeinslope

ChowTest
yt = + xt + d t + d t xt + et
Themodelhasconstantparametersif==0
Hypothesistest:
H0:=0and=0

ImplementasanF testafterestimation
Ifprob>.05,youdonotrejectthehypothesisof
stablecoefficients

Example:GDP

Chowtest

Thepvalueislargerthan0.05
Itisnotsignificant
Wedonotrejecthypothesisofconstantcoefficients

FishingforaBreakdate
AnimportanttroublewiththeChowtestis
thatitassumesthatthebreakdate isknown
beforelookingatthedata
Butweselectedthebreakdate byexamining
rollingandrecursiveestimates
Thismeansthataretoolikely tofind
misleadingevidenceofnonconstant
coefficients

Fishing
Wecouldconsiderpickingmultiplepossible
breakdates t*=[t1,t2,,tM]
Foreachbreakdate t*,wecouldestimatethe
regressionandcomputetheChowstatistic
F(t*)
Fishingforabreakdate issimilartosearching
forabig(significant)Chowstatistic.

TheQuandtLikelihodRatio(QLR)Statistic
(alsocalledthesupWaldstatistic)

The QLR statistic = the maximal Chow statistics

Let F() = the Chow test statistic testing the hypothesis of no

break at date .
The QLR test statistic is the maximum of all the Chow Fstatistics, over a range of , 0 1:
QLR = max[F(0), F(0+1) ,, F(11), F(1)]
A conventional choice for 0 and 1 are the inner 70% of the
sample (exclude the first and last 15%.
34

RichardQuandt
ProfessorRichardQuandt (1930)
PrincetonUniversity
Estimationofbreakdate (Quandt,1958)
QLRtest(Quandt,1960)

QLRCriticalValues
QLR = max[F(0), F(0+1) ,, F(11), F(1)]

Should you use the usual critical values?

The large-sample null distribution of F() for a given (fixed,
not estimated) is Fq,
But if you get to compute two Chow tests and choose the
biggest one, the critical value must be larger than the critical
value for a single Chow test.

If you compute very many Chow test statistics for example,

all dates in the central 70% of the sample the critical value
36
must be larger still!

 Get this: in large samples, QLR has the distribution,

1 q Bi ( s ) 2
max a s1a
,

q i =1 s(1 s )

where {Bi}, i =1,,n, are independent continuous-time

Brownian Bridges on 0 s 1 (a Brownian Bridge is a
Brownian motion deviated from its mean), and where a = .15
(exclude first and last 15% of the sample)

Critical values are tabulated in SW Table 14.6

37

Note that these critical values are larger than the Fq, critical
values for example, F1, 5% critical value is 3.84.
38

QLRTheory
DistributiontheoryfortheQLRstatistic
Developedby
ProfessorDonaldAndrews(Yale)

HasthepostwarU.S.PhillipsCurvebeen
stable?
Consider a model of Inft given Unempt the empirical
backwards-looking Phillips curve, estimated over (1962 2004):
Inf = 1.30 .42Inft1 .37Inft2 + .06Inft3 .04Inft4

(.44) (.08)

(.09)

(.08)

(.08)

2.64Unemt1 + 3.04Unemt2 0.38Unemt3 + .25Unempt4

(.46)
(.86)
(.89)
(.45)
Has this model been stable over the full period 1962-2004?
40

QLRtestsofstabilityofthePhillipscurve.
dependent variable: Inft
regressors: intercept, Inft1,, Inft4,
Unempt1,, Unempt4
test for constancy of intercept only (other coefficients are
assumed constant): QLR = 2.865 (q = 1).
10% critical value = 7.12 dont reject at 10% level
test for constancy of intercept and coefficients on Unempt,,
Unempt3 (coefficients on Inft1,, Inft4 are constant):
QLR = 5.158 (q = 5)
1% critical value = 4.53 reject at 1% level
Break date estimate: maximal F occurs in 1981:IV
Conclude that there is a break in the inflation
unemployment relation, with estimated date of 1981:IV
41

42

Implementation
ItisdifficulttocomputeQLRwithoutusing
someprogramming.
Butitiswellapproximatedby
Examiningrollingandrecursiveestimatesfor
possiblebreaks
ComputingChowtestatpotentialbreakdates.

DontuseSTATAspvalue!
UseTable14.6fromSW(orearlierslide).