Lectures on Structural Change

Eric Zivot Department of Economics, University of Washington April 5, 2003

Overview of Testing for and Estimating Structural Change in Econometric Models

1. Day 1: Tests of Parameter Constancy 2. Day 2: Estimation of Models with Structural Change 3. Day 3: Time Varying Parameter Models

Some Preliminary Asymptotic Theory

Reference: Stock, J.H. (1994) Unit Roots, Structural Breaks and Trends, in Handbook of Econometrics, Vol. IV.

3
3.1

Tests of Parameter Constancy in Linear Models

Motivation
Diagnostics for model adequacy Provide information about out-of-sample forecasting accuracy Within-sample parameter constancy is a necessary condition for superexogeneity

3.2
3.2.1

Example Data Sets

Simulated Data

Consider the linear regression model yt = + xt + t , t = 1, . . . , T = 200 xt iid N (0, 1) t iid N (0, 2 ) 1

No structural change parameterization: = 0, = 1, = 0.5 Structural change cases Break in intercept: = 1 for t > 100 Break in slope: = 3 for t > 100 Break in error variance: = 0.25 for t > 100 Random walk in slope: = t = t1 + t , t iid N (0, 0.1) and 0 = 1. (show simulated data) 3.2.2 Let st ft = = log of spot exchange rate in month t log of forward exchange rate in month t US/DM Monthly Exchange rate data

The forward rate unbiased hypothesis is typically investigated using the so-called dierences regression st+1 ft st = + (ft st ) + t+1 S = iU iDM = forward discount t t

If the forward rate ft is an unbiased forecast of the future spot rate st+1 then we should nd = 0 and = 1 The forward discount is often modeled as an AR(1) model ft st = + (ft1 st1 ) + ut Statistical Issues st+1 is close to random walk with large variance ft st behaves like highly persistent AR(1) with small variance ft st appears to be unstable over time

3.3

Chow Forecast Test

Reference: Chow, G.C. (1960). Tests of Equality between Sets of Coecients in Two Linear Regressions, Econometrica, 52, 211-22. Consider the linear regression model with k variables yt y
2 = x0 t + ut , ut (0, ), t = 1, . . . , n = X + u

Parameter constancy hypothesis H0 : is constant Intuition If parameters are constant then out-of-sample forecasts should be unbiased (forecast errors have mean zero) Test construction: Split sample into n1 > k and n2 = n n1 observations n1 n1 X1 y1 ,X = y= y2 n2 X2 n2 Fit model using rst n1 observations 1 u 1 2 1
1 0 = (X0 X1 y1 1 X1 ) = y1 X1 1 = u 0 1 /(n1 k ) 1u

1 and X2 to predict y2 using next n2 observations Use 1 y 2 = X2 Compute out-of-sample prediction errors 1 2 = y2 X2 u 2 = y2 y Under H0 : is constant 1 ) u 2 = u2 X2 ( and 1 0 var(u 2 ) = 2 In2 + X2 (X0 X2 1 X1 ) 2 )) u 2 N (0, var(u 0 1 u 2 var(u 2 ) u 2 2 (n2 ) 2 (n1 k) 1 / 2 2 (n1 k ) This motivates the Chow forecast test statistic 0 1 0 X2 u 2 u 0 2 In2 + X2 (X1 X1 ) F (n2 , n1 k ) ChowF CST (n2 ) = 2 n2 1 3 E [u 2 ] = 0

Further, If the errors u are Gaussian then

Decision: Reject H0 at 5% level if ChowF CST (n2 ) > cv0.05 Remarks: Test is a general specication test for unbiased forecasts Popular with LSE methodology Implementation requires a priori splitting of data into t and forecast samples 3.3.1 Application: Simulated Data Chow Forecast Test n2 Model 100 50 No SC 1.121 1.189 Mean shift 9.130*** 1.329* Slope shift 9.055*** 2.067*** Var shift 0.568 0.726 RW slope 2.183*** 1.302

25 1.331 1.061 1.545* 0.864 0.550

3.4

CUSUM and CUSUMSQ Tests

Reference: Brown, R.L., J. Durbin and J.M. Evans (1975). Techniques for Testing the Constancy of Regression Relationships over Time, Journal of the Royal Statistical Society, Series B, 35, 149-192. 3.4.1 Recursive least squares estimation

The recursive least squares (RLS) estimates of are based on estimating yt = 0 t xt + t , t = 1, . . . , n by least squares recursively for t = k + 1, . . . , n giving n k least squares (RLS) k+1 , . . . , T ). estimates ( RLS estimates may be eciently computed using the Kalman Filter t should quickly settle down near a com If is constant over time then mon value.

3.4.2

Recursive residuals

Formal tests for structural stability of the regression coecients may be computed from the standardized 1 step ahead recursive residuals wt ft Intuition: 0 yt vt t1 xt = = ft ft h i 0 1 = 2 1 + x0 xt t (Xt Xt )

If i changes in the next period then the forecast error will not have mean zero wt are recursive Chow Forecast t-statistics with n2 = 1 3.4.3 CUSUM statistic

The CUSUM statistic of Brown, Durbin and Evans (1975) is CU SU Mt 2 w =

t X w j w

j =k+1

n 1 X (wt w )2 n k t=1

Under the null hypothesis that is constant, CU SU Mt has mean zero and variance that is proportional to t k 1. 3.4.4 CUSUMSQ statistic Pt

THE CUSUMSQ statistic is

j =k+1 CU SU M SQt = Pn j =k+1 2 w j 2 w j

Under the null that is constant, CU SU M SQt behaves like a 2 (t) and condence bounds can be easily derived. 3.4.5 Application: Simulated Data (insert graphs here) Remarks Ploberger and Kramer (1990) show the CUSUM test can be constructed with OLS residuals instead of recursive residuals 5

 CUSUM Test is essentially a test to detect instability in intercept alone CUSUM Test has power only in direction of the mean regressors CUSUMSQ has power for changing variance There are tests with better power 3.4.6 Application: Exchange Rate Regression (insert graphs here)

3.5

Nybloms Parameter Stability Test

Reference:Nyblom, J. (1989). Testing for the Constancy of Parameters Over Time, Journal of the American Statistical Association, 84 (405), 223-230. Consider the linear regression model with k variables yt = x0 t + t , t = 1, . . . , n The time varying parameter (TVP) alternative model assumes = t = t1 + t , it (0, 2 i ), i = 1, . . . , k The hypotheses of interest are H0 H1 : is constant 2 i = 0 for all i : 2 i > 0 for some i

Nyblom (1989) derives the locally best invariant test as the Lagrange multiplier test. The score assuming Gaussian errors is
n X t=1

xt t = 0 t = yt x0 t = (X0 X)1 X0 y

Dene ft St V Note that = xt t t X = ft = cumulative sums

j =1

= n1 X0 X
n X j =1

ft = 0

Nyblom derives the LM statistic L = =

n 1 X St V1 St n 2 t=1 " # n X 1 1 0 tr V St St n 2 t=1

Under mild assumptions regarding the behavior of the regressors, the limiting distribution of L under the null is a Camer-von Mises distribution: Z 1 0 L B k ()Bk () d
0

B k () Wk ()

= =

Wk () Wk (1) k dimensional Brownian motion

Decision: Reject H0 at 5% level if L > cv0.05 Remarks: Distribution of L is non-standard and depends on k. Critical values are computed by simulation and are given in Nyblom, Hansen (1992) and Hansen (1997) Test is for constancy of all parameters Test is not informative about the date or type of structural change Test is applicable for models estimated by methods other than OLS Distribution of L is dierent if xt is non-stationary (unit root, deterministic trend). See Hansen (1992). 3.5.1 Application: Simulated Data Nyblom Model No SC Mean shift Slope shift var shift RW slope Test Lc .332 13.14 14.13 .351 9.77

3.5.2

Application: Exchange rate regression Nyblom Test Model Lc AR(1) 1.27 Di reg .413

3.6

Hansens Parameter Stability Tests

References 1. Hansen, B.E. (1992). Testing for Parameter Instability in Linear Models Journal of Policy Modeling, 14(4), 517-533. 2. Hansen, B.E. (1992). Tests for Parameter Instability in Regressions with I(1) Processes, Journal of Business and Economic Statistics, 10, 321-336. Idea: Extension of Nybloms LM test to individual coecients. Under the null of constant parameters, the score vector from the linear model with Gaussian errors is
n X t=1

xit t

= 0, i = 1, . . . , k

X 2 2) = 0 ( t t 2 Dene fit Sit Note that = = t xit 2 2 t

= yt x0 t n X = n1 2 t
t=1

i = 1, . . . k i=k+1

t X j =1

fij , i = 1, . . . , k + 1

n X i=1

fit = 0, i = 1, . . . , k + 1

3.6.1

Individual Coecient Tests

Hansens LM test for H0 : i is constant, i = 1, . . . , k 8

and for H0 : 2 is constant is Li Vi = =

n 1 X 2 S , i = 1, . . . , k nVi t=1 it 2 fit

n X t=1

Under H0 : i is constant or H0 : 2 is constant Li Z

1 B1 ()B1 ()d

Decision: Reject H0 at 5% level if Li > cv0.05 = 0.470 3.6.2 Joint Test for All Coecients

For testing the joint hypothesis H0 : and 2 are constant dene the (k + 1) 1 vectors ft St = (f1t , . . . , fk+1,t )0 = (S1t , . . . , Sk+1,t )0

Hansens LM statistic for testing the constancy of all parameters is ! n n X 1 X 0 1 1 1 0 S V St = tr V St St Lc = n t=1 t n t=1 V =
n X t=1

ft ft0

Under the null of no-structural change Z 1 B Lc k+1 ()Bk+1 ()d

0

Decision: Reject H0 at 5% level if Lc > cv0.05 Remarks 9

 Tests are very easy to compute and are robust to heteroskedasticity Null distribution is non-standard and depends upon number of parameters tested for stability Individual tests are informative about the type of structural change Tests are not informative about the date of structural change Hansens L1 test for constancy of intercept is analogous to the CUSUM test Hansens Lk+1 test for constancy of variance is analogous to CUSUMSQ test Hansens Lc test for constancy of all parameters is similar to Nyboloms test Distribution of tests is dierent if data are nonstationary (unit root, deterministic trend) - see Hansen (1992), JBES. 3.6.3 Application: Simulated Data Model No SC Mean shift Slope shift var shift RW slope 3.6.4 Hansen Tests .179 .134 13.19 .234 .588 5.11 .226 .119 .253 4.08 2 .248 .064 .067 .376 .196 Joint .503 13.3 5.25 .736 4.4

Application: Exchange rate regression contd Model AR(1) Di reg Hansen Tests intercept slope variance .382 .147 2.94 .104 .153 .186 Joint 3.90 .520

Tests for Single Structural Change

yt = x0 t t + t , t = 1 . . . , n

Consider the linear regression model with k variables

No structural change null hypothesis H0 : t =

10

Single break date alternative hypothesis t = , t m = break date H1 : t = + , t > m and 6= 0 k < m<nk m = = break fraction n Remarks: Under no break null = 0. Pure structural change model: all coecients change ( i 6= 0 for i = 1, . . . , k) Partial structural change model: some coecients change ( i 6= 0 for some i) m = [ n], [] = integer part

4.1

Chows Test with Known Break Date

Assume: m or is known For a data interval [r, . . . , s] such that s r > k dene r,s = OLS estimate of r,s = OLS residual vector 0 r,s = sum of squared residuals SSRr,s = r,s Chows breakpoint test for testing H0 vs. H1 with m known is m (SSR1,n (SSR1,m + SSRm+1,n ))/k Fn = Fn () = n (SSR1,m + SSRm+1,n ) /(n 2k ) The Chow test may also be computed as the F-statistic for testing = 0 from the dummy variable regression
0 yt = x0 t + Dt (m)xt + t Dt (m) = 1 if t > m; 0 otherwise

Under H0 : = 0 with m known Fn () F (k, n 2k)

d

k Fn () 2 (k) Decision: Reject H0 at 5% level if

Fn () > F0.95 (k, n k ) k Fn () > 2 0.95 (k ) 11

4.1.1

Application: Simulated Data Chow Breakpoint Test F200 (0.5) F200 (0.25) 0.808 0.081 377*** 13.03*** 374*** 10.97*** 1.071 0.117 80.14*** 4.058** F200 (0.75) 1.55 11.21*** 17.57*** 1.204 2.218

No SC Mean shift Slope shift Var shift RW slope

4.2

Quandts LR Test with Unknown Break Date

References: 1. Quandt, R.E. (1960). Tests of Hypotheses that a Linear System Obeys Two Separate Regimes, Journal of the American Statistical Association, 55, 324-330. 2. Davies, R.A. (1977). Hypothesis Testing When a Nuisance Parameter is Present only Under the Alternative, Biometrika, 64, 247-254. 3. Kim, H.-J., and D. Siegmund (1989). The Likelihood Ratio Test for a Change-Point in Simple Linear Regression, Biometrika, 76, 3, 409-23. 4. Andrews, D.W.K. (1993). Tests for Parameter Instability and Structural Change with Unknown Change Point, Econometrica, 59, 817-858. 5. Hansen, B.E. (1997). Approximate Asymptotic P Values for StructuralChange Tests, Journal of Business and Economic Statistics, 15, 60-67. Assume: m or is unknown. Quandt considered the LR statistic for testing H0 : = 0 vs. H1 : 6= 0 when m is unknown. This turns out to be the maximal Fn () statistic over a range of break dates m0 , . . . , m1 : m QLR = max Fn = max Fn () n m[m0 ,m1 ] [0 ,1 ] mi i = = trimming parameters, i = 0, 1 n Remarks QLR is also know as Andrews sup F statistic Trimming parameters 0 and 1 must be set Cannot have 0 = 1 and 1 = 1 because breaks are hard to identify near beginning and end of sample Information about location of break can be used to specify 0 and 1 12

 Andrews recommends 0 = 0.15 and 1 = 0.85 if there is no knowledge of break date Implicitly, the break data m and break fraction are estimated using m m = arg max Fn m n = m/n Under the null, m dened under the alternative is not identied. This is an example of the Davies problem. Davies (1977) showed that if estimated parameters are unidentied under the null, standard 2 inference does not obtain. Under H0 : = 0, Kim and Siegmund (1989) showed k QLR B k () =
0 B k () Bk () (1 ) [0 ,1 ]

sup

Wk () Wk (1) = Brownian Bridge k QLR > cv0.05

Decision: Reject H0 at 5% level if Remarks Distribution of QLR is non-standard and depends on the number of variables k and the trimming parameters 0 and 1 Critical values for various values of 0 and 1 computed by simulation are given in Andrews (1993), and are larger than 2 (k ) critical values. For 0 = 0.15 and 1 = 0.85 5% k 1 10 critical values 2 (k) QLR 3.84 8.85 18.3 27.03 k QLR

P-values can be computed using techniques from Hansen (1997) Graphical plot of Fn () statistics is informative to locate the break date 4.2.1 Application: Simulated data QLR or sup-F Test QLR m b No SC 2.87 142 Mean shift 377*** 101 Slope shift 374*** 101 Var shift 2.36 142 RW slope 113*** 77 (insert graphs here) 13 b 0.71 0.51 0.51 0.71 0.39

4.2.2

Application: Exchange rate data QLR or sup-F Test QLR m b AR(1) 12.13*** 1989:05 Di reg 4.08 1991:03 b 0.65 0.74

4.3

Optimal Tests with Unknown Break Date

References: 1. Andrews, D.W.K. and W. Ploberger (1994). Optimal Tests When a Nuisance Parameter Is Present Only Under the Alternative, Econometrica, 62, 1383-1414. Andrews and Ploberger (1994) derive tests for structural change with an unknown break date with optimal power. These tests turn out to be weighted averages of the Chow breakpoint statistics Fn ( m n ) used to compute the QLR statistic: ! m2 X 1 1 t ExpFn = ln exp k Fn m2 m1 + 1 t=m 2 n 1 m2 X 1 t k Fn AveFn = m2 m1 + 1 t=m n
1

number of regressors being tested

Remarks Asymptotic null distributions are non-standard and depend on k, 0 and 1 Critical values are given in Andrews and Ploberger; P-values can be computed using techniques of Hansen (1997) Tests can have higher power than QLR statistic Tests are not informative about location of break date

4.4

Empirical Application

Reference: Stock and Watson (199?), Journal of Business and Economic Statistics.

14