Structural modelling: Causality, exogeneity and unit roots Andrew P. Blake CCBS/HKMA May 2004 What do we need to do with our data? • Estimate structural equations (i.e. understand what’s happening now) • Forecast (i.e. say something about what’s likely to happen in the future) • Conduct scenario analysis (i.e. perform simulations) to inform policy What do we need to know? • Inter-relationships between variables – Causality in the Granger sense – Exogeneity • Concepts – Unit roots • Spurious regression • Role of pre-testing • Appropriate single equation methods 0.08 0.04 0.00 -0.04 -0.08 94 95 96 97 98 99 00 01 X Y Inter-relationships between variables Period t Period t+1 xt xt+1 yt yt+1 How best to estimate an equation? • Single equation structural model (estimated by OLS) • Single equation reduced form (IV/OLS) • Structural system (estimated by TSLS, 3SLS or by a system method - SUR, FIML) • Unrestricted VAR (OLS) • VECM (FIML) xt is autoregressive Period t Period t+1 xt xt+1 yt yt+1 xt has an autoregressive representation Period t Period t+1 xt xt+1 yt yt+1 xt has an ARMA representation xt yt t } yt yt 1 t yt 1 1 x t 1 Structural system t 1 yt xt 1 t 1 t so, xt xt 1 t t t 1 Reduced form Granger Causality Period t Period t+1 xt xt+1 yt yt+1 Vector autoregressions (VARs) Period t Period t+1 xt xt+1 yt yt+1 Needs to be modelled to have a structural interpretation Granger causality • If past values of y help to explain x, then y Granger causes x • Statistical concept • A lack of Granger causality does not imply no causal relationship GC tested by an unrestricted VAR xt a11 xt 1 a12 yt 1 b11 xt 2 b12 yt 2 ... t yt a21 xt 1 a22 yt 1 b21 xt 2 b22 yt 2 ... t • Definition of Granger Causality: – y does not Granger cause x if a12=b12=...=0 – x does not Granger cause y if a21=b21=...=0 • NB. x and y could still affect each other in the same period or via unmeasured common shocks to the error terms. Eviews Granger causality test result Null Hypothesis x does not Granger Cause y y does not Granger Cause x F-Statistic Probability F1 F2 P1 P2 • The closer P1 is to zero, the less the likelihood of accepting the null that x does not Granger cause y. • (P1<0.10 : at least 90% confident that s1 Granger causes s2). • P1 should be less than 0.10 for us to be reasonably confident that x Granger causes y. Leading indicators y is a leading indicator of x if • y Granger causes x; • x does not Granger cause y; • and y is weakly exogenous. 73 .2 74 .2 75 .2 76 .2 77 .2 78 .2 79 .2 80 .2 81 .2 82 .2 83 .2 84 .2 85 .2 86 .2 87 .2 88 .2 89 .2 90 .2 91 .2 92 .2 93 .2 94 .2 95 .2 96 .2 97 .2 % on year earlier, smoothed, prices lagged 6 quarters Long term trends of money and prices in UK 30.0 25.0 20.0 15.0 10.0 5.0 0.0 Broad Money Prices Criticisms of Granger causality • Granger causality can be assessed using an unrestricted VAR - not tied to any particular theory • How would you explain to your governor when it goes wrong? • It depends on the choice of lags, data frequency and variables in VAR Exogeneity • Engle et al. (1983) – Separate parameters into two groups – Those that matter, those that don’t • These are endogenous and weakly exogenous variables • In practice a bit more complicated than that Exogeneity (cont.) • Correct assumptions of exogeneity simplify modeling, reduce computational expense and aid interpretation • But incorrect assumptions may lead to inefficient or inconsistent estimates and misleading forecasts Exogeneity (cont.) • A variable is exogenous if it can be taken as given without losing information for the purpose at hand • This varies with the situation • We do not want the independent variables to be correlated with the regressors • If they are, the estimates will be biased Relationships between variables Period t Period t+1 xt xt+1 yt yt+1 • We do not want the black arrows • We need to understand the red arrows Both demand and supply shocks 14 12 P 10 8 6 4 2 0 0 1 2 3 4 5 6 Q OLS is unable to identify either the demand or supply curve Only supply shocks 14 12 P 10 8 6 4 2 0 0 1 2 3 4 5 Q We can identify the demand schedule using OLS 6 Weak exogeneity • Is y weakly exogenous with respect to x? • Do values of current x affect current y? • Are x and y both affected by a common unmeasured third variable? • Does the range of possible values for the parameters in the process that determines x affect the possible values of those that determine y Weak exogeneity: example 1 • Money demand function: mt yt rt • Would you estimate this as a single equation using OLS? • Very unlikely that money does not affect real output or the nominal interest rate Weak exogeneity: example 2 • Uncovered interest parity: E t et 1 rt rt * • Tests of UIP have performed very poorly, but ... • No risk premia and monetary policy might react to exchange rate changes Interest rate differentials Exchange rate change Question: how would you test for exogeneity in UIP? Weak exogeneity: example 3 • In UK consumption had been forecast using single-equation ECM • But relationship broke down in late 1980s • Problem was that possibility that wealth reactions to disequilibrium had been ignored Single Equation ECM yt yt 1 xt xt 1 ... Dynamic terms ... yt 1 xt 1 Long run Vector ECMS xt 1yt 11xt 1 12 yt 1 ... 1 ECM yt 2 xt 21xt 1 22 yt 1 ... 2 ECM Halfway between structural VARs and unrestricted VARs Strong exogeneity • Necessary for forecasting • Is y strongly exogenous to x? – Is y weakly exogenous to x – Does x Granger cause y? • Need the answers to be yes and no respectively Strong exogeneity: example First order VAR, ‘core’ and non-‘core’ inflation: zt Azt-1 t , zt xt , yt ' Given a forecast of {yt} can we forecast {xt}? • If y is not strongly exogenous to x, feedback problems Super exogeneity Necessary for policy/scenario analysis. Is y super exogenous to x? • Is y weakly exogenous to x? • Is the relationship between x and y invariant? Need the answers to be yes to both Invariance • The process driving a variable does not change in the face of shocks • Linked to ‘deep parameters’ • Example: the Lucas critique Testing for weak exogeneity: orthogonality test • Estimate a reduced form (marginal model) for x, regress x on any exogenous variables of the system • Take residuals from this reduced form and put them into the structural equation for y • If they are significant then x is not weakly exogenous with respect to the estimation of c10 Testing for weak exogeneity with respect to c(lr) • Estimate a reduced form (marginal model) for x: regress x on exogenous variables of system, including lagged ECM term involving x and y • Test if coefficient of ECM term is significant • If it is, then x is not weakly exogenous with respect to the estimation of long-run coeff, c(lr) • Consequence is that estimate is inefficient Stationarity • Why should we test whether series are stationary? • A non-stationary time series implies that shocks never die out • The mean, variance and higher moments depend on time • Standard statistics do not have standard distributions • Problem of spurious regression Non-stationarity • Start with the following expression yt = + yt-1 + ut u, 2 • Substitute recursively: yt = n + n yt-n + n-1jut-j • The variable will be non-stationary if = E(y)=t Var(y) = Var(n-1ut-j - t) = t 2 • Displays time dependency Non-stationarity (cont.) • t is a stochastic trend • The series drifts upwards or downwards depending on sign of ; increases if positive • Stationary series tend to return to its mean value and fluctuate around it within a more-or-less constant range • Non-stationary series has a different mean at different points in time and its variance increases with the sample size Non-stationarity (cont.) • • • • • • Mean and variance increase with time yt = n + n yt-n +n-1jut-j If = then shocks never die out If | |<1 as n, then y is like a finite MA What do non-stationary series look like? Could show made-up series (with and without drift) Difference vs trend stationarity • Compare previous equation with yt = a + b t + ut E(y) = a + b t var(y) = 2 • b t - deterministic trend • But stationary around a trend E(y - b t) = a Difference vs trend stationarity (2) • Compare two generated series • Stationary around trend • Difference stationary are non-constant around a trend • But can be difficult to tell apart • Also difficult to tell series with AR coefficients 1 and 0.95 Difference vs trend stationary 80 500 60 400 40 300 20 200 0 100 -20 0 00 10 20 30 40 50 X 60 70 Z 80 90 00 Difference vs trend stationarity • Can you tell the difference? xt = 1 + xt-1 + 0.6 ut zt = 1 + 0.15 t + 0.8 et • Can you tell the difference with a near-unit root? Unit root vs near-unit root 50 500 40 400 300 30 200 20 100 10 0 00 0 00 01 02 03 04 05 06 07 08 09 10 20 30 40 50 X X W 60 10 W 70 80 90 00 Testing for unit roots • Dickey-Fuller test • Write yt = yt-1 + et as yt - yt-1 = (-1)yt-1 + et Null: Coefficient on lagged value 0, vs < 0 Dickey-Fuller tests • • • • Test akin to t-test but distributions not standard Depends if series contains constant and/or trends Must incorporate this into DF test Augmented DF test - use lags of dependent variable to remove serial correlation • All of these must be checked against relevant DF statistic • But introducing extra variables reduces power Unit versus near-unit roots • Thus difficult to tell the difference between two series over small samples • Low power of ADF tests (sample of 400) x: ADF statistic -0.77048 p-value 0.8258 w: ADF statistic -6.90130 p-value 0.0000 • Small sample (40 observations) x: ADF statistic 0.39323 p-value 0.9804 w: ADF statistic -0.49216 p-value 0.8828 Stationarity in non-stationary time series • A variable is integrated of order d - I(d) - if it musto be differenced d times for stationarity • The required number of differences depends on the number of unit roots a series has • For example, an I(1) variable needs to be differenced once to achieve stationarity: it has only one unit root Spurious regressions • Trends in data can lead to spurious correlation between variables: there appears to be meaningful relationships • What is present are uncorrelated trends • Time trend in a trend-stationary variable can be removed by regressing variable on time • Regression model then operates with stationary series with constant means and variances (standard t and F test inferences) Spurious regressions • Regressing a non-stationary variable on a time trend generally does not yield a stationary variable (it must be differenced) i.e. taking trend away does not lead to stationarity • Using standard regression techniques with non-stationary data can lead to the problem of spurious regression involving invalid inference based on usual t and F tests Spurious regressions • Consider the following DGP: yt = yt-1 + ut u , 1 xt = xt-1 + et e , 1 • y and x are uncorrelated, but estimating y t = a + b xt + v t we find that we can reject b = 0. • Why? Non-stationary data => v nonstationary gives problems with t and F stats • Also find high R2 and low DW (G&N 1974) Spurious Regressions Dependent Variable: Y Method: Least Squares Date: 03/31/03 Time: 18:28 Sample: 1900:1 2003:4 Included observations: 416 Variable X R-squared Adjusted R-squared S.E. of regression Sum squared resid Log likelihood Coefficient Std. Error t-Statistic Prob. 0.964478 0.001112 867.6800 0.0000 0.997879 0.997879 5.543177 12751.63 -1302.206 Mean dependent var S.D. dependent var Akaike info criterion Schwarz criterion Durbin-Watson stat 202.9399 120.3730 6.265414 6.275103 0.023766 Spurious regression • Why do we find significant coefficients? • What will happen if we estimate a spurious regression with the variables in first differences? • What ‘economic problem’ do we encounter if we only use differenced variables in economics? • We lose information about the long-run Spurious Regression Dependent Variable: DY Method: Least Squares Date: 03/31/03 Time: 18:36 Sample(adjusted): 1900:2 2003:4 Included observations: 415 after adjusting endpoints Variable C DX R-squared Adjusted R-squared S.E. of regression Sum squared resid Log likelihood Durbin-Watson stat Coefficient Std. Error t-Statistic Prob. 0.989704 -0.005194 0.016085 0.012185 61.52980 -0.426235 0.0000 0.6702 0.000440 -0.001981 0.211922 18.54827 56.03014 1.752192 Mean dependent var S.D. dependent var Akaike info criterion Schwarz criterion F-statistic Prob(F-statistic) 0.984475 0.211713 -0.260386 -0.240973 0.181676 0.670159 Cointegration (definition) • In general, regressing two I(d) variables, d>0, leads to the problem of spurious regression • Assume two I(d) variables and estimate: yt xt t • If is a vector such that t is I(d-b) then we say that y and x are co-integrated of order CI(d,b) What is cointegration? • If two (or more) series have an equilibrium relationship in the long run even though the series contain stochastic trends they move together such that a (linear) combination of them is stationary • Cointegration resembles a long-run equilibrium and differences from the relationship are akin to disequilibrium • Trivially, a stationary model must be Modelling the short-run • Are we ever in the long run? • How do we model the short run? • Problem of using only differenced data and the loss of long-run information • Assume yt xt t • In steady state yt xt 0 has little meaning for the long run Modelling short run • Assume yt = xt + yt-1 + xt-1 + t, , 2 • If a LR relationship exists yt = + xt • We can write yt = xt - (1- )(yt-1 - - xt-1 ) + t • (1- ) is speed of adjustment • Implications for the sign of ECM Modelling the short-run • There are some issues about the estimation of • Stock (1987) shows that OLS is fine, is super-consistent; the estimator converges to its true value at a faster rate when a series is I(1) than when it is I(0) • However, there is significant of bias in small samples Testing strategies • Perron’s suggestion: – start with regression with constant and trend – proceed trying to reduce unnecessary paramaters – if we fail to reject parameters continue testing until we are able to reject the hypothesis of a unit root • In the end we should use common sense and economics – If there should not be a unit root - probably a break Cointegration and single equations • When looking at single equations it is easy to test for cointegration – Engle and Granger two-step procedure – Engle-Granger-Yoo three-step approach • What if there is more than a single cointerating relationship? – Need a system approach – VECMs Modelling strategies • Understand the data – Do whatever tests necessary to be sure of using appropriate models • Understand the limitations of individual methods – By not taking limitations into account a rejection does not necessarily imply that the hypothesis is false • Use appropriate methods for different problems EXOGENEITY • Banerjee, A, D.F. Hendry and G.E. Mizon (1996) “The econometric analysis of economic policy”, Oxford Bulletin of Economics and Statistics 58(4), 573-600 • Ericsson, N.R. and J.S. Irons (eds) (1994) Testing Exogeneity. Advanced Texts in Econometrics. Oxford University Press. • Lindé, J. (2001) “Testing for the Lucas Critique: A quantitative investigation”, American Economic Review 91(4), 986-1005. • Monfort, A and R. Rabemananjara (1990) “From a VAR model to a structural model, with an application to the wageprice spiral”, Journal of Applied Econometrics 5, 203-227 • Urbain, J.P. (1995) “Partial versus full system modelling of cointegrated systems: An empirical illustration”, Journal of Econometrics 69(1), 177-210. • Boswijk, P. and J.P. Urbain (1997) “Lagrange Multiplier tests for weak exogeneity: A synthesis”, Econometric Reviews 16(1), 21-38. • Charezma, W.W and D.F. Deadman, (1997) New Directions in Econometric Practice, Edward Elgar, Second Edition. • Urbain, J.P. (1992) “On weak exogeneity in error correction models”, Oxford Bulletin of Economics and Statistics 54(2), 187-207. MODELLING AND FORECASTING SHORT-TERM DATA • Jondeau, É., H. Le Bihan and F. Sédillot (1999) Modelling and Forecasting the French Consumer Price Index Components, Banque de France Working paper 68. • Clements, M. P. and D.F. Hendry (1999) Forecasting non-stationary economic time series. MIT Press. • Bardsen, G and P.G. Fisher (1996) On the roles of economic theory and equilibria in estimating dynamic econometric models-with an application to wages and prices in the United Kingdom, Essays in Honour of Ragnar Frisch. VARS • Levtchenkova, S., A.R. Pagan and J.C. Robertson (1998) “Shocking stories”, Journal of Economic Surveys 12(5), 507-532.