Models with Trend

Lecture 3
 Lecture Objectives

❖Material from Enders chapter 4.

❖Understand Trends
❖Unit root testing
 Deterministic and Stochastic Trends
❖Many series have a trend which may be deterministic or stochastic

❖𝑦𝑡 = trend + stationary component + noise

❖Deterministic trends are temporarily affected by an innovation (shock or
error)
❖Stochastic trends are permanently affected by an innovation
➢ RBC school argues change in technology permanently affects the level of GDP
 Trends
❖The key feature of a trend is that it has a permanent effect on a series
❖Suppose that a series changes by a fixed amount
❖∆𝑦𝑡 =𝑎0
❖Where the solution to this linear difference equation is;
❖𝑦𝑡 =𝑦0 +𝑎0 t
➢Changing by a fixed amt implies this is a deterministic linear time trend
 Trends
❖Adding a stationary error process A(L)𝜀𝑡 we obtain
❖𝑦𝑡 =𝑦0 + 𝑎0 𝑡 + 𝐴(𝐿)𝜀𝑡
❖𝑦𝑡 can differ from its trend value by the amount 𝐴(𝐿)𝜀𝑡
❖Since this deviation is stationary, the 𝑦𝑡 sequence will exhibit only temporary
departures from the trend.
❖This would be termed a trend stationary process
 Trends
❖Suppose the expected change in 𝑦𝑡 is 𝑎𝑜 units
❖Let ∆𝑦𝑡 =𝑎0 +𝜀𝑡
❖This has a profound impact on the trend where the solution is now
❖𝑦𝑡 =𝑦0 +𝑎0 t +σ𝑡𝑖=1 𝜀𝑡
➢The effect of each shock is now permanent (albeit random)
➢Such a sequence is termed a stochastic trend model.
 Trends
❖In stationary time series:
➢Shocks are temporary.
➢Revert back to long-run mean.
❖In non-stationary time series:
➢Includes permanent shock
➢Mean and/or variance are time dependent.
 Trends
❖To identify a non-stationary time series:
➢No long-run mean to which the series returns.
➢Variance is time dependent (increases with time)
 Random Walk model
❖∆𝑦𝑡 =𝑎0 + 𝜀𝑡
➢ Is the basic building block for modelling series containing stochastic trends.
❖When 𝑎0 =0, we have ∆𝑦𝑡 = 𝜀𝑡
➢ This is a random walk model
➢ Some formulations of the efficient market hypothesis posit that the change in the
price of a stock from one day to the next is completely random.
➢ As such, the current price (𝑦𝑡 ) should be equal to last period’s price plus a white-
noise term
 Random Walk model
❖So that
❖𝑦𝑡 =𝑦𝑡−1 +𝜀𝑡 or (∆𝑦𝑡 = 𝜀𝑡 )
❖The general solution for such a process is:
❖𝑦𝑡 =𝑦0 +σ𝑡𝑖=1 𝜀𝑖
❖Taking expected values, E[𝑦𝑡 ]=E[𝑦𝑡−𝑠 ]= 𝑦0
➢The mean of a random walk is a constant
 Random Walk model
❖However, the variance is time dependent
❖Var(𝑦𝑡 )=var(𝜀𝑡 + 𝜀𝑡−1 +…+ 𝜀1 )=t𝜎 2
❖Var(𝑦𝑡−𝑠 )=var(𝜀𝑡−𝑠 + 𝜀𝑡−𝑠−1 +…+ 𝜀1 )=(t-s) 𝜎 2
❖Since the variance is not constant, the process is non-stationary
 Random Walk plus Drift model
❖The random walk plus drift model augments the random walk model by
adding a constant term 𝑎0 , so that
❖𝑦𝑡 =𝑦𝑡−1 +𝑎0 +𝜀𝑡
❖The general solution for 𝑦𝑡 is given by
❖𝑦𝑡 =𝑦0 +𝑎0 t+σ𝑡𝑖=1 𝜀𝑖
➢This model includes two non-stationary components
➢Linear deterministic trend and stochastic trend
 Generalizations
❖A random walk plus noise model has a stochastic trend and a white-noise
component
❖𝑦𝑡 =𝑦0 +σ𝑡𝑖=1 𝜀𝑖 +𝑛𝑡
❖Where 𝑛𝑡 is a white-noise process
 Removing the trend
❖Shocks to a stationary time series are necessarily temporary.
❖A series containing a trend will not revert to a long-run level.
➢The trend can have deterministic and stochastic components.
❖The usual methods for eliminating the trend are:
➢Differencing
➢Detrending
 Removing the trend
❖Consider the solution for the random walk plus drift model
❖𝑦𝑡 =𝑦0 +𝑎0 t + σ𝑡𝑖=1 𝜀𝑖
❖Taking the first difference, we obtain ∆𝑦𝑡 = 𝑎0 + 𝜀𝑡
❖Its stationary.
 Removing the trend
❖The variable of interest ∆𝑦𝑡 has a constant mean and constant variance.
❖The first difference of a unit root process is stationary.
❖If a variable has two unit roots, the second difference of that variable is stationary.
❖The general point is that the dth difference of a process with d unit roots is
stationary.
 Removing the trend
❖First differencing a model with deterministic trend and a pure noise
component sometimes does not make it stationary.
❖Detrending then works.
 Unit Root Testing
❖Consider the regression equation: 𝑦𝑡 = 𝑎0 + 𝑎1 𝑧𝑡 + 𝑒𝑡
❖The assumptions of the classical regression model necessitate that both the
𝑦𝑡 and 𝑧𝑡 sequences be stationary and that the errors have a zero mean and
a finite variance.
❖In the presence of nonstationary variables, there might be a spurious
regression (Granger and Newbold, 1974)
 Unit Root Testing
❖A spurious regression has a high R^2 and t-statistics that appear to be
significant, but the results are without any economic meaning.
➢Estimated residuals exhibit a high degree of autocorrelation.
➢T-test, F-test or R^2 values are unreliable.
❖The econometrician has to be very careful in working with nonstationary
variables.
❖In terms of equation 𝑦𝑡 = 𝑎0 + 𝑎1 𝑧𝑡 + 𝑒𝑡 , there are four cases to consider.
 Unit Root Testing
CASE 1
❖Both 𝑦𝑡 and 𝑧𝑡 are stationary.

❖When both variables are stationary, the classical regression model is

appropriate.
 Unit Root Testing
CASE 2
❖The 𝑦𝑡 and 𝑧𝑡 sequences are integrated of different orders.
❖Regression equations using such variables are meaningless.
❖Because the error term sequence still contains a trend component.
 Unit Root Testing
CASE 3
❖The nonstationary 𝑦𝑡 and 𝑧𝑡 sequences are integrated of the same order
and the residual sequence contains a stochastic trend.
❖This is the case in which the regression is spurious.
❖The results from such spurious regressions are meaningless in that all errors
are permanent
 Unit Root Testing
CASE 4
❖The nonstationary 𝑦𝑡 and 𝑧𝑡 sequences are integrated of the same order,
and the residual sequence is stationary.
❖In this circumstance, 𝑦𝑡 and 𝑧𝑡 are cointegrated.
 Unit Root Testing
❖Pretesting the variables in a regression for nonstationarity is extremely
important.
❖Estimating a regression in the form of equation, 𝑦𝑡 = 𝑎0 + 𝑎1 𝑧𝑡 + 𝑒𝑡 is
meaningless if case 2 or 3 applies.
 Formal tests
Dickey-Fuller Tests
❖A critical first step in any econometric analysis is to visually inspect the data.
❖Dickey and Fuller (1979) consider three different types of regressions that
can be used to test for the presence of a unit root
❖∆𝑦𝑡 = 𝛾𝑦𝑡−1 + 𝜀𝑡 (1)
❖∆𝑦𝑡 = 𝑎0 + 𝛾𝑦𝑡−1 + 𝜀𝑡 (2)
❖∆𝑦𝑡 = 𝑎0 + 𝛾𝑦𝑡−1 + 𝑎2 𝑡 + 𝜀𝑡 (3)
 Formal tests
❖The parameter of interest in all equations is 𝛾
❖If 𝛾 =0 then 𝑦𝑡 contains a unit root.
❖Estimate these equations using OLS and compare the resulting t-statistic
with appropriate critical values.
 Formal tests
❖Equations (1), (2) and (3) can be replaced by the autoregressive processes to
be:
❖∆𝑦𝑡 = 𝛾𝑦𝑡−1 + σ𝑃𝑖=2 𝛽𝑖 ∆𝑦𝑡−𝑖+1 + 𝜀𝑡 (4)
❖∆𝑦𝑡 = 𝑎0 + 𝛾𝑦𝑡−1 + σ𝑃𝑖=2 𝛽𝑖 ∆𝑦𝑡−𝑖+1 + 𝜀𝑡 (5)
❖∆𝑦𝑡 = 𝑎0 + 𝛾𝑦𝑡−1 + 𝑎2 𝑡 + σ𝑃𝑖=2 𝛽𝑖 ∆𝑦𝑡−𝑖+1 + 𝜀𝑡 (6)
 Formal tests
❖Not all time series variables can be well represented by the first-order
autoregressive process:
❖∆𝑦𝑡 = 𝑎0 + 𝛾𝑦𝑡−1 + 𝑎2 𝑡 + 𝜀𝑡
❖It is possible to use the Dickey-Fuller tests in higher order equations such as
4-6.
❖We then have augmented Dickey-Fuller (ADF) tests.
 Selection of the Lag Length
❖It is important to use the correct number of lags in conducting a Dickey-
Fuller test.
❖Too few lags implies residuals don’t behave like white-noise process.
❖Too many lags implies it reduces the power of the test to reject the null of a
unit root since the increased number of lags necessitates the estimation of
additional parameters and a loss of degree of freedom
 Selection of the Lag Length
❖How does a researcher select the appropriate lag length:
(i)General to specific approach
➢ Start with many lags and pare down the model by the usual t-test and/or F-tests.
(ii)plot the residual
➢ There should not appear to be any strong evidence of structural change or serial
correlation.
(iii)Use the information criterion such as AIC or the BIC(remember it tends to be
parsimonious in practice)
Null Hypothesis: LRGDP has a unit root
Exogenous: Constant
Lag Length: 0 (Automatic - based on SIC, maxlag=6)

t-Statistic Prob.*

Augmented Dickey-Fuller test statistic -2.400823 0.1506

Test critical values: 1% level -3.689194
5% level -2.971853
10% level -2.625121

*MacKinnon (1996) one-sided p-values.

Null Hypothesis: LRGDP has a unit root
Exogenous: Constant, Linear Trend
Lag Length: 0 (Automatic - based on SIC, maxlag=6)

t-Statistic Prob.*

Augmented Dickey-Fuller test statistic -0.325836 0.9856

Test critical values: 1% level -4.323979
5% level -3.580622
10% level -3.225334

*MacKinnon (1996) one-sided p-values.

Null Hypothesis: D(LRGDP) has a unit root
Exogenous: None
Lag Length: 0 (Automatic - based on SIC, maxlag=6)

t-Statistic Prob.*

Augmented Dickey-Fuller test statistic -2.873205 0.0058

Test critical values: 1% level -2.653401
5% level -1.953858
10% level -1.609571

*MacKinnon (1996) one-sided p-values.

Null Hypothesis: LRGDP has a unit root
Exogenous: Constant
Bandwidth: 0 (Newey-West automatic) using Bartlett kernel

Adj. t-Stat Prob.*

Phillips-Perron test statistic -2.400823 0.1506

Test critical values: 1% level -3.689194
5% level -2.971853
10% level -2.625121

*MacKinnon (1996) one-sided p-values.

Null Hypothesis: LRGDP has a unit root
Exogenous: Constant, Linear Trend
Bandwidth: 1 (Newey-West automatic) using Bartlett kernel

Adj. t-Stat Prob.*

Phillips-Perron test statistic -0.355811 0.9844

Test critical values: 1% level -4.323979
5% level -3.580622
10% level -3.225334

*MacKinnon (1996) one-sided p-values.

Null Hypothesis: D(LRGDP) has a unit root
Exogenous: None
Bandwidth: 0 (Newey-West automatic) using Bartlett kernel

Adj. t-Stat Prob.*

Phillips-Perron test statistic -2.873205 0.0058

Test critical values: 1% level -2.653401
5% level -1.953858
10% level -1.609571

*MacKinnon (1996) one-sided p-values.