Title: Inference with Time Series Data: An Introduction Chapters 10, 11, 12
1Inference with Time Series Data An
Introduction(Chapters 10, 11, 12)
- Course Applied Econometrics
- Lecturer Zhigang Li
2Time-Series Modelling
- Time Series Process (or Stochastic Process)
- A sequence of random variables indexed by time.
- Why time-series modelling?
- Data availability
- Utilize past information to approximate
unobserved variables. - Natural process
- Population growth
- Capital accumulation
3Key Time Series Properties
- Covariance Stationary Process X
- E(xt), Var(xt), Cov(xt,xth) are all constant
over time - If any of the conditions are violated, the
process is nonstationary. - A process with time trend is nonstationary (its
mean changes over time). - Weakly Dependent Process X
- A covariance stationary process X is weakly
dependent if Cov(xt,xth) converges to zero fast
enough as h gets large. - Otherwise, the process is strongly dependent
(or highly persistent). - Random Walk
- ytyt-1ut (ut is i.i.d. with mean zero)
4Why is Inference with Time Series Data Typically
More Difficult?
- Serial Correlation in Errors
- More restrictive assumptions to achieve
consistent estimates - Arbitrary Model Structure
- Small sample
5Consistency of Small Sample Estimates
- With Strict Exogeneity
- If ut is uncorrelated with xs for all s and t,
coefficient estimates are generally unbiased
(linear model, no perfect collinearity). - This condition rules out feedback from y on
future values of x (e.g. crime rate and number of
policemen). - This condition also rules out models with lagged
dependent variables. - Random sampling assumption not needed.
- ut can be heteroskedastic and correlated over
time.
6Consistency of Large Sample Estimates
- Key Assumptions
- Xs and u are contemporaneously uncorrelated.
- All variables (Xs and y) are stationary and
weakly dependent. - Loosely speaking, xt and xth are almost
independent as h increases without bound. - A special case A covariance stationary time
series is weakly dependent if Corr(xt,xth) goes
to zero sufficiently quickly as h approaches
infinity. (pp. 362 or 382)
7Typical Time Series Models
- Moving Average Process of Order One MA(1)
- ytetaet-1 (et is i.i.d. with zero mean)
- MA(1) process is a stationary and weakly
dependent process. (Corr(yt,yth)?) - Autoregressive Process of Order One AR(1)
- yta?yt-1ut (ut is i.i.d.)
- If ?lt1, it is called a stable AR(1) process,
which is stationary and weakly dependent, with
Corr(yt,yth)?h - ARMA(1,1) is a combination of the above two
- A trending series, while certainly nonstationary,
can be weakly dependent (called trend-stationary
process).
8Inconsistency of Standard Errors with Serial
Correlation in Ut
- OLS standard errors are not valid with serially
correlated errors. (pp. 392 or 413)
9Testing for Serial Correlation
- Durbin-Watson test
- Test for AR(1) serial correlation based on the
OLS residuals - DW is approximately 2(1-?)
- When DW is significantly less than two, serial
correlation is present.
10Solution 1 Differencing (pp. 409 or 431)
- Differencing a time series model gives us
?yta?xt?ut - If the process u follows a random walk, then
differencing makes ?ut a serially uncorrelated
process. - Even if u does not follows a random walk, if ? is
positive and large, first differencing will
eliminate part of the serial correlation. (You
may check it by testing the serial correlation in
residuals.)
11Solution 2 Use Estimator that is Robust to
Serial Correlation I
- Davidson-MacKinnon Approach
- Estimate yta0a1xt1 a2xt2ut (we are interested
in estimates of a1.) - Regress xt1 on other independent variables and
get residuals rt. - Asymptotic standard error of a1 estimate is
SD(Sru)/SE(r2). - Assumption Once the terms are farther apart than
a few periods, the correlation is essentially
zero (like weak dependence). - Robust to both heteroskedasticity and serial
Correlation.
12Solution 2 Use Estimator that is Robust to
Serial Correlation II
- Newey-West-Wooldridge Approach
- The serial correlation-robust standard error of
a1 estimate is SE(a1)/s2(v)1/2 - Where v is a given function of an integer g,
which controls how much serial correlation we are
allowing in computing the standard error. - VSa22S1g1-h/(g1)(Satat-h)
- The larger the g, the more terms are included to
correct for serial correlation. (The choice of g
can be done by statistics software but there is
no strict rule.) - Also robust to arbitrary heteroskedasticity.
- For correction under the strict exogeneity
assumption (which is very restrictive), read
textbook pp. 402 or 424.
13Solution 2 Use Estimator that is Robust to
Serial Correlation III
- Why SC-robust standard errors has lagged behind
the use of heteroskedasticity-robust standard
errors? - Large cross sections are more common than large
time series. The SC-robust standard errors can
behave poorly even for time periods as large as
100. - The choice of g is arbitrary.
- In the presence of severe serial correlation, OLS
can be very inefficient. After correcting the
standard errors for serial correlation, the
coefficients are often insignificant. In this
case, the correction approach under strict
exogeneity assumption may have advantage because
it is more efficient.
14Solution 3 Re-specify Model until the Serial
Correlation Disappears
- A model is called dynamically complete if its
errors are serially uncorrelated. Therefore, with
serially correlated errors, we may re-specify
model by including lags such that it becomes
dynamically complete. - A model should be dynamically complete if our
purpose is to fit the data and for forecasting. - Example
- yta0a1yt-1ut and ut?ut-1et (?lt1)
- The model can be rewritten as
- yta0a1yt-1 a2yt-2et
15Testing for Serial Correlation (p. 395 or 416)
- Ideas (1) Estimate time series models and obtain
residuals. (2) Then estimate autoregressive
models with the residuals and test whether
coefficients of the autoregressive parts are
statistically significant. - The tests can usually be done with ready-to-use
econometrics packages. Read the textbook for more
detail. - T-test and Durbin-Watson under strict exogeneity
- Durbin-Watson test robust to the lack of strict
exogeneity - Regress y on Xs to obtain residuals u
- Regress ut on Xs and ut-1 to test whether the
coefficient of ut-1 is significant. - Note that the test is invalid for a unit root
process.
16Unit Root Process
- Unit Root Process
- ytyt-1ut (ut is a weakly dependent process)
- A random walk (ut is i.i.d. with mean zero) is a
special case of the unit root process. - No matter how far in the future we look and how
much information we have for the past, our best
prediction of future is todays value. - The expected value of a random walk does not
depend on t - The variance of a random walk increases as a
linear function of time (nonstationary). - High persistency Corr(yt, yth)t/(th)1/2
17Trends
- Fixed Time Trend
- yta0 a1tut (linear trend)
- Ln(yt)a0 a1tut (exponential trend)
- Stochastic Time Trend
- A random walk with Drift ytayt-1ut
- The best prediction of yth at time t is yt plus
the drift ah. - Ignoring the fact that two sequences y and x are
trending (possibly due to changes in other
variables) can lead to spurious relationship
between x and y. (actually an omitted variable
problem) - Solution Include t in the regression of y on x
for fixed time trend take first-difference of
equations for stochastic time trend
18Seasonal Adjustment
- Method 1 Include a set of seasonal dummy
variables to account for seasonality in the
dependent and independent variables. (p.353-4) - Method 2 Using autoregressive models.
19Infinite Distributed Lag Models
- ytaa0xta1xt-1ut
- Geometric Distributed Lag
- aj??j, ?lt1
- Therefore, we can rewrite the model as
- yta?xt?yt-1ut-?ut-1
- The errors are thus correlated with yt-1, causing
endogeneity problem. A natural IV is xt-1. - Rational Distributed Lag
- Augment the GDL model by adding a lag of z to
independent variables.