Title: Applying TimeSeries Analysis in Sport Management Research: Avoiding the Pitfalls
1Applying Time-Series Analysis in Sport Management
Research Avoiding the Pitfalls
- To insert your company logo on this slide
- From the Insert Menu
- Select Picture
- Locate your logo file
- Click OK
- To resize the logo
- Click anywhere inside the logo. The boxes that
appear outside the logo are known as resize
handles. - Use these to resize the object.
- If you hold down the shift key before using the
resize handles, you will maintain the proportions
of the object you wish to resize.
Research Methodology Seminar
2Focus of seminar
- Time-series analysis the statistical analysis of
a sample of time-ordered, periodic observations - Example annual performance data for a sporting
organisation - Focus of seminar how to avoid the various
potential pitfalls in the statistical analysis of
time-series data that can invalidate the
empirical results
3Potential pitfalls in time-series analysis
- Dynamic dependence current outcome depends on
past events (including past outcomes) - Structural instability the non-local causal
process (i.e. DGP) changes over time - Residual autocorrelation estimated residuals are
correlated over time - Non-stationarity the statistical properties of
the data (i.e. mean, variance and
autocovariances) are time-dependent such that
each observation has different distributional
characteristics
4Plan of workshop
- Introduction
- Dynamic dependence
- Structural instability
- The autocorrelation problem
- Non-stationarity and cointegration
- Application Man Utd average league gate
attendances, 1946/47 2002/03 - Practical laboratory session
5Dynamic Dependence in Time-Series Models
6Dynamic dependence
- Dynamic model
- Current outcome f(past outcomes, current and
past events) - Formal representation (single period lag, one
independent variable) - Yt a0 a1Yt-1 ß1Xt ß2Xt-1 et
- Autoregressive distributed lag (ADL) model
- Mathematics stochastic linear difference equation
7A simple dynamic model
- Yt a0 a1Yt-1 et
- 1st order difference equation
- Mathematically this can be solved to yield an
expression for Yt in terms of the parameters,
current and past disturbance terms and the
initial value of Y - Three components to solution
- long-run equilibrium position (Yt Yt-1, et
0,) - history of short-run disturbance effects
- current effect of initial deviation from long-run
equilibrium
8Dynamic dependence implies history matters
- Yt a0 a1Yt-1 et
- Solution
- Yt a0/(1- a1) ?(a1)iet-i (a1)tY0 - a0/(1
- a1) - a0/(1- a1) long-run equilibrium position
- ?(a1)iet-i history of short-run disturbance
effects - (a1)tY0 - a0/(1 - a1) current effect of
initial deviation from long-run equilibrium - a1 is known as the characteristic root
- Magnitude of a1 is crucial to determining the
behaviour of Yt over time
9The nature of the adjustment path
- Yt a0 a1Yt-1 et
- Yt a0/(1- a1) ?(a1)iet-i (a1)tY0 - a0/(1
- a1) - a1 lt 1 gt stable, equilibrium-convergent
adjustment path - a1 gt 1 gt explosive, non-convergent path
- a1 1 gt unit-root process non-convergent
- a1 gt 0 gt direct adjustment
- a1 lt 0 gt oscillating (i.e. cyclical) adjustment
- Vital to determine whether or not you are dealing
with a unit-root process
10The random walk model
- Yt a1Yt-1 et
- Random walk model a1 1
- Yt Yt-1 et
- gt ?Yt Yt Yt-1 et
- Solution
- Yt Y0 ? et-i
- Implications
- Yt follows a random, non-predictable path
- disturbances have permanent effects
- If a1lt1, disturbances have a decaying effect
- e, a1e, (a1)2 e, ...
11Short-run and long-run impacts
- Simple multivariate ADL
- Yt a0 a1Yt-1 ß1Xt et
- Example
- Dynamic win-wage benchmark model
- W/Lt a0 a1W/Lt-1 ß1WAGES t et
- Short-run impact of WAGES ß1
- Long-run impact (W/Lt W/Lt-1) ß1 /(1 - a1)
- a1Yt-1 represents learning/momentum effects
12Structural Instability
13Structural instability
- OLS estimation assumes that the DGP is
structurally stable (i.e. parameter constancy)
over the whole sample - But time-series data may cover periods of
structural change and regime shifts - Structural stability may also be an inappropriate
assumption for ordered cross-sectional data
14Consequences of structural instability
- Parameter estimates summarise the average effects
across alternative regimes - OLS estimates are biased for any individual
regime - Benchmarks and simulations based on the OLS
estimates are biased
15Methods of detecting structural instability
- Residuals
- Chow tests
- Chow test for structural instability
- Chows second test for predictive failure
- Recursive least squares
- Within-sample forecasting
- Hansen test
16The limitations of OLS residuals
- OLS residuals are based on full sample
information - Smoothing effect that may disguise the presence
of structural instability and outliers (i.e.
observations far removed from rest of sample) - Important to use other types of residuals to
detect structural instability and outliers
17Alternative residuals to detect instability and
outliers
- Predicted residuals estimated coefficient on
impulse dummy (i.e. predicted values obtained
from estimated regression with observation
omitted) - Studentised residuals t-ratios of impulse dummy
- Time-series residuals
- Estimated innovations determined using parameter
estimates that utilise only past sample
information - One-step residuals determined using parameter
estimates that utilise only current and past
sample information
18Chow test for structural stability
- Estimate pooled (i.e. full sample) regression
model - Estimate separate regression model for each
sub-sample - Apply F test for linear restrictions
- F(J,N-K) (RRSS URSS)/J
- URSS/(N K)
- RRSS restricted RSS
- URSS unrestricted RSS
- J no. of restrictions
- N sample size
- K no. of parameters in unrestricted model
19Recursive least squares
- Recursive estimation involves
- estimating the regression model for an
- initial sub-sample and then re-
- estimating the model repeatedly as the
- sub-sample is expanded to include the
- next (ordered) observation
20Dealing with structural instability
- Separate sample into structurally stable
sub-samples and estimate separate regression
models - Or
- Estimate single regression model for pooled
sample with dummy variables for intercept and
slope shifts - Both approaches are equivalent
21TheAutocorrelation Problem
22What is the autocorrelation problem?
- Residual autocorrelation OLS residuals, et,
serially correlated with lagged values, et-1,
et-2, - Common occurrence in estimated time-series models
- Autocorrelation problem multiple possible causes
of residual autocorrelation with different
solutions
23Tests for detecting residual autocorrelation
- Durbin-Watson (DW) test
- Berenblut-Webb test
- Durbins h test
- Durbins alternative test
- Breusch-Godfrey test
24The DW test for autocorrelation
- Traditional test for residual autocorrelation
- Test statistic
- d ?(et-et-1)2/?et2 ? 2(1 - ?)
- 0 lt d lt 4
- 0 lt d lt 2 ltgt positive residual autocorrelation
- 2 lt d lt 4 ltgt negative residual autocorrelation
- d 2 ltgt zero residual autocorrelation
- Appropriate only if 1st-order residual
autocorrelation and no lagged dependent variables
25The Breusch-Godfrey test for higher-order serial
correlation
- Regress residuals on explanatory variables and
lagged residuals - et f(X1t,..,Xkt, et-1,,et-j)
- Test the joint hypothesis that all the estimated
coefficients on the lagged residuals are zero - Test statistic, LM jF ?2(j)
- where F ? F statistic for joint significance of
lagged residuals - Alternatively test the overall significance of
the auxiliary regression using LM nR2 ?2(kj)
26One possible cause of residual autocorrelation
- One possible cause AR errors, et f(et-1, et-2,
) - Example AR(1) error process, et ?et-1 ut
- where
- autocorrelation parameter, ? ? corr(et, et-1)
- pure stochastic disturbance, ut IID(0, s2)
- AR error processes violate OLS assumptions
- Consequences
- OLS estimates unbiased but inefficient
- variance estimate biased rendering interval
estimation and statistical tests invalid - Solution error misspecification requires either
reparameterisation to produce IID error process
for OLS estimation (e.g. Cochrane-Orcutt
quasi-differencing procedure) or use of
alternative estimation methods e.g. RALS - Conventional approach to residual autocorrelation
27But other possible causes of residual
autocorrelation
- Residuals may be autoregressive for several
reasons - AR disturbances (i.e. error misspecification)
- Model misspecification (e.g. excluded variables)
- Misspecified dynamics (e.g. static model
estimated when dynamic dependence) - Structural instability
- ARCH effects (i.e. error variance is
autoregressive) - Solution re-specify structural model if DGP
misspecified - AR disturbances are a special restricted case of
misspecified dynamics
28The greatest non sequitur in econometrics?
- Perhaps the greatest non sequitur in the
history of econometrics is the assumption that
autocorrelated residuals entail autoregressive
errors, as is entailed in correcting serial
correlation using Cochrane-Orcutt. - (Doornik and Hendry)
29Test, test and test again
- Dont rely only on the DW test
- Use other tests for residual autocorrelation e.g.
Breusch-Godfrey test - Test for general model misspecification e.g.
Ramsey RESET test - Test for misspecified dynamics
- Test for structural instability and ARCH effects
- Test for validity of imposing restriction of AR
errors - Investigate robustness of parameter estimates to
alternative model strategies
30Non-Stationary Time Series and Cointegration
31Stationary, non-stationary and integrated
processes
- Stationary process distributional properties are
time-independent - Non-stationary process distributional properties
are time-dependent - Integrated process non-stationary process that
can be transformed into a stationary process by
differencing - Yt is integrated of order d, I(d), if it requires
to be differenced d times to become stationary - Conventional statistical tests are invalid for
non-stationary processes
32ARIMA(p,d,q) models
- General representation of I(d) process
- ARMA(p,q) model of dth difference
- Autoregressive AR(p) process
- Yt a0 ?paiYt-i et
- Moving-average MA(q) error process
- Yt a0 ?qdiet-i
- ARIMA(p,1,q) model
- ?Yt ?0 ?p?i?Yt-i ?t ?q dj?t-j
- Starting point for Box-Jenkins methodology for
modelling univariate time series
33Dickey-Fuller (DF) tests for unit roots
- Unit-root processes are non-stationary.
- Crucial to test for unit roots
- AR(1) process
- Yt ?1Yt-1 ?t
- Subtracting Yt-1 from both sides yields
- ?Yt ?Yt-1 ?t
- where ? ?1 - 1
- Unit root gt ?1 1
- Dickey-Fuller test for unit root ? 0
34Alternative forms of the DF test
- Pure random walk
- ?Yt ?Yt-1 ?t
- Random walk with drift
- ?Yt ?0 ?Yt-1 ?t
- Random walk with drift and time trend
- ?Yt ?0 ?Yt-1 ?2t ?t
35The DF test statistics
- t test of null hypothesis, ? 0
- DF test statistic does not have standard t
distribution - Critical values depend on model specification,
sample size and significance level - F test of null hypothesis, ? 0, jointly with ?0
0 and/or ?2 0 - F test has non-standard critical values
36Augmented Dickey-Fuller (ADF) tests
- DF test assumes that the error term is a
white-noise process - If ARMA error process, DF regressions should be
augmented with additional ?Yt-j terms - Test statistics and critical values for ADF tests
are the same as for the DF tests
37Cointegration definition and interpretation
- Suppose Yt, Xt are both I(1) processes
- Yt and Xt are cointegrated, CI(1,1), if there
exists ? such that Yt - ?Xt ?t I(0) - Cointegration implies that a long-run
relationship exists between Yt and Xt they do
not drift apart over time
38Spurious regression problem
- Regression of one random walk on another
independent random walk - Regression of integrated variables with no
cointegrating relationship - May have high goodness of fit
- Extremely low DW statistic indicative of spurious
regression problem
39Testing for cointegration
- Two principal tests for cointegration
- Engle-Granger test apply unit-root test to
residuals from cointegrating regression (but DF
critical values not valid) - Cointegrating regression Durbin-Watson (CRDW)
test no cointegrating relationship implies DW 0
40The Granger representation theorem
- If Yt and Xt are cointegrated, there is a
long-run relationship between the two time series
and the short-run dynamics can be described by
the error correction model (ECM).
41The ECM
- Long-run relationship
- Yt ?Xt ?t
- ECM (short-run dynamics)
- ?Yt ??Xt ?Yt-1 - ?Xt-1 ?t
- Two components of the current change in Yt
- (i) the short-run response to current changes in
Xt, ??Xt and - (ii) the partial correction of the previous
deviation of Yt from its optimal long-run level,
?Yt-1 - ?Xt-1
42Causality tests
- Granger causality the future cannot cause the
present or past - Xt fails to Granger-cause Yt if, in a regression
of Yt on lagged Ys and lagged Xs, the
coefficients on the lagged Xs are jointly
insignificant
43Application
44Application football gate attendances
- Theory Gate attendance f(current and past
playing performance, divisional demand) - Data Man Utd, average league gate,
1946/47-2002/03 (57 observations) - Key results
- ManUGate is stationary process (whole sample)
- Autocorrelated residuals and possible ARCH
effects - RALS ? D1AvGate and PtsGap as determinants
- Dynamic model ? possible FA Cup effect
- Evidence of structural instability, 1992/93
onwards - FA Cup effect significant pre-1992/93
- Possible unit-root process pre-1992/93