Panel time-series models

A practitioner's guide using Stata

Prof. Geoff Pugh (Feb.2015)

26/02/2025 1
 Terminology and limits
• Methods
1. Panel unit root tests
2. Cointegration tests
3. Empirical estimators
• Development
– 1st generation methods
• Assumed panel members (groups) to be
cross-sectionally independent
– 2nd generation
• Addressed correlation across panel members
26/02/2025 2
 Example
• Many examples use data from Eberhardt (2012)
• xtmg to download
– Either: Command Window: findit xtmg
– Or: Help – Search – xtmg
• Command Window, type:
– use manu_prod
– xtset nwcode year
• Testing a simple theory of labour productivity
– Constant returns Cobb-Douglas production function
• Y = AKiL1-I
• Y – value added (deflated)
• K – capital stock (deflated)
• L – labour force
• A - captures Total Factor Productivity (TFP)
– Systematic unobservable influences (technology parameter)
– Can be rewritten in per capita terms and log-linearised for estimation
• where y=Y/L and k=K/L (i.e. per worker)
• lyit = Ait + αilkit + it
– it = error term (idiosyncratic unobservable influences)
– α = elasticity of labour productivity with respect to output (to be estimated)
•
26/02/2025 Heterogeneous across countries 3
 Methods I: Panel unit root tests
• Start with plots of each data series
– xtline
• xtline ly
• Xtline lk
– Most series display a nonconstant mean
• Most series certainly not “textbook” stationary!
• Panel unit root tests
– Work best with “large” T and at least “moderate” N

26/02/2025 4
 Panel Unit Root testing (PURT) in Stata
• Help - search – multipurt
• Using existing xtfisher & pescadf commands
– If not installed
• help multipurt
– Click on xtfisher
» help xtfisher
– Click on pescadf
» help pescadf

• Also uses data handling practices from xtwest

– If not installed
• help multipurt
– Click on xtwest
» help xtwest

• Panel unit root tests for

– Multiple variables
• Maximum of 9
• Reflects suitability of panel TS modelling for small models
(i.e. limited number of variables)
– Not a limitation of “Micro panel” estimators (difference and system GMM)
– Multiple lags
– Unbalanced panels
– Panels with discontinuous time-series
• Two panel unit root tests building on Dickey-Fuller
& Augmented DF regressions
26/02/2025 5
– For models with and without a trend term
 User-written programme: Multipurt
• 1st generation test: Maddala and Wu (1999)
– Allows for heterogeneity in the autoregressive coefficient
of the Dickey-Fuller regression
– Ignores cross-section dependence in the data
• 1st generation characteristic
– Null: nonstationarity in all panel members/series
• Alternative: at least one series in the panel is stationary
– If implemented by xtfisher
• Drift indicates that the process under the null hypothesis is a random walk with nonzero drift
– May not be used with pp or trend
– pp indicates that the Phillips-Perron test is used rather than the ADF
• Trend specifies that a trend term be included in the associated regression.
– May not be used with the drift option

• 2nd generation test: Pesaran (2007) CIPS test

– Allows for heterogeneity in the autoregressive coefficient
of the Dickey-Fuller regression
– Allows for cross-section dependence in the data
• 2nd generation characteristic
– Test statistic is constructed from the results of panel-member-specific (A)DF regressions
• Averaging of the group-specific results follows Im, Pesaran and Shin (2003)
– Null: Nonstationarity (i.e. all panels contain a UR)
• Under H0 - test statistic has a non-standard distribution
– Consistent under HA that only a fraction of the series are stationary
26/02/2025 6
– Default test: includes a Constant
 Implementing PURTs
• All tests: required: lags(numlist)
– Identifies the maximum number of lagged differences to be included
in the group-specific ADF regressions
• To control for serial correlation in the errors
• Command window: multipurt ly lk, lags(2)
– ly
• Conclusion depends on with or without trend
– lk
• Preponderance of results suggest non-rejection of H0
• For robustness checks
– Try xtfischer options
• xtfisher ly, lag(2) trend display
– Display: presents the unit root tests for each individual group
» 4 non-rejections at the one per cent level
» 8 non-rejections at the five per cent level
• Group-level series overwhelmingly non-stationary
• Conclusion: Non-stationary series
– Real economy variables, hence most likely to be I(1)
– Check: multipurt D_ly D_lk , lags(1)
• One lag sufficient with differenced data
26/02/2025 7
• Uniform results: p=0.000
 Additional panel-data unit-root tests
• Stata 13: xtunitroot
– Wide range of tests
– Most require a balanced panel
• 2 work also with unbalanced panels
– Variants of the two multipurt tests
• e.g. IPS
– xtunitroot ips ly, lags(aic 2) demean
– xtunitroot ips ly, lags(aic 2) trend demean
• Demean: mitigates the impact of cross-sectional dependence
– for each time period xtunitroot computes the mean of the series across panels
and subtracts this mean from the series
• aic: number of lags of the series be chosen such that the AIC
for the regression is minimized
– Akaike Information Criteria – a measure of fit not of statistical validity
» Statistical validity require “white noise” residuals
26/02/2025
» Not guaranteed by “best fit” 8
 Methods II: Cointegration testing
• Conceptual concerns in testing for cointegration in
panels
– How much heterogeneity do we allow across
groups/countries?
– How do we combine the statistics we arrive at if we
opted for heterogeneous tests?
• Major approaches
1. Residual-based tests
• Run some regression, collect residuals
and test for stationarity
2. Error correction tests
• Construct an error correction model and investigate whether
the EC term is significant
26/02/2025 9
 1st generation tests
(Pedroni, 1999, 2000 & 2004)
• Introduces flexibility/heterogeneity in terms of
– Cointegrating vector
– Dynamics
• Output 1: Seven tests for cointegration
– Null of no cointegration
– Assume cross-section independence
• Residual tests in the Engle-Granger tradition
• Two groups of statistics:
– ‘group-mean’, allowing for heterogeity
1. Fit each panel separately
2. Average the resulting test statistics
– ‘panel’ (pooled), imposing a common parameter
1. Pool the data before fitting the model
26/02/2025 2. Compute based on the pooled regression results 10
 Estimate of the long-run cointegrating
vector (equilibrium relationship)
• Output 2: the cointegrating vector
– Implements Pedroni's group-mean panel-dynamic ordinary least-
squares (PDOLS) model
• Extends the dynamic ordinary least-squares (DOLS) technique of
estimating the cointegrating vector in a single equation to panel time-
series data
– Medium to large N, large T
• DOLS involves adding lags and leads of the regressors to eliminate
feedback effects and endogeneity.
• PDOLS
– A DOLS regression is conducted for each individual
– The results are combined for the entire panel following Pedroni's
group-mean approach.
– Variables must cointegrate for PDOLS
26/02/2025 11
 Implementation
• Help – Search – xtpedroni
• Command window:
– help xtpedroni
– use xtwestdata
• Analysing influence of log per capita GDP (loggdp)
on log per capita health expenditure (loghex)
– N = 20 (OECD countries)
– T = 32 (Annual data 1970-2001)
– xtset ctr year
• Panel defined by Country; Year
• Examples of syntax
– xtpedroni loghex loggdp, trend lagselect(aic)
– xtpedroni loghex loggdp, trend lagselect(aic) extraobs
notdum
• notdum suppresses default common time dummies
26/02/2025 – May be appropriate when averaging over the N dimension 12
may destroy the cointegrating relationship
 Interpretation
• Each of the statistics is distributed as standard normal when both the time series and the
cross-sectional dimensions of the panel grow large
• “Thus, to test the null of no cointegration, one simply computes the value … and
compares these to the appropriate tails of the normal distribution” (Pedroni, 1999,
pp.666-68)
– “For the panel variance statistic, large positive values imply that
the null of no cointegration is rejected.”
– “For each of the other six test statistics … the left tail of the normal distribution is used to reject
the null hypothesis … large negative values imply that the null of no cointegration is rejected.”
– Baltagi, 2013, p.296: “Rejection of the null hypothesis means that enough of the individual cross-
sections have statistics ‘far away’ from the means predicted by theory were they to be generated
under the null.”
• A relevant critical value:  1.96
– Critical value for a one-sided test at the 2.5 per cent significance level
– >|1,96|  Reject the null of no cointegration
• Different results according to inclusion/non-inclusion of period dummies
– Non-inclusion makes cross-group error correlation more likely
– But trend should be sufficient to pick up unmodelled growth processes
• If cointegration not rejected,
then strong evidence of a long-run equilibrium relationship
– Without period dummies, health expenditure a “luxury good”
–
26/02/2025Strong fiscal implications for growing economies 13
 2nd generation tests
• ECM approach
– Check whether an ECM does/does not have
error correction
• individual group or full panel

• a0i is the error correction/speed of adjustment term.

• Penultimate term includes lags and leads of x
– Otherwise need to assume exogeneity of x
• Estimate separately for each panel (group)
• If a0i = 0  no error-correction  y,x not cointegrated
• If a0i < 0 Error Correction (negative feedback)  cointegration.
• In total 4 tests, based on
– ‘group mean’
– ‘pooled panel’
• Large negative values reject H0 of no cointegration.
• If cross-group correlation of errors is suspected: use bootstrap to obtain robust SEs
• Long T is important
26/02/2025 14
– Often strong contrast (homogenous/heterogeneous) unless T is large
 Implementation
• Help – Search – xtwest
– Command window:
• help xtwest
• Command window
– use xtwestdata
• xtset ctr year
– xtwest loghex loggdp, lags(1 3) leads(0 3) lrwindow(3) constant
trend westerlund
• To allow for cross-group error correlation
– Bootstrap robust SEs for the test statistics
– xtwest loghex loggdp, lags(1 3) leads(0 3) lrwindow(3) constant
trend westerlund bootstrap(50)
26/02/2025 15
 Testing for cross-section dependence
• The xtcd command implements the Pesaran (2004) CD test for cross-
section independence in macro panel data.
– Not a post-estimation command
– Can apply xtcd to test ‘raw’ variables pre-estimation
• Employs the correlation-coefficients
between the time-series for each panel member
• Allows for
– Analysis of cross-section dependence in any variable and residual series
– Computation of the averaged correlation and absolute correlation coefficients for
up to nine series at a time
– Both balanced and unbalanced panels
• But not discontinuous data (i.e. time series with “gaps”)
• Null hypothesis: cross-section independence
• Robust to
– Nonstationarity (the spuriousness would show up in the averaging)
– Parameter heterogeneity or structural breaks
• Performs well even in small samples
26/02/2025 16
 xtcd: Implementation
• Help – Search – xtcd
– Command window: help xtcd
• Command window
– xtcd loghex loggdp
• Compare
– In this case, insufficient common observations
– Panel too small
– Panel too unbalanced
• What now?
– Assume that cross-section dependence exists
• Implication
– Either: Use estimators that address the problem
– Or: estimate with period dummies to attenuate the problem
26/02/2025 17
 Methods III: Estimators
1. Bias-Corrected Least Squares Dummy Variable (LSDV)
– Stata user-written programme: xtlsdvc
• Bruno (2005) SJ 5(4) 437-500
2. Panel-dynamic Ordinary Least-squares (PDOLS)
– Stata user-written programme: xtpedroni
• Neal (2014) SJ 14(3)
3. Mean Group (MG)
4. Pooled Mean Group (PMG)
– Stata user-written programme: xtpmg
• Blackburne and Frank (2007) SJ 7(2) 197-208
5. Common Correlated Effects Mean Group (CCEMG)
6. Augmented Mean Group (AMG)
– Stata user-written programme: xtmg
• Eberhardt (2012) 12(1) 61-71
26/02/2025 18
 Considerations for the choice of
estimator
• A challenge!
– Assumptions under which estimators deliver an optimal solution almost
never realised in practice
– Estimators typically do not satisfy all considerations
• Rarely a definitive econometric method
– Need to be aware of the trade offs
– Importance of robustness checking
• Otherwise choice of technique can introduce selection bias into estimation
• Review estimators available in Stata, taking into account
1. Panel dimensions
2. Dynamics
3. Parameter heterogeneity
4. Potential endogeneity
5. Cross-group correlation
26/02/2025 19
26/02/2025 20
 Consideration 1: Dimensions T &N: Where
are panel TS techniques appropriate?
• Panel data where both cross-section (N) and time series (T)
dimensions are at least “moderately large”
– Say, T>25 and N about the same
– Often T>N
– T large enough for separate estimation for each group
• e.g. datasets from international organisations (typically annual data back to 1950)
– Blackburne and Frank (2007): N=24; T= 34
– Eberhardt (2012): N=48; T=33 (unbalanced)
• e.g. Penn World Table
– Referred to a macro panels; “data fields”
• Default position should not be micro-panel estimators
• Different asymptotics
– Difference- and System GMM
• Pooled estimators allowing only the intercepts to differ across groups
• Do not explicitly address parameter heterogeneity
– But interaction terms can be used
– TS panel estimators allow for parameter heterogeneity
• Assumption of parameter homogeneity inappropriate for
26/02/2025
large N & large T panels 21
 A note on Seemingly Unrelated
Regression (SUR)
• Separate regressions
– e.g. same model for different countries (but can include equation-specific
observables
• Allows for common unobservable factors that affect all countries at the
same time
– Contemporaneous error correlation provides additional information to improve the
efficiency of estimates
• Advantages
– SUR generally more efficient than
OLS estimation of separate equations
– Allows for complete panel heterogeneity
• Disadvantages
– Long and narrow datasets
• Small N method: N<10 (Cameron and Trivedi, 2009)
• T relatively large: T/N > 3 (at least) (Beck and Katz, 1995)
– Time-series properties of the data not the focus
• Dynamic SUR models problematic (Greene, 2012 – 7th Ed. - pp.344-45)
26/02/2025 22
 Indicative dimensions of T and N
for dynamic panel estimation
Cross
-section
Large
Small Moderate
Time N>25-30
-series
N<10 10≤N≤25


Micro
• Group Mean estimators (but
Insufficient (?) applied to
Small data for • Bias- macro panels)
T<10 dynamic Corrected  Difference
analysis LSDV GMM
 System
GMM
Moderat • Group Mean
Unobserved
e
10≤T≤2
components
model (CFRs)
• Bias-
Corrected 
LSDV
26/02/2025 5 23
 Consideration 2: Dynamics
• In economics, history matters!
– Economic variables are not independently realised
in every period
• Instead: expectations; anticipations; persistence; adjustment; path
dependency; hysteresis effects;
and so on
• In macro and in microeconomics
• In econometrics, non-stationarity matters
– Problem of spurious regression unless
non-stationary variable cointegrate
– Requires different estimators
• Dynamics typically introduced into models
via the lagged dependent variable
26/02/2025
– Autoregressive models 24
 Introducing history via the lagged
dependent variable
• Starting with the simplified specification in Equation 2,
repeatedly substitute for the lagged dependent variable.
• Substitute for in (2):

• Substitute (3) into (2)

• Substitute for in (4):

• Substitute (5) into (4)

• Gather terms

• … and so on.
26/02/2025 25
 Lagged dependent variable introduces the entire
history of the independent variables
• Dynamic specifications
– Repeated substitution demonstrates that the lagged dependent variable introduces the
entire history of the independent variable(s)
• Equation 6’, current y influenced
– not only by current x
– but also by the cumulated effects from x one period back ()
and two periods back () … and so on
• Unobserved influences (shocks) u also accumulate
• Persistence effects attenuate the more remote the period
– Shown by the increasing exponent on
• Substitutions demonstrate that a dynamic specification includes
the whole history of both the observed and the unobserved influences
on the current value of the dependent variable
• By taking this history into account, we are able to identify the additional short-run,
contemporaneous effects of x on y
– Informative about the process of adjustment of y to x (short-run dynamics)
• Specifying and estimating a static model in the presence of dynamics
– Biased and inconsistent estimates
• Long-run equilibrium effects and short-run deviations not separately identified
26/02/2025 26
• Hence the need for dynamic estimators
 Requirements of a dynamic estimator

• Fully exploit the information content

of time series
• From the Granger Representation Theorem
1. The long-run equilibrium relationship between
variables y and x
2. Adjustment towards long-run equilibrium
• Plus
3. Short-run impact of current changes in x on y

26/02/2025 27
Consideration 3: Parameter heterogeneity
• Pesaran and Smith (1995 & 1996)
– Estimating LR relationships from dynamic heterogeneous
panels
Main conclusions:
 Random coefficients models
do not extend to dynamic panels
– Pooling dynamic heterogeneous panels can give misleading results
Þ Test for common slope coefficients
• usually rejected

 Cross-section estimates
give average LR effects
– The group mean regression
26/02/2025 • Coefficients robust to dynamic misspecification of the underlying time series 28
 Slope heterogeneity in models with panel
data: consequences
• Static models
– Each approach
• Unbiased and consistent estimators
of the average effect ()
– Nothing to choose between the methods
• Dynamic models
– Lagged dependent variable
– 4 approaches not all valid

29
 Initial four approaches to estimating LR
effects with heterogeneous panels
 Mean Group estimator
– Separate regressions for each time series
1
i = 1, 2, …, N  
N
 i
– Take the mean:
• Average effect of some exogenous variable
on the endogenous variable
 Effects models: FE and RE
– Restriction: i = 
 Aggregate time series estimator
– In each period, aggregate data over units
• Pooled across groups
– One time series
• Standard time series analysis
 Aggregate cross section estimator
– In each group, aggregate data over time
– One cross section of group means
• Standard cross-section analysis
30
 average LR effects
 Verdict on initial four approaches
 Separate time series regressions
– OK with large T
• Normal time series analysis
– Results for individual panel members (groups) unreliable
and can be difficult to interpret
• Unless T is large
– But panel averages establish a reliable mean estimate (Eberhardt, 2013, p.442)
• Basis of Mean Group approach
 FE and RE
– Estimates biased and inconsistent
 Aggregation over countries
– Estimates inconsistent
 Aggregation over time
– OK: estimates consistent
• With large T (>25)
– Maybe 15 or less?
• If T small, report that results
“valid approximately”
– Be cautious about conclusions!

For large N and large T

– Estimators from 1 and 4 both give 31
the average LR effect of x on y
Estimating time-series panel models in the presence
of slope heterogeneity: 2 basic approaches

1. Mean Group estimation

– Basis of modern time-series panel estimators
• PDOLS
• MG and PMG
• CCEMG and AMG
2. Group Mean estimation
– Aggregate cross section
• Model formulated in terms of LR coefficients
– No SR dynamics

32
 Implementing the Group-Mean estimator
• Heterogeneous dynamic model
yit = ixit + iyi,t-1 + it (2.1)
• From Smith and Pesaran (1995), p.87

• Aggregate over time

yi  xi  yi ,  1  vi (2.17)
– Where bar  group mean
• Averaged over time t = 1, …,T
– V-bar is a complex error term
• Containing terms from the random coefficients

• For large T (>25)

y
– Mean of lagged dependent variable can be omitted
i i  x  vi
– Hence, estimate:
• Valid asymptotically
• For a finite sample, estimators biased
– Even with large N
• However, in empirical investigation
“the estimates seem quite robust to variations in the number of years used to form
averages”
– Smith and Pesaran (1995), p.101

• Supports conventional wisdom!

– Times series estimates  short-run effects 33
 Why use the Group Mean
cross-section estimator?
• Statistical
– To overcome problems arising from dynamic
heterogeneous panels
– Cross-section analysis much more tractable
• Can implement in any package
• Economic
– To obtain long-run coefficients
• But discards dynamics!

34
 Mean group estimation of
time-series panel models
• Parameters vary across groups
– In “large T panels”, parameters cannot be estimated in a
model that imposes cross-group homogeneity
• Unless the slope coefficients are truly identical across groups
• But slope heterogeneity is pervasive
• “Traditional panel analysis” (“small T”)
– Pool the time series
• Assume parameter homogeneity
– Efficiency gain

• Time-series panel analysis

– Separate time series regressions for the N groups
– Estimates averaged
26/02/2025 • Allows parameter heterogeneity 35
 Consideration 4: Addressing endogeneity
• According to Pesaran (1997)
– The Autoregressive Distributive Lag (ARDL) approach to time-series panel
analysis “continues to be applicable even if the Xt s are endogenous,
irrespective of whether they are I(1) or not”.
• Although these results were obtained in a time-series context, they
have been applied routinely to panel analysis: for example,
according to Pesaran et al. (1999, p.624):
– … it is relatively straightforward to allow for the possible dependence of Xit
on εit when when estimating the long-run coefficients, as long as Xit have
finite-order autoregressive representations.
• Note the doubling subscripting, indicating panel analysis
• The practical implication
– Augmenting the ARDL specification with an adequate number of lags
makes the estimation of the long-run coefficients immune to endogeneity
problems
26/02/2025 • Irrespective of whether the regressors are stationary 36
 Pesaran’s argument
• The presence of endogenous regressors is not a cause for concern in
in estimating long-run parameters in the context of ARDL modelling

When the off-diagonal components of the Variance-Covariance matrix are non-zero

then the derivation of the long-run relationship between yt and xt should allow for the
contemporaneous feedback that exists between the variables
If ut and εt are jointly normally distributed, then:

and ƞt is distributed independently of εt

26/02/2025 37
 Use this result for Equation 5

The difference between equations (5) and (8) is that in (8) “xt can be treated as
strictly exogenous” (Pesaran, 1997, p.183), even if . In this case, the
long-run relationship between yt and xt is given by:

where * denotes long-run or equilibrium values.

• This long-run relationship has two important features:
– “it allows for the direct as well as the indirect effects of changes in xt on yt that take
place through the contemporaneous dependence between ut and εt; and
– “the long-run coefficients  and ϴ will be invariant to the parameters of the xt
process … “.
• And the consequences for estimation are that
– … valid asymptotic inferences on the short-run and long-run parameters can be
made, using the least squares estimates of the ARDL model (8), once the order of the
26/02/2025
ARDL is appropriately augmented to allow for possible contemporaneous correlations 38
between ut and εt
 Consideration 5: Cross-group
variable and/or error correlation
• Variable and/or residual correlation across panel members: due to
– common shocks (e.g. recession)
– spillover effects
• Cross-section dependence (CSD) can lead to
– imprecise estimates (error correlation)
– identification problem (variable correlation)
• Standard panel estimators assume cross-section independence
• Two approaches:
1. Spatial econometrics:
– Econometrician ‘knows’ how panel members are associated/correlated (e.g.
neighbourhood)
– Models this association explicitly employing a weight matrix
• ‘spatially lagged dependent variable’
2. Time-series panel analysis: Common factor models
– Models dependence with unobserved common factors ft
with heterogeneous impact i
– Trick
26/02/2025 is to estimate common factors 39
or blend out their impact on estimation.
 Estimator 1: Bias-Corrected Least Squares
Dummy Variable (LSDV)
• LSDV ( FE) not consistent for finite T in dynamic panel models
– Bias can be substantial even for T=30
• Bias correction of LSDV in dynamic panel models
– Accounts for >90% of the bias
• Advantages
– Suitable for unbalanced panels
• But gaps cause problems in bootstrapping the SEs
– Suitable for small N (e.g. when 10≤N≤20)
• With small N and strictly exogenous regressors,
outperforms Difference GMM and System GMM
• But some evidence of acceptable finite-sample properties
with weakly exogenous regressors (?)
– Suitable for small T
• Bruno (2005, p.487) gives an example with N=29 and T=9 (or even fewer; p.491)
– Addresses small-sample bias
• Disadvantages
– For strictly exogeneous regressors
– Not easy to address heteroskedasticity
• But bootstrapped SEs take “full account of the dependency in the DGP”
(data generating process)
– Does not address parameter heterogeneity or cross-group error
26/02/2025 correlation 40
• So estimate with a full set of period DVs
 Example 1
• Procedure (implemented by the software)
– Standard dynamic panel model

– Subtract bias approximation estimations from LSDV estimates

• Using in Stata
– Download: xtlsdvc
• Command Window
– help xtlsdvc
– use abdata
– xtlsdvc n w k yr1977-yr1984 if ind==4, initial(ab) first bias(3) vcov(100)
• cf Difference GMM (AB) and Bias-Corrected LSDV estimates
– Choice of initial estimator makes little difference, but choose System
GMM (BB) if the dependent variable is highly persistent (say, >0.8)
• Calculate long-run coefficient(s) and SEs, by post estimation
test for non-linear combinations: nlcom
– e.g., nlcom (_b[k]/(1-_b[l1.n]))
26/02/2025 41
 Example 2
• Using Everhardt’s data
– N=48, T=33
– xtlsdvc ly lk year_2-year_33, initial(ab) first bias(3)
vcov(100) lsdv
• Compare Difference GMM (AB), LSDV and
Bias-Corrected LSDV estimates
• Save in a log file for comparison

26/02/2025 42
 Estimator 2: Panel Dynamic
Ordinary Least Squares (PDOLS)
• Model: cointegrating regression
– Yit = αi + ixit + uit
• Cointegration tests of long-run hypotheses using aggregate panel data
– Extension of the individual time-series dynamic OLS
• Not clear why “dynamic” (does not estimate error-correction)
– Single-equation estimator of the cointegrating vector
– Applied to nonstationary data & cointegrated variables
• “Reasonably large” N and T (Pedroni, 2001: N=20, T=246 monthly observations)
• Regression on each panel group
– Yit = αi + ixit +i,jxit-j + u*it
• where j=-p…p and p=1,2,…P
– Cointegrating regression augmented with lead and lagged differences
of the regressor(s) to control for the endogenous feedback effect
• Output
– I coefficients and associated t-statistics average over the entire panel
– Individual group estimates of I coefficients and associated t-statistics
26/02/2025 43
 Features
• Allows for greater flexibility in the presence of heterogeneity
of the cointegrating vectors
• Point estimates from the between-dimension estimator interpreted as the
mean value for the cointegrating vectors
• Test statistics constructed from the between-dimension estimators test the
null hypothesis H0: βi = β0 for all i
against the alternative hypothesis HA: βi  β0
• Values for βi are not constrained to be the same under HA
– Important advantage over test statistics constructed from
the within-dimension estimators
• Averaging across groups to create a single time series
 βi the same under both H0 & HA
– Much lower small-sample size distortion
than the within-dimension estimators
• PDOLS does not take account of cross-group error correlation
– But by default does estimate with common time dummies

26/02/2025 44
 Examples
• Using Stata
– Download: xtpedroni
• Command Window
– help xtpedroni
– xtset country time
• Example 1: use pedronidata
– xtpedroni logexrate logratio, notest lags(5) mlags(5) b(1)
– xtpedroni logexrate logratio, notest lags(5) mlags(5) b(1) notdum
• Without time dummies
• b(1) – compute all t-stats against H0 slope coefficient=1 (Default: b(0))
– Appropriate for long-run PPP
– xtpedroni logexrate logratio, full notest lags(4) mlags(4) b(1) notdum
• Results for each country (no common time dummies)
• Example 2: use manu_prod (Eberhardt’s data)
– xtpedroni loghex loggdp, trend lagselect(aic)
• lags – default is 2
• mlags – default (chosen automatically for each individual)
26/02/2025 45
Estimators 3 and 4: Mean Group (MG) and
Pooled Mean group (PMG)
• For large N and large T panels
– Large N, small T panels
• Individual groups pooled
• FE or RE or RE+IV estimators (e.g. difference and system GMM)
– Only the intercepts vary across groups
– Large N, Large T
• T large enough to fit the model to each group separately
• Assumption of slope homogeneity not appropriate
• Nonstationarity also a concern
• Hence the need for models to estimate nonstationary,
heterogeneous dynamic panels
– MG
• Estimates N time-series regressions and average the coefficients
– PMG
26/02/2025 • Combination of averaging and pooling 46
 Procedure
• Start with an ARDL model
– p lagged values of the dependent variable (starting from 1)
– q current and lagged values of the independent variable(s) (starting from 0)

• Reparameterise (1) into an Error Correction Model (ECM)

26/02/2025 47
 Interpretation
• All terms in the ECM are stationary
– No need for unit root tests in ARDL approach
• ’i is the vector of coefficients that define the long-run
relationship between the variables
• i is the error-correcting speed of adjustment
– i of the form 1-Persistence (i.e. the inverse of persistence)
– If i = 0, then there is no evidence of a long-run relationship
– If i < 0, and statistically significant, then there is negative feedback,
which means that the LR relationship is stable
• i.e. the dependent variable returns to LR equilibrium over time
• ’ij is a vector of coefficients on the differenced values of the
independent variable(s)
– Defines the short-run (impact) relationship(s)
between the differenced dependent variable
and the differenced independent variable(s)
26/02/2025 48
 Three approaches to estimation
• Two extremes
1. Dynamic LSDV estimator (i.e. Dynamic FE)
• Time-series data pooled across groups
• Only the intercepts are allowed to vary by group
• Unless the slopes are identical, estimates are inconsistent and potentially misleading
2. Mean Group estimator
• Fitted separately for each group
• Mean of the estimated coefficients calculated
• Intercepts, slope coefficients, and error variances
are all allowed to differ across groups
• Compromise: Pooled Mean Group estimator
– Combines both pooling and averaging
– Intercepts, short-run coefficients and error variances
all allowed to differ across the groups
• As in MG estimation
– Long-run coefficients constrained to be equal across groups
• As in dynamic FE estimation
• Exploiting similarity between groups to gain efficiency over MG
– Fewer parameters to estimate than in MG
26/02/2025 49
 Key assumption
• Short-run parameter heterogeneity
– Unique features of cross-section groups – e.g.
countries – likely to bring about short-run differences
in response of dependent variable to changes in
independent variables
• Long-run parameter homogeneity
– Common features more likely to yield similar long-run
responses, especially if the groups are similar
• e.g. OECD countries
– To check this assumption: Hausman test
• If the true model is heterogeneous
– MG: always consistent
» But not most efficient if LR parameters homogeneous
26/02/2025 – PMG: inconsistent 50
 PMG: advantages and Disadvantages
• OK in an unbalanced panel
• Even one with gaps
• ARDL platform
– In I(1) panels allows for mix of cointegration I < 0 )
and noncointegration (i = 0)
– Potential endogeneity of X variable(s) addressed
• Assumption of a common long-run equilibrium relationship
is appealing for small sets of arguably ‘similar’ groups
rather than large diverse macro panels
• Error term it
– Assumed to be identically distributed across groups and time
and uncorrelated with the regressors
• Can use cluster-robust SEs with DFE
– But does not address cross-group error correlation
26/02/2025 51
 Example
• Using Everhardt’s data:
– use manu_prod
• Model will not converge with a full set
(minus one) of period dummies
• Estimate with time trend to take care of unmodelled growth
process(es)
– gen trend = year - 1970
– xtpmg d.ly d.lk , lr(l.ly lk trend ) ec(ec) replace pmg
– xtpmg d.ly d.lk , lr(l.ly lk trend ) ec(ec) replace mg
– hausman mg pmg, sigmamore
• Do not reject H0: mg and pmg estimates not systematically different
– Assumption of long-run parameter homogeneity not rejected
• pmg preferred
• Inclusion of the time trend in the CV brings pmg estimate
of the capital effect closer to ccemg estimate
– But not inclusion only of a constant in the CV
• Can add lags of d.ly and d.lk
– Do not make much difference in this case
26/02/2025 52
Estimators 5 and 6: Common Correlated Effects
Mean Group and Augmented Mean Group
• “2nd generation” models (Eberhardt)
– Allow for unobserved correlation across panel members (groups)
• Model
– One covariate and one unobserved common factor
• But generalises to multiple covariates
and multiple unobserved common factors

• Both N and T “moderate to large”

– Eberhardt et al. (2013) estimate on N=119, T=22
• CCE estimators “unlikely to perform as expected” in a short T panel
(e.g. average T=7.5) (p.445)
26/02/2025 53
 Explanation
• Observables
– yit and xit
– i country-specific slope coefficient on xit
• Unobservables uit
– Group fixed effects α1i
• Capture time-invariant heterogeneity across groups
– Unobserved common factor ft with heterogeneous factor loadings λi
to capture
• Time-invariant heterogeneity
• Cross-section dependence
• Unobserved common factor gt
– Allows the observable xit to be driven by common factors other than ft
• Both unobserved common factors ft and gt
– Can be nonlinear and nonstationary
– Influence any cointegrating relationship
• Additional complexity
– ft in both (2) and (3)
– Induces endogeneity in the estimation equation (1)
it and eit assumed white noise
•26/02/2025 54
 Estimation
• All Mean Group estimators follow the same
principle two-step methodology
1. Estimate N group-specific OLS regressions
2. Average the estimated coefficients across groups
• CCEMG and AMG estimators
– Step 1 augmented with additional covariates
– Step 2 the same
• Of interest is the average relationship across panel members
• But all N regression results can be listed
– Enabling analysis of patterns of heterogeneity
– Further insight into the ultimate source of heterogeneity

26/02/2025 55
 CCEMG – implementation
• CCEMG estimator allows for the set up in (1), (2) and (3), which induces
1. cross-section dependence
2. time invariant unobservables with heterogeneous impact across groups
3. Identification problems
• i cannot be identified if xit contains ft
• CCEMG Solution (Pesaran, 2006)
– Augment the group-specific regression equation (1) with cross-section averages of
both the dependent and independent variables
• Using the data for the entire panel
• yi-bar and xi-bar added an additional regressors
in each of the N regression equations
– Step 2 is the usual MG averaging across panel members
• Estimated coefficients on the cross-section average variables
– Not interpretable in any meaningful way
– Present to “blend out” the biasing impact of the unobservable common factors
• Both “weak” factors (e.g. local spillover effects) &“strong” factors (e.g. global shocks)
• 26/02/2025
CCEMG robust to nonstationary common factors 56
 Importance of the cross-section
dimension
• … as the cross-section dimension becomes large, the
unobserved common factor ft
can be captured by a combination of
cross-sectional averages of y and x
– Nicely explained in Eberhardt et al., 2013, p.443
• Parameters on the cross-sectional averages
and intercepts must be group specific
– Allows for both
1. Common factors (shocks, spillovers, etc) (ft)
2. Group-specific reactions to the common factors (λi)
– See equations (2) and (3)
– Achieved in Mean Group estimators by construction
• Each group estimated separately
26/02/2025 57
 AMG
• Alternative to the Pesaran (2006) CCEMG
– Similar performance with respect to bias and RMSE
in panels with nonstationary variables
and multifactor error terms
• Both cointegrated and non-cointegrated
• Main difference is the intended application to
cross-country production functions
– CCEMG
• Unobservable common factor ft treated as a “nuisance”
– To be accounted for but of no interest
– AMG
• In cross-country production functions, unobservables represent total
factor productivity (TFP), a variable of interest
26/02/2025 58
 Other features of CCE estimators
• Focus of CCEMG
– Consistent estimates of the parameters on the observed variable(s) (xit)
– i.e. the mean of the heterogeneous βi
• long-run coefficient(s) - see Equation 1
• Disadvantage
– Not informative about adjustment or short-run dynamics
– Does not address endogeneity of regressor(s)
• Except for endogeneity arising from common factors
• Does not address endogeneity of types common in macro panels
– E.g. simultaneity/feedback effects

• Advantages
– “ … can accommodate a fixed number of strong common factors and an
infinite number of weak common factors … where the former can be thought
of as common global shocks and the latter as local or regional spillover
effects.”
– “… remarkably robust to structural breaks, lack of cointegration, and certain
serial correlation.”
26/02/2025 • Eberhardt et al., 2013, p.443 59
 Example
• Command Window: help xtmg
• Using Everhardt’s data:
– use manu_prod
• N=48, T=33
• Command window
– xtset the data
• Cross-country productivity analysis
– ly and lk both I(1)
• Theory:
– Capital per worker coefficient should be around 1/3 rd
• Compare
– MG (default): xtmg ly lk, trend res(eMG)
– CCEMG: xtmg ly lk, cce trend res(eCMGt)
• Trend option
– Each group-specific regression augmented with a linear trend term
• Takes account of unmodelled growth process(es)
• Can test MG and CCEMG residuals for cross-group correlation
– But not in this case
26/02/2025 • xtcd does not work when gaps in the panel 60
 Choice of estimator: Summary
First check suitability for the dimensions of the data

Does the BCLSDV PDOLS MG PMG CCEMG AMG

estimator allow:
Unbalanced panel √ √ √ √ √ √
Gaps in time series √ √ √ √ √
Long-run equilibrium √ √ √ √ √
Error Correction √ √
Short-run dynamics √ √
Cointegrating vector
for each group √ √ √ √
EC and SR dynamics
for each group √ √
Parameter
heterogeneity √ √ √ √ √
Potential endogeneity √ √ √
Cross-group error
correlation √ √

26/02/2025 61
 So, which estimator?
• Lack of systematic comparisons
• One useful study: Banerjee et al. (2010)
– Monte Carlo simulations
• Findings
1. Simultaneity a much larger source of bias than cross-group dependence
2. As yet, no CCE-IV estimator
• Research ongoing
3. Common factors can - at least partially - be addressed by time dummies
• In the observable part of the model rather than
very complicated unobservable effects in the residuals
• Implications (implicit rather than the authors’ own conclusions)
– CCEMG may not always be preferable to PMG
– Maybe use PMG and/or MG augmented with year effects
• If a full set of year DVs infeasible, try period DVs (e.g. before/after GFC)
• No universally preferred estimator
• Recommendation
1. Consider the nature of the problem to be investigated
and match with the most appropriate estimator
2. Robustness check
26/02/2025 • Are your results an artefact of your chosen estimator? 62
 Using diagnostic tests to decide?
• Econometric models exist on two levels
– As statistical models
– As economic models
• Model has to be valid on both levels
• Statistical validity a necessary condition for a valid econometric model
– Example 1: Single equation time-series analysis using OLS/DOLS
• Regression residuals assumed NIID
– Normally, Individually and Independently Distributed
• Assumptions checked by standard diagnostic tests
– Serial (Auto)correlation in the residual
– Heteroskedasticity
– Autocorrelated Conditional Heteroskedasticity (ARCH)
» Especially for high-frequency data
– Linearity
– Normality
» Not essential for OLS estimation, as long as iid holds
» But informative about the underlying Data Generating process
– Example 2: “Micro” panel analysis (Wide N, Short T) using GMM
• Consistency depends on the validity of the instruments
• Assumptions checked by diagnostic tests
– m1/m2
– Sargan/Hansen

• Usual procedure when deciding between two models

– Give precedence to diagnostic tests
26/02/2025 63
– iff the model is statistically well specified, interpret for economic meaning
 Usual procedure not possible
in time-series panel analysis
• Mainstream panel analysis
– Collections of cross-section data over short periods of time
• Time-series panel analysis
– Collections of time-series data
• Compare
– Time-series analysis
• Highly developed and well-understood diagnostics
to test for statistical misspecification
– Time-series panel analysis (Banerjee et al. 2010, p.2)
• “By and large … a misspecification-test-free zone … issues central to time-
series estimation such as the investigation of model specification have not
been duly emphasised … the lack of residual diagnostics in macro panel
econometrics now more than ever represents a glaring omission.”
• To date, no diagnostic procedures that practitioners can use routinely and with
confidence
– “A great deal of work remains”! Banerjee et al. 2010, p.34
26/02/2025 64
 References
• Key resources: user-written programmes introduced in The Stata Journal
• mtpmg
– Blackburne, E. and Frank M. (2007). Estimation of nonstationary heterogeneous panels. The Stata Journal, 7(2) 197-208.
• xtlsdvc
– Bruno, G. (2005). Estimation and inference in dynamic unbalanced panel-data models with a small number of individuals. The Stata Journal, 5(4)
473-500.
• xtmg
– Eberhardt, M. (2012). Estimating panel time-series models with heterogeneous slopes. The Stata Journal, 12(1) 61-71.
• xtpedroni
– Neal, T. (2014). Panel cointegration analysis with xtpedroni. The Stata Journal, 14(3) 684-692.
• xtwest
– Persyn, D. and Westerlund, J. (2008). Error-Correction – based cointegration tests for panel data. The Stata Journal, 8(2) 232-241.
• Important articles
– Banerjee, A., Eberhardt, M. and Reade, J. (2010). Panel Estimation for Worriers. https://
sites.google.com/site/medevecon/publications-and-working-papers
– Eberhardt, M., Helmers, C. and Strauss H. (2013). Do spillovers mattter when estimating private returns to R&D? The Review of Economics and
Statistics, 95(2) 436-448.
– Maddala, G.S. and S. Wu (1999) 'A comparative study of unit root tests with panel data and a new simple test', Oxford Bulletin of Economics
and Statistics, Vol.61(Special Issue), pp.631-652.
– Pesaran, M.H. (2006). Estimation and inference in large heterogeneous panels with a multifactor error structure. Econometrica 74, 967-1012.
– Pesaran, M. H. (2007). A simple panel unit root test in the presence of cross-section dependence. Journal of Applied Econometrics, 22(2), 265-
312.
– Pesaran, M.H. and R. Smith, 1995, Estimating long-run relationships from dynamic heterogeneous panels, Journal of Econometrics, 68, 79-113.
– Pesaran, M.H., R. Smith and Kyung S.I., 1996, Dynamic linear models for heterogeneous panels, Chapter 8, in L. Mátyás and P. Sevestre, eds.,
The Econometrics of Panel Data: handbook of theory and applications, 2nd edition (Kluwer Academic Publishers, Dordrecht).
– Pesaran, M.H. (1997). The Role of Economic Theory in Modelling the Long-Run. The Economic Journal 107, 178-191.
– Pesaran, M.H., Shin, Y. and Smith, R. (1999). Pooled mean Group Estimation of Dynamic Heterogeneous Panels. Journal of the American
Statistical Association, 94(446) (June) 621-34.
– Pedroni, P. (1999). Critical values for cointegration tests in heterogeneous panels with multiple regressors. Oxford Bulletin of Economics and
Statistics, 61(Special Issue), 653-670.
– Pedroni, P. (2001). Purchasing Power Parity Tests in Cointegrated Panels. The Review of Economics and Statistics, 83(4) 727-731.
• Textbook
– Baltagi, B.H. (2013) Econometric Analysis of Panel Data. 5th Ed. New York: Wiley.
• 26/02/2025
Presentation 65
• MARKUS EBERHARDT, Panel time-series modelling: New tools for analyzing xt data.