Beck y Katz 2011

PL14CH16-Beck ARI 14 April 2011 17:11
Modeling Dynamics in
Time-SeriesCross-Section
ANNUAL
REVIEWS Further
Click here for quick links to
Annual Reviews content online,
Political Economy Data
including:
Other articles in this volume
Top cited articles
Nathaniel Beck1 and Jonathan N. Katz2
Top downloaded articles 1
Department of Politics, New York University, New York, New York 10003;
Annu. Rev. Polit. Sci. 2011.14:331-352. Downloaded from www.annualreviews.org
Our comprehensive search

email: nathaniel.beck@nyu.edu
by University of Massachusetts - Amherst on 10/06/12. For personal use only.
2
Division of the Humanities and Social Sciences, California Institute of Technology,
Pasadena, California 91125; email: jkatz@caltech.edu
Annu. Rev. Polit. Sci. 2011. 14:33152 Keywords

First published online as a Review in Advance on lagged dependent variables, correlated errors, error correction model,
March 17, 2011
non-stationarity, model specication
The Annual Review of Political Science is online at
polisci.annualreviews.org Abstract
This articles doi: This article deals with a variety of dynamic issues in the analysis of
10.1146/annurev-polisci-071510-103222
time-seriescross-section (TSCS) data. Although the issues raised are
Copyright c 2011 by Annual Reviews. general, we focus on applications to comparative political economy,
All rights reserved
which frequently uses TSCS data. We begin with a discussion of spec-
1094-2939/11/0615-0331$20.00 ication and lay out the theoretical differences implied by the various
types of dynamic models that can be estimated. It is shown that there
is nothing pernicious in using a lagged dependent variable and that all
dynamic models either implicitly or explicitly have such a variable; the
differences between the models relate to assumptions about the speeds
of adjustment of measured and unmeasured variables. When adjust-
ment is quick, it is hard to differentiate between the various models;
with slower speeds of adjustment, the various models make sufciently
different predictions that they can be tested against each other. As the
speed of adjustment gets slower and slower, specication (and estima-
tion) gets more and more tricky. We then turn to a discussion of esti-
mation. It is noted that models with both a lagged dependent variable
and serially correlated errors can easily be estimated; it is only ordi-
nary least squares that is inconsistent in this situation. There is a brief
discussion of lagged dependent variables combined with xed effects
and issues related to non-stationarity. We then show how our favored
method of modeling dynamics combines nicely with methods for deal-
ing with other TSCS issues, such as parameter heterogeneity and spatial
dependence. We conclude with two examples.
331
1. INTRODUCTION assumed that the dynamics apply similarly to all

the components of that vector.
Time-seriescross-section (TSCS) data are Where we need to differentiate dynamics,
iid: independent
identically distributed perhaps the most commonly used data in com- we use a second variable (or vector of variables),
parative politics (broadly dened) and particu- zi,t . Because the constant term is typically ir-
larly in comparative political economy.1 There relevant for what we do, we omit it from our
are various issues related to TSCS data; here notation (or include it in the vector of other
we focus on some important ones related to independent variables). We assume that these
the dynamic (time-series) properties of the data. independent variables are exogenous, which is
Obviously many of these issues are similar to clearly both a strong and critical assumption.
those for a single time series, but the context We postpone discussing specications for yi,t
of comparative political economy and the rel- until the next section.
atively short lengths of the TSCS time peri- TSCS data are most commonly used in com-
ods make for some interesting special issues. parative political economy, and so we refer to
We assume that the reader is familiar with time periods as years and units as countries.
the basic statistical issues of time-series data. Given the yearly frequency, we focus on models
Because various specication issues are covered where explicit lags are only of one or two years;
for political scientists elsewhere (Beck 1985, we would not expect to see the higher-order
1991; Keele & Kelly 2006; De Boef & Keele dynamic processes common in standard time-
2008), we go fairly quickly over the basic issues, series analysis of monthly or quarterly data.
spending more time on issues relating to the Although we focus on low-order processes, it
dynamic modeling and interpretation of those is trivial to extend the interpretation, tests, and
models in political economy. estimation to higher-order dynamics.
T must be large enough so that averaging
over time makes sense. In our prior work (Beck
1.1. Notation and Nomenclature & Katz 1995, 1996), we did not examine sit-
Our interest is in modeling the dynamics of uations where T < 15. Political economy data
TSCS models. By dynamics we mean any pro- often span a sample period of 30 or more years
cess where some variable has an impact that is and so there are no problems. This article does
distributed over time. Let yi,t be an observa- not discuss standard survey panels that typically
tion for unit i at time t where i = 1, . . . N and have ve or fewer waves, and there is no reason
t = 1, . . . T . We assume that y is measured as to believe that the methods discussed here ap-
a continuous variable, or at least can be taken ply to such survey panels. We make no assump-
as continuous. In what follows, we typically do tions about N, although in comparative political
not care if we have one or more than one inde- economy it is seldom less than 20 (advanced in-
pendent variable or variables, so let xi,t be either dustrial nations); it can be quite large (>100,000
an observation on a single independent variable for some studies in international relations using
or a vector of such variables; if the latter, it is dyadic data).
We distinguish two types of error terms:
1
i,t refers to an independent identically dis-
Comparative politics refers to any comparison of political
units, so it encompasses almost all of international relations, tributed (iid) error process, whereas i,t refers to
which has countries or country pairs as the unit of analysis, a generic error process that may or may not be
and any study that compares units (regions, states, or coun- iid. Unless specically stated, we restrict non-
ties) within a single country. Our language comes from com-
parative political economy, but everything applies to all other iid processes to simple rst-order autoregres-
types of TSCS studies (both within and beyond political sci- sions. This simplies notation with no loss to
ence) so long as the data are observed over a long enough our argument; it is simple to extend our ar-
period. Adolph et al. (2005) report that by the early 2000s
5% of all political science articles in JSTOR journals used gument to more complicated error processes.
TSCS data. Because coefcients are only interpretable in
332 Beck Katz

the context of a specic model, we superscript ning or the end of the period under study, and
coefcients to indicate the specication they re- the only cost of so doing would be an additional
fer to whenever confusion might arise. We use subscript in the notation.
the terms specication and model interchange- Missing data in the interior of the period
ably, and refer to both as equations when refer- under study is not benign. At a minimum, their
ring to a specic equation in the text. absence causes all the standard problems asso-
When relevant, we use L as the lag operator, ciated with missing data (Little & Rubin 1987).
so that The default solution, list-wise deletion, is well
Lyi,t = yi,t1 if t > 1 1. known to be an inferior solution to the missing-
data problem for cross-sectional data. But the
and Lyi,1 is missing. The rst difference opera-
problem is more severe for TSCS data because
tor is then = 1 L.
the specication invariably includes temporal
lags of the data; even if the model has only
1.2. Stationarity
rst-order lags, each observation with missing

We initially, and for most of the article, as- data leads to the deletion of two data points.
sume that the data are drawn from a station- Thus, even more than with cross-sectional data,
ary process. A univariate process is stationary multiple imputation techniques are de rigueur
if its various moments and cross-moments do for dynamic models that have more than a
not vary with the time period. In particular, trivial amount of missingness. Obviously the
the initial sections assume that the data are amount of missingness will vary as we move
drawn from a covariance stationary process, from studies of advanced industrial societies to
that is, studies of poorer nations, and so the attention
paid to missingness can vary.
E(yi,t ) = 2a.
Although it is easy to say that analysts
Var(yi,t ) = 2 2b. should use Little & Rubins multiple imputa-
tions, the standard methods for cross-sectional
E(yi,t yi,tk ) = k 2c. imputations (hot decking or assuming that
(and similarly for any other random variables). both missing and non-missing observations are
Stationary processes are mean reverting, and essentially multivariate normal) are not appro-
the best long-run forecast for a stationary pro- priate for TSCS data. This is because we know
cess is that mean. Thus, we can think of the a lot about TSCS data. Thus, for example, we
mean as the equilibrium of a stationary pro- would expect that missing economic variables
cess. Alternatively, we can think of the statistical are likely to be highly related to observed values
properties of the data as not varying simply as in nearby countries with similar economies, or
a function of time (so, for example, there are that observations on trending time series can
no trends in the data). We briey discuss non- be imputed with interpolated values. Honaker
stationary data in Section 4. & Kings (2010) Amelia II allows users this
kind of exibility. But there is no mechanistic
solution to the missing-data problem in
1.3. Missing Data political economy. To give but one example,
To simplify notation, we assume that the missingness is often related to civil wars; if
dataset is rectangular, that is, each country is we simply use some complicated averaging
observed for the same time period (which is method to impute missing economic data
called the sample period even though it is not during a civil war, our imputations are likely
a sample of anything). It is easy to extend ev- to be overly optimistic. Analysts using TSCS
erything to a world in which some countries are datasets with signicant missing data can only
not observed for a few years at either the begin- be warned that they must take extreme care.
www.annualreviews.org Time-SeriesCross-Section Data 333

1.4. Roadmap their effect dissipates instantaneously; there are

no delayed effects. (It may be that xi,t is mea-
The next section, on the interpretation of al-
sured with a lag, so the effect could be felt with
LDV: lagged ternative dynamic specication, is the heart of
dependent variable a lag, but the model is still inexible in that
the article. There we deal only with stationary
the effect is completely and only felt at the one
FDL: nite data. The third section briey examines com-
distributed lag model specied year.)
bining dynamics with cross-sectional issues, in
There are several ways to add dynamics to
AR1: rst-order particular accounting for heterogeneity across
autoregressive error the static specication. The simplest is the -
units. The fourth section extends the argument
process nite distributed lag (FDL) model, which as-
to slowly moving and non-stationary data. Two
SC: rst-order serially sumes that the impact of x sets in over two (or
examples of dynamic modeling with political
correlated error model a few) periods but then dissipates completely.
economy TSCS data are discussed in the fth
This specication has
section, and we offer some general conclusions
in the nal section. yi,t = f 1 xi,t + f 2 xi,t1 + i,t 4.
with the obvious generalization for higher-

2. DYNAMIC MODELS: ordered lags. Equation 3 is nested inside
STATIONARY DATA Equation 4 (Equation 3 is a special case of
Equation 4), so testing between the two is sim-
There are various specications for any time-
ple in principle (although the correlation of x
series model; for reviews considering applica-
and its lags makes for a number of practical
tions to political science, see Beck (1991) and
issues).
De Boef & Keele (2008). All time-series speci-
Another commonly used dynamic specica-
cations have identical counterparts in TSCS
tion is to assume that the errors follow a rst-
models. These specications appear in any
order autoregressive (AR1) process rather than
standard text, so we discuss general specica-
the iid process of Equation 3. If we assume that
tion issues without either citation or claim of
the errors follow an AR1 process, we have a se-
originality.
rially correlated (SC) error model:
In our prior work (Beck & Katz 1996) we
argued that a lagged dependent variable (LDV) yi,t = SC xi,t + i,t + i,t1 5a.
specication is often adequate; because that has
sparked some discussion (Achen 2000, Keele & i,t
= SC xi,t + 5b.
Kelly 2006), we spend some time on this issue. 1 L
After discussing a variety of specications, we = SC xi,t + yi,t1 SC xi,t1 + i,t . 5c.
discuss issues of interpretation and estimation.2 The formulation in Equation 5c makes the dy-
namics implied by the model clearer and also
makes it easier to compare various models.
2.1. Dynamic Specifications Another alternative is the LDV model (with
The generic static (non-dynamic) specication iid errors):
is
yi,t = LDV xi,t + yi,t1 + i,t 6a.
yi,t = xi,t + i,t .
s
3. xi,t i,t
= LDV + . 6b.
This specication is static because any changes 1 L 1 L
in x or the errors are felt instantaneously and As Equation 6b makes clear, the LDV model
simply assumes that the effect of x decays geo-
2
We often refer to Achens (2000) critiques of the use of metrically (and for a vector of independent vari-
LDVs. Although it is odd to spend time critiquing a decade- ables, all decay geometrically at the same rate).
old unpublished paper, this paper has been inuential (more Note also that the compound error term is an
than 300 Google Scholar citations as of this writing). We
only deal with the portions of Achens paper relevant to issues innite geometric sum (with the same decay pa-
raised here. rameter as for x); this error term is equivalent to
334 Beck Katz

a rst-order moving average (MA1) error pro- has run long enough for y to be at its equilib-
cess, again with its decay parameter constrained rium value (stationarity implies the existence of
to equal , the rate at which the effect of x on y such an equilibrium). We can then think of a
MA1: rst-order
decays. one-time shock in x (or ) of one unit, with a moving average
Both the SC and LDV specications are spe- subsequent return to equilibrium (zero for the process
cial cases of the autoregressive distributed lag error) the next year; if we then plot y against ADL: autoregressive
(ADL) model, this, we get an impulse response function (IRF). distributed lag model
Alternatively, we can shock x by one unit and
yi,t = ADL xi,t + yi,t1 + xi,t1 + i,t , 7. EC: error correction
let it stay at the new value; the plot of y against model
where Equation 5c imposes the constraint that x is a unit response function (URF). IRF: impulse
= ADL and Equation 6a assumes that The static specication assumes that all vari- response function
= 0. The nesting of both the LDV and SC ables have an instantaneous and only an instan- URF: unit response
specications within the ADL specication al- taneous impact. Thus, the IRF for either x or function
is a spike, associated with an instantaneous
lows for testing between the various models.

For interpretative purposes, it can be help- change in y, and if x or then returns to its pre-
ful to rewrite the ADL model in error correc- vious value in the next period, y immediately
tion (EC) form (Davidson et al. 1978). To do also returns to its previous value. The URF is
this, subtract yi,t1 from both sides of the ADL simply a step function, with the height of the
model to get a rst difference in y on the left- single step being s .
hand side, and add and subtract ADL xi,t1 from The FDL model generalizes this. The URF
the right-hand side to get a rst difference of x has two steps, of height f 1 and f 1 + f 2 , and
in the specication. This leads to the interval between the steps is one year. Thus,
unlike in the simple static model, if x changes, it
yi,t = EC xi,t (yi,t1 xi,t1 ) + i,t , 8.
takes two years for the full effect of the change
which allows for the nice interpretation that to be felt, but the effect is fully felt in those two
short-run changes in y are a function of both years. For example, it may take one year for a
short-run changes in x and how much x and party to have an impact on unemployment, but
y were out of equilibrium last year, where the it may be the case that after that year the new
equilibrium y and x are given by yi,t = xi,t party in ofce has done all it can and will do
and the speed of equilibration (per year) is . in terms of changing unemployment. Similarly,
The coefcients of the EC model can be easily an institutional change may not have all of its
derived from the corresponding ADL model: impact immediately, but the full impact may
ADL
EC = ADL , = 1 and = + 1
. For occur within the space of a year.
comparison with other models the ADL model We could add more lags to Equation 4,
works better, but for direct substantive inter- allowing for more complicated dynamics.
pretation of the coefcients the EC model is But time series within a country are often
easier to work with (since one can directly read temporally correlated, so multicolinearity
off the short-run impact of a change in x as well makes it difcult to get good estimates of the
as various long-run impacts). Because the two coefcients of the FDL specication with many
are identical, either one can be estimated. We lags of x. Given the annual frequency of much of
return to the EC model when we deal with non- the data seen in comparative political economy,
stationary data in Section 4. the problem of having to add too many lags to
Equation 4 (say more than one additional lag)
may not, in practice, be a problem. It may be
2.2. Interpretation unlikely that interesting institutional changes
To see how the various specications differ, we have only an immediate impact, but the FDL
turn to unit and impulse response functions. model might be appropriate. It surely should be
Since x itself is stochastic, assume the process borne in mind in thinking about appropriate

specications, and, as we shall see, it combines various models differ in the assumptions they
nicely with some other specications. impose on the dynamics that govern how x and
The SC model has a different IRF for x and the errors impact y. None of the dynamic speci-
the error. The IRF for x is a spike, identical cations can be more or less right a priori. Later,
to that of the static model; the IRF for the er- we discuss some estimation (see Section 2.5) and
ror has a declining geometric form with rate of testing (see Section 2.6) issues, but for now we
decay . It seems odd that all the omitted vari- can say that various theories would suggest var-
ables have a declining geometric IRF but the x ious specications. The most important issue
we are modeling has only an instantaneous im- is whether we think a change in some variable
pact. Maybe that is correct, but this is not the is felt only immediately or whether its impact is
rst specication that would occur to us. The distributed over time; in the latter case, do we
SC specication can be generalized by adding think that a simple structure, such as a declin-
more lags of x, but we would still have very dif- ing geometric form, is adequate? How would
ferent dynamics for x and the unobserved vari- we expect an institutional change to affect some
ables in the error term. One should clearly y of interest in terms of the timing of that ef-
have a reason to believe that dynamics are of fect? If only immediately or completely in one
this form before using the SC specication. or two years, the SC or FDL model seems right;
The LDV model has an IRF for both x and if we expect some initial effect that increases to
the error that has a declining geometric form; some limit over time, the LDV or ADL model
the initial response is LDV (or 1 for the error); would be used. But there is nothing atheoretical
this declines to zero geometrically at a rate . about the use of a lagged dependent variable,
Although the effect never completely dissipates, and there is nothing that should lead anyone
it becomes tiny fairly quickly unless is almost to think the use of a lagged dependent variable
one. The URF starts with an effect LDV im- causes incorrect harm. It may cause correct
LDV
mediately, increasing to 1 . If is close to harm, in that it may keep us from incorrectly
one, the long-run impact of x can be 10 or more concluding that x has a big effect when it does
times the immediate impact. not, but that cannot be a bad thing. As has been
Although the ADL specication appears to well known, and as Hibbs (1974) showed three
be much more exible, it actually has an IRF decades ago for political science, the correct
similar to that of the LDV specication, other modeling and estimation of time-series models
than in the rst year (and is identical for a shock often undoes seemingly obvious ndings.
to the error process). Initially, y changes by A related way to say the same thing is that
ADL units; in the next period the change is each of the models (for stationary data) implies
ADL + , which then dies out geometrically some long-run equilibrium and a speed with
at a rate . Thus, the ADL specication is only which that equilibrium is reached after some
a bit more general than the LDV specication. shock to the system. It is easy to solve for equi-
It does allow for the maximal impact of x to oc- libria (if they exist) by noting that in equilibrium
cur a year later, rather than instantaneously (or, both x and y are stable. Let yE and xE refer to
more generally, the effect of x after one period equilibrium x and y. Then for the ADL model
is not constrained to be the immediate impact it must be true that
with one years decay). This may be important
y E = ADL x E + y E + x E , 9.
in some applications. A comparison of the vari-
( ADL + x
ous IRFs and URFs is in Figure 1, which clearly yielding y E = 1
E
(|| < 1 by station-
shows that the difference between the speci- arity). This is easier to see in the EC form,
cations has simply to do with the timing of the where yi,t1 = xi,t1 in equilibrium and is the
adjustment to y after a change in x. rate (per year) at which y returns to this equi-
Before we get to slightly more complicated librium. All the models for stationary data im-
models, this analysis tells us several things. The ply both a long-run equilibrium and a speed of
336 Beck Katz

A
2.0 ADL
AR1
FDL
LDV
1.5
Response
1.0
0.5
0.0
2 4 6 8 10
Year
B 2.0
1.5
Response
1.0
ADL
0.5 AR1
FDL
LDV
2 4 6 8 10
Year
Figure 1
Comparison of impulse and unit response functions for four specications: autoregressive distributed lag
(ADL, yi,t = 0.2xi,t + 0.5yi,t1 + 0.8xi,t1 ), nite distributed lag (FDL, yi,t = 1.5xi,t + 0.5xi,t1 ), lagged
dependent variable (LDV, yi,t = 1.0xi,t + 0.5yi,t1 ), and autoregressive errors (Static, yi,t = 2xi,t ).
equilibration, with the different parameter con- 2.3. Higher-Order Processes

straints determining these long-run features. and Other Complications
Each of these models implies different short- We can generalize any of these models to allow
and long-run reactions of y to x, and standard for non-iid errors and higher-order dynamics.
econometric methods (see Section 2.6) can be However, because our applications typically use
used to discriminate between them. annual data, it is often the case that rst-order

error processes sufce, and it would be unusual 2.4. More Complicated

to have more than second-order processes. As DynamicsMultiple
we shall see, it is easy to test for higher-order Independent Variables
error processes, so there is no reason to sim-
We typically have more than one independent
ply assume that errors are iid or only follow
variable. How much dynamic generality can or
a rst-order process. For notational simplicity,
should be allowed for? One easy generalization
we restrict ourselves to second-order processes,
is to allow for two independent (or sets of in-
but the generalization is obvious.
dependent) variables, x and z. Allowing also for
Consider the LDV model with AR1 errors,
a separate speed of adjustment for the errors
in which
yields
i,t
yi,t = LDV xi,t + yi,t1 + . 10. xi,t zi,t i,t
1 L yi,t = + + . 12.
1 x L 1 z L 1 L
After multiplying through by (1 L), we get
a model with two lags of y, x and lagged x and Obviously each new variable now requires us
some constraints on the parameters; if we gen- to estimate two additional parameters. Also, on
eralize the ADL model similarly, we get a model multiplying out the lag structures, we see that
with two lags of both y and x and more con- with three separate speeds of adjustment we
straints. The interpretation of this model is sim- have a third-order lag polynomial multiplying
ilar to the model with iid errors. y, which means that we will have the rst three
We have already seen that the LDV model lags of y on the right-hand side of the specica-
with iid errors is equivalent to a model where tion (and two lags of both x and z) and a second-
the effect of all the independent variables and order moving average error process. Although
the error decline at the same geometric rate. But there are many constraints on the parameters
if we assume that the errors, that is, omitted or of this model, the need for three lags of y costs
unmeasured variables, follow an MA1 process us three years worth of observations (assuming
with the same decay rate, , as for the measured the original dataset contained as many obser-
variables (which may or may not be a good as- vations as were available). With k independent
sumption), then we have variables, we would lose k+1 years of data; for
a typical problem, where T is perhaps 30 and
yi,t = LDV xi,t + yi,t1 + (1 L)i,t , 11a. k is perhaps 5, this is nontrivial. Thus, we are
unlikely to ever be able to (or want to) estimate
which simplies to
a model where each variable has its own speed
xi,t of adjustment.
yi,t = LDV + i,t . 11b.
1 L But we might get some leverage by allowing
for two kinds of independent variables: those
That is, we have a model that combines a geo-
where adjustment (speed of return to equi-
metrically declining impact of x on y with iid er-
librium) is relatively fast (x) and those where
rors. It is surely more likely that the errors are
the system returns to equilibrium much more
correlated than that they are independent. Of
slowly. To simplify, assume the error process
course, the most likely case is that the errors are
shows the same slower adjustment speed as z;
neither iid nor MA1 with the same dynamics as
we can obviously build more complex models,
x, so we should entertain a more general spec-
but they bring nothing additional to this
ication, where the effects of both measured
discussion. We then would have
and unmeasured variables have a declining ge-
ometric impact with different rates of decline. zi,t i,t
yi,t = x xi,t + z + 13a.
The simplest such specication is Equation 10. 1 L 1 L
We return to this more general specication in
Section 2.5. = x xi,t x xi,t1 +z zi,t +yi,t1 +i,t . 13b.
338 Beck Katz

Thus, at the cost of one extra parameter, we can Remember that the error term is simply all the
allow some variables to have only an immediate omitted variables, that is, everything that de-
or very quick effect, while others have a slower termines y that is not explained by x. If we ad-
effect, with that effect setting in geometrically. join yi,t1 to the specication, the error in that
With enough years we could estimate more new specication is i,t yi,t1 , where i,t is
complex models, allowing for multiple dynamic the original error in Equation 14a, not some
processes, but such an opportunity is unlikely generic error term. Since the i,t are serially cor-
to present itself in studies of comparative related because they contain a common omitted
political economy. We could also generalize variable, and yi,t1 contains the omitted vari-
the model by allowing for the lags of x and z to ables at time t 1, including yi,t1 will almost
enter without constraint. It is possible to test certainly lower the degree of serial correlation,
for whether these various complications are and often will eliminate it. But there is no reason
supported by the data, or whether they simply to simply hope for this; we can estimate (using
ask too much of the data. As always, it is easy OLS) the LDV model assuming iid errors, and
enough to test and then make a decision. then test the null hypothesis that the errors are
independent using a Lagrange multiplier test.
(This only requires that OLS be consistent un-
2.5. Estimation Issues der the null of iid errors, which it is.) The test
As is well known, a specication with no lagged is trivial to implement, by simply regressing the
dependent variable but serially correlated er- residuals from the OLS regression on the lag(s)
rors is easy to estimate using any of several vari- of this residual and all the independent vari-
ants of feasible generalized least squares, with ables including the lagged dependent variable
the Cochrane-Orcutt iterated procedure being with NTR2 from this regression, being asymp-
the most well known. It is also easy to estimate totically distributed 2 with degrees of freedom
such a model via maximum likelihood, breaking equal to the number of lags tested. If, as often
up the full likelihood into a product of condi- happens, we do not reject the null that the re-
tional likelihoods. maining errors are iid, we can continue with
The LDV model with iid errors is optimally the OLS estimates. If we do reject that null, we
estimated by ordinary least squares regression should estimate a more complicated model.
(OLS). However, it is also well known that Obviously, failing to reject the null of no
OLS yields inconsistent estimates of the serial correlation of the errors is not the same
LDV model if the error process is serially thing as knowing there is no serial correlation
correlated. Perhaps less well known is that of the errors. Is this incorrect logic in inter-
Cochrane-Orcutt or maximum likelihood pro- preting a failure to reject the null likely to
vides consistent estimates of the LDV model cause problems? There are two reasons to be
with serially correlated errors by accounting sanguine here. First, the large amount of data
for that serial correlation (Hamilton 1994, in typical TSCS studies gives the Lagrange
p. 226). Thus, it is easy to correctly estimate multiplier test good power. In our rst example
the LDV model while allowing for serially (Section 5.1), with 300 total observations, the
correlated errors if analysts wish to do so. But Lagrange multiplier test detected a serial cor-
we hope that analysts will not wish to do so. relation of the errors of 0.10. It is also the case
It is often the case that the inclusion of a that ignoring a small amount of serial correla-
lagged dependent variable eliminates almost all tion (that is, estimating the LDV model with
serial correlation of the errors. To see this, start OLS as if there were no serial correlation) leads
with the SC equation: to only small amounts of bias. As Achen (2000,
p. 13) elegantly shows, the estimation bias in
yi,t = SC xi,t + i,t 14a. incorrectly using OLS to estimate the LDV
i,t = i,t + i,t1 . 14b. model with serially correlated errors is directly

proportional to that serial correlation. Applied specication better ts the data. The LDV
researchers make many assumptions to simplify model assumes = 0 (in Equation 7), whereas
analysis, assumptions that are never exactly cor- the SC model assumes = ADL . Thus,
ADLLDV2
specification: the rect. Ignoring a small serial correlation of the we can estimate the full ADL model and test
autoregressive errors is surely one of the more benign mistakes. whether = 0 or = ADL .3 If both sim-
distributed lag As we shall see in Section 3, a number of plications are rejected, we can retain the more
specication with a fruitful avenues of investigation are open if the complicated ADL model.4 Even in the absence
second lag of the
errors are either uncorrelated or sufciently of a precise test, the ADL estimates will often
dependent variable
uncorrelated that we can ignore their small indicate which simplication is not too costly
correlation. But what if a researcher is not so to impose.
sanguine? As we have seen in Section 2.5, one For fast dynamics (where is close to zero),
can easily estimate, using methods other than it will be hard to distinguish between the LDV
OLS, the ADL model with serially correlated and SC specicationsor, alternatively, it does
errors. But a more fruitful approach, as shown not make much difference which specication
in Section 2.3, is to include second-order lags we use. To see this, note that if the SC model
of the variables in the ADL specication; this is correct, but we estimate the LDV model, we
ADL2 specication can be appropriately are incorrectly omitting the lagged x variable
estimated by OLS, once again allowing the although it should be in the specication, but
researcher to more easily examine other inter- with a constrained coefcient . As goes to
esting features of the data. Of course the same zero, the bias from failing to include this term
Lagrange multiplier testing procedure should goes to zero. Similarly, if we incorrectly esti-
rst be used to test for remaining serial corre- mate the SC model when the LDV model is
lation, but with annual data we can be hopeful correct, we have incorrectly included in the
that we will not need highly complicated lag specication the lagged x variable, with co-
structures in the preferred specication. efcient . Again, as goes to zero, this
Obviously more parsimonious specications goes to zero. Thus, we might nd ourselves
are easier to interpret (and convey to the not rejecting either the LDV or SC speci-
reader), and so more complicated specications cations in favor of the more general specica-
with higher-order lags come at a cost. Thus, tion, but for small it matters little. As grows
we might want to consider models intermedi- larger the two models diverge, and so we have
ate between the ADL and ADL2 models. One a better chance of discriminating between the
obvious choice is to simply append a second lag specications.
of the dependent variable to the ADL speci- This view is different from the conventional
cation; this is analogous to moving from the wisdom on omitted-variable bias. It is normally
static to the LDV specications, as discussed thought to be worse to incorrectly exclude
above. This simpler specication, ADLLDV2, than to incorrectly include a variable. This
should be tested to see if the errors are iid. The
ADLLDV2 specication may be a good com-
3
promise between parsimony and delity to im- The rst test is an ordinary t-test. The second is easiest via
a linear approximation to the nonlinear constraint using a
portant features of the data; in our rst example Taylor series (Greene 2008, pp. 9698); this test is imple-
this is our preferred model. In other cases even mented in some common statistical packages such as Stata.
simpler models may provide a better tradeoff 4
Starting with the ADL model and then testing whether sim-
between the various goals. plications are consistent with the data is part of the idea of
general-to-simple testing (also called the encompassing ap-
proach) espoused by Hendry and his colleagues (Hendry &
Mizon 1978, Mizon 1984). This approach could start with
2.6. Discriminating Between Models a more complicated model with higher-order specications,
but given annual data, the ADL model with no more than two
We can use the fact that the ADL model lags is often the most complicated specication that need be
nests the LDV and SC models to test which considered.
340 Beck Katz

difference is because both models constrain the

coefcient of the lagged x, and so the SC model WHY NOT JUST CORRECT THE
forces the lagged x to be in the specication. STANDARD ERRORS?
But if we start with the ADL model and then
test for whether simplications are consistent Most standard programs that allow for modeling complicated
with the data, we will not be misled. This test- cross-sectional situations do not allow for temporally correlated
ing of simplications is easy to extend to more errors. Some researchers try to solve this problem by using sim-
complicated models, such as Equation 13b. ple models and then correcting the standard errors using some
variant of Hubers (1967) robust standard errors. This is the
reasoning behind our recommendation to use PCSEs (panel cor-
3. COMBINING DYNAMIC AND rected standard errors) to deal with some difcult cross-sectional
CROSS-SECTIONAL ISSUES complications of the error process. There are similar autocor-
Modeling dynamics with TSCS data is only relation consistent standard errors (Newey & West 1987). We
half the job; clearly analysts also need to model do not recommend these because failing to account for serially
the cross-sectional properties of the data. We correlated errors often leads to substantial inefciencies in esti-
have discussed various cross-sectional issues for mation as well as incorrect standard errors; failing to account for
TSCS data elsewhere (Beck & Katz 1995, Beck cross-sectional problems in the data is usually less serious. In any
2001). Here we discuss some issues that re- event, users of our preferred methods have no need to resort to
late to the interaction of modeling dynam- autocorrelation consistent standard errors.
ics and cross-sectional issues. For reasons of
space, we omit discussion of dynamics with
discrete dependent variables (see Beck et al. ables in one country impact other countries.
1998). Dynamics are no less important in mod- Models of the political causes of economic per-
els with discrete dependent variables, but the formance, for example, must take into account
recommended modeling is different for that that the economic performance of any country
situation. is a function of the economic performance of
its trading partners. These issues have been dis-
cussed in the context of TSCS data elsewhere
3.1. Independent Errors Simplify (Beck et al. 2006, Franzese & Hayes 2007), and
Cross-Sectional Modeling here we simply point out that our preferred
We have advocated modeling dynamics by in- approach to dynamics makes it easy for ana-
cluding appropriate current and lagged values lysts to deal with this critical cross-sectional
of the xs and lagged values of the dependent issue.
variable so that the resulting errors appear to Another cross-sectional feature that should
be serially independent, enabling easy inter- be considered (see Beck & Katz 2007 and the
pretation and estimation. This approach makes references cited there) is that parameters may
it much simpler to model cross-sectional sit- vary randomly by country, and possibly as a
uations. Most standard programs that allow function of country-level covariates. It is easy
for modeling complicated cross-sectional situ- to allow for this by using the random coef-
ations do not allow for temporally correlated cients model (which is equivalent to a mixed
errors. Although this is a practical rather than or hierarchical or multilevel model) if the
a theoretical issue, some estimation methods error process is iid. Note that one of the ran-
are sufciently complex that one really wants domly varying parameters can be that of the
to use a canned program (see Sidebar Why lagged dependent variable, the parameter that
Not Just Correct the Standard Errors?). controls the speed of adjustment in the model.
In particular, realistic political economy Perhaps countries differ in that speed of adjust-
models often should allow for spatial ef- ment. As we see in Section 5.1, this issue is easy
fects, that is, they should recognize that vari- to examine when errors are iid.

3.2. Fixed Effects and Lagged similar advice following a similar discussion of
Dependent Variables this issue.
Perhaps the most common cross-sectional is-
sue is heterogeneity of the intercepts. In the
TSCS context, this is usually dealt with by 4. NON-STATIONARITY IN
adding xed effects (country-specic inter- POLITICAL ECONOMY
cepts) to the specication. We would adjoin TSCS DATA
these country-specic intercepts to the pre- Before we look at some examples, one topic
ferred ADL specication. But here we get into remains: what to do with non-stationary data.
potential trouble, since it is well known that au- During the past two decades, with the pioneer-
toregressive models with xed effects lead to bi- ing work of Engle & Granger (1987), time-
ased parameter estimates (Nickell 1981). This series econometrics has been dominated by the
bias is induced because centering all variables study of non-stationary series. There are many
by country, which eliminates the heterogeneity ways to violate the assumptions of stationar-
of the constant term, induces a correlation be- ity presented in Equation 2, but most of the
tween the centered lagged dependent variable work has focused on the issue of unit roots
and the centered error term. or integrated series in which shocks to the se-
It is also well known that this bias is of order ries accumulate forever. These series are long-
1
T
, and almost all of the work on this prob- memoried; even distant shocks persist to the
lem has been in the context of small-T pan- present. The key question is how to estimate
els. When T is 2 or 3, the bias is indeed severe models where the data are integrated (we re-
(50% or so). But when T is 20 or more, the bias strict ourselves to integration of order one with
becomes small. no loss of generality). Such data, denoted I(1),
Various corrections for this bias are well are not stationary but their rst difference is sta-
known. Most of them involve the use of in- tionary. The simplest example of such an I(1)
strumental variables, building on the work of process is a random walk, where
Anderson & Hsiao (1982). As is often the case,
yi,t = yi,t1 + i,t , 15.
it is hard to nd good instruments, and so
the instrumental variable corrections often ob- with i,t being stationary by denition. Inte-
tain consistency at the price of rather poor - grated data look very different from data gener-
nite sample properties. Other estimators (Kiviet ated by a stationary process. Most importantly,
1995) are hard to combine with other methods they do not have equilibria (because there is no
and hard to generalize to even non-rectangular mean reversion), and the best prediction of an
data sets. integrated series many periods ahead is the cur-
We ran Monte Carlo experiments to com- rent value of that series.
pare OLS estimation of a simple LDV model There is a huge literature on estimating
with xed effects to the Kiviet and Anderson- models with integrated data. Such methods
Hsiao estimators (Beck & Katz 2009). For the must take into account that standard asymptotic
T s seen typically in TSCS analysis (20 or theory does not apply, and also that
more), OLS performs about as well as Kiviet
lim Var(yi,t ) = . 16.
and much better than Anderson-Hsiao. Given t
the advantages of the OLS method discussed Thus, if we wait long enough, any integrated
in the previous subsection, we do not hesitate series will wander innitely far from its mean.
to recommend OLS when country-specic in- Much work on both diagnosing and estimating
tercepts must be adjoined to the specication models with integrated series builds on both
of a TSCS model. Judson & Owen (1999) give these issues. Our interest is not in the estimation
342 Beck Katz

of single time series but rather TSCS political and to fall when high, or for Christian Demo-
economy data.5 cratic cabinet strength to exhibit a similar ten-
Political economy data is typically observed dency? In the Huber & Stephens data, social
annually for relatively short periods of time security spending ranges only between 3% and
(often 2040 years). Of most relevance, dur- 33% of the budget, and Christian Democratic
ing that time period we often observe very few cabinet strength ranges between 0% and 34%.
cycles. Thus, although the series may be very Even though these series are very persistent,
persistent, we have no idea if a longer time they simply cannot be I(1). The impressive ap-
period would show the series to be stationary paratus built over the past two decades to esti-
(though with a slow speed of adjustment) or mate models with I(1) series does not provide
non-stationary. These annual observations on, the tools needed for many, if not most, political
for example, GDP or left political control of economy TSCS datasets.
the economy are very different from the daily One possibility is to induce stationarity by
observations we may have on nancial rates. So rst differencing all slowly changing variables,
although it may appear from an autoregression leading to a model that explains changes in y by
that some political economy series have unit changes in x. In practice, rst-difference mod-
roots, is this the right characterization of these els often perform poorly (at least from the per-
series? For example, using Huber & Stephens spective of the researcher, for whom changes in
(2001) data, an autoregression of social secu- x appear unrelated to changes in y). Modeling
rity on its lag yields a point estimate of the rst differences also throws out any long-run
autoregressive coefcient of 1.003 with a stan- information about y and x, so the effect of a
dard error of 0.009; a similar autoregression of change in x is the same regardless of whether y
Christian Democratic party cabinet participa- is high or low by historical standards.
tion yields 1.03 with a standard error of 0.001. Fortunately, the modeling issue is not really
It does not take heavy-duty statistical testing to about the univariate properties of any series
see we cannot reject the null (that the autore- but the properties of the stochastic process
gressive coefcient is one) in favor of the alter- that generated the ys conditional on the
native (that it is less than one). But does this observed covariates. Even with data similar
mean that we think the series might be I(1)? to Huber & Stephens, the errors may appear
If these series had unit roots, they would stationary and so the methods of the previous
tend to wander far from their means, and the section can be used. In particular, whether the
variance of the observations would grow larger series are integrated or stationary but slowly
and larger over time (a similar argument is made moving, they may be well modeled by the EC
by Alvarez & Katz 2000). But by denition both specication (Equation 8), which, as we have
the proportion of the budget spent on social se- seen, is just an alternative parameterization of
curity and Christian Democratic cabinet par- the ADL model. The EC form is nice because
ticipation are bounded between 0% and 100%, it combines the short-run rst-differences
which then bounds how large their variances model with the long-run tendency for series to
can become. Further, if either series were I(1), be in equilibrium. If the estimated in the EC
then we would be equally likely to see an in- specication is zero, that indicates that y and x
crease or decrease in either variable regardless have no long-run equilibrium relationship. We
of its present value. Do we really believe that have already seen that if x and y are stationary,
there is no tendency for social security spend- they always have a long-run relationship, so this
ing to be more likely to rise when it is low is only a problem if the series are integrated.
In other words, if the series are stationary but
5
adjust very slowly, the EC (or equivalent ADL)
There is a literature on panel unit roots (Im et al. 2003, Levin
et al. 2002), but at this point the literature is still largely about model is a good place to start, and if the series
testing for unit roots. are integrated, either the EC model will work

(the series are said to be co-integrated) or the dynamics in political economy TSCS datasets.
residuals will appear highly correlated. Because The rst example, presented in some detail,
our preferred methodology chooses speci- looks at the political determinants of capital
cations with almost uncorrelated residuals, it taxation rates where adjustment speeds are
should never lead to choosing an incorrect EC fairly slow. The second example, presented
(or ADL) specication. more cursorily, looks at the impact of political
Why do we propose ignoring much of what variables on the growth of GDP. In the GDP
has dominated econometric argument for two example, where the dynamics are quite fast, the
decades? First, economists study many series specication choice has fewer consequences.
(such as interest or exchange rates) that inher- All computations were done using Stata 11.1,
ently are in levels, and so are likely to be inte- with data kindly provided by Geoff Garrett.
grated; variables in political economy are often Although our analysis is different from those
expressed as a proportion of GDP or the gov- of Garrett & Mitchell, we began by easily
ernment budget and hence are much less likely replicating their results.
to be integrated. Other political variables, such

as party control of government, may be per-
sistent, but cannot possibly be integrated (they 5.1. Capital Taxation Rates
take on values of zero and one only, and so have Our rst example models capital taxation rates
neither innite variance nor no tendency to re- in 16 OECD nations from 1961 through 1993
vert back toward the mean). A second differ- using the data and specication of Garrett &
ence is that economists have no theory about Mitchell (2001).7 Obviously tax rates move
whether one short-run exchange rate adjusts to relatively slowly over time; the autoregressive
a second rate, or the second rate adjusts to the coefcient of tax rates is 0.77. Thus, although
rst, or both; this leads to complicated estima- tax rates are clearly stationary, it will take some
tion issues. In many political economy mod- number of years for the system to get close to
els, it is clear that y adjusts to x but not vice fully adjusting; it takes about 2.65 years for any
versa. We think that left governments increase shock to dissipate.
spending but we do not think that low spending Before estimation, one should examine the
leads directly to a right-wing government (Beck data to see whether there is sufcient within-
1992). Thus, even with highly persistent data, country heterogeneity to make TSCS analy-
the EC (or ADL) model, estimated by OLS, will sis meaningful, to see whether there appears to
quite often work well, and, when it fails, simple be very much inter-country heterogeneity that
tests and a rigorous methodology will indicate might need to be modeled, and to see whether
that failure.6 there are any temporal pattens, such as trends,
that need to be modeled. For the rst two issues
a standard country-specic box plot of tax rates,
5. EXAMPLES shown in Figure 2, is appropriate; for the third
In this section we consider two examples question time-series plots by country, shown in
to explore the practical issues in estimating Figure 3, are more useful.
6
There is a slight technical problem in that the distribution
of the estimated is not normal if the series are not co- 7
The data set is not rectangular; some countries only report
integrated. Instead, they have a Dickey-Fuller type distribu- tax rates for a portion of the period under study. In total there
tion, which has fatter tails. Thus, there may be some cases were 322 observations after one drops missing data at the
where a standard test of the null hypothesis that = 0 yields beginning of a period (and omits observations with missing
incorrect conclusions. But given the large n and T of TSCS lagged data so that all results pertain to the same data). The
data, in many cases it is clear that the EC model is adequate extra lag in the ADLLDV2 model leads to the loss of the
or not, and if we incorrectly assume stationarity, consistent rst observation for each country, yielding 306 observations
application of appropriate standard methods will indicate the for that estimation. This loss of data points is an additional
problem. reason why analysts may not prefer this specication.
344 Beck Katz

Australia Austria Belgium Canada

80
60
40
20
Capital Taxation Rate
Denmark Finland France Germany

80
60
40
20
Italy Japan Netherlands Norway

80
60
40
20
Sweden Switzerland UK US
80
60
40
20
Figure 2
Box plots of capital taxation rates by country, 19671992.
Whereas some countries (Austria, be unexplained temporal and cross-sectional

Denmark, the Netherlands, and New Zealand) heterogeneity. Following Garrett & Mitchell,
show little if any variation in their tax rates, the we mean-centered all observations by country
other countries show enough variation over and year, which is equivalent to allowing for
time to make a TSCS analysis of interest. There year- and country-specic intercepts.8
is also some inter-country heterogeneity, with
France, Sweden, and the United Kingdom
having generally higher rates. Figure 3
8
shows that taxation rates in some countries Of course one can only decide whether these year- and
country-specic intercepts are needed after a specication
are strongly trending whereas others show is chosen, and because these intercepts are atheoretical, one
little trend; this gure also clearly shows the should attempt to nd specications where they are not nec-
beginning of period missingness pattern in the essary. Alas, this is often impossible. Here the intercepts were
signicant in all specications. We might have preferred a
data. A regression of tax rates on time shows a model with a time trend instead of year-specic intercepts,
trend of 0.33% (with a small standard error) but the difference between the two specications was neg-
per annum in those rates. Thus, it appears ligible, and we preferred to stay consistent with Garrett &
Mitchell. Obviously in actual research such decisions should
as though a TSCS analysis of these data is be made with care, and researchers should not simply do what
sensible, and it may be the case that there will others have done before.

Australia Austria Belgium Canada

80
60
40
20
Capital Taxation Rate
Denmark Finland France Germany

80
60
40
20
Italy Japan Netherlands Norway

80
60
40
20
Sweden Switzerland UK US
80
60
40
20
1970 1975 1980 1985 1990 1970 1975 1980 1985 1990 1970 1975 1980 1985 1990 1970 1975 1980 1985 1990
Year
Figure 3
Time-series plots of capital taxation rates by country.
Garrett & Mitchell (2001) wish to explain (CDEM). Because we mean-centered all vari-
capital taxation rates (this is only one of their ables, there are no intercepts in the model.
analyses) by variables related to the economy, Table 1 reports the results of the various dy-
the demand for services, and political factors. namic estimations. All standard errors are our
We work more or less with the Garrett & recommended panel-corrected standard errors
Mitchell specication, dropping a few vari- (Beck & Katz 1995), which are easy to compute
ables that were not substantively interesting with our recommended methodology.
nor statistically signicant in any specication. The static model (not shown) is clearly in-
We thus regress the capital tax rate (CAP- adequate; a Lagrange multiplier test for se-
TAX ) on unemployment (UNEM), economic rial correlation of the errors strongly rejects
growth (GDPPC), the proportion of the pop- the null hypothesis of serially independent er-
ulation that is elderly (AGED), vulnerability of rors. Because the static model is nested in-
the workforce as measured by low wage imports side both the LDV and SC models, standard
(VULN), foreign direct investment (FDI), and Wald tests (a t-test of either H 0 : = 0 or H 0 :
two political variables: the proportion of the TAX L = 0) clearly show that the static model
cabinet portfolios held by the left (LEFT ) and can be rejected in favor of either of these two
the proportion held by Christian Democrats models.
346 Beck Katz

Table 1 Comparison of SC, LDV, ADL, and ADLLDV2 estimates of Garrett & Mitchells (2001)
model of capital taxation in 16 OECD nations, 19671992 (country and year centered)
SC LDV ADL ADLLDV2
Variable PCSE PCSE PCSE PCSE
VULN 0.22 0.12 0.10 0.07 0.28 0.13 0.33 0.14
FDI 0.51 0.26 0.37 0.21 0.59 0.26 0.48 0.28
UNEM 0.18 0.22 0.34 0.14 0.68 0.27 0.68 0.30
AGED 1.42 0.51 0.35 0.24 0.26 0.71 0.27 0.87
GDPPC 0.69 0.11 0.62 0.12 0.80 0.13 0.81 0.14
LEFT 0.004 0.012 0.006 0.009 0.003 0.013 0.002 0.014
CDEM 0.018 0.022 0.015 0.012 0.015 0.025 0.031 0.024
TAXL 0.70 0.06 0.76 0.07 0.93 0.10
VULNL 0.21 0.14 0.24 0.15
FDIL 0.55 0.29 0.56 0.31

UNEML 0.48 0.26 0.62 0.28

AGEDL 0.24 0.76 0.98 0.94
GDPPCL 0.29 0.12 0.36 0.14
LEFTL 0.005 0.013 0.004 0.014
CDEML 0.005 0.024 0.010 0.025
TAXL2 0.26 0.09
0.66
Abbreviations: ADL, autoregressive distributed lag; ADLLDV2, the ADL specication with a second lag of the dependent
variable; LDV, lagged dependent variable; PCSE, panel corrected standard errors; SC, rst-order serially correlated error
model. See text for denitions of variables in the left-hand column.
But we must compare both the LDV and lag of capital taxation to the specication; re-
SC specications to the more general ADL sults of estimating this specication are in the
specication. Again, since both these speci- ADLLDV2 columns. We cannot reject the null
cations are nested inside the ADL specication hypothesis of independent errors for this re-
we can use standard Wald tests (in this case an gression (12 = 1.1). The ADLLDV2 specica-
F-test of the null hypothesis that the coefcients tion is both statistically superior to the simpler
on all the lagged xs are zero); that null is deci- specications and shows iid errors. There are,
sively rejected, so the more general ADL spec- of course, many other specications that a sub-
ication is preferred. stantive article would test (multiple speeds of
The ADL specication still shows serial cor- adjustment, for example), but we do not pursue
relation of the errors; a Lagrange multiplier test these here.
of the null hypothesis of independent errors All the models show that a one-time unit
shows we can reject that null of iid errors. (A shock to the error process dies out exponentially
regression of the residuals from the ADL spec- (or nearly exponentially for the ADLLDV2
ication on the lagged residuals and all the other model) with similar decay rates ranging from
independent variables has an R2 of 0.046, which, 24% to 34% per annum for the rst three mod-
multiplied by the number of observations in els; for the ADLLDV2 model, the initial de-
that regression, yields a statistic of 14; since cay rate is only 7% in the rst year but in-
this statistic is distributed 12 , the null hypoth- creases to 33% (one minus the sum of the coef-
esis of independent errors is clearly rejected.) cients on the lagged dependent variable) after
As discussed in Section 2.5, we added a second the rst year. Given the standard errors on these

coefcients, the decay rates are quite similar. which results make sense, and making sense
Thus, for example, a classical condence inter- is hardly a statistical criterion. Note also that
val for the decay rate in the ADL model is (11%, AGED is itself highly trending (its autoregres-
38%), and in the ADLLDV2 model after the sion has a coefcient of 0.93 with a standard er-
rst year it is (17%, 49%). ror of 0.01). Although we can reject the null that
Turning to the estimated effect of the vari- AGED has a unit root, it, like the capital tax rate,
ous independent variables (omitting the two po- changes very slowly. Thus, we might suspect
litical variables, which show almost no effect but that the simple contemporaneous relationship
huge standard errors), recall that the SC speci- between the two variables is spurious (in the
cation assumes that the effect of the variables sense of Granger & Newbold 1974). Of course
is only instantaneous, the LDV model assumes we cannot know the truth here, but it is not
the effect decays geometrically, and the ADL obvious that the ADL (or LDV) results on the
and ADLLDV2 models allow us to test those impact of AGED are somehow foolish or wrong.
assumptions. In those latter specications, the The moral so far is that researchers should
coefcients on the current and lagged values estimate a model exible enough to account
of VULN and FDI are close in absolute value for various types of dynamics; they should also
and of opposite sign. Thus, the impact of those try hard to make sure that the error process is
variables on capital taxation rates is more or less close to iid. The ADLLDV2 model performs
only instantaneous, and the ADL coefcient es- very well here, both in terms of its passing var-
timates of this instantaneous effect are similar ious tests and its interpretability (with the sim-
to the SC estimates but different from the LDV pler ADL model being easier to interpret but
estimates. Of course, the ADL specications al- not quite passing the statistical tests). While no
low us to study the speed of adjustment, whereas specication will be good in all situations, it is
the SC specication just assumes instantaneous clear that researchers should not consider more
adjustment. general specications before accepting highly
The coefcients on UNEM and GDPPC and constrained ones such as either the SC or LDV
their lags are of opposite sign but do not quite model.
offset each other. Here the ADL estimates are, Our focus is on dynamics, but no TSCS
as we would expect, much closer to the LDV analysis is complete without a nal assessment
estimates than to the SC estimates. But again, of heterogeneity over countries. Remember
we need not make the assumptions about de- that our analysis uses country-centered data, so
cay rates that the LDV specication imposes; there can be no heterogeneity in the various
instead, we can examine what the decay pro- centered means. But we can see if the model fails
cess looks like. Interestingly, and contrary to to work for some subset of countries by cross-
Achens notion of the lagged dependent vari- validation (Stone 1974), leaving out one coun-
able dominating a regression, the coefcients try at a time. Thus we reran the ADLLDV2
of all four of these substantive coefcients are specication, leaving out one country at a time
as large as or larger than the similar coefcients and then using the estimated values to pre-
in the SC specication. dict capital tax rates in the omitted country.
The variable AGED determines tax rates in The mean absolute prediction error was then
the SC specication but fails to show any impact computed for each country. For all observa-
in any of the other specications. Intuitively, tions, the absolute forecast error was about 2.3.
although AGED perhaps ought to affect tax Four countriesJapan, Norway, Sweden, and
rates, its coefcient in the SC specication the United Kingdomhad mean absolute fore-
seems a bit large; would a one-point increase cast errors above 3.5, indicating at least some
in the aged population be expected to lead to lack of homogeneity. We do not pursue this is-
a more-than-one-point increase in capital taxa- sue further here, but clearly this issue would be
tion rates? Perhaps it is not so simple to discuss pursued in a more complete analysis. (We also
348 Beck Katz

do not present other post-estimation analyses left parties (LEFT ), the degree of central-
that should be standard, such as residual plots ized labor bargaining as a measure of corpo-
by countries.) ratism (CORP), and the product of the lat-
We also assessed heterogeneity by test- ter two variables (LEFT x CORP). The eco-
ing for parameter heterogeneity (by country). nomic and control variables are a dummy mark-
Here, since we focus on dynamics, we t the ing the relatively prosperous period through
ADL specication allowing for the coefcient 1973 (PER73), overall OECD GDP growth
of the lagged dependent variable ( ) for each (DEMAND), trade openness (TRADE), capital
country to be a random draw from a normal mobility (CAPMOB), and a measure of oil im-
distribution with zero mean. This allows us to ports (OILD). All variables, following Garretts
see whether the general speed of adjustment use of country xed effects, were mean centered
varies by country. Results of this estimation re- by country. As before, all standard errors are
veal no statistically (or substantively) signicant panel corrected.
parameter heterogeneity (on the lagged depen- GDP growth appears stationary, with an au-
dent variable); the estimated standard deviation toregressive coefcient of 0.32. Thus, all spec-
on the normal from which the coefcients were ications are expected to show relatively fast
drawn was only 0.09. dynamics, with quick returns to equilibrium.
The standard error of the estimated stan- Turning to models with explanatory variables,
dard deviation was 0.07. A test of the null hy- results of estimating various specications are
pothesis that does not vary by country yields in Table 2.
a statistic of 0.70; this statistic is 12 , so far from The static model showed modest serial cor-
the critical value for rejection of the null. We relation of the errors; a Lagrange multiplier
can look at the individual country estimates of test showed we could clearly reject the null of
. Most are within 0.01 of the overall estimate serially independent errors (12 = 8.6). Sub-
of , with only the coefcient for the United stantively, the serial correlation of the errors
Kingdom really differing; the estimated for is small (0.12), so the OLS results are similar
the United Kingdom is 0.11 under the over- to the slightly more correct results in the two
all estimate for , though with a standard error dynamic specications.
of 0.07. Given all this, we prefer not to pur- Given the rapid speed of adjustment (the co-
sue whether further investigation of the speed efcient on the LDV is 0.16), it is not surprising
of adjustment in tax rates in the United King- that all three specications show similar esti-
dom is needed, but clearly this type of analy- mates. Very few coefcients are signicant in
sis in other situations might prove extremely any of the specications, but the two variables
useful. that show a strong impact in the static speci-
cation continue to show a strong impact in the
two dynamic specications.
5.2. The Growth of Gross The similarity of the SC and LDV estimates
Domestic Product is not surprising; because of the fast dynamics,
Our second example relates to political econ- the two models are not really very different. Af-
omy explanations of the growth of GDP in 14 ter one period, the various independent vari-
OECD nations observed from 1966 through ables in the LDV specication have only 3%
1990 (yielding 336 observations), using data of their original impact; the long-run effects in
from Garrett (1998). Our treatment is cur- the LDV specication are only 18% larger than
sory for reasons of space. We use one of his the immediate impacts. Thus, the two speci-
models, taking the growth in GDP as a lin- cations are saying more or less the same things,
ear additive function of political factors and and the estimated coefcients are quite simi-
economic controls. The political variables are lar. Substantively, it appears as though GDP
the proportion of cabinet posts occupied by growth in a country is largely determined by

Table 2 Comparison of SC and LDV estimates of Garretts (1998) model of economic growth in
14 OECD nations, 19661990 (country centered)
Static SC LDV
Variable PCSE PCSE PCSE
DEMAND 0.007 0.0012 0.007 0.002 0.007 0.001
TRADE 0.018 0.019 0.021 0.021 0.019 0.019
CAPMOB 0.20 0.21 0.25 0.23 0.24 0.21
OILD 7.86 7.34 6.69 7.89 5.85 7.08
PER73 1.76 0.42 1.76 0.45 1.45 0.43
CORP 0.54 0.56 0.43 0.61 0.30 0.56
LEFT 0.075 0.17 0.076 0.18 0.062 0.17
LEFTxCORP 0.10 0.53 0.10 0.56 0.17 0.52
GDP_L 0.16 0.07
0.12
Abbreviations: LDV, lagged dependent variable; PCSE, panel corrected standard errors; SC, rst-order serially correlated
error model. See text for denitions of variables in the left-hand column.
GDP growth in its trading partners, and poli- namics, this will always be the case when vari-
tics appears to play little if any role. ables adjust quickly.
Both specications were tested against the
full ADL specication that contained all the
one-year lags of the independent variables. 6. CONCLUSION
Standard hypothesis tests do not come close to There is no cookbook for modeling the dy-
rejection of the simpler SC or LDV models in namics of TSCS models. Instead, careful ex-
favor of the ADL model. Since the LDV and amination of the specications, and what they
AIC specications are not nested, discriminat- entail substantively, can allow TSCS analysts
ing between them is not so simple. Because both to think about how to model these dynamics.
specications have the same number of parame- Well-known econometric tests help in this pro-
ters, discrimination using standard information cess, and standard methods make it easy to es-
criteria (AIC or BIC) simplies to comparisons timate the appropriate dynamic model. Mod-
of goodness of t, on which criterion both spec- eling decisions are less critical where variables
ications perform almost equally well. This will equilibrate quickly; as the adjustment process
often be the case, since both the SC and LDV slows, the various specications imply more and
specications imply very quick adjustments to more different characteristics of the data. An-
equilibrium when the dynamic parameters are alysts should take advantage of this to choose
near zero. the appropriate model, namely, one that implies
In summary, the data are consistent with dynamics consistent with theoretical concerns.
very short-run impacts, and it does not much The specication chosen should of course be
matter how we exactly specify those dynam- exible enough to allow for testing against al-
ics. In terms of the Achen critique, there are ternative dynamic specications.
two predictors of GDP that are strong in the Being more specic, we have provided evi-
SC model; they remain about equally strong in dence that, contra Achen, there is nothing per-
the LDV model. As we have argued, there is nicious in the use of a model with a lagged de-
nothing about lagged dependent variables that pendent variable. Obviously attention to issues
makes them dominate a regression or makes of testing and specication are as important
real effects disappear. Given the nature of dy- here as anywhere, but there is nothing about
350 Beck Katz

lagged dependent variables that makes them ical and practical advantages to this. The theo-
generically harmful. As we have seen, there is retical advantage is that dynamic issues become
a variety of generic dynamic specications, and much more than nuisances for estimation. The
researchers should choose among them by us- practical advantage is that it is easy to estimate
ing the same general methodology they use models with (approximately) independent er-
in other cases. The ADL model (or its ADL2 ror processes via OLS, and easy then to esti-
complication) is a good place to start; at that mate these models with additional complicating
point, various specializations of the model can cross-sectional features.
be tested against this general specication. An- There are many important features of TSCS
alysts, should, of course, interpret the dynamic data, both in the temporal and spatial realms.
results in substantive terms, focusing on long- Both sets of features are interesting, and nei-
as well as short-run effects. ther should be swept under the rug. Fortu-
Instead of pushing dynamics into a compli- nately, the econometrics involved with good
cated error process that then must be xed TSCS modeling are not difcult, and a clear eye
up to allow for estimation, it is much better to on specication and testing allows researchers
model the dynamics directlythat is, in terms to nd substantively interesting features of the
of observable variables. There are both theoret- data.
DISCLOSURE STATEMENT
The authors are not aware of any afliations, memberships, funding, or nancial holdings that
might be perceived as affecting the objectivity of this review.
ACKNOWLEDGMENTS
For research support we thank the National Science Foundation. An earlier version was presented
at the Annual Meeting of the Society for Political Methodology, Stanford University, Stanford,
California, July 2931, 2004. We thank Geoff Garrett, Evelyne Huber, and John Stephens for
providing data. We also thank the many colleagues who have discussed TSCS issues with us and
allowed us to present in various forums, among whom Chris Achen and Simon Jackman (for efforts
well beyond the call of duty) must be singled out.
LITERATURE CITED
Achen C. 2000. Why lagged dependent variables can suppress the explanatory power of other independent variables.
Presented at Annu. Meet. Soc. Polit. Methodol., Jul. 2022, Univ. Calif. Los Angeles
Adolph C, Butler DM, Wilson SE. 2005. Which time-series cross-section estimator should I use now? Guidance from
Monte Carlo experiments. Presented at Annu. Meet. Am. Polit. Sci. Assoc., Sep. 14, Washington, DC
Alvarez RM, Katz JN. 2000. Aggregation and dynamics of survey responses: the case of presidential approval. Soc.
Sci. Work. Pap. 1103, Div. Hum. Soc. Sci., Calif. Inst. Technol.
Anderson TW, Hsiao C. 1982. Formulation and estimation of dynamic models using panel data. J. Econ.
18:4782
Beck N. 1985. Estimating dynamic models is not merely a matter of technique. Polit. Methodol. 11:7190
Beck N. 1991. Comparing dynamic specications: the case of presidential approval. Polit. Anal. 3:5187
Beck N. 1992. The methodology of cointegration. Polit. Anal. 4:23747
Beck N. 2001. Time-seriescross-section data: What have we learned in the past few years? Annu. Rev. Polit.
Sci. 4:27193
Beck N, Gleditsch KS, Beardsley K. 2006. Space is more than geography: using spatial econometrics in the
study of political economy. Int. Stud. Q. 50:2744

Beck N, Katz JN. 1995. What to do (and not to do) with time-series cross-section data. Am. Polit. Sci. Rev.
89:63447
Beck N, Katz JN. 1996. Nuisance versus substance: specifying and estimating time-seriescross-section mod-
els. Polit. Anal. 6:136
Beck N, Katz JN. 2007. Random coefcient models for time-seriescross-section data: Monte Carlo experi-
ments. Polit. Anal. 15:18295
Beck N, Katz JN. 2009. Modeling dynamics in time-seriescross-section political economy data. Soc. Sci. Work.
Pap. 1304, Div. Hum. Soc. Sci., Calif. Inst. Technol.
Beck N, Katz JN, Tucker R. 1998. Taking time seriously: time series cross section analysis with a binary
dependent variable. Am. J. Polit. Sci. 42:126088
Davidson JH, Hendry DF, Srba F, Yeo S. 1978. Econometric modelling of the aggregate time-series relation-
ship between consumers expenditures and income in the United Kingdom. Econ. J. 88:66192
De Boef S, Keele L. 2008. Taking time seriously: dynamic regression. Am. J. Polit. Sci. 52:184200
Engle R, Granger CW. 1987. Co-integration and error correction: representation, estimation and testing.
Econometrica 55:25176
Franzese RJ, Hayes JC. 2007. Spatial econometric models of cross-sectional interdependence in political
science panel and time-series-cross-section data. Polit. Anal. 15:14064
Garrett G. 1998. Partisan Politics in the Global Economy. New York: Cambridge Univ. Press
Garrett G, Mitchell D. 2001. Globalization, government spending and taxation in the OECD. Eur. J. Polit.
Res. 39:14477
Granger CW, Newbold P. 1974. Spurious regressions in econometrics. J. Econometrics 2:11120
Greene W. 2008. Econometric Analysis. Upper Saddle River, NJ: Pearson Prentice Hall. 6th ed.
Hamilton JD. 1994. Time Series Analysis. Princeton, NJ: Princeton Univ. Press
Hendry D, Mizon G. 1978. Serial correlation as a convenient simplication, not a nuisance: a comment on a
study of the demand for money by the Bank of England. Econ. J. 88:54963
Hibbs D. 1974. Problems of statistical estimation and causal inference in time-series regression models. In
Sociological Methodology 19731974, ed. H Costner, pp. 252308. San Francisco: Jossey-Bass
Honaker J, King G. 2010. What to do about missing values in time series cross-section data. Am. J. Polit. Sci.
54:56181
Huber E, Stephens JD. 2001. Development and Crisis of the Welfare State. Chicago: Univ. Chicago Press
Huber PJ. 1967. The behavior of maximum likelihood estimates under non-standard conditions. In Proc. Fifth
Annu. Berkeley Symp. Math. Stat. Probab., ed. LM LeCam, J Neyman, I:22133. Berkeley: Univ. Calif.
Press
Im KS, Pesaran MH, Shin Y. 2003. Testing for unit roots in heterogeneous panels. J. Econometrics 115:5374
Judson KA, Owen AL. 1999. Estimating dynamic panel data models: a guide for macroeconomists. Econ. Lett.
65:915
Keele L, Kelly NJ. 2006. Dynamic models for dynamic theories: the ins and outs of lagged dependent variables.
Polit. Anal. 14:186205
Kiviet JF. 1995. On bias, inconsistency, and efciency of various estimators in dynamic panel models.
J. Econometrics 68:5378
Levin A, Lin C-F, Chu C-SJ. 2002. Unit root tests in panel data: asymptotic and nite-sample properties.
J. Econometrics 108:124
Little RJA, Rubin DB. 1987. Statistical Analysis with Missing Data. New York: Wiley
Mizon G. 1984. The encompassing approach in econometrics. In Econometrics and Quantitative Economics, ed.
D Hendry, K Wallis, pp. 13572. Oxford, UK: Blackwell
Newey WK, West KD. 1987. A simple positive semi-denite heteroskedasticity and autocorrelation consistent
covariance matrix estimator. Econometrica 55:7038
Nickell S. 1981. Biases in dynamic models with xed effects. Econometrica 49:141726
Stone M. 1974. Crossvalidatory choice and assessment of statistical prediction. J. R. Stat. Soc. Ser. B 36:11133
352 Beck Katz

PL14-frontmatter ARI 11 April 2011 17:22
Annual Review of
Political Science
Contents Volume 14, 2011
A Life in Political Science

Sidney Verba p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p i
Leadership: What It Means, What It Does, and What We Want to
Know About It
John S. Ahlquist and Margaret Levi p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p 1

Examining the Electoral Connection Across Time
Jamie L. Carson and Jeffery A. Jenkins p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p25
Presidential Appointments and Personnel
David E. Lewis p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p47
Understanding the 20072008 Global Financial Crisis: Lessons for
Scholars of International Political Economy
Eric Helleiner p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p67
Presidential Power in War
William G. Howell p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p89
The Politics of Regulation: From New Institutionalism to New
Governance
Christopher Carrigan and Cary Coglianese p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p 107
The New Judicial Politics of Legal Doctrine
Jeffrey R. Lax p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p 131
The Rhetoric Revival in Political Theory
Bryan Garsten p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p 159
The Rhetoric of the Economy and the Polity
Deirdre Nansen McCloskey p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p 181
The Contribution of Behavioral Economics to Political Science
Rick K. Wilson p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p 201
The Causes of Nuclear Weapons Proliferation
Scott D. Sagan p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p 225
Network Analysis and Political Science
Michael D. Ward, Katherine Stovel, and Audrey Sacks p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p 245
v
PL14-frontmatter ARI 11 April 2011 17:22
The Big Five Personality Traits in the Political Arena

Alan S. Gerber, Gregory A. Huber, David Doherty, and Conor M. Dowling p p p p p p p p p p 265
Clientelism
Allen Hicken p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p 289
Political Economy Models of Elections
Torun Dewan and Kenneth A. Shepsle p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p 311
Modeling Dynamics in Time-SeriesCross-Section Political
Economy Data
Nathaniel Beck and Jonathan N. Katz p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p 331
Voting Technologies
Charles Stewart III p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p 353
Indexes
Cumulative Index of Contributing Authors, Volumes 1014 p p p p p p p p p p p p p p p p p p p p p p p p p p p 379

Cumulative Index of Chapter Titles, Volumes 1014 p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p 381
Errata
An online log of corrections to Annual Review of Political Science articles may be found
at http://polisci.annualreviews.org/
vi Contents

Beck y Katz 2011

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Beck y Katz 2011

Uploaded by

Copyright:

Available Formats

PL14CH16-Beck ARI 14 April 2011 17:11

Our comprehensive search

Annu. Rev. Polit. Sci. 2011. 14:33152 Keywords

1. INTRODUCTION assumed that the dynamics apply similarly to all

332 Beck Katz

rst-order lags, each observation with missing

www.annualreviews.org Time-SeriesCross-Section Data 333

1.4. Roadmap their effect dissipates instantaneously; there are

with the obvious generalization for higher-

334 Beck Katz

lows for testing between the various models.

www.annualreviews.org Time-SeriesCross-Section Data 335

336 Beck Katz

equilibration, with the different parameter con- 2.3. Higher-Order Processes

www.annualreviews.org Time-SeriesCross-Section Data 337

error processes sufce, and it would be unusual 2.4. More Complicated

338 Beck Katz

www.annualreviews.org Time-SeriesCross-Section Data 339

340 Beck Katz

difference is because both models constrain the

www.annualreviews.org Time-SeriesCross-Section Data 341

342 Beck Katz

www.annualreviews.org Time-SeriesCross-Section Data 343

to be integrated. Other political variables, such

344 Beck Katz

Australia Austria Belgium Canada

Denmark Finland France Germany

Italy Japan Netherlands Norway

Whereas some countries (Austria, be unexplained temporal and cross-sectional

www.annualreviews.org Time-SeriesCross-Section Data 345

Australia Austria Belgium Canada

Denmark Finland France Germany

Italy Japan Netherlands Norway

346 Beck Katz

FDIL 0.55 0.29 0.56 0.31

UNEML 0.48 0.26 0.62 0.28

www.annualreviews.org Time-SeriesCross-Section Data 347

348 Beck Katz

www.annualreviews.org Time-SeriesCross-Section Data 349

350 Beck Katz

www.annualreviews.org Time-SeriesCross-Section Data 351

352 Beck Katz

Contents Volume 14, 2011

A Life in Political Science

John S. Ahlquist and Margaret Levi p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p 1

The Big Five Personality Traits in the Political Arena

Cumulative Index of Contributing Authors, Volumes 1014 p p p p p p p p p p p p p p p p p p p p p p p p p p p 379

You might also like