Beck and Katz Series

Taking Time Seriously: Time-Series-Cross-Section Analysis with a Binary Dependent Variable
Author(s): Nathaniel Beck, Jonathan N. Katz and Richard Tucker

Source: American Journal of Political Science, Vol. 42, No. 4 (Oct., 1998), pp. 1260-1288
Published by: Midwest Political Science Association
Stable URL: http://www.jstor.org/stable/2991857 .
Accessed: 31/07/2014 10:24
Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at .
http://www.jstor.org/page/info/about/policies/terms.jsp
.
JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of
content in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms
of scholarship. For more information about JSTOR, please contact [email protected].
.
Midwest Political Science Association is collaborating with JSTOR to digitize, preserve and extend access to
American Journal of Political Science.
http://www.jstor.org
This content downloaded from 150.164.96.5 on Thu, 31 Jul 2014 10:24:37 AM
All use subject to JSTOR Terms and Conditions
Taking Time Seriously:
Time-Series-Cross-Section Analysis
with a Binary Dependent Variable
Nathaniel Beck, University of California, San Diego
Jonathan N. Katz, University of Chicago
Richard Tucker, Harvard University
Researchers typically analyze time-series-cross-section data with a binary dependent
variable (BTSCS) using ordinary logit or probit. However, BTSCS observations are
likely to violate the independence assumption of the ordinary logit or probit statistical
model. It is well known that if the observations are temporally related that the results of
an ordinary logit or probit analysis may be misleading. In this paper, we provide a simple
diagnostic for temporal dependence and a simple remedy. Our remedy is based on the
idea that BTSCS data are identical to grouped duration data. This remedy does not re-
quire the BTSCS analyst to acquire any further methodological skills, and it can be eas-
ily implemented in any standard statistical software package. While our approach is suit-
able for any type of BTSCS data, we provide examples and applications from the field of
Intemnational Relations, where BTSCS data are frequently used. We use our methodology
to reassess Oneal and Russett's (1997) findings regarding the relationship between eco-
nomic interdependence, democracy, and peace. Our analyses show that (1) their finding
that economic interdependence is associated with peace is an artifact of their failure to
account for temporal dependence yet (2) their finding that democracy inhibits conflict is
upheld even taking duration dependence into account.
1. INTRODUCTION
The analysis of time-series-cross-section data with a binary dependent
variable (BTSCS data) is becoming more common, particularly in the study
of international relations (IR). Moreover, the number of such studies appears
to be increasing exponentially.' Since it is unlikely that units are statistically
unrelated over time, BTSCS observations, like their continuous dependent
We thank John Oneal and Bruce Russett for providing their data and Robert Engle, Gary King,
Jonathan Nagler and Glenn Sueyoshi for helpful comments and conversations. A replication data set
may be found at ftp://weber.ucsd.edu:/pub/nbeck
'The vast majority of IR BTSCS analysts study militarized conflict or interstate war; others
study alliance and rivalry behavior. A brief list of IR BTSCS analyses using ordinary logit or probit
published in the previous eighteen months includes Barbieri (1996), Bennett (1996), Enterline
(1996, 1997), Farber and Gowa (1997), Gartzke (Nd.), Gleditsch and Hegre (1997), Henderson
(1997), Hermann and Kegley (1996), Ijuth (1996), Lemke and Reed (1996), Mansfield and Snyder
(1996, 1997), Maoz (1996, Nd.), Mousseau (1997), Oneal et al. (1996) and Oneal and Russett
(1997). We do not claim that these studies draw incorrect conclusions. However, the possibly faulty
(and untested) assumption of temporal independence, inherent in their respective logit/probit analy-
ses, casts some doubt about the validity of their substantive findings.
American Journal of Political Science, Vol. 42, No. 4, October 1998, Pp. 1260-1288 C 1998 by the
Board of Regents of the University of Wisconsin System
BINARY TIME-SERIES-CROSS-SECTION ANALYSIS 1261
variable TSCS cousins, are likely to be temporally dependent. It is well
known that violations of the assumption of independent observations can re-
sult in overly optimistic inferences (underestimates of variability leading to
inflated t-values). Nevertheless, BTSCS data are almost invariably analyzed
using ordinary logit or probit analysis, techniques that assume temporal in-
dependence.2 While analysts are certainly aware of the pitfalls of such as-
sumptions, they seem to have overlooked a very simple solution.
Our simple solution is to add a series of dummy variables to the logit
specification. These variables mark the number of periods (usually years)
since either the start of the sample period or the previous occurrence of an
"event" (such as war). A standard statistical test on whether these dummy
vanables belong in the specification is a test of whether the observations are
temporally independent. The addition of these dummy variables to the speci-
fication, if the test indicates they are needed, corrects for temporally depen-
dent observations. This simple solution, which can be implemented in any
software package, allows for accurate estimation of the parameters of tem-
porally dependent BTSCS models.3
This simple solution is based on the recognition that BTSCS data are
grouped duration data. Note that we do not say "like grouped duration data":
BTSCS data are grouped duration data. This recognition permits us to use
known and validated event history concepts explicitly
designed
for tempo-
rally dependent data.4
In the next section, we briefly discuss the prominence of BTSCS data in
international relations and why ordinary logit is inappropriate for BTSCS data
in most contexts. The subsequent section illustrates the equivalence of BTSCS
and grouped duration data. We also delineate our proposed method for ana-
lyzing temporally dependent BTSCS data and discuss its application to the
study of conflict/peace. The next section then uses our proposed method to
reanalyze one prominent study of conflict (Oneal and Russett 1997).
2. BTSCS DATA IN INTERNATIONAL RELATIONS
BTSCS data are most common in international relations.5 The IR
conflict processes literature has favored a theoretical emphasis on dyadic
2We freely mix logit and probit analyses here. In the context of this paper they suffer identi-
cal flaws which have identical remedies. For simplicity we refer to logit analysis throughout this
paper. Those committed to probit analysis should make our recommended changes to the probit
specification.
3While cross-sectional dependence also causes problems, our goal here is to address the prob-
lem of temporal dependence. Our proposed remedy is, however, sufficiently simple that it should be
easy to adjoin to any remedy for cross-sectional dependence.
4We use the terms event history methods and duration models interchangeably.
5Following the pioneering efforts of Berry and Berry (1990, 1992), American state politics re-
searchers also frequently use BTSCS data. Unlike IR researchers, however, they typically begin
with
1262 N. Beck, J. Katz, and R. Tucker
interstate interactions (e.g., Bueno de Mesquita and Lalman 1992; Goertz
and Diehl 1993; Vasquez 1993) and an empirical focus on the dyad-year as
the unit-of-analysis (e.g. Bremer 1992; Maoz and Russett 1993). Dyad-year
data sets typically contain yearly observations on conflict occurrence be-
tween pairs of nations (or engagement in some other interstate behavior
such as alliance formation or rivalry dissolution). These datasets also
include properties of the dyad (which may vary from year to year) to ex-
plain the presence or absence of conflict. While our argument generalizes
to all BTSCS data, we couch our discussion in terms of IR dyad-year
BTSCS data.
BTSCS data shares all the standard characteristics of continuous depen-
dent variable time-series-cross-section data.6 Formally, a BTSCS model
with binary dependent variable, y, and a vector of independent variables, x,
has
P(Yi,t
=
1)
=
t
(xi,t, Yi,l, *v, Yi,t-l, xi,lg, *,
xi,t-1),
i
=
1l,...,9Ng,t
=
1,
...,9T (1)
wheref
is any suitable function that has a range of the unit interval. The in-
clusion of the lagged values of y and x allows for a very general form of tem-
poral dependence of the observations.7 We assume the number of time
points (1) to be reasonably large (say at least 20). This is in contrast to bi-
nary panel data, where T may be as small as two or three. Panel methods are
also designed to handle enormous cross-section sample sizes (N), ranging
into the thousands. While N is not critical for our interests here, we do not
have to solve the problems brought about by large (and asymptotically un-
bounded) N's combined with small (and bounded) T's that have plagued
a discrete time event history model, which is then appropriately estimated using BTSCS methods.
Meier and McFarland (1992) and Mintrom (1997), for example, correctly allow for duration depen-
dence by adding yearly temporal dummy variables to the logit specification. Typical state politics
BTSCS datasets, however, are simpler than their IR counterparts; while IR datasets often contain
multiple failures per unit, state datasets typically only have a single failure per unit (i.e., no post-fail-
ure observations). The latter datasets also typically track all states for the same time frame. Thus
state researchers can use temporal dummy variables that correspond to years. IR researchers, as we
will subsequently illustrate, must create temporal dummy variables that track time since the previous
event occurrence.
6See Beck and Katz (1995, 1996) for a discussion of continuous dependent variable TSCS
methods.
7Equation 1 is very general. One possible specialization is a latent variable formulation, where
temporal dependence is induced by serially correlated errors in the latent variable (Beck and Katz
1997). Equation 1 does not imply that one should add a lagged dependent variable to the logit speci-
fication. The essential nonlinearity of BTSCS models makes their dynamics much more complex
than continuous TSCS models.
panel analysts. This contrast is important, since there are available estima-
tion techniques for interdependent binary panel data (see Diggle, Liang, and
Zeger 1994). While some of these techniques may prove useful for interde-
pendent BTSCS data, such utility has not yet been demonstrated. But in gen-
eral, the temporal dimension of BTSCS data are so much richer than its
panel counterpart that we would not be overly optimistic about the utility of
panel methods for BTSCS data.8
Analysts almost invariably simplify Equation 1 to
p
=Yi,t li
=
1
+
e-)i
(2)
and perform an "ordinary logit" analysis of their data.
BTSCS data, however, are simply a variant of TSCS data, and we know
that TSCS data often shows temporal dependence. Might we not expect
BTSCS data to show temporal dependence as well? The probability of dy-
adic conflict in a given year, for example, is likely to be dependent on the
conflict history of that dyad.9 Remedies for continuous dependent variable
TSCS data (Beck and Katz 1996), however, are inapplicable to BTSCS data
(Beck and Tucker 1997). It is well known that if the observations are tempo-
rally related, results of an ordinary logit or probit analysis may be mislead-
ing. Poirier and Ruud demonstrate that probitl0 standard errors are incorrect
for time series data with serially correlated errors. These time series results
hold for BTSCS data. Simulations reported in Beck and Katz (1997) indi-
cate the severity of these problems, with reported standard errors possibly
understating variability by 50 percent or more! While probit analysis of tem-
porally dependent data provides consistent parameter estimates, ignoring
this dependence may also lead to severe inefficiency. The incorrect assump-
tion of temporal independence leads to both inaccurate statistical tests and
the loss of valuable information in the data.
IR BTSCS analysts routinely acknowledge these problems, but in the
absence of better alternatives, continue to ignore temporal dependence and
use ordinary logit analysis. Farber and Gowa (1997, 397), for example,
agree that "the yearly observations for a dyad cannot be considered to be in-
dependent" but they "proceed ignoring this lack of independence. While
8IR datasets may have large N's; the one we reexamine has an N of almost 1000. The critical
issue is that IR datasets typically have reasonably large T. Most dyads in our reanalysis, for example,
are observed for over twenty years, with more observations per dyad becoming available as more
recent data are collected. Our proposed method would not work for datasets with very small Ps.
9As we discussed below, temporal dependence cannot provide a satisfactory explanation by it-
self, but must, instead, be the consequence of some important, but unobserved, variable.
I0Their conclusions hold for logit and any other standard binary dependent variable method.
[they] recognize that the power of [their] tests is somewhat overstated as a
result, a better solution is not obvious." Oneal and Russett (1997, 283) note
that the "greatest danger arises from autocorrelation, but that there are not
yet generally accepted means of testing for or correcting this problem in lo-
gistic regressions."11 Some BTSCS analysts have simply given up on logit
based methods, opting for less well-known event history methods. Bennett
(1997, 12), for example, argues that a "hazard [event history] model is the
most appropriate way to analyze alliance durations, and superior to the [or-
dinary logit] procedure, since hazard models allow corrections for censor-
ing, heterogeneity and duration dependence."
In this paper we will show that the logit, once corrected, is an event his-
tory method for BTSCS data. Moreover, we illustrate a simple and easy to
implement modification to the logit specification that allows it to handle
temporally dependent data. Thus our methodology allows logit oriented
BTSCS analysts to continue to use their familiar methods while deriving all
the benefits of event history analysis.
12
3. BTSCS DATA Is GROUPED DuRATIoN DATA
Our solution depends on the recognition that BTSCS data are identical to
grouped duration data. While we need very little of the specialized language
of event history analysis, a few concepts will prove helpful.13 Event history
analysts model the elapsed time until an "event" or "failure," or, equiva-
lently, the length of a non-eventful "spell." In our IR examples an event is
conflict, with the duration of spells of peace being modeled. A unit has "sur-
vived" or is "at risk" until it fails.14 The "hazard" rate is, loosely speaking,
an indication of how likely failure is to occur at any given time (or more pre-
cisely, the rate of failure in any small time interval), provided the unit has sur-
vived until that time. If the hazard rate is time invariant, that is, the risk of
failure does not depend on how long a unit has survived, the hazard is said to
be "duration independent"; if it varies with time, the hazard rate is said to
show "duration dependence.
"
Event history analysts model the hazard rate as
a function of independent variables, which may or may not be time invariant.
The most common event history methods assume continuous time, so
that durations are measured continuously and hazard rates vary continu-
"They attempt various ad hoc remedies, which we discuss in the reanalysis section.
'20ur only objection to Bennett's approach of using standard event history methods is that it
requires analysts to learn an entirely new methodology
13Introductions to event history methods for political scientists are in Beck (N.d.) and Box-
Steffensmeier and Jones (1997).
14For simplicity, we initially assume only one possible failure per unit. We relax this assump-
tion below.
ously. But duration data may be "grouped," so that we only know whether a
unit has failed in some discrete time interval (with independent variables
only measured to the fineness of that interval). This is usually a result of the
measurement process, so that instead of recording the exact time of failure,
we only record whether a unit failed in some fixed time interval. BTSCS
data, as coded, only allows us to know if a conflict occurred sometime dur-
ing a year 15
Annual BTSCS data are equivalent to grouped duration data with an ob-
servation interval of one year. 16 The dichotomous dependent variable is one
in a given year if there was a failure (for example, conflict) during that year,
with the independent variables also being measured yearly. 17 We stress that
BTSCS data are, by definition, grouped event history data; no sophisticated
mathematical, statistical nor computational argument is required to demon-
strate this.
3.1 The Grouped Duration Solution
Having noticed the equivalence, we also note that there are standard
methods for estimating models with grouped event history data where the ob-
servations may be temporally dependent. These methods begin with a con-
tinuous time event history model. They are derived under the assumption that
observations of this continuous process are only made at discrete intervals,
'5The beginning and end of conflicts can obviously be determined more accurately, often to the
day. Raknerud and Hegre (1997), for example, use the daily dating of wars to convert a BTSCS data
set into a continuous time event history data set (since they are interested in the order in which na-
tions join multilateral conflicts). But while it may be possible to more accurately date events, many
independent variables are only measured yearly. Our interest is in the use of event history methods
to analyze data that has already been coded as BTSCS data.
16Some analysts prefer to use the term discrete time duration data rather than grouped duration
data. BTSCS data, however, are grouped, ftot discrete time data. Grouped duration data allows for
exits at any time, but we only observe whether an exit has occurred in some time interval. Exits,
within the discrete time framework, only occur at discrete time intervals, We do not contend that
wars only occur on New Year's Eve! But this distinction has few, if any, practical implications, since
discrete time models are analyzed using grouped time concepts.
17Note that we are assuming there can be no more than one measured conflict in a year. This
may be due to a censoring process, where the only recorded information is whether at least one con-
flict occurred in a year, or it may be due to something about the conflict process which limits con-
flicts to one per year. BTSCS data are presented this way. Analysts may have a choice as to whether
to use a binary dependent variable or an event count dependent variable; our discussion assumes that
either the investigator or some outside data collector has previously decided to only collect informa-
tion about the binary dependent variable. Alt, King, and Signorino (1997) provide a very interesting
treatment of this entire issue. Our point here is much simpler than theirs, since we assume that ag-
gregation decisions have already been made, and so only BTSCS data are available. Event count
TSCS models must also take duration dependence into account.
with only one event possible per interval.18 The most common continuous
time duration model is the Cox (1975) proportional hazards model; this
model dominates applied work in the social and life sciences. 19
In this model the instantaneous hazard rate is
h(sxi s
) = ho(s)exi
s (3)
where x s is the vector of independent variables at (continuously measured)
time s. In this setup the hazard of exit depends both on the independent vari-
ables (via the eXi,sI term) and the length of time the unit has been at risk (via
ho(s),
the "baseline hazard"). The proportional hazards model is widely
used because it allows for estimation of the parameters of interest (,3) in the
presence of an unknown, and possibly complicated, time varying baseline
hazard.20 As we shall see, the ,3 in Equation 3 are what logit BTSCS model-
ers are estimating. Ordinary logit fails because it doesn't allow for a
(nonconstant) baseline hazard. The grouped duration model, although de-
rived from an underlying continuous time Cox proportional hazards model,
is easier to estimate, and does not suffer from some problems inherent in the
continuous time model.21 For notational simplicity, let us assume annual
data indexed by year t. The "discrete hazard" in year t for dyad i is simply
the probability that a dyad will experience conflict sometime during that
year. Letting
yit
be a binary indicator of conflict in dyad i sometime in year
'8The grouped duration model was first derived by Prentice and Gloeckler (1978). Readable
social science treatments are in Allison (1982), Singer and Willett (1993) and Jenkins (1995). For
completeness we lay out the basic argument in an appendix, although it is dependent on some dura-
tion results not contained in this paper. Han and Hausman (1990) and Sueyoshi (1995) provide a
modem econometric treatment of many of the issues discussed here. Katz and Sala (1996) have ap-
plied the grouped duration model to Congressional data.
'9The assumption of proportional hazards is not innocuous, and surely there are situations
where it is a bad assumption. For example, we implicitly assume that there is no heterogenity in the
baseline hazards across units, so that we may pool all observations. But Equation 3 is more general
than other common hazard specifications used in event history analysis. The Weibull model is the
most common fully parametric event history model. The Weibull model uses a hazard rate which is
a special case of Equation 3, with
ho(t)
assumed to follow a specific parametric form. In practice, the
proportional hazards model works well, but no one model is perfect for all situations. Since, as we
shall see, ordinary logit can be derived from a special case of the proportional hazards model, any
criticism of proportional hazards is at least as strong a criticism of ordinary logit.
20The grouped model could easily be adapted to fully parametric duration models. Given the
dominance of the semi-parametric Cox approach in applied work, we see no reason to pursue the
fully parametric approach here. Alt, King, and Signorino (1997) derive the grouped model for a con-
tinuous time gamma duration model.
2lIn particular, the continuous time model has problems if there are many units that exit at the
same time.
t, the discrete hazard is just
P(yit
=1). This is the probability estimated by
logit analysis. But the logit probability (Equation 2) is not the same as the
discrete hazard constructed by aggregating the continuous hazard rate of
Equation 3. The discrete hazard rate corresponding to Equation 3 is (as
shown in the Appendix)
P(yi,t = l|xi,t =
h(txi,t) = 1 - exp(-exi`+Kt-to) (4)
where
xi t
now represents the observed value of the independent variable for
the entire year t.
Kt_to
is a dummy variable marking the length of the se-
quence of zeros that precede the current observation; for first events, tO = 0.
We use t -
tO instead of the simpler t subscript because the notation must al-
low for multiple events; in that case tO marks the time of the previous event
and t - tO is the length of the spell of peace from tO until t. 22 We use the
more complicated notation even when tO
= 0 to remind us that the temporal
dummies mark the length of prior spells of peace, which will not always be
the current year index, t.23
3.2 The Logit Solution
The grouped duration model differs from ordinary logit in two ways.
First, it is a binary-dependent variable model using what is known as a
"complementary log-log (cloglog) link" instead of the more familiar logit
(or probit) link.24 Second, the specification contains the temporal dummy
variables,
Kt-to.
The distinction between the cloglog and logit links is trivial;
the inclusion of the temporal dummy variables is not. Let us eliminate the
trivia first.
22To be concrete we show one particular assignment of the i.
t I 1 2 3 4 5 6 7 8 9
y I O 0 0 1 0 1 1 0 0
K I KI
IK32
K3 K4
K
1
K2 K1 K K2
As with any saturated set of dummy variables we must either not estimate a constant term or drop
one dummy variable. For notational simplicity we assume the former, though most statistical pack-
ages do the latter. This should cause no problems.
23We should note that in constructing the discrete hazard rate in Equation 4 we have implicitly
assumed that the unit has "survived" to period t. That is, we should condition on the event that
yit-l
= 0 in the definition of the discrete hazard. Since we will allow for multiple events, we drop this
added notation. This assumption, however, is important. In particular, for the case of multiple fail-
ures (conflicts) this means that we are treating the second and subsequent events as if they were new
units in the data.
24This terminology is from the "generalized linear model" (GLM approach [McCullagh and
Nelder 1989]). A link function specifies the relationship between a linear predictor
(xifi)
and the de-
pendent variable. The logit and cloglog links are two common links for binary dependent variable
models.
The cloglog vs. logit link
The two links transform probabilities by
cloglog(P) = log(-log(l - P)) and (5)
logit(P)
=
log (j
PIJ
(6)
These are the inverses of the transforms used in Equations 4 and 7 and plot-
ted in Figure 1.
We see that the two links are almost identical when the probability of an
event is less than 25 percent, and are extremely similar so long as the prob-
ability of an event does not exceed 50 percent. If the probabilities of an event
are small, either a logit or a cloglog link can be used. For typical event his-
tory data (especially in IR) the probability of an event in any given time pe-
riod will be small. The two links will differ only in the unlikely case (for
event history data) that many observations have a probability of failure ex-
ceeding 50 percent. And even here, who is to say which is the best model to
use? While the cloglog model is the exact grouped duration analogue of the
most widely used Cox proportional hazards model, it may not be appropri-
ate in every instance. The logit link corresponds to a (complicated) continu-
ous time duration model (Sueyoshi 1995). While the Cox proportional haz-
ards model is computationally convenient, there is no reason to assume that
the data were generated in order to make computations simple! There ap-
pears to be little if any cost then to use the more familiar logit link for typi-
cal BTSCS data. There are clear benefits to using the logit link. It is well
understood by researchers, is estimable with any software package, does not
require learning new methods (generalized linear models), and most impor-
tantly, can be extended easily in a variety of interesting ways.25 We, there-
fore, recommend that researchers use
P(Y"
=
=
h(tlxit=
) =t 1 +
e(Xi',tV+'tto)
which is the logistic analogue of Equation 4.
Temporal dummy variables
Using the logit rather than the cloglog link allows us to focus on the sec-
ond way that Equation 7 differs from ordinary logit: the inclusion of the
25Allison (1982, 87-90) and Farhmier and Wagenpfeil (1996) show how the logit can be ex-
tended to the multinomial logit to handle multiple types of failures. Thus we could allow for data
where
Yi
t denotes a series of unordered outcomes, so long as the outcomes satisfy the independent
risks assumption underlying the "competing risks" model. Any remedies which allow logit to deal
with cross-sectional dependence will also be easy to combine with the logit link.
Figure 1. Comparison of cloglog and logit Transforms
0.25
- Spline
..
Dummies
0.20-
0.15-
0.10-
0.05-
0.0-
0 5 10 15 20 25 30 35
Duration of Peace
temporal dummies,
Ktt0.
These are the grouped duration analogue of the
continuous time baseline hazard function,
ho.
Omitting these dummies is
equivalent to assuming that the baseline hazard is constant, so that the model
shows duration independence. While such a situation can occur, event his-
tory analysts typically allow for duration dependence, at least initially, and
then test whether the model can be simplified by imposing duration indepen-
dence. The costs of incorrectly imposing duration dependence are, at a mini-
mum, inefficiency and incorrect standard errors, and in some complicated
cases may even lead to inconsistent parameter estimates. It is exactly these
problems that the Cox proportional hazard model avoids. It is simple enough
to include the temporal dummies in the logit specification. Before doing so,
however, one should determine whether they are required. Temporal dum-
mies should not be included in the specification if the observations are al-
ready temporally independent, since the temporal dummies might then in-
troduce unnecessary multicollinearity. The test of whether the temporal
dummies should be included is a standard likelihood ratio test of the hypoth-
esis that all the
Kt-to
=
0. If the null hypothesis of temporal independence is
rejected, then all the
K,-,0
should be included in the logit specification.
Thus Equation 7 is the generalization of ordinary logit that allows for tem-
porally interdependent observations. As we have just seen, it is easy to both
test and correct the logit for temporally dependent observations.
Cubic splines
Equation 7 requires the estimation of the coefficients of many dummy
variables. Unless N is large, estimates will not be precise. While this is not a
problem if our interest is in estimating 3, we may have some interest in the
K themselves. Note that the
Kt-t0
are easily interpretable as "baseline"
probabilities (or hazards) in that
P
yi,t
=
li,xt
=
0,
to =
1
(8)
1 1+ eK,ct-o
These baseline hazards give the probability of failure in each time interval
when all the independent variables are zero. If the independent variables are
measured so that these zeros are substantively meaningful, then the baseline
hazards are of substantive interest.
While the path traced out by the
Kxt_0
is easily interpretable, the impre-
cision with which the K are estimated may give a false impression that the
baseline hazard is jagged. We would expect it to be smooth, that is, baseline
hazard rates should change relatively slowly over time, rather than jumping
around from year to year.
One solution to this problem is to replace the dummy variables in Equa-
tion 7 with a smooth function of t - tO (we cannot directly use t - tO since
there is no reason to assume that the baseline hazard is a linear function of
time). In earlier work we recommended "cubic smoothing splines" (Beck
and Jackman 1997; Beck and Turner 1997). But while these work very
nicely, they do require software (such as S-Plus) that often is not readily ac-
cessible. One can obtain almost the same degree of smoothness with "natu-
ral cubic splines" (Eubank 1988), which are easy to implement with widely
available software packages (such as Stata). Natural cubic splines fit cubic
polynomials to a predetermined number of subintervals of a variable. These
polynomials are joined at "knots," with the number and placement of the
knots specified by the analyst. Smoothness is imposed by forcing the splines,
and their first and second derivatives, to agree at each of the knots. Thus each
knot only uses up one degree of freedom, so that we can flexibly fit a cubic
spline using up only a very few degrees of freedom. The estimated spline co-
efficients can then be used to trace out the path of duration dependence.
One advantage of the spline is that it facilitates a test of the hypothesis
of duration dependence. With many temporal dummy variables, the likeli-
hood ratio test for whether they are all zero may have poor finite sample
properties. The equivalent test on the spline formulation requires testing
only whether a small number of spline coefficients are zero.
Analysts can choose either the dummy variable or the spline formula-
tion; neither will have significant consequences for the estimation of
P3.
We
have a slight preference for the spline formulation. Users hesitant to deal
with natural splines can use the simpler dummy variable specification with
little loss if they are primarily interested in examining the effects of the sub-
stantive independent variables. We use both approaches in our replication,
though we rely primarily on the spline approach.
Since the logit with temporal dummy (or spline) variables is more gen-
eral than ordinary logit, and since we can easily test the null hypothesis of
duration independence, there is no reason not to undertake logit analysis of
BTSCS data, adding the temporal variables when they are required. This is
not to say that there might not be better methods for estimating some mod-
els. While the Cox proportional hazards model is widely used and works
well in practice, no model can be expected to be optimal for all problems.
We expect that logit analysis with temporal dummy or spline variables will
work well for most BTSCS data sets, and undoubtedly this approach is supe-
rior to ordinary logit.
3.3 Complications
Before turning to our reanalysis, several complications must be dis-
cussed. These complications would not arise if the data were independent,
but they are inherent if we are unwilling to make that assumption. The event
history approach simply makes these problems (and possible solutions)
clearer.
Multiple failures
The first problem is that BTSCS data allows for multiple failures per
unit. Many event history analyses simply model time until the first (or only)
failure, but the nature of BTSCS data allows for more than one failure per
unit.26 Ordinary logit avoids this problem by assuming that the probability of
failure in any year is the same as in any other year (conditional only on the
independent variables), so that second and subsequent failures are assumed
to be generated identically to first failures. In our construction of the K's we
have also used this assumption, since the only relevant information in the K
is time since the most recent event. However implausible the assumption
that second and subsequent events are independent of the number and timing
of previous events, this assumption is weaker than the ordinary logit
assumption that all observations are independent.
261f only one failure per unit were possible, we should discard all data after the first failure. But
in BTSCS data we have observations through a fixed time T.
Since the assumption that second spells are independent of first spells is
questionable, one solution might be to limit the analysis to the initial event.
While losing data on second events is inefficient, it does allow for consistent
estimation of ,1 without having to model the dependence of second and later
events on earlier events. Of course it would be better to correctly model re-
peated events. One easy way to do this is to include in the specification a
variable which counts the number of previous events. This approach, while
primitive, is better than ignoring the problem. A related issue common to IR
studies is that events may appear to take place over the course of several
years. If conflicts really are multi-year, we should simply drop all but the
first year of the conflict from the analysis. If we have a theory about the du-
ration of peace, we should not include spells of conflict in testing that
theory. However, since we can observe different conflicts in consecutive
years, this would be tantamount to discarding new, but very short, spells of
peace. A decision on how to proceed should be made on theoretical grounds.
But if we observe multi-year spells of conflict, it is difficult to maintain the
assumption that yearly observations are independent of each other. Duration
dependence may manifest itself in the finding that conflicts are more likely
to follow other conflicts.
Left censoring
The second concern has to do with what event history analysts call "left
censoring." Spells are left censored if we do not know when they began.27
For example, if our first dyadic observation is 1951, we do not know if a
spell of peace began in 1951, 1950, or before. This may not be a large prob-
lem in IR, since we can often begin analyses at the start of a new interna-
tional order or security regime (the Congress of Vienna or the beginning of
the Cold War). Our proposed method allows left censoring so long as all ob-
servations are equally left censored. For example, if the Cold War began in
1947, but our data starts in 1954, left censoring causes literally no problems
for our proposed method.28 All that is required is that the K for any given
year reflect the same length of prior peace spell length for all units.
This could cause problems, for example, for dyads that enter the data set
after the starting year. In our reanalysis, for example, some dyads enter the
data set after one of the members became independent. Suppose the data set
27Spells are "right censored" if we do not know when they ended. Units that are right censored
simply contribute a string of zeros, with no final one, to the logit likelihood. These are not a problem
for grouped duration logit analysis.
28AII we lose are estimates of the nuisance yearly dummies from 1948 through 1953. See
Jenkins (1995) for a formal proof and a good discussion of the interaction of grouped duration
analysis with various sampling designs. Jenkins shows that it is irrelevant, for the estimation of B, if
the sample period begins with t2 but that spells actually began at tl < t2.
begins in 1951 but a dyad enters the data set in 1962. Should the dummy
variable for that observation be K1 or K12? If our example (and in our re-
analysis) it seems reasonable to use
x,
here. But analysts will have to make
judgments before beginning their own data analysis. Results should be rela-
tively insensitive to a few differences in judgment on this issue.
Variables that are fixed across units
The third potential problem with our method is that it does not allow re-
searchers to use independent variables that vary by time but not across units.
In IR such variables are measured at the system level. Some examples of sys-
temic level variables are the concentration of power or the number of nation-
states in the world at any given time. These variables will be highly collinear
with the K.29 Inclusion of the K in the specification makes it unlikely that the
coefficients of these systemic variables will remain statistically significant.
This will cause problems for some, but not all, research agendas. 30 Systemic
level variables are, for example, rare in dyad-year studies of conflict.
If system level variables account for most of the duration dependence,
then our test for it will indicate that we cannot reject the hypothesis of dura-
tion independence. At that point researchers can confidently use ordinary
logit analysis, including system-level variables. This is the optimal situation,
since the system-level variables theoretically explain duration dependence.
However, we fear that this situation is rare.
There may be other situations that remain problematic. If the system
level variables are important, we might choose to ignore duration depen-
dence if it is not serious (as indicated by a baseline hazard function that
looks fairly flat). Sometimes the cure may be worse than the disease! Some
situations will occur where the researcher is faced with a choice between
two evils. The analysis of data is an art, not a science. No one method will
ever solve all possible problems. But a test for whether the temporal vari-
ables belong in the logit specification, even with the system level variables
included, at least alerts the researcher to the existence of potential problems
caused by temporally dependent observations.
Missing data
The fourth problem is that missing data becomes more troublesome in
the presence of duration dependence. The assumption of independence
29They are not perfectly collinear if there are multiple events per unit, since the K then no
longer simply mark t. They will also not be perfectly collinear with the temporal spline. But they
might be highly collinear.
30The problem is identical to that associated with fixed unit effects in models with independent
variables that are constant within units. However, this problem has not caused researchers to aban-
don fixed effects modeling.
allows the analyst to omit all observations with missing data-subject, of
course, to the usual caveats about missing data (Little and Rubin 1987).31
Our method also allows for the elimination of observations with missing
data so long as the correct time dummy variable is retained. Thus we cannot
allow missing observations on the dependent variable (or we must assume
that there were no missing years of conflict). In practice we will encounter
relatively little, if any, missing data in the conflict variable, since IR re-
searchers have gone to great lengths to code this data. But missing data on
the dependent variable could be a potential problem for other types of
BTSCS analyses. Keeping this in mind, we now turn to a reanalysis of one
prominent BTSCS study.
4. A REASSESSMENT OF THE LIBERAL PEACE
Russett and his colleagues (Russet 1990, 1993; Maoz and Russett 1992;
Maoz and Russett 1993; Oneal et al. 1996) have pioneered one of the most
important current research projects about the causes of militarized conflict.32
Their work on the "Liberal Peace" in particular has captivated IR research-
ers. To date, there have been two components of this liberal peace: a politi-
cal one (democracies are less likely to fight with other democracies) and an
economic one (trading partners are less likely to engage in militarized con-
flict). In fact, one of the most celebrated propositions in the IR/IPE literature
is that democracies do not wage war on one another.33 Moreover, the classi-
cal economic liberal argument that economic interdependence inhibits war
has received extensive empirical support for almost two decades.34 Russett
and his colleagues (Oneal et al. 1996) strengthened the confidence in these
findings by showing that the effects of economic interdependence and de-
mocracy are inversely related to the onset of military hostilities, even when
controlling for several important confounding factors. Oneal and Russett
(1997) claim to have improved upon the Oneal et al. (1996) specification to
further connect these two major strands of research on the causes of conflict.
Oneal and Russett, in exploring the interrelationship between liberalism (po-
3lOnly intra-unit missing data is a particular problem for our method. Obviously our proposed
method, like any method, is sensitive to the choice of which units to analyze. This choice is often
controversial in international relations (see, for example, Bremer 1992; Maoz and Russett 1993;
Lemke 1995).
32More than seventy-five articles have been published or presented at conferences in the last
five years that have relied on case selection criteria, variable measurement, or substantive foci origi-
nally developed or pursued by Russett and the members of his group.
33This has been confirmed in myriad empirical studies and has prompted Levy (1988) to imply
that this is the only law-like generalization in IR. For recent overviews of the democratic peace lit-
erature see Chan (1997) and Ray (1997).
34The first prominent statistical study was conducted by Polachek (1980) and the most recent
analysis can be found in Gartzke (1978). For a recent overview of the economic interdependence lit-
erature see McMillan (1997).
litical and economic) and militarized conflict, found that, during the Cold
War era, higher levels of democracy, as well as trade, lowered the probabil-
ity of hostilities between pairs of nations. These results, seemingly the most
robust of their genre, appear to have solidified conventional wisdom regard-
ing the relationships between economic interdependence, democracy, and
war. They therefore conclude that the classical liberal prescription for peace,
trade, and democracy, is sound.
The Oneal and Russett results are the Russett's research group's most
recently published and, arguably, most rigorous, empirical support for the
liberal economic and political peace research program. Their observations,
however, were presumed to be independent. That is, Oneal and Russett
(hereinafter O/R) performed ordinary logit analyses on BTSCS data without
accounting for temporal dependence. We use our proposed method to reana-
lyze their data to see whether their findings survive more appropriate statis-
tical tests.35
The dataset we use, generously provided by O/R, contains 20990 dyad-
years, comprised of 827 "politically relevant dyads" observed annually
from 1951 through 1985.36 Some dyads are observed for all thirty-five years,
while others are observed for a shorter subperiod. The median observation
length is twenty-two years.37 The dependent variable, militarized conflict, is
whether or not a dyad engaged in a militarized interstate dispute in a given
year. While earlier researchers typically used interstate war as a dependent
variable, recent research has frequently examined militarized interstate dis-
putes. Interstate wars are a small subset of militarized interstate disputes.
The latter include any event involving the threat or actual use of military
force, while the former require a substantial number of battle deaths.38
35Oneal and Russett propose, in a footnote, a series of methodological solutions. Initially, they
regress a variable that is the number of prior years of dyadic peace on the trade variable and then add
the residuals from the regression to the logit specification. However, this does not correct for tempo-
ral dependence. The temporal variables added to the logit specification to correct for duration depen-
dence may not be arbitrarily changed without undoing the correction for temporally dependent ob-
servations. They also claim that a modified Cochrane-Orcutt correction for temporal dependence did
not change their results. We know of no way, however, to modify the Cochrane-Orcutt procedure to
handle BTSCS data. Finally, O/R report that bootstrapped standard errors differed only slightly from
their reported standard errors. Standard bootstrapping, however, does not work with interdependent
observations (Freedman and Peters 1984).
36A dyad is "politically relevant" if the nations are geographically proximate or if one state is
a major power. The analysis of politically relevant dyad-years is a prominent IR BTSCS design. The
limitation of the dataset to politically relevant dyads, or a particular conception of relevancy, is not
without criticism, but is irrelevant to the methodological issues we are concerned with.
37Their data set has gaps in some dyadic observations. Although we did not attempt to fill these
in, we did correct the temporal variables for these gaps.
38We initially maintain O/R's coding decision to count every dispute year as a separate con-
flict, even when many of these were merely a continuation of the same event. As we shall see, this
decision turns out to have been crucial.
The two key independent variables, democracy and trade, are both dy-
adic measures. The dyadic democracy variable is constructed by creating de-
mocracy scores (using Polity III data) for each member of the dyad and tak-
ing the dyadic score as the lesser of the two (Oneal and Russett refer to this
as the "weak link" assumption). We rescaled democracy to run from -1 to 1.
The trade variable measures the importance of dyadic trade to the less trade-
oriented of the two partners. The importance of trade is measured by the ra-
tio (in percent) of dyadic trade to the GDP of each partner. Following O/R,
trade is lagged one year so that low trade does not proxy a current dispute.
O/R also use a series of control variables. Alliance is a dummy variable
measuring whether the dyad partners were allied (or both were allied with
the United States). Contiguous is a dummy variable indicating the geo-
graphical contiguity of both states. Capability Ratio measures the dyadic
balance of power. Using the Correlates of War material capabilities index, it
is the ratio (in percent) of the stronger nation's score to the weaker nation's.
Finally, economic growth measures the lesser of the rates of economic
growth (as a percent) of the partners. Detailed discussion of the O/R data set
and research design is contained in their original paper.39 The analyses
which correct for duration dependence either use a natural cubic spline in a
variable we call peace years or the set of dummy variables created from
peace years. Peace years counts the length of the spell of peace preceding
the current observation. For observations with no previous dyadic disputes,
this variable is simply t - 1, since the time index starts at zero; subsequent to
a dispute, this variable is t
-
tO (where tO is the time index of the most recent
dispute). The variable peace years ranges from zero to 34.
Column I of Table 1 shows the original O/R results. We were able to
replicate Oneal and Russett's (1997) original estimates exactly.40 We limited
our reanalysis to the temporal dependence issues discussed in our paper. Ex-
amining only their specification 1, we will not present alternative substan-
tive models of conflict.41 Results indicate that both democracy and trade
lower the probability of a militarized dispute; they appear to be both statisti-
cally and substantively significant. The control variables, as O/R predicted,
also exhibit substantively important effects.
A different picture emerges, however, when we correct for temporally
dependent observations using grouped duration methods. (See Columns II
39Since our interest is in examining the consequences of temporal dependence, we do not con-
sider issues of operationalization or case-selection, nor do we consider specifications outside the 0/
R framework.
40A11 of our analyses were done with Stata, Version 5. Note that some variables were rescaled
to simplify the reading of the tables.
41Their other specifications are similar to that examined here. We have applied our method to
their specifications two through six and obtained similar results.
Table 1. Comparison of Ordinary Logit
and Grouped Duration Analyses
Ordinary
Logit Grouped Duration
Logit Logit Cloglog
Dummya Spline Dummyb
Variable I II III IV
Democracy -0.50 -0.55 -0.54 -0.49
(0.07) (0.08) (0.08) (0.07)
Economic Growth -2.23 -1.15 -1.15 -0.81
(0.85) (0.92) (0.92) (0.76)
Alliance -0.82 -0.47 -0.47 -0.43
(0.08) (0.09) (0.09) (0.08)
Contiguous 1.31 0.70 0.69 0.55
(0.08) (0.09) (0.09) (0.08)
Capability Ratio -0.31 -0.30 -0.30 -0.30
(0.04) (0.04) (0.04) (0.04)
Trade -66.13 -12.67 -12.88 -12.50
(13.44) (10.50) (10.51) 9.96
Constant -3.29 -0.94 -0.96 -1.11
(0.08) (0.09) (0.09) (0.08)
Peace Years -1.82
(0.11)
Spline(l)
c
-.24
(0.03)
Spline(2)
c
-.08
(0.01)
Spline(3)c -.01
(0.003)
Log Likelihood -3477.6 -2554.7 -2582.9 -2554.1
df 20983 20036 20979 20949
N=20990
Standard errors in parentheses
a31 temporal dummy variables in specification not shown
3dummy variables and 916 observations dropped due to outcomes being perfectly predicted
b34 temporal dummy variables in specification not shown
CCoefficients of Peace Years cubic spline segments
and III for the logit link and Column IV for the cloglog link results.) A test
for whether the temporal dummies (a likelihood ratio test of I vs. II), or
the temporal splines (I vs. III), are required indicate strong duration
dependence.42 Likelihood ratio tests of specification II versus I yielded
x2
statistics of 1778 with 31 degrees of freedom; the test of III versus I yields
a statistic of 1789 with 4 degrees of freedom. The test of I versus II drops
916 perfectly predicted observations from both logits so that log likeli-
hoods are comparable. The probability of obtaining either result by chance
is, to computer precision, zero.43 Thus the O/R logits clearly show duration
dependence.
Estimation that accounts for duration dependence has dramatic conse-
quences for the O/R finding. The coefficient of trade, in particular, is reduced
by a factor of five (and becomes statistically insignificant). These results pro-
vide no evidence for a liberal economic peace. Not all coefficients, however,
are affected by controlling for duration dependence. Our reanalysis leaves the
democracy coefficient and standard error basically unchanged. Thus the
Oneal and Russett evidence in support of the liberal economic peace is an ar-
tifact of their incorrect assumption of temporal independence; their findings
about the liberal political peace, however, are upheld.44 The consequence of
controlling for duration dependence on any variable is difficult to predict in
advance. But, as we see here, the consequences of failing to correctly account
for duration dependence in typical logit estimation may be enormous.45
4.1 Links and Splines
A closer examination of Table 1 reveals that it appears to make little dif-
ference whether we use the logit or cloglog link. The estimates for the two
different links are even more similar than Table 1 (Columns II and IV) indi-
cates, since the transformation of the independent variables into probabili-
ties differs slightly between the two links. The mean difference in predicted
probabilities between the two models is 0.007 percent, with only about 2
percent of all dyad-years having predicted probabilities of a dispute differing
by more than 1 percent. Thus, as we recommended, subsequent analyses use
only the logit link.
42This can also be seen by looking at the t-ratios of the four terms that comprise the cubic
spline in peace years: 14, 9, 7, and 4.
43
Tests on specifications in subsequent tables reveal similar results and are not shown here.
4'The control variables are also differentially impacted. Although the coefficient of capability
ratio is almost unchanged, in magnitude and statistical significance, the coefficients of the three
other control variables are cut by half (with the economic growth coefficient even becoming statisti-
cally insignificant).
45We have applied these methods to their specifications 2 through 6 and obtained similar re-
sults. For example, Oneal and Russett (1997, 283) state that they "re-estimated [their] equation (6)
with indicator variables for all years but one [with] results consistent with those [they] report." We
reestimated their Equation 6 with temporal dummies and found that the coefficient on trade dropped
by a factor of four, (becoming statistically insignificant), and that the coefficient on the trend of trade
dropped by a factor of five (also becoming statistically insignificant). While the coefficients on the
two democracy variables declined by 30 percent, they remained strongly statistically significant.
Results using a natural cubic spline in peace years appear in Column III.
A comparison of Columns II and III shows that it makes no difference in
terms of estimating 3 whether we use temporal dummy variables or a cubic
spline in peace years. Since we prefer the spline setup, all subsequent analy-
ses are performed using the natural cubic spline in the length of prior spells
of peace.46
4.2 Why Duration Dependence Affects
the Findings on Economic Interdependence
Temporal dependence clearly has dramatic effects on Oneal and
Russett's finding that economic interdependence decreases conflict. Oneal
and Russett (1997, 283) claim that they theoretically expect a high correla-
tion between trade and length of spells of peace and hence conclude that
trade really does lessen conflict. But the problem in this explanation is that
it does not take into account the correlation between trade and lengths of
spells of conflict, particularly when combined with a higher than average
probability of a subsequent conflict immediately following the initial onset.
We can better understand why accounting for duration dependence so
strongly affects the O/R finding on trade by using some basic event history
techniques. We begin with an examination of the estimated hazard function
which is computed for the logit analyses by setting all independent variables
at their means (except for the two dummy variables which are set to their
modal value of zero). The estimated hazard function, plotted against the
length of peace spell, peace years, is shown in Figure 2.
The probability of a dispute immediately following a prior dispute is al-
most 25 percent. It immediately falls to about 5 percent the next year and to
about 2 percent the third year, where it remains for the rest of the spell of
peace. Thus, much of what the duration dependent logit highlights is the de-
pendence of the probability of a dispute on an immediately preceding dis-
pute. Counting the latter years of multi-year disputes as new disputes, and
failing to correct for dependence between these disputes, is what leads to the
Oneal and Russett finding that trade lowers the probability of the onset of a
dispute.
It appears that economic interdependence does not dampen the prob-
ability of a dispute, but it does diminish the duration of a dispute once it oc-
curs. Remember that trade is lagged one year so that the previous year's
46AII splines allowed for three knots, placed at 1, 4, and 7 years of peace. The number of knots
was chosen by a sequence of F-tests; a variety of knot placements were tried to ascertain the one
with the best performance. Small changes in the number or placement of the knots had no effect on
the results. The natural cubic spline estimated here is similar to the smoothing splines shown in Beck
and Jackman (1997) and Beck and Tucker (1997). We also reran the analyses for subsequent tables
using temporal dummy variables, obtaining almost identical results.
Figure 2. Discrete Hazard of Dispute
0.25-
Spline
- -Dummies
0.20-
0.15-
0.10-
0.05-
0.0- ....
0 5 10 15 20 25 30 35
Duration of Peace
trade predicts the current probability of a dispute. Trade averages 0.22 per-
cent of GDP prior to one year disputes. This is only slightly lower than the
0.23 percent of GDP that trade averages prior to a year of peace. But, in the
last year of peace prior to a multi-year dispute, trade averages only 0.15 per-
cent of GDP. Thus trade is not a good predictor of whether a dispute will oc-
cur, but if one does, it is a good predictor of whether it will be lengthy. Low
trade may prolong conflicts, but it does not appear to cause them.
4.3 The Effect of Multiple Disputes
The elimination of ongoing dispute years
We can further examine the contaminating effects of long spells of dis-
putes by eliminating ongoing years of a dispute from the analysis. Five hun-
dred forty-two dyad-years with a dispute are thus dropped.47 Results of this
analysis are in Table 2.
47All disputes that continue for more than one year are dropped, even if disputes in subsequent
years have different identification codes.
Table 2. Grouped Duration Analyses: No Continuing Dispute Years
I. Logit II. Logit
Ordinary Group Dur.
Variable se f se
Democracy -0.40 0.10 -0.39 0.10
Economic Growth -3.43 1.25 -4.01 1.25
Alliance -0.48 0.11 -0.37 0.11
Contiguous 1.35 0.12 0.99 0.12
Capability Ratio -0.20 0.05 -0.22 0.05
Trade -21.08 11.30 -3.81 9.68
Constant -4.33 0.11 -3.57 0.17
Peace Years 0.39 0.16
Spline(l)a 0.09 0.03
Spline(2)a -0.03 0.01
Spline(3)a 0.003 0.003
Log Likelihood -1846.9 -1751.4
df 20441 20437
N = 20448
aCoefficients of Peace Years cubic spline
Dropping the latter years of a dispute, even without accounting for du-
ration dependence, reduces the trade coefficient by a factor of three, leaving
it barely statistically significant. A likelihood ratio test, however, clearly
shows remaining duration dependence. When we account for this (Column
II), the effect of dyadic trade is again greatly reduced and is now not even
close to being statistically significant. The elimination of ongoing dispute
years, even accounting for duration dependence, has little effect on the de-
mocracy coefficient.48
Time
untilfirstfailure
We can also examine the contaminating effects of disputes on later dis-
putes by confining our analysis to first disputes (eliminating observations on
3999 dyad years which followed an initial dispute). This analysis avoids any
problems associated with the need to model the conditional probability of
second and later disputes. Results appear in Table 3, Column I.
Limiting our analyses to the onset of the first dispute eliminates about
20 percent of the data and results in an increase in all the standard errors.
The pacific effect of democracy remains almost unaffected by this limitation
48The estimated pacific impact of economic growth dramatically increases when ongoing dis-
pute years are omitted.
Table 3. Grouped Duration Analyses: Second Disputes Differ
I. First II. Prior
Disputes Disputes
Variable 1 se 1 se
Democracy -0.46 0.13 -0.41 0.08
Economic Growth -2.29 1.78 -2.09 0.97
Alliance -0.42 0.16 -0.25 0.09
Contiguous 1.11 0.17 0.69 0.09
Capability Ratio -0.19 0.06 -0.20 0.04
Trade -3.55 11.73 -9.39 10.19
Prior Disputes (#) 0.17 0.01
Constant -3.21 0.21 -1.60 0.10
Peace Years -1.08 0.24 -1.67 0.11
Spline(I)a -0.18 0.05 -0.22 0.03
Spline(2)a -0.07 0.02 0.07 0.01
Spline(3)a -0.01 0.05 -0.01 0.003
Log Likelihood -964.1 -2393.0
df 16980 20978
N 16691 20990
aCoefficients of Peace Years cubic spline segments
(despite a slight increase in its standard error). But once again, the estimated
impact of economic interdependence drastically decreases. Increased dyadic
trade does not reduce the likelihood of an initial dyadic dispute onset; dyadic
democracy does.
A less drastic way to allow for differing conditional probabilities of a
dispute given the number of prior disputes is to add to the logit a counter
measuring the number of prior dyadic disputes. These results are in Table 3,
Column II. While the results are not as dramatic as the limitation to first dis-
putes only, they clearly show the pacific effect of democracy but not of
trade. Accounting for temporal dependence clearly has dramatic effects on
O/R's finding that trade decreases conflict. When the O/R estimation is cor-
rected for that dependence, the finding that trade reduces conflict simply dis-
appears (although trade may reduce the length of conflicts once they occur).
Our reassessment of the O/R finding, however, leaves intact their conclusion
about the pacific effects of democracy.
5. CONCLUSION
The analysis of binary dependent variable time-series-cross-section
data are becoming more common, particularly in the study of international
conflict. Virtually all analyses of this type of data use ordinary logit, ignor-
ing issues of temporal interdependence of the data. We have shown that al-
lowing for temporal dependence in logit analysis is easy once we recognize
that BTSCS data are grouped duration data. Such data can then be analyzed
by adding temporal dummy variables (or a temporal spline) to the logit
specification. This can be done using any statistical software package.
This remedy has advantages over other attempts to correct for temporal
dependence in BTSCS data. Because it uses standard logit routines, it can be
combined with remedies that address other problems. In particular, it is
simple to combine our method with Huber (1967) standard errors, which
solve other problems inherent in BTSCS data. One can also allow for hetero-
geneity using our method (Jenkins 1995). Solutions to one problem should
be applicable to others; real data are seldom subject to only one problem. A
related advantage is that logit is well-known and well understood by re-
searchers.
Our proposed method forces logit analysts to think about some prob-
lems that naturally occur to the event history analyst (which do not naturally
occur to the logit analyst). In particular, our approach requires analysts to
think about whether they are modeling spells of peace, spells of conflict, or
both. It also requires pondering the modeling of second and subsequent
events for the same unit. These considerations are critical to the modeling
process.
There are clearly other possible ways to allow for temporal dependence
in BTSCS data. Ideally we would model that dependence as a function of
other variables. Our proposed method simply treats serial correlation as a
nuisance which impedes estimation of the
P.
Nothing is explained by noting
that hazard rates change with time.
Our treatment of duration dependence parallels the earlier methods for
estimating time series models with serially correlated errors that treats the
dependence as an estimation nuisance. While we agree with more modern
treatments that advocate directly modeling the dynamics instead of treating
it as a nuisance, we note that simply ignoring the nuisance can lead to se-
verely incorrect inferences. A theoretically based specification of this dura-
tion dependence would be best, but it is easier to give this advice than to
implement it. In the meantime we surely do not wish to continue the current
practice of ignoring potentially serious duration dependence.
The analogy to the estimation of time series models with serially corre-
lated errors is also helpful in understanding when duration dependence
might cause serious problems for BTSCS data. As with serially correlated
errors in a linear regression, a small amount of duration dependence, even if
statistically significant, will only cause a small amount of harm. Thus re-
searchers must assess not only whether they can reject the null hypothesis of
duration independence, but also the severity of duration dependence (by
ex-
amining the estimated baseline hazards). Also, as with serially correlated
errors, the degree of harm is related to the temporal structure of the indepen-
dent variables. Estimation problems caused by ignoring duration depen-
dence increase as the independent variables themselves trend or otherwise
show time variation. Duration dependence, then, will not always have enor-
mous consequence. For example, duration dependence should be less of a
problem in data where the units are observed less frequently, so that dyad-
year data will generally show more duration dependence than will dyad-de-
cade data. But, as with serially correlated errors, researchers cannot ignore
the potential problems that might be caused by duration dependence. Fortu-
nately our proposed test and correction are easy to implement, so there is no
reason for research to ignore these potential problems.
We have applied our methodological remedy to Oneal and Russett's
(1997) analysis of militarized conflict during the Cold War period, which
found that both political and economic liberalism inhibit conflict. Our re-
analyses show that democracy clearly inhibits conflict, but that trade (at least
as Oneal and Russett measure it) does not. But while trade may not inhibit
conflict, it does appear to shorten spells of conflict. The differences between
the original analysis and our reanalysis are considerable. Temporal depen-
dence in BTSCS models is not a minor problem that can be ignored at the
cost of a small error. And there is no reason to commit these errors. The in-
clusion of temporal variables in the specification is a simple solution, avail-
able to all researchers, providing a low cost cure to the problem of tempo-
rally dependent BTSCS data.
Manuscript submitted 3 September 1997.
Final manuscript received 12 December 1997.
APPENDIX
The Simple Math of Grouped Durations
This appendix derives the grouped duration model. We present it here for complete-
ness and because many standard event history texts do not present this result. This
appendix assumes familiarity with basic duration concepts.
Start with a continuous time Cox proportional hazards model, so
hi
(t)
=
ho
(t)exiI3 (9)
where i refers to units, t refers to continuous time,
xit
is a vector of independent vari-
ables, and
ho(t)
is the unspecified baseline hazard.
Letting S(t) be the probability of surviving beyond t, we use the basic identity that
t
S(t)
=
exp
-|
h(,r)dr
(10)
We only observe whether or not an event occurred between time tk
- 1 and tk (assuming
annual data) and are interested in the probability of this event,
P(yi,tk
= 1). This prob-
ability is one minus the probability of surviving beyond tk given survival up to tk
- 1.
Assuming no prior events (so tO = 0) and using Equation 10, we then get
P(Yi,tk
=
1)
=
1-
exp
-j
(c)dtJ
(
lla)
= 1- exp-, eXi*kfho (c)dt1 ( lib)
1- exp -exki jho (t)dtI (l Ic)
(Note that x is indexed by tk not t because we assume that the independent variables are
only measured for an entire interval and not for every instant in the interval tk
- 1 to tk.)
Since the baseline hazard is unspecified, we can just treat the integral of the baseline
hazard as an unknown constant. Defining
k=
jtk
o
()dt
and (12)
Ktk= log(atk) (13)
we then have
P(Yi,tk
=
1)
= 1-
exp(
eXitPa
tk)
(14a)
1
-
exp(_eXi,tkP+Ktk) (14b)
This is exactly a binary dependent variable model with a cloglog link
REFERENCES
Allison, Paul D. 1982. "Discrete-Time Methods for the Analysis of Event Histories." In Sociologi-
cal Methodology, ed. Samuel Leinhardt. San Francisco: Jossey-Bass. pp. 61-98.
Alt, James E., Gary King, and Curtis Signorino. 1997. Estimating the Same Quantities from Differ-
ent Levels of Data: Time Dependence and Aggregation in Event Process Models. Technical re-
port. Department of Government, Harvard University (http://Gking.Harvard.Edu/preprints.
shtml).
Barbieri, Katherine. 1996. "Economic Interdependence: A Path to Peace or a Source of Interstate
Conflict?" Journal of Peace Research 33:29-50.
Beck, Nathaniel. N.d. "Modelling Space and Time: The Event History Approach." In Research Strat-
egies in Social Science, ed. Elinor Scarbrough and Eric Tanenbaum. Oxford: Oxford Univer-
sity Press.
Beck, Nathaniel, and Jonathan N. Katz. 1995. "What To Do (and Not To Do) with Times-Series
Cross-Section Data." American Political Science Review 89:634-47.
Beck, Nathaniel, and Jonathan N. Katz. 1996. "Nuisance vs. Substance: Specifying and Estimating
Time-Series-Cross-Section Models." Political Analysis 6:1-36.
Beck, Nathaniel, and Jonathan N. Katz. 1997. "The Analysis of Binary Time-Series-Cross-Section
Data and/or The Democratic Peace." Presented at the annual meeting of the Political Method-
ology Group, Columbus, OH.
Beck, Nathaniel, and Richard Tucker. 1997. Conflict in Time and Space. Center for International
Affairs Working Paper, No. 97-8. Harvard University (http://wwwl.columbia.edu/sec/dlc/
ciao/wpsfrm.html).
Beck, Nathaniel, and Simon Jackman. 1997. Getting the Mean Right is a Good Thing: Generalized
Additive Models. Working paper. Political Methodology WWW Site (http://wizard.ucr.edu/
polmeth/working_papers97).
Bennett, D. Scott. 1996. "Security, Bargaining, and the End of Interstate Rivalry." International
Studies Quarterly 40:157-84.
Bennett, D. Scott. 1997. "Testing Alternative Models of Alliance Duration, 1816-1984." American
Journal of Political Science 41:846-78.
Berry, Frances Stokes, and William D. Berry. 1990. "State Lottery Adoptions as Policy Innovations:
An Event History Analysis." American Political Science Review 84:395-415.
Berry, Frances S., and William D. Berry. 1992. "Tax Innovation in the States: Capitalizing on Politi-
cal Opportunity." American Journal of Political Science 36:715-42.
Box-Steffensmeier, Janet M., and Bradford S. Jones. 1997. "Time is of the Essence: Event History
Models in Political Science." American Journal of Political Science 41:1414-61.
Bremer, Stuart A. 1992. "Dangerous Dyads: Conditions Affecting the Likelihood of Interstate War,
1816-1965." Journal of Conflict Resolution 36:309-41.
Bueno de Mesquita, Bruce, and David Lalman. 1992. War and Reason: Domestic and International
Imperatives. First ed. New Haven: Yale University Press.
Chan, Steve. 1997. "In Search of Democratic Peace: Problems and Promise." Mershon International
Studies Review 41:59-91.
Cox, D. R. 1975. "Partial Likelihood." Biometrika 62:269-76.
Diggle, Peter J., Kung-Yee Liang, and Scott L. Zeger. 1994. Analysis of Longitudinal Data. Oxford:
Oxford University Press.
Enterline, Andrew. 1996. "Driving While Democratizing." International Security 20:183-196.
Enterline, Andrew. 1997. "Fledgling Regimes: Is the Case of Inter-War Germany Generalizable?"
International Interactions 22:245-277.
Eubank, Randall L. 1988. Spline smoothing and nonparametric regression. New York: Marcel
Dekker.
Farber, Henry S., and Joanne Gowa. 1997. "Common Interests or Common Polities: Reinterpreting
the Democratic Peace." Journal of Politics 59:393-417.
Farhmier, Ludwig, and Stefan Wagenpfeil. 1996. "Smoothing Hazard Functions and Time-Varying
Effects in Discrete Duration and Competing Risks Models." Journal of the American Statisti-
cal Association 91:1584-94.
Freedman, David, and Stephen Peters. 1984. "Bootstrapping a Regression Equation: Some Empiri-
cal Results." Journal of the American Statistical Association 79:97-106.
Gartzke, Erik. 1998. "Kant We All Just Get Along?: Opportunity, Willingness and the Origins of the
Democratic Peace." American Journal of Political Science 42:1-27.
Gleditsch, Nils Petter, and H'avard Hegre. 1997. "Peace and Democracy: Three Levels of Analysis."
Journal of Conflict Resolution 41:283-3 10.
Goertz, Gary, and Paul Diehl. 1993. "Enduring Rivalries: Theoretical Constructs and Empirical Pat-
terns." International Studies Quarterly 37:147-71.
Han, Aaron, and Jerry A. Hausman. 1990. "Flexible Parametric Estimation of Duration and Compet-
ing Risk Models." Journal of Applied Econometrics 5:1-28.
Henderson, Errol A. 1997. "Ethnic Conflict, The Similarity of States, and the Onset of War, 1820-
1989." Journal of Conflict Resolution 4:649-68.
Hermann, Margaret G., and Charles W. Kegley. 1996. "Ballots, a Barrier against the Use of Bullets
and Bombs: Democratization and Military Intervention." Journal of Conflict Resolution
40:436-60.
Huber, Peter J. 1967. The Behavior of Maximum Likelihood Estimates Under Non-Standard Condi-
tions. In Proceedings of the Fifth Annual Berkeley Symposium on Mathematical Statistics and
Probability, ed. Lucien M. LeCam and Jerzy Neyman. Vol. I. Berkeley: University of Califor-
nia Press.
Huth, Paul. 1996. Standing Your Ground: Territorial Disputes and International Conflict. Ann Ar-
bor: University of Michigan Press.
Jenkins, Stephen P. 1995. "Easy Estimation Methods for Discrete-Time Duration Models." Oxford
Bulletin of Economics and Statistics 57:129-38.
Katz, Jonathan N., and Brian Sala. 1996. "Careerism, Committee Assignments and the Electoral
Connection." American Political Science Review 90:21-33.
Lemke, Douglas. 1995. "The Tyranny of Distance: Redefining Relevant Dyads." International Inter-
actions 21:23-38.
Lemke, Douglas, and William Reed. 1996. "Regime Types and Status Quo Evaluations: Power Tran-
sition Theory and the Democratic Peace." International Interactions 22:143-164.
Levy, Jack. 1988. "Domestic Politics and War." Journal of Interdisciplinary History 18:653-73.
Little, Roderick J. A., and Donald B. Rubin. 1987. Statistical Analysis with Missing Data. New
York: Wiley.
Mansfield, Edward, and Jack Snyder. 1996. "The Effects of Democratization on War." International
Security 20:196-207.
Mansfield, Edward, and Jack Snyder. 1997. "A Response to Thompson and Tucker." Journal of Con-
flict Resolution 41:457-461.
Maoz, Zeev. 1996. Domestic Sources of Global Change. Ann Arbor: University of Michigan Press.
Maoz, Zeev. N.d. "Realist and Cultural Critiques of the Democratic Peace: A Theoretical and Em-
pirical Re-Assessment." International Interactions. Forthcoming.
Maoz, Zeev, and Bruce Russett. 1992. "Alliance, Contiguity, Wealth, and Political Stability: Is the
Lack of Conflict Among Democracies A Statistical Artifact?" International Interactions
17:245-267.
Maoz, Zeev, and Bruce Russett. 1993. "Normative and Structural Causes of Democratic Peace,
1946-1986." American Political Science Review 87:639-56.
McCullagh, P., and J. A. Nelder. 1989. Generalized Linear Models. 2nd ed. London: Chapman and
Hall.
McMillan, Susan M. 1997. "Interdependence and Conflict." Mershon International Studies Review
41:33-58.
Meier, Kenneth J., and Deborah R. McFarlane. 1992. "State Policies on Funding of Abortions: A
Pooled Time Series Analysis." Social Science Quarterly 73:690-8.
Mintrom, Michael. 1997. "Policy Entrepreneurs and the Diffusion of Innovation." American Journal
of Political Science 41:738-70.
Mousseau, Michael. 1997. "Democracy and Militarized Interstate Collaboration." Journal of Peace
Research 34:73-87.
Oneal, John R., and Bruce Russett. 1997. "The Classical Liberals Were Right: Democracy, Interde-
pendence, and Conflict, 1950-1985." International Studies Quarterly. 41:267-94.
Oneal, John R., Francis H. Oneal, Zeev Maoz, and Bruce Russett. 1996. "The Liberal Peace: Inter-
dependence, Democracy and International Conflict." Journal of Peace Research 33:11-29.
Poirier, Dale J., and Paul A. Ruud. 1988. "Probit with Dependent Observations." Review of Eco-
nomic Studies 55:593-614.
Polachek, Solomon. 1980. "Conflict and Trade." Journal of Conflict Resolution. 24:55-78.
Prentice, R., and L. Gloeckler. 1978. "Regression Analysis of Grouped Survival Data with Applica-
tion to Breast Cancer Data'." Biometrics 34:57-67.
Raknerud, Arvid, and Hiavard Hegre. 1997. "The Hazard Of War: Reassessing Evidence of the
Democratic Peace." Journal of Peace Research 34:385-404.
Ray, James Lee. 1997. "The Democratic Path to Peace." Journal of Democracy. 8:49-64.
Russett, Bruce. 1990. "Economic Decline, Electoral Pressure, and the Initiation of International
Conflict." In The Prisoners of War, ed. Charles Gochman and Ned Allan Sabrosky. Lexington,
Mass.: Lexington Book.
Russett, Bruce. 1993. Grasping the Democratic Peace. Princeton: Princeton University Press.
Singer, Judith D., and John B. Willett. 1993. "It's About Time: Using Discrete-Time Survival Analy-
sis to Study Duration and the Timing of Events." Journal of Educational Statistics 18:155-95.
Sueyoshi, Glenn T. 1995. "A Class of Binary Response Models for Grouped Duration Data." Jour-
nal of Applied Econometrics 10:411-31.
Vasquez, John A. 1993. The War Puzzle. Cambridge: Cambridge University Press.

Beck and Katz Series

Uploaded by

Copyright:

Available Formats

Beck and Katz Series

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 and Katz Series

Uploaded by

Copyright:

Available Formats

Taking Time Seriously: Time-Series-Cross-Section Analysis with a Binary Dependent Variable

Author(s): Nathaniel Beck, Jonathan N. Katz and Richard Tucker

You might also like