Content uploaded by Katherine Barbieri
Author content
All content in this area was uploaded by Katherine Barbieri on Dec 28, 2013
Content may be subject to copyright.
471
Conflict Management and Peace Science
© The Author(s). 2009. Reprints and permissions:
http://www.sagepub.co.uk/journalsPermissions.nav
[DOI:10.1177/0738894209343887]
Vol 26(5): 471–491
Trading Data:
Evaluating our Assumptions and Coding Rules1
KATHERINE BARBIERI
University of South Carolina, USA
OMAR M. G. KESHK
The Ohio State University, USA
BRIAN M. POLLINS
The Ohio State University, USA
Some scholars have rushed to judgment about the nature of the relationship between
trade and conflict, making strong assumptions about the data upon which their
conclusions rest. In this paper, we test these assumptions, showing that they are often
not warranted and, thus, pose threats to many of our conclusions about trade’s impact
on conflict. We discuss official trade statistics; the treatment of missing trade data; and
problems with some decision rules being adopted within our research community. We
introduce the new Correlates of War (COW) Trade Data Set; discuss the rationale
behind our coding decisions; and compare this data set with other sets. The end result
is a series of findings that should help our field advance its understanding of the often
difficult issue of trade’s relationship with international conflict.
KEYWORDS: COW Trade Data Set; IMF data; imputed data; militarized interstate
disputes; missing data; trade; trade and conflict; trade data; trade statistics
Introduction
The emergence of the sub-field of international political economy (IPE) in the 1970s
reframed the attention of a discipline that had largely ignored economic variables.
Suddenly, economics, including relations of dependence and interdependence, became
central to our understanding of peace, security, and justice. Despite the theoretical
interest, scholars seeking to test hypotheses about economics and conflict faced
1 We would like to thank John Geer, Ranan Kuperman, Glenn Palmer, Paul Paust , and James
Lee Ray for their helpful comments.
Conflict Management and Peace Science
472
numerous hurdles in acquiring quality data. The field, which had previously paid too
much attention to Morgenthau’s Politics Among Nations (1948), now paid too little
heed to Morgenstern’s (1963) On the Accuracy of Economic Observations.
Too little consideration was given to the accuracy of economic data, the expertise
of the analysts generating it, coding decisions, or the rationale for combining dif-
ferent data types into aggregate variables.2 Unfortunately, these problems continue
to plague our research community, particularly in the area of international trade
statistics.3 In this paper, we discuss some of the general problems associated with
official trade statistics;4 the treatment of missing trade data; and problems with several
decision rules being employed in the field. Next, we introduce the new Correlates
of War (COW) Trade Data Set; discuss the rationale behind our coding decisions;
and compare these data with other available evidence. Our findings call into ques-
tion some of the conclusions that previous work has made about trade and conflict.
Common Problems with Official Trade Statistics
Scholars interested in accurately depicting trading relationships confront a series
of challenges. We identify and discuss what we believe are three of the most serious
problems scholars face when using official trade statistics: (1) one or both states in a
dyad file inaccurate reports; (2) two states report dyadic trade, but the values reported
differ considerably by state; or (3) one or both countries fail to report trade. These are
not mutually exclusive categories. For example, if the International Monetary Fund
(IMF) trade data set suggests that one state does not trade with another state (i.e.
a zero value is attached to the dyad), the report may be (1) accurate or inaccurate;
(2) similarly reported by both states in the dyad; or (3) the product of IMF adjust-
ments to missing national reports. Scholars who observe the zero value are unlikely
to know or question the origin and accuracy of the IMF’s figure. We will discuss, in
turn, the causes and consequences of several critical trade data problems. Later, we
discuss how we deal with these problems and how others handle them.
Inaccurate Trade Reports
Inaccurate trade reports may result from deliberate or non-deliberate acts by
governments who compile information; publish trade records; and submit these to
international organizations. The international organization itself might also intro-
duce errors. It is often difficult to compare national trade reports and assess their
2 For example, many scholars created ratio dependence measures where they relied on figures
that had the dollar value of the denominator (GDP) adjusted with a PPP index, but no index
was applied to the numerator (Trade).
3 Fully reviewing the trade–conflict literature is beyond the scope of this study. See Schneider
et al., 2003; and Mansfield and Pollins, 2003 for thorough reviews.
4 For problems with trade statistics, see Bhagwati, 1964, 1967; Bhagwati and Srinivasan, 1981;
Ely, 1961; Morgenstern, 1963; Pak and Zdanowicz, 1994a,b; Pak et al., 2003; Sheikh, 1974;
Yeats, 1978, 1990; de Wulf, 1981.
Barbieri, Keshk & Pollins: Trading Data
473
accuracy, since reporting procedures vary across countries. Despite the IMF’s efforts
to standardize reporting practices, differences remain among states and across time.
Researchers should keep this in mind when employing historical time series data.
For example, in the past, most states reported special trade (direct trade between
the point of origin and point of destination), but a few states reported only general
trade (special trade plus transit trade—goods that flow through an area on their way
to their final destination). Scholars typically ignore the distinction between special
and general trade. Yet, these figures may be significantly different and they may also
provide insight into the role of transit traders in global economics and politics.
For example, transit traders may serve as a conduit through which hostile states
trade and may provide a means for circumventing sanctions, quotas, and other trade
laws. This and other types of illegal trade may result in inaccurate trade figures.
Advances by scholars who research and estimate illegal trade are seldom incorpo-
rated into traditional trade research. Although we do not solve this problem here,
we believe it is important to consider what is missing in official trade figures and
to be cautious when making assumptions about trade data. Even when trade is
legal, it may be difficult to determine whether national trade statistics are accurate,
whether reporting errors are deliberate, and whether firms, states, or other actors
are responsible for the inaccuracies. Scholars typically assume that developed states
have more advanced techniques than developing states to compile data, monitor
firm activities, and ascertain the origin and destination of traded goods. Certainly,
one source of inaccurate reporting is suboptimal recording and monitoring systems.
There are great variations across countries in terms of customs administrations, law
enforcement, and border patrol. Here, the government might wish to report trade
accurately, but be unable to do so.
False reports may also be the product of politically motivated efforts to distort
trade figures. A state might inflate its trade figures with a given partner when con-
fronted with pressure to open its markets to goods from that country. States might
also inflate trade figures when they wish to project an image of economic growth
(e.g. for domestic audiences or to attract foreign investment). Transit traders might
also distort trade figures to hide sanctions-busting activities. Even when trade is
legal, both states and firms may wish to avoid public image problems that could
come with trading with pariah states. In addition, governments might not report all
trade in military or sensitive goods (IMF, 2008). Clearly, there are various reasons
why states might hide legal and illegal trade.
In some cases, governments may not be aware that trade exists between their
citizens and those other countries, or they may recognize the existence of trade ties
but be unable to record or prevent them. For example, the United States is aware
that illegal products cross its borders daily and may even know some actors involved
in this trade, but it is unable to stop illegal trade or record its value accurately. Even
when states develop estimates of illegal trade and transnational ties, based upon
seizures and information from violators caught in illegal activities, these estimates
are typically excluded from official trade statistics.5
5 According to the UN, some nations include illegal trade in their national trade statistics, but
the UN gives no explanation for how the estimates are derived (UN, 2008).
Conflict Management and Peace Science
474
Firms may also have economic and political incentives to mischaracterize real trade
values, which could result in inaccurate national reports. When economic and politi-
cal incentives increase the desirability for firms to mischaracterize point of origin,
destination, trade values, etc., national and dyadic trade figures may be inaccurate.
As mentioned, firms may be concerned with being publicly linked to repressive
regimes. They may also seek to avoid high tax rates, quotas, and other trade restric-
tions. If firms consistently engage in transfer pricing, where trade is underreported
with states that have high tax rates and underreported with those that have low rates,
there will be a systematic bias in the patterns of trade reported.
Understanding when and why states or firms might manipulate trade figures is
interesting for a number of reasons—the most important being that inaccurate fig-
ures undermine our ability to understand trading relationships and their causes and
consequences in international politics. Clearly, the problems of inaccurate reporting
are difficult to overcome, but understanding when and where problems may exist is
a necessary first step to solving them.
One way to assess the accuracy of a given nation’s trade figures is to compare
these with its trading partners’ figures. This may reveal whether one state tends to
over or underreport trade, relative to its partners, or whether it fails to report trade
with a given state or category of states (e.g. Israeli trade with its neighbors). This
certainly does not solve all of our problems, but could highlight the challenges fac-
ing researchers using dyadic trade figures. This brings us to the second category of
trade problems—inconsistent trade reports between partners.
Discrepant Partner Reports
Many of the strategies that scholars develop to address problems of underreporting
or nonreporting trade are based upon the assumption that two countries tend to
provide similar depictions of the same trading relationship. However, trade experts
have long recognized the problem of discrepant partner reports (see Morgenstern,
1963). In addition, we should not expect an importer and an exporter to report
identical trade values, since the importing state typically reports the value of goods
traded in c.i.f. (cost, insurance, and freight) values, which includes transportation
costs, while the exporter reports them in f.o.b. (free on board) values, which excludes
transport costs.
The IMF outlines several reasons for differences in partner reports. These in-
clude: differences in classification schemes, recording times, valuations, coverage,
and processing errors as causes of inconsistent reports (IMF, 2007). Differences are
also tied to “inconsistent currency version, evasion, anti-evasion procedures, values
not known at the time of consignment, and differences in treatment of particular
costs or procedures for assessment” (IMF, 2008). Among the things the IMF (2008)
describes as coverage problems are “shipments to and from free-trade zones and
bonded warehouses, exclusion of military and other confidential items and govern-
ment goods, value thresholds for customs registration of shipments, returned goods,
and other goods missed by customs (or surveys).”
The IMF estimates that the difference in values between the importer and exporter
is approximately ten percent, as a result of importers reporting trade in c.i.f. values,
Barbieri, Keshk & Pollins: Trading Data
475
while exporters typically use f.o.b. values.6 F.o.b. values include the price of goods at the
port of origin and include domestic transportation fees. The c.i.f. value, on the other
hand, is assessed at the destination port and includes both the f.o.b. price components
and the costs of international transportation and insurance. At times, the IMF adjusts
trade values up or down by ten percent in order to transform one state’s reported
c.i.f. values into f.o.b. values or vice versa or when the IMF decides to use one state’s
reported trade values to substitute for its partner’s missing trade report.
The IMF’s adjustment practices and their ten percent rule pose problems for
scholars interested in the degree to which trade partner reports coincide. First,
scholars are typically unaware if and when the IMF adjusts data before publishing
it. They may assume two states provide consistent dyadic trade reports, when the
reports are simply a product of IMF adjustments. Certainly, one would expect two
states’ trade reports to be highly correlated if they are based upon the same data
that come from one state.
The IMF’s ten percent rule may be particularly problematic for conflict researchers,
since it could grossly undervalue transportation fees and mask trade ties that exist
during conflicts. Using an exporter’s trade report to replace a missing importer’s
report assumes that the price of goods remains relatively constant from the time
goods leave the port of origin until they reach their destination, but delivery delays
and other factors could increase the gap between the exporter’s and importer’s price.
Periods of intense conflict and instability within a state, region, or relationship are
likely to result in delivery delays, shortages, increases in insurance premiums, and
higher transportation costs. Data are also less likely to be systematically collected
or reported during periods of intense internal or external conflict. There may be
political reasons why states or firms want to hide or distort trade ties or delay the
release of trade reports.
In addition, new security measures instituted after 11 September 2001 have placed
added burdens on shippers.7 The IMF’s ten percent figure may be less accurate today
than in the past and less accurate in conflict zones than peaceful areas. The IMF’s
decision rule cleanses trade data of important information for conflict researchers
and may mask the realities we hope to investigate. For example, variations in insur-
ance and shipping costs across space and time could provide valuable information
about changing political and economic conditions, rising tensions, and perceived
risk in international business.
The question is how often does one state report dyadic trade when its partner
does not? The problem here is that the IMF does not provide adequate informa-
tion to allow us to know when and how often the IMF uses the ten percent rule. We
only know when it does not happen—when we have a dyad where only one state
reports trade in the relationship. We found that in at least 14% of our unadjusted
IMF observations, only one of two states in the relationship reports trade. In the
6 The following countries report their imports in f.o.b.: Australia, Bermuda, Brazil, Canada,
Czechoslovakia, Dominican Republic, Mexico, Papua New Guinea, Paraguay, Peru, Romania,
South Africa, and Zimbabwe. The IMF adjusts these figures and reports them in c.i.f.
7 See Barbieri and Pathak (2007) for a discussion of post-9/11 trade security regulations.
Conflict Management and Peace Science
476
majority of cases—64% of observations—either both states report trade or one
state reports trade and the IMF uses that figure to replace the other state’s missing
report. In approximately 22% of our cases, both states in a dyad do not report trade
in their relationship.
Another important question for researchers is how great the variation is in most
partner reports and what impact this might have on our research. Once again, we
confront the problem of not knowing the impact of IMF adjustments on our figures.
Yet, we measured the magnitude of inconsistencies in partner reports, by dividing
the larger value of reported trade in a given dyadic flow by the smaller reported
trade value. We found that in approximately 73% of cases, the importer and exporter
reports differed by more than ten percent. The differences were ten percent or less
in approximately 26% of partner reports. As mentioned above, some of the similar
values are a function of IMF data adjustments, so we are likely to underestimate
the differences in reports.
Scholars must consider how they might want to portray a trading relationship when
dramatic differences exist in partner reports. There may be times when we want to
understand why differences exist between state reports. In reality, the dyad has one
trading relationship that we want to depict. Our desire to create one operational measure
of dyadic trade affects the decision rules we adopt to produce our data set.
Missing Trade Reports
One of the most difficult challenges occurs when one state fails to report its trade
with another. This produces what we typically describe as “missing data.” As men-
tioned, approximately 14% of all IMF observations are situations in which there
are no trade values reported for one side of a dyadic relationship, and there are no
reports for either state in the dyad in 22% of IMF cases. Researchers must deal with
the question of whether we should assume missing data indicate an absence of trade
or whether the missing values conform to a particular trend in the data series. Given
that we know there are instances in which one state reports trade and the other does
not, we know missing data does not always mean trade is absent.
If we know trade does not exist, we could replace missing data with zero values.
However, it is difficult to conclude that no trade exists between two states, unless
we dig deeper into why the data might be missing. In the past, the IMF frequently
reported zero trade between states when the value of dyadic trade fell below a cer-
tain minimum threshold. The reported value was available in print versions of the
IMF’s trade statistics and from other sources. Yet, most scholars opted not to look
beyond the zero values; some went even further and assumed missing data indicated
zero trade. In these and other cases, we opted to collect data, rather than generate it
artificially. Fortunately, the IMF has altered its practice of replacing minimum values
with zero values in its electronic data releases. Yet, many scholars continue to make
assumptions about missing data that could threaten the reliability of our data sets.
For example, one must consider how to deal with gaps in a data series. Suppose
two states appear to have an upward trend in their trade over time but the data
series have several years of missing data. Should we simply assume the missing years
conform to the same trend and interpolate to create data to replace missing data?
Barbieri, Keshk & Pollins: Trading Data
477
Suppose these two states stopped trading during those years or stopped reporting it,
because something usual happened. The states may have even fought a war; one may
have invaded and occupied the other; and they may have resumed trade and even
increased it afterwards. Data may be missing for many reasons—some of them tied to
conflict. If we simply manufacture data based upon flawed assumptions, we threaten
our ability to make accurate inferences about trade and conflict relationships.
Clearly, conflict scholars must make a number of decisions about how to handle
missing and other data problems. In creating the Correlates of War Trade Data Set,
we considered the issues described above and developed coding decisions designed
to minimize measurement errors. We recognize that any large data set is prone to
errors, but we believe we have gone further than most in reducing errors that might
be particularly problematic for conflict scholars.
The Correlates of War Bilateral Trade Data
The COW Trade Data Set, Version 2.0, builds upon several of the authors’ earlier proj-
ects, including BKP Version 1.0 (see Barbieri et al., 2003) and Barbieri’s International
Trade Data Base, Version 1.0 (Barbieri, 2002, Appendix A).8
The COW Trade Data Set
includes dyadic and national trade figures for state system members (COW Project,
2008) for the period 1870–2006. The set includes three files: (1) dyadic trade statistics;
(2) national trade statistics; and (3) a codebook that describes the variables and details
about the collection procedures. The majority of the post-WWII data for the COW
Trade Data Set were obtained from the IMF’s Direction of Trade Statistics (DOTS)
(IMF CD-ROM, 2007), while the pre-WWII data are from Barbieri (2002).
Today, the IMF is the most commonly used source of trade data for scholars,
so we focus our attention primarily on these data. The IMF obtains data from its
members. As its membership has expanded over time, data from previously excluded
countries and regions (e.g. Eastern Europe in the Cold War) have been updated.
This provides us with a better picture of trading relationships than we previously
had. The quality and precision of the IMF data have also improved over time. The
IMF’s monthly data releases include corrections to previous releases (e.g. typographi-
cal, decimal positioning, dropped numbers, member updates). The data errors, at
times, result in trade figures that differ from their correct value by a factor of one
thousand or more.
Over time, the IMF has employed a variety of decision rules to handle data
problems. Unfortunately, the IMF does not make it clear when it applies its rules.
For example, it replaces some missing data with partner reports, but not others. As
we discuss below, some of the coding rules may be problematic for conflict scholars.
Moreover, moving from IMF electronic data releases to a user-friendly dyadic format
requires additional coding decisions and considerable work.
To construct the COW Trade Data Set, we first constructed a data matrix of IMF
partner pairs, for the period 1948–2006, using the IMF’s DOTS CD-ROM (2006, 2007).
The IMF partner list includes state and non-state actors and aggregations of states by
8 BKP Version 1.0 used the IMF’s DOTS from ICPSR Study No. 7628, Computer File 1996.
See the IMF (1993, 2007, 2008) for information about their trade data.
Conflict Management and Peace Science
478
region, association, level of development, and other groups. We narrowed our focus to
dyads composed of state partners and made changes to the IMF state list and dates, so
it would match the COW state system list and dates. After producing the IMF trade
data matrix, we applied a series of coding rules to move to the next phase of our data
construction process. To understand our strategy for constructing the dyadic trade data,
it helps to consider an example. Figure 1 depicts a simplified version of the trading
relationship between the US and China.
In the US–China dyad and any other bilateral trading relationship, goods flow in
two directions (e.g. East–West, West–East, etc.). The exporter and importer report the
value of the goods traded. Thus, we have two elements to consider in our data. One
is the direction of the trade flow and the other is the source of the report (importer
versus exporter). Imagine a vessel docked in a Chinese port, loaded with containers
of Chinese goods destined for the US market. When the ship leaves its dock, China
records the value of these goods as Chinese exports to the US; when they arrive
at a US port, US Customs Officials report the value of these goods as US imports
from China. At the end of each year, the US and China each provide the IMF with
the aggregate value of all goods flowing between their two states.
In Figure 1, we see two directional trade flows, Flow 1 and Flow 2. For each, we
have an importer and an exporter report. This gives us four trade values: (1) Flow
1, importer report; (2) Flow 1, exporter report; (3) Flow 2, importer report; and (4)
Flow 2, exporter report. Again, this is clearly a simplification of the process, but it
should illuminate the issues we confront.
Scholars might reasonably conclude that the importer and exporter trade reports
would be roughly equal and that we could rely upon either state for our trade data.
Yet, partner reports are not always equal, as we discussed above. Moreover, we
should not expect them to be identical. As mentioned above, most states report
imports in c.i.f. values (i.e. cost, insurance, freight) and exports in f.o.b. values (i.e.
free on board); c.i.f. values should be larger than f.o.b. values. For example, if China
reports exports of one million dollars’ worth of goods to the US, using f.o.b. figures,
the US, as importer, should add shipping and insurance costs to the million dollar
figure to produce the c.i.f. import value. The US’s import value for Flow 1 should
be higher than China’s export report, all else being equal.
Figure 1. Simplified Version of the Dyadic Trading Relationship between the US and China
Barbieri, Keshk & Pollins: Trading Data
479
Ideally, a researcher would have data from the importer and exporter for each
directional flow (Flow 1 and Flow 2), but this ideal is not always realized. Even if
four data points are available for a given relationships, it may not be optimal to use
all of them. When possible, we relied upon the importer’s trade values, rather than
the exporter’s figures, because states have a greater tendency to monitor goods com-
ing into their country, rather than leaving it. For example, states tend to tax imports
more often than exports, and they tend to be more concerned that guns will flow
into, rather than out of, their country. By using the importer’s report to characterize
each directional flow, our dyad consists of information from both states. This should
help us reduce, although not eliminate, measurement error that could arise if one
state has a systematic bias in reporting.
In constructing the COW Trade Data Set, we started with a list of directed dyads
and used data from the importing state (the US, above) to characterize Flow 1 and
the importing state (China) to characterize Flow 2. We used the IMF data to fill in
all import cells for State 1, since State 2 is always the exporter in the directed dyad
structure.9 When one of the importer’s reports was missing, we used the exporter’s
trade report to replace missing and zero values. Within the directed dyadic format,
we really have only Flow 1 listed, along with the importer and exporter values,
since the US–China and China–US are considered two different dyads. These rows
were merged when we moved to the non-directed dyad format, so we had only the
China–US dyad and Flow 1 and Flow 2, where the US and China alternated posi-
tions as importer and exporter.
Once we had filled in all possible data cells using current IMF data, we turned
to alternative data sources to replace missing values. We started with the BKP data
set, which includes information from earlier IMF data tapes (1996). Next, we used
Barbieri’s trade data; this set includes IMF, non-IMF, and pre-WWII data. In several
special cases (Taiwan, Belgium–Luxembourg, and China), we used additional sources
or made adjustments to aggregate or disaggregated trade figures.
In the case of Taiwan, the IMF does not provide trade data, so we compiled
data from the Republic of China’s government websites. In the case of Belgium
and Luxembourg, the IMF reported one aggregate value for these states until
1996, after which time data are separated by state. We calculated disaggregated
trade data for the pre-1996 period, based on the relative size of Luxembourg
and Belgium’s GDPs. To do this, we obtained annual GDP data for Belgium
and Luxembourg (World Bank, 2005) and generated an annual ratio value of
the smaller to higher GDP values (i.e. Luxembourg to Belgium). We multiplied
Luxembourg’s dyadic trade figures by this ratio value and multiplied Belgium’s
trade figures by one minus the ratio value. While not an ideal solution, we rec-
ommend using the disaggregated values; but we also provide the IMF’s original
figures for interested scholars. 10
9In the COW Trade Data Set, states are arranged according to their COW country codes
(CCode), where CCode1 is the state with the lower value and CCode 2 is the higher value.
10Another strategy would be using post-1996 dyadic trade figures for these two states to
generate dyad specific ratios that are applied to the pre-1996 figures.
Conflict Management and Peace Science
480
For China, we confront the opposite problem—disaggregated, rather than aggregated
trade reports. Here, the IMF continues to report separate trade values for China, Macao,
and Hong Kong despite the fact that these areas were unified after 1998. Since this
might seriously undervalue China’s trade, we added Macao and Hong Kong to China’s
reported trade values after 1998. The average difference between China’s dyadic trade
values before and after Macao and Hong Kong are added is approximately 27% for
the period 1999–2006. Once again, we include the IMF’s original China figures in a
separate column.
Using all of the decision rules described above, the COW Trade Data Set has
approximately 19% more trade values than the original IMF data base for the
post-WWII period. We also added data for the pre-WWII period (1870–1947). We
summarize our data sources and procedures in Table 1. We still have a great deal of
missing data, which we code as missing. Our decision to code missing values as such
differs from scholars who assume missing data signifies an absence of trade between
states or those who assume that dyadic trade continues according to a linear trend,
even when values are missing, and that we could apply this trend to generate values
for missing data. We believe these assumptions and several others being adopted in
our research community are often faulty and could threaten our quest for greater
scientific understanding of the causes and consequences of trade. Next, we turn to
a discussion of some of these decision rules for handling missing data.
Decision Rules for Missing Data
Scholars have devised strategies to handle the missing data problem. We believe
some of these are more problematic than others for conflict researchers. We focus,
in particular, on the decision rules adopted by Gleditsch (2002), since his data are
frequently employed in the field. Oneal and Russett’s (1997) data sets are also widely
used and suffer from many of the same problems. The most recent versions of the
data sets discussed here contain far fewer problems than earlier versions; unfortu-
nately, most published studies are based on the earlier data sets.11
Gleditsch’s (2002) database provides trade figures for imports and exports and relies
primarily on IMF data. Gleditsch first replaces missing or zero trade values with IMF
export data and then uses the World Export Data set (Faber and Nierop, 1989). At this
point, Gleditsch adopts several coding decisions that we find problematic. First, he as-
sumes that states within a dyad have balanced trade—that the value of goods flowing
in one direction is roughly equal to those flowing in another direction (Gleditsch, 2002:
718).12 Based on this balanced trade assumptions, he replaces missing trade figures with
the value of the opposite flow. For Figure 1, this means that we may assume the value of
11Among the problems we identified were that Gleditsch (2002): (1) replaced Taiwan’s data, all of
which were missing from the IMF set, with zero values, suggesting Taiwan had zero trade with the
US and other major partners; (2) mistook South Vietnamese trade for North Vietnamese trade;
(3) had approximately 3000 observations that were off by a year; (4) approximately 4000 illogical
trade value for the United Kingdom; and (5) had incorrect values for his lags and leads, when
compared with the actual IMF reports. Gleditsch’s Version 4.1 corrects problems 1, 2, and 4.
12 Our comparisons are based on Gleditsch exptradegdpv4.1, uddtrade_cow.asc.
Barbieri, Keshk & Pollins: Trading Data
481
US imports from China are identical to China’s imports from the US (i.e. Flow 1 is equal
to Flow 2, in Figure 1) and use the value of China’s imports from the US to replace the
US’s missing import report. Most scholars realize that assuming China–US trade is bal-
anced is problematic, yet, they may be employing data generated with this decision rule.
Approximately 6% of Gleditsch’s data were generated with his balanced trade
assumption. This is particularly problematic for conflict scholars who seek to identify
and understand the sources of tension in interstate relationships. Trade imbalances
are one such source of conflict. By generating data with the problematic balanced
trade rule, scholars could erase vital information about why tensions may emerge in
a given trading relationship. Ironically, conflict scholars are the only ones we could
identify who have used the balanced trade assumption to generate data.
Table 1. Source Codes and Procedures for COW Trade Data Set 2.0, 1870–2006
Indicator
variable Share of observations Share of observations
code Procedure Source Flow 1 (% share) Flow 2 (% share)
1 Barbieri Trade Data See Barbieri 19,389 (2.99) 19,268 (2.97)
(1870–1947) (2002; appendix A)
2 Filled in c.i.f. DOTS c.i.f. import 431,254 (66.49) 417,096 (64.30)
import value value (IMF, 2007)
3 Filled in missing DOTS export value 34,614 (5.34) 49,471 (7.63)
import value (IMF, 2007)
4 Filled in zero DOTS export value 14,612 (2.25) 20,801 (3.21)
import value (IMF, 2007)
5 Filled in missing DOTS c.i.f. importer 1,443 (0.22) 1,037 (0.16)
import value value (IMF, 1996)
6 Filled in zero DOTS c.i.f. importer 4,403 (0.68) 2,729 (0.42)
import value value (IMF, 1996)
7 Filled in missing DOTS export value 371 (0.06) 418 (0.06)
import value (IMF, 1996)
8 Filled in zero DOTS export value 2,915 (0.45) 616 (0.09)
import value (IMF, 1996)
9 Filled in missing Barbieri value 274 (0.04) 423 (0.07)
import value
10 Filled in zero Barbieri value 1,113 (0.17) 1,320 (0.20)
import value
11 Belgium– DOTS c.i.f. import 9,610 (1.48) 3,071 (0.47)
Luxembourg value (IMF, 2007)
data 1948–1996
12 Taiwan data Taiwan mixed sources1 1,150 (0.18) 994 (0.15)
(1952–1988)
13 Taiwan data Taiwan Government 3,311 (0.51) 3,251 (0.50)
(1989–2006)
–9 Missing – 124,178 (19.14) 128,142 (19.76)
1 Taiwan data for 1951–1969 are from United Nations; national data for 1971–1972 are from the APEC
Study Center (2008); data for 1973–1988 are from the ROC, Council of Economic Planning and
Development (2002, 2004); and data for 1989–2006 are from the ROC, Bureau of Foreign Trade (2008).
Conflict Management and Peace Science
482
We examined the empirical accuracy of the balanced trade assumption, using
dyads with available IMF data for the importer and exporter for two-way trade.
To do this, we created a ratio variable that measures the larger to the smaller
directional trade flow value. Our findings appear in Table 2. As we see, the major-
ity of trade relationships are unbalanced. Balanced trade is the exception, rather
than the rule.
We explore this and other problematic decision rules further by comparing the
COW trade data based on IMF reports with the trade data Gleditsch generated,
using several decision rules we find to be problematic. Table 3 reveals the dramatic
differences between the IMF-based COW Trade figures and Gleditsch’s computer
generated data. First, Gleditsch generates data for Botswana’s imports from the
US for the period 1975–1983, using the questionable balanced trade assumption
(Gleditsch, 2002: 719). If we look at the period 1975–1983, we see that US imports
from Botswana far exceed Botswana’s imports from the US and that the balanced
trade assumption is inaccurate.
Next, Gleditsch generates additional values for the period 1984–2000, using a
second decision rule. He assumes uninterrupted trade—that trade conforms to
a particular trend that continues during years with missing data. Gleditsch uses
this common, but problematic, assumption to interpolate and extrapolate data for
Botswana’s imports from the US for the period 1984–2000. Some of these values
are generated with data produced with the balanced trade rule discussed above;
others use IMF data. Finally, Gleditsch assumes that other years with missing data
indicate an absence of trade between the US and Botswana; and he substitutes
these values with zeros.
Many scholars assume missing trade data indicates an absence of trade and adopt
the decision rule that missing data could be replaced with zeros. Yet, the “missing
trade equals zero” trade assumption lacks empirical support and may be problem-
atic for conflict scholars. While some states may engage in no trade, globalization
has linked many states that previously had little or no contact. It is also difficult
to assume that reports of zero trade are accurate. In the past, the IMF’s electronic
data files reported zero trade when trade values fell below a minimum threshold.
The actual trade values were contained in the IMF paper publications and today
appear in both paper and electronic data releases. Yet, today, some states continue
to report zero trade in their national accounts when values fall below a certain
threshold (UN, 2008).
From a statistical point of view, employing Gleditsch’s and other faulty decision
rules to generate data to replace missing observations could produce serious mea-
surement error. If scholars use these figures as dependent variables in their analysis
Table 2. The Accuracy of the Balanced Trade Assumption
Balanced trade Roughly balanced trade Unbalanced trade No trade
Flow 1 = Flow 2 Difference in flows < = 10% Difference in flows >10% Both flows = 0
1,763 9,129 178,556 92,731
Barbieri, Keshk & Pollins: Trading Data
483
the result will be inefficient estimates; if data are used as independent variables, the
estimates will be biased and inconsistent, even as N gets very large (Gujarati, 2003).
For these and the other reasons discussed above, we believe it is best to treat missing
data as such and to seek alternative sources for trade data and solutions to the miss-
ing data problem. We believe some of the faulty decision rules described here may
be particularly problematic for scholars who are interested in the periods associated
Table 3. Comparing Official and Manufactured Data for US–Botswana Trade
COW Trade Data Gleditsch Trade Data
Year US–Botswana Botswana–US US–Botswana Botswana–US
1966 0 – 0 0
1967 0 – 0 0
1968 0 – 0 0
1969 0 – 0 0
1970 0 – 0 0
1971 1 – 0 0
1972 2 2 0 0
1973 2 2 0 0
1974 3 2 3 3
1975 21 2 21 21
1976 54 3 48 48
1977 50.4 2.1 45 45
1978 63.8 2.2 58 58
1979 60.8 5.7 55 55
1980 88.7 6.1 80 80
1981 134.7 6.5 122 122
1982 18.8 5.1 17 17
1983 44.2 4.1 40 40
1984 58.2 18.5 39 40
1985 29.5 16.1 38 40
1986 2.6 20 36 40
1987 7 28.8 35 41
1988 9 41.3 34 41
1989 17.07 30.1 33 41
1990 14.3 19.1 32 41
1991 13.5 30.8 31 41
1992 12.4 46.6 30 42
1993 8.6 24.7 29 42
1994 13.8 22.7 28 42
1995 21.7 35.7 27 42
1996 27.5 28.9 26 42
1997 25.1 – 25 43
1998 20.4 – 20 36
1999 18.3 – 18 35
2000 42 – 42 33
Conflict Management and Peace Science
484
with intense conflicts and hope to understand how certain independent variables
affect and are affected by the onset, duration, and resolution of conflict. If scholars
generate data artificially with no consideration of the context of the missing data, par-
ticularly in relation to conflict, they are essentially saying that conflict is irrelevant for
understanding the trade patterns or that trade patterns exist independent of conflict.
Ironically, some scholars use such trade data to investigate the relationship between
trade and conflict, when their assumptions often suggest no relationship exists.
As mentioned, we believe trade data are more likely to be missing during periods
of intense conflict or in highly conflictual relationships. Given the importance of
having accurate trade data for periods associated with intense conflict, one should
strive to develop accurate assumptions about missing data and should seek mul-
tiple sources for data, rather than limiting their search. It is important to consider
how accurate the trade data are for analyzing periods of conflict and whether data
are more likely to be missing during these periods or in conflictual relationships. If
such a pattern exists and data are being manufactured with little or no attention to
conflict, the consequences could be quite serious.
Trade and Conflict
One important question is whether the problems we outline and alternative strate-
gies to address them may produce to different empirical findings. It is difficult to
answer this question definitively. First, trade data are used for a variety of purposes
and the impact of artificially produced data might vary, depending upon one’s total
sample size and the portion of the sample that relies upon manufactured data. For
example, someone might want to examine one dyad over time and might not realize
that the majority of their time series data were produced with problematic assump-
tions and not obtained from actual trade reports. Imagine a case in which states were
experiencing an upward trend in trade over time, but entered a period of intense
conflict in which they failed to report trade. Suppose further that this trade returns
to the pre-conflict level once reporting begins again. If we replace the missing trade
data using a linear trend, without considering the context in which data are miss-
ing, we are likely to produce inaccurate reports. Unfortunately, we often employ
data without fully understanding how it was generated or its limitations. While we
may not know how alternative data strategies impact our findings, we are able to
assess how frequently data are missing during periods of conflict and whether some
scholars tend to rely upon questionable data for these critical observations. We also
examine whether dyads with missing data are different in their conflict propensities
than dyads with available data. This test also reveals whether it is reasonable to use
standard assumptions to replace missing data points. To explore the relationship
between conflict and trade data availability, we perform some preliminary analyses.
First, we examine the extent to which data are missing during conflict periods.
Next, we examine whether dyads with missing data are more or less likely to en-
gage in conflict. In the first analysis, we compare the COW Trade Data Set with
Gleditsch’s trade data and examine the patterns of missing and manufactured
data. We use the COW Militarized Interstate Dispute (MID) data to divide our
sample into years in which a dyad experienced a conflict and those in which it did
Barbieri, Keshk & Pollins: Trading Data
485
not (Jones et al., 1996; Ghosn, et al., 2004; and Ghosn and Bennett, 2003). 13 We
then examine the frequency with which trade data are missing for conflict years;
the availability of official trade reports; and the extent to which some scholars rely
upon trade data generated with the problematic assumptions we critiqued above.
Our results appear in Table 4.
We see that for the COW Trade Data Set, trade data are missing for approximately
6% of the MID observations and about 11% of non-MID observations. For the
Gleditsch data set, more than 20% of the MID observations and 44% of non-MID
observations consist of computer generated trade values. As mentioned, if data are
missing as a result of conflict, it might be particularly problematic to apply some
of the assumptions we described above. The fact that Gleditsch has made assump-
tions about the trade values for approximately 20% of all MIDS could affect any
conflict analysis that relies upon this data set. The question of how much the results
are affected is less clear, since official statistics are not always available for these
observations. Yet, it is still possible to evaluate whether dyads with missing data are
somehow different than those without missing data.
Next, we examine whether dyads with missing data have a different propensity
toward conflict than those without missing data.14 If they do, the way we handle
missing data may be even more important than some believe. To investigate this is-
sue, we perform a set of statistical analyses that examines the probability of conflict
in dyads with missing data and compare this to cases with available data. We also
compare dyads with reports of zero trade to those with positive trade values, since
replacing missing trade data with zero values is a common way of addressing the
missing data problem.
Our goal here is not to resolve the question of which model is best for analyzing
the trade–conflict relationship. Instead, we simply wish to make comparisons about
the conflict propensity of different dyad types, based upon trade data availability,
rather than trade level. Thus, our measure of trade in a basic trade–conflict analysis
related to trade data type and not level.
Our model includes the few variables present in most trade–conflict analyses—
conflict, trade, GDP, and geographic proximity—and includes all non-directed dyads
13 We utilize EUGene Version 3.023 to generate the MID data and the DISTANCE data
described below (Bennett and Stam, 2000, 2007).
14 Future analysis will consider the simultaneous nature of the trade–conflict relationship.
Table 4. Trade Data Availability and MIDs
COW Trade Gleditsch Trade
Missing Not missing Artificial data Not missing
MIDs 145 2,176 463 1,798
6.27% 93.73% 20.48% 79.52%
No MIDs 61,601 487,457 231,735 293,306
11.22% 88.78% 44.14% 55.86%
Conflict Management and Peace Science
486
for the years 1948–2001. Our dependent variable is conflict and we measure that
using three different types of MIDs: all MIDs, FATAL MIDs (MIDs with at least 1
fatality), and WARS (MIDs with more than 1,000 battle deaths). We also examine
MID ONSET and MID INVOLVEMENT. MID ONSET is coded 1 for the first year
of the dispute and 0 otherwise. MID INVOLVEMENT is coded 1 for each year of
the conflict and coded 0 during years of peace.
We use three different data sets in our analysis: (1) the original IMF dyadic data
without the COW adjustments (IMF); (2) the COW Trade Data Set (COW); and (3)
the Gleditsch Trade Data Set (Gleditsch, 2002). For the central variable of interest,
trade data, we create two dichotomous variables per data set. The first is MISSING
DATA and is set to 1 if the trade values are missing and 0 otherwise. For Gleditsch, we
set his computer generated values to 1, since we assume these values were originally
missing.15 For the second variable, ZERO TRADE, the dummy variable is coded as
1 if dyadic trade equals zero and is coded 0 for non-zero trade values. For Gleditsch,
the dummy variable includes IMF zeros and his assumed zeros.
Our model includes the most common control variables found in trade–conflict
analysis: distance and GDP. DISTANCE is measured as capital to capital distance
with adjustments for contiguity.16 GDP data come primarily from the World Bank
(2008) and were converted to real values using a conversion factor index (Sahr, 2008).
DISTANCE is assumed to be negatively related to conflict, while GDP is assumed
to be positively associated with conflict. These variables were both statistically
significant and had the expected signs. For ease of presentation, we do not include
the control variables in Table 5, where we present the results of our analysis.
We see that for the IMF and COW Trade Data Sets, dyads with MISSING DATA
tend to be more likely to engage in MIDs and WARs and less likely to experience
FATAL MIDs. For Gleditsch, dyads with MISSING DATA are less likely to experience
15 For Gleditsch, this holds when variables giabo & gibao are greater than 2.
16 Gleditsch combines trade and GDP into one ratio variable, which was not appropriate for
this analysis.
Table 5. Conflict Analysis for Dyads with Missing, Zero, and Available Data
MIDs Fatal MIDs Wars
Onset Involvement Onset Involvement Onset Involvement
IMF Missing 0.059*** 0.121*** −0.286*** −0.283*** 0.271*** 0.409***
Zero −0.074** −0.126*** −0.104*** −0.105*** 0.199** −0.109*
COW Missing 0.054* 0.072*** −0.022*** −0.021*** 0.336*** 0.228***
Zero −0.023 0.027 −0.211*** −0.210*** 0.194*** 0.206***
Gleditsch Missing −0.141*** −0.101*** −2.765*** −2.597*** .0297*** 0.293***
Zero −0.127*** 0.157*** −3.261*** 2.987*** −0.051*** −0.232***
*p < .10; **p < .05; ***p < .01.
Barbieri, Keshk & Pollins: Trading Data
487
MIDs and FATAL MIDs, but more likely to experience WARS. The findings are
statistically significant, which means that there are significant differences in the con-
flict propensity of dyads with missing trade data and those with available data. The
differences are not consistent across conflict types, but we believe the results provide
sufficient evidence to question the usual assumptions about trade values when data
are missing, particularly if we hope to use these data to analyze conflict.
Our results comparing the conflict propensity of dyads with ZERO TRADE and
those with positive trade also vary across data sets. For the IMF, dyads with ZERO
TRADE are less likely to engage in conflict. The only exception is WAR ONSET,
where dyads with no trade are more likely to witness an outbreak of war. For COW
Trade, dyads with ZERO TRADE are less likely to engage in MIDs, but more likely
to see the outbreak of war. Finally, Gleditsch’s ZERO TRADE dyads are negatively
associated with MID and FATAL MID ONSET and positively associated with MID
and FATAL MID INVOVLEMENT. However, ZERO TRADE is negatively associ-
ated with the predicted probability of both WAR ONSET and INVOLVEMENT.
Given the variation in results for cases with ZERO TRADE, it seems unwise to
make assumptions about when ZERO TRADE values should be used in place of
missing data and when reports of no trade are accurate.
Overall, the results substantiate the concerns raised in this paper: that replacing
missing trade data with questionable values is problematic and may affect our re-
sults, particularly when we do not explore the reasons the data are missing. Missing
trade data tend to be associated with significant increases or decreases in different
types of conflict. We would want to know if missing data coincided with conflict in
a relationship. In addition, conflict might be short or long-term; it may or may not
affect trade and the reporting of trade; and trade may or may not vary over time.
Furthermore, we have only discussed direct trade and conflict and have not con-
sidered how conflict might affect trade relations with third parties (e.g. allies and
adversaries of each state in the conflict).
If we conclude that missing trade data indicates certain values of trade, then we
assume away other conditions. The assumptions we make could determine the results
we obtain and produce tautologies. For example, if we assume dyadic trade conforms
to a particular trend and use that to predict conflict, we might produce values that
determine our findings about trade and conflict. The assumption that there is a
trend in the data will impact the results and, therefore, the results are driven by the
assumption about missing trade values. This is clearly not how we want to study the
trade–conflict relationship or many other facets of trade.
Furthermore, our results should not be used to justify filling in missing trade values,
because there are situations where missing trade data may be positively or negatively
related to conflict. Findings of a negative result between conflict and trade may be
overestimated, if we assume missing data are zero and the missing cases tend to be more
conflict prone. Findings of a positive relationship between missing trade and conflict
may be underestimated, since we do not know whether trade was going up or down or
was nonexistent before, during, and after the conflict. Thus, filling in missing trade data
using assumptions is not a solution to the underestimation or overestimation problem.
In fact, a finding that filling in missing values increases the strength of the pre-filled-in
missing results, as some contend (Gleditsch, 2002), is not a finding at all, since we have
Conflict Management and Peace Science
488
no information about why the trade values are missing. Filling in missing trade values
becomes even more problematic if conflict is impacting trade. For all these reasons,
we are dubious about current methods for filling in missing trade values.
Conclusions
This paper reveals the importance of understanding the trade data that we employ
in our research. The issue of missing data and how it impacts our analysis is one
that must be addressed with greater care than conflict researchers have done thus
far. While some have sought remedies to the problem of missing data, we believe
no one has found the optimal solution. Until that solution comes, we must consider
how missing data patterns affect the answers we receive about questions that involve
international trade.
More importantly, when data are generated with problematic decision rules, it is
difficult to have confidence in the results they generate. Our goal should be to devise
better strategies to obtain valid and reliable data. To this end, we believe that the new
Correlates of War Trade Data Set is a first step in this process. Future efforts need to
move beyond the realm of IMF trade statistics and official national reports. We must
seek sources and, when possible, build data-based measures that capture the most
accurate picture of world trade, including the actors involved, the goods they supply
and demand, the linkages and dependencies that exist across countries, the means by
which laws are circumvented, and the ways trade data are falsified or distorted.
We would prefer to see more cooperation among scholars interested in trading
relationships. When dealing with extraordinarily large numbers of observations,
the possibility of machine-induced coding errors is greater than with smaller data
sets where cross-checking of every individual observation is practical. We hope that
researchers employing the new Correlates of War Trade Data Set will report any
problems they discover to its creators. Any trade data set will benefit from input
from a larger community of scholars. We have identified and begun working on
several issue areas that we believe will augment our understanding of global trade
relations and their myriad effects upon the lives of nations.
As mentioned, our goal is to provide an accurate depiction of trading relationships.
Yet, as is true of any quantitative indicators, we must recognize that our measures
may paint a picture that stands in sharp contrast from reality. We must understand the
shortcomings of certain research strategies, as we seek solutions to our data problems
and interpret empirical findings tied to different data sets. Future research should be
better able to produce valid, reproducible findings if we generate more reliable data
on interstate trade. While this article and the trade data project it describes strive to
bring us one step closer to meeting our scientific objectives of properly measuring
trading relationships, it is by no means the final step.
References
APEC Study Center, City University of Hong Kong Data Bank. 2008. http://fbweb.cityu.edu.
hk/hkapec/DataBank/Taiwan-Trade-Balance.pdf.
Barbieri, Keshk & Pollins: Trading Data
489
Barbieri, Katherine. 2002. The liberal illusion: Does trade promote peace? Ann Arbor:
University of Michigan Press.
Barbieri, Katherine, and Swapna Pathak. 2007. Free trade and terrorism. In Terrornomics,
eds Sean S. Costigan and David Gold, ch. 7. Burlington, VT: Ashgate.
Barbieri, Katherine, Omar Keshk, and Brian M. Pollins. 2003. The Correlates of War Trade
Data Set: A progress report. Paper presented at the Annual Meeting of the Peace Science
Society (International), Ann Arbor, Michigan, 13–16 November.
Bennett, D. Scott, and Allan Stam. 2000. EUGene: A conceptual manual. International
Interactions 26:179–204. Website: http://eugenesoftware.org.
Bennett, D. Scott, and Allan Stam. 2007. EUGene: Expected Utility Generation and Data
management program. Version 3.203, 29 October. http://www.eugenesoftware.org/.
Bhagwati, Jagdish N. 1964. “On the Underinvoicing of Imports.” Oxford Bulletin of Economics
and Statistics 27(4): 389–397.
Bhagwati, Jagdish N. 1967. Fiscal policies, the faking of foreign trade declarations, and the
balance of payments. Oxford Bulletin of Economics and Statistics 29(1): 61–77.
Bhagwati, Jagdish N., and T. N. Srinivasan. 1981 The evaluation of projects at world prices
under trade distortions: Quantitative restrictions, monopoly power in trade and nontraded
goods. International Economic Review 22(2): 385–399.
Correlates of War Project. 2008. State System Membership List, v2008.1. Online, http://
correlatesofwar.org.
de Wulf, Luc. 1981. Customs valuation and the faking of invoices. Economia Internazionale
34(1): 13–33.
Ely, J. Edward. 1961. Variations between U.S. and its trading partner import and export
statistics. American Statistician 15(2): 23–26.
Faber, Jan, and Tom Nierop. 1989. World export data, 1948–1983 (ICPSR no. 9116). Amsterdam:
University of Amsterdam (producers); Ann Arbor, MI: Inter-university Consortium for
Political and Social Research (distributors).
Ghosn, Faten, and Scott Bennett. 2003. Codebook for the Dyadic Militarized Interstate
Incident Data, Version 3.10. Online: http://cow2.la.psu.edu
Ghosn, Faten, Glenn Palmer, and Stuart Bremer. 2004. The MID3 Data Set, 1993–2001:
Procedures, coding rules, and description. Conflict Management and Peace Science 21:
133–154.
Ghosn, Faten, and Scott Bennett. 2003. Codebook for the Dyadic Militarized Interstate
Incident Data, Version 3.10. Online: http://cow2.la.psu.edu.
Gleditsch, Kristian S. 2002. Expanded trade and GDP data. Journal of Conflict Resolution
46: 712–24.
Gujarati, Damador N. 2003. Basic econometrics. Boston, MA: McGraw Hill.
International Monetary Fund. 1993. A guide to direction of trade statistics (http://www.imf.
org/external/pubs/cat/longres.cfm?sk=154.0).
International Monetary Fund. 2006. Direction of trade statistics: Historical database and
browser on CD-ROM.
International Monetary Fund. 2007. Direction of trade statistics on CD-ROM. Monthly
Subscription.
International Monetary Fund. 2008. Data and statistics: Direction of trade statistics
consistency between partner country data (http://www.imfstatistics.org/dot/DOTConsi.
htm).
Jones, Daniel M., Stuart A. Bremer, and J. David Singer. 1996. Militarized interstate disputes,
1816–1992: Rationale, coding rules, and empirical patterns. Conflict Management and Peace
Science 15(2): 163–213.
Conflict Management and Peace Science
490
Keshk, Omar M. G., Brian M. Pollins, and Rafael Reuveny. 2004. Trade still follows the flag:
The primacy of politics in a simultaneous model of interdependence and armed conflict.
Journal of Politics 66: 1155–1179.
Mansfield, Edward D., and Brian M. Pollins, eds. 2003. Economic interdependence and
international conflict: New perspectives on an enduring debate. Ann Arbor: University of
Michigan Press.
Morgenstern, Oskar. 1963. On the accuracy of economic observations. Princeton, NJ: Princeton
University Press.
Morgenthau, Hans J. 1948. Politics among nations: The struggle for power and peace.
New York: Knopf.
Oneal, John R., and Bruce M. Russett. 1997. The classical liberals were right: Democracy,
interdependence, and conflict, 1950–85. International Studies Quarterly 41(2): 267–294.
Pak, Simon J., and John S. Zdanowicz. 1994a. An estimate of lost US federal income tax
revenues due to over-invoiced imports and under-invoiced exports. Report 9404–01, Trade
Research Institute, Miami, FL.
Pak, Simon J., and John S. Zdanowicz. 1994b. A statistical analysis of the US merchandise trade
data base and its uses in transfer pricing compliance and enforcement. Tax Management
Transfer Pricing Report 3(1) 50–57.
Pak, Simon J., Stelios H. Zanakis, and John S. Zdanowicz. 2003. Detecting abnormal pricing
in international trade: The Greece–USA Case. Interfaces 33(2): 54–64. Linthicum, MD:
Institute for Operations Research and the Management Sciences (INFORMS).
Republic of China (ROC), Bureau of Foreign Trade Website. 2008. Data source, Directorate
General of Customs, Ministry of Finance (http://cus93.trade.gov.tw/english/FSCE/
FSC0011E.ASP).
Republic of China (ROC), Council for Economic Planning and Development. 2002, 2004.
Taiwan Statistical Data Book (http://www.cepd.gov.tw/english/file/2004DataBook.pdf?C
FID=87002&CFTOKEN=50209707).
Russett, Bruce, and John Oneal. 2001. Triangular peace: Democracy, interdependence, and
international organizations. New York: W. W. Norton.
Sahr, Robert. 2008. Inflation conversion factors for dollars 1774 to estimated 2018. Corvallis, OR:
Oregon State University. (http://oregonstate.edu/cla/polisci/faculty-research/sahr/sahr.htm).
Schneider, Gerald, Katherine Barbieri, and Nils Petter Gleditsch. 2003. Does globalization
contribute to peace? A critical survey of the theoretical and formal literature. In
Globalization and armed conflict, eds Gerald Schneider, Katherine Barbieri and Nils
Petter Gleditsch, pp. 3–29. Lanham, MD: Rowman & Littlefield.
Sheikh, Munir. 1974. Underinvoicing of imports in Pakistan. Oxford Bulletin of Economics
and Statistics 36(4): 287–296.
United Nations, Statistical Office, Department of Economic Affairs. (1951–1984). Yearbook
of international trade statistics. New York: United Nations.
United Nations. 2008. International merchandise trade statistics national compilation and
reporting practices. (UN Website http://unstats.un.org/unsd/tradereport/question.asp).
World Bank. 2005, 2008. World development indicators. On-Line Version. Washington, DC:
World Bank.
Yeats, Alexander. 1978. On the accuracy of partner country trade statistics. Oxford Bulletin
of Economics and Statistics 40(1): 341–361.
Yeats, Alexander. 1990. On the accuracy of economic observations: Do Sub-Saharan trade
statistics mean anything? World Bank Economic Review 4(2): 135–156.
Barbieri, Keshk & Pollins: Trading Data
491
KATHERINE BARBIERI is Associate Professor of Political Science at the University of
South Carolina and author of The Liberal Illusion: Does Trade Promote Peace? Her data set
on trade and conflict is available to scholars and she now serves as a co-host for the
Correlates of War Trade Data Set.
OMAR M. G. KESHK is senior lecturer in the Political Science Department and the
Undergraduate International Studies Program at the Ohio State University. He has
published in the Journal of Politics, Conflict Management and Peace Studies, and the Stata Journal.
He co-hosts the COW Trade Data Set.
BRIAN M. POLLINS is Emeritus Associate Professor of Political Science at the Ohio State
University. He now teaches and writes at the Naval Postgraduate School in Monterey, CA.